]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // vim: ts=8 sw=2 smarttab |
2 | /* | |
3 | * Ceph - scalable distributed file system | |
4 | * | |
5 | * Copyright (C) 2014 Red Hat | |
6 | * | |
7 | * This is free software; you can redistribute it and/or | |
8 | * modify it under the terms of the GNU Lesser General Public | |
9 | * License version 2.1, as published by the Free Software | |
10 | * Foundation. See file COPYING. | |
11 | * | |
12 | */ | |
13 | ||
14 | #include <unistd.h> | |
15 | #include <stdlib.h> | |
16 | #include <sys/types.h> | |
17 | #include <sys/stat.h> | |
18 | #include <fcntl.h> | |
19 | ||
31f18b77 FG |
20 | #include "include/cpp-btree/btree_set.h" |
21 | ||
7c673cae FG |
22 | #include "BlueStore.h" |
23 | #include "os/kv.h" | |
24 | #include "include/compat.h" | |
25 | #include "include/intarith.h" | |
26 | #include "include/stringify.h" | |
27 | #include "common/errno.h" | |
28 | #include "common/safe_io.h" | |
29 | #include "Allocator.h" | |
30 | #include "FreelistManager.h" | |
31 | #include "BlueFS.h" | |
32 | #include "BlueRocksEnv.h" | |
33 | #include "auth/Crypto.h" | |
34 | #include "common/EventTrace.h" | |
35 | ||
36 | #define dout_context cct | |
37 | #define dout_subsys ceph_subsys_bluestore | |
38 | ||
31f18b77 FG |
39 | using bid_t = decltype(BlueStore::Blob::id); |
40 | ||
41 | // bluestore_cache_onode | |
7c673cae | 42 | MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Onode, bluestore_onode, |
31f18b77 | 43 | bluestore_cache_onode); |
7c673cae | 44 | |
31f18b77 | 45 | // bluestore_cache_other |
7c673cae | 46 | MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Buffer, bluestore_buffer, |
31f18b77 | 47 | bluestore_cache_other); |
7c673cae | 48 | MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Extent, bluestore_extent, |
31f18b77 | 49 | bluestore_cache_other); |
7c673cae | 50 | MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Blob, bluestore_blob, |
31f18b77 | 51 | bluestore_cache_other); |
7c673cae | 52 | MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::SharedBlob, bluestore_shared_blob, |
31f18b77 FG |
53 | bluestore_cache_other); |
54 | ||
55 | // bluestore_txc | |
56 | MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::TransContext, bluestore_transcontext, | |
57 | bluestore_txc); | |
58 | ||
7c673cae FG |
59 | |
60 | // kv store prefixes | |
61 | const string PREFIX_SUPER = "S"; // field -> value | |
62 | const string PREFIX_STAT = "T"; // field -> value(int64 array) | |
63 | const string PREFIX_COLL = "C"; // collection name -> cnode_t | |
64 | const string PREFIX_OBJ = "O"; // object name -> onode_t | |
65 | const string PREFIX_OMAP = "M"; // u64 + keyname -> value | |
66 | const string PREFIX_DEFERRED = "L"; // id -> deferred_transaction_t | |
67 | const string PREFIX_ALLOC = "B"; // u64 offset -> u64 length (freelist) | |
68 | const string PREFIX_SHARED_BLOB = "X"; // u64 offset -> shared_blob_t | |
69 | ||
70 | // write a label in the first block. always use this size. note that | |
71 | // bluefs makes a matching assumption about the location of its | |
72 | // superblock (always the second block of the device). | |
73 | #define BDEV_LABEL_BLOCK_SIZE 4096 | |
74 | ||
75 | // reserve: label (4k) + bluefs super (4k), which means we start at 8k. | |
76 | #define SUPER_RESERVED 8192 | |
77 | ||
78 | #define OBJECT_MAX_SIZE 0xffffffff // 32 bits | |
79 | ||
80 | ||
81 | /* | |
82 | * extent map blob encoding | |
83 | * | |
84 | * we use the low bits of the blobid field to indicate some common scenarios | |
85 | * and spanning vs local ids. See ExtentMap::{encode,decode}_some(). | |
86 | */ | |
87 | #define BLOBID_FLAG_CONTIGUOUS 0x1 // this extent starts at end of previous | |
88 | #define BLOBID_FLAG_ZEROOFFSET 0x2 // blob_offset is 0 | |
89 | #define BLOBID_FLAG_SAMELENGTH 0x4 // length matches previous extent | |
90 | #define BLOBID_FLAG_SPANNING 0x8 // has spanning blob id | |
91 | #define BLOBID_SHIFT_BITS 4 | |
92 | ||
93 | /* | |
94 | * object name key structure | |
95 | * | |
96 | * encoded u8: shard + 2^7 (so that it sorts properly) | |
97 | * encoded u64: poolid + 2^63 (so that it sorts properly) | |
98 | * encoded u32: hash (bit reversed) | |
99 | * | |
100 | * escaped string: namespace | |
101 | * | |
102 | * escaped string: key or object name | |
103 | * 1 char: '<', '=', or '>'. if =, then object key == object name, and | |
104 | * we are done. otherwise, we are followed by the object name. | |
105 | * escaped string: object name (unless '=' above) | |
106 | * | |
107 | * encoded u64: snap | |
108 | * encoded u64: generation | |
109 | * 'o' | |
110 | */ | |
111 | #define ONODE_KEY_SUFFIX 'o' | |
112 | ||
113 | /* | |
114 | * extent shard key | |
115 | * | |
116 | * object prefix key | |
117 | * u32 | |
118 | * 'x' | |
119 | */ | |
120 | #define EXTENT_SHARD_KEY_SUFFIX 'x' | |
121 | ||
122 | /* | |
123 | * string encoding in the key | |
124 | * | |
125 | * The key string needs to lexicographically sort the same way that | |
126 | * ghobject_t does. We do this by escaping anything <= to '#' with # | |
127 | * plus a 2 digit hex string, and anything >= '~' with ~ plus the two | |
128 | * hex digits. | |
129 | * | |
130 | * We use ! as a terminator for strings; this works because it is < # | |
131 | * and will get escaped if it is present in the string. | |
132 | * | |
133 | */ | |
134 | template<typename S> | |
135 | static void append_escaped(const string &in, S *out) | |
136 | { | |
224ce89b WB |
137 | char hexbyte[in.length() * 3 + 1]; |
138 | char* ptr = &hexbyte[0]; | |
7c673cae FG |
139 | for (string::const_iterator i = in.begin(); i != in.end(); ++i) { |
140 | if (*i <= '#') { | |
224ce89b WB |
141 | *ptr++ = '#'; |
142 | *ptr++ = "0123456789abcdef"[(*i >> 4) & 0x0f]; | |
143 | *ptr++ = "0123456789abcdef"[*i & 0x0f]; | |
7c673cae | 144 | } else if (*i >= '~') { |
224ce89b WB |
145 | *ptr++ = '~'; |
146 | *ptr++ = "0123456789abcdef"[(*i >> 4) & 0x0f]; | |
147 | *ptr++ = "0123456789abcdef"[*i & 0x0f]; | |
7c673cae | 148 | } else { |
224ce89b | 149 | *ptr++ = *i; |
7c673cae FG |
150 | } |
151 | } | |
224ce89b WB |
152 | *ptr++ = '!'; |
153 | out->append(hexbyte, ptr - &hexbyte[0]); | |
154 | } | |
155 | ||
156 | inline unsigned h2i(char c) | |
157 | { | |
158 | if ((c >= '0') && (c <= '9')) { | |
159 | return c - 0x30; | |
160 | } else if ((c >= 'a') && (c <= 'f')) { | |
161 | return c - 'a' + 10; | |
162 | } else if ((c >= 'A') && (c <= 'F')) { | |
163 | return c - 'A' + 10; | |
164 | } else { | |
165 | return 256; // make it always larger than 255 | |
166 | } | |
7c673cae FG |
167 | } |
168 | ||
169 | static int decode_escaped(const char *p, string *out) | |
170 | { | |
224ce89b WB |
171 | char buff[256]; |
172 | char* ptr = &buff[0]; | |
173 | char* max = &buff[252]; | |
7c673cae FG |
174 | const char *orig_p = p; |
175 | while (*p && *p != '!') { | |
176 | if (*p == '#' || *p == '~') { | |
224ce89b WB |
177 | unsigned hex = 0; |
178 | p++; | |
179 | hex = h2i(*p++) << 4; | |
180 | if (hex > 255) { | |
181 | return -EINVAL; | |
182 | } | |
183 | hex |= h2i(*p++); | |
184 | if (hex > 255) { | |
185 | return -EINVAL; | |
186 | } | |
187 | *ptr++ = hex; | |
7c673cae | 188 | } else { |
224ce89b WB |
189 | *ptr++ = *p++; |
190 | } | |
191 | if (ptr > max) { | |
192 | out->append(buff, ptr-buff); | |
193 | ptr = &buff[0]; | |
7c673cae FG |
194 | } |
195 | } | |
224ce89b WB |
196 | if (ptr != buff) { |
197 | out->append(buff, ptr-buff); | |
198 | } | |
7c673cae FG |
199 | return p - orig_p; |
200 | } | |
201 | ||
202 | // some things we encode in binary (as le32 or le64); print the | |
203 | // resulting key strings nicely | |
204 | template<typename S> | |
205 | static string pretty_binary_string(const S& in) | |
206 | { | |
207 | char buf[10]; | |
208 | string out; | |
209 | out.reserve(in.length() * 3); | |
210 | enum { NONE, HEX, STRING } mode = NONE; | |
211 | unsigned from = 0, i; | |
212 | for (i=0; i < in.length(); ++i) { | |
213 | if ((in[i] < 32 || (unsigned char)in[i] > 126) || | |
214 | (mode == HEX && in.length() - i >= 4 && | |
215 | ((in[i] < 32 || (unsigned char)in[i] > 126) || | |
216 | (in[i+1] < 32 || (unsigned char)in[i+1] > 126) || | |
217 | (in[i+2] < 32 || (unsigned char)in[i+2] > 126) || | |
218 | (in[i+3] < 32 || (unsigned char)in[i+3] > 126)))) { | |
219 | if (mode == STRING) { | |
220 | out.append(in.c_str() + from, i - from); | |
221 | out.push_back('\''); | |
222 | } | |
223 | if (mode != HEX) { | |
224 | out.append("0x"); | |
225 | mode = HEX; | |
226 | } | |
227 | if (in.length() - i >= 4) { | |
228 | // print a whole u32 at once | |
229 | snprintf(buf, sizeof(buf), "%08x", | |
230 | (uint32_t)(((unsigned char)in[i] << 24) | | |
231 | ((unsigned char)in[i+1] << 16) | | |
232 | ((unsigned char)in[i+2] << 8) | | |
233 | ((unsigned char)in[i+3] << 0))); | |
234 | i += 3; | |
235 | } else { | |
236 | snprintf(buf, sizeof(buf), "%02x", (int)(unsigned char)in[i]); | |
237 | } | |
238 | out.append(buf); | |
239 | } else { | |
240 | if (mode != STRING) { | |
241 | out.push_back('\''); | |
242 | mode = STRING; | |
243 | from = i; | |
244 | } | |
245 | } | |
246 | } | |
247 | if (mode == STRING) { | |
248 | out.append(in.c_str() + from, i - from); | |
249 | out.push_back('\''); | |
250 | } | |
251 | return out; | |
252 | } | |
253 | ||
254 | template<typename T> | |
255 | static void _key_encode_shard(shard_id_t shard, T *key) | |
256 | { | |
257 | key->push_back((char)((uint8_t)shard.id + (uint8_t)0x80)); | |
258 | } | |
259 | ||
260 | static const char *_key_decode_shard(const char *key, shard_id_t *pshard) | |
261 | { | |
262 | pshard->id = (uint8_t)*key - (uint8_t)0x80; | |
263 | return key + 1; | |
264 | } | |
265 | ||
266 | static void get_coll_key_range(const coll_t& cid, int bits, | |
267 | string *temp_start, string *temp_end, | |
268 | string *start, string *end) | |
269 | { | |
270 | temp_start->clear(); | |
271 | temp_end->clear(); | |
272 | start->clear(); | |
273 | end->clear(); | |
274 | ||
275 | spg_t pgid; | |
276 | if (cid.is_pg(&pgid)) { | |
277 | _key_encode_shard(pgid.shard, start); | |
278 | *temp_start = *start; | |
279 | ||
280 | _key_encode_u64(pgid.pool() + 0x8000000000000000ull, start); | |
281 | _key_encode_u64((-2ll - pgid.pool()) + 0x8000000000000000ull, temp_start); | |
282 | ||
283 | *end = *start; | |
284 | *temp_end = *temp_start; | |
285 | ||
286 | uint32_t reverse_hash = hobject_t::_reverse_bits(pgid.ps()); | |
287 | _key_encode_u32(reverse_hash, start); | |
288 | _key_encode_u32(reverse_hash, temp_start); | |
289 | ||
290 | uint64_t end_hash = reverse_hash + (1ull << (32 - bits)); | |
291 | if (end_hash > 0xffffffffull) | |
292 | end_hash = 0xffffffffull; | |
293 | ||
294 | _key_encode_u32(end_hash, end); | |
295 | _key_encode_u32(end_hash, temp_end); | |
296 | } else { | |
297 | _key_encode_shard(shard_id_t::NO_SHARD, start); | |
298 | _key_encode_u64(-1ull + 0x8000000000000000ull, start); | |
299 | *end = *start; | |
300 | _key_encode_u32(0, start); | |
301 | _key_encode_u32(0xffffffff, end); | |
302 | ||
303 | // no separate temp section | |
304 | *temp_start = *end; | |
305 | *temp_end = *end; | |
306 | } | |
307 | } | |
308 | ||
309 | static void get_shared_blob_key(uint64_t sbid, string *key) | |
310 | { | |
311 | key->clear(); | |
312 | _key_encode_u64(sbid, key); | |
313 | } | |
314 | ||
315 | static int get_key_shared_blob(const string& key, uint64_t *sbid) | |
316 | { | |
317 | const char *p = key.c_str(); | |
318 | if (key.length() < sizeof(uint64_t)) | |
319 | return -1; | |
224ce89b | 320 | _key_decode_u64(p, sbid); |
7c673cae FG |
321 | return 0; |
322 | } | |
323 | ||
324 | template<typename S> | |
325 | static int get_key_object(const S& key, ghobject_t *oid) | |
326 | { | |
327 | int r; | |
328 | const char *p = key.c_str(); | |
329 | ||
330 | if (key.length() < 1 + 8 + 4) | |
331 | return -1; | |
332 | p = _key_decode_shard(p, &oid->shard_id); | |
333 | ||
334 | uint64_t pool; | |
335 | p = _key_decode_u64(p, &pool); | |
336 | oid->hobj.pool = pool - 0x8000000000000000ull; | |
337 | ||
338 | unsigned hash; | |
339 | p = _key_decode_u32(p, &hash); | |
340 | ||
341 | oid->hobj.set_bitwise_key_u32(hash); | |
342 | ||
343 | r = decode_escaped(p, &oid->hobj.nspace); | |
344 | if (r < 0) | |
345 | return -2; | |
346 | p += r + 1; | |
347 | ||
348 | string k; | |
349 | r = decode_escaped(p, &k); | |
350 | if (r < 0) | |
351 | return -3; | |
352 | p += r + 1; | |
353 | if (*p == '=') { | |
354 | // no key | |
355 | ++p; | |
356 | oid->hobj.oid.name = k; | |
357 | } else if (*p == '<' || *p == '>') { | |
358 | // key + name | |
359 | ++p; | |
360 | r = decode_escaped(p, &oid->hobj.oid.name); | |
361 | if (r < 0) | |
362 | return -5; | |
363 | p += r + 1; | |
364 | oid->hobj.set_key(k); | |
365 | } else { | |
366 | // malformed | |
367 | return -6; | |
368 | } | |
369 | ||
370 | p = _key_decode_u64(p, &oid->hobj.snap.val); | |
371 | p = _key_decode_u64(p, &oid->generation); | |
372 | ||
373 | if (*p != ONODE_KEY_SUFFIX) { | |
374 | return -7; | |
375 | } | |
376 | p++; | |
377 | if (*p) { | |
378 | // if we get something other than a null terminator here, | |
379 | // something goes wrong. | |
380 | return -8; | |
381 | } | |
382 | ||
383 | return 0; | |
384 | } | |
385 | ||
386 | template<typename S> | |
387 | static void get_object_key(CephContext *cct, const ghobject_t& oid, S *key) | |
388 | { | |
389 | key->clear(); | |
390 | ||
391 | size_t max_len = 1 + 8 + 4 + | |
392 | (oid.hobj.nspace.length() * 3 + 1) + | |
393 | (oid.hobj.get_key().length() * 3 + 1) + | |
394 | 1 + // for '<', '=', or '>' | |
395 | (oid.hobj.oid.name.length() * 3 + 1) + | |
396 | 8 + 8 + 1; | |
397 | key->reserve(max_len); | |
398 | ||
399 | _key_encode_shard(oid.shard_id, key); | |
400 | _key_encode_u64(oid.hobj.pool + 0x8000000000000000ull, key); | |
401 | _key_encode_u32(oid.hobj.get_bitwise_key_u32(), key); | |
402 | ||
403 | append_escaped(oid.hobj.nspace, key); | |
404 | ||
405 | if (oid.hobj.get_key().length()) { | |
406 | // is a key... could be < = or >. | |
407 | append_escaped(oid.hobj.get_key(), key); | |
408 | // (ASCII chars < = and > sort in that order, yay) | |
409 | int r = oid.hobj.get_key().compare(oid.hobj.oid.name); | |
410 | if (r) { | |
411 | key->append(r > 0 ? ">" : "<"); | |
412 | append_escaped(oid.hobj.oid.name, key); | |
413 | } else { | |
414 | // same as no key | |
415 | key->append("="); | |
416 | } | |
417 | } else { | |
418 | // no key | |
419 | append_escaped(oid.hobj.oid.name, key); | |
420 | key->append("="); | |
421 | } | |
422 | ||
423 | _key_encode_u64(oid.hobj.snap, key); | |
424 | _key_encode_u64(oid.generation, key); | |
425 | ||
426 | key->push_back(ONODE_KEY_SUFFIX); | |
427 | ||
428 | // sanity check | |
429 | if (true) { | |
430 | ghobject_t t; | |
431 | int r = get_key_object(*key, &t); | |
432 | if (r || t != oid) { | |
433 | derr << " r " << r << dendl; | |
434 | derr << "key " << pretty_binary_string(*key) << dendl; | |
435 | derr << "oid " << oid << dendl; | |
436 | derr << " t " << t << dendl; | |
437 | assert(r == 0 && t == oid); | |
438 | } | |
439 | } | |
440 | } | |
441 | ||
442 | ||
443 | // extent shard keys are the onode key, plus a u32, plus 'x'. the trailing | |
444 | // char lets us quickly test whether it is a shard key without decoding any | |
445 | // of the prefix bytes. | |
446 | template<typename S> | |
447 | static void get_extent_shard_key(const S& onode_key, uint32_t offset, | |
448 | string *key) | |
449 | { | |
450 | key->clear(); | |
451 | key->reserve(onode_key.length() + 4 + 1); | |
452 | key->append(onode_key.c_str(), onode_key.size()); | |
453 | _key_encode_u32(offset, key); | |
454 | key->push_back(EXTENT_SHARD_KEY_SUFFIX); | |
455 | } | |
456 | ||
457 | static void rewrite_extent_shard_key(uint32_t offset, string *key) | |
458 | { | |
459 | assert(key->size() > sizeof(uint32_t) + 1); | |
460 | assert(*key->rbegin() == EXTENT_SHARD_KEY_SUFFIX); | |
461 | _key_encode_u32(offset, key->size() - sizeof(uint32_t) - 1, key); | |
462 | } | |
463 | ||
464 | template<typename S> | |
465 | static void generate_extent_shard_key_and_apply( | |
466 | const S& onode_key, | |
467 | uint32_t offset, | |
468 | string *key, | |
469 | std::function<void(const string& final_key)> apply) | |
470 | { | |
471 | if (key->empty()) { // make full key | |
472 | assert(!onode_key.empty()); | |
473 | get_extent_shard_key(onode_key, offset, key); | |
474 | } else { | |
475 | rewrite_extent_shard_key(offset, key); | |
476 | } | |
477 | apply(*key); | |
478 | } | |
479 | ||
480 | int get_key_extent_shard(const string& key, string *onode_key, uint32_t *offset) | |
481 | { | |
482 | assert(key.size() > sizeof(uint32_t) + 1); | |
483 | assert(*key.rbegin() == EXTENT_SHARD_KEY_SUFFIX); | |
484 | int okey_len = key.size() - sizeof(uint32_t) - 1; | |
485 | *onode_key = key.substr(0, okey_len); | |
486 | const char *p = key.data() + okey_len; | |
224ce89b | 487 | _key_decode_u32(p, offset); |
7c673cae FG |
488 | return 0; |
489 | } | |
490 | ||
491 | static bool is_extent_shard_key(const string& key) | |
492 | { | |
493 | return *key.rbegin() == EXTENT_SHARD_KEY_SUFFIX; | |
494 | } | |
495 | ||
496 | // '-' < '.' < '~' | |
497 | static void get_omap_header(uint64_t id, string *out) | |
498 | { | |
499 | _key_encode_u64(id, out); | |
500 | out->push_back('-'); | |
501 | } | |
502 | ||
503 | // hmm, I don't think there's any need to escape the user key since we | |
504 | // have a clean prefix. | |
505 | static void get_omap_key(uint64_t id, const string& key, string *out) | |
506 | { | |
507 | _key_encode_u64(id, out); | |
508 | out->push_back('.'); | |
509 | out->append(key); | |
510 | } | |
511 | ||
512 | static void rewrite_omap_key(uint64_t id, string old, string *out) | |
513 | { | |
514 | _key_encode_u64(id, out); | |
515 | out->append(old.c_str() + out->length(), old.size() - out->length()); | |
516 | } | |
517 | ||
518 | static void decode_omap_key(const string& key, string *user_key) | |
519 | { | |
520 | *user_key = key.substr(sizeof(uint64_t) + 1); | |
521 | } | |
522 | ||
523 | static void get_omap_tail(uint64_t id, string *out) | |
524 | { | |
525 | _key_encode_u64(id, out); | |
526 | out->push_back('~'); | |
527 | } | |
528 | ||
529 | static void get_deferred_key(uint64_t seq, string *out) | |
530 | { | |
531 | _key_encode_u64(seq, out); | |
532 | } | |
533 | ||
534 | ||
535 | // merge operators | |
536 | ||
537 | struct Int64ArrayMergeOperator : public KeyValueDB::MergeOperator { | |
538 | void merge_nonexistent( | |
539 | const char *rdata, size_t rlen, std::string *new_value) override { | |
540 | *new_value = std::string(rdata, rlen); | |
541 | } | |
542 | void merge( | |
543 | const char *ldata, size_t llen, | |
544 | const char *rdata, size_t rlen, | |
545 | std::string *new_value) override { | |
546 | assert(llen == rlen); | |
547 | assert((rlen % 8) == 0); | |
548 | new_value->resize(rlen); | |
549 | const __le64* lv = (const __le64*)ldata; | |
550 | const __le64* rv = (const __le64*)rdata; | |
551 | __le64* nv = &(__le64&)new_value->at(0); | |
552 | for (size_t i = 0; i < rlen >> 3; ++i) { | |
553 | nv[i] = lv[i] + rv[i]; | |
554 | } | |
555 | } | |
556 | // We use each operator name and each prefix to construct the | |
557 | // overall RocksDB operator name for consistency check at open time. | |
558 | string name() const override { | |
559 | return "int64_array"; | |
560 | } | |
561 | }; | |
562 | ||
563 | ||
564 | // Buffer | |
565 | ||
566 | ostream& operator<<(ostream& out, const BlueStore::Buffer& b) | |
567 | { | |
568 | out << "buffer(" << &b << " space " << b.space << " 0x" << std::hex | |
569 | << b.offset << "~" << b.length << std::dec | |
570 | << " " << BlueStore::Buffer::get_state_name(b.state); | |
571 | if (b.flags) | |
572 | out << " " << BlueStore::Buffer::get_flag_name(b.flags); | |
573 | return out << ")"; | |
574 | } | |
575 | ||
576 | // Garbage Collector | |
577 | ||
578 | void BlueStore::GarbageCollector::process_protrusive_extents( | |
579 | const BlueStore::ExtentMap& extent_map, | |
580 | uint64_t start_offset, | |
581 | uint64_t end_offset, | |
582 | uint64_t start_touch_offset, | |
583 | uint64_t end_touch_offset, | |
584 | uint64_t min_alloc_size) | |
585 | { | |
586 | assert(start_offset <= start_touch_offset && end_offset>= end_touch_offset); | |
587 | ||
588 | uint64_t lookup_start_offset = P2ALIGN(start_offset, min_alloc_size); | |
589 | uint64_t lookup_end_offset = ROUND_UP_TO(end_offset, min_alloc_size); | |
590 | ||
591 | dout(30) << __func__ << " (hex): [" << std::hex | |
592 | << lookup_start_offset << ", " << lookup_end_offset | |
593 | << ")" << std::dec << dendl; | |
594 | ||
595 | for (auto it = extent_map.seek_lextent(lookup_start_offset); | |
596 | it != extent_map.extent_map.end() && | |
597 | it->logical_offset < lookup_end_offset; | |
598 | ++it) { | |
599 | uint64_t alloc_unit_start = it->logical_offset / min_alloc_size; | |
600 | uint64_t alloc_unit_end = (it->logical_end() - 1) / min_alloc_size; | |
601 | ||
602 | dout(30) << __func__ << " " << *it | |
603 | << "alloc_units: " << alloc_unit_start << ".." << alloc_unit_end | |
604 | << dendl; | |
605 | ||
606 | Blob* b = it->blob.get(); | |
607 | ||
608 | if (it->logical_offset >=start_touch_offset && | |
609 | it->logical_end() <= end_touch_offset) { | |
610 | // Process extents within the range affected by | |
611 | // the current write request. | |
612 | // Need to take into account if existing extents | |
613 | // can be merged with them (uncompressed case) | |
614 | if (!b->get_blob().is_compressed()) { | |
615 | if (blob_info_counted && used_alloc_unit == alloc_unit_start) { | |
616 | --blob_info_counted->expected_allocations; // don't need to allocate | |
617 | // new AU for compressed | |
618 | // data since another | |
619 | // collocated uncompressed | |
620 | // blob already exists | |
621 | dout(30) << __func__ << " --expected:" | |
622 | << alloc_unit_start << dendl; | |
623 | } | |
624 | used_alloc_unit = alloc_unit_end; | |
625 | blob_info_counted = nullptr; | |
626 | } | |
627 | } else if (b->get_blob().is_compressed()) { | |
628 | ||
629 | // additionally we take compressed blobs that were not impacted | |
630 | // by the write into account too | |
631 | BlobInfo& bi = | |
632 | affected_blobs.emplace( | |
633 | b, BlobInfo(b->get_referenced_bytes())).first->second; | |
634 | ||
635 | int adjust = | |
636 | (used_alloc_unit && used_alloc_unit == alloc_unit_start) ? 0 : 1; | |
637 | bi.expected_allocations += alloc_unit_end - alloc_unit_start + adjust; | |
638 | dout(30) << __func__ << " expected_allocations=" | |
639 | << bi.expected_allocations << " end_au:" | |
640 | << alloc_unit_end << dendl; | |
641 | ||
642 | blob_info_counted = &bi; | |
643 | used_alloc_unit = alloc_unit_end; | |
644 | ||
645 | assert(it->length <= bi.referenced_bytes); | |
646 | bi.referenced_bytes -= it->length; | |
647 | dout(30) << __func__ << " affected_blob:" << *b | |
648 | << " unref 0x" << std::hex << it->length | |
649 | << " referenced = 0x" << bi.referenced_bytes | |
650 | << std::dec << dendl; | |
651 | // NOTE: we can't move specific blob to resulting GC list here | |
652 | // when reference counter == 0 since subsequent extents might | |
653 | // decrement its expected_allocation. | |
654 | // Hence need to enumerate all the extents first. | |
655 | if (!bi.collect_candidate) { | |
656 | bi.first_lextent = it; | |
657 | bi.collect_candidate = true; | |
658 | } | |
659 | bi.last_lextent = it; | |
660 | } else { | |
661 | if (blob_info_counted && used_alloc_unit == alloc_unit_start) { | |
662 | // don't need to allocate new AU for compressed data since another | |
663 | // collocated uncompressed blob already exists | |
664 | --blob_info_counted->expected_allocations; | |
665 | dout(30) << __func__ << " --expected_allocations:" | |
666 | << alloc_unit_start << dendl; | |
667 | } | |
668 | used_alloc_unit = alloc_unit_end; | |
669 | blob_info_counted = nullptr; | |
670 | } | |
671 | } | |
672 | ||
673 | for (auto b_it = affected_blobs.begin(); | |
674 | b_it != affected_blobs.end(); | |
675 | ++b_it) { | |
676 | Blob* b = b_it->first; | |
677 | BlobInfo& bi = b_it->second; | |
678 | if (bi.referenced_bytes == 0) { | |
679 | uint64_t len_on_disk = b_it->first->get_blob().get_ondisk_length(); | |
680 | int64_t blob_expected_for_release = | |
681 | ROUND_UP_TO(len_on_disk, min_alloc_size) / min_alloc_size; | |
682 | ||
683 | dout(30) << __func__ << " " << *(b_it->first) | |
684 | << " expected4release=" << blob_expected_for_release | |
685 | << " expected_allocations=" << bi.expected_allocations | |
686 | << dendl; | |
687 | int64_t benefit = blob_expected_for_release - bi.expected_allocations; | |
688 | if (benefit >= g_conf->bluestore_gc_enable_blob_threshold) { | |
689 | if (bi.collect_candidate) { | |
690 | auto it = bi.first_lextent; | |
691 | bool bExit = false; | |
692 | do { | |
693 | if (it->blob.get() == b) { | |
694 | extents_to_collect.emplace_back(it->logical_offset, it->length); | |
695 | } | |
696 | bExit = it == bi.last_lextent; | |
697 | ++it; | |
31f18b77 | 698 | } while (!bExit); |
7c673cae FG |
699 | } |
700 | expected_for_release += blob_expected_for_release; | |
701 | expected_allocations += bi.expected_allocations; | |
702 | } | |
703 | } | |
704 | } | |
705 | } | |
706 | ||
707 | int64_t BlueStore::GarbageCollector::estimate( | |
708 | uint64_t start_offset, | |
709 | uint64_t length, | |
710 | const BlueStore::ExtentMap& extent_map, | |
711 | const BlueStore::old_extent_map_t& old_extents, | |
712 | uint64_t min_alloc_size) | |
713 | { | |
714 | ||
715 | affected_blobs.clear(); | |
716 | extents_to_collect.clear(); | |
717 | used_alloc_unit = boost::optional<uint64_t >(); | |
718 | blob_info_counted = nullptr; | |
719 | ||
720 | gc_start_offset = start_offset; | |
721 | gc_end_offset = start_offset + length; | |
722 | ||
723 | uint64_t end_offset = start_offset + length; | |
724 | ||
725 | for (auto it = old_extents.begin(); it != old_extents.end(); ++it) { | |
726 | Blob* b = it->e.blob.get(); | |
727 | if (b->get_blob().is_compressed()) { | |
728 | ||
729 | // update gc_start_offset/gc_end_offset if needed | |
730 | gc_start_offset = min(gc_start_offset, (uint64_t)it->e.blob_start()); | |
731 | gc_end_offset = max(gc_end_offset, (uint64_t)it->e.blob_end()); | |
732 | ||
733 | auto o = it->e.logical_offset; | |
734 | auto l = it->e.length; | |
735 | ||
736 | uint64_t ref_bytes = b->get_referenced_bytes(); | |
737 | // micro optimization to bypass blobs that have no more references | |
738 | if (ref_bytes != 0) { | |
739 | dout(30) << __func__ << " affected_blob:" << *b | |
740 | << " unref 0x" << std::hex << o << "~" << l | |
741 | << std::dec << dendl; | |
742 | affected_blobs.emplace(b, BlobInfo(ref_bytes)); | |
743 | } | |
744 | } | |
745 | } | |
746 | dout(30) << __func__ << " gc range(hex): [" << std::hex | |
747 | << gc_start_offset << ", " << gc_end_offset | |
748 | << ")" << std::dec << dendl; | |
749 | ||
750 | // enumerate preceeding extents to check if they reference affected blobs | |
751 | if (gc_start_offset < start_offset || gc_end_offset > end_offset) { | |
752 | process_protrusive_extents(extent_map, | |
753 | gc_start_offset, | |
754 | gc_end_offset, | |
755 | start_offset, | |
756 | end_offset, | |
757 | min_alloc_size); | |
758 | } | |
759 | return expected_for_release - expected_allocations; | |
760 | } | |
761 | ||
762 | // Cache | |
763 | ||
764 | BlueStore::Cache *BlueStore::Cache::create(CephContext* cct, string type, | |
765 | PerfCounters *logger) | |
766 | { | |
767 | Cache *c = nullptr; | |
768 | ||
769 | if (type == "lru") | |
770 | c = new LRUCache(cct); | |
771 | else if (type == "2q") | |
772 | c = new TwoQCache(cct); | |
773 | else | |
774 | assert(0 == "unrecognized cache type"); | |
775 | ||
776 | c->logger = logger; | |
777 | return c; | |
778 | } | |
779 | ||
780 | void BlueStore::Cache::trim_all() | |
781 | { | |
782 | std::lock_guard<std::recursive_mutex> l(lock); | |
783 | _trim(0, 0); | |
7c673cae FG |
784 | } |
785 | ||
786 | void BlueStore::Cache::trim( | |
787 | uint64_t target_bytes, | |
788 | float target_meta_ratio, | |
31f18b77 | 789 | float target_data_ratio, |
7c673cae FG |
790 | float bytes_per_onode) |
791 | { | |
792 | std::lock_guard<std::recursive_mutex> l(lock); | |
793 | uint64_t current_meta = _get_num_onodes() * bytes_per_onode; | |
794 | uint64_t current_buffer = _get_buffer_bytes(); | |
795 | uint64_t current = current_meta + current_buffer; | |
796 | ||
31f18b77 FG |
797 | uint64_t target_meta = target_bytes * target_meta_ratio; |
798 | uint64_t target_buffer = target_bytes * target_data_ratio; | |
7c673cae | 799 | |
31f18b77 FG |
800 | // correct for overflow or float imprecision |
801 | target_meta = min(target_bytes, target_meta); | |
802 | target_buffer = min(target_bytes - target_meta, target_buffer); | |
7c673cae FG |
803 | |
804 | if (current <= target_bytes) { | |
805 | dout(10) << __func__ | |
806 | << " shard target " << pretty_si_t(target_bytes) | |
31f18b77 FG |
807 | << " meta/data ratios " << target_meta_ratio |
808 | << " + " << target_data_ratio << " (" | |
7c673cae FG |
809 | << pretty_si_t(target_meta) << " + " |
810 | << pretty_si_t(target_buffer) << "), " | |
811 | << " current " << pretty_si_t(current) << " (" | |
812 | << pretty_si_t(current_meta) << " + " | |
813 | << pretty_si_t(current_buffer) << ")" | |
814 | << dendl; | |
815 | return; | |
816 | } | |
817 | ||
818 | uint64_t need_to_free = current - target_bytes; | |
819 | uint64_t free_buffer = 0; | |
820 | uint64_t free_meta = 0; | |
821 | if (current_buffer > target_buffer) { | |
822 | free_buffer = current_buffer - target_buffer; | |
823 | if (free_buffer > need_to_free) { | |
824 | free_buffer = need_to_free; | |
825 | } | |
826 | } | |
827 | free_meta = need_to_free - free_buffer; | |
828 | ||
829 | // start bounds at what we have now | |
830 | uint64_t max_buffer = current_buffer - free_buffer; | |
831 | uint64_t max_meta = current_meta - free_meta; | |
832 | uint64_t max_onodes = max_meta / bytes_per_onode; | |
833 | ||
834 | dout(10) << __func__ | |
835 | << " shard target " << pretty_si_t(target_bytes) | |
836 | << " ratio " << target_meta_ratio << " (" | |
837 | << pretty_si_t(target_meta) << " + " | |
838 | << pretty_si_t(target_buffer) << "), " | |
839 | << " current " << pretty_si_t(current) << " (" | |
840 | << pretty_si_t(current_meta) << " + " | |
841 | << pretty_si_t(current_buffer) << ")," | |
842 | << " need_to_free " << pretty_si_t(need_to_free) << " (" | |
843 | << pretty_si_t(free_meta) << " + " | |
844 | << pretty_si_t(free_buffer) << ")" | |
845 | << " -> max " << max_onodes << " onodes + " | |
846 | << max_buffer << " buffer" | |
847 | << dendl; | |
848 | _trim(max_onodes, max_buffer); | |
849 | } | |
850 | ||
851 | ||
852 | // LRUCache | |
853 | #undef dout_prefix | |
854 | #define dout_prefix *_dout << "bluestore.LRUCache(" << this << ") " | |
855 | ||
856 | void BlueStore::LRUCache::_touch_onode(OnodeRef& o) | |
857 | { | |
858 | auto p = onode_lru.iterator_to(*o); | |
859 | onode_lru.erase(p); | |
860 | onode_lru.push_front(*o); | |
861 | } | |
862 | ||
863 | void BlueStore::LRUCache::_trim(uint64_t onode_max, uint64_t buffer_max) | |
864 | { | |
865 | dout(20) << __func__ << " onodes " << onode_lru.size() << " / " << onode_max | |
866 | << " buffers " << buffer_size << " / " << buffer_max | |
867 | << dendl; | |
868 | ||
869 | _audit("trim start"); | |
870 | ||
871 | // buffers | |
872 | while (buffer_size > buffer_max) { | |
873 | auto i = buffer_lru.rbegin(); | |
874 | if (i == buffer_lru.rend()) { | |
875 | // stop if buffer_lru is now empty | |
876 | break; | |
877 | } | |
878 | ||
879 | Buffer *b = &*i; | |
880 | assert(b->is_clean()); | |
881 | dout(20) << __func__ << " rm " << *b << dendl; | |
882 | b->space->_rm_buffer(this, b); | |
883 | } | |
884 | ||
885 | // onodes | |
886 | int num = onode_lru.size() - onode_max; | |
887 | if (num <= 0) | |
888 | return; // don't even try | |
889 | ||
890 | auto p = onode_lru.end(); | |
891 | assert(p != onode_lru.begin()); | |
892 | --p; | |
893 | int skipped = 0; | |
894 | int max_skipped = g_conf->bluestore_cache_trim_max_skip_pinned; | |
895 | while (num > 0) { | |
896 | Onode *o = &*p; | |
897 | int refs = o->nref.load(); | |
898 | if (refs > 1) { | |
899 | dout(20) << __func__ << " " << o->oid << " has " << refs | |
900 | << " refs, skipping" << dendl; | |
901 | if (++skipped >= max_skipped) { | |
902 | dout(20) << __func__ << " maximum skip pinned reached; stopping with " | |
903 | << num << " left to trim" << dendl; | |
904 | break; | |
905 | } | |
906 | ||
907 | if (p == onode_lru.begin()) { | |
908 | break; | |
909 | } else { | |
910 | p--; | |
911 | num--; | |
912 | continue; | |
913 | } | |
914 | } | |
915 | dout(30) << __func__ << " rm " << o->oid << dendl; | |
916 | if (p != onode_lru.begin()) { | |
917 | onode_lru.erase(p--); | |
918 | } else { | |
919 | onode_lru.erase(p); | |
920 | assert(num == 1); | |
921 | } | |
922 | o->get(); // paranoia | |
923 | o->c->onode_map.remove(o->oid); | |
924 | o->put(); | |
925 | --num; | |
926 | } | |
927 | } | |
928 | ||
929 | #ifdef DEBUG_CACHE | |
930 | void BlueStore::LRUCache::_audit(const char *when) | |
931 | { | |
932 | dout(10) << __func__ << " " << when << " start" << dendl; | |
933 | uint64_t s = 0; | |
934 | for (auto i = buffer_lru.begin(); i != buffer_lru.end(); ++i) { | |
935 | s += i->length; | |
936 | } | |
937 | if (s != buffer_size) { | |
938 | derr << __func__ << " buffer_size " << buffer_size << " actual " << s | |
939 | << dendl; | |
940 | for (auto i = buffer_lru.begin(); i != buffer_lru.end(); ++i) { | |
941 | derr << __func__ << " " << *i << dendl; | |
942 | } | |
943 | assert(s == buffer_size); | |
944 | } | |
945 | dout(20) << __func__ << " " << when << " buffer_size " << buffer_size | |
946 | << " ok" << dendl; | |
947 | } | |
948 | #endif | |
949 | ||
950 | // TwoQCache | |
951 | #undef dout_prefix | |
952 | #define dout_prefix *_dout << "bluestore.2QCache(" << this << ") " | |
953 | ||
954 | ||
955 | void BlueStore::TwoQCache::_touch_onode(OnodeRef& o) | |
956 | { | |
957 | auto p = onode_lru.iterator_to(*o); | |
958 | onode_lru.erase(p); | |
959 | onode_lru.push_front(*o); | |
960 | } | |
961 | ||
962 | void BlueStore::TwoQCache::_add_buffer(Buffer *b, int level, Buffer *near) | |
963 | { | |
964 | dout(20) << __func__ << " level " << level << " near " << near | |
965 | << " on " << *b | |
966 | << " which has cache_private " << b->cache_private << dendl; | |
967 | if (near) { | |
968 | b->cache_private = near->cache_private; | |
969 | switch (b->cache_private) { | |
970 | case BUFFER_WARM_IN: | |
971 | buffer_warm_in.insert(buffer_warm_in.iterator_to(*near), *b); | |
972 | break; | |
973 | case BUFFER_WARM_OUT: | |
974 | assert(b->is_empty()); | |
975 | buffer_warm_out.insert(buffer_warm_out.iterator_to(*near), *b); | |
976 | break; | |
977 | case BUFFER_HOT: | |
978 | buffer_hot.insert(buffer_hot.iterator_to(*near), *b); | |
979 | break; | |
980 | default: | |
981 | assert(0 == "bad cache_private"); | |
982 | } | |
983 | } else if (b->cache_private == BUFFER_NEW) { | |
984 | b->cache_private = BUFFER_WARM_IN; | |
985 | if (level > 0) { | |
986 | buffer_warm_in.push_front(*b); | |
987 | } else { | |
988 | // take caller hint to start at the back of the warm queue | |
989 | buffer_warm_in.push_back(*b); | |
990 | } | |
991 | } else { | |
992 | // we got a hint from discard | |
993 | switch (b->cache_private) { | |
994 | case BUFFER_WARM_IN: | |
995 | // stay in warm_in. move to front, even though 2Q doesn't actually | |
996 | // do this. | |
997 | dout(20) << __func__ << " move to front of warm " << *b << dendl; | |
998 | buffer_warm_in.push_front(*b); | |
999 | break; | |
1000 | case BUFFER_WARM_OUT: | |
1001 | b->cache_private = BUFFER_HOT; | |
1002 | // move to hot. fall-thru | |
1003 | case BUFFER_HOT: | |
1004 | dout(20) << __func__ << " move to front of hot " << *b << dendl; | |
1005 | buffer_hot.push_front(*b); | |
1006 | break; | |
1007 | default: | |
1008 | assert(0 == "bad cache_private"); | |
1009 | } | |
1010 | } | |
1011 | if (!b->is_empty()) { | |
1012 | buffer_bytes += b->length; | |
1013 | buffer_list_bytes[b->cache_private] += b->length; | |
1014 | } | |
1015 | } | |
1016 | ||
1017 | void BlueStore::TwoQCache::_rm_buffer(Buffer *b) | |
1018 | { | |
1019 | dout(20) << __func__ << " " << *b << dendl; | |
1020 | if (!b->is_empty()) { | |
1021 | assert(buffer_bytes >= b->length); | |
1022 | buffer_bytes -= b->length; | |
1023 | assert(buffer_list_bytes[b->cache_private] >= b->length); | |
1024 | buffer_list_bytes[b->cache_private] -= b->length; | |
1025 | } | |
1026 | switch (b->cache_private) { | |
1027 | case BUFFER_WARM_IN: | |
1028 | buffer_warm_in.erase(buffer_warm_in.iterator_to(*b)); | |
1029 | break; | |
1030 | case BUFFER_WARM_OUT: | |
1031 | buffer_warm_out.erase(buffer_warm_out.iterator_to(*b)); | |
1032 | break; | |
1033 | case BUFFER_HOT: | |
1034 | buffer_hot.erase(buffer_hot.iterator_to(*b)); | |
1035 | break; | |
1036 | default: | |
1037 | assert(0 == "bad cache_private"); | |
1038 | } | |
1039 | } | |
1040 | ||
1041 | void BlueStore::TwoQCache::_move_buffer(Cache *srcc, Buffer *b) | |
1042 | { | |
1043 | TwoQCache *src = static_cast<TwoQCache*>(srcc); | |
1044 | src->_rm_buffer(b); | |
1045 | ||
1046 | // preserve which list we're on (even if we can't preserve the order!) | |
1047 | switch (b->cache_private) { | |
1048 | case BUFFER_WARM_IN: | |
1049 | assert(!b->is_empty()); | |
1050 | buffer_warm_in.push_back(*b); | |
1051 | break; | |
1052 | case BUFFER_WARM_OUT: | |
1053 | assert(b->is_empty()); | |
1054 | buffer_warm_out.push_back(*b); | |
1055 | break; | |
1056 | case BUFFER_HOT: | |
1057 | assert(!b->is_empty()); | |
1058 | buffer_hot.push_back(*b); | |
1059 | break; | |
1060 | default: | |
1061 | assert(0 == "bad cache_private"); | |
1062 | } | |
1063 | if (!b->is_empty()) { | |
1064 | buffer_bytes += b->length; | |
1065 | buffer_list_bytes[b->cache_private] += b->length; | |
1066 | } | |
1067 | } | |
1068 | ||
1069 | void BlueStore::TwoQCache::_adjust_buffer_size(Buffer *b, int64_t delta) | |
1070 | { | |
1071 | dout(20) << __func__ << " delta " << delta << " on " << *b << dendl; | |
1072 | if (!b->is_empty()) { | |
1073 | assert((int64_t)buffer_bytes + delta >= 0); | |
1074 | buffer_bytes += delta; | |
1075 | assert((int64_t)buffer_list_bytes[b->cache_private] + delta >= 0); | |
1076 | buffer_list_bytes[b->cache_private] += delta; | |
1077 | } | |
1078 | } | |
1079 | ||
1080 | void BlueStore::TwoQCache::_trim(uint64_t onode_max, uint64_t buffer_max) | |
1081 | { | |
1082 | dout(20) << __func__ << " onodes " << onode_lru.size() << " / " << onode_max | |
1083 | << " buffers " << buffer_bytes << " / " << buffer_max | |
1084 | << dendl; | |
1085 | ||
1086 | _audit("trim start"); | |
1087 | ||
1088 | // buffers | |
1089 | if (buffer_bytes > buffer_max) { | |
1090 | uint64_t kin = buffer_max * cct->_conf->bluestore_2q_cache_kin_ratio; | |
1091 | uint64_t khot = buffer_max - kin; | |
1092 | ||
1093 | // pre-calculate kout based on average buffer size too, | |
1094 | // which is typical(the warm_in and hot lists may change later) | |
1095 | uint64_t kout = 0; | |
1096 | uint64_t buffer_num = buffer_hot.size() + buffer_warm_in.size(); | |
1097 | if (buffer_num) { | |
1098 | uint64_t buffer_avg_size = buffer_bytes / buffer_num; | |
1099 | assert(buffer_avg_size); | |
1100 | uint64_t calculated_buffer_num = buffer_max / buffer_avg_size; | |
1101 | kout = calculated_buffer_num * cct->_conf->bluestore_2q_cache_kout_ratio; | |
1102 | } | |
1103 | ||
1104 | if (buffer_list_bytes[BUFFER_HOT] < khot) { | |
1105 | // hot is small, give slack to warm_in | |
1106 | kin += khot - buffer_list_bytes[BUFFER_HOT]; | |
1107 | } else if (buffer_list_bytes[BUFFER_WARM_IN] < kin) { | |
1108 | // warm_in is small, give slack to hot | |
1109 | khot += kin - buffer_list_bytes[BUFFER_WARM_IN]; | |
1110 | } | |
1111 | ||
1112 | // adjust warm_in list | |
1113 | int64_t to_evict_bytes = buffer_list_bytes[BUFFER_WARM_IN] - kin; | |
1114 | uint64_t evicted = 0; | |
1115 | ||
1116 | while (to_evict_bytes > 0) { | |
1117 | auto p = buffer_warm_in.rbegin(); | |
1118 | if (p == buffer_warm_in.rend()) { | |
1119 | // stop if warm_in list is now empty | |
1120 | break; | |
1121 | } | |
1122 | ||
1123 | Buffer *b = &*p; | |
1124 | assert(b->is_clean()); | |
1125 | dout(20) << __func__ << " buffer_warm_in -> out " << *b << dendl; | |
1126 | assert(buffer_bytes >= b->length); | |
1127 | buffer_bytes -= b->length; | |
1128 | assert(buffer_list_bytes[BUFFER_WARM_IN] >= b->length); | |
1129 | buffer_list_bytes[BUFFER_WARM_IN] -= b->length; | |
1130 | to_evict_bytes -= b->length; | |
1131 | evicted += b->length; | |
1132 | b->state = Buffer::STATE_EMPTY; | |
1133 | b->data.clear(); | |
1134 | buffer_warm_in.erase(buffer_warm_in.iterator_to(*b)); | |
1135 | buffer_warm_out.push_front(*b); | |
1136 | b->cache_private = BUFFER_WARM_OUT; | |
1137 | } | |
1138 | ||
1139 | if (evicted > 0) { | |
1140 | dout(20) << __func__ << " evicted " << prettybyte_t(evicted) | |
1141 | << " from warm_in list, done evicting warm_in buffers" | |
1142 | << dendl; | |
1143 | } | |
1144 | ||
1145 | // adjust hot list | |
1146 | to_evict_bytes = buffer_list_bytes[BUFFER_HOT] - khot; | |
1147 | evicted = 0; | |
1148 | ||
1149 | while (to_evict_bytes > 0) { | |
1150 | auto p = buffer_hot.rbegin(); | |
1151 | if (p == buffer_hot.rend()) { | |
1152 | // stop if hot list is now empty | |
1153 | break; | |
1154 | } | |
1155 | ||
1156 | Buffer *b = &*p; | |
1157 | dout(20) << __func__ << " buffer_hot rm " << *b << dendl; | |
1158 | assert(b->is_clean()); | |
1159 | // adjust evict size before buffer goes invalid | |
1160 | to_evict_bytes -= b->length; | |
1161 | evicted += b->length; | |
1162 | b->space->_rm_buffer(this, b); | |
1163 | } | |
1164 | ||
1165 | if (evicted > 0) { | |
1166 | dout(20) << __func__ << " evicted " << prettybyte_t(evicted) | |
1167 | << " from hot list, done evicting hot buffers" | |
1168 | << dendl; | |
1169 | } | |
1170 | ||
1171 | // adjust warm out list too, if necessary | |
1172 | int64_t num = buffer_warm_out.size() - kout; | |
1173 | while (num-- > 0) { | |
1174 | Buffer *b = &*buffer_warm_out.rbegin(); | |
1175 | assert(b->is_empty()); | |
1176 | dout(20) << __func__ << " buffer_warm_out rm " << *b << dendl; | |
1177 | b->space->_rm_buffer(this, b); | |
1178 | } | |
1179 | } | |
1180 | ||
1181 | // onodes | |
1182 | int num = onode_lru.size() - onode_max; | |
1183 | if (num <= 0) | |
1184 | return; // don't even try | |
1185 | ||
1186 | auto p = onode_lru.end(); | |
1187 | assert(p != onode_lru.begin()); | |
1188 | --p; | |
1189 | int skipped = 0; | |
1190 | int max_skipped = g_conf->bluestore_cache_trim_max_skip_pinned; | |
1191 | while (num > 0) { | |
1192 | Onode *o = &*p; | |
1193 | dout(20) << __func__ << " considering " << o << dendl; | |
1194 | int refs = o->nref.load(); | |
1195 | if (refs > 1) { | |
1196 | dout(20) << __func__ << " " << o->oid << " has " << refs | |
1197 | << " refs; skipping" << dendl; | |
1198 | if (++skipped >= max_skipped) { | |
1199 | dout(20) << __func__ << " maximum skip pinned reached; stopping with " | |
1200 | << num << " left to trim" << dendl; | |
1201 | break; | |
1202 | } | |
1203 | ||
1204 | if (p == onode_lru.begin()) { | |
1205 | break; | |
1206 | } else { | |
1207 | p--; | |
1208 | num--; | |
1209 | continue; | |
1210 | } | |
1211 | } | |
1212 | dout(30) << __func__ << " " << o->oid << " num=" << num <<" lru size="<<onode_lru.size()<< dendl; | |
1213 | if (p != onode_lru.begin()) { | |
1214 | onode_lru.erase(p--); | |
1215 | } else { | |
1216 | onode_lru.erase(p); | |
1217 | assert(num == 1); | |
1218 | } | |
1219 | o->get(); // paranoia | |
1220 | o->c->onode_map.remove(o->oid); | |
1221 | o->put(); | |
1222 | --num; | |
1223 | } | |
1224 | } | |
1225 | ||
1226 | #ifdef DEBUG_CACHE | |
1227 | void BlueStore::TwoQCache::_audit(const char *when) | |
1228 | { | |
1229 | dout(10) << __func__ << " " << when << " start" << dendl; | |
1230 | uint64_t s = 0; | |
1231 | for (auto i = buffer_hot.begin(); i != buffer_hot.end(); ++i) { | |
1232 | s += i->length; | |
1233 | } | |
1234 | ||
1235 | uint64_t hot_bytes = s; | |
1236 | if (hot_bytes != buffer_list_bytes[BUFFER_HOT]) { | |
1237 | derr << __func__ << " hot_list_bytes " | |
1238 | << buffer_list_bytes[BUFFER_HOT] | |
1239 | << " != actual " << hot_bytes | |
1240 | << dendl; | |
1241 | assert(hot_bytes == buffer_list_bytes[BUFFER_HOT]); | |
1242 | } | |
1243 | ||
1244 | for (auto i = buffer_warm_in.begin(); i != buffer_warm_in.end(); ++i) { | |
1245 | s += i->length; | |
1246 | } | |
1247 | ||
1248 | uint64_t warm_in_bytes = s - hot_bytes; | |
1249 | if (warm_in_bytes != buffer_list_bytes[BUFFER_WARM_IN]) { | |
1250 | derr << __func__ << " warm_in_list_bytes " | |
1251 | << buffer_list_bytes[BUFFER_WARM_IN] | |
1252 | << " != actual " << warm_in_bytes | |
1253 | << dendl; | |
1254 | assert(warm_in_bytes == buffer_list_bytes[BUFFER_WARM_IN]); | |
1255 | } | |
1256 | ||
1257 | if (s != buffer_bytes) { | |
1258 | derr << __func__ << " buffer_bytes " << buffer_bytes << " actual " << s | |
1259 | << dendl; | |
1260 | assert(s == buffer_bytes); | |
1261 | } | |
1262 | ||
1263 | dout(20) << __func__ << " " << when << " buffer_bytes " << buffer_bytes | |
1264 | << " ok" << dendl; | |
1265 | } | |
1266 | #endif | |
1267 | ||
1268 | ||
1269 | // BufferSpace | |
1270 | ||
1271 | #undef dout_prefix | |
1272 | #define dout_prefix *_dout << "bluestore.BufferSpace(" << this << " in " << cache << ") " | |
1273 | ||
1274 | void BlueStore::BufferSpace::_clear(Cache* cache) | |
1275 | { | |
1276 | // note: we already hold cache->lock | |
1277 | ldout(cache->cct, 20) << __func__ << dendl; | |
1278 | while (!buffer_map.empty()) { | |
1279 | _rm_buffer(cache, buffer_map.begin()); | |
1280 | } | |
1281 | } | |
1282 | ||
1283 | int BlueStore::BufferSpace::_discard(Cache* cache, uint32_t offset, uint32_t length) | |
1284 | { | |
1285 | // note: we already hold cache->lock | |
1286 | ldout(cache->cct, 20) << __func__ << std::hex << " 0x" << offset << "~" << length | |
1287 | << std::dec << dendl; | |
1288 | int cache_private = 0; | |
1289 | cache->_audit("discard start"); | |
1290 | auto i = _data_lower_bound(offset); | |
1291 | uint32_t end = offset + length; | |
1292 | while (i != buffer_map.end()) { | |
1293 | Buffer *b = i->second.get(); | |
1294 | if (b->offset >= end) { | |
1295 | break; | |
1296 | } | |
1297 | if (b->cache_private > cache_private) { | |
1298 | cache_private = b->cache_private; | |
1299 | } | |
1300 | if (b->offset < offset) { | |
1301 | int64_t front = offset - b->offset; | |
1302 | if (b->end() > end) { | |
1303 | // drop middle (split) | |
1304 | uint32_t tail = b->end() - end; | |
1305 | if (b->data.length()) { | |
1306 | bufferlist bl; | |
1307 | bl.substr_of(b->data, b->length - tail, tail); | |
31f18b77 FG |
1308 | Buffer *nb = new Buffer(this, b->state, b->seq, end, bl); |
1309 | nb->maybe_rebuild(); | |
1310 | _add_buffer(cache, nb, 0, b); | |
7c673cae | 1311 | } else { |
31f18b77 FG |
1312 | _add_buffer(cache, new Buffer(this, b->state, b->seq, end, tail), |
1313 | 0, b); | |
7c673cae FG |
1314 | } |
1315 | if (!b->is_writing()) { | |
1316 | cache->_adjust_buffer_size(b, front - (int64_t)b->length); | |
1317 | } | |
1318 | b->truncate(front); | |
31f18b77 | 1319 | b->maybe_rebuild(); |
7c673cae FG |
1320 | cache->_audit("discard end 1"); |
1321 | break; | |
1322 | } else { | |
1323 | // drop tail | |
1324 | if (!b->is_writing()) { | |
1325 | cache->_adjust_buffer_size(b, front - (int64_t)b->length); | |
1326 | } | |
1327 | b->truncate(front); | |
31f18b77 | 1328 | b->maybe_rebuild(); |
7c673cae FG |
1329 | ++i; |
1330 | continue; | |
1331 | } | |
1332 | } | |
1333 | if (b->end() <= end) { | |
1334 | // drop entire buffer | |
1335 | _rm_buffer(cache, i++); | |
1336 | continue; | |
1337 | } | |
1338 | // drop front | |
1339 | uint32_t keep = b->end() - end; | |
1340 | if (b->data.length()) { | |
1341 | bufferlist bl; | |
1342 | bl.substr_of(b->data, b->length - keep, keep); | |
31f18b77 FG |
1343 | Buffer *nb = new Buffer(this, b->state, b->seq, end, bl); |
1344 | nb->maybe_rebuild(); | |
1345 | _add_buffer(cache, nb, 0, b); | |
7c673cae FG |
1346 | } else { |
1347 | _add_buffer(cache, new Buffer(this, b->state, b->seq, end, keep), 0, b); | |
1348 | } | |
1349 | _rm_buffer(cache, i); | |
1350 | cache->_audit("discard end 2"); | |
1351 | break; | |
1352 | } | |
1353 | return cache_private; | |
1354 | } | |
1355 | ||
1356 | void BlueStore::BufferSpace::read( | |
1357 | Cache* cache, | |
224ce89b WB |
1358 | uint32_t offset, |
1359 | uint32_t length, | |
7c673cae FG |
1360 | BlueStore::ready_regions_t& res, |
1361 | interval_set<uint32_t>& res_intervals) | |
1362 | { | |
7c673cae FG |
1363 | res.clear(); |
1364 | res_intervals.clear(); | |
1365 | uint32_t want_bytes = length; | |
1366 | uint32_t end = offset + length; | |
224ce89b WB |
1367 | |
1368 | { | |
1369 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1370 | for (auto i = _data_lower_bound(offset); | |
1371 | i != buffer_map.end() && offset < end && i->first < end; | |
1372 | ++i) { | |
1373 | Buffer *b = i->second.get(); | |
1374 | assert(b->end() > offset); | |
1375 | if (b->is_writing() || b->is_clean()) { | |
1376 | if (b->offset < offset) { | |
1377 | uint32_t skip = offset - b->offset; | |
1378 | uint32_t l = MIN(length, b->length - skip); | |
1379 | res[offset].substr_of(b->data, skip, l); | |
1380 | res_intervals.insert(offset, l); | |
1381 | offset += l; | |
1382 | length -= l; | |
1383 | if (!b->is_writing()) { | |
1384 | cache->_touch_buffer(b); | |
1385 | } | |
1386 | continue; | |
1387 | } | |
1388 | if (b->offset > offset) { | |
1389 | uint32_t gap = b->offset - offset; | |
1390 | if (length <= gap) { | |
1391 | break; | |
1392 | } | |
1393 | offset += gap; | |
1394 | length -= gap; | |
1395 | } | |
1396 | if (!b->is_writing()) { | |
7c673cae | 1397 | cache->_touch_buffer(b); |
224ce89b WB |
1398 | } |
1399 | if (b->length > length) { | |
1400 | res[offset].substr_of(b->data, 0, length); | |
1401 | res_intervals.insert(offset, length); | |
7c673cae | 1402 | break; |
224ce89b WB |
1403 | } else { |
1404 | res[offset].append(b->data); | |
1405 | res_intervals.insert(offset, b->length); | |
1406 | if (b->length == length) | |
1407 | break; | |
1408 | offset += b->length; | |
1409 | length -= b->length; | |
1410 | } | |
7c673cae FG |
1411 | } |
1412 | } | |
1413 | } | |
1414 | ||
1415 | uint64_t hit_bytes = res_intervals.size(); | |
1416 | assert(hit_bytes <= want_bytes); | |
1417 | uint64_t miss_bytes = want_bytes - hit_bytes; | |
1418 | cache->logger->inc(l_bluestore_buffer_hit_bytes, hit_bytes); | |
1419 | cache->logger->inc(l_bluestore_buffer_miss_bytes, miss_bytes); | |
1420 | } | |
1421 | ||
1422 | void BlueStore::BufferSpace::finish_write(Cache* cache, uint64_t seq) | |
1423 | { | |
1424 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1425 | ||
1426 | auto i = writing.begin(); | |
1427 | while (i != writing.end()) { | |
1428 | if (i->seq > seq) { | |
1429 | break; | |
1430 | } | |
1431 | if (i->seq < seq) { | |
1432 | ++i; | |
1433 | continue; | |
1434 | } | |
1435 | ||
1436 | Buffer *b = &*i; | |
1437 | assert(b->is_writing()); | |
1438 | ||
1439 | if (b->flags & Buffer::FLAG_NOCACHE) { | |
1440 | writing.erase(i++); | |
1441 | ldout(cache->cct, 20) << __func__ << " discard " << *b << dendl; | |
1442 | buffer_map.erase(b->offset); | |
1443 | } else { | |
1444 | b->state = Buffer::STATE_CLEAN; | |
1445 | writing.erase(i++); | |
31f18b77 FG |
1446 | b->maybe_rebuild(); |
1447 | b->data.reassign_to_mempool(mempool::mempool_bluestore_cache_data); | |
7c673cae FG |
1448 | cache->_add_buffer(b, 1, nullptr); |
1449 | ldout(cache->cct, 20) << __func__ << " added " << *b << dendl; | |
1450 | } | |
1451 | } | |
1452 | ||
1453 | cache->_audit("finish_write end"); | |
1454 | } | |
1455 | ||
1456 | void BlueStore::BufferSpace::split(Cache* cache, size_t pos, BlueStore::BufferSpace &r) | |
1457 | { | |
1458 | std::lock_guard<std::recursive_mutex> lk(cache->lock); | |
1459 | if (buffer_map.empty()) | |
1460 | return; | |
1461 | ||
1462 | auto p = --buffer_map.end(); | |
1463 | while (true) { | |
1464 | if (p->second->end() <= pos) | |
1465 | break; | |
1466 | ||
1467 | if (p->second->offset < pos) { | |
1468 | ldout(cache->cct, 30) << __func__ << " cut " << *p->second << dendl; | |
1469 | size_t left = pos - p->second->offset; | |
1470 | size_t right = p->second->length - left; | |
1471 | if (p->second->data.length()) { | |
1472 | bufferlist bl; | |
1473 | bl.substr_of(p->second->data, left, right); | |
1474 | r._add_buffer(cache, new Buffer(&r, p->second->state, p->second->seq, 0, bl), | |
1475 | 0, p->second.get()); | |
1476 | } else { | |
1477 | r._add_buffer(cache, new Buffer(&r, p->second->state, p->second->seq, 0, right), | |
1478 | 0, p->second.get()); | |
1479 | } | |
1480 | cache->_adjust_buffer_size(p->second.get(), -right); | |
1481 | p->second->truncate(left); | |
1482 | break; | |
1483 | } | |
1484 | ||
1485 | assert(p->second->end() > pos); | |
1486 | ldout(cache->cct, 30) << __func__ << " move " << *p->second << dendl; | |
1487 | if (p->second->data.length()) { | |
1488 | r._add_buffer(cache, new Buffer(&r, p->second->state, p->second->seq, | |
1489 | p->second->offset - pos, p->second->data), | |
1490 | 0, p->second.get()); | |
1491 | } else { | |
1492 | r._add_buffer(cache, new Buffer(&r, p->second->state, p->second->seq, | |
1493 | p->second->offset - pos, p->second->length), | |
1494 | 0, p->second.get()); | |
1495 | } | |
1496 | if (p == buffer_map.begin()) { | |
1497 | _rm_buffer(cache, p); | |
1498 | break; | |
1499 | } else { | |
1500 | _rm_buffer(cache, p--); | |
1501 | } | |
1502 | } | |
1503 | assert(writing.empty()); | |
1504 | } | |
1505 | ||
1506 | // OnodeSpace | |
1507 | ||
1508 | #undef dout_prefix | |
1509 | #define dout_prefix *_dout << "bluestore.OnodeSpace(" << this << " in " << cache << ") " | |
1510 | ||
1511 | BlueStore::OnodeRef BlueStore::OnodeSpace::add(const ghobject_t& oid, OnodeRef o) | |
1512 | { | |
1513 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1514 | auto p = onode_map.find(oid); | |
1515 | if (p != onode_map.end()) { | |
1516 | ldout(cache->cct, 30) << __func__ << " " << oid << " " << o | |
1517 | << " raced, returning existing " << p->second | |
1518 | << dendl; | |
1519 | return p->second; | |
1520 | } | |
1521 | ldout(cache->cct, 30) << __func__ << " " << oid << " " << o << dendl; | |
1522 | onode_map[oid] = o; | |
1523 | cache->_add_onode(o, 1); | |
1524 | return o; | |
1525 | } | |
1526 | ||
1527 | BlueStore::OnodeRef BlueStore::OnodeSpace::lookup(const ghobject_t& oid) | |
1528 | { | |
7c673cae | 1529 | ldout(cache->cct, 30) << __func__ << dendl; |
224ce89b WB |
1530 | OnodeRef o; |
1531 | bool hit = false; | |
1532 | ||
1533 | { | |
1534 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1535 | ceph::unordered_map<ghobject_t,OnodeRef>::iterator p = onode_map.find(oid); | |
1536 | if (p == onode_map.end()) { | |
1537 | ldout(cache->cct, 30) << __func__ << " " << oid << " miss" << dendl; | |
1538 | } else { | |
1539 | ldout(cache->cct, 30) << __func__ << " " << oid << " hit " << p->second | |
1540 | << dendl; | |
1541 | cache->_touch_onode(p->second); | |
1542 | hit = true; | |
1543 | o = p->second; | |
1544 | } | |
1545 | } | |
1546 | ||
1547 | if (hit) { | |
1548 | cache->logger->inc(l_bluestore_onode_hits); | |
1549 | } else { | |
7c673cae | 1550 | cache->logger->inc(l_bluestore_onode_misses); |
7c673cae | 1551 | } |
224ce89b | 1552 | return o; |
7c673cae FG |
1553 | } |
1554 | ||
1555 | void BlueStore::OnodeSpace::clear() | |
1556 | { | |
1557 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1558 | ldout(cache->cct, 10) << __func__ << dendl; | |
1559 | for (auto &p : onode_map) { | |
1560 | cache->_rm_onode(p.second); | |
1561 | } | |
1562 | onode_map.clear(); | |
1563 | } | |
1564 | ||
1565 | bool BlueStore::OnodeSpace::empty() | |
1566 | { | |
1567 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1568 | return onode_map.empty(); | |
1569 | } | |
1570 | ||
1571 | void BlueStore::OnodeSpace::rename( | |
1572 | OnodeRef& oldo, | |
1573 | const ghobject_t& old_oid, | |
1574 | const ghobject_t& new_oid, | |
31f18b77 | 1575 | const mempool::bluestore_cache_other::string& new_okey) |
7c673cae FG |
1576 | { |
1577 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1578 | ldout(cache->cct, 30) << __func__ << " " << old_oid << " -> " << new_oid | |
1579 | << dendl; | |
1580 | ceph::unordered_map<ghobject_t,OnodeRef>::iterator po, pn; | |
1581 | po = onode_map.find(old_oid); | |
1582 | pn = onode_map.find(new_oid); | |
1583 | assert(po != pn); | |
1584 | ||
1585 | assert(po != onode_map.end()); | |
1586 | if (pn != onode_map.end()) { | |
1587 | ldout(cache->cct, 30) << __func__ << " removing target " << pn->second | |
1588 | << dendl; | |
1589 | cache->_rm_onode(pn->second); | |
1590 | onode_map.erase(pn); | |
1591 | } | |
1592 | OnodeRef o = po->second; | |
1593 | ||
1594 | // install a non-existent onode at old location | |
1595 | oldo.reset(new Onode(o->c, old_oid, o->key)); | |
1596 | po->second = oldo; | |
1597 | cache->_add_onode(po->second, 1); | |
1598 | ||
1599 | // add at new position and fix oid, key | |
1600 | onode_map.insert(make_pair(new_oid, o)); | |
1601 | cache->_touch_onode(o); | |
1602 | o->oid = new_oid; | |
1603 | o->key = new_okey; | |
1604 | } | |
1605 | ||
1606 | bool BlueStore::OnodeSpace::map_any(std::function<bool(OnodeRef)> f) | |
1607 | { | |
1608 | std::lock_guard<std::recursive_mutex> l(cache->lock); | |
1609 | ldout(cache->cct, 20) << __func__ << dendl; | |
1610 | for (auto& i : onode_map) { | |
1611 | if (f(i.second)) { | |
1612 | return true; | |
1613 | } | |
1614 | } | |
1615 | return false; | |
1616 | } | |
1617 | ||
1618 | ||
1619 | // SharedBlob | |
1620 | ||
1621 | #undef dout_prefix | |
1622 | #define dout_prefix *_dout << "bluestore.sharedblob(" << this << ") " | |
1623 | ||
1624 | ostream& operator<<(ostream& out, const BlueStore::SharedBlob& sb) | |
1625 | { | |
1626 | out << "SharedBlob(" << &sb; | |
1627 | ||
1628 | if (sb.loaded) { | |
1629 | out << " loaded " << *sb.persistent; | |
1630 | } else { | |
1631 | out << " sbid 0x" << std::hex << sb.sbid_unloaded << std::dec; | |
1632 | } | |
1633 | return out << ")"; | |
1634 | } | |
1635 | ||
1636 | BlueStore::SharedBlob::SharedBlob(uint64_t i, Collection *_coll) | |
1637 | : coll(_coll), sbid_unloaded(i) | |
1638 | { | |
1639 | assert(sbid_unloaded > 0); | |
1640 | if (get_cache()) { | |
1641 | get_cache()->add_blob(); | |
1642 | } | |
1643 | } | |
1644 | ||
1645 | BlueStore::SharedBlob::~SharedBlob() | |
1646 | { | |
1647 | if (get_cache()) { // the dummy instances have a nullptr | |
1648 | std::lock_guard<std::recursive_mutex> l(get_cache()->lock); | |
1649 | bc._clear(get_cache()); | |
1650 | get_cache()->rm_blob(); | |
1651 | } | |
1652 | if (loaded && persistent) { | |
1653 | delete persistent; | |
1654 | } | |
1655 | } | |
1656 | ||
1657 | void BlueStore::SharedBlob::put() | |
1658 | { | |
1659 | if (--nref == 0) { | |
1660 | ldout(coll->store->cct, 20) << __func__ << " " << this | |
1661 | << " removing self from set " << get_parent() | |
1662 | << dendl; | |
1663 | if (get_parent()) { | |
1664 | if (get_parent()->remove(this)) { | |
1665 | delete this; | |
1666 | } else { | |
1667 | ldout(coll->store->cct, 20) | |
1668 | << __func__ << " " << this << " lost race to remove myself from set" | |
1669 | << dendl; | |
1670 | } | |
1671 | } else { | |
1672 | delete this; | |
1673 | } | |
1674 | } | |
1675 | } | |
1676 | ||
1677 | void BlueStore::SharedBlob::get_ref(uint64_t offset, uint32_t length) | |
1678 | { | |
1679 | assert(persistent); | |
1680 | persistent->ref_map.get(offset, length); | |
1681 | } | |
1682 | ||
1683 | void BlueStore::SharedBlob::put_ref(uint64_t offset, uint32_t length, | |
31f18b77 FG |
1684 | PExtentVector *r, |
1685 | set<SharedBlob*> *maybe_unshared) | |
7c673cae FG |
1686 | { |
1687 | assert(persistent); | |
31f18b77 FG |
1688 | bool maybe = false; |
1689 | persistent->ref_map.put(offset, length, r, maybe_unshared ? &maybe : nullptr); | |
1690 | if (maybe_unshared && maybe) { | |
1691 | maybe_unshared->insert(this); | |
1692 | } | |
7c673cae FG |
1693 | } |
1694 | ||
1695 | // Blob | |
1696 | ||
1697 | #undef dout_prefix | |
1698 | #define dout_prefix *_dout << "bluestore.blob(" << this << ") " | |
1699 | ||
1700 | ostream& operator<<(ostream& out, const BlueStore::Blob& b) | |
1701 | { | |
1702 | out << "Blob(" << &b; | |
1703 | if (b.is_spanning()) { | |
1704 | out << " spanning " << b.id; | |
1705 | } | |
35e4c445 FG |
1706 | out << " " << b.get_blob() << " " << b.get_blob_use_tracker(); |
1707 | if (b.shared_blob) { | |
1708 | out << " " << *b.shared_blob; | |
1709 | } else { | |
1710 | out << " (shared_blob=NULL)"; | |
1711 | } | |
1712 | out << ")"; | |
7c673cae FG |
1713 | return out; |
1714 | } | |
1715 | ||
1716 | void BlueStore::Blob::discard_unallocated(Collection *coll) | |
1717 | { | |
224ce89b | 1718 | if (get_blob().is_shared()) { |
7c673cae FG |
1719 | return; |
1720 | } | |
224ce89b | 1721 | if (get_blob().is_compressed()) { |
7c673cae FG |
1722 | bool discard = false; |
1723 | bool all_invalid = true; | |
224ce89b | 1724 | for (auto e : get_blob().get_extents()) { |
7c673cae FG |
1725 | if (!e.is_valid()) { |
1726 | discard = true; | |
1727 | } else { | |
1728 | all_invalid = false; | |
1729 | } | |
1730 | } | |
1731 | assert(discard == all_invalid); // in case of compressed blob all | |
1732 | // or none pextents are invalid. | |
1733 | if (discard) { | |
224ce89b WB |
1734 | shared_blob->bc.discard(shared_blob->get_cache(), 0, |
1735 | get_blob().get_logical_length()); | |
7c673cae FG |
1736 | } |
1737 | } else { | |
1738 | size_t pos = 0; | |
224ce89b | 1739 | for (auto e : get_blob().get_extents()) { |
7c673cae FG |
1740 | if (!e.is_valid()) { |
1741 | ldout(coll->store->cct, 20) << __func__ << " 0x" << std::hex << pos | |
1742 | << "~" << e.length | |
1743 | << std::dec << dendl; | |
1744 | shared_blob->bc.discard(shared_blob->get_cache(), pos, e.length); | |
1745 | } | |
1746 | pos += e.length; | |
1747 | } | |
224ce89b WB |
1748 | if (get_blob().can_prune_tail()) { |
1749 | dirty_blob().prune_tail(); | |
1750 | used_in_blob.prune_tail(get_blob().get_ondisk_length()); | |
7c673cae | 1751 | auto cct = coll->store->cct; //used by dout |
224ce89b | 1752 | dout(20) << __func__ << " pruned tail, now " << get_blob() << dendl; |
7c673cae FG |
1753 | } |
1754 | } | |
1755 | } | |
1756 | ||
1757 | void BlueStore::Blob::get_ref( | |
1758 | Collection *coll, | |
1759 | uint32_t offset, | |
1760 | uint32_t length) | |
1761 | { | |
1762 | // Caller has to initialize Blob's logical length prior to increment | |
1763 | // references. Otherwise one is neither unable to determine required | |
1764 | // amount of counters in case of per-au tracking nor obtain min_release_size | |
1765 | // for single counter mode. | |
1766 | assert(get_blob().get_logical_length() != 0); | |
1767 | auto cct = coll->store->cct; | |
1768 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
1769 | << std::dec << " " << *this << dendl; | |
1770 | ||
1771 | if (used_in_blob.is_empty()) { | |
1772 | uint32_t min_release_size = | |
224ce89b WB |
1773 | get_blob().get_release_size(coll->store->min_alloc_size); |
1774 | uint64_t l = get_blob().get_logical_length(); | |
1775 | dout(20) << __func__ << " init 0x" << std::hex << l << ", " | |
1776 | << min_release_size << std::dec << dendl; | |
7c673cae FG |
1777 | used_in_blob.init(l, min_release_size); |
1778 | } | |
1779 | used_in_blob.get( | |
1780 | offset, | |
1781 | length); | |
1782 | } | |
1783 | ||
1784 | bool BlueStore::Blob::put_ref( | |
1785 | Collection *coll, | |
1786 | uint32_t offset, | |
1787 | uint32_t length, | |
1788 | PExtentVector *r) | |
1789 | { | |
1790 | PExtentVector logical; | |
1791 | ||
1792 | auto cct = coll->store->cct; | |
1793 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
1794 | << std::dec << " " << *this << dendl; | |
1795 | ||
1796 | bool empty = used_in_blob.put( | |
1797 | offset, | |
1798 | length, | |
1799 | &logical); | |
1800 | r->clear(); | |
1801 | // nothing to release | |
1802 | if (!empty && logical.empty()) { | |
1803 | return false; | |
1804 | } | |
1805 | ||
1806 | bluestore_blob_t& b = dirty_blob(); | |
1807 | return b.release_extents(empty, logical, r); | |
1808 | } | |
1809 | ||
224ce89b | 1810 | bool BlueStore::Blob::can_reuse_blob(uint32_t min_alloc_size, |
7c673cae FG |
1811 | uint32_t target_blob_size, |
1812 | uint32_t b_offset, | |
1813 | uint32_t *length0) { | |
1814 | assert(min_alloc_size); | |
1815 | assert(target_blob_size); | |
1816 | if (!get_blob().is_mutable()) { | |
1817 | return false; | |
1818 | } | |
1819 | ||
1820 | uint32_t length = *length0; | |
1821 | uint32_t end = b_offset + length; | |
1822 | ||
1823 | // Currently for the sake of simplicity we omit blob reuse if data is | |
1824 | // unaligned with csum chunk. Later we can perform padding if needed. | |
1825 | if (get_blob().has_csum() && | |
1826 | ((b_offset % get_blob().get_csum_chunk_size()) != 0 || | |
1827 | (end % get_blob().get_csum_chunk_size()) != 0)) { | |
1828 | return false; | |
1829 | } | |
1830 | ||
1831 | auto blen = get_blob().get_logical_length(); | |
1832 | uint32_t new_blen = blen; | |
1833 | ||
1834 | // make sure target_blob_size isn't less than current blob len | |
1835 | target_blob_size = MAX(blen, target_blob_size); | |
1836 | ||
1837 | if (b_offset >= blen) { | |
224ce89b WB |
1838 | // new data totally stands out of the existing blob |
1839 | new_blen = end; | |
7c673cae | 1840 | } else { |
224ce89b WB |
1841 | // new data overlaps with the existing blob |
1842 | new_blen = MAX(blen, end); | |
1843 | ||
1844 | uint32_t overlap = 0; | |
1845 | if (new_blen > blen) { | |
1846 | overlap = blen - b_offset; | |
1847 | } else { | |
1848 | overlap = length; | |
1849 | } | |
1850 | ||
1851 | if (!get_blob().is_unallocated(b_offset, overlap)) { | |
1852 | // abort if any piece of the overlap has already been allocated | |
1853 | return false; | |
7c673cae FG |
1854 | } |
1855 | } | |
224ce89b | 1856 | |
7c673cae FG |
1857 | if (new_blen > blen) { |
1858 | int64_t overflow = int64_t(new_blen) - target_blob_size; | |
1859 | // Unable to decrease the provided length to fit into max_blob_size | |
1860 | if (overflow >= length) { | |
1861 | return false; | |
1862 | } | |
1863 | ||
1864 | // FIXME: in some cases we could reduce unused resolution | |
1865 | if (get_blob().has_unused()) { | |
1866 | return false; | |
1867 | } | |
1868 | ||
1869 | if (overflow > 0) { | |
1870 | new_blen -= overflow; | |
1871 | length -= overflow; | |
1872 | *length0 = length; | |
1873 | } | |
224ce89b | 1874 | |
7c673cae FG |
1875 | if (new_blen > blen) { |
1876 | dirty_blob().add_tail(new_blen); | |
1877 | used_in_blob.add_tail(new_blen, | |
224ce89b | 1878 | get_blob().get_release_size(min_alloc_size)); |
7c673cae FG |
1879 | } |
1880 | } | |
1881 | return true; | |
1882 | } | |
1883 | ||
1884 | void BlueStore::Blob::split(Collection *coll, uint32_t blob_offset, Blob *r) | |
1885 | { | |
1886 | auto cct = coll->store->cct; //used by dout | |
1887 | dout(10) << __func__ << " 0x" << std::hex << blob_offset << std::dec | |
1888 | << " start " << *this << dendl; | |
1889 | assert(blob.can_split()); | |
1890 | assert(used_in_blob.can_split()); | |
1891 | bluestore_blob_t &lb = dirty_blob(); | |
1892 | bluestore_blob_t &rb = r->dirty_blob(); | |
1893 | ||
1894 | used_in_blob.split( | |
1895 | blob_offset, | |
1896 | &(r->used_in_blob)); | |
1897 | ||
1898 | lb.split(blob_offset, rb); | |
1899 | shared_blob->bc.split(shared_blob->get_cache(), blob_offset, r->shared_blob->bc); | |
1900 | ||
1901 | dout(10) << __func__ << " 0x" << std::hex << blob_offset << std::dec | |
1902 | << " finish " << *this << dendl; | |
1903 | dout(10) << __func__ << " 0x" << std::hex << blob_offset << std::dec | |
1904 | << " and " << *r << dendl; | |
1905 | } | |
1906 | ||
1907 | #ifndef CACHE_BLOB_BL | |
1908 | void BlueStore::Blob::decode( | |
1909 | Collection *coll, | |
1910 | bufferptr::iterator& p, | |
1911 | uint64_t struct_v, | |
1912 | uint64_t* sbid, | |
1913 | bool include_ref_map) | |
1914 | { | |
1915 | denc(blob, p, struct_v); | |
1916 | if (blob.is_shared()) { | |
1917 | denc(*sbid, p); | |
1918 | } | |
1919 | if (include_ref_map) { | |
1920 | if (struct_v > 1) { | |
1921 | used_in_blob.decode(p); | |
1922 | } else { | |
1923 | used_in_blob.clear(); | |
1924 | bluestore_extent_ref_map_t legacy_ref_map; | |
1925 | legacy_ref_map.decode(p); | |
1926 | for (auto r : legacy_ref_map.ref_map) { | |
1927 | get_ref( | |
1928 | coll, | |
1929 | r.first, | |
1930 | r.second.refs * r.second.length); | |
1931 | } | |
1932 | } | |
1933 | } | |
1934 | } | |
1935 | #endif | |
1936 | ||
1937 | // Extent | |
1938 | ||
1939 | ostream& operator<<(ostream& out, const BlueStore::Extent& e) | |
1940 | { | |
1941 | return out << std::hex << "0x" << e.logical_offset << "~" << e.length | |
1942 | << ": 0x" << e.blob_offset << "~" << e.length << std::dec | |
1943 | << " " << *e.blob; | |
1944 | } | |
1945 | ||
1946 | // OldExtent | |
1947 | BlueStore::OldExtent* BlueStore::OldExtent::create(CollectionRef c, | |
1948 | uint32_t lo, | |
1949 | uint32_t o, | |
1950 | uint32_t l, | |
1951 | BlobRef& b) { | |
1952 | OldExtent* oe = new OldExtent(lo, o, l, b); | |
1953 | b->put_ref(c.get(), o, l, &(oe->r)); | |
1954 | oe->blob_empty = b->get_referenced_bytes() == 0; | |
1955 | return oe; | |
1956 | } | |
1957 | ||
1958 | // ExtentMap | |
1959 | ||
1960 | #undef dout_prefix | |
1961 | #define dout_prefix *_dout << "bluestore.extentmap(" << this << ") " | |
1962 | ||
1963 | BlueStore::ExtentMap::ExtentMap(Onode *o) | |
1964 | : onode(o), | |
1965 | inline_bl( | |
1966 | o->c->store->cct->_conf->bluestore_extent_map_inline_shard_prealloc_size) { | |
1967 | } | |
1968 | ||
1969 | void BlueStore::ExtentMap::update(KeyValueDB::Transaction t, | |
1970 | bool force) | |
1971 | { | |
1972 | auto cct = onode->c->store->cct; //used by dout | |
1973 | dout(20) << __func__ << " " << onode->oid << (force ? " force" : "") << dendl; | |
1974 | if (onode->onode.extent_map_shards.empty()) { | |
1975 | if (inline_bl.length() == 0) { | |
1976 | unsigned n; | |
1977 | // we need to encode inline_bl to measure encoded length | |
1978 | bool never_happen = encode_some(0, OBJECT_MAX_SIZE, inline_bl, &n); | |
1979 | assert(!never_happen); | |
1980 | size_t len = inline_bl.length(); | |
1981 | dout(20) << __func__ << " inline shard " << len << " bytes from " << n | |
1982 | << " extents" << dendl; | |
1983 | if (!force && len > cct->_conf->bluestore_extent_map_shard_max_size) { | |
1984 | request_reshard(0, OBJECT_MAX_SIZE); | |
1985 | return; | |
1986 | } | |
1987 | } | |
1988 | // will persist in the onode key. | |
1989 | } else { | |
1990 | // pending shard update | |
1991 | struct dirty_shard_t { | |
1992 | Shard *shard; | |
1993 | bufferlist bl; | |
1994 | dirty_shard_t(Shard *s) : shard(s) {} | |
1995 | }; | |
1996 | vector<dirty_shard_t> encoded_shards; | |
1997 | // allocate slots for all shards in a single call instead of | |
1998 | // doing multiple allocations - one per each dirty shard | |
1999 | encoded_shards.reserve(shards.size()); | |
2000 | ||
2001 | auto p = shards.begin(); | |
2002 | auto prev_p = p; | |
2003 | while (p != shards.end()) { | |
31f18b77 | 2004 | assert(p->shard_info->offset >= prev_p->shard_info->offset); |
7c673cae FG |
2005 | auto n = p; |
2006 | ++n; | |
2007 | if (p->dirty) { | |
2008 | uint32_t endoff; | |
2009 | if (n == shards.end()) { | |
2010 | endoff = OBJECT_MAX_SIZE; | |
2011 | } else { | |
2012 | endoff = n->shard_info->offset; | |
2013 | } | |
2014 | encoded_shards.emplace_back(dirty_shard_t(&(*p))); | |
2015 | bufferlist& bl = encoded_shards.back().bl; | |
2016 | if (encode_some(p->shard_info->offset, endoff - p->shard_info->offset, | |
2017 | bl, &p->extents)) { | |
2018 | if (force) { | |
2019 | derr << __func__ << " encode_some needs reshard" << dendl; | |
2020 | assert(!force); | |
2021 | } | |
2022 | } | |
2023 | size_t len = bl.length(); | |
2024 | ||
2025 | dout(20) << __func__ << " shard 0x" << std::hex | |
2026 | << p->shard_info->offset << std::dec << " is " << len | |
2027 | << " bytes (was " << p->shard_info->bytes << ") from " | |
2028 | << p->extents << " extents" << dendl; | |
2029 | ||
2030 | if (!force) { | |
2031 | if (len > cct->_conf->bluestore_extent_map_shard_max_size) { | |
2032 | // we are big; reshard ourselves | |
2033 | request_reshard(p->shard_info->offset, endoff); | |
2034 | } | |
2035 | // avoid resharding the trailing shard, even if it is small | |
2036 | else if (n != shards.end() && | |
2037 | len < g_conf->bluestore_extent_map_shard_min_size) { | |
31f18b77 FG |
2038 | assert(endoff != OBJECT_MAX_SIZE); |
2039 | if (p == shards.begin()) { | |
2040 | // we are the first shard, combine with next shard | |
7c673cae | 2041 | request_reshard(p->shard_info->offset, endoff + 1); |
7c673cae | 2042 | } else { |
31f18b77 FG |
2043 | // combine either with the previous shard or the next, |
2044 | // whichever is smaller | |
7c673cae FG |
2045 | if (prev_p->shard_info->bytes > n->shard_info->bytes) { |
2046 | request_reshard(p->shard_info->offset, endoff + 1); | |
2047 | } else { | |
2048 | request_reshard(prev_p->shard_info->offset, endoff); | |
2049 | } | |
2050 | } | |
2051 | } | |
2052 | } | |
2053 | } | |
2054 | prev_p = p; | |
2055 | p = n; | |
2056 | } | |
2057 | if (needs_reshard()) { | |
2058 | return; | |
2059 | } | |
2060 | ||
2061 | // schedule DB update for dirty shards | |
2062 | string key; | |
2063 | for (auto& it : encoded_shards) { | |
2064 | it.shard->dirty = false; | |
2065 | it.shard->shard_info->bytes = it.bl.length(); | |
2066 | generate_extent_shard_key_and_apply( | |
2067 | onode->key, | |
2068 | it.shard->shard_info->offset, | |
2069 | &key, | |
2070 | [&](const string& final_key) { | |
2071 | t->set(PREFIX_OBJ, final_key, it.bl); | |
2072 | } | |
2073 | ); | |
2074 | } | |
2075 | } | |
2076 | } | |
2077 | ||
31f18b77 FG |
2078 | bid_t BlueStore::ExtentMap::allocate_spanning_blob_id() |
2079 | { | |
2080 | if (spanning_blob_map.empty()) | |
2081 | return 0; | |
2082 | bid_t bid = spanning_blob_map.rbegin()->first + 1; | |
2083 | // bid is valid and available. | |
2084 | if (bid >= 0) | |
2085 | return bid; | |
2086 | // Find next unused bid; | |
2087 | bid = rand() % (numeric_limits<bid_t>::max() + 1); | |
2088 | const auto begin_bid = bid; | |
2089 | do { | |
2090 | if (!spanning_blob_map.count(bid)) | |
2091 | return bid; | |
2092 | else { | |
2093 | bid++; | |
2094 | if (bid < 0) bid = 0; | |
2095 | } | |
2096 | } while (bid != begin_bid); | |
2097 | assert(0 == "no available blob id"); | |
2098 | } | |
2099 | ||
7c673cae FG |
2100 | void BlueStore::ExtentMap::reshard( |
2101 | KeyValueDB *db, | |
2102 | KeyValueDB::Transaction t) | |
2103 | { | |
2104 | auto cct = onode->c->store->cct; // used by dout | |
2105 | ||
2106 | dout(10) << __func__ << " 0x[" << std::hex << needs_reshard_begin << "," | |
2107 | << needs_reshard_end << ")" << std::dec | |
2108 | << " of " << onode->onode.extent_map_shards.size() | |
2109 | << " shards on " << onode->oid << dendl; | |
2110 | for (auto& p : spanning_blob_map) { | |
2111 | dout(20) << __func__ << " spanning blob " << p.first << " " << *p.second | |
2112 | << dendl; | |
2113 | } | |
2114 | // determine shard index range | |
2115 | unsigned si_begin = 0, si_end = 0; | |
2116 | if (!shards.empty()) { | |
2117 | while (si_begin + 1 < shards.size() && | |
2118 | shards[si_begin + 1].shard_info->offset <= needs_reshard_begin) { | |
2119 | ++si_begin; | |
2120 | } | |
2121 | needs_reshard_begin = shards[si_begin].shard_info->offset; | |
2122 | for (si_end = si_begin; si_end < shards.size(); ++si_end) { | |
2123 | if (shards[si_end].shard_info->offset >= needs_reshard_end) { | |
2124 | needs_reshard_end = shards[si_end].shard_info->offset; | |
2125 | break; | |
2126 | } | |
2127 | } | |
2128 | if (si_end == shards.size()) { | |
2129 | needs_reshard_end = OBJECT_MAX_SIZE; | |
2130 | } | |
2131 | dout(20) << __func__ << " shards [" << si_begin << "," << si_end << ")" | |
2132 | << " over 0x[" << std::hex << needs_reshard_begin << "," | |
2133 | << needs_reshard_end << ")" << std::dec << dendl; | |
2134 | } | |
2135 | ||
181888fb | 2136 | fault_range(db, needs_reshard_begin, (needs_reshard_end - needs_reshard_begin)); |
7c673cae FG |
2137 | |
2138 | // we may need to fault in a larger interval later must have all | |
2139 | // referring extents for spanning blobs loaded in order to have | |
2140 | // accurate use_tracker values. | |
2141 | uint32_t spanning_scan_begin = needs_reshard_begin; | |
2142 | uint32_t spanning_scan_end = needs_reshard_end; | |
2143 | ||
2144 | // remove old keys | |
2145 | string key; | |
2146 | for (unsigned i = si_begin; i < si_end; ++i) { | |
2147 | generate_extent_shard_key_and_apply( | |
2148 | onode->key, shards[i].shard_info->offset, &key, | |
2149 | [&](const string& final_key) { | |
2150 | t->rmkey(PREFIX_OBJ, final_key); | |
2151 | } | |
2152 | ); | |
2153 | } | |
2154 | ||
2155 | // calculate average extent size | |
2156 | unsigned bytes = 0; | |
2157 | unsigned extents = 0; | |
2158 | if (onode->onode.extent_map_shards.empty()) { | |
2159 | bytes = inline_bl.length(); | |
2160 | extents = extent_map.size(); | |
2161 | } else { | |
2162 | for (unsigned i = si_begin; i < si_end; ++i) { | |
2163 | bytes += shards[i].shard_info->bytes; | |
2164 | extents += shards[i].extents; | |
2165 | } | |
2166 | } | |
2167 | unsigned target = cct->_conf->bluestore_extent_map_shard_target_size; | |
2168 | unsigned slop = target * | |
2169 | cct->_conf->bluestore_extent_map_shard_target_size_slop; | |
2170 | unsigned extent_avg = bytes / MAX(1, extents); | |
2171 | dout(20) << __func__ << " extent_avg " << extent_avg << ", target " << target | |
2172 | << ", slop " << slop << dendl; | |
2173 | ||
2174 | // reshard | |
2175 | unsigned estimate = 0; | |
31f18b77 | 2176 | unsigned offset = needs_reshard_begin; |
7c673cae FG |
2177 | vector<bluestore_onode_t::shard_info> new_shard_info; |
2178 | unsigned max_blob_end = 0; | |
2179 | Extent dummy(needs_reshard_begin); | |
2180 | for (auto e = extent_map.lower_bound(dummy); | |
2181 | e != extent_map.end(); | |
2182 | ++e) { | |
2183 | if (e->logical_offset >= needs_reshard_end) { | |
2184 | break; | |
2185 | } | |
2186 | dout(30) << " extent " << *e << dendl; | |
2187 | ||
2188 | // disfavor shard boundaries that span a blob | |
2189 | bool would_span = (e->logical_offset < max_blob_end) || e->blob_offset; | |
2190 | if (estimate && | |
2191 | estimate + extent_avg > target + (would_span ? slop : 0)) { | |
2192 | // new shard | |
31f18b77 | 2193 | if (offset == needs_reshard_begin) { |
7c673cae FG |
2194 | new_shard_info.emplace_back(bluestore_onode_t::shard_info()); |
2195 | new_shard_info.back().offset = offset; | |
2196 | dout(20) << __func__ << " new shard 0x" << std::hex << offset | |
31f18b77 | 2197 | << std::dec << dendl; |
7c673cae FG |
2198 | } |
2199 | offset = e->logical_offset; | |
2200 | new_shard_info.emplace_back(bluestore_onode_t::shard_info()); | |
2201 | new_shard_info.back().offset = offset; | |
2202 | dout(20) << __func__ << " new shard 0x" << std::hex << offset | |
2203 | << std::dec << dendl; | |
2204 | estimate = 0; | |
2205 | } | |
2206 | estimate += extent_avg; | |
31f18b77 FG |
2207 | unsigned bs = e->blob_start(); |
2208 | if (bs < spanning_scan_begin) { | |
2209 | spanning_scan_begin = bs; | |
7c673cae FG |
2210 | } |
2211 | uint32_t be = e->blob_end(); | |
2212 | if (be > max_blob_end) { | |
2213 | max_blob_end = be; | |
2214 | } | |
2215 | if (be > spanning_scan_end) { | |
2216 | spanning_scan_end = be; | |
2217 | } | |
2218 | } | |
2219 | if (new_shard_info.empty() && (si_begin > 0 || | |
2220 | si_end < shards.size())) { | |
2221 | // we resharded a partial range; we must produce at least one output | |
2222 | // shard | |
2223 | new_shard_info.emplace_back(bluestore_onode_t::shard_info()); | |
2224 | new_shard_info.back().offset = needs_reshard_begin; | |
2225 | dout(20) << __func__ << " new shard 0x" << std::hex << needs_reshard_begin | |
2226 | << std::dec << " (singleton degenerate case)" << dendl; | |
2227 | } | |
2228 | ||
2229 | auto& sv = onode->onode.extent_map_shards; | |
2230 | dout(20) << __func__ << " new " << new_shard_info << dendl; | |
2231 | dout(20) << __func__ << " old " << sv << dendl; | |
2232 | if (sv.empty()) { | |
2233 | // no old shards to keep | |
2234 | sv.swap(new_shard_info); | |
2235 | init_shards(true, true); | |
2236 | } else { | |
2237 | // splice in new shards | |
2238 | sv.erase(sv.begin() + si_begin, sv.begin() + si_end); | |
2239 | shards.erase(shards.begin() + si_begin, shards.begin() + si_end); | |
2240 | sv.insert( | |
2241 | sv.begin() + si_begin, | |
2242 | new_shard_info.begin(), | |
2243 | new_shard_info.end()); | |
2244 | shards.insert(shards.begin() + si_begin, new_shard_info.size(), Shard()); | |
7c673cae | 2245 | si_end = si_begin + new_shard_info.size(); |
31f18b77 FG |
2246 | |
2247 | assert(sv.size() == shards.size()); | |
2248 | ||
2249 | // note that we need to update every shard_info of shards here, | |
2250 | // as sv might have been totally re-allocated above | |
2251 | for (unsigned i = 0; i < shards.size(); i++) { | |
7c673cae | 2252 | shards[i].shard_info = &sv[i]; |
31f18b77 FG |
2253 | } |
2254 | ||
2255 | // mark newly added shards as dirty | |
2256 | for (unsigned i = si_begin; i < si_end; ++i) { | |
7c673cae FG |
2257 | shards[i].loaded = true; |
2258 | shards[i].dirty = true; | |
2259 | } | |
7c673cae FG |
2260 | } |
2261 | dout(20) << __func__ << " fin " << sv << dendl; | |
2262 | inline_bl.clear(); | |
2263 | ||
2264 | if (sv.empty()) { | |
2265 | // no more shards; unspan all previously spanning blobs | |
2266 | auto p = spanning_blob_map.begin(); | |
2267 | while (p != spanning_blob_map.end()) { | |
2268 | p->second->id = -1; | |
2269 | dout(30) << __func__ << " un-spanning " << *p->second << dendl; | |
2270 | p = spanning_blob_map.erase(p); | |
2271 | } | |
2272 | } else { | |
2273 | // identify new spanning blobs | |
2274 | dout(20) << __func__ << " checking spanning blobs 0x[" << std::hex | |
2275 | << spanning_scan_begin << "," << spanning_scan_end << ")" << dendl; | |
2276 | if (spanning_scan_begin < needs_reshard_begin) { | |
2277 | fault_range(db, spanning_scan_begin, | |
2278 | needs_reshard_begin - spanning_scan_begin); | |
2279 | } | |
2280 | if (spanning_scan_end > needs_reshard_end) { | |
2281 | fault_range(db, needs_reshard_end, | |
31f18b77 | 2282 | spanning_scan_end - needs_reshard_end); |
7c673cae FG |
2283 | } |
2284 | auto sp = sv.begin() + si_begin; | |
2285 | auto esp = sv.end(); | |
2286 | unsigned shard_start = sp->offset; | |
2287 | unsigned shard_end; | |
2288 | ++sp; | |
2289 | if (sp == esp) { | |
2290 | shard_end = OBJECT_MAX_SIZE; | |
2291 | } else { | |
2292 | shard_end = sp->offset; | |
2293 | } | |
7c673cae FG |
2294 | Extent dummy(needs_reshard_begin); |
2295 | for (auto e = extent_map.lower_bound(dummy); e != extent_map.end(); ++e) { | |
2296 | if (e->logical_offset >= needs_reshard_end) { | |
2297 | break; | |
2298 | } | |
2299 | dout(30) << " extent " << *e << dendl; | |
2300 | while (e->logical_offset >= shard_end) { | |
2301 | shard_start = shard_end; | |
2302 | assert(sp != esp); | |
2303 | ++sp; | |
2304 | if (sp == esp) { | |
2305 | shard_end = OBJECT_MAX_SIZE; | |
2306 | } else { | |
2307 | shard_end = sp->offset; | |
2308 | } | |
2309 | dout(30) << __func__ << " shard 0x" << std::hex << shard_start | |
2310 | << " to 0x" << shard_end << std::dec << dendl; | |
2311 | } | |
2312 | if (e->blob_escapes_range(shard_start, shard_end - shard_start)) { | |
2313 | if (!e->blob->is_spanning()) { | |
2314 | // We have two options: (1) split the blob into pieces at the | |
2315 | // shard boundaries (and adjust extents accordingly), or (2) | |
2316 | // mark it spanning. We prefer to cut the blob if we can. Note that | |
2317 | // we may have to split it multiple times--potentially at every | |
2318 | // shard boundary. | |
2319 | bool must_span = false; | |
2320 | BlobRef b = e->blob; | |
2321 | if (b->can_split()) { | |
2322 | uint32_t bstart = e->blob_start(); | |
2323 | uint32_t bend = e->blob_end(); | |
2324 | for (const auto& sh : shards) { | |
2325 | if (bstart < sh.shard_info->offset && | |
2326 | bend > sh.shard_info->offset) { | |
2327 | uint32_t blob_offset = sh.shard_info->offset - bstart; | |
2328 | if (b->can_split_at(blob_offset)) { | |
2329 | dout(20) << __func__ << " splitting blob, bstart 0x" | |
2330 | << std::hex << bstart << " blob_offset 0x" | |
2331 | << blob_offset << std::dec << " " << *b << dendl; | |
2332 | b = split_blob(b, blob_offset, sh.shard_info->offset); | |
2333 | // switch b to the new right-hand side, in case it | |
2334 | // *also* has to get split. | |
2335 | bstart += blob_offset; | |
2336 | onode->c->store->logger->inc(l_bluestore_blob_split); | |
2337 | } else { | |
2338 | must_span = true; | |
2339 | break; | |
2340 | } | |
2341 | } | |
2342 | } | |
2343 | } else { | |
2344 | must_span = true; | |
2345 | } | |
2346 | if (must_span) { | |
31f18b77 FG |
2347 | auto bid = allocate_spanning_blob_id(); |
2348 | b->id = bid; | |
7c673cae FG |
2349 | spanning_blob_map[b->id] = b; |
2350 | dout(20) << __func__ << " adding spanning " << *b << dendl; | |
2351 | } | |
2352 | } | |
2353 | } else { | |
2354 | if (e->blob->is_spanning()) { | |
2355 | spanning_blob_map.erase(e->blob->id); | |
2356 | e->blob->id = -1; | |
2357 | dout(30) << __func__ << " un-spanning " << *e->blob << dendl; | |
2358 | } | |
2359 | } | |
2360 | } | |
2361 | } | |
2362 | ||
2363 | clear_needs_reshard(); | |
2364 | } | |
2365 | ||
2366 | bool BlueStore::ExtentMap::encode_some( | |
2367 | uint32_t offset, | |
2368 | uint32_t length, | |
2369 | bufferlist& bl, | |
2370 | unsigned *pn) | |
2371 | { | |
2372 | auto cct = onode->c->store->cct; //used by dout | |
2373 | Extent dummy(offset); | |
2374 | auto start = extent_map.lower_bound(dummy); | |
2375 | uint32_t end = offset + length; | |
2376 | ||
2377 | __u8 struct_v = 2; // Version 2 differs from v1 in blob's ref_map | |
2378 | // serialization only. Hence there is no specific | |
2379 | // handling at ExtentMap level. | |
2380 | ||
2381 | unsigned n = 0; | |
2382 | size_t bound = 0; | |
7c673cae FG |
2383 | bool must_reshard = false; |
2384 | for (auto p = start; | |
2385 | p != extent_map.end() && p->logical_offset < end; | |
2386 | ++p, ++n) { | |
2387 | assert(p->logical_offset >= offset); | |
2388 | p->blob->last_encoded_id = -1; | |
2389 | if (!p->blob->is_spanning() && p->blob_escapes_range(offset, length)) { | |
2390 | dout(30) << __func__ << " 0x" << std::hex << offset << "~" << length | |
2391 | << std::dec << " hit new spanning blob " << *p << dendl; | |
2392 | request_reshard(p->blob_start(), p->blob_end()); | |
2393 | must_reshard = true; | |
2394 | } | |
31f18b77 FG |
2395 | if (!must_reshard) { |
2396 | denc_varint(0, bound); // blobid | |
2397 | denc_varint(0, bound); // logical_offset | |
2398 | denc_varint(0, bound); // len | |
2399 | denc_varint(0, bound); // blob_offset | |
7c673cae | 2400 | |
31f18b77 FG |
2401 | p->blob->bound_encode( |
2402 | bound, | |
2403 | struct_v, | |
2404 | p->blob->shared_blob->get_sbid(), | |
2405 | false); | |
2406 | } | |
7c673cae FG |
2407 | } |
2408 | if (must_reshard) { | |
2409 | return true; | |
2410 | } | |
2411 | ||
31f18b77 FG |
2412 | denc(struct_v, bound); |
2413 | denc_varint(0, bound); // number of extents | |
2414 | ||
7c673cae FG |
2415 | { |
2416 | auto app = bl.get_contiguous_appender(bound); | |
2417 | denc(struct_v, app); | |
2418 | denc_varint(n, app); | |
2419 | if (pn) { | |
2420 | *pn = n; | |
2421 | } | |
2422 | ||
2423 | n = 0; | |
2424 | uint64_t pos = 0; | |
2425 | uint64_t prev_len = 0; | |
2426 | for (auto p = start; | |
2427 | p != extent_map.end() && p->logical_offset < end; | |
2428 | ++p, ++n) { | |
2429 | unsigned blobid; | |
2430 | bool include_blob = false; | |
2431 | if (p->blob->is_spanning()) { | |
2432 | blobid = p->blob->id << BLOBID_SHIFT_BITS; | |
2433 | blobid |= BLOBID_FLAG_SPANNING; | |
2434 | } else if (p->blob->last_encoded_id < 0) { | |
2435 | p->blob->last_encoded_id = n + 1; // so it is always non-zero | |
2436 | include_blob = true; | |
2437 | blobid = 0; // the decoder will infer the id from n | |
2438 | } else { | |
2439 | blobid = p->blob->last_encoded_id << BLOBID_SHIFT_BITS; | |
2440 | } | |
2441 | if (p->logical_offset == pos) { | |
2442 | blobid |= BLOBID_FLAG_CONTIGUOUS; | |
2443 | } | |
2444 | if (p->blob_offset == 0) { | |
2445 | blobid |= BLOBID_FLAG_ZEROOFFSET; | |
2446 | } | |
2447 | if (p->length == prev_len) { | |
2448 | blobid |= BLOBID_FLAG_SAMELENGTH; | |
2449 | } else { | |
2450 | prev_len = p->length; | |
2451 | } | |
2452 | denc_varint(blobid, app); | |
2453 | if ((blobid & BLOBID_FLAG_CONTIGUOUS) == 0) { | |
2454 | denc_varint_lowz(p->logical_offset - pos, app); | |
2455 | } | |
2456 | if ((blobid & BLOBID_FLAG_ZEROOFFSET) == 0) { | |
2457 | denc_varint_lowz(p->blob_offset, app); | |
2458 | } | |
2459 | if ((blobid & BLOBID_FLAG_SAMELENGTH) == 0) { | |
2460 | denc_varint_lowz(p->length, app); | |
2461 | } | |
2462 | pos = p->logical_end(); | |
2463 | if (include_blob) { | |
2464 | p->blob->encode(app, struct_v, p->blob->shared_blob->get_sbid(), false); | |
2465 | } | |
2466 | } | |
2467 | } | |
2468 | /*derr << __func__ << bl << dendl; | |
2469 | derr << __func__ << ":"; | |
2470 | bl.hexdump(*_dout); | |
2471 | *_dout << dendl; | |
2472 | */ | |
2473 | return false; | |
2474 | } | |
2475 | ||
2476 | unsigned BlueStore::ExtentMap::decode_some(bufferlist& bl) | |
2477 | { | |
2478 | auto cct = onode->c->store->cct; //used by dout | |
2479 | /* | |
2480 | derr << __func__ << ":"; | |
2481 | bl.hexdump(*_dout); | |
2482 | *_dout << dendl; | |
2483 | */ | |
2484 | ||
2485 | assert(bl.get_num_buffers() <= 1); | |
2486 | auto p = bl.front().begin_deep(); | |
2487 | __u8 struct_v; | |
2488 | denc(struct_v, p); | |
2489 | // Version 2 differs from v1 in blob's ref_map | |
2490 | // serialization only. Hence there is no specific | |
2491 | // handling at ExtentMap level below. | |
2492 | assert(struct_v == 1 || struct_v == 2); | |
2493 | ||
2494 | uint32_t num; | |
2495 | denc_varint(num, p); | |
2496 | vector<BlobRef> blobs(num); | |
2497 | uint64_t pos = 0; | |
2498 | uint64_t prev_len = 0; | |
2499 | unsigned n = 0; | |
2500 | ||
2501 | while (!p.end()) { | |
2502 | Extent *le = new Extent(); | |
2503 | uint64_t blobid; | |
2504 | denc_varint(blobid, p); | |
2505 | if ((blobid & BLOBID_FLAG_CONTIGUOUS) == 0) { | |
2506 | uint64_t gap; | |
2507 | denc_varint_lowz(gap, p); | |
2508 | pos += gap; | |
2509 | } | |
2510 | le->logical_offset = pos; | |
2511 | if ((blobid & BLOBID_FLAG_ZEROOFFSET) == 0) { | |
2512 | denc_varint_lowz(le->blob_offset, p); | |
2513 | } else { | |
2514 | le->blob_offset = 0; | |
2515 | } | |
2516 | if ((blobid & BLOBID_FLAG_SAMELENGTH) == 0) { | |
2517 | denc_varint_lowz(prev_len, p); | |
2518 | } | |
2519 | le->length = prev_len; | |
2520 | ||
2521 | if (blobid & BLOBID_FLAG_SPANNING) { | |
2522 | dout(30) << __func__ << " getting spanning blob " | |
2523 | << (blobid >> BLOBID_SHIFT_BITS) << dendl; | |
2524 | le->assign_blob(get_spanning_blob(blobid >> BLOBID_SHIFT_BITS)); | |
2525 | } else { | |
2526 | blobid >>= BLOBID_SHIFT_BITS; | |
2527 | if (blobid) { | |
2528 | le->assign_blob(blobs[blobid - 1]); | |
2529 | assert(le->blob); | |
2530 | } else { | |
2531 | Blob *b = new Blob(); | |
2532 | uint64_t sbid = 0; | |
2533 | b->decode(onode->c, p, struct_v, &sbid, false); | |
2534 | blobs[n] = b; | |
2535 | onode->c->open_shared_blob(sbid, b); | |
2536 | le->assign_blob(b); | |
2537 | } | |
2538 | // we build ref_map dynamically for non-spanning blobs | |
2539 | le->blob->get_ref( | |
2540 | onode->c, | |
2541 | le->blob_offset, | |
2542 | le->length); | |
2543 | } | |
2544 | pos += prev_len; | |
2545 | ++n; | |
2546 | extent_map.insert(*le); | |
2547 | } | |
2548 | ||
2549 | assert(n == num); | |
2550 | return num; | |
2551 | } | |
2552 | ||
2553 | void BlueStore::ExtentMap::bound_encode_spanning_blobs(size_t& p) | |
2554 | { | |
2555 | // Version 2 differs from v1 in blob's ref_map | |
2556 | // serialization only. Hence there is no specific | |
2557 | // handling at ExtentMap level. | |
2558 | __u8 struct_v = 2; | |
2559 | ||
2560 | denc(struct_v, p); | |
2561 | denc_varint((uint32_t)0, p); | |
2562 | size_t key_size = 0; | |
2563 | denc_varint((uint32_t)0, key_size); | |
2564 | p += spanning_blob_map.size() * key_size; | |
2565 | for (const auto& i : spanning_blob_map) { | |
2566 | i.second->bound_encode(p, struct_v, i.second->shared_blob->get_sbid(), true); | |
2567 | } | |
2568 | } | |
2569 | ||
2570 | void BlueStore::ExtentMap::encode_spanning_blobs( | |
2571 | bufferlist::contiguous_appender& p) | |
2572 | { | |
2573 | // Version 2 differs from v1 in blob's ref_map | |
2574 | // serialization only. Hence there is no specific | |
2575 | // handling at ExtentMap level. | |
2576 | __u8 struct_v = 2; | |
2577 | ||
2578 | denc(struct_v, p); | |
2579 | denc_varint(spanning_blob_map.size(), p); | |
2580 | for (auto& i : spanning_blob_map) { | |
2581 | denc_varint(i.second->id, p); | |
2582 | i.second->encode(p, struct_v, i.second->shared_blob->get_sbid(), true); | |
2583 | } | |
2584 | } | |
2585 | ||
2586 | void BlueStore::ExtentMap::decode_spanning_blobs( | |
2587 | bufferptr::iterator& p) | |
2588 | { | |
2589 | __u8 struct_v; | |
2590 | denc(struct_v, p); | |
2591 | // Version 2 differs from v1 in blob's ref_map | |
2592 | // serialization only. Hence there is no specific | |
2593 | // handling at ExtentMap level. | |
2594 | assert(struct_v == 1 || struct_v == 2); | |
2595 | ||
2596 | unsigned n; | |
2597 | denc_varint(n, p); | |
2598 | while (n--) { | |
2599 | BlobRef b(new Blob()); | |
2600 | denc_varint(b->id, p); | |
2601 | spanning_blob_map[b->id] = b; | |
2602 | uint64_t sbid = 0; | |
2603 | b->decode(onode->c, p, struct_v, &sbid, true); | |
2604 | onode->c->open_shared_blob(sbid, b); | |
2605 | } | |
2606 | } | |
2607 | ||
2608 | void BlueStore::ExtentMap::init_shards(bool loaded, bool dirty) | |
2609 | { | |
2610 | shards.resize(onode->onode.extent_map_shards.size()); | |
2611 | unsigned i = 0; | |
2612 | for (auto &s : onode->onode.extent_map_shards) { | |
2613 | shards[i].shard_info = &s; | |
2614 | shards[i].loaded = loaded; | |
2615 | shards[i].dirty = dirty; | |
2616 | ++i; | |
2617 | } | |
2618 | } | |
2619 | ||
2620 | void BlueStore::ExtentMap::fault_range( | |
2621 | KeyValueDB *db, | |
2622 | uint32_t offset, | |
2623 | uint32_t length) | |
2624 | { | |
2625 | auto cct = onode->c->store->cct; //used by dout | |
2626 | dout(30) << __func__ << " 0x" << std::hex << offset << "~" << length | |
2627 | << std::dec << dendl; | |
2628 | auto start = seek_shard(offset); | |
2629 | auto last = seek_shard(offset + length); | |
2630 | ||
2631 | if (start < 0) | |
2632 | return; | |
2633 | ||
2634 | assert(last >= start); | |
2635 | string key; | |
2636 | while (start <= last) { | |
2637 | assert((size_t)start < shards.size()); | |
2638 | auto p = &shards[start]; | |
2639 | if (!p->loaded) { | |
2640 | dout(30) << __func__ << " opening shard 0x" << std::hex | |
2641 | << p->shard_info->offset << std::dec << dendl; | |
2642 | bufferlist v; | |
2643 | generate_extent_shard_key_and_apply( | |
2644 | onode->key, p->shard_info->offset, &key, | |
2645 | [&](const string& final_key) { | |
2646 | int r = db->get(PREFIX_OBJ, final_key, &v); | |
2647 | if (r < 0) { | |
2648 | derr << __func__ << " missing shard 0x" << std::hex | |
2649 | << p->shard_info->offset << std::dec << " for " << onode->oid | |
2650 | << dendl; | |
2651 | assert(r >= 0); | |
2652 | } | |
2653 | } | |
2654 | ); | |
2655 | p->extents = decode_some(v); | |
2656 | p->loaded = true; | |
2657 | dout(20) << __func__ << " open shard 0x" << std::hex | |
2658 | << p->shard_info->offset << std::dec | |
2659 | << " (" << v.length() << " bytes)" << dendl; | |
2660 | assert(p->dirty == false); | |
2661 | assert(v.length() == p->shard_info->bytes); | |
2662 | onode->c->store->logger->inc(l_bluestore_onode_shard_misses); | |
2663 | } else { | |
2664 | onode->c->store->logger->inc(l_bluestore_onode_shard_hits); | |
2665 | } | |
2666 | ++start; | |
2667 | } | |
2668 | } | |
2669 | ||
2670 | void BlueStore::ExtentMap::dirty_range( | |
7c673cae FG |
2671 | uint32_t offset, |
2672 | uint32_t length) | |
2673 | { | |
2674 | auto cct = onode->c->store->cct; //used by dout | |
2675 | dout(30) << __func__ << " 0x" << std::hex << offset << "~" << length | |
2676 | << std::dec << dendl; | |
2677 | if (shards.empty()) { | |
2678 | dout(20) << __func__ << " mark inline shard dirty" << dendl; | |
2679 | inline_bl.clear(); | |
2680 | return; | |
2681 | } | |
2682 | auto start = seek_shard(offset); | |
2683 | auto last = seek_shard(offset + length); | |
2684 | if (start < 0) | |
2685 | return; | |
2686 | ||
2687 | assert(last >= start); | |
2688 | while (start <= last) { | |
2689 | assert((size_t)start < shards.size()); | |
2690 | auto p = &shards[start]; | |
2691 | if (!p->loaded) { | |
2692 | dout(20) << __func__ << " shard 0x" << std::hex << p->shard_info->offset | |
2693 | << std::dec << " is not loaded, can't mark dirty" << dendl; | |
2694 | assert(0 == "can't mark unloaded shard dirty"); | |
2695 | } | |
2696 | if (!p->dirty) { | |
2697 | dout(20) << __func__ << " mark shard 0x" << std::hex | |
2698 | << p->shard_info->offset << std::dec << " dirty" << dendl; | |
2699 | p->dirty = true; | |
2700 | } | |
2701 | ++start; | |
2702 | } | |
2703 | } | |
2704 | ||
2705 | BlueStore::extent_map_t::iterator BlueStore::ExtentMap::find( | |
2706 | uint64_t offset) | |
2707 | { | |
2708 | Extent dummy(offset); | |
2709 | return extent_map.find(dummy); | |
2710 | } | |
2711 | ||
7c673cae FG |
2712 | BlueStore::extent_map_t::iterator BlueStore::ExtentMap::seek_lextent( |
2713 | uint64_t offset) | |
2714 | { | |
2715 | Extent dummy(offset); | |
2716 | auto fp = extent_map.lower_bound(dummy); | |
2717 | if (fp != extent_map.begin()) { | |
2718 | --fp; | |
2719 | if (fp->logical_end() <= offset) { | |
2720 | ++fp; | |
2721 | } | |
2722 | } | |
2723 | return fp; | |
2724 | } | |
2725 | ||
2726 | BlueStore::extent_map_t::const_iterator BlueStore::ExtentMap::seek_lextent( | |
2727 | uint64_t offset) const | |
2728 | { | |
2729 | Extent dummy(offset); | |
2730 | auto fp = extent_map.lower_bound(dummy); | |
2731 | if (fp != extent_map.begin()) { | |
2732 | --fp; | |
2733 | if (fp->logical_end() <= offset) { | |
2734 | ++fp; | |
2735 | } | |
2736 | } | |
2737 | return fp; | |
2738 | } | |
2739 | ||
2740 | bool BlueStore::ExtentMap::has_any_lextents(uint64_t offset, uint64_t length) | |
2741 | { | |
2742 | auto fp = seek_lextent(offset); | |
2743 | if (fp == extent_map.end() || fp->logical_offset >= offset + length) { | |
2744 | return false; | |
2745 | } | |
2746 | return true; | |
2747 | } | |
2748 | ||
2749 | int BlueStore::ExtentMap::compress_extent_map( | |
2750 | uint64_t offset, | |
2751 | uint64_t length) | |
2752 | { | |
2753 | auto cct = onode->c->store->cct; //used by dout | |
2754 | if (extent_map.empty()) | |
2755 | return 0; | |
2756 | int removed = 0; | |
2757 | auto p = seek_lextent(offset); | |
2758 | if (p != extent_map.begin()) { | |
2759 | --p; // start to the left of offset | |
2760 | } | |
2761 | // the caller should have just written to this region | |
2762 | assert(p != extent_map.end()); | |
2763 | ||
2764 | // identify the *next* shard | |
2765 | auto pshard = shards.begin(); | |
2766 | while (pshard != shards.end() && | |
2767 | p->logical_offset >= pshard->shard_info->offset) { | |
2768 | ++pshard; | |
2769 | } | |
2770 | uint64_t shard_end; | |
2771 | if (pshard != shards.end()) { | |
2772 | shard_end = pshard->shard_info->offset; | |
2773 | } else { | |
2774 | shard_end = OBJECT_MAX_SIZE; | |
2775 | } | |
2776 | ||
2777 | auto n = p; | |
2778 | for (++n; n != extent_map.end(); p = n++) { | |
2779 | if (n->logical_offset > offset + length) { | |
2780 | break; // stop after end | |
2781 | } | |
2782 | while (n != extent_map.end() && | |
2783 | p->logical_end() == n->logical_offset && | |
2784 | p->blob == n->blob && | |
2785 | p->blob_offset + p->length == n->blob_offset && | |
2786 | n->logical_offset < shard_end) { | |
2787 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
2788 | << " next shard 0x" << shard_end << std::dec | |
2789 | << " merging " << *p << " and " << *n << dendl; | |
2790 | p->length += n->length; | |
2791 | rm(n++); | |
2792 | ++removed; | |
2793 | } | |
2794 | if (n == extent_map.end()) { | |
2795 | break; | |
2796 | } | |
2797 | if (n->logical_offset >= shard_end) { | |
2798 | assert(pshard != shards.end()); | |
2799 | ++pshard; | |
2800 | if (pshard != shards.end()) { | |
2801 | shard_end = pshard->shard_info->offset; | |
2802 | } else { | |
2803 | shard_end = OBJECT_MAX_SIZE; | |
2804 | } | |
2805 | } | |
2806 | } | |
2807 | if (removed && onode) { | |
2808 | onode->c->store->logger->inc(l_bluestore_extent_compress, removed); | |
2809 | } | |
2810 | return removed; | |
2811 | } | |
2812 | ||
2813 | void BlueStore::ExtentMap::punch_hole( | |
2814 | CollectionRef &c, | |
2815 | uint64_t offset, | |
2816 | uint64_t length, | |
2817 | old_extent_map_t *old_extents) | |
2818 | { | |
2819 | auto p = seek_lextent(offset); | |
2820 | uint64_t end = offset + length; | |
2821 | while (p != extent_map.end()) { | |
2822 | if (p->logical_offset >= end) { | |
2823 | break; | |
2824 | } | |
2825 | if (p->logical_offset < offset) { | |
2826 | if (p->logical_end() > end) { | |
2827 | // split and deref middle | |
2828 | uint64_t front = offset - p->logical_offset; | |
2829 | OldExtent* oe = OldExtent::create(c, offset, p->blob_offset + front, | |
2830 | length, p->blob); | |
2831 | old_extents->push_back(*oe); | |
2832 | add(end, | |
2833 | p->blob_offset + front + length, | |
2834 | p->length - front - length, | |
2835 | p->blob); | |
2836 | p->length = front; | |
2837 | break; | |
2838 | } else { | |
2839 | // deref tail | |
2840 | assert(p->logical_end() > offset); // else seek_lextent bug | |
2841 | uint64_t keep = offset - p->logical_offset; | |
2842 | OldExtent* oe = OldExtent::create(c, offset, p->blob_offset + keep, | |
2843 | p->length - keep, p->blob); | |
2844 | old_extents->push_back(*oe); | |
2845 | p->length = keep; | |
2846 | ++p; | |
2847 | continue; | |
2848 | } | |
2849 | } | |
2850 | if (p->logical_offset + p->length <= end) { | |
2851 | // deref whole lextent | |
2852 | OldExtent* oe = OldExtent::create(c, p->logical_offset, p->blob_offset, | |
2853 | p->length, p->blob); | |
2854 | old_extents->push_back(*oe); | |
2855 | rm(p++); | |
2856 | continue; | |
2857 | } | |
2858 | // deref head | |
2859 | uint64_t keep = p->logical_end() - end; | |
2860 | OldExtent* oe = OldExtent::create(c, p->logical_offset, p->blob_offset, | |
2861 | p->length - keep, p->blob); | |
2862 | old_extents->push_back(*oe); | |
2863 | ||
2864 | add(end, p->blob_offset + p->length - keep, keep, p->blob); | |
2865 | rm(p); | |
2866 | break; | |
2867 | } | |
2868 | } | |
2869 | ||
2870 | BlueStore::Extent *BlueStore::ExtentMap::set_lextent( | |
2871 | CollectionRef &c, | |
2872 | uint64_t logical_offset, | |
2873 | uint64_t blob_offset, uint64_t length, BlobRef b, | |
2874 | old_extent_map_t *old_extents) | |
2875 | { | |
2876 | // We need to have completely initialized Blob to increment its ref counters. | |
2877 | assert(b->get_blob().get_logical_length() != 0); | |
2878 | ||
2879 | // Do get_ref prior to punch_hole to prevent from putting reused blob into | |
2880 | // old_extents list if we overwre the blob totally | |
2881 | // This might happen during WAL overwrite. | |
2882 | b->get_ref(onode->c, blob_offset, length); | |
2883 | ||
2884 | if (old_extents) { | |
2885 | punch_hole(c, logical_offset, length, old_extents); | |
2886 | } | |
2887 | ||
2888 | Extent *le = new Extent(logical_offset, blob_offset, length, b); | |
2889 | extent_map.insert(*le); | |
2890 | if (spans_shard(logical_offset, length)) { | |
2891 | request_reshard(logical_offset, logical_offset + length); | |
2892 | } | |
2893 | return le; | |
2894 | } | |
2895 | ||
2896 | BlueStore::BlobRef BlueStore::ExtentMap::split_blob( | |
2897 | BlobRef lb, | |
2898 | uint32_t blob_offset, | |
2899 | uint32_t pos) | |
2900 | { | |
2901 | auto cct = onode->c->store->cct; //used by dout | |
2902 | ||
2903 | uint32_t end_pos = pos + lb->get_blob().get_logical_length() - blob_offset; | |
2904 | dout(20) << __func__ << " 0x" << std::hex << pos << " end 0x" << end_pos | |
2905 | << " blob_offset 0x" << blob_offset << std::dec << " " << *lb | |
2906 | << dendl; | |
2907 | BlobRef rb = onode->c->new_blob(); | |
2908 | lb->split(onode->c, blob_offset, rb.get()); | |
2909 | ||
2910 | for (auto ep = seek_lextent(pos); | |
2911 | ep != extent_map.end() && ep->logical_offset < end_pos; | |
2912 | ++ep) { | |
2913 | if (ep->blob != lb) { | |
2914 | continue; | |
2915 | } | |
2916 | if (ep->logical_offset < pos) { | |
2917 | // split extent | |
2918 | size_t left = pos - ep->logical_offset; | |
2919 | Extent *ne = new Extent(pos, 0, ep->length - left, rb); | |
2920 | extent_map.insert(*ne); | |
2921 | ep->length = left; | |
2922 | dout(30) << __func__ << " split " << *ep << dendl; | |
2923 | dout(30) << __func__ << " to " << *ne << dendl; | |
2924 | } else { | |
2925 | // switch blob | |
2926 | assert(ep->blob_offset >= blob_offset); | |
2927 | ||
2928 | ep->blob = rb; | |
2929 | ep->blob_offset -= blob_offset; | |
2930 | dout(30) << __func__ << " adjusted " << *ep << dendl; | |
2931 | } | |
2932 | } | |
2933 | return rb; | |
2934 | } | |
2935 | ||
2936 | // Onode | |
2937 | ||
2938 | #undef dout_prefix | |
2939 | #define dout_prefix *_dout << "bluestore.onode(" << this << ")." << __func__ << " " | |
2940 | ||
2941 | void BlueStore::Onode::flush() | |
2942 | { | |
2943 | if (flushing_count.load()) { | |
2944 | ldout(c->store->cct, 20) << __func__ << " cnt:" << flushing_count << dendl; | |
2945 | std::unique_lock<std::mutex> l(flush_lock); | |
2946 | while (flushing_count.load()) { | |
2947 | flush_cond.wait(l); | |
2948 | } | |
2949 | } | |
2950 | ldout(c->store->cct, 20) << __func__ << " done" << dendl; | |
2951 | } | |
2952 | ||
2953 | // ======================================================= | |
2954 | // WriteContext | |
2955 | ||
2956 | /// Checks for writes to the same pextent within a blob | |
2957 | bool BlueStore::WriteContext::has_conflict( | |
2958 | BlobRef b, | |
2959 | uint64_t loffs, | |
2960 | uint64_t loffs_end, | |
2961 | uint64_t min_alloc_size) | |
2962 | { | |
2963 | assert((loffs % min_alloc_size) == 0); | |
2964 | assert((loffs_end % min_alloc_size) == 0); | |
2965 | for (auto w : writes) { | |
2966 | if (b == w.b) { | |
2967 | auto loffs2 = P2ALIGN(w.logical_offset, min_alloc_size); | |
224ce89b | 2968 | auto loffs2_end = P2ROUNDUP(w.logical_offset + w.length0, min_alloc_size); |
7c673cae | 2969 | if ((loffs <= loffs2 && loffs_end > loffs2) || |
224ce89b | 2970 | (loffs >= loffs2 && loffs < loffs2_end)) { |
7c673cae FG |
2971 | return true; |
2972 | } | |
2973 | } | |
2974 | } | |
2975 | return false; | |
2976 | } | |
2977 | ||
2978 | // ======================================================= | |
2979 | ||
2980 | // DeferredBatch | |
2981 | #undef dout_prefix | |
2982 | #define dout_prefix *_dout << "bluestore.DeferredBatch(" << this << ") " | |
2983 | ||
2984 | void BlueStore::DeferredBatch::prepare_write( | |
2985 | CephContext *cct, | |
2986 | uint64_t seq, uint64_t offset, uint64_t length, | |
2987 | bufferlist::const_iterator& blp) | |
2988 | { | |
2989 | _discard(cct, offset, length); | |
2990 | auto i = iomap.insert(make_pair(offset, deferred_io())); | |
2991 | assert(i.second); // this should be a new insertion | |
2992 | i.first->second.seq = seq; | |
2993 | blp.copy(length, i.first->second.bl); | |
31f18b77 FG |
2994 | i.first->second.bl.reassign_to_mempool( |
2995 | mempool::mempool_bluestore_writing_deferred); | |
7c673cae FG |
2996 | dout(20) << __func__ << " seq " << seq |
2997 | << " 0x" << std::hex << offset << "~" << length | |
2998 | << " crc " << i.first->second.bl.crc32c(-1) | |
2999 | << std::dec << dendl; | |
3000 | seq_bytes[seq] += length; | |
3001 | #ifdef DEBUG_DEFERRED | |
3002 | _audit(cct); | |
3003 | #endif | |
3004 | } | |
3005 | ||
3006 | void BlueStore::DeferredBatch::_discard( | |
3007 | CephContext *cct, uint64_t offset, uint64_t length) | |
3008 | { | |
3009 | generic_dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
3010 | << std::dec << dendl; | |
3011 | auto p = iomap.lower_bound(offset); | |
3012 | if (p != iomap.begin()) { | |
3013 | --p; | |
3014 | auto end = p->first + p->second.bl.length(); | |
3015 | if (end > offset) { | |
3016 | bufferlist head; | |
3017 | head.substr_of(p->second.bl, 0, offset - p->first); | |
3018 | dout(20) << __func__ << " keep head " << p->second.seq | |
3019 | << " 0x" << std::hex << p->first << "~" << p->second.bl.length() | |
3020 | << " -> 0x" << head.length() << std::dec << dendl; | |
3021 | auto i = seq_bytes.find(p->second.seq); | |
224ce89b | 3022 | assert(i != seq_bytes.end()); |
7c673cae FG |
3023 | if (end > offset + length) { |
3024 | bufferlist tail; | |
3025 | tail.substr_of(p->second.bl, offset + length - p->first, | |
3026 | end - (offset + length)); | |
3027 | dout(20) << __func__ << " keep tail " << p->second.seq | |
3028 | << " 0x" << std::hex << p->first << "~" << p->second.bl.length() | |
3029 | << " -> 0x" << tail.length() << std::dec << dendl; | |
3030 | auto &n = iomap[offset + length]; | |
3031 | n.bl.swap(tail); | |
3032 | n.seq = p->second.seq; | |
3033 | i->second -= length; | |
3034 | } else { | |
3035 | i->second -= end - offset; | |
3036 | } | |
224ce89b | 3037 | assert(i->second >= 0); |
7c673cae FG |
3038 | p->second.bl.swap(head); |
3039 | } | |
3040 | ++p; | |
3041 | } | |
3042 | while (p != iomap.end()) { | |
3043 | if (p->first >= offset + length) { | |
3044 | break; | |
3045 | } | |
3046 | auto i = seq_bytes.find(p->second.seq); | |
224ce89b | 3047 | assert(i != seq_bytes.end()); |
7c673cae FG |
3048 | auto end = p->first + p->second.bl.length(); |
3049 | if (end > offset + length) { | |
3050 | unsigned drop_front = offset + length - p->first; | |
3051 | unsigned keep_tail = end - (offset + length); | |
3052 | dout(20) << __func__ << " truncate front " << p->second.seq | |
3053 | << " 0x" << std::hex << p->first << "~" << p->second.bl.length() | |
3054 | << " drop_front 0x" << drop_front << " keep_tail 0x" << keep_tail | |
3055 | << " to 0x" << (offset + length) << "~" << keep_tail | |
3056 | << std::dec << dendl; | |
3057 | auto &s = iomap[offset + length]; | |
3058 | s.seq = p->second.seq; | |
3059 | s.bl.substr_of(p->second.bl, drop_front, keep_tail); | |
3060 | i->second -= drop_front; | |
3061 | } else { | |
3062 | dout(20) << __func__ << " drop " << p->second.seq | |
3063 | << " 0x" << std::hex << p->first << "~" << p->second.bl.length() | |
3064 | << std::dec << dendl; | |
3065 | i->second -= p->second.bl.length(); | |
3066 | } | |
224ce89b | 3067 | assert(i->second >= 0); |
7c673cae FG |
3068 | p = iomap.erase(p); |
3069 | } | |
3070 | } | |
3071 | ||
3072 | void BlueStore::DeferredBatch::_audit(CephContext *cct) | |
3073 | { | |
3074 | map<uint64_t,int> sb; | |
3075 | for (auto p : seq_bytes) { | |
3076 | sb[p.first] = 0; // make sure we have the same set of keys | |
3077 | } | |
3078 | uint64_t pos = 0; | |
3079 | for (auto& p : iomap) { | |
3080 | assert(p.first >= pos); | |
3081 | sb[p.second.seq] += p.second.bl.length(); | |
3082 | pos = p.first + p.second.bl.length(); | |
3083 | } | |
3084 | assert(sb == seq_bytes); | |
3085 | } | |
3086 | ||
3087 | ||
3088 | // Collection | |
3089 | ||
3090 | #undef dout_prefix | |
3091 | #define dout_prefix *_dout << "bluestore(" << store->path << ").collection(" << cid << " " << this << ") " | |
3092 | ||
3093 | BlueStore::Collection::Collection(BlueStore *ns, Cache *c, coll_t cid) | |
3094 | : store(ns), | |
3095 | cache(c), | |
3096 | cid(cid), | |
3097 | lock("BlueStore::Collection::lock", true, false), | |
3098 | exists(true), | |
3099 | onode_map(c) | |
3100 | { | |
3101 | } | |
3102 | ||
3103 | void BlueStore::Collection::open_shared_blob(uint64_t sbid, BlobRef b) | |
3104 | { | |
3105 | assert(!b->shared_blob); | |
3106 | const bluestore_blob_t& blob = b->get_blob(); | |
3107 | if (!blob.is_shared()) { | |
3108 | b->shared_blob = new SharedBlob(this); | |
3109 | return; | |
3110 | } | |
3111 | ||
3112 | b->shared_blob = shared_blob_set.lookup(sbid); | |
3113 | if (b->shared_blob) { | |
3114 | ldout(store->cct, 10) << __func__ << " sbid 0x" << std::hex << sbid | |
3115 | << std::dec << " had " << *b->shared_blob << dendl; | |
3116 | } else { | |
3117 | b->shared_blob = new SharedBlob(sbid, this); | |
3118 | shared_blob_set.add(this, b->shared_blob.get()); | |
3119 | ldout(store->cct, 10) << __func__ << " sbid 0x" << std::hex << sbid | |
3120 | << std::dec << " opened " << *b->shared_blob | |
3121 | << dendl; | |
3122 | } | |
3123 | } | |
3124 | ||
3125 | void BlueStore::Collection::load_shared_blob(SharedBlobRef sb) | |
3126 | { | |
3127 | if (!sb->is_loaded()) { | |
3128 | ||
3129 | bufferlist v; | |
3130 | string key; | |
3131 | auto sbid = sb->get_sbid(); | |
3132 | get_shared_blob_key(sbid, &key); | |
3133 | int r = store->db->get(PREFIX_SHARED_BLOB, key, &v); | |
3134 | if (r < 0) { | |
3135 | lderr(store->cct) << __func__ << " sbid 0x" << std::hex << sbid | |
3136 | << std::dec << " not found at key " | |
3137 | << pretty_binary_string(key) << dendl; | |
3138 | assert(0 == "uh oh, missing shared_blob"); | |
3139 | } | |
3140 | ||
3141 | sb->loaded = true; | |
3142 | sb->persistent = new bluestore_shared_blob_t(sbid); | |
3143 | bufferlist::iterator p = v.begin(); | |
3144 | ::decode(*(sb->persistent), p); | |
3145 | ldout(store->cct, 10) << __func__ << " sbid 0x" << std::hex << sbid | |
3146 | << std::dec << " loaded shared_blob " << *sb << dendl; | |
3147 | } | |
3148 | } | |
3149 | ||
3150 | void BlueStore::Collection::make_blob_shared(uint64_t sbid, BlobRef b) | |
3151 | { | |
7c673cae | 3152 | ldout(store->cct, 10) << __func__ << " " << *b << dendl; |
31f18b77 | 3153 | assert(!b->shared_blob->is_loaded()); |
7c673cae FG |
3154 | |
3155 | // update blob | |
31f18b77 | 3156 | bluestore_blob_t& blob = b->dirty_blob(); |
7c673cae | 3157 | blob.set_flag(bluestore_blob_t::FLAG_SHARED); |
7c673cae FG |
3158 | |
3159 | // update shared blob | |
3160 | b->shared_blob->loaded = true; | |
3161 | b->shared_blob->persistent = new bluestore_shared_blob_t(sbid); | |
3162 | shared_blob_set.add(this, b->shared_blob.get()); | |
3163 | for (auto p : blob.get_extents()) { | |
3164 | if (p.is_valid()) { | |
3165 | b->shared_blob->get_ref( | |
3166 | p.offset, | |
3167 | p.length); | |
3168 | } | |
3169 | } | |
3170 | ldout(store->cct, 20) << __func__ << " now " << *b << dendl; | |
3171 | } | |
3172 | ||
31f18b77 FG |
3173 | uint64_t BlueStore::Collection::make_blob_unshared(SharedBlob *sb) |
3174 | { | |
3175 | ldout(store->cct, 10) << __func__ << " " << *sb << dendl; | |
3176 | assert(sb->is_loaded()); | |
3177 | ||
3178 | uint64_t sbid = sb->get_sbid(); | |
3179 | shared_blob_set.remove(sb); | |
3180 | sb->loaded = false; | |
3181 | delete sb->persistent; | |
3182 | sb->sbid_unloaded = 0; | |
3183 | ldout(store->cct, 20) << __func__ << " now " << *sb << dendl; | |
3184 | return sbid; | |
3185 | } | |
3186 | ||
7c673cae FG |
3187 | BlueStore::OnodeRef BlueStore::Collection::get_onode( |
3188 | const ghobject_t& oid, | |
3189 | bool create) | |
3190 | { | |
3191 | assert(create ? lock.is_wlocked() : lock.is_locked()); | |
3192 | ||
3193 | spg_t pgid; | |
3194 | if (cid.is_pg(&pgid)) { | |
3195 | if (!oid.match(cnode.bits, pgid.ps())) { | |
3196 | lderr(store->cct) << __func__ << " oid " << oid << " not part of " | |
3197 | << pgid << " bits " << cnode.bits << dendl; | |
3198 | ceph_abort(); | |
3199 | } | |
3200 | } | |
3201 | ||
3202 | OnodeRef o = onode_map.lookup(oid); | |
3203 | if (o) | |
3204 | return o; | |
3205 | ||
31f18b77 | 3206 | mempool::bluestore_cache_other::string key; |
7c673cae FG |
3207 | get_object_key(store->cct, oid, &key); |
3208 | ||
3209 | ldout(store->cct, 20) << __func__ << " oid " << oid << " key " | |
3210 | << pretty_binary_string(key) << dendl; | |
3211 | ||
3212 | bufferlist v; | |
3213 | int r = store->db->get(PREFIX_OBJ, key.c_str(), key.size(), &v); | |
3214 | ldout(store->cct, 20) << " r " << r << " v.len " << v.length() << dendl; | |
3215 | Onode *on; | |
3216 | if (v.length() == 0) { | |
3217 | assert(r == -ENOENT); | |
3218 | if (!store->cct->_conf->bluestore_debug_misc && | |
3219 | !create) | |
3220 | return OnodeRef(); | |
3221 | ||
3222 | // new object, new onode | |
3223 | on = new Onode(this, oid, key); | |
3224 | } else { | |
3225 | // loaded | |
3226 | assert(r >= 0); | |
3227 | on = new Onode(this, oid, key); | |
3228 | on->exists = true; | |
31f18b77 | 3229 | bufferptr::iterator p = v.front().begin_deep(); |
7c673cae FG |
3230 | on->onode.decode(p); |
3231 | ||
3232 | // initialize extent_map | |
3233 | on->extent_map.decode_spanning_blobs(p); | |
3234 | if (on->onode.extent_map_shards.empty()) { | |
3235 | denc(on->extent_map.inline_bl, p); | |
3236 | on->extent_map.decode_some(on->extent_map.inline_bl); | |
3237 | } else { | |
3238 | on->extent_map.init_shards(false, false); | |
3239 | } | |
3240 | } | |
3241 | o.reset(on); | |
3242 | return onode_map.add(oid, o); | |
3243 | } | |
3244 | ||
3245 | void BlueStore::Collection::split_cache( | |
3246 | Collection *dest) | |
3247 | { | |
3248 | ldout(store->cct, 10) << __func__ << " to " << dest << dendl; | |
3249 | ||
3250 | // lock (one or both) cache shards | |
3251 | std::lock(cache->lock, dest->cache->lock); | |
3252 | std::lock_guard<std::recursive_mutex> l(cache->lock, std::adopt_lock); | |
3253 | std::lock_guard<std::recursive_mutex> l2(dest->cache->lock, std::adopt_lock); | |
3254 | ||
3255 | int destbits = dest->cnode.bits; | |
3256 | spg_t destpg; | |
3257 | bool is_pg = dest->cid.is_pg(&destpg); | |
3258 | assert(is_pg); | |
3259 | ||
3260 | auto p = onode_map.onode_map.begin(); | |
3261 | while (p != onode_map.onode_map.end()) { | |
3262 | if (!p->second->oid.match(destbits, destpg.pgid.ps())) { | |
3263 | // onode does not belong to this child | |
3264 | ++p; | |
3265 | } else { | |
3266 | OnodeRef o = p->second; | |
3267 | ldout(store->cct, 20) << __func__ << " moving " << o << " " << o->oid | |
3268 | << dendl; | |
3269 | ||
3270 | cache->_rm_onode(p->second); | |
3271 | p = onode_map.onode_map.erase(p); | |
3272 | ||
3273 | o->c = dest; | |
3274 | dest->cache->_add_onode(o, 1); | |
3275 | dest->onode_map.onode_map[o->oid] = o; | |
3276 | dest->onode_map.cache = dest->cache; | |
3277 | ||
3278 | // move over shared blobs and buffers. cover shared blobs from | |
3279 | // both extent map and spanning blob map (the full extent map | |
3280 | // may not be faulted in) | |
3281 | vector<SharedBlob*> sbvec; | |
3282 | for (auto& e : o->extent_map.extent_map) { | |
3283 | sbvec.push_back(e.blob->shared_blob.get()); | |
3284 | } | |
3285 | for (auto& b : o->extent_map.spanning_blob_map) { | |
3286 | sbvec.push_back(b.second->shared_blob.get()); | |
3287 | } | |
3288 | for (auto sb : sbvec) { | |
3289 | if (sb->coll == dest) { | |
3290 | ldout(store->cct, 20) << __func__ << " already moved " << *sb | |
3291 | << dendl; | |
3292 | continue; | |
3293 | } | |
3294 | ldout(store->cct, 20) << __func__ << " moving " << *sb << dendl; | |
3295 | sb->coll = dest; | |
31f18b77 FG |
3296 | if (sb->get_sbid()) { |
3297 | ldout(store->cct, 20) << __func__ | |
3298 | << " moving registration " << *sb << dendl; | |
3299 | shared_blob_set.remove(sb); | |
3300 | dest->shared_blob_set.add(dest, sb); | |
3301 | } | |
7c673cae | 3302 | if (dest->cache != cache) { |
7c673cae FG |
3303 | for (auto& i : sb->bc.buffer_map) { |
3304 | if (!i.second->is_writing()) { | |
3305 | ldout(store->cct, 20) << __func__ << " moving " << *i.second | |
3306 | << dendl; | |
3307 | dest->cache->_move_buffer(cache, i.second.get()); | |
3308 | } | |
3309 | } | |
3310 | } | |
3311 | } | |
7c673cae FG |
3312 | } |
3313 | } | |
3314 | } | |
3315 | ||
7c673cae FG |
3316 | // ======================================================= |
3317 | ||
3318 | void *BlueStore::MempoolThread::entry() | |
3319 | { | |
3320 | Mutex::Locker l(lock); | |
3321 | while (!stop) { | |
31f18b77 FG |
3322 | uint64_t meta_bytes = |
3323 | mempool::bluestore_cache_other::allocated_bytes() + | |
3324 | mempool::bluestore_cache_onode::allocated_bytes(); | |
3325 | uint64_t onode_num = | |
3326 | mempool::bluestore_cache_onode::allocated_items(); | |
3327 | ||
3328 | if (onode_num < 2) { | |
3329 | onode_num = 2; | |
3330 | } | |
3331 | ||
3332 | float bytes_per_onode = (float)meta_bytes / (float)onode_num; | |
3333 | size_t num_shards = store->cache_shards.size(); | |
3334 | float target_ratio = store->cache_meta_ratio + store->cache_data_ratio; | |
3335 | // A little sloppy but should be close enough | |
224ce89b | 3336 | uint64_t shard_target = target_ratio * (store->cache_size / num_shards); |
31f18b77 FG |
3337 | |
3338 | for (auto i : store->cache_shards) { | |
3339 | i->trim(shard_target, | |
3340 | store->cache_meta_ratio, | |
3341 | store->cache_data_ratio, | |
3342 | bytes_per_onode); | |
3343 | } | |
3344 | ||
3345 | store->_update_cache_logger(); | |
3346 | ||
7c673cae FG |
3347 | utime_t wait; |
3348 | wait += store->cct->_conf->bluestore_cache_trim_interval; | |
3349 | cond.WaitInterval(lock, wait); | |
3350 | } | |
3351 | stop = false; | |
3352 | return NULL; | |
3353 | } | |
3354 | ||
3355 | // ======================================================= | |
3356 | ||
31f18b77 FG |
3357 | // OmapIteratorImpl |
3358 | ||
3359 | #undef dout_prefix | |
3360 | #define dout_prefix *_dout << "bluestore.OmapIteratorImpl(" << this << ") " | |
3361 | ||
3362 | BlueStore::OmapIteratorImpl::OmapIteratorImpl( | |
3363 | CollectionRef c, OnodeRef o, KeyValueDB::Iterator it) | |
3364 | : c(c), o(o), it(it) | |
3365 | { | |
3366 | RWLock::RLocker l(c->lock); | |
3367 | if (o->onode.has_omap()) { | |
3368 | get_omap_key(o->onode.nid, string(), &head); | |
3369 | get_omap_tail(o->onode.nid, &tail); | |
3370 | it->lower_bound(head); | |
3371 | } | |
3372 | } | |
3373 | ||
3374 | int BlueStore::OmapIteratorImpl::seek_to_first() | |
3375 | { | |
3376 | RWLock::RLocker l(c->lock); | |
3377 | if (o->onode.has_omap()) { | |
3378 | it->lower_bound(head); | |
3379 | } else { | |
3380 | it = KeyValueDB::Iterator(); | |
3381 | } | |
3382 | return 0; | |
3383 | } | |
3384 | ||
3385 | int BlueStore::OmapIteratorImpl::upper_bound(const string& after) | |
3386 | { | |
3387 | RWLock::RLocker l(c->lock); | |
3388 | if (o->onode.has_omap()) { | |
3389 | string key; | |
3390 | get_omap_key(o->onode.nid, after, &key); | |
3391 | ldout(c->store->cct,20) << __func__ << " after " << after << " key " | |
3392 | << pretty_binary_string(key) << dendl; | |
3393 | it->upper_bound(key); | |
3394 | } else { | |
3395 | it = KeyValueDB::Iterator(); | |
3396 | } | |
3397 | return 0; | |
3398 | } | |
3399 | ||
3400 | int BlueStore::OmapIteratorImpl::lower_bound(const string& to) | |
3401 | { | |
3402 | RWLock::RLocker l(c->lock); | |
3403 | if (o->onode.has_omap()) { | |
3404 | string key; | |
3405 | get_omap_key(o->onode.nid, to, &key); | |
3406 | ldout(c->store->cct,20) << __func__ << " to " << to << " key " | |
3407 | << pretty_binary_string(key) << dendl; | |
3408 | it->lower_bound(key); | |
3409 | } else { | |
3410 | it = KeyValueDB::Iterator(); | |
3411 | } | |
3412 | return 0; | |
3413 | } | |
3414 | ||
3415 | bool BlueStore::OmapIteratorImpl::valid() | |
3416 | { | |
3417 | RWLock::RLocker l(c->lock); | |
3418 | bool r = o->onode.has_omap() && it && it->valid() && | |
3419 | it->raw_key().second <= tail; | |
3420 | if (it && it->valid()) { | |
3421 | ldout(c->store->cct,20) << __func__ << " is at " | |
3422 | << pretty_binary_string(it->raw_key().second) | |
3423 | << dendl; | |
3424 | } | |
3425 | return r; | |
3426 | } | |
3427 | ||
3428 | int BlueStore::OmapIteratorImpl::next(bool validate) | |
3429 | { | |
3430 | RWLock::RLocker l(c->lock); | |
3431 | if (o->onode.has_omap()) { | |
3432 | it->next(); | |
3433 | return 0; | |
3434 | } else { | |
3435 | return -1; | |
3436 | } | |
3437 | } | |
3438 | ||
3439 | string BlueStore::OmapIteratorImpl::key() | |
3440 | { | |
3441 | RWLock::RLocker l(c->lock); | |
3442 | assert(it->valid()); | |
3443 | string db_key = it->raw_key().second; | |
3444 | string user_key; | |
3445 | decode_omap_key(db_key, &user_key); | |
3446 | return user_key; | |
3447 | } | |
3448 | ||
3449 | bufferlist BlueStore::OmapIteratorImpl::value() | |
3450 | { | |
3451 | RWLock::RLocker l(c->lock); | |
3452 | assert(it->valid()); | |
3453 | return it->value(); | |
3454 | } | |
3455 | ||
3456 | ||
3457 | // ===================================== | |
3458 | ||
7c673cae FG |
3459 | #undef dout_prefix |
3460 | #define dout_prefix *_dout << "bluestore(" << path << ") " | |
3461 | ||
3462 | ||
3463 | static void aio_cb(void *priv, void *priv2) | |
3464 | { | |
3465 | BlueStore *store = static_cast<BlueStore*>(priv); | |
3466 | BlueStore::AioContext *c = static_cast<BlueStore::AioContext*>(priv2); | |
3467 | c->aio_finish(store); | |
3468 | } | |
3469 | ||
3470 | BlueStore::BlueStore(CephContext *cct, const string& path) | |
3471 | : ObjectStore(cct, path), | |
3472 | throttle_bytes(cct, "bluestore_throttle_bytes", | |
3473 | cct->_conf->bluestore_throttle_bytes), | |
3474 | throttle_deferred_bytes(cct, "bluestore_throttle_deferred_bytes", | |
3475 | cct->_conf->bluestore_throttle_bytes + | |
3476 | cct->_conf->bluestore_throttle_deferred_bytes), | |
181888fb | 3477 | deferred_finisher(cct, "defered_finisher", "dfin"), |
7c673cae | 3478 | kv_sync_thread(this), |
31f18b77 | 3479 | kv_finalize_thread(this), |
7c673cae FG |
3480 | mempool_thread(this) |
3481 | { | |
3482 | _init_logger(); | |
3483 | cct->_conf->add_observer(this); | |
3484 | set_cache_shards(1); | |
7c673cae FG |
3485 | } |
3486 | ||
3487 | BlueStore::BlueStore(CephContext *cct, | |
3488 | const string& path, | |
3489 | uint64_t _min_alloc_size) | |
3490 | : ObjectStore(cct, path), | |
3491 | throttle_bytes(cct, "bluestore_throttle_bytes", | |
3492 | cct->_conf->bluestore_throttle_bytes), | |
3493 | throttle_deferred_bytes(cct, "bluestore_throttle_deferred_bytes", | |
3494 | cct->_conf->bluestore_throttle_bytes + | |
3495 | cct->_conf->bluestore_throttle_deferred_bytes), | |
181888fb | 3496 | deferred_finisher(cct, "defered_finisher", "dfin"), |
7c673cae | 3497 | kv_sync_thread(this), |
31f18b77 | 3498 | kv_finalize_thread(this), |
7c673cae FG |
3499 | min_alloc_size(_min_alloc_size), |
3500 | min_alloc_size_order(ctz(_min_alloc_size)), | |
3501 | mempool_thread(this) | |
3502 | { | |
3503 | _init_logger(); | |
3504 | cct->_conf->add_observer(this); | |
3505 | set_cache_shards(1); | |
7c673cae FG |
3506 | } |
3507 | ||
3508 | BlueStore::~BlueStore() | |
3509 | { | |
3510 | for (auto f : finishers) { | |
3511 | delete f; | |
3512 | } | |
3513 | finishers.clear(); | |
3514 | ||
3515 | cct->_conf->remove_observer(this); | |
3516 | _shutdown_logger(); | |
3517 | assert(!mounted); | |
3518 | assert(db == NULL); | |
3519 | assert(bluefs == NULL); | |
3520 | assert(fsid_fd < 0); | |
3521 | assert(path_fd < 0); | |
3522 | for (auto i : cache_shards) { | |
3523 | delete i; | |
3524 | } | |
3525 | cache_shards.clear(); | |
3526 | } | |
3527 | ||
3528 | const char **BlueStore::get_tracked_conf_keys() const | |
3529 | { | |
3530 | static const char* KEYS[] = { | |
3531 | "bluestore_csum_type", | |
3532 | "bluestore_compression_mode", | |
3533 | "bluestore_compression_algorithm", | |
3534 | "bluestore_compression_min_blob_size", | |
3535 | "bluestore_compression_min_blob_size_ssd", | |
3536 | "bluestore_compression_min_blob_size_hdd", | |
3537 | "bluestore_compression_max_blob_size", | |
3538 | "bluestore_compression_max_blob_size_ssd", | |
3539 | "bluestore_compression_max_blob_size_hdd", | |
c07f9fc5 | 3540 | "bluestore_compression_required_ratio", |
7c673cae FG |
3541 | "bluestore_max_alloc_size", |
3542 | "bluestore_prefer_deferred_size", | |
181888fb FG |
3543 | "bluestore_prefer_deferred_size_hdd", |
3544 | "bluestore_prefer_deferred_size_ssd", | |
31f18b77 FG |
3545 | "bluestore_deferred_batch_ops", |
3546 | "bluestore_deferred_batch_ops_hdd", | |
3547 | "bluestore_deferred_batch_ops_ssd", | |
7c673cae FG |
3548 | "bluestore_throttle_bytes", |
3549 | "bluestore_throttle_deferred_bytes", | |
3550 | "bluestore_throttle_cost_per_io_hdd", | |
3551 | "bluestore_throttle_cost_per_io_ssd", | |
3552 | "bluestore_throttle_cost_per_io", | |
3553 | "bluestore_max_blob_size", | |
3554 | "bluestore_max_blob_size_ssd", | |
3555 | "bluestore_max_blob_size_hdd", | |
3556 | NULL | |
3557 | }; | |
3558 | return KEYS; | |
3559 | } | |
3560 | ||
3561 | void BlueStore::handle_conf_change(const struct md_config_t *conf, | |
3562 | const std::set<std::string> &changed) | |
3563 | { | |
3564 | if (changed.count("bluestore_csum_type")) { | |
3565 | _set_csum(); | |
3566 | } | |
3567 | if (changed.count("bluestore_compression_mode") || | |
3568 | changed.count("bluestore_compression_algorithm") || | |
3569 | changed.count("bluestore_compression_min_blob_size") || | |
3570 | changed.count("bluestore_compression_max_blob_size")) { | |
3571 | if (bdev) { | |
3572 | _set_compression(); | |
3573 | } | |
3574 | } | |
3575 | if (changed.count("bluestore_max_blob_size") || | |
3576 | changed.count("bluestore_max_blob_size_ssd") || | |
3577 | changed.count("bluestore_max_blob_size_hdd")) { | |
3578 | if (bdev) { | |
3579 | // only after startup | |
3580 | _set_blob_size(); | |
3581 | } | |
3582 | } | |
3583 | if (changed.count("bluestore_prefer_deferred_size") || | |
181888fb FG |
3584 | changed.count("bluestore_prefer_deferred_size_hdd") || |
3585 | changed.count("bluestore_prefer_deferred_size_ssd") || | |
7c673cae FG |
3586 | changed.count("bluestore_max_alloc_size") || |
3587 | changed.count("bluestore_deferred_batch_ops") || | |
3588 | changed.count("bluestore_deferred_batch_ops_hdd") || | |
3589 | changed.count("bluestore_deferred_batch_ops_ssd")) { | |
3590 | if (bdev) { | |
3591 | // only after startup | |
3592 | _set_alloc_sizes(); | |
3593 | } | |
3594 | } | |
3595 | if (changed.count("bluestore_throttle_cost_per_io") || | |
3596 | changed.count("bluestore_throttle_cost_per_io_hdd") || | |
3597 | changed.count("bluestore_throttle_cost_per_io_ssd")) { | |
3598 | if (bdev) { | |
3599 | _set_throttle_params(); | |
3600 | } | |
3601 | } | |
3602 | if (changed.count("bluestore_throttle_bytes")) { | |
3603 | throttle_bytes.reset_max(conf->bluestore_throttle_bytes); | |
3604 | throttle_deferred_bytes.reset_max( | |
3605 | conf->bluestore_throttle_bytes + conf->bluestore_throttle_deferred_bytes); | |
3606 | } | |
3607 | if (changed.count("bluestore_throttle_deferred_bytes")) { | |
3608 | throttle_deferred_bytes.reset_max( | |
3609 | conf->bluestore_throttle_bytes + conf->bluestore_throttle_deferred_bytes); | |
3610 | } | |
3611 | } | |
3612 | ||
3613 | void BlueStore::_set_compression() | |
3614 | { | |
224ce89b WB |
3615 | auto m = Compressor::get_comp_mode_type(cct->_conf->bluestore_compression_mode); |
3616 | if (m) { | |
3617 | comp_mode = *m; | |
3618 | } else { | |
3619 | derr << __func__ << " unrecognized value '" | |
3620 | << cct->_conf->bluestore_compression_mode | |
3621 | << "' for bluestore_compression_mode, reverting to 'none'" | |
3622 | << dendl; | |
3623 | comp_mode = Compressor::COMP_NONE; | |
3624 | } | |
3625 | ||
3626 | compressor = nullptr; | |
3627 | ||
3628 | if (comp_mode == Compressor::COMP_NONE) { | |
3629 | dout(10) << __func__ << " compression mode set to 'none', " | |
3630 | << "ignore other compression setttings" << dendl; | |
3631 | return; | |
3632 | } | |
3633 | ||
7c673cae FG |
3634 | if (cct->_conf->bluestore_compression_max_blob_size) { |
3635 | comp_min_blob_size = cct->_conf->bluestore_compression_max_blob_size; | |
3636 | } else { | |
3637 | assert(bdev); | |
3638 | if (bdev->is_rotational()) { | |
3639 | comp_min_blob_size = cct->_conf->bluestore_compression_min_blob_size_hdd; | |
3640 | } else { | |
3641 | comp_min_blob_size = cct->_conf->bluestore_compression_min_blob_size_ssd; | |
3642 | } | |
3643 | } | |
3644 | ||
3645 | if (cct->_conf->bluestore_compression_max_blob_size) { | |
3646 | comp_max_blob_size = cct->_conf->bluestore_compression_max_blob_size; | |
3647 | } else { | |
3648 | assert(bdev); | |
3649 | if (bdev->is_rotational()) { | |
3650 | comp_max_blob_size = cct->_conf->bluestore_compression_max_blob_size_hdd; | |
3651 | } else { | |
3652 | comp_max_blob_size = cct->_conf->bluestore_compression_max_blob_size_ssd; | |
3653 | } | |
3654 | } | |
3655 | ||
7c673cae FG |
3656 | auto& alg_name = cct->_conf->bluestore_compression_algorithm; |
3657 | if (!alg_name.empty()) { | |
3658 | compressor = Compressor::create(cct, alg_name); | |
3659 | if (!compressor) { | |
3660 | derr << __func__ << " unable to initialize " << alg_name.c_str() << " compressor" | |
3661 | << dendl; | |
3662 | } | |
3663 | } | |
3664 | ||
3665 | dout(10) << __func__ << " mode " << Compressor::get_comp_mode_name(comp_mode) | |
3666 | << " alg " << (compressor ? compressor->get_type_name() : "(none)") | |
3667 | << dendl; | |
3668 | } | |
3669 | ||
3670 | void BlueStore::_set_csum() | |
3671 | { | |
3672 | csum_type = Checksummer::CSUM_NONE; | |
3673 | int t = Checksummer::get_csum_string_type(cct->_conf->bluestore_csum_type); | |
3674 | if (t > Checksummer::CSUM_NONE) | |
3675 | csum_type = t; | |
3676 | ||
3677 | dout(10) << __func__ << " csum_type " | |
3678 | << Checksummer::get_csum_type_string(csum_type) | |
3679 | << dendl; | |
3680 | } | |
3681 | ||
3682 | void BlueStore::_set_throttle_params() | |
3683 | { | |
3684 | if (cct->_conf->bluestore_throttle_cost_per_io) { | |
3685 | throttle_cost_per_io = cct->_conf->bluestore_throttle_cost_per_io; | |
3686 | } else { | |
3687 | assert(bdev); | |
3688 | if (bdev->is_rotational()) { | |
3689 | throttle_cost_per_io = cct->_conf->bluestore_throttle_cost_per_io_hdd; | |
3690 | } else { | |
3691 | throttle_cost_per_io = cct->_conf->bluestore_throttle_cost_per_io_ssd; | |
3692 | } | |
3693 | } | |
3694 | ||
3695 | dout(10) << __func__ << " throttle_cost_per_io " << throttle_cost_per_io | |
3696 | << dendl; | |
3697 | } | |
3698 | void BlueStore::_set_blob_size() | |
3699 | { | |
3700 | if (cct->_conf->bluestore_max_blob_size) { | |
3701 | max_blob_size = cct->_conf->bluestore_max_blob_size; | |
3702 | } else { | |
3703 | assert(bdev); | |
3704 | if (bdev->is_rotational()) { | |
3705 | max_blob_size = cct->_conf->bluestore_max_blob_size_hdd; | |
3706 | } else { | |
3707 | max_blob_size = cct->_conf->bluestore_max_blob_size_ssd; | |
3708 | } | |
3709 | } | |
3710 | dout(10) << __func__ << " max_blob_size 0x" << std::hex << max_blob_size | |
3711 | << std::dec << dendl; | |
3712 | } | |
3713 | ||
31f18b77 FG |
3714 | int BlueStore::_set_cache_sizes() |
3715 | { | |
224ce89b WB |
3716 | assert(bdev); |
3717 | if (cct->_conf->bluestore_cache_size) { | |
3718 | cache_size = cct->_conf->bluestore_cache_size; | |
3719 | } else { | |
3720 | // choose global cache size based on backend type | |
3721 | if (bdev->is_rotational()) { | |
3722 | cache_size = cct->_conf->bluestore_cache_size_hdd; | |
3723 | } else { | |
3724 | cache_size = cct->_conf->bluestore_cache_size_ssd; | |
3725 | } | |
3726 | } | |
31f18b77 FG |
3727 | cache_meta_ratio = cct->_conf->bluestore_cache_meta_ratio; |
3728 | cache_kv_ratio = cct->_conf->bluestore_cache_kv_ratio; | |
224ce89b WB |
3729 | |
3730 | double cache_kv_max = cct->_conf->bluestore_cache_kv_max; | |
3731 | double cache_kv_max_ratio = 0; | |
3732 | ||
3733 | // if cache_kv_max is negative, disable it | |
3734 | if (cache_size > 0 && cache_kv_max >= 0) { | |
3735 | cache_kv_max_ratio = (double) cache_kv_max / (double) cache_size; | |
3736 | if (cache_kv_max_ratio < 1.0 && cache_kv_max_ratio < cache_kv_ratio) { | |
3737 | dout(1) << __func__ << " max " << cache_kv_max_ratio | |
3738 | << " < ratio " << cache_kv_ratio | |
3739 | << dendl; | |
3740 | cache_meta_ratio = cache_meta_ratio + cache_kv_ratio - cache_kv_max_ratio; | |
3741 | cache_kv_ratio = cache_kv_max_ratio; | |
3742 | } | |
3743 | } | |
3744 | ||
31f18b77 FG |
3745 | cache_data_ratio = |
3746 | (double)1.0 - (double)cache_meta_ratio - (double)cache_kv_ratio; | |
3747 | ||
224ce89b | 3748 | if (cache_meta_ratio < 0 || cache_meta_ratio > 1.0) { |
d2e6a577 | 3749 | derr << __func__ << " bluestore_cache_meta_ratio (" << cache_meta_ratio |
224ce89b | 3750 | << ") must be in range [0,1.0]" << dendl; |
31f18b77 FG |
3751 | return -EINVAL; |
3752 | } | |
224ce89b | 3753 | if (cache_kv_ratio < 0 || cache_kv_ratio > 1.0) { |
d2e6a577 | 3754 | derr << __func__ << " bluestore_cache_kv_ratio (" << cache_kv_ratio |
224ce89b | 3755 | << ") must be in range [0,1.0]" << dendl; |
31f18b77 FG |
3756 | return -EINVAL; |
3757 | } | |
3758 | if (cache_meta_ratio + cache_kv_ratio > 1.0) { | |
d2e6a577 | 3759 | derr << __func__ << " bluestore_cache_meta_ratio (" << cache_meta_ratio |
31f18b77 FG |
3760 | << ") + bluestore_cache_kv_ratio (" << cache_kv_ratio |
3761 | << ") = " << cache_meta_ratio + cache_kv_ratio << "; must be <= 1.0" | |
3762 | << dendl; | |
3763 | return -EINVAL; | |
3764 | } | |
3765 | if (cache_data_ratio < 0) { | |
3766 | // deal with floating point imprecision | |
3767 | cache_data_ratio = 0; | |
3768 | } | |
224ce89b WB |
3769 | dout(1) << __func__ << " cache_size " << cache_size |
3770 | << " meta " << cache_meta_ratio | |
31f18b77 FG |
3771 | << " kv " << cache_kv_ratio |
3772 | << " data " << cache_data_ratio | |
3773 | << dendl; | |
3774 | return 0; | |
3775 | } | |
3776 | ||
7c673cae FG |
3777 | void BlueStore::_init_logger() |
3778 | { | |
3779 | PerfCountersBuilder b(cct, "bluestore", | |
3780 | l_bluestore_first, l_bluestore_last); | |
3781 | b.add_time_avg(l_bluestore_kv_flush_lat, "kv_flush_lat", | |
3782 | "Average kv_thread flush latency", | |
3783 | "fl_l", PerfCountersBuilder::PRIO_INTERESTING); | |
3784 | b.add_time_avg(l_bluestore_kv_commit_lat, "kv_commit_lat", | |
3785 | "Average kv_thread commit latency"); | |
3786 | b.add_time_avg(l_bluestore_kv_lat, "kv_lat", | |
3787 | "Average kv_thread sync latency", | |
3788 | "k_l", PerfCountersBuilder::PRIO_INTERESTING); | |
3789 | b.add_time_avg(l_bluestore_state_prepare_lat, "state_prepare_lat", | |
3790 | "Average prepare state latency"); | |
3791 | b.add_time_avg(l_bluestore_state_aio_wait_lat, "state_aio_wait_lat", | |
3792 | "Average aio_wait state latency", | |
3793 | "io_l", PerfCountersBuilder::PRIO_INTERESTING); | |
3794 | b.add_time_avg(l_bluestore_state_io_done_lat, "state_io_done_lat", | |
3795 | "Average io_done state latency"); | |
3796 | b.add_time_avg(l_bluestore_state_kv_queued_lat, "state_kv_queued_lat", | |
3797 | "Average kv_queued state latency"); | |
3798 | b.add_time_avg(l_bluestore_state_kv_committing_lat, "state_kv_commiting_lat", | |
3799 | "Average kv_commiting state latency"); | |
3800 | b.add_time_avg(l_bluestore_state_kv_done_lat, "state_kv_done_lat", | |
3801 | "Average kv_done state latency"); | |
3802 | b.add_time_avg(l_bluestore_state_deferred_queued_lat, "state_deferred_queued_lat", | |
3803 | "Average deferred_queued state latency"); | |
3804 | b.add_time_avg(l_bluestore_state_deferred_aio_wait_lat, "state_deferred_aio_wait_lat", | |
3805 | "Average aio_wait state latency"); | |
3806 | b.add_time_avg(l_bluestore_state_deferred_cleanup_lat, "state_deferred_cleanup_lat", | |
3807 | "Average cleanup state latency"); | |
3808 | b.add_time_avg(l_bluestore_state_finishing_lat, "state_finishing_lat", | |
3809 | "Average finishing state latency"); | |
3810 | b.add_time_avg(l_bluestore_state_done_lat, "state_done_lat", | |
3811 | "Average done state latency"); | |
3812 | b.add_time_avg(l_bluestore_throttle_lat, "throttle_lat", | |
3813 | "Average submit throttle latency", | |
3814 | "th_l", PerfCountersBuilder::PRIO_CRITICAL); | |
3815 | b.add_time_avg(l_bluestore_submit_lat, "submit_lat", | |
3816 | "Average submit latency", | |
3817 | "s_l", PerfCountersBuilder::PRIO_CRITICAL); | |
3818 | b.add_time_avg(l_bluestore_commit_lat, "commit_lat", | |
3819 | "Average commit latency", | |
3820 | "c_l", PerfCountersBuilder::PRIO_CRITICAL); | |
3821 | b.add_time_avg(l_bluestore_read_lat, "read_lat", | |
3822 | "Average read latency", | |
3823 | "r_l", PerfCountersBuilder::PRIO_CRITICAL); | |
3824 | b.add_time_avg(l_bluestore_read_onode_meta_lat, "read_onode_meta_lat", | |
3825 | "Average read onode metadata latency"); | |
3826 | b.add_time_avg(l_bluestore_read_wait_aio_lat, "read_wait_aio_lat", | |
3827 | "Average read latency"); | |
3828 | b.add_time_avg(l_bluestore_compress_lat, "compress_lat", | |
3829 | "Average compress latency"); | |
3830 | b.add_time_avg(l_bluestore_decompress_lat, "decompress_lat", | |
3831 | "Average decompress latency"); | |
3832 | b.add_time_avg(l_bluestore_csum_lat, "csum_lat", | |
3833 | "Average checksum latency"); | |
3834 | b.add_u64_counter(l_bluestore_compress_success_count, "compress_success_count", | |
3835 | "Sum for beneficial compress ops"); | |
3836 | b.add_u64_counter(l_bluestore_compress_rejected_count, "compress_rejected_count", | |
3837 | "Sum for compress ops rejected due to low net gain of space"); | |
3838 | b.add_u64_counter(l_bluestore_write_pad_bytes, "write_pad_bytes", | |
3839 | "Sum for write-op padded bytes"); | |
3840 | b.add_u64_counter(l_bluestore_deferred_write_ops, "deferred_write_ops", | |
3841 | "Sum for deferred write op"); | |
3842 | b.add_u64_counter(l_bluestore_deferred_write_bytes, "deferred_write_bytes", | |
3843 | "Sum for deferred write bytes", "def"); | |
3844 | b.add_u64_counter(l_bluestore_write_penalty_read_ops, "write_penalty_read_ops", | |
3845 | "Sum for write penalty read ops"); | |
3846 | b.add_u64(l_bluestore_allocated, "bluestore_allocated", | |
3847 | "Sum for allocated bytes"); | |
3848 | b.add_u64(l_bluestore_stored, "bluestore_stored", | |
3849 | "Sum for stored bytes"); | |
3850 | b.add_u64(l_bluestore_compressed, "bluestore_compressed", | |
3851 | "Sum for stored compressed bytes"); | |
3852 | b.add_u64(l_bluestore_compressed_allocated, "bluestore_compressed_allocated", | |
3853 | "Sum for bytes allocated for compressed data"); | |
3854 | b.add_u64(l_bluestore_compressed_original, "bluestore_compressed_original", | |
3855 | "Sum for original bytes that were compressed"); | |
3856 | ||
3857 | b.add_u64(l_bluestore_onodes, "bluestore_onodes", | |
3858 | "Number of onodes in cache"); | |
3859 | b.add_u64_counter(l_bluestore_onode_hits, "bluestore_onode_hits", | |
3860 | "Sum for onode-lookups hit in the cache"); | |
3861 | b.add_u64_counter(l_bluestore_onode_misses, "bluestore_onode_misses", | |
3862 | "Sum for onode-lookups missed in the cache"); | |
3863 | b.add_u64_counter(l_bluestore_onode_shard_hits, "bluestore_onode_shard_hits", | |
3864 | "Sum for onode-shard lookups hit in the cache"); | |
3865 | b.add_u64_counter(l_bluestore_onode_shard_misses, | |
3866 | "bluestore_onode_shard_misses", | |
3867 | "Sum for onode-shard lookups missed in the cache"); | |
3868 | b.add_u64(l_bluestore_extents, "bluestore_extents", | |
3869 | "Number of extents in cache"); | |
3870 | b.add_u64(l_bluestore_blobs, "bluestore_blobs", | |
3871 | "Number of blobs in cache"); | |
3872 | b.add_u64(l_bluestore_buffers, "bluestore_buffers", | |
3873 | "Number of buffers in cache"); | |
3874 | b.add_u64(l_bluestore_buffer_bytes, "bluestore_buffer_bytes", | |
3875 | "Number of buffer bytes in cache"); | |
3876 | b.add_u64(l_bluestore_buffer_hit_bytes, "bluestore_buffer_hit_bytes", | |
3877 | "Sum for bytes of read hit in the cache"); | |
3878 | b.add_u64(l_bluestore_buffer_miss_bytes, "bluestore_buffer_miss_bytes", | |
3879 | "Sum for bytes of read missed in the cache"); | |
3880 | ||
3881 | b.add_u64_counter(l_bluestore_write_big, "bluestore_write_big", | |
3882 | "Large aligned writes into fresh blobs"); | |
3883 | b.add_u64_counter(l_bluestore_write_big_bytes, "bluestore_write_big_bytes", | |
3884 | "Large aligned writes into fresh blobs (bytes)"); | |
3885 | b.add_u64_counter(l_bluestore_write_big_blobs, "bluestore_write_big_blobs", | |
3886 | "Large aligned writes into fresh blobs (blobs)"); | |
3887 | b.add_u64_counter(l_bluestore_write_small, "bluestore_write_small", | |
3888 | "Small writes into existing or sparse small blobs"); | |
3889 | b.add_u64_counter(l_bluestore_write_small_bytes, "bluestore_write_small_bytes", | |
3890 | "Small writes into existing or sparse small blobs (bytes)"); | |
3891 | b.add_u64_counter(l_bluestore_write_small_unused, | |
3892 | "bluestore_write_small_unused", | |
3893 | "Small writes into unused portion of existing blob"); | |
3894 | b.add_u64_counter(l_bluestore_write_small_deferred, | |
3895 | "bluestore_write_small_deferred", | |
3896 | "Small overwrites using deferred"); | |
3897 | b.add_u64_counter(l_bluestore_write_small_pre_read, | |
3898 | "bluestore_write_small_pre_read", | |
3899 | "Small writes that required we read some data (possibly " | |
3900 | "cached) to fill out the block"); | |
3901 | b.add_u64_counter(l_bluestore_write_small_new, "bluestore_write_small_new", | |
3902 | "Small write into new (sparse) blob"); | |
3903 | ||
3904 | b.add_u64_counter(l_bluestore_txc, "bluestore_txc", "Transactions committed"); | |
3905 | b.add_u64_counter(l_bluestore_onode_reshard, "bluestore_onode_reshard", | |
3906 | "Onode extent map reshard events"); | |
3907 | b.add_u64_counter(l_bluestore_blob_split, "bluestore_blob_split", | |
3908 | "Sum for blob splitting due to resharding"); | |
3909 | b.add_u64_counter(l_bluestore_extent_compress, "bluestore_extent_compress", | |
3910 | "Sum for extents that have been removed due to compression"); | |
3911 | b.add_u64_counter(l_bluestore_gc_merged, "bluestore_gc_merged", | |
3912 | "Sum for extents that have been merged due to garbage " | |
3913 | "collection"); | |
3914 | logger = b.create_perf_counters(); | |
3915 | cct->get_perfcounters_collection()->add(logger); | |
3916 | } | |
3917 | ||
3918 | int BlueStore::_reload_logger() | |
3919 | { | |
3920 | struct store_statfs_t store_statfs; | |
3921 | ||
3922 | int r = statfs(&store_statfs); | |
3923 | if(r >= 0) { | |
3924 | logger->set(l_bluestore_allocated, store_statfs.allocated); | |
3925 | logger->set(l_bluestore_stored, store_statfs.stored); | |
3926 | logger->set(l_bluestore_compressed, store_statfs.compressed); | |
3927 | logger->set(l_bluestore_compressed_allocated, store_statfs.compressed_allocated); | |
3928 | logger->set(l_bluestore_compressed_original, store_statfs.compressed_original); | |
3929 | } | |
3930 | return r; | |
3931 | } | |
3932 | ||
3933 | void BlueStore::_shutdown_logger() | |
3934 | { | |
3935 | cct->get_perfcounters_collection()->remove(logger); | |
3936 | delete logger; | |
3937 | } | |
3938 | ||
3939 | int BlueStore::get_block_device_fsid(CephContext* cct, const string& path, | |
3940 | uuid_d *fsid) | |
3941 | { | |
3942 | bluestore_bdev_label_t label; | |
3943 | int r = _read_bdev_label(cct, path, &label); | |
3944 | if (r < 0) | |
3945 | return r; | |
3946 | *fsid = label.osd_uuid; | |
3947 | return 0; | |
3948 | } | |
3949 | ||
3950 | int BlueStore::_open_path() | |
3951 | { | |
3952 | assert(path_fd < 0); | |
224ce89b | 3953 | path_fd = TEMP_FAILURE_RETRY(::open(path.c_str(), O_DIRECTORY)); |
7c673cae FG |
3954 | if (path_fd < 0) { |
3955 | int r = -errno; | |
3956 | derr << __func__ << " unable to open " << path << ": " << cpp_strerror(r) | |
3957 | << dendl; | |
3958 | return r; | |
3959 | } | |
3960 | return 0; | |
3961 | } | |
3962 | ||
3963 | void BlueStore::_close_path() | |
3964 | { | |
3965 | VOID_TEMP_FAILURE_RETRY(::close(path_fd)); | |
3966 | path_fd = -1; | |
3967 | } | |
3968 | ||
3969 | int BlueStore::_write_bdev_label(string path, bluestore_bdev_label_t label) | |
3970 | { | |
3971 | dout(10) << __func__ << " path " << path << " label " << label << dendl; | |
3972 | bufferlist bl; | |
3973 | ::encode(label, bl); | |
3974 | uint32_t crc = bl.crc32c(-1); | |
3975 | ::encode(crc, bl); | |
3976 | assert(bl.length() <= BDEV_LABEL_BLOCK_SIZE); | |
3977 | bufferptr z(BDEV_LABEL_BLOCK_SIZE - bl.length()); | |
3978 | z.zero(); | |
3979 | bl.append(std::move(z)); | |
3980 | ||
224ce89b | 3981 | int fd = TEMP_FAILURE_RETRY(::open(path.c_str(), O_WRONLY)); |
7c673cae FG |
3982 | if (fd < 0) { |
3983 | fd = -errno; | |
3984 | derr << __func__ << " failed to open " << path << ": " << cpp_strerror(fd) | |
3985 | << dendl; | |
3986 | return fd; | |
3987 | } | |
3988 | int r = bl.write_fd(fd); | |
3989 | if (r < 0) { | |
3990 | derr << __func__ << " failed to write to " << path | |
3991 | << ": " << cpp_strerror(r) << dendl; | |
3992 | } | |
3993 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
3994 | return r; | |
3995 | } | |
3996 | ||
3997 | int BlueStore::_read_bdev_label(CephContext* cct, string path, | |
3998 | bluestore_bdev_label_t *label) | |
3999 | { | |
4000 | dout(10) << __func__ << dendl; | |
224ce89b | 4001 | int fd = TEMP_FAILURE_RETRY(::open(path.c_str(), O_RDONLY)); |
7c673cae FG |
4002 | if (fd < 0) { |
4003 | fd = -errno; | |
4004 | derr << __func__ << " failed to open " << path << ": " << cpp_strerror(fd) | |
4005 | << dendl; | |
4006 | return fd; | |
4007 | } | |
4008 | bufferlist bl; | |
4009 | int r = bl.read_fd(fd, BDEV_LABEL_BLOCK_SIZE); | |
4010 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
4011 | if (r < 0) { | |
4012 | derr << __func__ << " failed to read from " << path | |
4013 | << ": " << cpp_strerror(r) << dendl; | |
4014 | return r; | |
4015 | } | |
4016 | ||
4017 | uint32_t crc, expected_crc; | |
4018 | bufferlist::iterator p = bl.begin(); | |
4019 | try { | |
4020 | ::decode(*label, p); | |
4021 | bufferlist t; | |
4022 | t.substr_of(bl, 0, p.get_off()); | |
4023 | crc = t.crc32c(-1); | |
4024 | ::decode(expected_crc, p); | |
4025 | } | |
4026 | catch (buffer::error& e) { | |
4027 | derr << __func__ << " unable to decode label at offset " << p.get_off() | |
4028 | << ": " << e.what() | |
4029 | << dendl; | |
4030 | return -EINVAL; | |
4031 | } | |
4032 | if (crc != expected_crc) { | |
4033 | derr << __func__ << " bad crc on label, expected " << expected_crc | |
4034 | << " != actual " << crc << dendl; | |
4035 | return -EIO; | |
4036 | } | |
4037 | dout(10) << __func__ << " got " << *label << dendl; | |
4038 | return 0; | |
4039 | } | |
4040 | ||
4041 | int BlueStore::_check_or_set_bdev_label( | |
4042 | string path, uint64_t size, string desc, bool create) | |
4043 | { | |
4044 | bluestore_bdev_label_t label; | |
4045 | if (create) { | |
4046 | label.osd_uuid = fsid; | |
4047 | label.size = size; | |
4048 | label.btime = ceph_clock_now(); | |
4049 | label.description = desc; | |
4050 | int r = _write_bdev_label(path, label); | |
4051 | if (r < 0) | |
4052 | return r; | |
4053 | } else { | |
4054 | int r = _read_bdev_label(cct, path, &label); | |
4055 | if (r < 0) | |
4056 | return r; | |
31f18b77 FG |
4057 | if (cct->_conf->bluestore_debug_permit_any_bdev_label) { |
4058 | dout(20) << __func__ << " bdev " << path << " fsid " << label.osd_uuid | |
4059 | << " and fsid " << fsid << " check bypassed" << dendl; | |
4060 | } | |
4061 | else if (label.osd_uuid != fsid) { | |
7c673cae FG |
4062 | derr << __func__ << " bdev " << path << " fsid " << label.osd_uuid |
4063 | << " does not match our fsid " << fsid << dendl; | |
4064 | return -EIO; | |
4065 | } | |
4066 | } | |
4067 | return 0; | |
4068 | } | |
4069 | ||
4070 | void BlueStore::_set_alloc_sizes(void) | |
4071 | { | |
7c673cae FG |
4072 | max_alloc_size = cct->_conf->bluestore_max_alloc_size; |
4073 | ||
4074 | if (cct->_conf->bluestore_prefer_deferred_size) { | |
4075 | prefer_deferred_size = cct->_conf->bluestore_prefer_deferred_size; | |
4076 | } else { | |
4077 | assert(bdev); | |
4078 | if (bdev->is_rotational()) { | |
4079 | prefer_deferred_size = cct->_conf->bluestore_prefer_deferred_size_hdd; | |
4080 | } else { | |
4081 | prefer_deferred_size = cct->_conf->bluestore_prefer_deferred_size_ssd; | |
4082 | } | |
4083 | } | |
4084 | ||
4085 | if (cct->_conf->bluestore_deferred_batch_ops) { | |
4086 | deferred_batch_ops = cct->_conf->bluestore_deferred_batch_ops; | |
4087 | } else { | |
4088 | assert(bdev); | |
4089 | if (bdev->is_rotational()) { | |
4090 | deferred_batch_ops = cct->_conf->bluestore_deferred_batch_ops_hdd; | |
4091 | } else { | |
4092 | deferred_batch_ops = cct->_conf->bluestore_deferred_batch_ops_ssd; | |
4093 | } | |
4094 | } | |
4095 | ||
4096 | dout(10) << __func__ << " min_alloc_size 0x" << std::hex << min_alloc_size | |
4097 | << std::dec << " order " << min_alloc_size_order | |
4098 | << " max_alloc_size 0x" << std::hex << max_alloc_size | |
4099 | << " prefer_deferred_size 0x" << prefer_deferred_size | |
4100 | << std::dec | |
4101 | << " deferred_batch_ops " << deferred_batch_ops | |
4102 | << dendl; | |
4103 | } | |
4104 | ||
4105 | int BlueStore::_open_bdev(bool create) | |
4106 | { | |
4107 | assert(bdev == NULL); | |
4108 | string p = path + "/block"; | |
4109 | bdev = BlockDevice::create(cct, p, aio_cb, static_cast<void*>(this)); | |
4110 | int r = bdev->open(p); | |
4111 | if (r < 0) | |
4112 | goto fail; | |
4113 | ||
4114 | if (bdev->supported_bdev_label()) { | |
4115 | r = _check_or_set_bdev_label(p, bdev->get_size(), "main", create); | |
4116 | if (r < 0) | |
4117 | goto fail_close; | |
4118 | } | |
4119 | ||
4120 | // initialize global block parameters | |
4121 | block_size = bdev->get_block_size(); | |
4122 | block_mask = ~(block_size - 1); | |
4123 | block_size_order = ctz(block_size); | |
4124 | assert(block_size == 1u << block_size_order); | |
224ce89b WB |
4125 | // and set cache_size based on device type |
4126 | r = _set_cache_sizes(); | |
4127 | if (r < 0) { | |
4128 | goto fail_close; | |
4129 | } | |
7c673cae FG |
4130 | return 0; |
4131 | ||
4132 | fail_close: | |
4133 | bdev->close(); | |
4134 | fail: | |
4135 | delete bdev; | |
4136 | bdev = NULL; | |
4137 | return r; | |
4138 | } | |
4139 | ||
4140 | void BlueStore::_close_bdev() | |
4141 | { | |
4142 | assert(bdev); | |
4143 | bdev->close(); | |
4144 | delete bdev; | |
4145 | bdev = NULL; | |
4146 | } | |
4147 | ||
4148 | int BlueStore::_open_fm(bool create) | |
4149 | { | |
4150 | assert(fm == NULL); | |
4151 | fm = FreelistManager::create(cct, freelist_type, db, PREFIX_ALLOC); | |
4152 | ||
4153 | if (create) { | |
4154 | // initialize freespace | |
4155 | dout(20) << __func__ << " initializing freespace" << dendl; | |
4156 | KeyValueDB::Transaction t = db->get_transaction(); | |
4157 | { | |
4158 | bufferlist bl; | |
4159 | bl.append(freelist_type); | |
4160 | t->set(PREFIX_SUPER, "freelist_type", bl); | |
4161 | } | |
4162 | fm->create(bdev->get_size(), t); | |
4163 | ||
4164 | // allocate superblock reserved space. note that we do not mark | |
4165 | // bluefs space as allocated in the freelist; we instead rely on | |
4166 | // bluefs_extents. | |
4167 | fm->allocate(0, SUPER_RESERVED, t); | |
4168 | ||
4169 | uint64_t reserved = 0; | |
4170 | if (cct->_conf->bluestore_bluefs) { | |
4171 | assert(bluefs_extents.num_intervals() == 1); | |
4172 | interval_set<uint64_t>::iterator p = bluefs_extents.begin(); | |
4173 | reserved = p.get_start() + p.get_len(); | |
4174 | dout(20) << __func__ << " reserved 0x" << std::hex << reserved << std::dec | |
4175 | << " for bluefs" << dendl; | |
4176 | bufferlist bl; | |
4177 | ::encode(bluefs_extents, bl); | |
4178 | t->set(PREFIX_SUPER, "bluefs_extents", bl); | |
4179 | dout(20) << __func__ << " bluefs_extents 0x" << std::hex << bluefs_extents | |
4180 | << std::dec << dendl; | |
4181 | } else { | |
4182 | reserved = SUPER_RESERVED; | |
4183 | } | |
4184 | ||
4185 | if (cct->_conf->bluestore_debug_prefill > 0) { | |
4186 | uint64_t end = bdev->get_size() - reserved; | |
4187 | dout(1) << __func__ << " pre-fragmenting freespace, using " | |
4188 | << cct->_conf->bluestore_debug_prefill << " with max free extent " | |
4189 | << cct->_conf->bluestore_debug_prefragment_max << dendl; | |
4190 | uint64_t start = P2ROUNDUP(reserved, min_alloc_size); | |
4191 | uint64_t max_b = cct->_conf->bluestore_debug_prefragment_max / min_alloc_size; | |
4192 | float r = cct->_conf->bluestore_debug_prefill; | |
4193 | r /= 1.0 - r; | |
4194 | bool stop = false; | |
4195 | ||
4196 | while (!stop && start < end) { | |
4197 | uint64_t l = (rand() % max_b + 1) * min_alloc_size; | |
4198 | if (start + l > end) { | |
4199 | l = end - start; | |
4200 | l = P2ALIGN(l, min_alloc_size); | |
4201 | } | |
4202 | assert(start + l <= end); | |
4203 | ||
4204 | uint64_t u = 1 + (uint64_t)(r * (double)l); | |
4205 | u = P2ROUNDUP(u, min_alloc_size); | |
4206 | if (start + l + u > end) { | |
4207 | u = end - (start + l); | |
4208 | // trim to align so we don't overflow again | |
4209 | u = P2ALIGN(u, min_alloc_size); | |
4210 | stop = true; | |
4211 | } | |
4212 | assert(start + l + u <= end); | |
4213 | ||
4214 | dout(20) << " free 0x" << std::hex << start << "~" << l | |
4215 | << " use 0x" << u << std::dec << dendl; | |
4216 | ||
4217 | if (u == 0) { | |
4218 | // break if u has been trimmed to nothing | |
4219 | break; | |
4220 | } | |
4221 | ||
4222 | fm->allocate(start + l, u, t); | |
4223 | start += l + u; | |
4224 | } | |
4225 | } | |
4226 | db->submit_transaction_sync(t); | |
4227 | } | |
4228 | ||
4229 | int r = fm->init(); | |
4230 | if (r < 0) { | |
4231 | derr << __func__ << " freelist init failed: " << cpp_strerror(r) << dendl; | |
4232 | delete fm; | |
4233 | fm = NULL; | |
4234 | return r; | |
4235 | } | |
4236 | return 0; | |
4237 | } | |
4238 | ||
4239 | void BlueStore::_close_fm() | |
4240 | { | |
4241 | dout(10) << __func__ << dendl; | |
4242 | assert(fm); | |
4243 | fm->shutdown(); | |
4244 | delete fm; | |
4245 | fm = NULL; | |
4246 | } | |
4247 | ||
4248 | int BlueStore::_open_alloc() | |
4249 | { | |
4250 | assert(alloc == NULL); | |
4251 | assert(bdev->get_size()); | |
4252 | alloc = Allocator::create(cct, cct->_conf->bluestore_allocator, | |
4253 | bdev->get_size(), | |
4254 | min_alloc_size); | |
4255 | if (!alloc) { | |
4256 | lderr(cct) << __func__ << " Allocator::unknown alloc type " | |
4257 | << cct->_conf->bluestore_allocator | |
4258 | << dendl; | |
4259 | return -EINVAL; | |
4260 | } | |
4261 | ||
4262 | uint64_t num = 0, bytes = 0; | |
4263 | ||
4264 | dout(1) << __func__ << " opening allocation metadata" << dendl; | |
4265 | // initialize from freelist | |
4266 | fm->enumerate_reset(); | |
4267 | uint64_t offset, length; | |
4268 | while (fm->enumerate_next(&offset, &length)) { | |
4269 | alloc->init_add_free(offset, length); | |
4270 | ++num; | |
4271 | bytes += length; | |
4272 | } | |
224ce89b | 4273 | fm->enumerate_reset(); |
7c673cae FG |
4274 | dout(1) << __func__ << " loaded " << pretty_si_t(bytes) |
4275 | << " in " << num << " extents" | |
4276 | << dendl; | |
4277 | ||
4278 | // also mark bluefs space as allocated | |
4279 | for (auto e = bluefs_extents.begin(); e != bluefs_extents.end(); ++e) { | |
4280 | alloc->init_rm_free(e.get_start(), e.get_len()); | |
4281 | } | |
4282 | dout(10) << __func__ << " marked bluefs_extents 0x" << std::hex | |
4283 | << bluefs_extents << std::dec << " as allocated" << dendl; | |
4284 | ||
4285 | return 0; | |
4286 | } | |
4287 | ||
4288 | void BlueStore::_close_alloc() | |
4289 | { | |
4290 | assert(alloc); | |
4291 | alloc->shutdown(); | |
4292 | delete alloc; | |
4293 | alloc = NULL; | |
4294 | } | |
4295 | ||
4296 | int BlueStore::_open_fsid(bool create) | |
4297 | { | |
4298 | assert(fsid_fd < 0); | |
4299 | int flags = O_RDWR; | |
4300 | if (create) | |
4301 | flags |= O_CREAT; | |
4302 | fsid_fd = ::openat(path_fd, "fsid", flags, 0644); | |
4303 | if (fsid_fd < 0) { | |
4304 | int err = -errno; | |
4305 | derr << __func__ << " " << cpp_strerror(err) << dendl; | |
4306 | return err; | |
4307 | } | |
4308 | return 0; | |
4309 | } | |
4310 | ||
4311 | int BlueStore::_read_fsid(uuid_d *uuid) | |
4312 | { | |
4313 | char fsid_str[40]; | |
4314 | memset(fsid_str, 0, sizeof(fsid_str)); | |
4315 | int ret = safe_read(fsid_fd, fsid_str, sizeof(fsid_str)); | |
4316 | if (ret < 0) { | |
4317 | derr << __func__ << " failed: " << cpp_strerror(ret) << dendl; | |
4318 | return ret; | |
4319 | } | |
4320 | if (ret > 36) | |
4321 | fsid_str[36] = 0; | |
4322 | else | |
4323 | fsid_str[ret] = 0; | |
4324 | if (!uuid->parse(fsid_str)) { | |
4325 | derr << __func__ << " unparsable uuid " << fsid_str << dendl; | |
4326 | return -EINVAL; | |
4327 | } | |
4328 | return 0; | |
4329 | } | |
4330 | ||
4331 | int BlueStore::_write_fsid() | |
4332 | { | |
4333 | int r = ::ftruncate(fsid_fd, 0); | |
4334 | if (r < 0) { | |
4335 | r = -errno; | |
4336 | derr << __func__ << " fsid truncate failed: " << cpp_strerror(r) << dendl; | |
4337 | return r; | |
4338 | } | |
4339 | string str = stringify(fsid) + "\n"; | |
4340 | r = safe_write(fsid_fd, str.c_str(), str.length()); | |
4341 | if (r < 0) { | |
4342 | derr << __func__ << " fsid write failed: " << cpp_strerror(r) << dendl; | |
4343 | return r; | |
4344 | } | |
4345 | r = ::fsync(fsid_fd); | |
4346 | if (r < 0) { | |
4347 | r = -errno; | |
4348 | derr << __func__ << " fsid fsync failed: " << cpp_strerror(r) << dendl; | |
4349 | return r; | |
4350 | } | |
4351 | return 0; | |
4352 | } | |
4353 | ||
4354 | void BlueStore::_close_fsid() | |
4355 | { | |
4356 | VOID_TEMP_FAILURE_RETRY(::close(fsid_fd)); | |
4357 | fsid_fd = -1; | |
4358 | } | |
4359 | ||
4360 | int BlueStore::_lock_fsid() | |
4361 | { | |
4362 | struct flock l; | |
4363 | memset(&l, 0, sizeof(l)); | |
4364 | l.l_type = F_WRLCK; | |
4365 | l.l_whence = SEEK_SET; | |
4366 | int r = ::fcntl(fsid_fd, F_SETLK, &l); | |
4367 | if (r < 0) { | |
4368 | int err = errno; | |
4369 | derr << __func__ << " failed to lock " << path << "/fsid" | |
4370 | << " (is another ceph-osd still running?)" | |
4371 | << cpp_strerror(err) << dendl; | |
4372 | return -err; | |
4373 | } | |
4374 | return 0; | |
4375 | } | |
4376 | ||
31f18b77 FG |
4377 | bool BlueStore::is_rotational() |
4378 | { | |
4379 | if (bdev) { | |
4380 | return bdev->is_rotational(); | |
4381 | } | |
4382 | ||
4383 | bool rotational = true; | |
4384 | int r = _open_path(); | |
4385 | if (r < 0) | |
4386 | goto out; | |
4387 | r = _open_fsid(false); | |
4388 | if (r < 0) | |
4389 | goto out_path; | |
4390 | r = _read_fsid(&fsid); | |
4391 | if (r < 0) | |
4392 | goto out_fsid; | |
4393 | r = _lock_fsid(); | |
4394 | if (r < 0) | |
4395 | goto out_fsid; | |
4396 | r = _open_bdev(false); | |
4397 | if (r < 0) | |
4398 | goto out_fsid; | |
4399 | rotational = bdev->is_rotational(); | |
4400 | _close_bdev(); | |
4401 | out_fsid: | |
4402 | _close_fsid(); | |
4403 | out_path: | |
4404 | _close_path(); | |
4405 | out: | |
4406 | return rotational; | |
4407 | } | |
4408 | ||
d2e6a577 FG |
4409 | bool BlueStore::is_journal_rotational() |
4410 | { | |
4411 | if (!bluefs) { | |
4412 | dout(5) << __func__ << " bluefs disabled, default to store media type" | |
4413 | << dendl; | |
4414 | return is_rotational(); | |
4415 | } | |
4416 | dout(10) << __func__ << " " << (int)bluefs->wal_is_rotational() << dendl; | |
4417 | return bluefs->wal_is_rotational(); | |
4418 | } | |
4419 | ||
7c673cae FG |
4420 | bool BlueStore::test_mount_in_use() |
4421 | { | |
4422 | // most error conditions mean the mount is not in use (e.g., because | |
4423 | // it doesn't exist). only if we fail to lock do we conclude it is | |
4424 | // in use. | |
4425 | bool ret = false; | |
4426 | int r = _open_path(); | |
4427 | if (r < 0) | |
4428 | return false; | |
4429 | r = _open_fsid(false); | |
4430 | if (r < 0) | |
4431 | goto out_path; | |
4432 | r = _lock_fsid(); | |
4433 | if (r < 0) | |
4434 | ret = true; // if we can't lock, it is in use | |
4435 | _close_fsid(); | |
4436 | out_path: | |
4437 | _close_path(); | |
4438 | return ret; | |
4439 | } | |
4440 | ||
4441 | int BlueStore::_open_db(bool create) | |
4442 | { | |
4443 | int r; | |
4444 | assert(!db); | |
4445 | string fn = path + "/db"; | |
4446 | string options; | |
4447 | stringstream err; | |
4448 | ceph::shared_ptr<Int64ArrayMergeOperator> merge_op(new Int64ArrayMergeOperator); | |
4449 | ||
4450 | string kv_backend; | |
4451 | if (create) { | |
4452 | kv_backend = cct->_conf->bluestore_kvbackend; | |
4453 | } else { | |
4454 | r = read_meta("kv_backend", &kv_backend); | |
4455 | if (r < 0) { | |
4456 | derr << __func__ << " unable to read 'kv_backend' meta" << dendl; | |
4457 | return -EIO; | |
4458 | } | |
4459 | } | |
4460 | dout(10) << __func__ << " kv_backend = " << kv_backend << dendl; | |
4461 | ||
4462 | bool do_bluefs; | |
4463 | if (create) { | |
4464 | do_bluefs = cct->_conf->bluestore_bluefs; | |
4465 | } else { | |
4466 | string s; | |
4467 | r = read_meta("bluefs", &s); | |
4468 | if (r < 0) { | |
4469 | derr << __func__ << " unable to read 'bluefs' meta" << dendl; | |
4470 | return -EIO; | |
4471 | } | |
4472 | if (s == "1") { | |
4473 | do_bluefs = true; | |
4474 | } else if (s == "0") { | |
4475 | do_bluefs = false; | |
4476 | } else { | |
4477 | derr << __func__ << " bluefs = " << s << " : not 0 or 1, aborting" | |
4478 | << dendl; | |
4479 | return -EIO; | |
4480 | } | |
4481 | } | |
4482 | dout(10) << __func__ << " do_bluefs = " << do_bluefs << dendl; | |
4483 | ||
4484 | rocksdb::Env *env = NULL; | |
4485 | if (do_bluefs) { | |
4486 | dout(10) << __func__ << " initializing bluefs" << dendl; | |
4487 | if (kv_backend != "rocksdb") { | |
4488 | derr << " backend must be rocksdb to use bluefs" << dendl; | |
4489 | return -EINVAL; | |
4490 | } | |
4491 | bluefs = new BlueFS(cct); | |
4492 | ||
4493 | string bfn; | |
4494 | struct stat st; | |
4495 | ||
4496 | bfn = path + "/block.db"; | |
4497 | if (::stat(bfn.c_str(), &st) == 0) { | |
4498 | r = bluefs->add_block_device(BlueFS::BDEV_DB, bfn); | |
4499 | if (r < 0) { | |
4500 | derr << __func__ << " add block device(" << bfn << ") returned: " | |
4501 | << cpp_strerror(r) << dendl; | |
4502 | goto free_bluefs; | |
4503 | } | |
4504 | ||
4505 | if (bluefs->bdev_support_label(BlueFS::BDEV_DB)) { | |
4506 | r = _check_or_set_bdev_label( | |
4507 | bfn, | |
4508 | bluefs->get_block_device_size(BlueFS::BDEV_DB), | |
4509 | "bluefs db", create); | |
4510 | if (r < 0) { | |
4511 | derr << __func__ | |
4512 | << " check block device(" << bfn << ") label returned: " | |
4513 | << cpp_strerror(r) << dendl; | |
4514 | goto free_bluefs; | |
4515 | } | |
4516 | } | |
4517 | if (create) { | |
4518 | bluefs->add_block_extent( | |
4519 | BlueFS::BDEV_DB, | |
4520 | SUPER_RESERVED, | |
4521 | bluefs->get_block_device_size(BlueFS::BDEV_DB) - SUPER_RESERVED); | |
4522 | } | |
4523 | bluefs_shared_bdev = BlueFS::BDEV_SLOW; | |
4524 | bluefs_single_shared_device = false; | |
31f18b77 | 4525 | } else if (::lstat(bfn.c_str(), &st) == -1) { |
7c673cae | 4526 | bluefs_shared_bdev = BlueFS::BDEV_DB; |
31f18b77 FG |
4527 | } else { |
4528 | //symlink exist is bug | |
4529 | derr << __func__ << " " << bfn << " link target doesn't exist" << dendl; | |
4530 | r = -errno; | |
4531 | goto free_bluefs; | |
7c673cae FG |
4532 | } |
4533 | ||
4534 | // shared device | |
4535 | bfn = path + "/block"; | |
4536 | r = bluefs->add_block_device(bluefs_shared_bdev, bfn); | |
4537 | if (r < 0) { | |
4538 | derr << __func__ << " add block device(" << bfn << ") returned: " | |
4539 | << cpp_strerror(r) << dendl; | |
4540 | goto free_bluefs; | |
4541 | } | |
4542 | if (create) { | |
4543 | // note: we always leave the first SUPER_RESERVED (8k) of the device unused | |
4544 | uint64_t initial = | |
4545 | bdev->get_size() * (cct->_conf->bluestore_bluefs_min_ratio + | |
4546 | cct->_conf->bluestore_bluefs_gift_ratio); | |
4547 | initial = MAX(initial, cct->_conf->bluestore_bluefs_min); | |
4548 | // align to bluefs's alloc_size | |
4549 | initial = P2ROUNDUP(initial, cct->_conf->bluefs_alloc_size); | |
31f18b77 FG |
4550 | // put bluefs in the middle of the device in case it is an HDD |
4551 | uint64_t start = P2ALIGN((bdev->get_size() - initial) / 2, | |
4552 | cct->_conf->bluefs_alloc_size); | |
4553 | bluefs->add_block_extent(bluefs_shared_bdev, start, initial); | |
4554 | bluefs_extents.insert(start, initial); | |
7c673cae FG |
4555 | } |
4556 | ||
4557 | bfn = path + "/block.wal"; | |
4558 | if (::stat(bfn.c_str(), &st) == 0) { | |
4559 | r = bluefs->add_block_device(BlueFS::BDEV_WAL, bfn); | |
4560 | if (r < 0) { | |
4561 | derr << __func__ << " add block device(" << bfn << ") returned: " | |
4562 | << cpp_strerror(r) << dendl; | |
4563 | goto free_bluefs; | |
4564 | } | |
4565 | ||
4566 | if (bluefs->bdev_support_label(BlueFS::BDEV_WAL)) { | |
4567 | r = _check_or_set_bdev_label( | |
4568 | bfn, | |
4569 | bluefs->get_block_device_size(BlueFS::BDEV_WAL), | |
4570 | "bluefs wal", create); | |
4571 | if (r < 0) { | |
4572 | derr << __func__ << " check block device(" << bfn | |
4573 | << ") label returned: " << cpp_strerror(r) << dendl; | |
4574 | goto free_bluefs; | |
4575 | } | |
4576 | } | |
4577 | ||
4578 | if (create) { | |
4579 | bluefs->add_block_extent( | |
4580 | BlueFS::BDEV_WAL, BDEV_LABEL_BLOCK_SIZE, | |
4581 | bluefs->get_block_device_size(BlueFS::BDEV_WAL) - | |
4582 | BDEV_LABEL_BLOCK_SIZE); | |
4583 | } | |
4584 | cct->_conf->set_val("rocksdb_separate_wal_dir", "true"); | |
4585 | bluefs_single_shared_device = false; | |
31f18b77 | 4586 | } else if (::lstat(bfn.c_str(), &st) == -1) { |
7c673cae | 4587 | cct->_conf->set_val("rocksdb_separate_wal_dir", "false"); |
31f18b77 FG |
4588 | } else { |
4589 | //symlink exist is bug | |
4590 | derr << __func__ << " " << bfn << " link target doesn't exist" << dendl; | |
4591 | r = -errno; | |
4592 | goto free_bluefs; | |
7c673cae FG |
4593 | } |
4594 | ||
4595 | if (create) { | |
4596 | bluefs->mkfs(fsid); | |
4597 | } | |
4598 | r = bluefs->mount(); | |
4599 | if (r < 0) { | |
4600 | derr << __func__ << " failed bluefs mount: " << cpp_strerror(r) << dendl; | |
4601 | goto free_bluefs; | |
4602 | } | |
4603 | if (cct->_conf->bluestore_bluefs_env_mirror) { | |
4604 | rocksdb::Env *a = new BlueRocksEnv(bluefs); | |
4605 | rocksdb::Env *b = rocksdb::Env::Default(); | |
4606 | if (create) { | |
4607 | string cmd = "rm -rf " + path + "/db " + | |
4608 | path + "/db.slow " + | |
4609 | path + "/db.wal"; | |
4610 | int r = system(cmd.c_str()); | |
4611 | (void)r; | |
4612 | } | |
4613 | env = new rocksdb::EnvMirror(b, a, false, true); | |
4614 | } else { | |
4615 | env = new BlueRocksEnv(bluefs); | |
4616 | ||
4617 | // simplify the dir names, too, as "seen" by rocksdb | |
4618 | fn = "db"; | |
4619 | } | |
4620 | ||
4621 | if (bluefs_shared_bdev == BlueFS::BDEV_SLOW) { | |
4622 | // we have both block.db and block; tell rocksdb! | |
4623 | // note: the second (last) size value doesn't really matter | |
4624 | ostringstream db_paths; | |
4625 | uint64_t db_size = bluefs->get_block_device_size(BlueFS::BDEV_DB); | |
4626 | uint64_t slow_size = bluefs->get_block_device_size(BlueFS::BDEV_SLOW); | |
4627 | db_paths << fn << "," | |
4628 | << (uint64_t)(db_size * 95 / 100) << " " | |
4629 | << fn + ".slow" << "," | |
4630 | << (uint64_t)(slow_size * 95 / 100); | |
4631 | cct->_conf->set_val("rocksdb_db_paths", db_paths.str(), false); | |
4632 | dout(10) << __func__ << " set rocksdb_db_paths to " | |
4633 | << cct->_conf->get_val<std::string>("rocksdb_db_paths") << dendl; | |
4634 | } | |
4635 | ||
4636 | if (create) { | |
4637 | env->CreateDir(fn); | |
4638 | if (cct->_conf->rocksdb_separate_wal_dir) | |
4639 | env->CreateDir(fn + ".wal"); | |
4640 | if (cct->_conf->get_val<std::string>("rocksdb_db_paths").length()) | |
4641 | env->CreateDir(fn + ".slow"); | |
4642 | } | |
4643 | } else if (create) { | |
4644 | int r = ::mkdir(fn.c_str(), 0755); | |
4645 | if (r < 0) | |
4646 | r = -errno; | |
4647 | if (r < 0 && r != -EEXIST) { | |
4648 | derr << __func__ << " failed to create " << fn << ": " << cpp_strerror(r) | |
4649 | << dendl; | |
4650 | return r; | |
4651 | } | |
4652 | ||
4653 | // wal_dir, too! | |
4654 | if (cct->_conf->rocksdb_separate_wal_dir) { | |
4655 | string walfn = path + "/db.wal"; | |
4656 | r = ::mkdir(walfn.c_str(), 0755); | |
4657 | if (r < 0) | |
4658 | r = -errno; | |
4659 | if (r < 0 && r != -EEXIST) { | |
4660 | derr << __func__ << " failed to create " << walfn | |
4661 | << ": " << cpp_strerror(r) | |
4662 | << dendl; | |
4663 | return r; | |
4664 | } | |
4665 | } | |
4666 | } | |
4667 | ||
4668 | db = KeyValueDB::create(cct, | |
4669 | kv_backend, | |
4670 | fn, | |
4671 | static_cast<void*>(env)); | |
4672 | if (!db) { | |
4673 | derr << __func__ << " error creating db" << dendl; | |
4674 | if (bluefs) { | |
4675 | bluefs->umount(); | |
4676 | delete bluefs; | |
4677 | bluefs = NULL; | |
4678 | } | |
4679 | // delete env manually here since we can't depend on db to do this | |
4680 | // under this case | |
4681 | delete env; | |
4682 | env = NULL; | |
4683 | return -EIO; | |
4684 | } | |
4685 | ||
4686 | FreelistManager::setup_merge_operators(db); | |
4687 | db->set_merge_operator(PREFIX_STAT, merge_op); | |
4688 | ||
224ce89b | 4689 | db->set_cache_size(cache_size * cache_kv_ratio); |
31f18b77 | 4690 | |
7c673cae FG |
4691 | if (kv_backend == "rocksdb") |
4692 | options = cct->_conf->bluestore_rocksdb_options; | |
4693 | db->init(options); | |
4694 | if (create) | |
4695 | r = db->create_and_open(err); | |
4696 | else | |
4697 | r = db->open(err); | |
4698 | if (r) { | |
4699 | derr << __func__ << " erroring opening db: " << err.str() << dendl; | |
4700 | if (bluefs) { | |
4701 | bluefs->umount(); | |
4702 | delete bluefs; | |
4703 | bluefs = NULL; | |
4704 | } | |
4705 | delete db; | |
4706 | db = NULL; | |
4707 | return -EIO; | |
4708 | } | |
4709 | dout(1) << __func__ << " opened " << kv_backend | |
4710 | << " path " << fn << " options " << options << dendl; | |
4711 | return 0; | |
4712 | ||
4713 | free_bluefs: | |
4714 | assert(bluefs); | |
4715 | delete bluefs; | |
4716 | bluefs = NULL; | |
4717 | return r; | |
4718 | } | |
4719 | ||
4720 | void BlueStore::_close_db() | |
4721 | { | |
4722 | assert(db); | |
4723 | delete db; | |
4724 | db = NULL; | |
4725 | if (bluefs) { | |
4726 | bluefs->umount(); | |
4727 | delete bluefs; | |
4728 | bluefs = NULL; | |
4729 | } | |
4730 | } | |
4731 | ||
4732 | int BlueStore::_reconcile_bluefs_freespace() | |
4733 | { | |
4734 | dout(10) << __func__ << dendl; | |
4735 | interval_set<uint64_t> bset; | |
4736 | int r = bluefs->get_block_extents(bluefs_shared_bdev, &bset); | |
4737 | assert(r == 0); | |
4738 | if (bset == bluefs_extents) { | |
4739 | dout(10) << __func__ << " we agree bluefs has 0x" << std::hex << bset | |
4740 | << std::dec << dendl; | |
4741 | return 0; | |
4742 | } | |
4743 | dout(10) << __func__ << " bluefs says 0x" << std::hex << bset << std::dec | |
4744 | << dendl; | |
4745 | dout(10) << __func__ << " super says 0x" << std::hex << bluefs_extents | |
4746 | << std::dec << dendl; | |
4747 | ||
4748 | interval_set<uint64_t> overlap; | |
4749 | overlap.intersection_of(bset, bluefs_extents); | |
4750 | ||
4751 | bset.subtract(overlap); | |
4752 | if (!bset.empty()) { | |
4753 | derr << __func__ << " bluefs extra 0x" << std::hex << bset << std::dec | |
4754 | << dendl; | |
4755 | return -EIO; | |
4756 | } | |
4757 | ||
4758 | interval_set<uint64_t> super_extra; | |
4759 | super_extra = bluefs_extents; | |
4760 | super_extra.subtract(overlap); | |
4761 | if (!super_extra.empty()) { | |
4762 | // This is normal: it can happen if we commit to give extents to | |
4763 | // bluefs and we crash before bluefs commits that it owns them. | |
4764 | dout(10) << __func__ << " super extra " << super_extra << dendl; | |
4765 | for (interval_set<uint64_t>::iterator p = super_extra.begin(); | |
4766 | p != super_extra.end(); | |
4767 | ++p) { | |
4768 | bluefs->add_block_extent(bluefs_shared_bdev, p.get_start(), p.get_len()); | |
4769 | } | |
4770 | } | |
4771 | ||
4772 | return 0; | |
4773 | } | |
4774 | ||
4775 | int BlueStore::_balance_bluefs_freespace(PExtentVector *extents) | |
4776 | { | |
4777 | int ret = 0; | |
4778 | assert(bluefs); | |
4779 | ||
4780 | vector<pair<uint64_t,uint64_t>> bluefs_usage; // <free, total> ... | |
4781 | bluefs->get_usage(&bluefs_usage); | |
4782 | assert(bluefs_usage.size() > bluefs_shared_bdev); | |
4783 | ||
4784 | // fixme: look at primary bdev only for now | |
4785 | uint64_t bluefs_free = bluefs_usage[bluefs_shared_bdev].first; | |
4786 | uint64_t bluefs_total = bluefs_usage[bluefs_shared_bdev].second; | |
4787 | float bluefs_free_ratio = (float)bluefs_free / (float)bluefs_total; | |
4788 | ||
4789 | uint64_t my_free = alloc->get_free(); | |
4790 | uint64_t total = bdev->get_size(); | |
4791 | float my_free_ratio = (float)my_free / (float)total; | |
4792 | ||
4793 | uint64_t total_free = bluefs_free + my_free; | |
4794 | ||
4795 | float bluefs_ratio = (float)bluefs_free / (float)total_free; | |
4796 | ||
4797 | dout(10) << __func__ | |
4798 | << " bluefs " << pretty_si_t(bluefs_free) | |
4799 | << " free (" << bluefs_free_ratio | |
4800 | << ") bluestore " << pretty_si_t(my_free) | |
4801 | << " free (" << my_free_ratio | |
4802 | << "), bluefs_ratio " << bluefs_ratio | |
4803 | << dendl; | |
4804 | ||
4805 | uint64_t gift = 0; | |
4806 | uint64_t reclaim = 0; | |
4807 | if (bluefs_ratio < cct->_conf->bluestore_bluefs_min_ratio) { | |
4808 | gift = cct->_conf->bluestore_bluefs_gift_ratio * total_free; | |
4809 | dout(10) << __func__ << " bluefs_ratio " << bluefs_ratio | |
4810 | << " < min_ratio " << cct->_conf->bluestore_bluefs_min_ratio | |
4811 | << ", should gift " << pretty_si_t(gift) << dendl; | |
4812 | } else if (bluefs_ratio > cct->_conf->bluestore_bluefs_max_ratio) { | |
4813 | reclaim = cct->_conf->bluestore_bluefs_reclaim_ratio * total_free; | |
4814 | if (bluefs_total - reclaim < cct->_conf->bluestore_bluefs_min) | |
4815 | reclaim = bluefs_total - cct->_conf->bluestore_bluefs_min; | |
4816 | dout(10) << __func__ << " bluefs_ratio " << bluefs_ratio | |
4817 | << " > max_ratio " << cct->_conf->bluestore_bluefs_max_ratio | |
4818 | << ", should reclaim " << pretty_si_t(reclaim) << dendl; | |
4819 | } | |
4820 | if (bluefs_total < cct->_conf->bluestore_bluefs_min && | |
4821 | cct->_conf->bluestore_bluefs_min < | |
4822 | (uint64_t)(cct->_conf->bluestore_bluefs_max_ratio * total_free)) { | |
4823 | uint64_t g = cct->_conf->bluestore_bluefs_min - bluefs_total; | |
4824 | dout(10) << __func__ << " bluefs_total " << bluefs_total | |
4825 | << " < min " << cct->_conf->bluestore_bluefs_min | |
4826 | << ", should gift " << pretty_si_t(g) << dendl; | |
4827 | if (g > gift) | |
4828 | gift = g; | |
4829 | reclaim = 0; | |
4830 | } | |
4831 | ||
4832 | if (gift) { | |
4833 | // round up to alloc size | |
4834 | gift = P2ROUNDUP(gift, cct->_conf->bluefs_alloc_size); | |
4835 | ||
4836 | // hard cap to fit into 32 bits | |
4837 | gift = MIN(gift, 1ull<<31); | |
4838 | dout(10) << __func__ << " gifting " << gift | |
4839 | << " (" << pretty_si_t(gift) << ")" << dendl; | |
4840 | ||
4841 | // fixme: just do one allocation to start... | |
4842 | int r = alloc->reserve(gift); | |
4843 | assert(r == 0); | |
4844 | ||
4845 | AllocExtentVector exts; | |
4846 | int64_t alloc_len = alloc->allocate(gift, cct->_conf->bluefs_alloc_size, | |
4847 | 0, 0, &exts); | |
4848 | ||
4849 | if (alloc_len < (int64_t)gift) { | |
4850 | derr << __func__ << " allocate failed on 0x" << std::hex << gift | |
4851 | << " min_alloc_size 0x" << min_alloc_size << std::dec << dendl; | |
4852 | alloc->dump(); | |
4853 | assert(0 == "allocate failed, wtf"); | |
4854 | return -ENOSPC; | |
4855 | } | |
4856 | for (auto& p : exts) { | |
4857 | bluestore_pextent_t e = bluestore_pextent_t(p); | |
4858 | dout(1) << __func__ << " gifting " << e << " to bluefs" << dendl; | |
4859 | extents->push_back(e); | |
4860 | } | |
4861 | gift = 0; | |
4862 | ||
4863 | ret = 1; | |
4864 | } | |
4865 | ||
4866 | // reclaim from bluefs? | |
4867 | if (reclaim) { | |
4868 | // round up to alloc size | |
4869 | reclaim = P2ROUNDUP(reclaim, cct->_conf->bluefs_alloc_size); | |
4870 | ||
4871 | // hard cap to fit into 32 bits | |
4872 | reclaim = MIN(reclaim, 1ull<<31); | |
4873 | dout(10) << __func__ << " reclaiming " << reclaim | |
4874 | << " (" << pretty_si_t(reclaim) << ")" << dendl; | |
4875 | ||
4876 | while (reclaim > 0) { | |
4877 | // NOTE: this will block and do IO. | |
4878 | AllocExtentVector extents; | |
4879 | int r = bluefs->reclaim_blocks(bluefs_shared_bdev, reclaim, | |
4880 | &extents); | |
4881 | if (r < 0) { | |
4882 | derr << __func__ << " failed to reclaim space from bluefs" | |
4883 | << dendl; | |
4884 | break; | |
4885 | } | |
4886 | for (auto e : extents) { | |
4887 | bluefs_extents.erase(e.offset, e.length); | |
4888 | bluefs_extents_reclaiming.insert(e.offset, e.length); | |
4889 | reclaim -= e.length; | |
4890 | } | |
4891 | } | |
4892 | ||
4893 | ret = 1; | |
4894 | } | |
4895 | ||
4896 | return ret; | |
4897 | } | |
4898 | ||
4899 | void BlueStore::_commit_bluefs_freespace( | |
4900 | const PExtentVector& bluefs_gift_extents) | |
4901 | { | |
4902 | dout(10) << __func__ << dendl; | |
4903 | for (auto& p : bluefs_gift_extents) { | |
4904 | bluefs->add_block_extent(bluefs_shared_bdev, p.offset, p.length); | |
4905 | } | |
4906 | } | |
4907 | ||
4908 | int BlueStore::_open_collections(int *errors) | |
4909 | { | |
4910 | assert(coll_map.empty()); | |
4911 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_COLL); | |
4912 | for (it->upper_bound(string()); | |
4913 | it->valid(); | |
4914 | it->next()) { | |
4915 | coll_t cid; | |
4916 | if (cid.parse(it->key())) { | |
4917 | CollectionRef c( | |
4918 | new Collection( | |
4919 | this, | |
4920 | cache_shards[cid.hash_to_shard(cache_shards.size())], | |
4921 | cid)); | |
4922 | bufferlist bl = it->value(); | |
4923 | bufferlist::iterator p = bl.begin(); | |
4924 | try { | |
4925 | ::decode(c->cnode, p); | |
4926 | } catch (buffer::error& e) { | |
4927 | derr << __func__ << " failed to decode cnode, key:" | |
4928 | << pretty_binary_string(it->key()) << dendl; | |
4929 | return -EIO; | |
4930 | } | |
4931 | dout(20) << __func__ << " opened " << cid << " " << c << dendl; | |
4932 | coll_map[cid] = c; | |
4933 | } else { | |
4934 | derr << __func__ << " unrecognized collection " << it->key() << dendl; | |
4935 | if (errors) | |
4936 | (*errors)++; | |
4937 | } | |
4938 | } | |
4939 | return 0; | |
4940 | } | |
4941 | ||
224ce89b | 4942 | void BlueStore::_open_statfs() |
31f18b77 FG |
4943 | { |
4944 | bufferlist bl; | |
4945 | int r = db->get(PREFIX_STAT, "bluestore_statfs", &bl); | |
4946 | if (r >= 0) { | |
4947 | if (size_t(bl.length()) >= sizeof(vstatfs.values)) { | |
4948 | auto it = bl.begin(); | |
4949 | vstatfs.decode(it); | |
224ce89b | 4950 | } else { |
31f18b77 FG |
4951 | dout(10) << __func__ << " store_statfs is corrupt, using empty" << dendl; |
4952 | } | |
4953 | } | |
4954 | else { | |
4955 | dout(10) << __func__ << " store_statfs missed, using empty" << dendl; | |
4956 | } | |
4957 | } | |
4958 | ||
7c673cae FG |
4959 | int BlueStore::_setup_block_symlink_or_file( |
4960 | string name, | |
4961 | string epath, | |
4962 | uint64_t size, | |
4963 | bool create) | |
4964 | { | |
4965 | dout(20) << __func__ << " name " << name << " path " << epath | |
4966 | << " size " << size << " create=" << (int)create << dendl; | |
4967 | int r = 0; | |
4968 | int flags = O_RDWR; | |
4969 | if (create) | |
4970 | flags |= O_CREAT; | |
4971 | if (epath.length()) { | |
4972 | r = ::symlinkat(epath.c_str(), path_fd, name.c_str()); | |
4973 | if (r < 0) { | |
4974 | r = -errno; | |
4975 | derr << __func__ << " failed to create " << name << " symlink to " | |
4976 | << epath << ": " << cpp_strerror(r) << dendl; | |
4977 | return r; | |
4978 | } | |
4979 | ||
4980 | if (!epath.compare(0, strlen(SPDK_PREFIX), SPDK_PREFIX)) { | |
4981 | int fd = ::openat(path_fd, epath.c_str(), flags, 0644); | |
4982 | if (fd < 0) { | |
4983 | r = -errno; | |
4984 | derr << __func__ << " failed to open " << epath << " file: " | |
4985 | << cpp_strerror(r) << dendl; | |
4986 | return r; | |
4987 | } | |
4988 | string serial_number = epath.substr(strlen(SPDK_PREFIX)); | |
4989 | r = ::write(fd, serial_number.c_str(), serial_number.size()); | |
4990 | assert(r == (int)serial_number.size()); | |
4991 | dout(1) << __func__ << " created " << name << " symlink to " | |
4992 | << epath << dendl; | |
4993 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
4994 | } | |
4995 | } | |
4996 | if (size) { | |
4997 | int fd = ::openat(path_fd, name.c_str(), flags, 0644); | |
4998 | if (fd >= 0) { | |
4999 | // block file is present | |
5000 | struct stat st; | |
5001 | int r = ::fstat(fd, &st); | |
5002 | if (r == 0 && | |
5003 | S_ISREG(st.st_mode) && // if it is a regular file | |
5004 | st.st_size == 0) { // and is 0 bytes | |
5005 | r = ::ftruncate(fd, size); | |
5006 | if (r < 0) { | |
5007 | r = -errno; | |
5008 | derr << __func__ << " failed to resize " << name << " file to " | |
5009 | << size << ": " << cpp_strerror(r) << dendl; | |
5010 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
5011 | return r; | |
5012 | } | |
5013 | ||
5014 | if (cct->_conf->bluestore_block_preallocate_file) { | |
5015 | #ifdef HAVE_POSIX_FALLOCATE | |
5016 | r = ::posix_fallocate(fd, 0, size); | |
5017 | if (r) { | |
5018 | derr << __func__ << " failed to prefallocate " << name << " file to " | |
5019 | << size << ": " << cpp_strerror(r) << dendl; | |
5020 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
5021 | return -r; | |
5022 | } | |
5023 | #else | |
5024 | char data[1024*128]; | |
5025 | for (uint64_t off = 0; off < size; off += sizeof(data)) { | |
5026 | if (off + sizeof(data) > size) | |
5027 | r = ::write(fd, data, size - off); | |
5028 | else | |
5029 | r = ::write(fd, data, sizeof(data)); | |
5030 | if (r < 0) { | |
5031 | r = -errno; | |
5032 | derr << __func__ << " failed to prefallocate w/ write " << name << " file to " | |
5033 | << size << ": " << cpp_strerror(r) << dendl; | |
5034 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
5035 | return r; | |
5036 | } | |
5037 | } | |
5038 | #endif | |
5039 | } | |
5040 | dout(1) << __func__ << " resized " << name << " file to " | |
5041 | << pretty_si_t(size) << "B" << dendl; | |
5042 | } | |
5043 | VOID_TEMP_FAILURE_RETRY(::close(fd)); | |
5044 | } else { | |
5045 | int r = -errno; | |
5046 | if (r != -ENOENT) { | |
5047 | derr << __func__ << " failed to open " << name << " file: " | |
5048 | << cpp_strerror(r) << dendl; | |
5049 | return r; | |
5050 | } | |
5051 | } | |
5052 | } | |
5053 | return 0; | |
5054 | } | |
5055 | ||
5056 | int BlueStore::mkfs() | |
5057 | { | |
5058 | dout(1) << __func__ << " path " << path << dendl; | |
5059 | int r; | |
5060 | uuid_d old_fsid; | |
5061 | ||
5062 | { | |
5063 | string done; | |
5064 | r = read_meta("mkfs_done", &done); | |
5065 | if (r == 0) { | |
5066 | dout(1) << __func__ << " already created" << dendl; | |
5067 | if (cct->_conf->bluestore_fsck_on_mkfs) { | |
5068 | r = fsck(cct->_conf->bluestore_fsck_on_mkfs_deep); | |
5069 | if (r < 0) { | |
5070 | derr << __func__ << " fsck found fatal error: " << cpp_strerror(r) | |
5071 | << dendl; | |
5072 | return r; | |
5073 | } | |
5074 | if (r > 0) { | |
5075 | derr << __func__ << " fsck found " << r << " errors" << dendl; | |
5076 | r = -EIO; | |
5077 | } | |
5078 | } | |
5079 | return r; // idempotent | |
5080 | } | |
5081 | } | |
5082 | ||
5083 | { | |
5084 | string type; | |
5085 | r = read_meta("type", &type); | |
5086 | if (r == 0) { | |
5087 | if (type != "bluestore") { | |
5088 | derr << __func__ << " expected bluestore, but type is " << type << dendl; | |
5089 | return -EIO; | |
5090 | } | |
5091 | } else { | |
5092 | r = write_meta("type", "bluestore"); | |
5093 | if (r < 0) | |
5094 | return r; | |
5095 | } | |
5096 | } | |
5097 | ||
5098 | freelist_type = "bitmap"; | |
5099 | ||
5100 | r = _open_path(); | |
5101 | if (r < 0) | |
5102 | return r; | |
5103 | ||
5104 | r = _open_fsid(true); | |
5105 | if (r < 0) | |
5106 | goto out_path_fd; | |
5107 | ||
5108 | r = _lock_fsid(); | |
5109 | if (r < 0) | |
5110 | goto out_close_fsid; | |
5111 | ||
5112 | r = _read_fsid(&old_fsid); | |
5113 | if (r < 0 || old_fsid.is_zero()) { | |
5114 | if (fsid.is_zero()) { | |
5115 | fsid.generate_random(); | |
5116 | dout(1) << __func__ << " generated fsid " << fsid << dendl; | |
5117 | } else { | |
5118 | dout(1) << __func__ << " using provided fsid " << fsid << dendl; | |
5119 | } | |
5120 | // we'll write it later. | |
5121 | } else { | |
5122 | if (!fsid.is_zero() && fsid != old_fsid) { | |
5123 | derr << __func__ << " on-disk fsid " << old_fsid | |
5124 | << " != provided " << fsid << dendl; | |
5125 | r = -EINVAL; | |
5126 | goto out_close_fsid; | |
5127 | } | |
5128 | fsid = old_fsid; | |
5129 | } | |
5130 | ||
5131 | r = _setup_block_symlink_or_file("block", cct->_conf->bluestore_block_path, | |
5132 | cct->_conf->bluestore_block_size, | |
5133 | cct->_conf->bluestore_block_create); | |
5134 | if (r < 0) | |
5135 | goto out_close_fsid; | |
5136 | if (cct->_conf->bluestore_bluefs) { | |
5137 | r = _setup_block_symlink_or_file("block.wal", cct->_conf->bluestore_block_wal_path, | |
5138 | cct->_conf->bluestore_block_wal_size, | |
5139 | cct->_conf->bluestore_block_wal_create); | |
5140 | if (r < 0) | |
5141 | goto out_close_fsid; | |
5142 | r = _setup_block_symlink_or_file("block.db", cct->_conf->bluestore_block_db_path, | |
5143 | cct->_conf->bluestore_block_db_size, | |
5144 | cct->_conf->bluestore_block_db_create); | |
5145 | if (r < 0) | |
5146 | goto out_close_fsid; | |
5147 | } | |
5148 | ||
5149 | r = _open_bdev(true); | |
5150 | if (r < 0) | |
5151 | goto out_close_fsid; | |
5152 | ||
5153 | r = _open_db(true); | |
5154 | if (r < 0) | |
5155 | goto out_close_bdev; | |
5156 | ||
5157 | r = _open_fm(true); | |
5158 | if (r < 0) | |
5159 | goto out_close_db; | |
5160 | ||
5161 | { | |
5162 | KeyValueDB::Transaction t = db->get_transaction(); | |
5163 | { | |
5164 | bufferlist bl; | |
5165 | ::encode((uint64_t)0, bl); | |
5166 | t->set(PREFIX_SUPER, "nid_max", bl); | |
5167 | t->set(PREFIX_SUPER, "blobid_max", bl); | |
5168 | } | |
5169 | ||
5170 | // choose min_alloc_size | |
5171 | if (cct->_conf->bluestore_min_alloc_size) { | |
5172 | min_alloc_size = cct->_conf->bluestore_min_alloc_size; | |
5173 | } else { | |
5174 | assert(bdev); | |
5175 | if (bdev->is_rotational()) { | |
5176 | min_alloc_size = cct->_conf->bluestore_min_alloc_size_hdd; | |
5177 | } else { | |
5178 | min_alloc_size = cct->_conf->bluestore_min_alloc_size_ssd; | |
5179 | } | |
5180 | } | |
224ce89b WB |
5181 | |
5182 | // make sure min_alloc_size is power of 2 aligned. | |
5183 | if (!ISP2(min_alloc_size)) { | |
5184 | derr << __func__ << " min_alloc_size 0x" | |
5185 | << std::hex << min_alloc_size << std::dec | |
5186 | << " is not power of 2 aligned!" | |
5187 | << dendl; | |
5188 | r = -EINVAL; | |
5189 | goto out_close_fm; | |
5190 | } | |
5191 | ||
7c673cae FG |
5192 | { |
5193 | bufferlist bl; | |
5194 | ::encode((uint64_t)min_alloc_size, bl); | |
5195 | t->set(PREFIX_SUPER, "min_alloc_size", bl); | |
5196 | } | |
5197 | ||
5198 | ondisk_format = latest_ondisk_format; | |
5199 | _prepare_ondisk_format_super(t); | |
5200 | db->submit_transaction_sync(t); | |
5201 | } | |
5202 | ||
7c673cae FG |
5203 | |
5204 | r = write_meta("kv_backend", cct->_conf->bluestore_kvbackend); | |
5205 | if (r < 0) | |
224ce89b WB |
5206 | goto out_close_fm; |
5207 | ||
7c673cae FG |
5208 | r = write_meta("bluefs", stringify((int)cct->_conf->bluestore_bluefs)); |
5209 | if (r < 0) | |
224ce89b | 5210 | goto out_close_fm; |
7c673cae FG |
5211 | |
5212 | if (fsid != old_fsid) { | |
5213 | r = _write_fsid(); | |
5214 | if (r < 0) { | |
5215 | derr << __func__ << " error writing fsid: " << cpp_strerror(r) << dendl; | |
224ce89b | 5216 | goto out_close_fm; |
7c673cae FG |
5217 | } |
5218 | } | |
5219 | ||
7c673cae FG |
5220 | out_close_fm: |
5221 | _close_fm(); | |
5222 | out_close_db: | |
5223 | _close_db(); | |
5224 | out_close_bdev: | |
5225 | _close_bdev(); | |
5226 | out_close_fsid: | |
5227 | _close_fsid(); | |
5228 | out_path_fd: | |
5229 | _close_path(); | |
5230 | ||
5231 | if (r == 0 && | |
5232 | cct->_conf->bluestore_fsck_on_mkfs) { | |
5233 | int rc = fsck(cct->_conf->bluestore_fsck_on_mkfs_deep); | |
5234 | if (rc < 0) | |
5235 | return rc; | |
5236 | if (rc > 0) { | |
5237 | derr << __func__ << " fsck found " << rc << " errors" << dendl; | |
5238 | r = -EIO; | |
5239 | } | |
5240 | } | |
31f18b77 FG |
5241 | |
5242 | if (r == 0) { | |
5243 | // indicate success by writing the 'mkfs_done' file | |
5244 | r = write_meta("mkfs_done", "yes"); | |
5245 | } | |
5246 | ||
7c673cae FG |
5247 | if (r < 0) { |
5248 | derr << __func__ << " failed, " << cpp_strerror(r) << dendl; | |
31f18b77 FG |
5249 | } else { |
5250 | dout(0) << __func__ << " success" << dendl; | |
7c673cae FG |
5251 | } |
5252 | return r; | |
5253 | } | |
5254 | ||
5255 | void BlueStore::set_cache_shards(unsigned num) | |
5256 | { | |
5257 | dout(10) << __func__ << " " << num << dendl; | |
5258 | size_t old = cache_shards.size(); | |
5259 | assert(num >= old); | |
5260 | cache_shards.resize(num); | |
5261 | for (unsigned i = old; i < num; ++i) { | |
5262 | cache_shards[i] = Cache::create(cct, cct->_conf->bluestore_cache_type, | |
5263 | logger); | |
5264 | } | |
5265 | } | |
5266 | ||
5267 | int BlueStore::_mount(bool kv_only) | |
5268 | { | |
5269 | dout(1) << __func__ << " path " << path << dendl; | |
5270 | ||
5271 | { | |
5272 | string type; | |
5273 | int r = read_meta("type", &type); | |
5274 | if (r < 0) { | |
5275 | derr << __func__ << " failed to load os-type: " << cpp_strerror(r) | |
5276 | << dendl; | |
5277 | return r; | |
5278 | } | |
5279 | ||
5280 | if (type != "bluestore") { | |
5281 | derr << __func__ << " expected bluestore, but type is " << type << dendl; | |
5282 | return -EIO; | |
5283 | } | |
5284 | } | |
5285 | ||
5286 | if (cct->_conf->bluestore_fsck_on_mount) { | |
5287 | int rc = fsck(cct->_conf->bluestore_fsck_on_mount_deep); | |
5288 | if (rc < 0) | |
5289 | return rc; | |
5290 | if (rc > 0) { | |
5291 | derr << __func__ << " fsck found " << rc << " errors" << dendl; | |
5292 | return -EIO; | |
5293 | } | |
5294 | } | |
5295 | ||
5296 | int r = _open_path(); | |
5297 | if (r < 0) | |
5298 | return r; | |
5299 | r = _open_fsid(false); | |
5300 | if (r < 0) | |
5301 | goto out_path; | |
5302 | ||
5303 | r = _read_fsid(&fsid); | |
5304 | if (r < 0) | |
5305 | goto out_fsid; | |
5306 | ||
5307 | r = _lock_fsid(); | |
5308 | if (r < 0) | |
5309 | goto out_fsid; | |
5310 | ||
5311 | r = _open_bdev(false); | |
5312 | if (r < 0) | |
5313 | goto out_fsid; | |
5314 | ||
5315 | r = _open_db(false); | |
5316 | if (r < 0) | |
5317 | goto out_bdev; | |
5318 | ||
5319 | if (kv_only) | |
5320 | return 0; | |
5321 | ||
5322 | r = _open_super_meta(); | |
5323 | if (r < 0) | |
5324 | goto out_db; | |
5325 | ||
5326 | r = _open_fm(false); | |
5327 | if (r < 0) | |
5328 | goto out_db; | |
5329 | ||
5330 | r = _open_alloc(); | |
5331 | if (r < 0) | |
5332 | goto out_fm; | |
5333 | ||
5334 | r = _open_collections(); | |
5335 | if (r < 0) | |
5336 | goto out_alloc; | |
5337 | ||
5338 | r = _reload_logger(); | |
5339 | if (r < 0) | |
5340 | goto out_coll; | |
5341 | ||
5342 | if (bluefs) { | |
5343 | r = _reconcile_bluefs_freespace(); | |
5344 | if (r < 0) | |
5345 | goto out_coll; | |
5346 | } | |
5347 | ||
31f18b77 | 5348 | _kv_start(); |
7c673cae FG |
5349 | |
5350 | r = _deferred_replay(); | |
5351 | if (r < 0) | |
5352 | goto out_stop; | |
5353 | ||
5354 | mempool_thread.init(); | |
5355 | ||
5356 | ||
5357 | mounted = true; | |
5358 | return 0; | |
5359 | ||
5360 | out_stop: | |
5361 | _kv_stop(); | |
7c673cae | 5362 | out_coll: |
31f18b77 | 5363 | _flush_cache(); |
7c673cae FG |
5364 | out_alloc: |
5365 | _close_alloc(); | |
5366 | out_fm: | |
5367 | _close_fm(); | |
5368 | out_db: | |
5369 | _close_db(); | |
5370 | out_bdev: | |
5371 | _close_bdev(); | |
5372 | out_fsid: | |
5373 | _close_fsid(); | |
5374 | out_path: | |
5375 | _close_path(); | |
5376 | return r; | |
5377 | } | |
5378 | ||
5379 | int BlueStore::umount() | |
5380 | { | |
5381 | assert(mounted); | |
5382 | dout(1) << __func__ << dendl; | |
5383 | ||
5384 | _osr_drain_all(); | |
5385 | _osr_unregister_all(); | |
5386 | ||
5387 | mempool_thread.shutdown(); | |
5388 | ||
5389 | dout(20) << __func__ << " stopping kv thread" << dendl; | |
5390 | _kv_stop(); | |
7c673cae | 5391 | _reap_collections(); |
31f18b77 | 5392 | _flush_cache(); |
7c673cae FG |
5393 | dout(20) << __func__ << " closing" << dendl; |
5394 | ||
5395 | mounted = false; | |
5396 | _close_alloc(); | |
5397 | _close_fm(); | |
5398 | _close_db(); | |
5399 | _close_bdev(); | |
5400 | _close_fsid(); | |
5401 | _close_path(); | |
5402 | ||
5403 | if (cct->_conf->bluestore_fsck_on_umount) { | |
5404 | int rc = fsck(cct->_conf->bluestore_fsck_on_umount_deep); | |
5405 | if (rc < 0) | |
5406 | return rc; | |
5407 | if (rc > 0) { | |
5408 | derr << __func__ << " fsck found " << rc << " errors" << dendl; | |
5409 | return -EIO; | |
5410 | } | |
5411 | } | |
5412 | return 0; | |
5413 | } | |
5414 | ||
5415 | static void apply(uint64_t off, | |
5416 | uint64_t len, | |
5417 | uint64_t granularity, | |
5418 | BlueStore::mempool_dynamic_bitset &bitset, | |
5419 | const char *what, | |
5420 | std::function<void(uint64_t, | |
5421 | BlueStore::mempool_dynamic_bitset &)> f) { | |
5422 | auto end = ROUND_UP_TO(off + len, granularity); | |
5423 | while (off < end) { | |
5424 | uint64_t pos = off / granularity; | |
5425 | f(pos, bitset); | |
5426 | off += granularity; | |
5427 | } | |
5428 | } | |
5429 | ||
5430 | int BlueStore::_fsck_check_extents( | |
5431 | const ghobject_t& oid, | |
5432 | const PExtentVector& extents, | |
5433 | bool compressed, | |
5434 | mempool_dynamic_bitset &used_blocks, | |
5435 | store_statfs_t& expected_statfs) | |
5436 | { | |
5437 | dout(30) << __func__ << " oid " << oid << " extents " << extents << dendl; | |
5438 | int errors = 0; | |
5439 | for (auto e : extents) { | |
5440 | if (!e.is_valid()) | |
5441 | continue; | |
5442 | expected_statfs.allocated += e.length; | |
5443 | if (compressed) { | |
5444 | expected_statfs.compressed_allocated += e.length; | |
5445 | } | |
5446 | bool already = false; | |
5447 | apply( | |
5448 | e.offset, e.length, block_size, used_blocks, __func__, | |
5449 | [&](uint64_t pos, mempool_dynamic_bitset &bs) { | |
5450 | if (bs.test(pos)) | |
5451 | already = true; | |
5452 | else | |
5453 | bs.set(pos); | |
5454 | }); | |
5455 | if (already) { | |
5456 | derr << " " << oid << " extent " << e | |
5457 | << " or a subset is already allocated" << dendl; | |
5458 | ++errors; | |
5459 | } | |
5460 | if (e.end() > bdev->get_size()) { | |
5461 | derr << " " << oid << " extent " << e | |
5462 | << " past end of block device" << dendl; | |
5463 | ++errors; | |
5464 | } | |
5465 | } | |
5466 | return errors; | |
5467 | } | |
5468 | ||
5469 | int BlueStore::fsck(bool deep) | |
5470 | { | |
5471 | dout(1) << __func__ << (deep ? " (deep)" : " (shallow)") << " start" << dendl; | |
5472 | int errors = 0; | |
31f18b77 FG |
5473 | |
5474 | typedef btree::btree_set< | |
5475 | uint64_t,std::less<uint64_t>, | |
5476 | mempool::bluestore_fsck::pool_allocator<uint64_t>> uint64_t_btree_t; | |
5477 | uint64_t_btree_t used_nids; | |
5478 | uint64_t_btree_t used_omap_head; | |
5479 | uint64_t_btree_t used_sbids; | |
5480 | ||
7c673cae | 5481 | mempool_dynamic_bitset used_blocks; |
7c673cae FG |
5482 | KeyValueDB::Iterator it; |
5483 | store_statfs_t expected_statfs, actual_statfs; | |
5484 | struct sb_info_t { | |
5485 | list<ghobject_t> oids; | |
5486 | SharedBlobRef sb; | |
5487 | bluestore_extent_ref_map_t ref_map; | |
5488 | bool compressed; | |
5489 | }; | |
5490 | mempool::bluestore_fsck::map<uint64_t,sb_info_t> sb_info; | |
5491 | ||
5492 | uint64_t num_objects = 0; | |
5493 | uint64_t num_extents = 0; | |
5494 | uint64_t num_blobs = 0; | |
5495 | uint64_t num_spanning_blobs = 0; | |
5496 | uint64_t num_shared_blobs = 0; | |
5497 | uint64_t num_sharded_objects = 0; | |
5498 | uint64_t num_object_shards = 0; | |
5499 | ||
5500 | utime_t start = ceph_clock_now(); | |
5501 | ||
5502 | int r = _open_path(); | |
5503 | if (r < 0) | |
5504 | return r; | |
5505 | r = _open_fsid(false); | |
5506 | if (r < 0) | |
5507 | goto out_path; | |
5508 | ||
5509 | r = _read_fsid(&fsid); | |
5510 | if (r < 0) | |
5511 | goto out_fsid; | |
5512 | ||
5513 | r = _lock_fsid(); | |
5514 | if (r < 0) | |
5515 | goto out_fsid; | |
5516 | ||
5517 | r = _open_bdev(false); | |
5518 | if (r < 0) | |
5519 | goto out_fsid; | |
5520 | ||
5521 | r = _open_db(false); | |
5522 | if (r < 0) | |
5523 | goto out_bdev; | |
5524 | ||
5525 | r = _open_super_meta(); | |
5526 | if (r < 0) | |
5527 | goto out_db; | |
5528 | ||
5529 | r = _open_fm(false); | |
5530 | if (r < 0) | |
5531 | goto out_db; | |
5532 | ||
5533 | r = _open_alloc(); | |
5534 | if (r < 0) | |
5535 | goto out_fm; | |
5536 | ||
5537 | r = _open_collections(&errors); | |
5538 | if (r < 0) | |
5539 | goto out_alloc; | |
5540 | ||
5541 | mempool_thread.init(); | |
5542 | ||
31f18b77 FG |
5543 | // we need finishers and kv_{sync,finalize}_thread *just* for replay |
5544 | _kv_start(); | |
7c673cae | 5545 | r = _deferred_replay(); |
31f18b77 | 5546 | _kv_stop(); |
7c673cae FG |
5547 | if (r < 0) |
5548 | goto out_scan; | |
5549 | ||
5550 | used_blocks.resize(bdev->get_size() / block_size); | |
5551 | apply( | |
5552 | 0, SUPER_RESERVED, block_size, used_blocks, "0~SUPER_RESERVED", | |
5553 | [&](uint64_t pos, mempool_dynamic_bitset &bs) { | |
5554 | bs.set(pos); | |
5555 | } | |
5556 | ); | |
5557 | ||
5558 | if (bluefs) { | |
5559 | for (auto e = bluefs_extents.begin(); e != bluefs_extents.end(); ++e) { | |
5560 | apply( | |
5561 | e.get_start(), e.get_len(), block_size, used_blocks, "bluefs", | |
5562 | [&](uint64_t pos, mempool_dynamic_bitset &bs) { | |
5563 | bs.set(pos); | |
5564 | } | |
5565 | ); | |
5566 | } | |
5567 | r = bluefs->fsck(); | |
5568 | if (r < 0) { | |
5569 | goto out_scan; | |
5570 | } | |
5571 | if (r > 0) | |
5572 | errors += r; | |
5573 | } | |
5574 | ||
5575 | // get expected statfs; fill unaffected fields to be able to compare | |
5576 | // structs | |
5577 | statfs(&actual_statfs); | |
5578 | expected_statfs.total = actual_statfs.total; | |
5579 | expected_statfs.available = actual_statfs.available; | |
5580 | ||
5581 | // walk PREFIX_OBJ | |
5582 | dout(1) << __func__ << " walking object keyspace" << dendl; | |
5583 | it = db->get_iterator(PREFIX_OBJ); | |
5584 | if (it) { | |
5585 | CollectionRef c; | |
5586 | spg_t pgid; | |
5587 | mempool::bluestore_fsck::list<string> expecting_shards; | |
5588 | for (it->lower_bound(string()); it->valid(); it->next()) { | |
31f18b77 FG |
5589 | if (g_conf->bluestore_debug_fsck_abort) { |
5590 | goto out_scan; | |
5591 | } | |
7c673cae FG |
5592 | dout(30) << " key " << pretty_binary_string(it->key()) << dendl; |
5593 | if (is_extent_shard_key(it->key())) { | |
5594 | while (!expecting_shards.empty() && | |
5595 | expecting_shards.front() < it->key()) { | |
5596 | derr << __func__ << " error: missing shard key " | |
5597 | << pretty_binary_string(expecting_shards.front()) | |
5598 | << dendl; | |
5599 | ++errors; | |
5600 | expecting_shards.pop_front(); | |
5601 | } | |
5602 | if (!expecting_shards.empty() && | |
5603 | expecting_shards.front() == it->key()) { | |
5604 | // all good | |
5605 | expecting_shards.pop_front(); | |
5606 | continue; | |
5607 | } | |
5608 | ||
5609 | uint32_t offset; | |
5610 | string okey; | |
5611 | get_key_extent_shard(it->key(), &okey, &offset); | |
5612 | derr << __func__ << " error: stray shard 0x" << std::hex << offset | |
5613 | << std::dec << dendl; | |
5614 | if (expecting_shards.empty()) { | |
5615 | derr << __func__ << " error: " << pretty_binary_string(it->key()) | |
5616 | << " is unexpected" << dendl; | |
5617 | ++errors; | |
5618 | continue; | |
5619 | } | |
5620 | while (expecting_shards.front() > it->key()) { | |
5621 | derr << __func__ << " error: saw " << pretty_binary_string(it->key()) | |
5622 | << dendl; | |
5623 | derr << __func__ << " error: exp " | |
5624 | << pretty_binary_string(expecting_shards.front()) << dendl; | |
5625 | ++errors; | |
5626 | expecting_shards.pop_front(); | |
5627 | if (expecting_shards.empty()) { | |
5628 | break; | |
5629 | } | |
5630 | } | |
5631 | continue; | |
5632 | } | |
5633 | ||
5634 | ghobject_t oid; | |
5635 | int r = get_key_object(it->key(), &oid); | |
5636 | if (r < 0) { | |
5637 | derr << __func__ << " error: bad object key " | |
5638 | << pretty_binary_string(it->key()) << dendl; | |
5639 | ++errors; | |
5640 | continue; | |
5641 | } | |
5642 | if (!c || | |
5643 | oid.shard_id != pgid.shard || | |
5644 | oid.hobj.pool != (int64_t)pgid.pool() || | |
5645 | !c->contains(oid)) { | |
5646 | c = nullptr; | |
5647 | for (ceph::unordered_map<coll_t, CollectionRef>::iterator p = | |
5648 | coll_map.begin(); | |
5649 | p != coll_map.end(); | |
5650 | ++p) { | |
5651 | if (p->second->contains(oid)) { | |
5652 | c = p->second; | |
5653 | break; | |
5654 | } | |
5655 | } | |
5656 | if (!c) { | |
5657 | derr << __func__ << " error: stray object " << oid | |
5658 | << " not owned by any collection" << dendl; | |
5659 | ++errors; | |
5660 | continue; | |
5661 | } | |
5662 | c->cid.is_pg(&pgid); | |
5663 | dout(20) << __func__ << " collection " << c->cid << dendl; | |
5664 | } | |
5665 | ||
5666 | if (!expecting_shards.empty()) { | |
5667 | for (auto &k : expecting_shards) { | |
5668 | derr << __func__ << " error: missing shard key " | |
5669 | << pretty_binary_string(k) << dendl; | |
5670 | } | |
5671 | ++errors; | |
5672 | expecting_shards.clear(); | |
5673 | } | |
5674 | ||
5675 | dout(10) << __func__ << " " << oid << dendl; | |
5676 | RWLock::RLocker l(c->lock); | |
5677 | OnodeRef o = c->get_onode(oid, false); | |
5678 | if (o->onode.nid) { | |
5679 | if (o->onode.nid > nid_max) { | |
5680 | derr << __func__ << " error: " << oid << " nid " << o->onode.nid | |
5681 | << " > nid_max " << nid_max << dendl; | |
5682 | ++errors; | |
5683 | } | |
5684 | if (used_nids.count(o->onode.nid)) { | |
5685 | derr << __func__ << " error: " << oid << " nid " << o->onode.nid | |
5686 | << " already in use" << dendl; | |
5687 | ++errors; | |
5688 | continue; // go for next object | |
5689 | } | |
5690 | used_nids.insert(o->onode.nid); | |
5691 | } | |
5692 | ++num_objects; | |
5693 | num_spanning_blobs += o->extent_map.spanning_blob_map.size(); | |
5694 | o->extent_map.fault_range(db, 0, OBJECT_MAX_SIZE); | |
5695 | _dump_onode(o, 30); | |
5696 | // shards | |
5697 | if (!o->extent_map.shards.empty()) { | |
5698 | ++num_sharded_objects; | |
5699 | num_object_shards += o->extent_map.shards.size(); | |
5700 | } | |
5701 | for (auto& s : o->extent_map.shards) { | |
5702 | dout(20) << __func__ << " shard " << *s.shard_info << dendl; | |
5703 | expecting_shards.push_back(string()); | |
5704 | get_extent_shard_key(o->key, s.shard_info->offset, | |
5705 | &expecting_shards.back()); | |
5706 | if (s.shard_info->offset >= o->onode.size) { | |
5707 | derr << __func__ << " error: " << oid << " shard 0x" << std::hex | |
5708 | << s.shard_info->offset << " past EOF at 0x" << o->onode.size | |
5709 | << std::dec << dendl; | |
5710 | ++errors; | |
5711 | } | |
5712 | } | |
5713 | // lextents | |
5714 | map<BlobRef,bluestore_blob_t::unused_t> referenced; | |
5715 | uint64_t pos = 0; | |
5716 | mempool::bluestore_fsck::map<BlobRef, | |
5717 | bluestore_blob_use_tracker_t> ref_map; | |
5718 | for (auto& l : o->extent_map.extent_map) { | |
5719 | dout(20) << __func__ << " " << l << dendl; | |
5720 | if (l.logical_offset < pos) { | |
5721 | derr << __func__ << " error: " << oid << " lextent at 0x" | |
5722 | << std::hex << l.logical_offset | |
5723 | << " overlaps with the previous, which ends at 0x" << pos | |
5724 | << std::dec << dendl; | |
5725 | ++errors; | |
5726 | } | |
5727 | if (o->extent_map.spans_shard(l.logical_offset, l.length)) { | |
5728 | derr << __func__ << " error: " << oid << " lextent at 0x" | |
5729 | << std::hex << l.logical_offset << "~" << l.length | |
5730 | << " spans a shard boundary" | |
5731 | << std::dec << dendl; | |
5732 | ++errors; | |
5733 | } | |
5734 | pos = l.logical_offset + l.length; | |
5735 | expected_statfs.stored += l.length; | |
5736 | assert(l.blob); | |
5737 | const bluestore_blob_t& blob = l.blob->get_blob(); | |
5738 | ||
5739 | auto& ref = ref_map[l.blob]; | |
5740 | if (ref.is_empty()) { | |
5741 | uint32_t min_release_size = blob.get_release_size(min_alloc_size); | |
5742 | uint32_t l = blob.get_logical_length(); | |
5743 | ref.init(l, min_release_size); | |
5744 | } | |
5745 | ref.get( | |
5746 | l.blob_offset, | |
5747 | l.length); | |
5748 | ++num_extents; | |
5749 | if (blob.has_unused()) { | |
5750 | auto p = referenced.find(l.blob); | |
5751 | bluestore_blob_t::unused_t *pu; | |
5752 | if (p == referenced.end()) { | |
5753 | pu = &referenced[l.blob]; | |
5754 | } else { | |
5755 | pu = &p->second; | |
5756 | } | |
5757 | uint64_t blob_len = blob.get_logical_length(); | |
5758 | assert((blob_len % (sizeof(*pu)*8)) == 0); | |
5759 | assert(l.blob_offset + l.length <= blob_len); | |
5760 | uint64_t chunk_size = blob_len / (sizeof(*pu)*8); | |
5761 | uint64_t start = l.blob_offset / chunk_size; | |
5762 | uint64_t end = | |
5763 | ROUND_UP_TO(l.blob_offset + l.length, chunk_size) / chunk_size; | |
5764 | for (auto i = start; i < end; ++i) { | |
5765 | (*pu) |= (1u << i); | |
5766 | } | |
5767 | } | |
5768 | } | |
5769 | for (auto &i : referenced) { | |
5770 | dout(20) << __func__ << " referenced 0x" << std::hex << i.second | |
5771 | << std::dec << " for " << *i.first << dendl; | |
5772 | const bluestore_blob_t& blob = i.first->get_blob(); | |
5773 | if (i.second & blob.unused) { | |
5774 | derr << __func__ << " error: " << oid << " blob claims unused 0x" | |
5775 | << std::hex << blob.unused | |
5776 | << " but extents reference 0x" << i.second | |
5777 | << " on blob " << *i.first << dendl; | |
5778 | ++errors; | |
5779 | } | |
5780 | if (blob.has_csum()) { | |
5781 | uint64_t blob_len = blob.get_logical_length(); | |
5782 | uint64_t unused_chunk_size = blob_len / (sizeof(blob.unused)*8); | |
5783 | unsigned csum_count = blob.get_csum_count(); | |
5784 | unsigned csum_chunk_size = blob.get_csum_chunk_size(); | |
5785 | for (unsigned p = 0; p < csum_count; ++p) { | |
5786 | unsigned pos = p * csum_chunk_size; | |
5787 | unsigned firstbit = pos / unused_chunk_size; // [firstbit,lastbit] | |
5788 | unsigned lastbit = (pos + csum_chunk_size - 1) / unused_chunk_size; | |
5789 | unsigned mask = 1u << firstbit; | |
5790 | for (unsigned b = firstbit + 1; b <= lastbit; ++b) { | |
5791 | mask |= 1u << b; | |
5792 | } | |
5793 | if ((blob.unused & mask) == mask) { | |
5794 | // this csum chunk region is marked unused | |
5795 | if (blob.get_csum_item(p) != 0) { | |
5796 | derr << __func__ << " error: " << oid | |
5797 | << " blob claims csum chunk 0x" << std::hex << pos | |
5798 | << "~" << csum_chunk_size | |
5799 | << " is unused (mask 0x" << mask << " of unused 0x" | |
5800 | << blob.unused << ") but csum is non-zero 0x" | |
5801 | << blob.get_csum_item(p) << std::dec << " on blob " | |
5802 | << *i.first << dendl; | |
5803 | ++errors; | |
5804 | } | |
5805 | } | |
5806 | } | |
5807 | } | |
5808 | } | |
5809 | for (auto &i : ref_map) { | |
5810 | ++num_blobs; | |
5811 | const bluestore_blob_t& blob = i.first->get_blob(); | |
5812 | bool equal = i.first->get_blob_use_tracker().equal(i.second); | |
5813 | if (!equal) { | |
5814 | derr << __func__ << " error: " << oid << " blob " << *i.first | |
5815 | << " doesn't match expected ref_map " << i.second << dendl; | |
5816 | ++errors; | |
5817 | } | |
5818 | if (blob.is_compressed()) { | |
5819 | expected_statfs.compressed += blob.get_compressed_payload_length(); | |
5820 | expected_statfs.compressed_original += | |
5821 | i.first->get_referenced_bytes(); | |
5822 | } | |
5823 | if (blob.is_shared()) { | |
5824 | if (i.first->shared_blob->get_sbid() > blobid_max) { | |
5825 | derr << __func__ << " error: " << oid << " blob " << blob | |
5826 | << " sbid " << i.first->shared_blob->get_sbid() << " > blobid_max " | |
5827 | << blobid_max << dendl; | |
5828 | ++errors; | |
5829 | } else if (i.first->shared_blob->get_sbid() == 0) { | |
5830 | derr << __func__ << " error: " << oid << " blob " << blob | |
5831 | << " marked as shared but has uninitialized sbid" | |
5832 | << dendl; | |
5833 | ++errors; | |
5834 | } | |
5835 | sb_info_t& sbi = sb_info[i.first->shared_blob->get_sbid()]; | |
5836 | sbi.sb = i.first->shared_blob; | |
5837 | sbi.oids.push_back(oid); | |
5838 | sbi.compressed = blob.is_compressed(); | |
5839 | for (auto e : blob.get_extents()) { | |
5840 | if (e.is_valid()) { | |
5841 | sbi.ref_map.get(e.offset, e.length); | |
5842 | } | |
5843 | } | |
5844 | } else { | |
5845 | errors += _fsck_check_extents(oid, blob.get_extents(), | |
5846 | blob.is_compressed(), | |
5847 | used_blocks, | |
5848 | expected_statfs); | |
5849 | } | |
5850 | } | |
5851 | if (deep) { | |
5852 | bufferlist bl; | |
5853 | int r = _do_read(c.get(), o, 0, o->onode.size, bl, 0); | |
5854 | if (r < 0) { | |
5855 | ++errors; | |
5856 | derr << __func__ << " error: " << oid << " error during read: " | |
5857 | << cpp_strerror(r) << dendl; | |
5858 | } | |
5859 | } | |
5860 | // omap | |
5861 | if (o->onode.has_omap()) { | |
5862 | if (used_omap_head.count(o->onode.nid)) { | |
5863 | derr << __func__ << " error: " << oid << " omap_head " << o->onode.nid | |
5864 | << " already in use" << dendl; | |
5865 | ++errors; | |
5866 | } else { | |
5867 | used_omap_head.insert(o->onode.nid); | |
5868 | } | |
5869 | } | |
7c673cae FG |
5870 | } |
5871 | } | |
5872 | dout(1) << __func__ << " checking shared_blobs" << dendl; | |
5873 | it = db->get_iterator(PREFIX_SHARED_BLOB); | |
5874 | if (it) { | |
5875 | for (it->lower_bound(string()); it->valid(); it->next()) { | |
5876 | string key = it->key(); | |
5877 | uint64_t sbid; | |
5878 | if (get_key_shared_blob(key, &sbid)) { | |
5879 | derr << __func__ << " error: bad key '" << key | |
5880 | << "' in shared blob namespace" << dendl; | |
5881 | ++errors; | |
5882 | continue; | |
5883 | } | |
5884 | auto p = sb_info.find(sbid); | |
5885 | if (p == sb_info.end()) { | |
5886 | derr << __func__ << " error: found stray shared blob data for sbid 0x" | |
5887 | << std::hex << sbid << std::dec << dendl; | |
5888 | ++errors; | |
5889 | } else { | |
5890 | ++num_shared_blobs; | |
5891 | sb_info_t& sbi = p->second; | |
5892 | bluestore_shared_blob_t shared_blob(sbid); | |
5893 | bufferlist bl = it->value(); | |
5894 | bufferlist::iterator blp = bl.begin(); | |
5895 | ::decode(shared_blob, blp); | |
5896 | dout(20) << __func__ << " " << *sbi.sb << " " << shared_blob << dendl; | |
5897 | if (shared_blob.ref_map != sbi.ref_map) { | |
5898 | derr << __func__ << " error: shared blob 0x" << std::hex << sbid | |
5899 | << std::dec << " ref_map " << shared_blob.ref_map | |
5900 | << " != expected " << sbi.ref_map << dendl; | |
5901 | ++errors; | |
5902 | } | |
5903 | PExtentVector extents; | |
5904 | for (auto &r : shared_blob.ref_map.ref_map) { | |
5905 | extents.emplace_back(bluestore_pextent_t(r.first, r.second.length)); | |
5906 | } | |
5907 | errors += _fsck_check_extents(p->second.oids.front(), | |
5908 | extents, | |
5909 | p->second.compressed, | |
5910 | used_blocks, expected_statfs); | |
5911 | sb_info.erase(p); | |
5912 | } | |
5913 | } | |
5914 | } | |
5915 | for (auto &p : sb_info) { | |
5916 | derr << __func__ << " error: shared_blob 0x" << p.first | |
5917 | << " key is missing (" << *p.second.sb << ")" << dendl; | |
5918 | ++errors; | |
5919 | } | |
5920 | if (!(actual_statfs == expected_statfs)) { | |
5921 | derr << __func__ << " error: actual " << actual_statfs | |
5922 | << " != expected " << expected_statfs << dendl; | |
5923 | ++errors; | |
5924 | } | |
5925 | ||
5926 | dout(1) << __func__ << " checking for stray omap data" << dendl; | |
5927 | it = db->get_iterator(PREFIX_OMAP); | |
5928 | if (it) { | |
5929 | for (it->lower_bound(string()); it->valid(); it->next()) { | |
5930 | uint64_t omap_head; | |
5931 | _key_decode_u64(it->key().c_str(), &omap_head); | |
5932 | if (used_omap_head.count(omap_head) == 0) { | |
5933 | derr << __func__ << " error: found stray omap data on omap_head " | |
5934 | << omap_head << dendl; | |
5935 | ++errors; | |
5936 | } | |
5937 | } | |
5938 | } | |
5939 | ||
5940 | dout(1) << __func__ << " checking deferred events" << dendl; | |
5941 | it = db->get_iterator(PREFIX_DEFERRED); | |
5942 | if (it) { | |
5943 | for (it->lower_bound(string()); it->valid(); it->next()) { | |
5944 | bufferlist bl = it->value(); | |
5945 | bufferlist::iterator p = bl.begin(); | |
5946 | bluestore_deferred_transaction_t wt; | |
5947 | try { | |
5948 | ::decode(wt, p); | |
5949 | } catch (buffer::error& e) { | |
5950 | derr << __func__ << " error: failed to decode deferred txn " | |
5951 | << pretty_binary_string(it->key()) << dendl; | |
5952 | r = -EIO; | |
5953 | goto out_scan; | |
5954 | } | |
5955 | dout(20) << __func__ << " deferred " << wt.seq | |
5956 | << " ops " << wt.ops.size() | |
5957 | << " released 0x" << std::hex << wt.released << std::dec << dendl; | |
5958 | for (auto e = wt.released.begin(); e != wt.released.end(); ++e) { | |
5959 | apply( | |
5960 | e.get_start(), e.get_len(), block_size, used_blocks, "deferred", | |
5961 | [&](uint64_t pos, mempool_dynamic_bitset &bs) { | |
5962 | bs.set(pos); | |
5963 | } | |
5964 | ); | |
5965 | } | |
5966 | } | |
5967 | } | |
5968 | ||
5969 | dout(1) << __func__ << " checking freelist vs allocated" << dendl; | |
5970 | { | |
5971 | // remove bluefs_extents from used set since the freelist doesn't | |
5972 | // know they are allocated. | |
5973 | for (auto e = bluefs_extents.begin(); e != bluefs_extents.end(); ++e) { | |
5974 | apply( | |
5975 | e.get_start(), e.get_len(), block_size, used_blocks, "bluefs_extents", | |
5976 | [&](uint64_t pos, mempool_dynamic_bitset &bs) { | |
5977 | bs.reset(pos); | |
5978 | } | |
5979 | ); | |
5980 | } | |
5981 | fm->enumerate_reset(); | |
5982 | uint64_t offset, length; | |
5983 | while (fm->enumerate_next(&offset, &length)) { | |
5984 | bool intersects = false; | |
5985 | apply( | |
5986 | offset, length, block_size, used_blocks, "free", | |
5987 | [&](uint64_t pos, mempool_dynamic_bitset &bs) { | |
5988 | if (bs.test(pos)) { | |
5989 | intersects = true; | |
5990 | } else { | |
5991 | bs.set(pos); | |
5992 | } | |
5993 | } | |
5994 | ); | |
5995 | if (intersects) { | |
5996 | derr << __func__ << " error: free extent 0x" << std::hex << offset | |
5997 | << "~" << length << std::dec | |
5998 | << " intersects allocated blocks" << dendl; | |
5999 | ++errors; | |
6000 | } | |
6001 | } | |
224ce89b | 6002 | fm->enumerate_reset(); |
7c673cae | 6003 | size_t count = used_blocks.count(); |
b5b8bbf5 FG |
6004 | if (used_blocks.size() == count + 1) { |
6005 | // this due to http://tracker.ceph.com/issues/21089 | |
6006 | bufferlist fm_bpb_bl, fm_blocks_bl, fm_bpk_bl; | |
6007 | db->get(PREFIX_ALLOC, "bytes_per_block", &fm_bpb_bl); | |
6008 | db->get(PREFIX_ALLOC, "blocks", &fm_blocks_bl); | |
6009 | db->get(PREFIX_ALLOC, "blocks_per_key", &fm_bpk_bl); | |
6010 | uint64_t fm_blocks = 0; | |
6011 | uint64_t fm_bsize = 1; | |
6012 | uint64_t fm_blocks_per_key = 1; | |
6013 | try { | |
6014 | auto p = fm_blocks_bl.begin(); | |
6015 | ::decode(fm_blocks, p); | |
6016 | auto q = fm_bpb_bl.begin(); | |
6017 | ::decode(fm_bsize, q); | |
6018 | auto r = fm_bpk_bl.begin(); | |
6019 | ::decode(fm_blocks_per_key, r); | |
6020 | } catch (buffer::error& e) { | |
6021 | } | |
6022 | uint64_t dev_bsize = bdev->get_block_size(); | |
6023 | uint64_t bad_size = bdev->get_size() & ~fm_bsize; | |
6024 | if (used_blocks.test(bad_size / dev_bsize) == 0) { | |
6025 | // this is the last block of the device that we previously | |
6026 | // (incorrectly) truncated off of the effective device size. this | |
6027 | // prevented BitmapFreelistManager from marking it as used along with | |
6028 | // the other "past-eof" blocks in the last key slot. mark it used | |
6029 | // now. | |
6030 | derr << __func__ << " warning: fixing leaked block 0x" << std::hex | |
6031 | << bad_size << "~" << fm_bsize << std::dec << " due to old bug" | |
6032 | << dendl; | |
6033 | KeyValueDB::Transaction t = db->get_transaction(); | |
6034 | // fix freelistmanager metadata (the internal 'blocks' count is | |
6035 | // rounded up to include the trailing key, past eof) | |
6036 | uint64_t new_blocks = bdev->get_size() / fm_bsize; | |
6037 | if (new_blocks / fm_blocks_per_key * fm_blocks_per_key != new_blocks) { | |
6038 | new_blocks = (new_blocks / fm_blocks_per_key + 1) * | |
6039 | fm_blocks_per_key; | |
6040 | } | |
6041 | if (new_blocks != fm_blocks) { | |
6042 | // the fm block count increased | |
6043 | derr << __func__ << " freelist block and key count changed, fixing 0x" | |
6044 | << std::hex << bdev->get_size() << "~" | |
6045 | << ((new_blocks * fm_bsize) - bdev->get_size()) << std::dec | |
6046 | << dendl; | |
6047 | bufferlist bl; | |
6048 | ::encode(new_blocks, bl); | |
6049 | t->set(PREFIX_ALLOC, "blocks", bl); | |
6050 | fm->allocate(bdev->get_size(), | |
6051 | (new_blocks * fm_bsize) - bdev->get_size(), | |
6052 | t); | |
6053 | } else { | |
6054 | // block count is the same, but size changed; fix just the size | |
6055 | derr << __func__ << " fixing just the stray block at 0x" | |
6056 | << std::hex << bad_size << "~" << fm_bsize << std::dec << dendl; | |
6057 | fm->allocate(bad_size, fm_bsize, t); | |
6058 | } | |
6059 | bufferlist sizebl; | |
6060 | ::encode(bdev->get_size(), sizebl); | |
6061 | t->set(PREFIX_ALLOC, "size", sizebl); | |
6062 | int r = db->submit_transaction_sync(t); | |
6063 | assert(r == 0); | |
6064 | ||
6065 | used_blocks.set(bad_size / dev_bsize); | |
6066 | ++count; | |
6067 | } | |
6068 | } | |
7c673cae FG |
6069 | if (used_blocks.size() != count) { |
6070 | assert(used_blocks.size() > count); | |
7c673cae | 6071 | ++errors; |
b5b8bbf5 FG |
6072 | used_blocks.flip(); |
6073 | size_t start = used_blocks.find_first(); | |
6074 | while (start != decltype(used_blocks)::npos) { | |
6075 | size_t cur = start; | |
6076 | while (true) { | |
6077 | size_t next = used_blocks.find_next(cur); | |
6078 | if (next != cur + 1) { | |
6079 | derr << __func__ << " error: leaked extent 0x" << std::hex | |
6080 | << ((uint64_t)start * block_size) << "~" | |
6081 | << ((cur + 1 - start) * block_size) << std::dec | |
6082 | << dendl; | |
6083 | start = next; | |
6084 | break; | |
6085 | } | |
6086 | cur = next; | |
6087 | } | |
6088 | } | |
6089 | used_blocks.flip(); | |
7c673cae FG |
6090 | } |
6091 | } | |
6092 | ||
6093 | out_scan: | |
6094 | mempool_thread.shutdown(); | |
31f18b77 | 6095 | _flush_cache(); |
7c673cae FG |
6096 | out_alloc: |
6097 | _close_alloc(); | |
6098 | out_fm: | |
6099 | _close_fm(); | |
6100 | out_db: | |
6101 | it.reset(); // before db is closed | |
6102 | _close_db(); | |
6103 | out_bdev: | |
6104 | _close_bdev(); | |
6105 | out_fsid: | |
6106 | _close_fsid(); | |
6107 | out_path: | |
6108 | _close_path(); | |
6109 | ||
6110 | // fatal errors take precedence | |
6111 | if (r < 0) | |
6112 | return r; | |
6113 | ||
6114 | dout(2) << __func__ << " " << num_objects << " objects, " | |
6115 | << num_sharded_objects << " of them sharded. " | |
6116 | << dendl; | |
6117 | dout(2) << __func__ << " " << num_extents << " extents to " | |
6118 | << num_blobs << " blobs, " | |
6119 | << num_spanning_blobs << " spanning, " | |
6120 | << num_shared_blobs << " shared." | |
6121 | << dendl; | |
6122 | ||
6123 | utime_t duration = ceph_clock_now() - start; | |
6124 | dout(1) << __func__ << " finish with " << errors << " errors in " | |
6125 | << duration << " seconds" << dendl; | |
6126 | return errors; | |
6127 | } | |
6128 | ||
6129 | void BlueStore::collect_metadata(map<string,string> *pm) | |
6130 | { | |
6131 | dout(10) << __func__ << dendl; | |
6132 | bdev->collect_metadata("bluestore_bdev_", pm); | |
6133 | if (bluefs) { | |
6134 | (*pm)["bluefs"] = "1"; | |
6135 | (*pm)["bluefs_single_shared_device"] = stringify((int)bluefs_single_shared_device); | |
6136 | bluefs->collect_metadata(pm); | |
6137 | } else { | |
6138 | (*pm)["bluefs"] = "0"; | |
6139 | } | |
6140 | } | |
6141 | ||
6142 | int BlueStore::statfs(struct store_statfs_t *buf) | |
6143 | { | |
6144 | buf->reset(); | |
6145 | buf->total = bdev->get_size(); | |
6146 | buf->available = alloc->get_free(); | |
6147 | ||
6148 | if (bluefs) { | |
6149 | // part of our shared device is "free" according to BlueFS | |
6150 | // Don't include bluestore_bluefs_min because that space can't | |
6151 | // be used for any other purpose. | |
6152 | buf->available += bluefs->get_free(bluefs_shared_bdev) - cct->_conf->bluestore_bluefs_min; | |
6153 | ||
6154 | // include dedicated db, too, if that isn't the shared device. | |
6155 | if (bluefs_shared_bdev != BlueFS::BDEV_DB) { | |
6156 | buf->total += bluefs->get_total(BlueFS::BDEV_DB); | |
6157 | } | |
6158 | } | |
6159 | ||
31f18b77 FG |
6160 | { |
6161 | std::lock_guard<std::mutex> l(vstatfs_lock); | |
6162 | ||
6163 | buf->allocated = vstatfs.allocated(); | |
6164 | buf->stored = vstatfs.stored(); | |
6165 | buf->compressed = vstatfs.compressed(); | |
6166 | buf->compressed_original = vstatfs.compressed_original(); | |
6167 | buf->compressed_allocated = vstatfs.compressed_allocated(); | |
7c673cae FG |
6168 | } |
6169 | ||
7c673cae FG |
6170 | dout(20) << __func__ << *buf << dendl; |
6171 | return 0; | |
6172 | } | |
6173 | ||
6174 | // --------------- | |
6175 | // cache | |
6176 | ||
6177 | BlueStore::CollectionRef BlueStore::_get_collection(const coll_t& cid) | |
6178 | { | |
6179 | RWLock::RLocker l(coll_lock); | |
6180 | ceph::unordered_map<coll_t,CollectionRef>::iterator cp = coll_map.find(cid); | |
6181 | if (cp == coll_map.end()) | |
6182 | return CollectionRef(); | |
6183 | return cp->second; | |
6184 | } | |
6185 | ||
6186 | void BlueStore::_queue_reap_collection(CollectionRef& c) | |
6187 | { | |
6188 | dout(10) << __func__ << " " << c << " " << c->cid << dendl; | |
6189 | std::lock_guard<std::mutex> l(reap_lock); | |
6190 | removed_collections.push_back(c); | |
6191 | } | |
6192 | ||
6193 | void BlueStore::_reap_collections() | |
6194 | { | |
6195 | list<CollectionRef> removed_colls; | |
6196 | { | |
6197 | std::lock_guard<std::mutex> l(reap_lock); | |
6198 | removed_colls.swap(removed_collections); | |
6199 | } | |
6200 | ||
6201 | bool all_reaped = true; | |
6202 | ||
6203 | for (list<CollectionRef>::iterator p = removed_colls.begin(); | |
6204 | p != removed_colls.end(); | |
6205 | ++p) { | |
6206 | CollectionRef c = *p; | |
6207 | dout(10) << __func__ << " " << c << " " << c->cid << dendl; | |
6208 | if (c->onode_map.map_any([&](OnodeRef o) { | |
6209 | assert(!o->exists); | |
6210 | if (o->flushing_count.load()) { | |
6211 | dout(10) << __func__ << " " << c << " " << c->cid << " " << o->oid | |
6212 | << " flush_txns " << o->flushing_count << dendl; | |
6213 | return false; | |
6214 | } | |
6215 | return true; | |
6216 | })) { | |
6217 | all_reaped = false; | |
6218 | continue; | |
6219 | } | |
6220 | c->onode_map.clear(); | |
6221 | dout(10) << __func__ << " " << c << " " << c->cid << " done" << dendl; | |
6222 | } | |
6223 | ||
6224 | if (all_reaped) { | |
6225 | dout(10) << __func__ << " all reaped" << dendl; | |
6226 | } | |
6227 | } | |
6228 | ||
6229 | void BlueStore::_update_cache_logger() | |
6230 | { | |
6231 | uint64_t num_onodes = 0; | |
6232 | uint64_t num_extents = 0; | |
6233 | uint64_t num_blobs = 0; | |
6234 | uint64_t num_buffers = 0; | |
6235 | uint64_t num_buffer_bytes = 0; | |
6236 | for (auto c : cache_shards) { | |
6237 | c->add_stats(&num_onodes, &num_extents, &num_blobs, | |
6238 | &num_buffers, &num_buffer_bytes); | |
6239 | } | |
6240 | logger->set(l_bluestore_onodes, num_onodes); | |
6241 | logger->set(l_bluestore_extents, num_extents); | |
6242 | logger->set(l_bluestore_blobs, num_blobs); | |
6243 | logger->set(l_bluestore_buffers, num_buffers); | |
6244 | logger->set(l_bluestore_buffer_bytes, num_buffer_bytes); | |
6245 | } | |
6246 | ||
6247 | // --------------- | |
6248 | // read operations | |
6249 | ||
6250 | ObjectStore::CollectionHandle BlueStore::open_collection(const coll_t& cid) | |
6251 | { | |
6252 | return _get_collection(cid); | |
6253 | } | |
6254 | ||
6255 | bool BlueStore::exists(const coll_t& cid, const ghobject_t& oid) | |
6256 | { | |
6257 | CollectionHandle c = _get_collection(cid); | |
6258 | if (!c) | |
6259 | return false; | |
6260 | return exists(c, oid); | |
6261 | } | |
6262 | ||
6263 | bool BlueStore::exists(CollectionHandle &c_, const ghobject_t& oid) | |
6264 | { | |
6265 | Collection *c = static_cast<Collection *>(c_.get()); | |
6266 | dout(10) << __func__ << " " << c->cid << " " << oid << dendl; | |
6267 | if (!c->exists) | |
6268 | return false; | |
6269 | ||
6270 | bool r = true; | |
6271 | ||
6272 | { | |
6273 | RWLock::RLocker l(c->lock); | |
6274 | OnodeRef o = c->get_onode(oid, false); | |
6275 | if (!o || !o->exists) | |
6276 | r = false; | |
6277 | } | |
6278 | ||
7c673cae FG |
6279 | return r; |
6280 | } | |
6281 | ||
6282 | int BlueStore::stat( | |
6283 | const coll_t& cid, | |
6284 | const ghobject_t& oid, | |
6285 | struct stat *st, | |
6286 | bool allow_eio) | |
6287 | { | |
6288 | CollectionHandle c = _get_collection(cid); | |
6289 | if (!c) | |
6290 | return -ENOENT; | |
6291 | return stat(c, oid, st, allow_eio); | |
6292 | } | |
6293 | ||
6294 | int BlueStore::stat( | |
6295 | CollectionHandle &c_, | |
6296 | const ghobject_t& oid, | |
6297 | struct stat *st, | |
6298 | bool allow_eio) | |
6299 | { | |
6300 | Collection *c = static_cast<Collection *>(c_.get()); | |
6301 | if (!c->exists) | |
6302 | return -ENOENT; | |
6303 | dout(10) << __func__ << " " << c->get_cid() << " " << oid << dendl; | |
6304 | ||
6305 | { | |
6306 | RWLock::RLocker l(c->lock); | |
6307 | OnodeRef o = c->get_onode(oid, false); | |
6308 | if (!o || !o->exists) | |
6309 | return -ENOENT; | |
6310 | st->st_size = o->onode.size; | |
6311 | st->st_blksize = 4096; | |
6312 | st->st_blocks = (st->st_size + st->st_blksize - 1) / st->st_blksize; | |
6313 | st->st_nlink = 1; | |
6314 | } | |
6315 | ||
7c673cae FG |
6316 | int r = 0; |
6317 | if (_debug_mdata_eio(oid)) { | |
6318 | r = -EIO; | |
6319 | derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl; | |
6320 | } | |
6321 | return r; | |
6322 | } | |
6323 | int BlueStore::set_collection_opts( | |
6324 | const coll_t& cid, | |
6325 | const pool_opts_t& opts) | |
6326 | { | |
6327 | CollectionHandle ch = _get_collection(cid); | |
6328 | if (!ch) | |
6329 | return -ENOENT; | |
6330 | Collection *c = static_cast<Collection *>(ch.get()); | |
6331 | dout(15) << __func__ << " " << cid << " options " << opts << dendl; | |
6332 | if (!c->exists) | |
6333 | return -ENOENT; | |
6334 | RWLock::WLocker l(c->lock); | |
6335 | c->pool_opts = opts; | |
6336 | return 0; | |
6337 | } | |
6338 | ||
6339 | int BlueStore::read( | |
6340 | const coll_t& cid, | |
6341 | const ghobject_t& oid, | |
6342 | uint64_t offset, | |
6343 | size_t length, | |
6344 | bufferlist& bl, | |
224ce89b | 6345 | uint32_t op_flags) |
7c673cae FG |
6346 | { |
6347 | CollectionHandle c = _get_collection(cid); | |
6348 | if (!c) | |
6349 | return -ENOENT; | |
224ce89b | 6350 | return read(c, oid, offset, length, bl, op_flags); |
7c673cae FG |
6351 | } |
6352 | ||
6353 | int BlueStore::read( | |
6354 | CollectionHandle &c_, | |
6355 | const ghobject_t& oid, | |
6356 | uint64_t offset, | |
6357 | size_t length, | |
6358 | bufferlist& bl, | |
224ce89b | 6359 | uint32_t op_flags) |
7c673cae FG |
6360 | { |
6361 | utime_t start = ceph_clock_now(); | |
6362 | Collection *c = static_cast<Collection *>(c_.get()); | |
6363 | const coll_t &cid = c->get_cid(); | |
6364 | dout(15) << __func__ << " " << cid << " " << oid | |
6365 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
6366 | << dendl; | |
6367 | if (!c->exists) | |
6368 | return -ENOENT; | |
6369 | ||
6370 | bl.clear(); | |
6371 | int r; | |
6372 | { | |
6373 | RWLock::RLocker l(c->lock); | |
6374 | utime_t start1 = ceph_clock_now(); | |
6375 | OnodeRef o = c->get_onode(oid, false); | |
6376 | logger->tinc(l_bluestore_read_onode_meta_lat, ceph_clock_now() - start1); | |
6377 | if (!o || !o->exists) { | |
6378 | r = -ENOENT; | |
6379 | goto out; | |
6380 | } | |
6381 | ||
6382 | if (offset == length && offset == 0) | |
6383 | length = o->onode.size; | |
6384 | ||
6385 | r = _do_read(c, o, offset, length, bl, op_flags); | |
6386 | } | |
6387 | ||
6388 | out: | |
7c673cae FG |
6389 | if (r == 0 && _debug_data_eio(oid)) { |
6390 | r = -EIO; | |
6391 | derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl; | |
224ce89b WB |
6392 | } else if (cct->_conf->bluestore_debug_random_read_err && |
6393 | (rand() % (int)(cct->_conf->bluestore_debug_random_read_err * 100.0)) == 0) { | |
6394 | dout(0) << __func__ << ": inject random EIO" << dendl; | |
6395 | r = -EIO; | |
7c673cae FG |
6396 | } |
6397 | dout(10) << __func__ << " " << cid << " " << oid | |
6398 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
6399 | << " = " << r << dendl; | |
6400 | logger->tinc(l_bluestore_read_lat, ceph_clock_now() - start); | |
6401 | return r; | |
6402 | } | |
6403 | ||
6404 | // -------------------------------------------------------- | |
6405 | // intermediate data structures used while reading | |
6406 | struct region_t { | |
6407 | uint64_t logical_offset; | |
6408 | uint64_t blob_xoffset; //region offset within the blob | |
6409 | uint64_t length; | |
6410 | bufferlist bl; | |
6411 | ||
6412 | // used later in read process | |
6413 | uint64_t front = 0; | |
6414 | uint64_t r_off = 0; | |
6415 | ||
6416 | region_t(uint64_t offset, uint64_t b_offs, uint64_t len) | |
6417 | : logical_offset(offset), | |
6418 | blob_xoffset(b_offs), | |
6419 | length(len){} | |
6420 | region_t(const region_t& from) | |
6421 | : logical_offset(from.logical_offset), | |
6422 | blob_xoffset(from.blob_xoffset), | |
6423 | length(from.length){} | |
6424 | ||
6425 | friend ostream& operator<<(ostream& out, const region_t& r) { | |
6426 | return out << "0x" << std::hex << r.logical_offset << ":" | |
6427 | << r.blob_xoffset << "~" << r.length << std::dec; | |
6428 | } | |
6429 | }; | |
6430 | ||
6431 | typedef list<region_t> regions2read_t; | |
6432 | typedef map<BlueStore::BlobRef, regions2read_t> blobs2read_t; | |
6433 | ||
6434 | int BlueStore::_do_read( | |
6435 | Collection *c, | |
6436 | OnodeRef o, | |
6437 | uint64_t offset, | |
6438 | size_t length, | |
6439 | bufferlist& bl, | |
6440 | uint32_t op_flags) | |
6441 | { | |
6442 | FUNCTRACE(); | |
7c673cae FG |
6443 | int r = 0; |
6444 | ||
6445 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
6446 | << " size 0x" << o->onode.size << " (" << std::dec | |
6447 | << o->onode.size << ")" << dendl; | |
6448 | bl.clear(); | |
6449 | ||
6450 | if (offset >= o->onode.size) { | |
6451 | return r; | |
6452 | } | |
6453 | ||
6454 | // generally, don't buffer anything, unless the client explicitly requests | |
6455 | // it. | |
6456 | bool buffered = false; | |
6457 | if (op_flags & CEPH_OSD_OP_FLAG_FADVISE_WILLNEED) { | |
6458 | dout(20) << __func__ << " will do buffered read" << dendl; | |
6459 | buffered = true; | |
6460 | } else if (cct->_conf->bluestore_default_buffered_read && | |
6461 | (op_flags & (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED | | |
6462 | CEPH_OSD_OP_FLAG_FADVISE_NOCACHE)) == 0) { | |
6463 | dout(20) << __func__ << " defaulting to buffered read" << dendl; | |
6464 | buffered = true; | |
6465 | } | |
6466 | ||
6467 | if (offset + length > o->onode.size) { | |
6468 | length = o->onode.size - offset; | |
6469 | } | |
6470 | ||
6471 | utime_t start = ceph_clock_now(); | |
6472 | o->extent_map.fault_range(db, offset, length); | |
6473 | logger->tinc(l_bluestore_read_onode_meta_lat, ceph_clock_now() - start); | |
6474 | _dump_onode(o); | |
6475 | ||
6476 | ready_regions_t ready_regions; | |
6477 | ||
6478 | // build blob-wise list to of stuff read (that isn't cached) | |
6479 | blobs2read_t blobs2read; | |
6480 | unsigned left = length; | |
6481 | uint64_t pos = offset; | |
6482 | unsigned num_regions = 0; | |
6483 | auto lp = o->extent_map.seek_lextent(offset); | |
6484 | while (left > 0 && lp != o->extent_map.extent_map.end()) { | |
6485 | if (pos < lp->logical_offset) { | |
6486 | unsigned hole = lp->logical_offset - pos; | |
6487 | if (hole >= left) { | |
6488 | break; | |
6489 | } | |
6490 | dout(30) << __func__ << " hole 0x" << std::hex << pos << "~" << hole | |
6491 | << std::dec << dendl; | |
6492 | pos += hole; | |
6493 | left -= hole; | |
6494 | } | |
6495 | BlobRef bptr = lp->blob; | |
6496 | unsigned l_off = pos - lp->logical_offset; | |
6497 | unsigned b_off = l_off + lp->blob_offset; | |
6498 | unsigned b_len = std::min(left, lp->length - l_off); | |
6499 | ||
6500 | ready_regions_t cache_res; | |
6501 | interval_set<uint32_t> cache_interval; | |
6502 | bptr->shared_blob->bc.read( | |
6503 | bptr->shared_blob->get_cache(), b_off, b_len, cache_res, cache_interval); | |
6504 | dout(20) << __func__ << " blob " << *bptr << std::hex | |
6505 | << " need 0x" << b_off << "~" << b_len | |
6506 | << " cache has 0x" << cache_interval | |
6507 | << std::dec << dendl; | |
6508 | ||
6509 | auto pc = cache_res.begin(); | |
6510 | while (b_len > 0) { | |
6511 | unsigned l; | |
6512 | if (pc != cache_res.end() && | |
6513 | pc->first == b_off) { | |
6514 | l = pc->second.length(); | |
6515 | ready_regions[pos].claim(pc->second); | |
6516 | dout(30) << __func__ << " use cache 0x" << std::hex << pos << ": 0x" | |
6517 | << b_off << "~" << l << std::dec << dendl; | |
6518 | ++pc; | |
6519 | } else { | |
6520 | l = b_len; | |
6521 | if (pc != cache_res.end()) { | |
6522 | assert(pc->first > b_off); | |
6523 | l = pc->first - b_off; | |
6524 | } | |
6525 | dout(30) << __func__ << " will read 0x" << std::hex << pos << ": 0x" | |
6526 | << b_off << "~" << l << std::dec << dendl; | |
6527 | blobs2read[bptr].emplace_back(region_t(pos, b_off, l)); | |
6528 | ++num_regions; | |
6529 | } | |
6530 | pos += l; | |
6531 | b_off += l; | |
6532 | left -= l; | |
6533 | b_len -= l; | |
6534 | } | |
6535 | ++lp; | |
6536 | } | |
6537 | ||
6538 | // read raw blob data. use aio if we have >1 blobs to read. | |
6539 | start = ceph_clock_now(); // for the sake of simplicity | |
6540 | // measure the whole block below. | |
6541 | // The error isn't that much... | |
6542 | vector<bufferlist> compressed_blob_bls; | |
6543 | IOContext ioc(cct, NULL); | |
6544 | for (auto& p : blobs2read) { | |
6545 | BlobRef bptr = p.first; | |
6546 | dout(20) << __func__ << " blob " << *bptr << std::hex | |
6547 | << " need " << p.second << std::dec << dendl; | |
6548 | if (bptr->get_blob().is_compressed()) { | |
6549 | // read the whole thing | |
6550 | if (compressed_blob_bls.empty()) { | |
6551 | // ensure we avoid any reallocation on subsequent blobs | |
6552 | compressed_blob_bls.reserve(blobs2read.size()); | |
6553 | } | |
6554 | compressed_blob_bls.push_back(bufferlist()); | |
6555 | bufferlist& bl = compressed_blob_bls.back(); | |
6556 | r = bptr->get_blob().map( | |
6557 | 0, bptr->get_blob().get_ondisk_length(), | |
6558 | [&](uint64_t offset, uint64_t length) { | |
6559 | int r; | |
6560 | // use aio if there are more regions to read than those in this blob | |
6561 | if (num_regions > p.second.size()) { | |
6562 | r = bdev->aio_read(offset, length, &bl, &ioc); | |
6563 | } else { | |
6564 | r = bdev->read(offset, length, &bl, &ioc, false); | |
6565 | } | |
6566 | if (r < 0) | |
6567 | return r; | |
6568 | return 0; | |
6569 | }); | |
6570 | assert(r == 0); | |
6571 | } else { | |
6572 | // read the pieces | |
6573 | for (auto& reg : p.second) { | |
6574 | // determine how much of the blob to read | |
6575 | uint64_t chunk_size = bptr->get_blob().get_chunk_size(block_size); | |
6576 | reg.r_off = reg.blob_xoffset; | |
6577 | uint64_t r_len = reg.length; | |
6578 | reg.front = reg.r_off % chunk_size; | |
6579 | if (reg.front) { | |
6580 | reg.r_off -= reg.front; | |
6581 | r_len += reg.front; | |
6582 | } | |
6583 | unsigned tail = r_len % chunk_size; | |
6584 | if (tail) { | |
6585 | r_len += chunk_size - tail; | |
6586 | } | |
6587 | dout(20) << __func__ << " region 0x" << std::hex | |
6588 | << reg.logical_offset | |
6589 | << ": 0x" << reg.blob_xoffset << "~" << reg.length | |
6590 | << " reading 0x" << reg.r_off << "~" << r_len << std::dec | |
6591 | << dendl; | |
6592 | ||
6593 | // read it | |
6594 | r = bptr->get_blob().map( | |
6595 | reg.r_off, r_len, | |
6596 | [&](uint64_t offset, uint64_t length) { | |
6597 | int r; | |
6598 | // use aio if there is more than one region to read | |
6599 | if (num_regions > 1) { | |
6600 | r = bdev->aio_read(offset, length, ®.bl, &ioc); | |
6601 | } else { | |
6602 | r = bdev->read(offset, length, ®.bl, &ioc, false); | |
6603 | } | |
6604 | if (r < 0) | |
6605 | return r; | |
6606 | return 0; | |
6607 | }); | |
6608 | assert(r == 0); | |
6609 | assert(reg.bl.length() == r_len); | |
6610 | } | |
6611 | } | |
6612 | } | |
6613 | if (ioc.has_pending_aios()) { | |
6614 | bdev->aio_submit(&ioc); | |
6615 | dout(20) << __func__ << " waiting for aio" << dendl; | |
6616 | ioc.aio_wait(); | |
6617 | } | |
6618 | logger->tinc(l_bluestore_read_wait_aio_lat, ceph_clock_now() - start); | |
6619 | ||
6620 | // enumerate and decompress desired blobs | |
6621 | auto p = compressed_blob_bls.begin(); | |
6622 | blobs2read_t::iterator b2r_it = blobs2read.begin(); | |
6623 | while (b2r_it != blobs2read.end()) { | |
6624 | BlobRef bptr = b2r_it->first; | |
6625 | dout(20) << __func__ << " blob " << *bptr << std::hex | |
6626 | << " need 0x" << b2r_it->second << std::dec << dendl; | |
6627 | if (bptr->get_blob().is_compressed()) { | |
6628 | assert(p != compressed_blob_bls.end()); | |
6629 | bufferlist& compressed_bl = *p++; | |
6630 | if (_verify_csum(o, &bptr->get_blob(), 0, compressed_bl, | |
6631 | b2r_it->second.front().logical_offset) < 0) { | |
6632 | return -EIO; | |
6633 | } | |
6634 | bufferlist raw_bl; | |
6635 | r = _decompress(compressed_bl, &raw_bl); | |
6636 | if (r < 0) | |
6637 | return r; | |
6638 | if (buffered) { | |
6639 | bptr->shared_blob->bc.did_read(bptr->shared_blob->get_cache(), 0, | |
6640 | raw_bl); | |
6641 | } | |
6642 | for (auto& i : b2r_it->second) { | |
6643 | ready_regions[i.logical_offset].substr_of( | |
6644 | raw_bl, i.blob_xoffset, i.length); | |
6645 | } | |
6646 | } else { | |
6647 | for (auto& reg : b2r_it->second) { | |
6648 | if (_verify_csum(o, &bptr->get_blob(), reg.r_off, reg.bl, | |
6649 | reg.logical_offset) < 0) { | |
6650 | return -EIO; | |
6651 | } | |
6652 | if (buffered) { | |
6653 | bptr->shared_blob->bc.did_read(bptr->shared_blob->get_cache(), | |
6654 | reg.r_off, reg.bl); | |
6655 | } | |
6656 | ||
6657 | // prune and keep result | |
6658 | ready_regions[reg.logical_offset].substr_of( | |
6659 | reg.bl, reg.front, reg.length); | |
6660 | } | |
6661 | } | |
6662 | ++b2r_it; | |
6663 | } | |
6664 | ||
6665 | // generate a resulting buffer | |
6666 | auto pr = ready_regions.begin(); | |
6667 | auto pr_end = ready_regions.end(); | |
6668 | pos = 0; | |
6669 | while (pos < length) { | |
6670 | if (pr != pr_end && pr->first == pos + offset) { | |
6671 | dout(30) << __func__ << " assemble 0x" << std::hex << pos | |
6672 | << ": data from 0x" << pr->first << "~" << pr->second.length() | |
6673 | << std::dec << dendl; | |
6674 | pos += pr->second.length(); | |
6675 | bl.claim_append(pr->second); | |
6676 | ++pr; | |
6677 | } else { | |
6678 | uint64_t l = length - pos; | |
6679 | if (pr != pr_end) { | |
6680 | assert(pr->first > pos + offset); | |
6681 | l = pr->first - (pos + offset); | |
6682 | } | |
6683 | dout(30) << __func__ << " assemble 0x" << std::hex << pos | |
6684 | << ": zeros for 0x" << (pos + offset) << "~" << l | |
6685 | << std::dec << dendl; | |
6686 | bl.append_zero(l); | |
6687 | pos += l; | |
6688 | } | |
6689 | } | |
6690 | assert(bl.length() == length); | |
6691 | assert(pos == length); | |
6692 | assert(pr == pr_end); | |
6693 | r = bl.length(); | |
6694 | return r; | |
6695 | } | |
6696 | ||
6697 | int BlueStore::_verify_csum(OnodeRef& o, | |
6698 | const bluestore_blob_t* blob, uint64_t blob_xoffset, | |
6699 | const bufferlist& bl, | |
6700 | uint64_t logical_offset) const | |
6701 | { | |
6702 | int bad; | |
6703 | uint64_t bad_csum; | |
6704 | utime_t start = ceph_clock_now(); | |
6705 | int r = blob->verify_csum(blob_xoffset, bl, &bad, &bad_csum); | |
6706 | if (r < 0) { | |
6707 | if (r == -1) { | |
6708 | PExtentVector pex; | |
6709 | blob->map( | |
6710 | bad, | |
6711 | blob->get_csum_chunk_size(), | |
6712 | [&](uint64_t offset, uint64_t length) { | |
6713 | pex.emplace_back(bluestore_pextent_t(offset, length)); | |
6714 | return 0; | |
6715 | }); | |
6716 | derr << __func__ << " bad " | |
6717 | << Checksummer::get_csum_type_string(blob->csum_type) | |
6718 | << "/0x" << std::hex << blob->get_csum_chunk_size() | |
6719 | << " checksum at blob offset 0x" << bad | |
6720 | << ", got 0x" << bad_csum << ", expected 0x" | |
6721 | << blob->get_csum_item(bad / blob->get_csum_chunk_size()) << std::dec | |
6722 | << ", device location " << pex | |
6723 | << ", logical extent 0x" << std::hex | |
6724 | << (logical_offset + bad - blob_xoffset) << "~" | |
6725 | << blob->get_csum_chunk_size() << std::dec | |
6726 | << ", object " << o->oid | |
6727 | << dendl; | |
6728 | } else { | |
6729 | derr << __func__ << " failed with exit code: " << cpp_strerror(r) << dendl; | |
6730 | } | |
6731 | } | |
6732 | logger->tinc(l_bluestore_csum_lat, ceph_clock_now() - start); | |
6733 | return r; | |
6734 | } | |
6735 | ||
6736 | int BlueStore::_decompress(bufferlist& source, bufferlist* result) | |
6737 | { | |
6738 | int r = 0; | |
6739 | utime_t start = ceph_clock_now(); | |
6740 | bufferlist::iterator i = source.begin(); | |
6741 | bluestore_compression_header_t chdr; | |
6742 | ::decode(chdr, i); | |
6743 | int alg = int(chdr.type); | |
6744 | CompressorRef cp = compressor; | |
6745 | if (!cp || (int)cp->get_type() != alg) { | |
6746 | cp = Compressor::create(cct, alg); | |
6747 | } | |
6748 | ||
6749 | if (!cp.get()) { | |
6750 | // if compressor isn't available - error, because cannot return | |
6751 | // decompressed data? | |
6752 | derr << __func__ << " can't load decompressor " << alg << dendl; | |
6753 | r = -EIO; | |
6754 | } else { | |
6755 | r = cp->decompress(i, chdr.length, *result); | |
6756 | if (r < 0) { | |
6757 | derr << __func__ << " decompression failed with exit code " << r << dendl; | |
6758 | r = -EIO; | |
6759 | } | |
6760 | } | |
6761 | logger->tinc(l_bluestore_decompress_lat, ceph_clock_now() - start); | |
6762 | return r; | |
6763 | } | |
6764 | ||
6765 | // this stores fiemap into interval_set, other variations | |
6766 | // use it internally | |
6767 | int BlueStore::_fiemap( | |
6768 | CollectionHandle &c_, | |
6769 | const ghobject_t& oid, | |
6770 | uint64_t offset, | |
6771 | size_t length, | |
6772 | interval_set<uint64_t>& destset) | |
6773 | { | |
6774 | Collection *c = static_cast<Collection *>(c_.get()); | |
6775 | if (!c->exists) | |
6776 | return -ENOENT; | |
6777 | { | |
6778 | RWLock::RLocker l(c->lock); | |
6779 | ||
6780 | OnodeRef o = c->get_onode(oid, false); | |
6781 | if (!o || !o->exists) { | |
6782 | return -ENOENT; | |
6783 | } | |
6784 | _dump_onode(o); | |
6785 | ||
6786 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length | |
6787 | << " size 0x" << o->onode.size << std::dec << dendl; | |
6788 | ||
6789 | boost::intrusive::set<Extent>::iterator ep, eend; | |
6790 | if (offset >= o->onode.size) | |
6791 | goto out; | |
6792 | ||
6793 | if (offset + length > o->onode.size) { | |
6794 | length = o->onode.size - offset; | |
6795 | } | |
6796 | ||
6797 | o->extent_map.fault_range(db, offset, length); | |
6798 | eend = o->extent_map.extent_map.end(); | |
6799 | ep = o->extent_map.seek_lextent(offset); | |
6800 | while (length > 0) { | |
6801 | dout(20) << __func__ << " offset " << offset << dendl; | |
6802 | if (ep != eend && ep->logical_offset + ep->length <= offset) { | |
6803 | ++ep; | |
6804 | continue; | |
6805 | } | |
6806 | ||
6807 | uint64_t x_len = length; | |
6808 | if (ep != eend && ep->logical_offset <= offset) { | |
6809 | uint64_t x_off = offset - ep->logical_offset; | |
6810 | x_len = MIN(x_len, ep->length - x_off); | |
6811 | dout(30) << __func__ << " lextent 0x" << std::hex << offset << "~" | |
6812 | << x_len << std::dec << " blob " << ep->blob << dendl; | |
6813 | destset.insert(offset, x_len); | |
6814 | length -= x_len; | |
6815 | offset += x_len; | |
6816 | if (x_off + x_len == ep->length) | |
6817 | ++ep; | |
6818 | continue; | |
6819 | } | |
6820 | if (ep != eend && | |
6821 | ep->logical_offset > offset && | |
6822 | ep->logical_offset - offset < x_len) { | |
6823 | x_len = ep->logical_offset - offset; | |
6824 | } | |
6825 | offset += x_len; | |
6826 | length -= x_len; | |
6827 | } | |
6828 | } | |
6829 | ||
6830 | out: | |
7c673cae FG |
6831 | dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length |
6832 | << " size = 0x(" << destset << ")" << std::dec << dendl; | |
6833 | return 0; | |
6834 | } | |
6835 | ||
6836 | int BlueStore::fiemap( | |
6837 | const coll_t& cid, | |
6838 | const ghobject_t& oid, | |
6839 | uint64_t offset, | |
6840 | size_t len, | |
6841 | bufferlist& bl) | |
6842 | { | |
6843 | CollectionHandle c = _get_collection(cid); | |
6844 | if (!c) | |
6845 | return -ENOENT; | |
6846 | return fiemap(c, oid, offset, len, bl); | |
6847 | } | |
6848 | ||
6849 | int BlueStore::fiemap( | |
6850 | CollectionHandle &c_, | |
6851 | const ghobject_t& oid, | |
6852 | uint64_t offset, | |
6853 | size_t length, | |
6854 | bufferlist& bl) | |
6855 | { | |
6856 | interval_set<uint64_t> m; | |
6857 | int r = _fiemap(c_, oid, offset, length, m); | |
6858 | if (r >= 0) { | |
6859 | ::encode(m, bl); | |
6860 | } | |
6861 | return r; | |
6862 | } | |
6863 | ||
6864 | int BlueStore::fiemap( | |
6865 | const coll_t& cid, | |
6866 | const ghobject_t& oid, | |
6867 | uint64_t offset, | |
6868 | size_t len, | |
6869 | map<uint64_t, uint64_t>& destmap) | |
6870 | { | |
6871 | CollectionHandle c = _get_collection(cid); | |
6872 | if (!c) | |
6873 | return -ENOENT; | |
6874 | return fiemap(c, oid, offset, len, destmap); | |
6875 | } | |
6876 | ||
6877 | int BlueStore::fiemap( | |
6878 | CollectionHandle &c_, | |
6879 | const ghobject_t& oid, | |
6880 | uint64_t offset, | |
6881 | size_t length, | |
6882 | map<uint64_t, uint64_t>& destmap) | |
6883 | { | |
6884 | interval_set<uint64_t> m; | |
6885 | int r = _fiemap(c_, oid, offset, length, m); | |
6886 | if (r >= 0) { | |
6887 | m.move_into(destmap); | |
6888 | } | |
6889 | return r; | |
6890 | } | |
6891 | ||
6892 | int BlueStore::getattr( | |
6893 | const coll_t& cid, | |
6894 | const ghobject_t& oid, | |
6895 | const char *name, | |
6896 | bufferptr& value) | |
6897 | { | |
6898 | CollectionHandle c = _get_collection(cid); | |
6899 | if (!c) | |
6900 | return -ENOENT; | |
6901 | return getattr(c, oid, name, value); | |
6902 | } | |
6903 | ||
6904 | int BlueStore::getattr( | |
6905 | CollectionHandle &c_, | |
6906 | const ghobject_t& oid, | |
6907 | const char *name, | |
6908 | bufferptr& value) | |
6909 | { | |
6910 | Collection *c = static_cast<Collection *>(c_.get()); | |
6911 | dout(15) << __func__ << " " << c->cid << " " << oid << " " << name << dendl; | |
6912 | if (!c->exists) | |
6913 | return -ENOENT; | |
6914 | ||
6915 | int r; | |
6916 | { | |
6917 | RWLock::RLocker l(c->lock); | |
31f18b77 | 6918 | mempool::bluestore_cache_other::string k(name); |
7c673cae FG |
6919 | |
6920 | OnodeRef o = c->get_onode(oid, false); | |
6921 | if (!o || !o->exists) { | |
6922 | r = -ENOENT; | |
6923 | goto out; | |
6924 | } | |
6925 | ||
6926 | if (!o->onode.attrs.count(k)) { | |
6927 | r = -ENODATA; | |
6928 | goto out; | |
6929 | } | |
6930 | value = o->onode.attrs[k]; | |
6931 | r = 0; | |
6932 | } | |
6933 | out: | |
7c673cae FG |
6934 | if (r == 0 && _debug_mdata_eio(oid)) { |
6935 | r = -EIO; | |
6936 | derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl; | |
6937 | } | |
6938 | dout(10) << __func__ << " " << c->cid << " " << oid << " " << name | |
6939 | << " = " << r << dendl; | |
6940 | return r; | |
6941 | } | |
6942 | ||
6943 | ||
6944 | int BlueStore::getattrs( | |
6945 | const coll_t& cid, | |
6946 | const ghobject_t& oid, | |
6947 | map<string,bufferptr>& aset) | |
6948 | { | |
6949 | CollectionHandle c = _get_collection(cid); | |
6950 | if (!c) | |
6951 | return -ENOENT; | |
6952 | return getattrs(c, oid, aset); | |
6953 | } | |
6954 | ||
6955 | int BlueStore::getattrs( | |
6956 | CollectionHandle &c_, | |
6957 | const ghobject_t& oid, | |
6958 | map<string,bufferptr>& aset) | |
6959 | { | |
6960 | Collection *c = static_cast<Collection *>(c_.get()); | |
6961 | dout(15) << __func__ << " " << c->cid << " " << oid << dendl; | |
6962 | if (!c->exists) | |
6963 | return -ENOENT; | |
6964 | ||
6965 | int r; | |
6966 | { | |
6967 | RWLock::RLocker l(c->lock); | |
6968 | ||
6969 | OnodeRef o = c->get_onode(oid, false); | |
6970 | if (!o || !o->exists) { | |
6971 | r = -ENOENT; | |
6972 | goto out; | |
6973 | } | |
6974 | for (auto& i : o->onode.attrs) { | |
6975 | aset.emplace(i.first.c_str(), i.second); | |
6976 | } | |
6977 | r = 0; | |
6978 | } | |
6979 | ||
6980 | out: | |
7c673cae FG |
6981 | if (r == 0 && _debug_mdata_eio(oid)) { |
6982 | r = -EIO; | |
6983 | derr << __func__ << " " << c->cid << " " << oid << " INJECT EIO" << dendl; | |
6984 | } | |
6985 | dout(10) << __func__ << " " << c->cid << " " << oid | |
6986 | << " = " << r << dendl; | |
6987 | return r; | |
6988 | } | |
6989 | ||
6990 | int BlueStore::list_collections(vector<coll_t>& ls) | |
6991 | { | |
6992 | RWLock::RLocker l(coll_lock); | |
6993 | for (ceph::unordered_map<coll_t, CollectionRef>::iterator p = coll_map.begin(); | |
6994 | p != coll_map.end(); | |
6995 | ++p) | |
6996 | ls.push_back(p->first); | |
6997 | return 0; | |
6998 | } | |
6999 | ||
7000 | bool BlueStore::collection_exists(const coll_t& c) | |
7001 | { | |
7002 | RWLock::RLocker l(coll_lock); | |
7003 | return coll_map.count(c); | |
7004 | } | |
7005 | ||
7006 | int BlueStore::collection_empty(const coll_t& cid, bool *empty) | |
7007 | { | |
7008 | dout(15) << __func__ << " " << cid << dendl; | |
7009 | vector<ghobject_t> ls; | |
7010 | ghobject_t next; | |
7011 | int r = collection_list(cid, ghobject_t(), ghobject_t::get_max(), 1, | |
7012 | &ls, &next); | |
7013 | if (r < 0) { | |
7014 | derr << __func__ << " collection_list returned: " << cpp_strerror(r) | |
7015 | << dendl; | |
7016 | return r; | |
7017 | } | |
7018 | *empty = ls.empty(); | |
7019 | dout(10) << __func__ << " " << cid << " = " << (int)(*empty) << dendl; | |
7020 | return 0; | |
7021 | } | |
7022 | ||
7023 | int BlueStore::collection_bits(const coll_t& cid) | |
7024 | { | |
7025 | dout(15) << __func__ << " " << cid << dendl; | |
7026 | CollectionRef c = _get_collection(cid); | |
7027 | if (!c) | |
7028 | return -ENOENT; | |
7029 | RWLock::RLocker l(c->lock); | |
7030 | dout(10) << __func__ << " " << cid << " = " << c->cnode.bits << dendl; | |
7031 | return c->cnode.bits; | |
7032 | } | |
7033 | ||
7034 | int BlueStore::collection_list( | |
7035 | const coll_t& cid, const ghobject_t& start, const ghobject_t& end, int max, | |
7036 | vector<ghobject_t> *ls, ghobject_t *pnext) | |
7037 | { | |
7038 | CollectionHandle c = _get_collection(cid); | |
7039 | if (!c) | |
7040 | return -ENOENT; | |
7041 | return collection_list(c, start, end, max, ls, pnext); | |
7042 | } | |
7043 | ||
7044 | int BlueStore::collection_list( | |
7045 | CollectionHandle &c_, const ghobject_t& start, const ghobject_t& end, int max, | |
7046 | vector<ghobject_t> *ls, ghobject_t *pnext) | |
7047 | { | |
7048 | Collection *c = static_cast<Collection *>(c_.get()); | |
7049 | dout(15) << __func__ << " " << c->cid | |
7050 | << " start " << start << " end " << end << " max " << max << dendl; | |
7051 | int r; | |
7052 | { | |
7053 | RWLock::RLocker l(c->lock); | |
7054 | r = _collection_list(c, start, end, max, ls, pnext); | |
7055 | } | |
7056 | ||
7c673cae FG |
7057 | dout(10) << __func__ << " " << c->cid |
7058 | << " start " << start << " end " << end << " max " << max | |
7059 | << " = " << r << ", ls.size() = " << ls->size() | |
7060 | << ", next = " << (pnext ? *pnext : ghobject_t()) << dendl; | |
7061 | return r; | |
7062 | } | |
7063 | ||
7064 | int BlueStore::_collection_list( | |
7065 | Collection *c, const ghobject_t& start, const ghobject_t& end, int max, | |
7066 | vector<ghobject_t> *ls, ghobject_t *pnext) | |
7067 | { | |
7068 | ||
7069 | if (!c->exists) | |
7070 | return -ENOENT; | |
7071 | ||
7072 | int r = 0; | |
7073 | ghobject_t static_next; | |
7074 | KeyValueDB::Iterator it; | |
7075 | string temp_start_key, temp_end_key; | |
7076 | string start_key, end_key; | |
7077 | bool set_next = false; | |
7078 | string pend; | |
7079 | bool temp; | |
7080 | ||
7081 | if (!pnext) | |
7082 | pnext = &static_next; | |
7083 | ||
7084 | if (start == ghobject_t::get_max() || | |
7085 | start.hobj.is_max()) { | |
7086 | goto out; | |
7087 | } | |
7088 | get_coll_key_range(c->cid, c->cnode.bits, &temp_start_key, &temp_end_key, | |
7089 | &start_key, &end_key); | |
7090 | dout(20) << __func__ | |
7091 | << " range " << pretty_binary_string(temp_start_key) | |
7092 | << " to " << pretty_binary_string(temp_end_key) | |
7093 | << " and " << pretty_binary_string(start_key) | |
7094 | << " to " << pretty_binary_string(end_key) | |
7095 | << " start " << start << dendl; | |
7096 | it = db->get_iterator(PREFIX_OBJ); | |
7097 | if (start == ghobject_t() || | |
7098 | start.hobj == hobject_t() || | |
7099 | start == c->cid.get_min_hobj()) { | |
7100 | it->upper_bound(temp_start_key); | |
7101 | temp = true; | |
7102 | } else { | |
7103 | string k; | |
7104 | get_object_key(cct, start, &k); | |
7105 | if (start.hobj.is_temp()) { | |
7106 | temp = true; | |
7107 | assert(k >= temp_start_key && k < temp_end_key); | |
7108 | } else { | |
7109 | temp = false; | |
7110 | assert(k >= start_key && k < end_key); | |
7111 | } | |
7112 | dout(20) << " start from " << pretty_binary_string(k) | |
7113 | << " temp=" << (int)temp << dendl; | |
7114 | it->lower_bound(k); | |
7115 | } | |
7116 | if (end.hobj.is_max()) { | |
7117 | pend = temp ? temp_end_key : end_key; | |
7118 | } else { | |
7119 | get_object_key(cct, end, &end_key); | |
7120 | if (end.hobj.is_temp()) { | |
7121 | if (temp) | |
7122 | pend = end_key; | |
7123 | else | |
7124 | goto out; | |
7125 | } else { | |
7126 | pend = temp ? temp_end_key : end_key; | |
7127 | } | |
7128 | } | |
7129 | dout(20) << __func__ << " pend " << pretty_binary_string(pend) << dendl; | |
7130 | while (true) { | |
7131 | if (!it->valid() || it->key() >= pend) { | |
7132 | if (!it->valid()) | |
7133 | dout(20) << __func__ << " iterator not valid (end of db?)" << dendl; | |
7134 | else | |
7135 | dout(20) << __func__ << " key " << pretty_binary_string(it->key()) | |
7136 | << " >= " << end << dendl; | |
7137 | if (temp) { | |
7138 | if (end.hobj.is_temp()) { | |
7139 | break; | |
7140 | } | |
7141 | dout(30) << __func__ << " switch to non-temp namespace" << dendl; | |
7142 | temp = false; | |
7143 | it->upper_bound(start_key); | |
7144 | pend = end_key; | |
7145 | dout(30) << __func__ << " pend " << pretty_binary_string(pend) << dendl; | |
7146 | continue; | |
7147 | } | |
7148 | break; | |
7149 | } | |
7150 | dout(30) << __func__ << " key " << pretty_binary_string(it->key()) << dendl; | |
7151 | if (is_extent_shard_key(it->key())) { | |
7152 | it->next(); | |
7153 | continue; | |
7154 | } | |
7155 | ghobject_t oid; | |
7156 | int r = get_key_object(it->key(), &oid); | |
7157 | assert(r == 0); | |
7158 | dout(20) << __func__ << " oid " << oid << " end " << end << dendl; | |
7159 | if (ls->size() >= (unsigned)max) { | |
7160 | dout(20) << __func__ << " reached max " << max << dendl; | |
7161 | *pnext = oid; | |
7162 | set_next = true; | |
7163 | break; | |
7164 | } | |
7165 | ls->push_back(oid); | |
7166 | it->next(); | |
7167 | } | |
7168 | out: | |
7169 | if (!set_next) { | |
7170 | *pnext = ghobject_t::get_max(); | |
7171 | } | |
7172 | ||
7173 | return r; | |
7174 | } | |
7175 | ||
7c673cae FG |
7176 | int BlueStore::omap_get( |
7177 | const coll_t& cid, ///< [in] Collection containing oid | |
7178 | const ghobject_t &oid, ///< [in] Object containing omap | |
7179 | bufferlist *header, ///< [out] omap header | |
7180 | map<string, bufferlist> *out /// < [out] Key to value map | |
7181 | ) | |
7182 | { | |
7183 | CollectionHandle c = _get_collection(cid); | |
7184 | if (!c) | |
7185 | return -ENOENT; | |
7186 | return omap_get(c, oid, header, out); | |
7187 | } | |
7188 | ||
7189 | int BlueStore::omap_get( | |
7190 | CollectionHandle &c_, ///< [in] Collection containing oid | |
7191 | const ghobject_t &oid, ///< [in] Object containing omap | |
7192 | bufferlist *header, ///< [out] omap header | |
7193 | map<string, bufferlist> *out /// < [out] Key to value map | |
7194 | ) | |
7195 | { | |
7196 | Collection *c = static_cast<Collection *>(c_.get()); | |
7197 | dout(15) << __func__ << " " << c->get_cid() << " oid " << oid << dendl; | |
7198 | if (!c->exists) | |
7199 | return -ENOENT; | |
7200 | RWLock::RLocker l(c->lock); | |
7201 | int r = 0; | |
7202 | OnodeRef o = c->get_onode(oid, false); | |
7203 | if (!o || !o->exists) { | |
7204 | r = -ENOENT; | |
7205 | goto out; | |
7206 | } | |
7207 | if (!o->onode.has_omap()) | |
7208 | goto out; | |
7209 | o->flush(); | |
7210 | { | |
7211 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_OMAP); | |
7212 | string head, tail; | |
7213 | get_omap_header(o->onode.nid, &head); | |
7214 | get_omap_tail(o->onode.nid, &tail); | |
7215 | it->lower_bound(head); | |
7216 | while (it->valid()) { | |
7217 | if (it->key() == head) { | |
7218 | dout(30) << __func__ << " got header" << dendl; | |
7219 | *header = it->value(); | |
7220 | } else if (it->key() >= tail) { | |
7221 | dout(30) << __func__ << " reached tail" << dendl; | |
7222 | break; | |
7223 | } else { | |
7224 | string user_key; | |
7225 | decode_omap_key(it->key(), &user_key); | |
7226 | dout(30) << __func__ << " got " << pretty_binary_string(it->key()) | |
7227 | << " -> " << user_key << dendl; | |
7228 | (*out)[user_key] = it->value(); | |
7229 | } | |
7230 | it->next(); | |
7231 | } | |
7232 | } | |
7233 | out: | |
7234 | dout(10) << __func__ << " " << c->get_cid() << " oid " << oid << " = " << r | |
7235 | << dendl; | |
7236 | return r; | |
7237 | } | |
7238 | ||
7239 | int BlueStore::omap_get_header( | |
7240 | const coll_t& cid, ///< [in] Collection containing oid | |
7241 | const ghobject_t &oid, ///< [in] Object containing omap | |
7242 | bufferlist *header, ///< [out] omap header | |
7243 | bool allow_eio ///< [in] don't assert on eio | |
7244 | ) | |
7245 | { | |
7246 | CollectionHandle c = _get_collection(cid); | |
7247 | if (!c) | |
7248 | return -ENOENT; | |
7249 | return omap_get_header(c, oid, header, allow_eio); | |
7250 | } | |
7251 | ||
7252 | int BlueStore::omap_get_header( | |
7253 | CollectionHandle &c_, ///< [in] Collection containing oid | |
7254 | const ghobject_t &oid, ///< [in] Object containing omap | |
7255 | bufferlist *header, ///< [out] omap header | |
7256 | bool allow_eio ///< [in] don't assert on eio | |
7257 | ) | |
7258 | { | |
7259 | Collection *c = static_cast<Collection *>(c_.get()); | |
7260 | dout(15) << __func__ << " " << c->get_cid() << " oid " << oid << dendl; | |
7261 | if (!c->exists) | |
7262 | return -ENOENT; | |
7263 | RWLock::RLocker l(c->lock); | |
7264 | int r = 0; | |
7265 | OnodeRef o = c->get_onode(oid, false); | |
7266 | if (!o || !o->exists) { | |
7267 | r = -ENOENT; | |
7268 | goto out; | |
7269 | } | |
7270 | if (!o->onode.has_omap()) | |
7271 | goto out; | |
7272 | o->flush(); | |
7273 | { | |
7274 | string head; | |
7275 | get_omap_header(o->onode.nid, &head); | |
7276 | if (db->get(PREFIX_OMAP, head, header) >= 0) { | |
7277 | dout(30) << __func__ << " got header" << dendl; | |
7278 | } else { | |
7279 | dout(30) << __func__ << " no header" << dendl; | |
7280 | } | |
7281 | } | |
7282 | out: | |
7283 | dout(10) << __func__ << " " << c->get_cid() << " oid " << oid << " = " << r | |
7284 | << dendl; | |
7285 | return r; | |
7286 | } | |
7287 | ||
7288 | int BlueStore::omap_get_keys( | |
7289 | const coll_t& cid, ///< [in] Collection containing oid | |
7290 | const ghobject_t &oid, ///< [in] Object containing omap | |
7291 | set<string> *keys ///< [out] Keys defined on oid | |
7292 | ) | |
7293 | { | |
7294 | CollectionHandle c = _get_collection(cid); | |
7295 | if (!c) | |
7296 | return -ENOENT; | |
7297 | return omap_get_keys(c, oid, keys); | |
7298 | } | |
7299 | ||
7300 | int BlueStore::omap_get_keys( | |
7301 | CollectionHandle &c_, ///< [in] Collection containing oid | |
7302 | const ghobject_t &oid, ///< [in] Object containing omap | |
7303 | set<string> *keys ///< [out] Keys defined on oid | |
7304 | ) | |
7305 | { | |
7306 | Collection *c = static_cast<Collection *>(c_.get()); | |
7307 | dout(15) << __func__ << " " << c->get_cid() << " oid " << oid << dendl; | |
7308 | if (!c->exists) | |
7309 | return -ENOENT; | |
7310 | RWLock::RLocker l(c->lock); | |
7311 | int r = 0; | |
7312 | OnodeRef o = c->get_onode(oid, false); | |
7313 | if (!o || !o->exists) { | |
7314 | r = -ENOENT; | |
7315 | goto out; | |
7316 | } | |
7317 | if (!o->onode.has_omap()) | |
7318 | goto out; | |
7319 | o->flush(); | |
7320 | { | |
7321 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_OMAP); | |
7322 | string head, tail; | |
7323 | get_omap_key(o->onode.nid, string(), &head); | |
7324 | get_omap_tail(o->onode.nid, &tail); | |
7325 | it->lower_bound(head); | |
7326 | while (it->valid()) { | |
7327 | if (it->key() >= tail) { | |
7328 | dout(30) << __func__ << " reached tail" << dendl; | |
7329 | break; | |
7330 | } | |
7331 | string user_key; | |
7332 | decode_omap_key(it->key(), &user_key); | |
7333 | dout(30) << __func__ << " got " << pretty_binary_string(it->key()) | |
7334 | << " -> " << user_key << dendl; | |
7335 | keys->insert(user_key); | |
7336 | it->next(); | |
7337 | } | |
7338 | } | |
7339 | out: | |
7340 | dout(10) << __func__ << " " << c->get_cid() << " oid " << oid << " = " << r | |
7341 | << dendl; | |
7342 | return r; | |
7343 | } | |
7344 | ||
7345 | int BlueStore::omap_get_values( | |
7346 | const coll_t& cid, ///< [in] Collection containing oid | |
7347 | const ghobject_t &oid, ///< [in] Object containing omap | |
7348 | const set<string> &keys, ///< [in] Keys to get | |
7349 | map<string, bufferlist> *out ///< [out] Returned keys and values | |
7350 | ) | |
7351 | { | |
7352 | CollectionHandle c = _get_collection(cid); | |
7353 | if (!c) | |
7354 | return -ENOENT; | |
7355 | return omap_get_values(c, oid, keys, out); | |
7356 | } | |
7357 | ||
7358 | int BlueStore::omap_get_values( | |
7359 | CollectionHandle &c_, ///< [in] Collection containing oid | |
7360 | const ghobject_t &oid, ///< [in] Object containing omap | |
7361 | const set<string> &keys, ///< [in] Keys to get | |
7362 | map<string, bufferlist> *out ///< [out] Returned keys and values | |
7363 | ) | |
7364 | { | |
7365 | Collection *c = static_cast<Collection *>(c_.get()); | |
7366 | dout(15) << __func__ << " " << c->get_cid() << " oid " << oid << dendl; | |
7367 | if (!c->exists) | |
7368 | return -ENOENT; | |
7369 | RWLock::RLocker l(c->lock); | |
7370 | int r = 0; | |
7371 | string final_key; | |
7372 | OnodeRef o = c->get_onode(oid, false); | |
7373 | if (!o || !o->exists) { | |
7374 | r = -ENOENT; | |
7375 | goto out; | |
7376 | } | |
7377 | if (!o->onode.has_omap()) | |
7378 | goto out; | |
7379 | o->flush(); | |
7380 | _key_encode_u64(o->onode.nid, &final_key); | |
7381 | final_key.push_back('.'); | |
7382 | for (set<string>::const_iterator p = keys.begin(); p != keys.end(); ++p) { | |
7383 | final_key.resize(9); // keep prefix | |
7384 | final_key += *p; | |
7385 | bufferlist val; | |
7386 | if (db->get(PREFIX_OMAP, final_key, &val) >= 0) { | |
7387 | dout(30) << __func__ << " got " << pretty_binary_string(final_key) | |
7388 | << " -> " << *p << dendl; | |
7389 | out->insert(make_pair(*p, val)); | |
7390 | } | |
7391 | } | |
7392 | out: | |
7393 | dout(10) << __func__ << " " << c->get_cid() << " oid " << oid << " = " << r | |
7394 | << dendl; | |
7395 | return r; | |
7396 | } | |
7397 | ||
7398 | int BlueStore::omap_check_keys( | |
7399 | const coll_t& cid, ///< [in] Collection containing oid | |
7400 | const ghobject_t &oid, ///< [in] Object containing omap | |
7401 | const set<string> &keys, ///< [in] Keys to check | |
7402 | set<string> *out ///< [out] Subset of keys defined on oid | |
7403 | ) | |
7404 | { | |
7405 | CollectionHandle c = _get_collection(cid); | |
7406 | if (!c) | |
7407 | return -ENOENT; | |
7408 | return omap_check_keys(c, oid, keys, out); | |
7409 | } | |
7410 | ||
7411 | int BlueStore::omap_check_keys( | |
7412 | CollectionHandle &c_, ///< [in] Collection containing oid | |
7413 | const ghobject_t &oid, ///< [in] Object containing omap | |
7414 | const set<string> &keys, ///< [in] Keys to check | |
7415 | set<string> *out ///< [out] Subset of keys defined on oid | |
7416 | ) | |
7417 | { | |
7418 | Collection *c = static_cast<Collection *>(c_.get()); | |
7419 | dout(15) << __func__ << " " << c->get_cid() << " oid " << oid << dendl; | |
7420 | if (!c->exists) | |
7421 | return -ENOENT; | |
7422 | RWLock::RLocker l(c->lock); | |
7423 | int r = 0; | |
7424 | string final_key; | |
7425 | OnodeRef o = c->get_onode(oid, false); | |
7426 | if (!o || !o->exists) { | |
7427 | r = -ENOENT; | |
7428 | goto out; | |
7429 | } | |
7430 | if (!o->onode.has_omap()) | |
7431 | goto out; | |
7432 | o->flush(); | |
7433 | _key_encode_u64(o->onode.nid, &final_key); | |
7434 | final_key.push_back('.'); | |
7435 | for (set<string>::const_iterator p = keys.begin(); p != keys.end(); ++p) { | |
7436 | final_key.resize(9); // keep prefix | |
7437 | final_key += *p; | |
7438 | bufferlist val; | |
7439 | if (db->get(PREFIX_OMAP, final_key, &val) >= 0) { | |
7440 | dout(30) << __func__ << " have " << pretty_binary_string(final_key) | |
7441 | << " -> " << *p << dendl; | |
7442 | out->insert(*p); | |
7443 | } else { | |
7444 | dout(30) << __func__ << " miss " << pretty_binary_string(final_key) | |
7445 | << " -> " << *p << dendl; | |
7446 | } | |
7447 | } | |
7448 | out: | |
7449 | dout(10) << __func__ << " " << c->get_cid() << " oid " << oid << " = " << r | |
7450 | << dendl; | |
7451 | return r; | |
7452 | } | |
7453 | ||
7454 | ObjectMap::ObjectMapIterator BlueStore::get_omap_iterator( | |
7455 | const coll_t& cid, ///< [in] collection | |
7456 | const ghobject_t &oid ///< [in] object | |
7457 | ) | |
7458 | { | |
7459 | CollectionHandle c = _get_collection(cid); | |
7460 | if (!c) { | |
7461 | dout(10) << __func__ << " " << cid << "doesn't exist" <<dendl; | |
7462 | return ObjectMap::ObjectMapIterator(); | |
7463 | } | |
7464 | return get_omap_iterator(c, oid); | |
7465 | } | |
7466 | ||
7467 | ObjectMap::ObjectMapIterator BlueStore::get_omap_iterator( | |
7468 | CollectionHandle &c_, ///< [in] collection | |
7469 | const ghobject_t &oid ///< [in] object | |
7470 | ) | |
7471 | { | |
7472 | Collection *c = static_cast<Collection *>(c_.get()); | |
7473 | dout(10) << __func__ << " " << c->get_cid() << " " << oid << dendl; | |
7474 | if (!c->exists) { | |
7475 | return ObjectMap::ObjectMapIterator(); | |
7476 | } | |
7477 | RWLock::RLocker l(c->lock); | |
7478 | OnodeRef o = c->get_onode(oid, false); | |
7479 | if (!o || !o->exists) { | |
7480 | dout(10) << __func__ << " " << oid << "doesn't exist" <<dendl; | |
7481 | return ObjectMap::ObjectMapIterator(); | |
7482 | } | |
7483 | o->flush(); | |
7484 | dout(10) << __func__ << " has_omap = " << (int)o->onode.has_omap() <<dendl; | |
7485 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_OMAP); | |
7486 | return ObjectMap::ObjectMapIterator(new OmapIteratorImpl(c, o, it)); | |
7487 | } | |
7488 | ||
7489 | // ----------------- | |
7490 | // write helpers | |
7491 | ||
7492 | void BlueStore::_prepare_ondisk_format_super(KeyValueDB::Transaction& t) | |
7493 | { | |
7494 | dout(10) << __func__ << " ondisk_format " << ondisk_format | |
7495 | << " min_compat_ondisk_format " << min_compat_ondisk_format | |
7496 | << dendl; | |
7497 | assert(ondisk_format == latest_ondisk_format); | |
7498 | { | |
7499 | bufferlist bl; | |
7500 | ::encode(ondisk_format, bl); | |
7501 | t->set(PREFIX_SUPER, "ondisk_format", bl); | |
7502 | } | |
7503 | { | |
7504 | bufferlist bl; | |
7505 | ::encode(min_compat_ondisk_format, bl); | |
7506 | t->set(PREFIX_SUPER, "min_compat_ondisk_format", bl); | |
7507 | } | |
7508 | } | |
7509 | ||
7510 | int BlueStore::_open_super_meta() | |
7511 | { | |
7512 | // nid | |
7513 | { | |
7514 | nid_max = 0; | |
7515 | bufferlist bl; | |
7516 | db->get(PREFIX_SUPER, "nid_max", &bl); | |
7517 | bufferlist::iterator p = bl.begin(); | |
7518 | try { | |
7519 | uint64_t v; | |
7520 | ::decode(v, p); | |
7521 | nid_max = v; | |
7522 | } catch (buffer::error& e) { | |
7523 | derr << __func__ << " unable to read nid_max" << dendl; | |
7524 | return -EIO; | |
7525 | } | |
7526 | dout(10) << __func__ << " old nid_max " << nid_max << dendl; | |
7527 | nid_last = nid_max.load(); | |
7528 | } | |
7529 | ||
7530 | // blobid | |
7531 | { | |
7532 | blobid_max = 0; | |
7533 | bufferlist bl; | |
7534 | db->get(PREFIX_SUPER, "blobid_max", &bl); | |
7535 | bufferlist::iterator p = bl.begin(); | |
7536 | try { | |
7537 | uint64_t v; | |
7538 | ::decode(v, p); | |
7539 | blobid_max = v; | |
7540 | } catch (buffer::error& e) { | |
7541 | derr << __func__ << " unable to read blobid_max" << dendl; | |
7542 | return -EIO; | |
7543 | } | |
7544 | dout(10) << __func__ << " old blobid_max " << blobid_max << dendl; | |
7545 | blobid_last = blobid_max.load(); | |
7546 | } | |
7547 | ||
7548 | // freelist | |
7549 | { | |
7550 | bufferlist bl; | |
7551 | db->get(PREFIX_SUPER, "freelist_type", &bl); | |
7552 | if (bl.length()) { | |
7553 | freelist_type = std::string(bl.c_str(), bl.length()); | |
7554 | dout(10) << __func__ << " freelist_type " << freelist_type << dendl; | |
7555 | } else { | |
7556 | assert("Not Support extent freelist manager" == 0); | |
7557 | } | |
7558 | } | |
7559 | ||
7560 | // bluefs alloc | |
7561 | if (cct->_conf->bluestore_bluefs) { | |
7562 | bluefs_extents.clear(); | |
7563 | bufferlist bl; | |
7564 | db->get(PREFIX_SUPER, "bluefs_extents", &bl); | |
7565 | bufferlist::iterator p = bl.begin(); | |
7566 | try { | |
7567 | ::decode(bluefs_extents, p); | |
7568 | } | |
7569 | catch (buffer::error& e) { | |
7570 | derr << __func__ << " unable to read bluefs_extents" << dendl; | |
7571 | return -EIO; | |
7572 | } | |
7573 | dout(10) << __func__ << " bluefs_extents 0x" << std::hex << bluefs_extents | |
7574 | << std::dec << dendl; | |
7575 | } | |
7576 | ||
7577 | // ondisk format | |
7578 | int32_t compat_ondisk_format = 0; | |
7579 | { | |
7580 | bufferlist bl; | |
7581 | int r = db->get(PREFIX_SUPER, "ondisk_format", &bl); | |
7582 | if (r < 0) { | |
7583 | // base case: kraken bluestore is v1 and readable by v1 | |
7584 | dout(20) << __func__ << " missing ondisk_format; assuming kraken" | |
7585 | << dendl; | |
7586 | ondisk_format = 1; | |
7587 | compat_ondisk_format = 1; | |
7588 | } else { | |
7589 | auto p = bl.begin(); | |
7590 | try { | |
7591 | ::decode(ondisk_format, p); | |
7592 | } catch (buffer::error& e) { | |
7593 | derr << __func__ << " unable to read ondisk_format" << dendl; | |
7594 | return -EIO; | |
7595 | } | |
7596 | bl.clear(); | |
7597 | { | |
7598 | r = db->get(PREFIX_SUPER, "min_compat_ondisk_format", &bl); | |
7599 | assert(!r); | |
7600 | auto p = bl.begin(); | |
7601 | try { | |
7602 | ::decode(compat_ondisk_format, p); | |
7603 | } catch (buffer::error& e) { | |
7604 | derr << __func__ << " unable to read compat_ondisk_format" << dendl; | |
7605 | return -EIO; | |
7606 | } | |
7607 | } | |
7608 | } | |
7609 | dout(10) << __func__ << " ondisk_format " << ondisk_format | |
7610 | << " compat_ondisk_format " << compat_ondisk_format | |
7611 | << dendl; | |
7612 | } | |
7613 | ||
7614 | if (latest_ondisk_format < compat_ondisk_format) { | |
7615 | derr << __func__ << " compat_ondisk_format is " | |
7616 | << compat_ondisk_format << " but we only understand version " | |
7617 | << latest_ondisk_format << dendl; | |
7618 | return -EPERM; | |
7619 | } | |
7620 | if (ondisk_format < latest_ondisk_format) { | |
7621 | int r = _upgrade_super(); | |
7622 | if (r < 0) { | |
7623 | return r; | |
7624 | } | |
7625 | } | |
7626 | ||
7627 | { | |
7628 | bufferlist bl; | |
7629 | db->get(PREFIX_SUPER, "min_alloc_size", &bl); | |
7630 | auto p = bl.begin(); | |
7631 | try { | |
7632 | uint64_t val; | |
7633 | ::decode(val, p); | |
7634 | min_alloc_size = val; | |
224ce89b WB |
7635 | min_alloc_size_order = ctz(val); |
7636 | assert(min_alloc_size == 1u << min_alloc_size_order); | |
7c673cae FG |
7637 | } catch (buffer::error& e) { |
7638 | derr << __func__ << " unable to read min_alloc_size" << dendl; | |
7639 | return -EIO; | |
7640 | } | |
7641 | dout(10) << __func__ << " min_alloc_size 0x" << std::hex << min_alloc_size | |
7642 | << std::dec << dendl; | |
7643 | } | |
224ce89b | 7644 | _open_statfs(); |
7c673cae FG |
7645 | _set_alloc_sizes(); |
7646 | _set_throttle_params(); | |
7647 | ||
7648 | _set_csum(); | |
7649 | _set_compression(); | |
7650 | _set_blob_size(); | |
7651 | ||
7652 | return 0; | |
7653 | } | |
7654 | ||
7655 | int BlueStore::_upgrade_super() | |
7656 | { | |
7657 | dout(1) << __func__ << " from " << ondisk_format << ", latest " | |
7658 | << latest_ondisk_format << dendl; | |
7659 | assert(ondisk_format > 0); | |
7660 | assert(ondisk_format < latest_ondisk_format); | |
7661 | ||
7662 | if (ondisk_format == 1) { | |
7663 | // changes: | |
7664 | // - super: added ondisk_format | |
7665 | // - super: added min_readable_ondisk_format | |
7666 | // - super: added min_compat_ondisk_format | |
7667 | // - super: added min_alloc_size | |
7668 | // - super: removed min_min_alloc_size | |
7669 | KeyValueDB::Transaction t = db->get_transaction(); | |
7670 | { | |
7671 | bufferlist bl; | |
7672 | db->get(PREFIX_SUPER, "min_min_alloc_size", &bl); | |
7673 | auto p = bl.begin(); | |
7674 | try { | |
7675 | uint64_t val; | |
7676 | ::decode(val, p); | |
7677 | min_alloc_size = val; | |
7678 | } catch (buffer::error& e) { | |
7679 | derr << __func__ << " failed to read min_min_alloc_size" << dendl; | |
7680 | return -EIO; | |
7681 | } | |
7682 | t->set(PREFIX_SUPER, "min_alloc_size", bl); | |
7683 | t->rmkey(PREFIX_SUPER, "min_min_alloc_size"); | |
7684 | } | |
7685 | ondisk_format = 2; | |
7686 | _prepare_ondisk_format_super(t); | |
7687 | int r = db->submit_transaction_sync(t); | |
7688 | assert(r == 0); | |
7689 | } | |
7690 | ||
7691 | // done | |
7692 | dout(1) << __func__ << " done" << dendl; | |
7693 | return 0; | |
7694 | } | |
7695 | ||
7696 | void BlueStore::_assign_nid(TransContext *txc, OnodeRef o) | |
7697 | { | |
224ce89b WB |
7698 | if (o->onode.nid) { |
7699 | assert(o->exists); | |
7c673cae | 7700 | return; |
224ce89b | 7701 | } |
7c673cae FG |
7702 | uint64_t nid = ++nid_last; |
7703 | dout(20) << __func__ << " " << nid << dendl; | |
7704 | o->onode.nid = nid; | |
7705 | txc->last_nid = nid; | |
224ce89b | 7706 | o->exists = true; |
7c673cae FG |
7707 | } |
7708 | ||
7709 | uint64_t BlueStore::_assign_blobid(TransContext *txc) | |
7710 | { | |
7711 | uint64_t bid = ++blobid_last; | |
7712 | dout(20) << __func__ << " " << bid << dendl; | |
7713 | txc->last_blobid = bid; | |
7714 | return bid; | |
7715 | } | |
7716 | ||
7717 | void BlueStore::get_db_statistics(Formatter *f) | |
7718 | { | |
7719 | db->get_statistics(f); | |
7720 | } | |
7721 | ||
7722 | BlueStore::TransContext *BlueStore::_txc_create(OpSequencer *osr) | |
7723 | { | |
7724 | TransContext *txc = new TransContext(cct, osr); | |
7725 | txc->t = db->get_transaction(); | |
7726 | osr->queue_new(txc); | |
7727 | dout(20) << __func__ << " osr " << osr << " = " << txc | |
7728 | << " seq " << txc->seq << dendl; | |
7729 | return txc; | |
7730 | } | |
7731 | ||
7732 | void BlueStore::_txc_calc_cost(TransContext *txc) | |
7733 | { | |
7734 | // this is about the simplest model for transaction cost you can | |
7735 | // imagine. there is some fixed overhead cost by saying there is a | |
7736 | // minimum of one "io". and then we have some cost per "io" that is | |
7737 | // a configurable (with different hdd and ssd defaults), and add | |
7738 | // that to the bytes value. | |
7739 | int ios = 1; // one "io" for the kv commit | |
7740 | for (auto& p : txc->ioc.pending_aios) { | |
7741 | ios += p.iov.size(); | |
7742 | } | |
7743 | auto cost = throttle_cost_per_io.load(); | |
7744 | txc->cost = ios * cost + txc->bytes; | |
7745 | dout(10) << __func__ << " " << txc << " cost " << txc->cost << " (" | |
7746 | << ios << " ios * " << cost << " + " << txc->bytes | |
7747 | << " bytes)" << dendl; | |
7748 | } | |
7749 | ||
7750 | void BlueStore::_txc_update_store_statfs(TransContext *txc) | |
7751 | { | |
7752 | if (txc->statfs_delta.is_empty()) | |
7753 | return; | |
7754 | ||
7755 | logger->inc(l_bluestore_allocated, txc->statfs_delta.allocated()); | |
7756 | logger->inc(l_bluestore_stored, txc->statfs_delta.stored()); | |
7757 | logger->inc(l_bluestore_compressed, txc->statfs_delta.compressed()); | |
7758 | logger->inc(l_bluestore_compressed_allocated, txc->statfs_delta.compressed_allocated()); | |
7759 | logger->inc(l_bluestore_compressed_original, txc->statfs_delta.compressed_original()); | |
7760 | ||
31f18b77 FG |
7761 | { |
7762 | std::lock_guard<std::mutex> l(vstatfs_lock); | |
7763 | vstatfs += txc->statfs_delta; | |
7764 | } | |
7765 | ||
7c673cae FG |
7766 | bufferlist bl; |
7767 | txc->statfs_delta.encode(bl); | |
7768 | ||
7769 | txc->t->merge(PREFIX_STAT, "bluestore_statfs", bl); | |
7770 | txc->statfs_delta.reset(); | |
7771 | } | |
7772 | ||
7773 | void BlueStore::_txc_state_proc(TransContext *txc) | |
7774 | { | |
7775 | while (true) { | |
7776 | dout(10) << __func__ << " txc " << txc | |
7777 | << " " << txc->get_state_name() << dendl; | |
7778 | switch (txc->state) { | |
7779 | case TransContext::STATE_PREPARE: | |
7780 | txc->log_state_latency(logger, l_bluestore_state_prepare_lat); | |
7781 | if (txc->ioc.has_pending_aios()) { | |
7782 | txc->state = TransContext::STATE_AIO_WAIT; | |
7783 | txc->had_ios = true; | |
7784 | _txc_aio_submit(txc); | |
7785 | return; | |
7786 | } | |
7787 | // ** fall-thru ** | |
7788 | ||
7789 | case TransContext::STATE_AIO_WAIT: | |
7790 | txc->log_state_latency(logger, l_bluestore_state_aio_wait_lat); | |
7791 | _txc_finish_io(txc); // may trigger blocked txc's too | |
7792 | return; | |
7793 | ||
7794 | case TransContext::STATE_IO_DONE: | |
7795 | //assert(txc->osr->qlock.is_locked()); // see _txc_finish_io | |
7796 | if (txc->had_ios) { | |
7797 | ++txc->osr->txc_with_unstable_io; | |
7798 | } | |
7799 | txc->log_state_latency(logger, l_bluestore_state_io_done_lat); | |
7800 | txc->state = TransContext::STATE_KV_QUEUED; | |
7801 | if (cct->_conf->bluestore_sync_submit_transaction) { | |
7802 | if (txc->last_nid >= nid_max || | |
7803 | txc->last_blobid >= blobid_max) { | |
7804 | dout(20) << __func__ | |
7805 | << " last_{nid,blobid} exceeds max, submit via kv thread" | |
7806 | << dendl; | |
7807 | } else if (txc->osr->kv_committing_serially) { | |
7808 | dout(20) << __func__ << " prior txc submitted via kv thread, us too" | |
7809 | << dendl; | |
7810 | // note: this is starvation-prone. once we have a txc in a busy | |
7811 | // sequencer that is committing serially it is possible to keep | |
7812 | // submitting new transactions fast enough that we get stuck doing | |
7813 | // so. the alternative is to block here... fixme? | |
7814 | } else if (txc->osr->txc_with_unstable_io) { | |
7815 | dout(20) << __func__ << " prior txc(s) with unstable ios " | |
7816 | << txc->osr->txc_with_unstable_io.load() << dendl; | |
7817 | } else if (cct->_conf->bluestore_debug_randomize_serial_transaction && | |
7818 | rand() % cct->_conf->bluestore_debug_randomize_serial_transaction | |
7819 | == 0) { | |
7820 | dout(20) << __func__ << " DEBUG randomly forcing submit via kv thread" | |
7821 | << dendl; | |
7822 | } else { | |
7823 | txc->state = TransContext::STATE_KV_SUBMITTED; | |
31f18b77 | 7824 | int r = cct->_conf->bluestore_debug_omit_kv_commit ? 0 : db->submit_transaction(txc->t); |
7c673cae FG |
7825 | assert(r == 0); |
7826 | _txc_applied_kv(txc); | |
7827 | } | |
7828 | } | |
7829 | { | |
7830 | std::lock_guard<std::mutex> l(kv_lock); | |
7831 | kv_queue.push_back(txc); | |
7832 | kv_cond.notify_one(); | |
7833 | if (txc->state != TransContext::STATE_KV_SUBMITTED) { | |
7834 | kv_queue_unsubmitted.push_back(txc); | |
7835 | ++txc->osr->kv_committing_serially; | |
7836 | } | |
31f18b77 FG |
7837 | if (txc->had_ios) |
7838 | kv_ios++; | |
7839 | kv_throttle_costs += txc->cost; | |
7c673cae FG |
7840 | } |
7841 | return; | |
7842 | case TransContext::STATE_KV_SUBMITTED: | |
7843 | txc->log_state_latency(logger, l_bluestore_state_kv_committing_lat); | |
7844 | txc->state = TransContext::STATE_KV_DONE; | |
7845 | _txc_committed_kv(txc); | |
7846 | // ** fall-thru ** | |
7847 | ||
7848 | case TransContext::STATE_KV_DONE: | |
7849 | txc->log_state_latency(logger, l_bluestore_state_kv_done_lat); | |
7850 | if (txc->deferred_txn) { | |
7851 | txc->state = TransContext::STATE_DEFERRED_QUEUED; | |
7852 | _deferred_queue(txc); | |
7853 | return; | |
7854 | } | |
7855 | txc->state = TransContext::STATE_FINISHING; | |
7856 | break; | |
7857 | ||
7858 | case TransContext::STATE_DEFERRED_CLEANUP: | |
7859 | txc->log_state_latency(logger, l_bluestore_state_deferred_cleanup_lat); | |
7860 | txc->state = TransContext::STATE_FINISHING; | |
7861 | // ** fall-thru ** | |
7862 | ||
7863 | case TransContext::STATE_FINISHING: | |
7864 | txc->log_state_latency(logger, l_bluestore_state_finishing_lat); | |
7865 | _txc_finish(txc); | |
7866 | return; | |
7867 | ||
7868 | default: | |
7869 | derr << __func__ << " unexpected txc " << txc | |
7870 | << " state " << txc->get_state_name() << dendl; | |
7871 | assert(0 == "unexpected txc state"); | |
7872 | return; | |
7873 | } | |
7874 | } | |
7875 | } | |
7876 | ||
7877 | void BlueStore::_txc_finish_io(TransContext *txc) | |
7878 | { | |
7879 | dout(20) << __func__ << " " << txc << dendl; | |
7880 | ||
7881 | /* | |
7882 | * we need to preserve the order of kv transactions, | |
7883 | * even though aio will complete in any order. | |
7884 | */ | |
7885 | ||
7886 | OpSequencer *osr = txc->osr.get(); | |
7887 | std::lock_guard<std::mutex> l(osr->qlock); | |
7888 | txc->state = TransContext::STATE_IO_DONE; | |
7889 | ||
31f18b77 FG |
7890 | // release aio contexts (including pinned buffers). |
7891 | txc->ioc.running_aios.clear(); | |
7892 | ||
7c673cae FG |
7893 | OpSequencer::q_list_t::iterator p = osr->q.iterator_to(*txc); |
7894 | while (p != osr->q.begin()) { | |
7895 | --p; | |
7896 | if (p->state < TransContext::STATE_IO_DONE) { | |
7897 | dout(20) << __func__ << " " << txc << " blocked by " << &*p << " " | |
7898 | << p->get_state_name() << dendl; | |
7899 | return; | |
7900 | } | |
7901 | if (p->state > TransContext::STATE_IO_DONE) { | |
7902 | ++p; | |
7903 | break; | |
7904 | } | |
7905 | } | |
7906 | do { | |
7907 | _txc_state_proc(&*p++); | |
7908 | } while (p != osr->q.end() && | |
7909 | p->state == TransContext::STATE_IO_DONE); | |
7910 | ||
7911 | if (osr->kv_submitted_waiters && | |
7912 | osr->_is_all_kv_submitted()) { | |
7913 | osr->qcond.notify_all(); | |
7914 | } | |
7915 | } | |
7916 | ||
7917 | void BlueStore::_txc_write_nodes(TransContext *txc, KeyValueDB::Transaction t) | |
7918 | { | |
7919 | dout(20) << __func__ << " txc " << txc | |
7920 | << " onodes " << txc->onodes | |
7921 | << " shared_blobs " << txc->shared_blobs | |
7922 | << dendl; | |
7923 | ||
7924 | // finalize onodes | |
7925 | for (auto o : txc->onodes) { | |
7926 | // finalize extent_map shards | |
7927 | o->extent_map.update(t, false); | |
7928 | if (o->extent_map.needs_reshard()) { | |
7929 | o->extent_map.reshard(db, t); | |
7930 | o->extent_map.update(t, true); | |
7931 | if (o->extent_map.needs_reshard()) { | |
7932 | dout(20) << __func__ << " warning: still wants reshard, check options?" | |
7933 | << dendl; | |
7934 | o->extent_map.clear_needs_reshard(); | |
7935 | } | |
7936 | logger->inc(l_bluestore_onode_reshard); | |
7937 | } | |
7938 | ||
7939 | // bound encode | |
7940 | size_t bound = 0; | |
7941 | denc(o->onode, bound); | |
7942 | o->extent_map.bound_encode_spanning_blobs(bound); | |
7943 | if (o->onode.extent_map_shards.empty()) { | |
7944 | denc(o->extent_map.inline_bl, bound); | |
7945 | } | |
7946 | ||
7947 | // encode | |
7948 | bufferlist bl; | |
7949 | unsigned onode_part, blob_part, extent_part; | |
7950 | { | |
7951 | auto p = bl.get_contiguous_appender(bound, true); | |
7952 | denc(o->onode, p); | |
7953 | onode_part = p.get_logical_offset(); | |
7954 | o->extent_map.encode_spanning_blobs(p); | |
7955 | blob_part = p.get_logical_offset() - onode_part; | |
7956 | if (o->onode.extent_map_shards.empty()) { | |
7957 | denc(o->extent_map.inline_bl, p); | |
7958 | } | |
7959 | extent_part = p.get_logical_offset() - onode_part - blob_part; | |
7960 | } | |
7961 | ||
7962 | dout(20) << " onode " << o->oid << " is " << bl.length() | |
7963 | << " (" << onode_part << " bytes onode + " | |
7964 | << blob_part << " bytes spanning blobs + " | |
7965 | << extent_part << " bytes inline extents)" | |
7966 | << dendl; | |
7967 | t->set(PREFIX_OBJ, o->key.c_str(), o->key.size(), bl); | |
7968 | o->flushing_count++; | |
7969 | } | |
7970 | ||
7971 | // objects we modified but didn't affect the onode | |
7972 | auto p = txc->modified_objects.begin(); | |
7973 | while (p != txc->modified_objects.end()) { | |
7974 | if (txc->onodes.count(*p) == 0) { | |
7975 | (*p)->flushing_count++; | |
7976 | ++p; | |
7977 | } else { | |
7978 | // remove dups with onodes list to avoid problems in _txc_finish | |
7979 | p = txc->modified_objects.erase(p); | |
7980 | } | |
7981 | } | |
7982 | ||
7983 | // finalize shared_blobs | |
7984 | for (auto sb : txc->shared_blobs) { | |
7985 | string key; | |
7986 | auto sbid = sb->get_sbid(); | |
7987 | get_shared_blob_key(sbid, &key); | |
7988 | if (sb->persistent->empty()) { | |
7989 | dout(20) << " shared_blob 0x" << std::hex << sbid << std::dec | |
7990 | << " is empty" << dendl; | |
7991 | t->rmkey(PREFIX_SHARED_BLOB, key); | |
7992 | } else { | |
7993 | bufferlist bl; | |
7994 | ::encode(*(sb->persistent), bl); | |
7995 | dout(20) << " shared_blob 0x" << std::hex << sbid << std::dec | |
31f18b77 | 7996 | << " is " << bl.length() << " " << *sb << dendl; |
7c673cae FG |
7997 | t->set(PREFIX_SHARED_BLOB, key, bl); |
7998 | } | |
7999 | } | |
8000 | } | |
8001 | ||
8002 | void BlueStore::BSPerfTracker::update_from_perfcounters( | |
8003 | PerfCounters &logger) | |
8004 | { | |
8005 | os_commit_latency.consume_next( | |
8006 | logger.get_tavg_ms( | |
8007 | l_bluestore_commit_lat)); | |
8008 | os_apply_latency.consume_next( | |
8009 | logger.get_tavg_ms( | |
8010 | l_bluestore_commit_lat)); | |
8011 | } | |
8012 | ||
8013 | void BlueStore::_txc_finalize_kv(TransContext *txc, KeyValueDB::Transaction t) | |
8014 | { | |
8015 | dout(20) << __func__ << " txc " << txc << std::hex | |
8016 | << " allocated 0x" << txc->allocated | |
8017 | << " released 0x" << txc->released | |
8018 | << std::dec << dendl; | |
8019 | ||
8020 | // We have to handle the case where we allocate *and* deallocate the | |
8021 | // same region in this transaction. The freelist doesn't like that. | |
8022 | // (Actually, the only thing that cares is the BitmapFreelistManager | |
8023 | // debug check. But that's important.) | |
8024 | interval_set<uint64_t> tmp_allocated, tmp_released; | |
8025 | interval_set<uint64_t> *pallocated = &txc->allocated; | |
8026 | interval_set<uint64_t> *preleased = &txc->released; | |
8027 | if (!txc->allocated.empty() && !txc->released.empty()) { | |
8028 | interval_set<uint64_t> overlap; | |
8029 | overlap.intersection_of(txc->allocated, txc->released); | |
8030 | if (!overlap.empty()) { | |
8031 | tmp_allocated = txc->allocated; | |
8032 | tmp_allocated.subtract(overlap); | |
8033 | tmp_released = txc->released; | |
8034 | tmp_released.subtract(overlap); | |
8035 | dout(20) << __func__ << " overlap 0x" << std::hex << overlap | |
8036 | << ", new allocated 0x" << tmp_allocated | |
8037 | << " released 0x" << tmp_released << std::dec | |
8038 | << dendl; | |
8039 | pallocated = &tmp_allocated; | |
8040 | preleased = &tmp_released; | |
8041 | } | |
8042 | } | |
8043 | ||
8044 | // update freelist with non-overlap sets | |
8045 | for (interval_set<uint64_t>::iterator p = pallocated->begin(); | |
8046 | p != pallocated->end(); | |
8047 | ++p) { | |
8048 | fm->allocate(p.get_start(), p.get_len(), t); | |
8049 | } | |
8050 | for (interval_set<uint64_t>::iterator p = preleased->begin(); | |
8051 | p != preleased->end(); | |
8052 | ++p) { | |
8053 | dout(20) << __func__ << " release 0x" << std::hex << p.get_start() | |
8054 | << "~" << p.get_len() << std::dec << dendl; | |
8055 | fm->release(p.get_start(), p.get_len(), t); | |
8056 | } | |
8057 | ||
8058 | _txc_update_store_statfs(txc); | |
8059 | } | |
8060 | ||
8061 | void BlueStore::_txc_applied_kv(TransContext *txc) | |
8062 | { | |
8063 | for (auto ls : { &txc->onodes, &txc->modified_objects }) { | |
8064 | for (auto& o : *ls) { | |
8065 | dout(20) << __func__ << " onode " << o << " had " << o->flushing_count | |
8066 | << dendl; | |
8067 | if (--o->flushing_count == 0) { | |
8068 | std::lock_guard<std::mutex> l(o->flush_lock); | |
8069 | o->flush_cond.notify_all(); | |
8070 | } | |
8071 | } | |
8072 | } | |
8073 | } | |
8074 | ||
8075 | void BlueStore::_txc_committed_kv(TransContext *txc) | |
8076 | { | |
8077 | dout(20) << __func__ << " txc " << txc << dendl; | |
8078 | ||
8079 | // warning: we're calling onreadable_sync inside the sequencer lock | |
8080 | if (txc->onreadable_sync) { | |
8081 | txc->onreadable_sync->complete(0); | |
8082 | txc->onreadable_sync = NULL; | |
8083 | } | |
8084 | unsigned n = txc->osr->parent->shard_hint.hash_to_shard(m_finisher_num); | |
8085 | if (txc->oncommit) { | |
8086 | logger->tinc(l_bluestore_commit_lat, ceph_clock_now() - txc->start); | |
8087 | finishers[n]->queue(txc->oncommit); | |
8088 | txc->oncommit = NULL; | |
8089 | } | |
8090 | if (txc->onreadable) { | |
8091 | finishers[n]->queue(txc->onreadable); | |
8092 | txc->onreadable = NULL; | |
8093 | } | |
8094 | ||
8095 | if (!txc->oncommits.empty()) { | |
8096 | finishers[n]->queue(txc->oncommits); | |
8097 | } | |
8098 | } | |
8099 | ||
8100 | void BlueStore::_txc_finish(TransContext *txc) | |
8101 | { | |
8102 | dout(20) << __func__ << " " << txc << " onodes " << txc->onodes << dendl; | |
8103 | assert(txc->state == TransContext::STATE_FINISHING); | |
8104 | ||
8105 | for (auto& sb : txc->shared_blobs_written) { | |
8106 | sb->bc.finish_write(sb->get_cache(), txc->seq); | |
8107 | } | |
8108 | txc->shared_blobs_written.clear(); | |
8109 | ||
8110 | while (!txc->removed_collections.empty()) { | |
8111 | _queue_reap_collection(txc->removed_collections.front()); | |
8112 | txc->removed_collections.pop_front(); | |
8113 | } | |
8114 | ||
8115 | OpSequencerRef osr = txc->osr; | |
7c673cae | 8116 | bool empty = false; |
31f18b77 | 8117 | bool submit_deferred = false; |
7c673cae FG |
8118 | OpSequencer::q_list_t releasing_txc; |
8119 | { | |
8120 | std::lock_guard<std::mutex> l(osr->qlock); | |
8121 | txc->state = TransContext::STATE_DONE; | |
8122 | bool notify = false; | |
8123 | while (!osr->q.empty()) { | |
8124 | TransContext *txc = &osr->q.front(); | |
8125 | dout(20) << __func__ << " txc " << txc << " " << txc->get_state_name() | |
8126 | << dendl; | |
8127 | if (txc->state != TransContext::STATE_DONE) { | |
8128 | if (txc->state == TransContext::STATE_PREPARE && | |
8129 | deferred_aggressive) { | |
8130 | // for _osr_drain_preceding() | |
8131 | notify = true; | |
8132 | } | |
31f18b77 FG |
8133 | if (txc->state == TransContext::STATE_DEFERRED_QUEUED && |
8134 | osr->q.size() > g_conf->bluestore_max_deferred_txc) { | |
8135 | submit_deferred = true; | |
8136 | } | |
7c673cae FG |
8137 | break; |
8138 | } | |
8139 | ||
7c673cae FG |
8140 | osr->q.pop_front(); |
8141 | releasing_txc.push_back(*txc); | |
8142 | notify = true; | |
8143 | } | |
8144 | if (notify) { | |
8145 | osr->qcond.notify_all(); | |
8146 | } | |
8147 | if (osr->q.empty()) { | |
8148 | dout(20) << __func__ << " osr " << osr << " q now empty" << dendl; | |
8149 | empty = true; | |
8150 | } | |
8151 | } | |
8152 | while (!releasing_txc.empty()) { | |
8153 | // release to allocator only after all preceding txc's have also | |
8154 | // finished any deferred writes that potentially land in these | |
8155 | // blocks | |
8156 | auto txc = &releasing_txc.front(); | |
8157 | _txc_release_alloc(txc); | |
8158 | releasing_txc.pop_front(); | |
8159 | txc->log_state_latency(logger, l_bluestore_state_done_lat); | |
8160 | delete txc; | |
8161 | } | |
8162 | ||
31f18b77 FG |
8163 | if (submit_deferred) { |
8164 | // we're pinning memory; flush! we could be more fine-grained here but | |
8165 | // i'm not sure it's worth the bother. | |
8166 | deferred_try_submit(); | |
7c673cae FG |
8167 | } |
8168 | ||
7c673cae FG |
8169 | if (empty && osr->zombie) { |
8170 | dout(10) << __func__ << " reaping empty zombie osr " << osr << dendl; | |
8171 | osr->_unregister(); | |
8172 | } | |
8173 | } | |
8174 | ||
8175 | void BlueStore::_txc_release_alloc(TransContext *txc) | |
8176 | { | |
8177 | // update allocator with full released set | |
8178 | if (!cct->_conf->bluestore_debug_no_reuse_blocks) { | |
8179 | dout(10) << __func__ << " " << txc << " " << txc->released << dendl; | |
8180 | for (interval_set<uint64_t>::iterator p = txc->released.begin(); | |
8181 | p != txc->released.end(); | |
8182 | ++p) { | |
8183 | alloc->release(p.get_start(), p.get_len()); | |
8184 | } | |
8185 | } | |
8186 | ||
8187 | txc->allocated.clear(); | |
8188 | txc->released.clear(); | |
8189 | } | |
8190 | ||
8191 | void BlueStore::_osr_drain_preceding(TransContext *txc) | |
8192 | { | |
8193 | OpSequencer *osr = txc->osr.get(); | |
8194 | dout(10) << __func__ << " " << txc << " osr " << osr << dendl; | |
8195 | ++deferred_aggressive; // FIXME: maybe osr-local aggressive flag? | |
8196 | { | |
8197 | // submit anything pending | |
224ce89b | 8198 | deferred_lock.lock(); |
7c673cae | 8199 | if (osr->deferred_pending) { |
224ce89b WB |
8200 | _deferred_submit_unlock(osr); |
8201 | } else { | |
8202 | deferred_lock.unlock(); | |
7c673cae FG |
8203 | } |
8204 | } | |
8205 | { | |
8206 | // wake up any previously finished deferred events | |
8207 | std::lock_guard<std::mutex> l(kv_lock); | |
8208 | kv_cond.notify_one(); | |
8209 | } | |
8210 | osr->drain_preceding(txc); | |
8211 | --deferred_aggressive; | |
8212 | dout(10) << __func__ << " " << osr << " done" << dendl; | |
8213 | } | |
8214 | ||
8215 | void BlueStore::_osr_drain_all() | |
8216 | { | |
8217 | dout(10) << __func__ << dendl; | |
8218 | ||
8219 | set<OpSequencerRef> s; | |
8220 | { | |
8221 | std::lock_guard<std::mutex> l(osr_lock); | |
8222 | s = osr_set; | |
8223 | } | |
8224 | dout(20) << __func__ << " osr_set " << s << dendl; | |
8225 | ||
8226 | ++deferred_aggressive; | |
8227 | { | |
8228 | // submit anything pending | |
224ce89b | 8229 | deferred_try_submit(); |
7c673cae FG |
8230 | } |
8231 | { | |
8232 | // wake up any previously finished deferred events | |
8233 | std::lock_guard<std::mutex> l(kv_lock); | |
8234 | kv_cond.notify_one(); | |
8235 | } | |
31f18b77 FG |
8236 | { |
8237 | std::lock_guard<std::mutex> l(kv_finalize_lock); | |
8238 | kv_finalize_cond.notify_one(); | |
8239 | } | |
7c673cae FG |
8240 | for (auto osr : s) { |
8241 | dout(20) << __func__ << " drain " << osr << dendl; | |
8242 | osr->drain(); | |
8243 | } | |
8244 | --deferred_aggressive; | |
8245 | ||
8246 | dout(10) << __func__ << " done" << dendl; | |
8247 | } | |
8248 | ||
8249 | void BlueStore::_osr_unregister_all() | |
8250 | { | |
8251 | set<OpSequencerRef> s; | |
8252 | { | |
8253 | std::lock_guard<std::mutex> l(osr_lock); | |
8254 | s = osr_set; | |
8255 | } | |
8256 | dout(10) << __func__ << " " << s << dendl; | |
8257 | for (auto osr : s) { | |
8258 | osr->_unregister(); | |
8259 | ||
8260 | if (!osr->zombie) { | |
8261 | // break link from Sequencer to us so that this OpSequencer | |
8262 | // instance can die with this mount/umount cycle. note that | |
8263 | // we assume umount() will not race against ~Sequencer. | |
8264 | assert(osr->parent); | |
8265 | osr->parent->p.reset(); | |
8266 | } | |
8267 | } | |
8268 | // nobody should be creating sequencers during umount either. | |
8269 | { | |
8270 | std::lock_guard<std::mutex> l(osr_lock); | |
8271 | assert(osr_set.empty()); | |
8272 | } | |
8273 | } | |
8274 | ||
31f18b77 FG |
8275 | void BlueStore::_kv_start() |
8276 | { | |
8277 | dout(10) << __func__ << dendl; | |
8278 | ||
8279 | if (cct->_conf->bluestore_shard_finishers) { | |
8280 | if (cct->_conf->osd_op_num_shards) { | |
8281 | m_finisher_num = cct->_conf->osd_op_num_shards; | |
8282 | } else { | |
8283 | assert(bdev); | |
8284 | if (bdev->is_rotational()) { | |
8285 | m_finisher_num = cct->_conf->osd_op_num_shards_hdd; | |
8286 | } else { | |
8287 | m_finisher_num = cct->_conf->osd_op_num_shards_ssd; | |
8288 | } | |
8289 | } | |
8290 | } | |
8291 | ||
8292 | assert(m_finisher_num != 0); | |
8293 | ||
8294 | for (int i = 0; i < m_finisher_num; ++i) { | |
8295 | ostringstream oss; | |
8296 | oss << "finisher-" << i; | |
8297 | Finisher *f = new Finisher(cct, oss.str(), "finisher"); | |
8298 | finishers.push_back(f); | |
8299 | } | |
8300 | ||
181888fb | 8301 | deferred_finisher.start(); |
31f18b77 FG |
8302 | for (auto f : finishers) { |
8303 | f->start(); | |
8304 | } | |
8305 | kv_sync_thread.create("bstore_kv_sync"); | |
8306 | kv_finalize_thread.create("bstore_kv_final"); | |
8307 | } | |
8308 | ||
8309 | void BlueStore::_kv_stop() | |
8310 | { | |
8311 | dout(10) << __func__ << dendl; | |
8312 | { | |
8313 | std::unique_lock<std::mutex> l(kv_lock); | |
8314 | while (!kv_sync_started) { | |
8315 | kv_cond.wait(l); | |
8316 | } | |
8317 | kv_stop = true; | |
8318 | kv_cond.notify_all(); | |
8319 | } | |
8320 | { | |
8321 | std::unique_lock<std::mutex> l(kv_finalize_lock); | |
8322 | while (!kv_finalize_started) { | |
8323 | kv_finalize_cond.wait(l); | |
8324 | } | |
8325 | kv_finalize_stop = true; | |
8326 | kv_finalize_cond.notify_all(); | |
8327 | } | |
8328 | kv_sync_thread.join(); | |
8329 | kv_finalize_thread.join(); | |
8330 | { | |
8331 | std::lock_guard<std::mutex> l(kv_lock); | |
8332 | kv_stop = false; | |
8333 | } | |
8334 | { | |
8335 | std::lock_guard<std::mutex> l(kv_finalize_lock); | |
8336 | kv_finalize_stop = false; | |
8337 | } | |
8338 | dout(10) << __func__ << " stopping finishers" << dendl; | |
181888fb FG |
8339 | deferred_finisher.wait_for_empty(); |
8340 | deferred_finisher.stop(); | |
31f18b77 FG |
8341 | for (auto f : finishers) { |
8342 | f->wait_for_empty(); | |
8343 | f->stop(); | |
8344 | } | |
8345 | dout(10) << __func__ << " stopped" << dendl; | |
8346 | } | |
8347 | ||
7c673cae FG |
8348 | void BlueStore::_kv_sync_thread() |
8349 | { | |
8350 | dout(10) << __func__ << " start" << dendl; | |
8351 | std::unique_lock<std::mutex> l(kv_lock); | |
31f18b77 FG |
8352 | assert(!kv_sync_started); |
8353 | kv_sync_started = true; | |
8354 | kv_cond.notify_all(); | |
7c673cae FG |
8355 | while (true) { |
8356 | assert(kv_committing.empty()); | |
8357 | if (kv_queue.empty() && | |
8358 | ((deferred_done_queue.empty() && deferred_stable_queue.empty()) || | |
8359 | !deferred_aggressive)) { | |
8360 | if (kv_stop) | |
8361 | break; | |
8362 | dout(20) << __func__ << " sleep" << dendl; | |
8363 | kv_cond.wait(l); | |
8364 | dout(20) << __func__ << " wake" << dendl; | |
8365 | } else { | |
8366 | deque<TransContext*> kv_submitting; | |
8367 | deque<DeferredBatch*> deferred_done, deferred_stable; | |
31f18b77 FG |
8368 | uint64_t aios = 0, costs = 0; |
8369 | ||
7c673cae FG |
8370 | dout(20) << __func__ << " committing " << kv_queue.size() |
8371 | << " submitting " << kv_queue_unsubmitted.size() | |
8372 | << " deferred done " << deferred_done_queue.size() | |
8373 | << " stable " << deferred_stable_queue.size() | |
8374 | << dendl; | |
8375 | kv_committing.swap(kv_queue); | |
8376 | kv_submitting.swap(kv_queue_unsubmitted); | |
8377 | deferred_done.swap(deferred_done_queue); | |
8378 | deferred_stable.swap(deferred_stable_queue); | |
31f18b77 FG |
8379 | aios = kv_ios; |
8380 | costs = kv_throttle_costs; | |
8381 | kv_ios = 0; | |
8382 | kv_throttle_costs = 0; | |
7c673cae FG |
8383 | utime_t start = ceph_clock_now(); |
8384 | l.unlock(); | |
8385 | ||
8386 | dout(30) << __func__ << " committing " << kv_committing << dendl; | |
8387 | dout(30) << __func__ << " submitting " << kv_submitting << dendl; | |
8388 | dout(30) << __func__ << " deferred_done " << deferred_done << dendl; | |
8389 | dout(30) << __func__ << " deferred_stable " << deferred_stable << dendl; | |
8390 | ||
7c673cae FG |
8391 | bool force_flush = false; |
8392 | // if bluefs is sharing the same device as data (only), then we | |
8393 | // can rely on the bluefs commit to flush the device and make | |
8394 | // deferred aios stable. that means that if we do have done deferred | |
8395 | // txcs AND we are not on a single device, we need to force a flush. | |
8396 | if (bluefs_single_shared_device && bluefs) { | |
31f18b77 | 8397 | if (aios) { |
7c673cae FG |
8398 | force_flush = true; |
8399 | } else if (kv_committing.empty() && kv_submitting.empty() && | |
8400 | deferred_stable.empty()) { | |
8401 | force_flush = true; // there's nothing else to commit! | |
8402 | } else if (deferred_aggressive) { | |
8403 | force_flush = true; | |
8404 | } | |
8405 | } else | |
8406 | force_flush = true; | |
8407 | ||
8408 | if (force_flush) { | |
31f18b77 | 8409 | dout(20) << __func__ << " num_aios=" << aios |
7c673cae FG |
8410 | << " force_flush=" << (int)force_flush |
8411 | << ", flushing, deferred done->stable" << dendl; | |
8412 | // flush/barrier on block device | |
8413 | bdev->flush(); | |
8414 | ||
8415 | // if we flush then deferred done are now deferred stable | |
8416 | deferred_stable.insert(deferred_stable.end(), deferred_done.begin(), | |
8417 | deferred_done.end()); | |
8418 | deferred_done.clear(); | |
8419 | } | |
8420 | utime_t after_flush = ceph_clock_now(); | |
8421 | ||
8422 | // we will use one final transaction to force a sync | |
8423 | KeyValueDB::Transaction synct = db->get_transaction(); | |
8424 | ||
8425 | // increase {nid,blobid}_max? note that this covers both the | |
8426 | // case where we are approaching the max and the case we passed | |
8427 | // it. in either case, we increase the max in the earlier txn | |
8428 | // we submit. | |
8429 | uint64_t new_nid_max = 0, new_blobid_max = 0; | |
8430 | if (nid_last + cct->_conf->bluestore_nid_prealloc/2 > nid_max) { | |
8431 | KeyValueDB::Transaction t = | |
8432 | kv_submitting.empty() ? synct : kv_submitting.front()->t; | |
8433 | new_nid_max = nid_last + cct->_conf->bluestore_nid_prealloc; | |
8434 | bufferlist bl; | |
8435 | ::encode(new_nid_max, bl); | |
8436 | t->set(PREFIX_SUPER, "nid_max", bl); | |
8437 | dout(10) << __func__ << " new_nid_max " << new_nid_max << dendl; | |
8438 | } | |
8439 | if (blobid_last + cct->_conf->bluestore_blobid_prealloc/2 > blobid_max) { | |
8440 | KeyValueDB::Transaction t = | |
8441 | kv_submitting.empty() ? synct : kv_submitting.front()->t; | |
8442 | new_blobid_max = blobid_last + cct->_conf->bluestore_blobid_prealloc; | |
8443 | bufferlist bl; | |
8444 | ::encode(new_blobid_max, bl); | |
8445 | t->set(PREFIX_SUPER, "blobid_max", bl); | |
8446 | dout(10) << __func__ << " new_blobid_max " << new_blobid_max << dendl; | |
8447 | } | |
c07f9fc5 FG |
8448 | |
8449 | for (auto txc : kv_committing) { | |
8450 | if (txc->state == TransContext::STATE_KV_QUEUED) { | |
8451 | txc->log_state_latency(logger, l_bluestore_state_kv_queued_lat); | |
8452 | int r = cct->_conf->bluestore_debug_omit_kv_commit ? 0 : db->submit_transaction(txc->t); | |
8453 | assert(r == 0); | |
8454 | _txc_applied_kv(txc); | |
8455 | --txc->osr->kv_committing_serially; | |
8456 | txc->state = TransContext::STATE_KV_SUBMITTED; | |
8457 | if (txc->osr->kv_submitted_waiters) { | |
8458 | std::lock_guard<std::mutex> l(txc->osr->qlock); | |
8459 | if (txc->osr->_is_all_kv_submitted()) { | |
8460 | txc->osr->qcond.notify_all(); | |
8461 | } | |
7c673cae | 8462 | } |
c07f9fc5 FG |
8463 | |
8464 | } else { | |
8465 | assert(txc->state == TransContext::STATE_KV_SUBMITTED); | |
8466 | txc->log_state_latency(logger, l_bluestore_state_kv_queued_lat); | |
7c673cae | 8467 | } |
7c673cae FG |
8468 | if (txc->had_ios) { |
8469 | --txc->osr->txc_with_unstable_io; | |
8470 | } | |
7c673cae FG |
8471 | } |
8472 | ||
31f18b77 FG |
8473 | // release throttle *before* we commit. this allows new ops |
8474 | // to be prepared and enter pipeline while we are waiting on | |
8475 | // the kv commit sync/flush. then hopefully on the next | |
8476 | // iteration there will already be ops awake. otherwise, we | |
8477 | // end up going to sleep, and then wake up when the very first | |
8478 | // transaction is ready for commit. | |
8479 | throttle_bytes.put(costs); | |
8480 | ||
7c673cae FG |
8481 | PExtentVector bluefs_gift_extents; |
8482 | if (bluefs && | |
8483 | after_flush - bluefs_last_balance > | |
8484 | cct->_conf->bluestore_bluefs_balance_interval) { | |
8485 | bluefs_last_balance = after_flush; | |
8486 | int r = _balance_bluefs_freespace(&bluefs_gift_extents); | |
8487 | assert(r >= 0); | |
8488 | if (r > 0) { | |
8489 | for (auto& p : bluefs_gift_extents) { | |
8490 | bluefs_extents.insert(p.offset, p.length); | |
8491 | } | |
8492 | bufferlist bl; | |
8493 | ::encode(bluefs_extents, bl); | |
8494 | dout(10) << __func__ << " bluefs_extents now 0x" << std::hex | |
8495 | << bluefs_extents << std::dec << dendl; | |
8496 | synct->set(PREFIX_SUPER, "bluefs_extents", bl); | |
8497 | } | |
8498 | } | |
8499 | ||
8500 | // cleanup sync deferred keys | |
8501 | for (auto b : deferred_stable) { | |
8502 | for (auto& txc : b->txcs) { | |
8503 | bluestore_deferred_transaction_t& wt = *txc.deferred_txn; | |
8504 | if (!wt.released.empty()) { | |
8505 | // kraken replay compat only | |
8506 | txc.released = wt.released; | |
8507 | dout(10) << __func__ << " deferred txn has released " | |
8508 | << txc.released | |
8509 | << " (we just upgraded from kraken) on " << &txc << dendl; | |
8510 | _txc_finalize_kv(&txc, synct); | |
8511 | } | |
8512 | // cleanup the deferred | |
8513 | string key; | |
8514 | get_deferred_key(wt.seq, &key); | |
8515 | synct->rm_single_key(PREFIX_DEFERRED, key); | |
8516 | } | |
8517 | } | |
8518 | ||
8519 | // submit synct synchronously (block and wait for it to commit) | |
31f18b77 | 8520 | int r = cct->_conf->bluestore_debug_omit_kv_commit ? 0 : db->submit_transaction_sync(synct); |
7c673cae FG |
8521 | assert(r == 0); |
8522 | ||
8523 | if (new_nid_max) { | |
8524 | nid_max = new_nid_max; | |
8525 | dout(10) << __func__ << " nid_max now " << nid_max << dendl; | |
8526 | } | |
8527 | if (new_blobid_max) { | |
8528 | blobid_max = new_blobid_max; | |
8529 | dout(10) << __func__ << " blobid_max now " << blobid_max << dendl; | |
8530 | } | |
8531 | ||
224ce89b WB |
8532 | { |
8533 | utime_t finish = ceph_clock_now(); | |
8534 | utime_t dur_flush = after_flush - start; | |
8535 | utime_t dur_kv = finish - after_flush; | |
8536 | utime_t dur = finish - start; | |
8537 | dout(20) << __func__ << " committed " << kv_committing.size() | |
8538 | << " cleaned " << deferred_stable.size() | |
8539 | << " in " << dur | |
8540 | << " (" << dur_flush << " flush + " << dur_kv << " kv commit)" | |
8541 | << dendl; | |
7c673cae FG |
8542 | logger->tinc(l_bluestore_kv_flush_lat, dur_flush); |
8543 | logger->tinc(l_bluestore_kv_commit_lat, dur_kv); | |
8544 | logger->tinc(l_bluestore_kv_lat, dur); | |
8545 | } | |
31f18b77 FG |
8546 | |
8547 | if (bluefs) { | |
8548 | if (!bluefs_gift_extents.empty()) { | |
8549 | _commit_bluefs_freespace(bluefs_gift_extents); | |
8550 | } | |
8551 | for (auto p = bluefs_extents_reclaiming.begin(); | |
8552 | p != bluefs_extents_reclaiming.end(); | |
8553 | ++p) { | |
8554 | dout(20) << __func__ << " releasing old bluefs 0x" << std::hex | |
8555 | << p.get_start() << "~" << p.get_len() << std::dec | |
8556 | << dendl; | |
8557 | alloc->release(p.get_start(), p.get_len()); | |
8558 | } | |
8559 | bluefs_extents_reclaiming.clear(); | |
8560 | } | |
8561 | ||
8562 | { | |
8563 | std::unique_lock<std::mutex> m(kv_finalize_lock); | |
8564 | if (kv_committing_to_finalize.empty()) { | |
8565 | kv_committing_to_finalize.swap(kv_committing); | |
8566 | } else { | |
8567 | kv_committing_to_finalize.insert( | |
8568 | kv_committing_to_finalize.end(), | |
8569 | kv_committing.begin(), | |
8570 | kv_committing.end()); | |
8571 | kv_committing.clear(); | |
8572 | } | |
8573 | if (deferred_stable_to_finalize.empty()) { | |
8574 | deferred_stable_to_finalize.swap(deferred_stable); | |
8575 | } else { | |
8576 | deferred_stable_to_finalize.insert( | |
8577 | deferred_stable_to_finalize.end(), | |
8578 | deferred_stable.begin(), | |
8579 | deferred_stable.end()); | |
8580 | deferred_stable.clear(); | |
8581 | } | |
8582 | kv_finalize_cond.notify_one(); | |
8583 | } | |
8584 | ||
8585 | l.lock(); | |
8586 | // previously deferred "done" are now "stable" by virtue of this | |
8587 | // commit cycle. | |
8588 | deferred_stable_queue.swap(deferred_done); | |
8589 | } | |
8590 | } | |
8591 | dout(10) << __func__ << " finish" << dendl; | |
8592 | kv_sync_started = false; | |
8593 | } | |
8594 | ||
8595 | void BlueStore::_kv_finalize_thread() | |
8596 | { | |
8597 | deque<TransContext*> kv_committed; | |
8598 | deque<DeferredBatch*> deferred_stable; | |
8599 | dout(10) << __func__ << " start" << dendl; | |
8600 | std::unique_lock<std::mutex> l(kv_finalize_lock); | |
8601 | assert(!kv_finalize_started); | |
8602 | kv_finalize_started = true; | |
8603 | kv_finalize_cond.notify_all(); | |
8604 | while (true) { | |
8605 | assert(kv_committed.empty()); | |
8606 | assert(deferred_stable.empty()); | |
8607 | if (kv_committing_to_finalize.empty() && | |
8608 | deferred_stable_to_finalize.empty()) { | |
8609 | if (kv_finalize_stop) | |
8610 | break; | |
8611 | dout(20) << __func__ << " sleep" << dendl; | |
8612 | kv_finalize_cond.wait(l); | |
8613 | dout(20) << __func__ << " wake" << dendl; | |
8614 | } else { | |
8615 | kv_committed.swap(kv_committing_to_finalize); | |
8616 | deferred_stable.swap(deferred_stable_to_finalize); | |
8617 | l.unlock(); | |
8618 | dout(20) << __func__ << " kv_committed " << kv_committed << dendl; | |
8619 | dout(20) << __func__ << " deferred_stable " << deferred_stable << dendl; | |
8620 | ||
8621 | while (!kv_committed.empty()) { | |
8622 | TransContext *txc = kv_committed.front(); | |
7c673cae FG |
8623 | assert(txc->state == TransContext::STATE_KV_SUBMITTED); |
8624 | _txc_state_proc(txc); | |
31f18b77 | 8625 | kv_committed.pop_front(); |
7c673cae | 8626 | } |
31f18b77 | 8627 | |
7c673cae FG |
8628 | for (auto b : deferred_stable) { |
8629 | auto p = b->txcs.begin(); | |
8630 | while (p != b->txcs.end()) { | |
8631 | TransContext *txc = &*p; | |
8632 | p = b->txcs.erase(p); // unlink here because | |
8633 | _txc_state_proc(txc); // this may destroy txc | |
8634 | } | |
8635 | delete b; | |
8636 | } | |
31f18b77 | 8637 | deferred_stable.clear(); |
7c673cae FG |
8638 | |
8639 | if (!deferred_aggressive) { | |
31f18b77 | 8640 | if (deferred_queue_size >= deferred_batch_ops.load() || |
7c673cae | 8641 | throttle_deferred_bytes.past_midpoint()) { |
224ce89b | 8642 | deferred_try_submit(); |
7c673cae FG |
8643 | } |
8644 | } | |
8645 | ||
8646 | // this is as good a place as any ... | |
8647 | _reap_collections(); | |
8648 | ||
7c673cae | 8649 | l.lock(); |
7c673cae FG |
8650 | } |
8651 | } | |
8652 | dout(10) << __func__ << " finish" << dendl; | |
31f18b77 | 8653 | kv_finalize_started = false; |
7c673cae FG |
8654 | } |
8655 | ||
8656 | bluestore_deferred_op_t *BlueStore::_get_deferred_op( | |
8657 | TransContext *txc, OnodeRef o) | |
8658 | { | |
8659 | if (!txc->deferred_txn) { | |
8660 | txc->deferred_txn = new bluestore_deferred_transaction_t; | |
8661 | } | |
8662 | txc->deferred_txn->ops.push_back(bluestore_deferred_op_t()); | |
8663 | return &txc->deferred_txn->ops.back(); | |
8664 | } | |
8665 | ||
8666 | void BlueStore::_deferred_queue(TransContext *txc) | |
8667 | { | |
8668 | dout(20) << __func__ << " txc " << txc << " osr " << txc->osr << dendl; | |
224ce89b | 8669 | deferred_lock.lock(); |
7c673cae FG |
8670 | if (!txc->osr->deferred_pending && |
8671 | !txc->osr->deferred_running) { | |
8672 | deferred_queue.push_back(*txc->osr); | |
8673 | } | |
8674 | if (!txc->osr->deferred_pending) { | |
8675 | txc->osr->deferred_pending = new DeferredBatch(cct, txc->osr.get()); | |
8676 | } | |
8677 | ++deferred_queue_size; | |
8678 | txc->osr->deferred_pending->txcs.push_back(*txc); | |
8679 | bluestore_deferred_transaction_t& wt = *txc->deferred_txn; | |
8680 | for (auto opi = wt.ops.begin(); opi != wt.ops.end(); ++opi) { | |
8681 | const auto& op = *opi; | |
8682 | assert(op.op == bluestore_deferred_op_t::OP_WRITE); | |
8683 | bufferlist::const_iterator p = op.data.begin(); | |
8684 | for (auto e : op.extents) { | |
8685 | txc->osr->deferred_pending->prepare_write( | |
8686 | cct, wt.seq, e.offset, e.length, p); | |
8687 | } | |
8688 | } | |
8689 | if (deferred_aggressive && | |
8690 | !txc->osr->deferred_running) { | |
224ce89b WB |
8691 | _deferred_submit_unlock(txc->osr.get()); |
8692 | } else { | |
8693 | deferred_lock.unlock(); | |
7c673cae FG |
8694 | } |
8695 | } | |
8696 | ||
224ce89b | 8697 | void BlueStore::deferred_try_submit() |
7c673cae FG |
8698 | { |
8699 | dout(20) << __func__ << " " << deferred_queue.size() << " osrs, " | |
8700 | << deferred_queue_size << " txcs" << dendl; | |
224ce89b WB |
8701 | std::lock_guard<std::mutex> l(deferred_lock); |
8702 | vector<OpSequencerRef> osrs; | |
8703 | osrs.reserve(deferred_queue.size()); | |
7c673cae | 8704 | for (auto& osr : deferred_queue) { |
224ce89b WB |
8705 | osrs.push_back(&osr); |
8706 | } | |
8707 | for (auto& osr : osrs) { | |
181888fb FG |
8708 | if (osr->deferred_pending) { |
8709 | if (!osr->deferred_running) { | |
8710 | _deferred_submit_unlock(osr.get()); | |
8711 | deferred_lock.lock(); | |
8712 | } else { | |
8713 | dout(20) << __func__ << " osr " << osr << " already has running" | |
8714 | << dendl; | |
8715 | } | |
8716 | } else { | |
8717 | dout(20) << __func__ << " osr " << osr << " has no pending" << dendl; | |
7c673cae FG |
8718 | } |
8719 | } | |
8720 | } | |
8721 | ||
224ce89b | 8722 | void BlueStore::_deferred_submit_unlock(OpSequencer *osr) |
7c673cae FG |
8723 | { |
8724 | dout(10) << __func__ << " osr " << osr | |
8725 | << " " << osr->deferred_pending->iomap.size() << " ios pending " | |
8726 | << dendl; | |
8727 | assert(osr->deferred_pending); | |
8728 | assert(!osr->deferred_running); | |
8729 | ||
8730 | auto b = osr->deferred_pending; | |
8731 | deferred_queue_size -= b->seq_bytes.size(); | |
8732 | assert(deferred_queue_size >= 0); | |
8733 | ||
8734 | osr->deferred_running = osr->deferred_pending; | |
8735 | osr->deferred_pending = nullptr; | |
8736 | ||
8737 | uint64_t start = 0, pos = 0; | |
8738 | bufferlist bl; | |
8739 | auto i = b->iomap.begin(); | |
8740 | while (true) { | |
8741 | if (i == b->iomap.end() || i->first != pos) { | |
8742 | if (bl.length()) { | |
8743 | dout(20) << __func__ << " write 0x" << std::hex | |
8744 | << start << "~" << bl.length() | |
8745 | << " crc " << bl.crc32c(-1) << std::dec << dendl; | |
8746 | if (!g_conf->bluestore_debug_omit_block_device_write) { | |
8747 | logger->inc(l_bluestore_deferred_write_ops); | |
8748 | logger->inc(l_bluestore_deferred_write_bytes, bl.length()); | |
8749 | int r = bdev->aio_write(start, bl, &b->ioc, false); | |
8750 | assert(r == 0); | |
8751 | } | |
8752 | } | |
8753 | if (i == b->iomap.end()) { | |
8754 | break; | |
8755 | } | |
8756 | start = 0; | |
8757 | pos = i->first; | |
8758 | bl.clear(); | |
8759 | } | |
8760 | dout(20) << __func__ << " seq " << i->second.seq << " 0x" | |
8761 | << std::hex << pos << "~" << i->second.bl.length() << std::dec | |
8762 | << dendl; | |
8763 | if (!bl.length()) { | |
8764 | start = pos; | |
8765 | } | |
8766 | pos += i->second.bl.length(); | |
8767 | bl.claim_append(i->second.bl); | |
8768 | ++i; | |
8769 | } | |
224ce89b | 8770 | |
224ce89b | 8771 | deferred_lock.unlock(); |
7c673cae FG |
8772 | bdev->aio_submit(&b->ioc); |
8773 | } | |
8774 | ||
8775 | void BlueStore::_deferred_aio_finish(OpSequencer *osr) | |
8776 | { | |
8777 | dout(10) << __func__ << " osr " << osr << dendl; | |
8778 | assert(osr->deferred_running); | |
8779 | DeferredBatch *b = osr->deferred_running; | |
8780 | ||
8781 | { | |
8782 | std::lock_guard<std::mutex> l(deferred_lock); | |
8783 | assert(osr->deferred_running == b); | |
8784 | osr->deferred_running = nullptr; | |
8785 | if (!osr->deferred_pending) { | |
181888fb | 8786 | dout(20) << __func__ << " dequeueing" << dendl; |
7c673cae FG |
8787 | auto q = deferred_queue.iterator_to(*osr); |
8788 | deferred_queue.erase(q); | |
8789 | } else if (deferred_aggressive) { | |
224ce89b | 8790 | dout(20) << __func__ << " queuing async deferred_try_submit" << dendl; |
181888fb | 8791 | deferred_finisher.queue(new FunctionContext([&](int) { |
224ce89b WB |
8792 | deferred_try_submit(); |
8793 | })); | |
181888fb FG |
8794 | } else { |
8795 | dout(20) << __func__ << " leaving queued, more pending" << dendl; | |
7c673cae FG |
8796 | } |
8797 | } | |
8798 | ||
8799 | { | |
31f18b77 | 8800 | uint64_t costs = 0; |
7c673cae FG |
8801 | std::lock_guard<std::mutex> l2(osr->qlock); |
8802 | for (auto& i : b->txcs) { | |
8803 | TransContext *txc = &i; | |
8804 | txc->state = TransContext::STATE_DEFERRED_CLEANUP; | |
31f18b77 | 8805 | costs += txc->cost; |
7c673cae | 8806 | } |
31f18b77 FG |
8807 | osr->qcond.notify_all(); |
8808 | throttle_deferred_bytes.put(costs); | |
7c673cae FG |
8809 | std::lock_guard<std::mutex> l(kv_lock); |
8810 | deferred_done_queue.emplace_back(b); | |
8811 | } | |
8812 | ||
8813 | // in the normal case, do not bother waking up the kv thread; it will | |
8814 | // catch us on the next commit anyway. | |
8815 | if (deferred_aggressive) { | |
8816 | std::lock_guard<std::mutex> l(kv_lock); | |
8817 | kv_cond.notify_one(); | |
8818 | } | |
8819 | } | |
8820 | ||
8821 | int BlueStore::_deferred_replay() | |
8822 | { | |
8823 | dout(10) << __func__ << " start" << dendl; | |
8824 | OpSequencerRef osr = new OpSequencer(cct, this); | |
8825 | int count = 0; | |
8826 | int r = 0; | |
8827 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_DEFERRED); | |
8828 | for (it->lower_bound(string()); it->valid(); it->next(), ++count) { | |
8829 | dout(20) << __func__ << " replay " << pretty_binary_string(it->key()) | |
8830 | << dendl; | |
8831 | bluestore_deferred_transaction_t *deferred_txn = | |
8832 | new bluestore_deferred_transaction_t; | |
8833 | bufferlist bl = it->value(); | |
8834 | bufferlist::iterator p = bl.begin(); | |
8835 | try { | |
8836 | ::decode(*deferred_txn, p); | |
8837 | } catch (buffer::error& e) { | |
8838 | derr << __func__ << " failed to decode deferred txn " | |
8839 | << pretty_binary_string(it->key()) << dendl; | |
8840 | delete deferred_txn; | |
8841 | r = -EIO; | |
8842 | goto out; | |
8843 | } | |
8844 | TransContext *txc = _txc_create(osr.get()); | |
8845 | txc->deferred_txn = deferred_txn; | |
8846 | txc->state = TransContext::STATE_KV_DONE; | |
8847 | _txc_state_proc(txc); | |
8848 | } | |
8849 | out: | |
8850 | dout(20) << __func__ << " draining osr" << dendl; | |
8851 | _osr_drain_all(); | |
8852 | osr->discard(); | |
8853 | dout(10) << __func__ << " completed " << count << " events" << dendl; | |
8854 | return r; | |
8855 | } | |
8856 | ||
8857 | // --------------------------- | |
8858 | // transactions | |
8859 | ||
8860 | int BlueStore::queue_transactions( | |
8861 | Sequencer *posr, | |
8862 | vector<Transaction>& tls, | |
8863 | TrackedOpRef op, | |
8864 | ThreadPool::TPHandle *handle) | |
8865 | { | |
8866 | FUNCTRACE(); | |
8867 | Context *onreadable; | |
8868 | Context *ondisk; | |
8869 | Context *onreadable_sync; | |
8870 | ObjectStore::Transaction::collect_contexts( | |
8871 | tls, &onreadable, &ondisk, &onreadable_sync); | |
8872 | ||
8873 | if (cct->_conf->objectstore_blackhole) { | |
8874 | dout(0) << __func__ << " objectstore_blackhole = TRUE, dropping transaction" | |
8875 | << dendl; | |
8876 | delete ondisk; | |
8877 | delete onreadable; | |
8878 | delete onreadable_sync; | |
8879 | return 0; | |
8880 | } | |
8881 | utime_t start = ceph_clock_now(); | |
8882 | // set up the sequencer | |
8883 | OpSequencer *osr; | |
8884 | assert(posr); | |
8885 | if (posr->p) { | |
8886 | osr = static_cast<OpSequencer *>(posr->p.get()); | |
8887 | dout(10) << __func__ << " existing " << osr << " " << *osr << dendl; | |
8888 | } else { | |
8889 | osr = new OpSequencer(cct, this); | |
8890 | osr->parent = posr; | |
8891 | posr->p = osr; | |
8892 | dout(10) << __func__ << " new " << osr << " " << *osr << dendl; | |
8893 | } | |
8894 | ||
8895 | // prepare | |
8896 | TransContext *txc = _txc_create(osr); | |
8897 | txc->onreadable = onreadable; | |
8898 | txc->onreadable_sync = onreadable_sync; | |
8899 | txc->oncommit = ondisk; | |
8900 | ||
8901 | for (vector<Transaction>::iterator p = tls.begin(); p != tls.end(); ++p) { | |
8902 | (*p).set_osr(osr); | |
8903 | txc->bytes += (*p).get_num_bytes(); | |
8904 | _txc_add_transaction(txc, &(*p)); | |
8905 | } | |
8906 | _txc_calc_cost(txc); | |
8907 | ||
8908 | _txc_write_nodes(txc, txc->t); | |
8909 | ||
8910 | // journal deferred items | |
8911 | if (txc->deferred_txn) { | |
8912 | txc->deferred_txn->seq = ++deferred_seq; | |
8913 | bufferlist bl; | |
8914 | ::encode(*txc->deferred_txn, bl); | |
8915 | string key; | |
8916 | get_deferred_key(txc->deferred_txn->seq, &key); | |
8917 | txc->t->set(PREFIX_DEFERRED, key, bl); | |
8918 | } | |
8919 | ||
8920 | _txc_finalize_kv(txc, txc->t); | |
8921 | if (handle) | |
8922 | handle->suspend_tp_timeout(); | |
8923 | ||
8924 | utime_t tstart = ceph_clock_now(); | |
8925 | throttle_bytes.get(txc->cost); | |
8926 | if (txc->deferred_txn) { | |
8927 | // ensure we do not block here because of deferred writes | |
8928 | if (!throttle_deferred_bytes.get_or_fail(txc->cost)) { | |
d2e6a577 FG |
8929 | dout(10) << __func__ << " failed get throttle_deferred_bytes, aggressive" |
8930 | << dendl; | |
8931 | ++deferred_aggressive; | |
7c673cae FG |
8932 | deferred_try_submit(); |
8933 | throttle_deferred_bytes.get(txc->cost); | |
d2e6a577 FG |
8934 | --deferred_aggressive; |
8935 | } | |
7c673cae FG |
8936 | } |
8937 | utime_t tend = ceph_clock_now(); | |
8938 | ||
8939 | if (handle) | |
8940 | handle->reset_tp_timeout(); | |
8941 | ||
8942 | logger->inc(l_bluestore_txc); | |
8943 | ||
8944 | // execute (start) | |
8945 | _txc_state_proc(txc); | |
8946 | ||
8947 | logger->tinc(l_bluestore_submit_lat, ceph_clock_now() - start); | |
8948 | logger->tinc(l_bluestore_throttle_lat, tend - tstart); | |
8949 | return 0; | |
8950 | } | |
8951 | ||
8952 | void BlueStore::_txc_aio_submit(TransContext *txc) | |
8953 | { | |
8954 | dout(10) << __func__ << " txc " << txc << dendl; | |
8955 | bdev->aio_submit(&txc->ioc); | |
8956 | } | |
8957 | ||
8958 | void BlueStore::_txc_add_transaction(TransContext *txc, Transaction *t) | |
8959 | { | |
8960 | Transaction::iterator i = t->begin(); | |
8961 | ||
8962 | _dump_transaction(t); | |
8963 | ||
8964 | vector<CollectionRef> cvec(i.colls.size()); | |
8965 | unsigned j = 0; | |
8966 | for (vector<coll_t>::iterator p = i.colls.begin(); p != i.colls.end(); | |
8967 | ++p, ++j) { | |
8968 | cvec[j] = _get_collection(*p); | |
7c673cae FG |
8969 | } |
8970 | vector<OnodeRef> ovec(i.objects.size()); | |
8971 | ||
8972 | for (int pos = 0; i.have_op(); ++pos) { | |
8973 | Transaction::Op *op = i.decode_op(); | |
8974 | int r = 0; | |
8975 | ||
8976 | // no coll or obj | |
8977 | if (op->op == Transaction::OP_NOP) | |
8978 | continue; | |
8979 | ||
8980 | // collection operations | |
8981 | CollectionRef &c = cvec[op->cid]; | |
8982 | switch (op->op) { | |
8983 | case Transaction::OP_RMCOLL: | |
8984 | { | |
8985 | const coll_t &cid = i.get_cid(op->cid); | |
8986 | r = _remove_collection(txc, cid, &c); | |
8987 | if (!r) | |
8988 | continue; | |
8989 | } | |
8990 | break; | |
8991 | ||
8992 | case Transaction::OP_MKCOLL: | |
8993 | { | |
8994 | assert(!c); | |
8995 | const coll_t &cid = i.get_cid(op->cid); | |
8996 | r = _create_collection(txc, cid, op->split_bits, &c); | |
8997 | if (!r) | |
8998 | continue; | |
8999 | } | |
9000 | break; | |
9001 | ||
9002 | case Transaction::OP_SPLIT_COLLECTION: | |
9003 | assert(0 == "deprecated"); | |
9004 | break; | |
9005 | ||
9006 | case Transaction::OP_SPLIT_COLLECTION2: | |
9007 | { | |
9008 | uint32_t bits = op->split_bits; | |
9009 | uint32_t rem = op->split_rem; | |
9010 | r = _split_collection(txc, c, cvec[op->dest_cid], bits, rem); | |
9011 | if (!r) | |
9012 | continue; | |
9013 | } | |
9014 | break; | |
9015 | ||
9016 | case Transaction::OP_COLL_HINT: | |
9017 | { | |
9018 | uint32_t type = op->hint_type; | |
9019 | bufferlist hint; | |
9020 | i.decode_bl(hint); | |
9021 | bufferlist::iterator hiter = hint.begin(); | |
9022 | if (type == Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS) { | |
9023 | uint32_t pg_num; | |
9024 | uint64_t num_objs; | |
9025 | ::decode(pg_num, hiter); | |
9026 | ::decode(num_objs, hiter); | |
9027 | dout(10) << __func__ << " collection hint objects is a no-op, " | |
9028 | << " pg_num " << pg_num << " num_objects " << num_objs | |
9029 | << dendl; | |
9030 | } else { | |
9031 | // Ignore the hint | |
9032 | dout(10) << __func__ << " unknown collection hint " << type << dendl; | |
9033 | } | |
9034 | continue; | |
9035 | } | |
9036 | break; | |
9037 | ||
9038 | case Transaction::OP_COLL_SETATTR: | |
9039 | r = -EOPNOTSUPP; | |
9040 | break; | |
9041 | ||
9042 | case Transaction::OP_COLL_RMATTR: | |
9043 | r = -EOPNOTSUPP; | |
9044 | break; | |
9045 | ||
9046 | case Transaction::OP_COLL_RENAME: | |
9047 | assert(0 == "not implemented"); | |
9048 | break; | |
9049 | } | |
9050 | if (r < 0) { | |
9051 | derr << __func__ << " error " << cpp_strerror(r) | |
9052 | << " not handled on operation " << op->op | |
9053 | << " (op " << pos << ", counting from 0)" << dendl; | |
9054 | _dump_transaction(t, 0); | |
9055 | assert(0 == "unexpected error"); | |
9056 | } | |
9057 | ||
9058 | // these operations implicity create the object | |
9059 | bool create = false; | |
9060 | if (op->op == Transaction::OP_TOUCH || | |
9061 | op->op == Transaction::OP_WRITE || | |
9062 | op->op == Transaction::OP_ZERO) { | |
9063 | create = true; | |
9064 | } | |
9065 | ||
9066 | // object operations | |
9067 | RWLock::WLocker l(c->lock); | |
9068 | OnodeRef &o = ovec[op->oid]; | |
9069 | if (!o) { | |
9070 | ghobject_t oid = i.get_oid(op->oid); | |
9071 | o = c->get_onode(oid, create); | |
9072 | } | |
9073 | if (!create && (!o || !o->exists)) { | |
9074 | dout(10) << __func__ << " op " << op->op << " got ENOENT on " | |
9075 | << i.get_oid(op->oid) << dendl; | |
9076 | r = -ENOENT; | |
9077 | goto endop; | |
9078 | } | |
9079 | ||
9080 | switch (op->op) { | |
9081 | case Transaction::OP_TOUCH: | |
9082 | r = _touch(txc, c, o); | |
9083 | break; | |
9084 | ||
9085 | case Transaction::OP_WRITE: | |
9086 | { | |
9087 | uint64_t off = op->off; | |
9088 | uint64_t len = op->len; | |
9089 | uint32_t fadvise_flags = i.get_fadvise_flags(); | |
9090 | bufferlist bl; | |
9091 | i.decode_bl(bl); | |
9092 | r = _write(txc, c, o, off, len, bl, fadvise_flags); | |
9093 | } | |
9094 | break; | |
9095 | ||
9096 | case Transaction::OP_ZERO: | |
9097 | { | |
9098 | uint64_t off = op->off; | |
9099 | uint64_t len = op->len; | |
9100 | r = _zero(txc, c, o, off, len); | |
9101 | } | |
9102 | break; | |
9103 | ||
9104 | case Transaction::OP_TRIMCACHE: | |
9105 | { | |
9106 | // deprecated, no-op | |
9107 | } | |
9108 | break; | |
9109 | ||
9110 | case Transaction::OP_TRUNCATE: | |
9111 | { | |
9112 | uint64_t off = op->off; | |
35e4c445 | 9113 | r = _truncate(txc, c, o, off); |
7c673cae FG |
9114 | } |
9115 | break; | |
9116 | ||
9117 | case Transaction::OP_REMOVE: | |
9118 | { | |
9119 | r = _remove(txc, c, o); | |
9120 | } | |
9121 | break; | |
9122 | ||
9123 | case Transaction::OP_SETATTR: | |
9124 | { | |
9125 | string name = i.decode_string(); | |
9126 | bufferptr bp; | |
9127 | i.decode_bp(bp); | |
9128 | r = _setattr(txc, c, o, name, bp); | |
9129 | } | |
9130 | break; | |
9131 | ||
9132 | case Transaction::OP_SETATTRS: | |
9133 | { | |
9134 | map<string, bufferptr> aset; | |
9135 | i.decode_attrset(aset); | |
9136 | r = _setattrs(txc, c, o, aset); | |
9137 | } | |
9138 | break; | |
9139 | ||
9140 | case Transaction::OP_RMATTR: | |
9141 | { | |
9142 | string name = i.decode_string(); | |
9143 | r = _rmattr(txc, c, o, name); | |
9144 | } | |
9145 | break; | |
9146 | ||
9147 | case Transaction::OP_RMATTRS: | |
9148 | { | |
9149 | r = _rmattrs(txc, c, o); | |
9150 | } | |
9151 | break; | |
9152 | ||
9153 | case Transaction::OP_CLONE: | |
9154 | { | |
9155 | OnodeRef& no = ovec[op->dest_oid]; | |
9156 | if (!no) { | |
9157 | const ghobject_t& noid = i.get_oid(op->dest_oid); | |
9158 | no = c->get_onode(noid, true); | |
9159 | } | |
9160 | r = _clone(txc, c, o, no); | |
9161 | } | |
9162 | break; | |
9163 | ||
9164 | case Transaction::OP_CLONERANGE: | |
9165 | assert(0 == "deprecated"); | |
9166 | break; | |
9167 | ||
9168 | case Transaction::OP_CLONERANGE2: | |
9169 | { | |
9170 | OnodeRef& no = ovec[op->dest_oid]; | |
9171 | if (!no) { | |
9172 | const ghobject_t& noid = i.get_oid(op->dest_oid); | |
9173 | no = c->get_onode(noid, true); | |
9174 | } | |
9175 | uint64_t srcoff = op->off; | |
9176 | uint64_t len = op->len; | |
9177 | uint64_t dstoff = op->dest_off; | |
9178 | r = _clone_range(txc, c, o, no, srcoff, len, dstoff); | |
9179 | } | |
9180 | break; | |
9181 | ||
9182 | case Transaction::OP_COLL_ADD: | |
9183 | assert(0 == "not implemented"); | |
9184 | break; | |
9185 | ||
9186 | case Transaction::OP_COLL_REMOVE: | |
9187 | assert(0 == "not implemented"); | |
9188 | break; | |
9189 | ||
9190 | case Transaction::OP_COLL_MOVE: | |
9191 | assert(0 == "deprecated"); | |
9192 | break; | |
9193 | ||
9194 | case Transaction::OP_COLL_MOVE_RENAME: | |
9195 | case Transaction::OP_TRY_RENAME: | |
9196 | { | |
9197 | assert(op->cid == op->dest_cid); | |
9198 | const ghobject_t& noid = i.get_oid(op->dest_oid); | |
9199 | OnodeRef& no = ovec[op->dest_oid]; | |
9200 | if (!no) { | |
9201 | no = c->get_onode(noid, false); | |
9202 | } | |
9203 | r = _rename(txc, c, o, no, noid); | |
9204 | } | |
9205 | break; | |
9206 | ||
9207 | case Transaction::OP_OMAP_CLEAR: | |
9208 | { | |
9209 | r = _omap_clear(txc, c, o); | |
9210 | } | |
9211 | break; | |
9212 | case Transaction::OP_OMAP_SETKEYS: | |
9213 | { | |
9214 | bufferlist aset_bl; | |
9215 | i.decode_attrset_bl(&aset_bl); | |
9216 | r = _omap_setkeys(txc, c, o, aset_bl); | |
9217 | } | |
9218 | break; | |
9219 | case Transaction::OP_OMAP_RMKEYS: | |
9220 | { | |
9221 | bufferlist keys_bl; | |
9222 | i.decode_keyset_bl(&keys_bl); | |
9223 | r = _omap_rmkeys(txc, c, o, keys_bl); | |
9224 | } | |
9225 | break; | |
9226 | case Transaction::OP_OMAP_RMKEYRANGE: | |
9227 | { | |
9228 | string first, last; | |
9229 | first = i.decode_string(); | |
9230 | last = i.decode_string(); | |
9231 | r = _omap_rmkey_range(txc, c, o, first, last); | |
9232 | } | |
9233 | break; | |
9234 | case Transaction::OP_OMAP_SETHEADER: | |
9235 | { | |
9236 | bufferlist bl; | |
9237 | i.decode_bl(bl); | |
9238 | r = _omap_setheader(txc, c, o, bl); | |
9239 | } | |
9240 | break; | |
9241 | ||
9242 | case Transaction::OP_SETALLOCHINT: | |
9243 | { | |
9244 | r = _set_alloc_hint(txc, c, o, | |
9245 | op->expected_object_size, | |
9246 | op->expected_write_size, | |
9247 | op->alloc_hint_flags); | |
9248 | } | |
9249 | break; | |
9250 | ||
9251 | default: | |
9252 | derr << __func__ << "bad op " << op->op << dendl; | |
9253 | ceph_abort(); | |
9254 | } | |
9255 | ||
9256 | endop: | |
9257 | if (r < 0) { | |
9258 | bool ok = false; | |
9259 | ||
9260 | if (r == -ENOENT && !(op->op == Transaction::OP_CLONERANGE || | |
9261 | op->op == Transaction::OP_CLONE || | |
9262 | op->op == Transaction::OP_CLONERANGE2 || | |
9263 | op->op == Transaction::OP_COLL_ADD || | |
9264 | op->op == Transaction::OP_SETATTR || | |
9265 | op->op == Transaction::OP_SETATTRS || | |
9266 | op->op == Transaction::OP_RMATTR || | |
9267 | op->op == Transaction::OP_OMAP_SETKEYS || | |
9268 | op->op == Transaction::OP_OMAP_RMKEYS || | |
9269 | op->op == Transaction::OP_OMAP_RMKEYRANGE || | |
9270 | op->op == Transaction::OP_OMAP_SETHEADER)) | |
9271 | // -ENOENT is usually okay | |
9272 | ok = true; | |
9273 | if (r == -ENODATA) | |
9274 | ok = true; | |
9275 | ||
9276 | if (!ok) { | |
9277 | const char *msg = "unexpected error code"; | |
9278 | ||
9279 | if (r == -ENOENT && (op->op == Transaction::OP_CLONERANGE || | |
9280 | op->op == Transaction::OP_CLONE || | |
9281 | op->op == Transaction::OP_CLONERANGE2)) | |
9282 | msg = "ENOENT on clone suggests osd bug"; | |
9283 | ||
9284 | if (r == -ENOSPC) | |
9285 | // For now, if we hit _any_ ENOSPC, crash, before we do any damage | |
9286 | // by partially applying transactions. | |
9287 | msg = "ENOSPC from bluestore, misconfigured cluster"; | |
9288 | ||
9289 | if (r == -ENOTEMPTY) { | |
9290 | msg = "ENOTEMPTY suggests garbage data in osd data dir"; | |
9291 | } | |
9292 | ||
9293 | derr << __func__ << " error " << cpp_strerror(r) | |
9294 | << " not handled on operation " << op->op | |
9295 | << " (op " << pos << ", counting from 0)" | |
9296 | << dendl; | |
9297 | derr << msg << dendl; | |
9298 | _dump_transaction(t, 0); | |
9299 | assert(0 == "unexpected error"); | |
9300 | } | |
9301 | } | |
9302 | } | |
9303 | } | |
9304 | ||
9305 | ||
9306 | ||
9307 | // ----------------- | |
9308 | // write operations | |
9309 | ||
9310 | int BlueStore::_touch(TransContext *txc, | |
9311 | CollectionRef& c, | |
9312 | OnodeRef &o) | |
9313 | { | |
9314 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
9315 | int r = 0; | |
7c673cae FG |
9316 | _assign_nid(txc, o); |
9317 | txc->write_onode(o); | |
9318 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
9319 | return r; | |
9320 | } | |
9321 | ||
9322 | void BlueStore::_dump_onode(OnodeRef o, int log_level) | |
9323 | { | |
9324 | if (!cct->_conf->subsys.should_gather(ceph_subsys_bluestore, log_level)) | |
9325 | return; | |
9326 | dout(log_level) << __func__ << " " << o << " " << o->oid | |
9327 | << " nid " << o->onode.nid | |
9328 | << " size 0x" << std::hex << o->onode.size | |
9329 | << " (" << std::dec << o->onode.size << ")" | |
9330 | << " expected_object_size " << o->onode.expected_object_size | |
9331 | << " expected_write_size " << o->onode.expected_write_size | |
9332 | << " in " << o->onode.extent_map_shards.size() << " shards" | |
9333 | << ", " << o->extent_map.spanning_blob_map.size() | |
9334 | << " spanning blobs" | |
9335 | << dendl; | |
9336 | for (auto p = o->onode.attrs.begin(); | |
9337 | p != o->onode.attrs.end(); | |
9338 | ++p) { | |
9339 | dout(log_level) << __func__ << " attr " << p->first | |
9340 | << " len " << p->second.length() << dendl; | |
9341 | } | |
9342 | _dump_extent_map(o->extent_map, log_level); | |
9343 | } | |
9344 | ||
9345 | void BlueStore::_dump_extent_map(ExtentMap &em, int log_level) | |
9346 | { | |
9347 | uint64_t pos = 0; | |
9348 | for (auto& s : em.shards) { | |
9349 | dout(log_level) << __func__ << " shard " << *s.shard_info | |
9350 | << (s.loaded ? " (loaded)" : "") | |
9351 | << (s.dirty ? " (dirty)" : "") | |
9352 | << dendl; | |
9353 | } | |
9354 | for (auto& e : em.extent_map) { | |
9355 | dout(log_level) << __func__ << " " << e << dendl; | |
9356 | assert(e.logical_offset >= pos); | |
9357 | pos = e.logical_offset + e.length; | |
9358 | const bluestore_blob_t& blob = e.blob->get_blob(); | |
9359 | if (blob.has_csum()) { | |
9360 | vector<uint64_t> v; | |
9361 | unsigned n = blob.get_csum_count(); | |
9362 | for (unsigned i = 0; i < n; ++i) | |
9363 | v.push_back(blob.get_csum_item(i)); | |
9364 | dout(log_level) << __func__ << " csum: " << std::hex << v << std::dec | |
9365 | << dendl; | |
9366 | } | |
9367 | std::lock_guard<std::recursive_mutex> l(e.blob->shared_blob->get_cache()->lock); | |
9368 | for (auto& i : e.blob->shared_blob->bc.buffer_map) { | |
9369 | dout(log_level) << __func__ << " 0x" << std::hex << i.first | |
9370 | << "~" << i.second->length << std::dec | |
9371 | << " " << *i.second << dendl; | |
9372 | } | |
9373 | } | |
9374 | } | |
9375 | ||
9376 | void BlueStore::_dump_transaction(Transaction *t, int log_level) | |
9377 | { | |
9378 | dout(log_level) << " transaction dump:\n"; | |
9379 | JSONFormatter f(true); | |
9380 | f.open_object_section("transaction"); | |
9381 | t->dump(&f); | |
9382 | f.close_section(); | |
9383 | f.flush(*_dout); | |
9384 | *_dout << dendl; | |
9385 | } | |
9386 | ||
9387 | void BlueStore::_pad_zeros( | |
9388 | bufferlist *bl, uint64_t *offset, | |
9389 | uint64_t chunk_size) | |
9390 | { | |
9391 | auto length = bl->length(); | |
9392 | dout(30) << __func__ << " 0x" << std::hex << *offset << "~" << length | |
9393 | << " chunk_size 0x" << chunk_size << std::dec << dendl; | |
9394 | dout(40) << "before:\n"; | |
9395 | bl->hexdump(*_dout); | |
9396 | *_dout << dendl; | |
9397 | // front | |
9398 | size_t front_pad = *offset % chunk_size; | |
9399 | size_t back_pad = 0; | |
9400 | size_t pad_count = 0; | |
9401 | if (front_pad) { | |
9402 | size_t front_copy = MIN(chunk_size - front_pad, length); | |
9403 | bufferptr z = buffer::create_page_aligned(chunk_size); | |
224ce89b | 9404 | z.zero(0, front_pad, false); |
7c673cae | 9405 | pad_count += front_pad; |
224ce89b | 9406 | bl->copy(0, front_copy, z.c_str() + front_pad); |
7c673cae FG |
9407 | if (front_copy + front_pad < chunk_size) { |
9408 | back_pad = chunk_size - (length + front_pad); | |
224ce89b | 9409 | z.zero(front_pad + length, back_pad, false); |
7c673cae FG |
9410 | pad_count += back_pad; |
9411 | } | |
9412 | bufferlist old, t; | |
9413 | old.swap(*bl); | |
9414 | t.substr_of(old, front_copy, length - front_copy); | |
9415 | bl->append(z); | |
9416 | bl->claim_append(t); | |
9417 | *offset -= front_pad; | |
224ce89b | 9418 | length += pad_count; |
7c673cae FG |
9419 | } |
9420 | ||
9421 | // back | |
9422 | uint64_t end = *offset + length; | |
9423 | unsigned back_copy = end % chunk_size; | |
9424 | if (back_copy) { | |
9425 | assert(back_pad == 0); | |
9426 | back_pad = chunk_size - back_copy; | |
9427 | assert(back_copy <= length); | |
9428 | bufferptr tail(chunk_size); | |
224ce89b WB |
9429 | bl->copy(length - back_copy, back_copy, tail.c_str()); |
9430 | tail.zero(back_copy, back_pad, false); | |
7c673cae FG |
9431 | bufferlist old; |
9432 | old.swap(*bl); | |
9433 | bl->substr_of(old, 0, length - back_copy); | |
9434 | bl->append(tail); | |
9435 | length += back_pad; | |
9436 | pad_count += back_pad; | |
9437 | } | |
9438 | dout(20) << __func__ << " pad 0x" << std::hex << front_pad << " + 0x" | |
9439 | << back_pad << " on front/back, now 0x" << *offset << "~" | |
9440 | << length << std::dec << dendl; | |
9441 | dout(40) << "after:\n"; | |
9442 | bl->hexdump(*_dout); | |
9443 | *_dout << dendl; | |
9444 | if (pad_count) | |
9445 | logger->inc(l_bluestore_write_pad_bytes, pad_count); | |
9446 | assert(bl->length() == length); | |
9447 | } | |
9448 | ||
9449 | void BlueStore::_do_write_small( | |
9450 | TransContext *txc, | |
9451 | CollectionRef &c, | |
9452 | OnodeRef o, | |
9453 | uint64_t offset, uint64_t length, | |
9454 | bufferlist::iterator& blp, | |
9455 | WriteContext *wctx) | |
9456 | { | |
9457 | dout(10) << __func__ << " 0x" << std::hex << offset << "~" << length | |
9458 | << std::dec << dendl; | |
9459 | assert(length < min_alloc_size); | |
9460 | uint64_t end_offs = offset + length; | |
9461 | ||
9462 | logger->inc(l_bluestore_write_small); | |
9463 | logger->inc(l_bluestore_write_small_bytes, length); | |
9464 | ||
9465 | bufferlist bl; | |
9466 | blp.copy(length, bl); | |
9467 | ||
9468 | // Look for an existing mutable blob we can use. | |
9469 | auto begin = o->extent_map.extent_map.begin(); | |
9470 | auto end = o->extent_map.extent_map.end(); | |
9471 | auto ep = o->extent_map.seek_lextent(offset); | |
9472 | if (ep != begin) { | |
9473 | --ep; | |
9474 | if (ep->blob_end() <= offset) { | |
9475 | ++ep; | |
9476 | } | |
9477 | } | |
9478 | auto prev_ep = ep; | |
9479 | if (prev_ep != begin) { | |
9480 | --prev_ep; | |
9481 | } else { | |
9482 | prev_ep = end; // to avoid this extent check as it's a duplicate | |
9483 | } | |
9484 | ||
9485 | auto max_bsize = MAX(wctx->target_blob_size, min_alloc_size); | |
9486 | auto min_off = offset >= max_bsize ? offset - max_bsize : 0; | |
9487 | uint32_t alloc_len = min_alloc_size; | |
9488 | auto offset0 = P2ALIGN(offset, alloc_len); | |
9489 | ||
9490 | bool any_change; | |
9491 | ||
9492 | // search suitable extent in both forward and reverse direction in | |
9493 | // [offset - target_max_blob_size, offset + target_max_blob_size] range | |
224ce89b | 9494 | // then check if blob can be reused via can_reuse_blob func or apply |
7c673cae FG |
9495 | // direct/deferred write (the latter for extents including or higher |
9496 | // than 'offset' only). | |
9497 | do { | |
9498 | any_change = false; | |
9499 | ||
9500 | if (ep != end && ep->logical_offset < offset + max_bsize) { | |
9501 | BlobRef b = ep->blob; | |
9502 | auto bstart = ep->blob_start(); | |
9503 | dout(20) << __func__ << " considering " << *b | |
9504 | << " bstart 0x" << std::hex << bstart << std::dec << dendl; | |
9505 | if (bstart >= end_offs) { | |
9506 | dout(20) << __func__ << " ignoring distant " << *b << dendl; | |
9507 | } else if (!b->get_blob().is_mutable()) { | |
9508 | dout(20) << __func__ << " ignoring immutable " << *b << dendl; | |
9509 | } else if (ep->logical_offset % min_alloc_size != | |
9510 | ep->blob_offset % min_alloc_size) { | |
9511 | dout(20) << __func__ << " ignoring offset-skewed " << *b << dendl; | |
9512 | } else { | |
9513 | uint64_t chunk_size = b->get_blob().get_chunk_size(block_size); | |
9514 | // can we pad our head/tail out with zeros? | |
9515 | uint64_t head_pad, tail_pad; | |
9516 | head_pad = P2PHASE(offset, chunk_size); | |
9517 | tail_pad = P2NPHASE(end_offs, chunk_size); | |
9518 | if (head_pad || tail_pad) { | |
9519 | o->extent_map.fault_range(db, offset - head_pad, | |
9520 | end_offs - offset + head_pad + tail_pad); | |
9521 | } | |
9522 | if (head_pad && | |
9523 | o->extent_map.has_any_lextents(offset - head_pad, chunk_size)) { | |
9524 | head_pad = 0; | |
9525 | } | |
9526 | if (tail_pad && o->extent_map.has_any_lextents(end_offs, tail_pad)) { | |
9527 | tail_pad = 0; | |
9528 | } | |
9529 | ||
9530 | uint64_t b_off = offset - head_pad - bstart; | |
9531 | uint64_t b_len = length + head_pad + tail_pad; | |
9532 | ||
9533 | // direct write into unused blocks of an existing mutable blob? | |
9534 | if ((b_off % chunk_size == 0 && b_len % chunk_size == 0) && | |
9535 | b->get_blob().get_ondisk_length() >= b_off + b_len && | |
9536 | b->get_blob().is_unused(b_off, b_len) && | |
9537 | b->get_blob().is_allocated(b_off, b_len)) { | |
224ce89b | 9538 | _apply_padding(head_pad, tail_pad, bl); |
7c673cae FG |
9539 | |
9540 | dout(20) << __func__ << " write to unused 0x" << std::hex | |
9541 | << b_off << "~" << b_len | |
9542 | << " pad 0x" << head_pad << " + 0x" << tail_pad | |
9543 | << std::dec << " of mutable " << *b << dendl; | |
224ce89b | 9544 | _buffer_cache_write(txc, b, b_off, bl, |
7c673cae FG |
9545 | wctx->buffered ? 0 : Buffer::FLAG_NOCACHE); |
9546 | ||
9547 | if (!g_conf->bluestore_debug_omit_block_device_write) { | |
9548 | if (b_len <= prefer_deferred_size) { | |
9549 | dout(20) << __func__ << " deferring small 0x" << std::hex | |
9550 | << b_len << std::dec << " unused write via deferred" << dendl; | |
9551 | bluestore_deferred_op_t *op = _get_deferred_op(txc, o); | |
9552 | op->op = bluestore_deferred_op_t::OP_WRITE; | |
9553 | b->get_blob().map( | |
9554 | b_off, b_len, | |
9555 | [&](uint64_t offset, uint64_t length) { | |
9556 | op->extents.emplace_back(bluestore_pextent_t(offset, length)); | |
9557 | return 0; | |
9558 | }); | |
224ce89b | 9559 | op->data = bl; |
7c673cae FG |
9560 | } else { |
9561 | b->get_blob().map_bl( | |
224ce89b | 9562 | b_off, bl, |
7c673cae FG |
9563 | [&](uint64_t offset, bufferlist& t) { |
9564 | bdev->aio_write(offset, t, | |
9565 | &txc->ioc, wctx->buffered); | |
9566 | }); | |
9567 | } | |
9568 | } | |
224ce89b | 9569 | b->dirty_blob().calc_csum(b_off, bl); |
7c673cae FG |
9570 | dout(20) << __func__ << " lex old " << *ep << dendl; |
9571 | Extent *le = o->extent_map.set_lextent(c, offset, b_off + head_pad, length, | |
9572 | b, | |
9573 | &wctx->old_extents); | |
9574 | b->dirty_blob().mark_used(le->blob_offset, le->length); | |
9575 | txc->statfs_delta.stored() += le->length; | |
9576 | dout(20) << __func__ << " lex " << *le << dendl; | |
9577 | logger->inc(l_bluestore_write_small_unused); | |
9578 | return; | |
9579 | } | |
9580 | // read some data to fill out the chunk? | |
9581 | uint64_t head_read = P2PHASE(b_off, chunk_size); | |
9582 | uint64_t tail_read = P2NPHASE(b_off + b_len, chunk_size); | |
9583 | if ((head_read || tail_read) && | |
9584 | (b->get_blob().get_ondisk_length() >= b_off + b_len + tail_read) && | |
9585 | head_read + tail_read < min_alloc_size) { | |
9586 | b_off -= head_read; | |
9587 | b_len += head_read + tail_read; | |
9588 | ||
9589 | } else { | |
9590 | head_read = tail_read = 0; | |
9591 | } | |
9592 | ||
9593 | // chunk-aligned deferred overwrite? | |
9594 | if (b->get_blob().get_ondisk_length() >= b_off + b_len && | |
9595 | b_off % chunk_size == 0 && | |
9596 | b_len % chunk_size == 0 && | |
9597 | b->get_blob().is_allocated(b_off, b_len)) { | |
9598 | ||
224ce89b | 9599 | _apply_padding(head_pad, tail_pad, bl); |
7c673cae FG |
9600 | |
9601 | dout(20) << __func__ << " reading head 0x" << std::hex << head_read | |
9602 | << " and tail 0x" << tail_read << std::dec << dendl; | |
9603 | if (head_read) { | |
9604 | bufferlist head_bl; | |
9605 | int r = _do_read(c.get(), o, offset - head_pad - head_read, head_read, | |
9606 | head_bl, 0); | |
9607 | assert(r >= 0 && r <= (int)head_read); | |
9608 | size_t zlen = head_read - r; | |
9609 | if (zlen) { | |
9610 | head_bl.append_zero(zlen); | |
9611 | logger->inc(l_bluestore_write_pad_bytes, zlen); | |
9612 | } | |
224ce89b | 9613 | bl.claim_prepend(head_bl); |
7c673cae FG |
9614 | logger->inc(l_bluestore_write_penalty_read_ops); |
9615 | } | |
9616 | if (tail_read) { | |
9617 | bufferlist tail_bl; | |
9618 | int r = _do_read(c.get(), o, offset + length + tail_pad, tail_read, | |
9619 | tail_bl, 0); | |
9620 | assert(r >= 0 && r <= (int)tail_read); | |
9621 | size_t zlen = tail_read - r; | |
9622 | if (zlen) { | |
9623 | tail_bl.append_zero(zlen); | |
9624 | logger->inc(l_bluestore_write_pad_bytes, zlen); | |
9625 | } | |
224ce89b | 9626 | bl.claim_append(tail_bl); |
7c673cae FG |
9627 | logger->inc(l_bluestore_write_penalty_read_ops); |
9628 | } | |
9629 | logger->inc(l_bluestore_write_small_pre_read); | |
9630 | ||
9631 | bluestore_deferred_op_t *op = _get_deferred_op(txc, o); | |
9632 | op->op = bluestore_deferred_op_t::OP_WRITE; | |
224ce89b | 9633 | _buffer_cache_write(txc, b, b_off, bl, |
7c673cae FG |
9634 | wctx->buffered ? 0 : Buffer::FLAG_NOCACHE); |
9635 | ||
9636 | int r = b->get_blob().map( | |
9637 | b_off, b_len, | |
9638 | [&](uint64_t offset, uint64_t length) { | |
9639 | op->extents.emplace_back(bluestore_pextent_t(offset, length)); | |
9640 | return 0; | |
9641 | }); | |
9642 | assert(r == 0); | |
9643 | if (b->get_blob().csum_type) { | |
224ce89b | 9644 | b->dirty_blob().calc_csum(b_off, bl); |
7c673cae | 9645 | } |
224ce89b | 9646 | op->data.claim(bl); |
7c673cae FG |
9647 | dout(20) << __func__ << " deferred write 0x" << std::hex << b_off << "~" |
9648 | << b_len << std::dec << " of mutable " << *b | |
9649 | << " at " << op->extents << dendl; | |
9650 | Extent *le = o->extent_map.set_lextent(c, offset, offset - bstart, length, | |
9651 | b, &wctx->old_extents); | |
9652 | b->dirty_blob().mark_used(le->blob_offset, le->length); | |
9653 | txc->statfs_delta.stored() += le->length; | |
9654 | dout(20) << __func__ << " lex " << *le << dendl; | |
9655 | logger->inc(l_bluestore_write_small_deferred); | |
9656 | return; | |
9657 | } | |
224ce89b WB |
9658 | // try to reuse blob if we can |
9659 | if (b->can_reuse_blob(min_alloc_size, | |
7c673cae FG |
9660 | max_bsize, |
9661 | offset0 - bstart, | |
9662 | &alloc_len)) { | |
9663 | assert(alloc_len == min_alloc_size); // expecting data always | |
9664 | // fit into reused blob | |
9665 | // Need to check for pending writes desiring to | |
9666 | // reuse the same pextent. The rationale is that during GC two chunks | |
9667 | // from garbage blobs(compressed?) can share logical space within the same | |
9668 | // AU. That's in turn might be caused by unaligned len in clone_range2. | |
9669 | // Hence the second write will fail in an attempt to reuse blob at | |
9670 | // do_alloc_write(). | |
9671 | if (!wctx->has_conflict(b, | |
9672 | offset0, | |
9673 | offset0 + alloc_len, | |
9674 | min_alloc_size)) { | |
9675 | ||
9676 | // we can't reuse pad_head/pad_tail since they might be truncated | |
9677 | // due to existent extents | |
9678 | uint64_t b_off = offset - bstart; | |
9679 | uint64_t b_off0 = b_off; | |
9680 | _pad_zeros(&bl, &b_off0, chunk_size); | |
9681 | ||
9682 | dout(20) << __func__ << " reuse blob " << *b << std::hex | |
224ce89b WB |
9683 | << " (0x" << b_off0 << "~" << bl.length() << ")" |
9684 | << " (0x" << b_off << "~" << length << ")" | |
7c673cae FG |
9685 | << std::dec << dendl; |
9686 | ||
9687 | o->extent_map.punch_hole(c, offset, length, &wctx->old_extents); | |
9688 | wctx->write(offset, b, alloc_len, b_off0, bl, b_off, length, | |
9689 | false, false); | |
9690 | logger->inc(l_bluestore_write_small_unused); | |
9691 | return; | |
9692 | } | |
9693 | } | |
9694 | } | |
9695 | ++ep; | |
9696 | any_change = true; | |
9697 | } // if (ep != end && ep->logical_offset < offset + max_bsize) | |
9698 | ||
9699 | // check extent for reuse in reverse order | |
9700 | if (prev_ep != end && prev_ep->logical_offset >= min_off) { | |
9701 | BlobRef b = prev_ep->blob; | |
9702 | auto bstart = prev_ep->blob_start(); | |
9703 | dout(20) << __func__ << " considering " << *b | |
9704 | << " bstart 0x" << std::hex << bstart << std::dec << dendl; | |
224ce89b | 9705 | if (b->can_reuse_blob(min_alloc_size, |
7c673cae FG |
9706 | max_bsize, |
9707 | offset0 - bstart, | |
9708 | &alloc_len)) { | |
9709 | assert(alloc_len == min_alloc_size); // expecting data always | |
9710 | // fit into reused blob | |
9711 | // Need to check for pending writes desiring to | |
9712 | // reuse the same pextent. The rationale is that during GC two chunks | |
9713 | // from garbage blobs(compressed?) can share logical space within the same | |
9714 | // AU. That's in turn might be caused by unaligned len in clone_range2. | |
9715 | // Hence the second write will fail in an attempt to reuse blob at | |
9716 | // do_alloc_write(). | |
9717 | if (!wctx->has_conflict(b, | |
9718 | offset0, | |
9719 | offset0 + alloc_len, | |
9720 | min_alloc_size)) { | |
9721 | ||
9722 | uint64_t chunk_size = b->get_blob().get_chunk_size(block_size); | |
9723 | uint64_t b_off = offset - bstart; | |
9724 | uint64_t b_off0 = b_off; | |
9725 | _pad_zeros(&bl, &b_off0, chunk_size); | |
9726 | ||
9727 | dout(20) << __func__ << " reuse blob " << *b << std::hex | |
224ce89b WB |
9728 | << " (0x" << b_off0 << "~" << bl.length() << ")" |
9729 | << " (0x" << b_off << "~" << length << ")" | |
7c673cae FG |
9730 | << std::dec << dendl; |
9731 | ||
9732 | o->extent_map.punch_hole(c, offset, length, &wctx->old_extents); | |
9733 | wctx->write(offset, b, alloc_len, b_off0, bl, b_off, length, | |
9734 | false, false); | |
9735 | logger->inc(l_bluestore_write_small_unused); | |
9736 | return; | |
9737 | } | |
9738 | } | |
9739 | if (prev_ep != begin) { | |
9740 | --prev_ep; | |
9741 | any_change = true; | |
9742 | } else { | |
9743 | prev_ep = end; // to avoid useless first extent re-check | |
9744 | } | |
9745 | } // if (prev_ep != end && prev_ep->logical_offset >= min_off) | |
9746 | } while (any_change); | |
9747 | ||
9748 | // new blob. | |
9749 | ||
9750 | BlobRef b = c->new_blob(); | |
9751 | uint64_t b_off = P2PHASE(offset, alloc_len); | |
9752 | uint64_t b_off0 = b_off; | |
9753 | _pad_zeros(&bl, &b_off0, block_size); | |
9754 | o->extent_map.punch_hole(c, offset, length, &wctx->old_extents); | |
9755 | wctx->write(offset, b, alloc_len, b_off0, bl, b_off, length, true, true); | |
9756 | logger->inc(l_bluestore_write_small_new); | |
9757 | ||
9758 | return; | |
9759 | } | |
9760 | ||
9761 | void BlueStore::_do_write_big( | |
9762 | TransContext *txc, | |
9763 | CollectionRef &c, | |
9764 | OnodeRef o, | |
9765 | uint64_t offset, uint64_t length, | |
9766 | bufferlist::iterator& blp, | |
9767 | WriteContext *wctx) | |
9768 | { | |
9769 | dout(10) << __func__ << " 0x" << std::hex << offset << "~" << length | |
9770 | << " target_blob_size 0x" << wctx->target_blob_size << std::dec | |
9771 | << " compress " << (int)wctx->compress | |
9772 | << dendl; | |
9773 | logger->inc(l_bluestore_write_big); | |
9774 | logger->inc(l_bluestore_write_big_bytes, length); | |
9775 | o->extent_map.punch_hole(c, offset, length, &wctx->old_extents); | |
9776 | auto max_bsize = MAX(wctx->target_blob_size, min_alloc_size); | |
9777 | while (length > 0) { | |
9778 | bool new_blob = false; | |
9779 | uint32_t l = MIN(max_bsize, length); | |
9780 | BlobRef b; | |
9781 | uint32_t b_off = 0; | |
9782 | ||
9783 | //attempting to reuse existing blob | |
9784 | if (!wctx->compress) { | |
9785 | // look for an existing mutable blob we can reuse | |
9786 | auto begin = o->extent_map.extent_map.begin(); | |
9787 | auto end = o->extent_map.extent_map.end(); | |
9788 | auto ep = o->extent_map.seek_lextent(offset); | |
9789 | auto prev_ep = ep; | |
9790 | if (prev_ep != begin) { | |
9791 | --prev_ep; | |
9792 | } else { | |
9793 | prev_ep = end; // to avoid this extent check as it's a duplicate | |
9794 | } | |
9795 | auto min_off = offset >= max_bsize ? offset - max_bsize : 0; | |
9796 | // search suitable extent in both forward and reverse direction in | |
9797 | // [offset - target_max_blob_size, offset + target_max_blob_size] range | |
224ce89b | 9798 | // then check if blob can be reused via can_reuse_blob func. |
7c673cae FG |
9799 | bool any_change; |
9800 | do { | |
9801 | any_change = false; | |
9802 | if (ep != end && ep->logical_offset < offset + max_bsize) { | |
9803 | if (offset >= ep->blob_start() && | |
224ce89b | 9804 | ep->blob->can_reuse_blob(min_alloc_size, max_bsize, |
7c673cae FG |
9805 | offset - ep->blob_start(), |
9806 | &l)) { | |
9807 | b = ep->blob; | |
9808 | b_off = offset - ep->blob_start(); | |
9809 | prev_ep = end; // to avoid check below | |
9810 | dout(20) << __func__ << " reuse blob " << *b << std::hex | |
224ce89b | 9811 | << " (0x" << b_off << "~" << l << ")" << std::dec << dendl; |
7c673cae FG |
9812 | } else { |
9813 | ++ep; | |
9814 | any_change = true; | |
9815 | } | |
9816 | } | |
9817 | ||
9818 | if (prev_ep != end && prev_ep->logical_offset >= min_off) { | |
224ce89b | 9819 | if (prev_ep->blob->can_reuse_blob(min_alloc_size, max_bsize, |
7c673cae FG |
9820 | offset - prev_ep->blob_start(), |
9821 | &l)) { | |
9822 | b = prev_ep->blob; | |
9823 | b_off = offset - prev_ep->blob_start(); | |
9824 | dout(20) << __func__ << " reuse blob " << *b << std::hex | |
224ce89b | 9825 | << " (0x" << b_off << "~" << l << ")" << std::dec << dendl; |
7c673cae FG |
9826 | } else if (prev_ep != begin) { |
9827 | --prev_ep; | |
9828 | any_change = true; | |
9829 | } else { | |
9830 | prev_ep = end; // to avoid useless first extent re-check | |
9831 | } | |
9832 | } | |
9833 | } while (b == nullptr && any_change); | |
9834 | } | |
9835 | if (b == nullptr) { | |
9836 | b = c->new_blob(); | |
9837 | b_off = 0; | |
9838 | new_blob = true; | |
9839 | } | |
9840 | ||
9841 | bufferlist t; | |
9842 | blp.copy(l, t); | |
9843 | wctx->write(offset, b, l, b_off, t, b_off, l, false, new_blob); | |
9844 | offset += l; | |
9845 | length -= l; | |
9846 | logger->inc(l_bluestore_write_big_blobs); | |
9847 | } | |
9848 | } | |
9849 | ||
9850 | int BlueStore::_do_alloc_write( | |
9851 | TransContext *txc, | |
9852 | CollectionRef coll, | |
9853 | OnodeRef o, | |
9854 | WriteContext *wctx) | |
9855 | { | |
9856 | dout(20) << __func__ << " txc " << txc | |
9857 | << " " << wctx->writes.size() << " blobs" | |
9858 | << dendl; | |
9859 | ||
9860 | uint64_t need = 0; | |
9861 | auto max_bsize = MAX(wctx->target_blob_size, min_alloc_size); | |
9862 | for (auto &wi : wctx->writes) { | |
9863 | need += wi.blob_length; | |
9864 | } | |
9865 | int r = alloc->reserve(need); | |
9866 | if (r < 0) { | |
9867 | derr << __func__ << " failed to reserve 0x" << std::hex << need << std::dec | |
9868 | << dendl; | |
9869 | return r; | |
9870 | } | |
9871 | ||
9872 | uint64_t hint = 0; | |
9873 | CompressorRef c; | |
9874 | double crr = 0; | |
9875 | if (wctx->compress) { | |
9876 | c = select_option( | |
9877 | "compression_algorithm", | |
9878 | compressor, | |
9879 | [&]() { | |
9880 | string val; | |
9881 | if (coll->pool_opts.get(pool_opts_t::COMPRESSION_ALGORITHM, &val)) { | |
9882 | CompressorRef cp = compressor; | |
9883 | if (!cp || cp->get_type_name() != val) { | |
9884 | cp = Compressor::create(cct, val); | |
9885 | } | |
9886 | return boost::optional<CompressorRef>(cp); | |
9887 | } | |
9888 | return boost::optional<CompressorRef>(); | |
9889 | } | |
9890 | ); | |
9891 | ||
9892 | crr = select_option( | |
9893 | "compression_required_ratio", | |
9894 | cct->_conf->bluestore_compression_required_ratio, | |
9895 | [&]() { | |
9896 | double val; | |
9897 | if(coll->pool_opts.get(pool_opts_t::COMPRESSION_REQUIRED_RATIO, &val)) { | |
9898 | return boost::optional<double>(val); | |
9899 | } | |
9900 | return boost::optional<double>(); | |
9901 | } | |
9902 | ); | |
9903 | } | |
9904 | ||
9905 | // checksum | |
9906 | int csum = csum_type.load(); | |
9907 | csum = select_option( | |
9908 | "csum_type", | |
9909 | csum, | |
9910 | [&]() { | |
9911 | int val; | |
9912 | if(coll->pool_opts.get(pool_opts_t::CSUM_TYPE, &val)) { | |
9913 | return boost::optional<int>(val); | |
9914 | } | |
9915 | return boost::optional<int>(); | |
9916 | } | |
9917 | ); | |
9918 | ||
9919 | for (auto& wi : wctx->writes) { | |
9920 | BlobRef b = wi.b; | |
9921 | bluestore_blob_t& dblob = b->dirty_blob(); | |
9922 | uint64_t b_off = wi.b_off; | |
9923 | bufferlist *l = &wi.bl; | |
9924 | uint64_t final_length = wi.blob_length; | |
9925 | uint64_t csum_length = wi.blob_length; | |
9926 | unsigned csum_order = block_size_order; | |
9927 | bufferlist compressed_bl; | |
9928 | bool compressed = false; | |
9929 | if(c && wi.blob_length > min_alloc_size) { | |
9930 | ||
9931 | utime_t start = ceph_clock_now(); | |
9932 | ||
9933 | // compress | |
9934 | assert(b_off == 0); | |
9935 | assert(wi.blob_length == l->length()); | |
9936 | bluestore_compression_header_t chdr; | |
9937 | chdr.type = c->get_type(); | |
9938 | // FIXME: memory alignment here is bad | |
9939 | bufferlist t; | |
9940 | ||
9941 | r = c->compress(*l, t); | |
9942 | assert(r == 0); | |
9943 | ||
9944 | chdr.length = t.length(); | |
9945 | ::encode(chdr, compressed_bl); | |
9946 | compressed_bl.claim_append(t); | |
9947 | uint64_t rawlen = compressed_bl.length(); | |
9948 | uint64_t newlen = P2ROUNDUP(rawlen, min_alloc_size); | |
9949 | uint64_t want_len_raw = final_length * crr; | |
9950 | uint64_t want_len = P2ROUNDUP(want_len_raw, min_alloc_size); | |
9951 | if (newlen <= want_len && newlen < final_length) { | |
9952 | // Cool. We compressed at least as much as we were hoping to. | |
9953 | // pad out to min_alloc_size | |
9954 | compressed_bl.append_zero(newlen - rawlen); | |
9955 | logger->inc(l_bluestore_write_pad_bytes, newlen - rawlen); | |
9956 | dout(20) << __func__ << std::hex << " compressed 0x" << wi.blob_length | |
9957 | << " -> 0x" << rawlen << " => 0x" << newlen | |
9958 | << " with " << c->get_type() | |
9959 | << std::dec << dendl; | |
9960 | txc->statfs_delta.compressed() += rawlen; | |
9961 | txc->statfs_delta.compressed_original() += l->length(); | |
9962 | txc->statfs_delta.compressed_allocated() += newlen; | |
9963 | l = &compressed_bl; | |
9964 | final_length = newlen; | |
9965 | csum_length = newlen; | |
9966 | csum_order = ctz(newlen); | |
9967 | dblob.set_compressed(wi.blob_length, rawlen); | |
9968 | compressed = true; | |
9969 | logger->inc(l_bluestore_compress_success_count); | |
9970 | } else { | |
9971 | dout(20) << __func__ << std::hex << " 0x" << l->length() | |
9972 | << " compressed to 0x" << rawlen << " -> 0x" << newlen | |
9973 | << " with " << c->get_type() | |
9974 | << ", which is more than required 0x" << want_len_raw | |
9975 | << " -> 0x" << want_len | |
9976 | << ", leaving uncompressed" | |
9977 | << std::dec << dendl; | |
9978 | logger->inc(l_bluestore_compress_rejected_count); | |
9979 | } | |
9980 | logger->tinc(l_bluestore_compress_lat, | |
9981 | ceph_clock_now() - start); | |
9982 | } | |
9983 | if (!compressed && wi.new_blob) { | |
9984 | // initialize newly created blob only | |
31f18b77 | 9985 | assert(dblob.is_mutable()); |
7c673cae FG |
9986 | if (l->length() != wi.blob_length) { |
9987 | // hrm, maybe we could do better here, but let's not bother. | |
9988 | dout(20) << __func__ << " forcing csum_order to block_size_order " | |
9989 | << block_size_order << dendl; | |
31f18b77 | 9990 | csum_order = block_size_order; |
7c673cae FG |
9991 | } else { |
9992 | csum_order = std::min(wctx->csum_order, ctz(l->length())); | |
9993 | } | |
9994 | // try to align blob with max_blob_size to improve | |
9995 | // its reuse ratio, e.g. in case of reverse write | |
9996 | uint32_t suggested_boff = | |
9997 | (wi.logical_offset - (wi.b_off0 - wi.b_off)) % max_bsize; | |
9998 | if ((suggested_boff % (1 << csum_order)) == 0 && | |
9999 | suggested_boff + final_length <= max_bsize && | |
10000 | suggested_boff > b_off) { | |
181888fb | 10001 | dout(20) << __func__ << " forcing blob_offset to 0x" |
7c673cae FG |
10002 | << std::hex << suggested_boff << std::dec << dendl; |
10003 | assert(suggested_boff >= b_off); | |
10004 | csum_length += suggested_boff - b_off; | |
10005 | b_off = suggested_boff; | |
10006 | } | |
181888fb FG |
10007 | if (csum != Checksummer::CSUM_NONE) { |
10008 | dout(20) << __func__ << " initialize csum setting for new blob " << *b | |
10009 | << " csum_type " << Checksummer::get_csum_type_string(csum) | |
10010 | << " csum_order " << csum_order | |
10011 | << " csum_length 0x" << std::hex << csum_length << std::dec | |
10012 | << dendl; | |
10013 | dblob.init_csum(csum, csum_order, csum_length); | |
10014 | } | |
7c673cae FG |
10015 | } |
10016 | ||
10017 | AllocExtentVector extents; | |
10018 | extents.reserve(4); // 4 should be (more than) enough for most allocations | |
10019 | int64_t got = alloc->allocate(final_length, min_alloc_size, | |
10020 | max_alloc_size.load(), | |
10021 | hint, &extents); | |
10022 | assert(got == (int64_t)final_length); | |
10023 | need -= got; | |
10024 | txc->statfs_delta.allocated() += got; | |
10025 | for (auto& p : extents) { | |
10026 | bluestore_pextent_t e = bluestore_pextent_t(p); | |
10027 | txc->allocated.insert(e.offset, e.length); | |
10028 | hint = p.end(); | |
10029 | } | |
10030 | dblob.allocated(P2ALIGN(b_off, min_alloc_size), final_length, extents); | |
10031 | ||
181888fb FG |
10032 | dout(20) << __func__ << " blob " << *b << dendl; |
10033 | if (dblob.has_csum()) { | |
7c673cae FG |
10034 | dblob.calc_csum(b_off, *l); |
10035 | } | |
181888fb | 10036 | |
7c673cae FG |
10037 | if (wi.mark_unused) { |
10038 | auto b_end = b_off + wi.bl.length(); | |
10039 | if (b_off) { | |
10040 | dblob.add_unused(0, b_off); | |
10041 | } | |
10042 | if (b_end < wi.blob_length) { | |
10043 | dblob.add_unused(b_end, wi.blob_length - b_end); | |
10044 | } | |
10045 | } | |
10046 | ||
10047 | Extent *le = o->extent_map.set_lextent(coll, wi.logical_offset, | |
10048 | b_off + (wi.b_off0 - wi.b_off), | |
10049 | wi.length0, | |
10050 | wi.b, | |
10051 | nullptr); | |
10052 | wi.b->dirty_blob().mark_used(le->blob_offset, le->length); | |
10053 | txc->statfs_delta.stored() += le->length; | |
10054 | dout(20) << __func__ << " lex " << *le << dendl; | |
10055 | _buffer_cache_write(txc, wi.b, b_off, wi.bl, | |
10056 | wctx->buffered ? 0 : Buffer::FLAG_NOCACHE); | |
10057 | ||
10058 | // queue io | |
10059 | if (!g_conf->bluestore_debug_omit_block_device_write) { | |
10060 | if (l->length() <= prefer_deferred_size.load()) { | |
10061 | dout(20) << __func__ << " deferring small 0x" << std::hex | |
10062 | << l->length() << std::dec << " write via deferred" << dendl; | |
10063 | bluestore_deferred_op_t *op = _get_deferred_op(txc, o); | |
10064 | op->op = bluestore_deferred_op_t::OP_WRITE; | |
10065 | int r = b->get_blob().map( | |
10066 | b_off, l->length(), | |
10067 | [&](uint64_t offset, uint64_t length) { | |
10068 | op->extents.emplace_back(bluestore_pextent_t(offset, length)); | |
10069 | return 0; | |
10070 | }); | |
10071 | assert(r == 0); | |
10072 | op->data = *l; | |
10073 | } else { | |
10074 | b->get_blob().map_bl( | |
10075 | b_off, *l, | |
10076 | [&](uint64_t offset, bufferlist& t) { | |
10077 | bdev->aio_write(offset, t, &txc->ioc, false); | |
10078 | }); | |
10079 | } | |
10080 | } | |
10081 | } | |
10082 | if (need > 0) { | |
10083 | alloc->unreserve(need); | |
10084 | } | |
10085 | return 0; | |
10086 | } | |
10087 | ||
10088 | void BlueStore::_wctx_finish( | |
10089 | TransContext *txc, | |
10090 | CollectionRef& c, | |
10091 | OnodeRef o, | |
31f18b77 FG |
10092 | WriteContext *wctx, |
10093 | set<SharedBlob*> *maybe_unshared_blobs) | |
7c673cae FG |
10094 | { |
10095 | auto oep = wctx->old_extents.begin(); | |
10096 | while (oep != wctx->old_extents.end()) { | |
10097 | auto &lo = *oep; | |
10098 | oep = wctx->old_extents.erase(oep); | |
10099 | dout(20) << __func__ << " lex_old " << lo.e << dendl; | |
10100 | BlobRef b = lo.e.blob; | |
10101 | const bluestore_blob_t& blob = b->get_blob(); | |
10102 | if (blob.is_compressed()) { | |
10103 | if (lo.blob_empty) { | |
10104 | txc->statfs_delta.compressed() -= blob.get_compressed_payload_length(); | |
10105 | } | |
10106 | txc->statfs_delta.compressed_original() -= lo.e.length; | |
10107 | } | |
10108 | auto& r = lo.r; | |
10109 | txc->statfs_delta.stored() -= lo.e.length; | |
10110 | if (!r.empty()) { | |
10111 | dout(20) << __func__ << " blob release " << r << dendl; | |
10112 | if (blob.is_shared()) { | |
10113 | PExtentVector final; | |
10114 | c->load_shared_blob(b->shared_blob); | |
10115 | for (auto e : r) { | |
31f18b77 FG |
10116 | b->shared_blob->put_ref( |
10117 | e.offset, e.length, &final, | |
10118 | b->is_referenced() ? nullptr : maybe_unshared_blobs); | |
7c673cae FG |
10119 | } |
10120 | dout(20) << __func__ << " shared_blob release " << final | |
10121 | << " from " << *b->shared_blob << dendl; | |
10122 | txc->write_shared_blob(b->shared_blob); | |
10123 | r.clear(); | |
10124 | r.swap(final); | |
10125 | } | |
10126 | } | |
10127 | // we can't invalidate our logical extents as we drop them because | |
10128 | // other lextents (either in our onode or others) may still | |
10129 | // reference them. but we can throw out anything that is no | |
10130 | // longer allocated. Note that this will leave behind edge bits | |
10131 | // that are no longer referenced but not deallocated (until they | |
10132 | // age out of the cache naturally). | |
10133 | b->discard_unallocated(c.get()); | |
10134 | for (auto e : r) { | |
10135 | dout(20) << __func__ << " release " << e << dendl; | |
10136 | txc->released.insert(e.offset, e.length); | |
10137 | txc->statfs_delta.allocated() -= e.length; | |
10138 | if (blob.is_compressed()) { | |
10139 | txc->statfs_delta.compressed_allocated() -= e.length; | |
10140 | } | |
10141 | } | |
10142 | delete &lo; | |
10143 | if (b->is_spanning() && !b->is_referenced()) { | |
10144 | dout(20) << __func__ << " spanning_blob_map removing empty " << *b | |
10145 | << dendl; | |
10146 | o->extent_map.spanning_blob_map.erase(b->id); | |
10147 | } | |
10148 | } | |
10149 | } | |
10150 | ||
10151 | void BlueStore::_do_write_data( | |
10152 | TransContext *txc, | |
10153 | CollectionRef& c, | |
10154 | OnodeRef o, | |
10155 | uint64_t offset, | |
10156 | uint64_t length, | |
10157 | bufferlist& bl, | |
10158 | WriteContext *wctx) | |
10159 | { | |
10160 | uint64_t end = offset + length; | |
10161 | bufferlist::iterator p = bl.begin(); | |
10162 | ||
10163 | if (offset / min_alloc_size == (end - 1) / min_alloc_size && | |
10164 | (length != min_alloc_size)) { | |
10165 | // we fall within the same block | |
10166 | _do_write_small(txc, c, o, offset, length, p, wctx); | |
10167 | } else { | |
10168 | uint64_t head_offset, head_length; | |
10169 | uint64_t middle_offset, middle_length; | |
10170 | uint64_t tail_offset, tail_length; | |
10171 | ||
10172 | head_offset = offset; | |
10173 | head_length = P2NPHASE(offset, min_alloc_size); | |
10174 | ||
10175 | tail_offset = P2ALIGN(end, min_alloc_size); | |
10176 | tail_length = P2PHASE(end, min_alloc_size); | |
10177 | ||
10178 | middle_offset = head_offset + head_length; | |
10179 | middle_length = length - head_length - tail_length; | |
10180 | ||
10181 | if (head_length) { | |
10182 | _do_write_small(txc, c, o, head_offset, head_length, p, wctx); | |
10183 | } | |
10184 | ||
10185 | if (middle_length) { | |
10186 | _do_write_big(txc, c, o, middle_offset, middle_length, p, wctx); | |
10187 | } | |
10188 | ||
10189 | if (tail_length) { | |
10190 | _do_write_small(txc, c, o, tail_offset, tail_length, p, wctx); | |
10191 | } | |
10192 | } | |
10193 | } | |
10194 | ||
31f18b77 FG |
10195 | void BlueStore::_choose_write_options( |
10196 | CollectionRef& c, | |
10197 | OnodeRef o, | |
10198 | uint32_t fadvise_flags, | |
10199 | WriteContext *wctx) | |
7c673cae | 10200 | { |
7c673cae FG |
10201 | if (fadvise_flags & CEPH_OSD_OP_FLAG_FADVISE_WILLNEED) { |
10202 | dout(20) << __func__ << " will do buffered write" << dendl; | |
31f18b77 | 10203 | wctx->buffered = true; |
7c673cae FG |
10204 | } else if (cct->_conf->bluestore_default_buffered_write && |
10205 | (fadvise_flags & (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED | | |
10206 | CEPH_OSD_OP_FLAG_FADVISE_NOCACHE)) == 0) { | |
10207 | dout(20) << __func__ << " defaulting to buffered write" << dendl; | |
31f18b77 | 10208 | wctx->buffered = true; |
7c673cae FG |
10209 | } |
10210 | ||
31f18b77 FG |
10211 | // apply basic csum block size |
10212 | wctx->csum_order = block_size_order; | |
7c673cae FG |
10213 | |
10214 | // compression parameters | |
10215 | unsigned alloc_hints = o->onode.alloc_hint_flags; | |
10216 | auto cm = select_option( | |
10217 | "compression_mode", | |
31f18b77 | 10218 | comp_mode.load(), |
7c673cae FG |
10219 | [&]() { |
10220 | string val; | |
10221 | if(c->pool_opts.get(pool_opts_t::COMPRESSION_MODE, &val)) { | |
31f18b77 FG |
10222 | return boost::optional<Compressor::CompressionMode>( |
10223 | Compressor::get_comp_mode_type(val)); | |
7c673cae FG |
10224 | } |
10225 | return boost::optional<Compressor::CompressionMode>(); | |
10226 | } | |
10227 | ); | |
31f18b77 FG |
10228 | |
10229 | wctx->compress = (cm != Compressor::COMP_NONE) && | |
7c673cae FG |
10230 | ((cm == Compressor::COMP_FORCE) || |
10231 | (cm == Compressor::COMP_AGGRESSIVE && | |
10232 | (alloc_hints & CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE) == 0) || | |
10233 | (cm == Compressor::COMP_PASSIVE && | |
10234 | (alloc_hints & CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE))); | |
10235 | ||
10236 | if ((alloc_hints & CEPH_OSD_ALLOC_HINT_FLAG_SEQUENTIAL_READ) && | |
10237 | (alloc_hints & CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_READ) == 0 && | |
31f18b77 FG |
10238 | (alloc_hints & (CEPH_OSD_ALLOC_HINT_FLAG_IMMUTABLE | |
10239 | CEPH_OSD_ALLOC_HINT_FLAG_APPEND_ONLY)) && | |
7c673cae | 10240 | (alloc_hints & CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_WRITE) == 0) { |
31f18b77 | 10241 | |
7c673cae | 10242 | dout(20) << __func__ << " will prefer large blob and csum sizes" << dendl; |
31f18b77 | 10243 | |
7c673cae | 10244 | if (o->onode.expected_write_size) { |
224ce89b | 10245 | wctx->csum_order = std::max(min_alloc_size_order, |
31f18b77 | 10246 | (uint8_t)ctz(o->onode.expected_write_size)); |
7c673cae | 10247 | } else { |
224ce89b | 10248 | wctx->csum_order = min_alloc_size_order; |
7c673cae FG |
10249 | } |
10250 | ||
31f18b77 FG |
10251 | if (wctx->compress) { |
10252 | wctx->target_blob_size = select_option( | |
7c673cae | 10253 | "compression_max_blob_size", |
31f18b77 | 10254 | comp_max_blob_size.load(), |
7c673cae FG |
10255 | [&]() { |
10256 | int val; | |
10257 | if(c->pool_opts.get(pool_opts_t::COMPRESSION_MAX_BLOB_SIZE, &val)) { | |
10258 | return boost::optional<uint64_t>((uint64_t)val); | |
10259 | } | |
10260 | return boost::optional<uint64_t>(); | |
10261 | } | |
10262 | ); | |
10263 | } | |
10264 | } else { | |
31f18b77 FG |
10265 | if (wctx->compress) { |
10266 | wctx->target_blob_size = select_option( | |
7c673cae | 10267 | "compression_min_blob_size", |
31f18b77 | 10268 | comp_min_blob_size.load(), |
7c673cae FG |
10269 | [&]() { |
10270 | int val; | |
10271 | if(c->pool_opts.get(pool_opts_t::COMPRESSION_MIN_BLOB_SIZE, &val)) { | |
10272 | return boost::optional<uint64_t>((uint64_t)val); | |
10273 | } | |
10274 | return boost::optional<uint64_t>(); | |
10275 | } | |
10276 | ); | |
10277 | } | |
10278 | } | |
31f18b77 | 10279 | |
7c673cae | 10280 | uint64_t max_bsize = max_blob_size.load(); |
31f18b77 FG |
10281 | if (wctx->target_blob_size == 0 || wctx->target_blob_size > max_bsize) { |
10282 | wctx->target_blob_size = max_bsize; | |
7c673cae | 10283 | } |
31f18b77 | 10284 | |
7c673cae FG |
10285 | // set the min blob size floor at 2x the min_alloc_size, or else we |
10286 | // won't be able to allocate a smaller extent for the compressed | |
10287 | // data. | |
31f18b77 FG |
10288 | if (wctx->compress && |
10289 | wctx->target_blob_size < min_alloc_size * 2) { | |
10290 | wctx->target_blob_size = min_alloc_size * 2; | |
7c673cae | 10291 | } |
31f18b77 FG |
10292 | |
10293 | dout(20) << __func__ << " prefer csum_order " << wctx->csum_order | |
10294 | << " target_blob_size 0x" << std::hex << wctx->target_blob_size | |
10295 | << std::dec << dendl; | |
10296 | } | |
10297 | ||
10298 | int BlueStore::_do_gc( | |
10299 | TransContext *txc, | |
10300 | CollectionRef& c, | |
10301 | OnodeRef o, | |
10302 | const GarbageCollector& gc, | |
10303 | const WriteContext& wctx, | |
10304 | uint64_t *dirty_start, | |
10305 | uint64_t *dirty_end) | |
10306 | { | |
10307 | auto& extents_to_collect = gc.get_extents_to_collect(); | |
10308 | ||
10309 | WriteContext wctx_gc; | |
7c673cae | 10310 | wctx_gc.fork(wctx); // make a clone for garbage collection |
7c673cae | 10311 | |
31f18b77 FG |
10312 | for (auto it = extents_to_collect.begin(); |
10313 | it != extents_to_collect.end(); | |
10314 | ++it) { | |
10315 | bufferlist bl; | |
10316 | int r = _do_read(c.get(), o, it->offset, it->length, bl, 0); | |
10317 | assert(r == (int)it->length); | |
10318 | ||
10319 | o->extent_map.fault_range(db, it->offset, it->length); | |
10320 | _do_write_data(txc, c, o, it->offset, it->length, bl, &wctx_gc); | |
10321 | logger->inc(l_bluestore_gc_merged, it->length); | |
10322 | ||
10323 | if (*dirty_start > it->offset) { | |
10324 | *dirty_start = it->offset; | |
10325 | } | |
10326 | ||
10327 | if (*dirty_end < it->offset + it->length) { | |
10328 | *dirty_end = it->offset + it->length; | |
10329 | } | |
10330 | } | |
10331 | ||
10332 | dout(30) << __func__ << " alloc write" << dendl; | |
10333 | int r = _do_alloc_write(txc, c, o, &wctx_gc); | |
10334 | if (r < 0) { | |
10335 | derr << __func__ << " _do_alloc_write failed with " << cpp_strerror(r) | |
10336 | << dendl; | |
10337 | return r; | |
10338 | } | |
10339 | ||
10340 | _wctx_finish(txc, c, o, &wctx_gc); | |
10341 | return 0; | |
10342 | } | |
10343 | ||
10344 | int BlueStore::_do_write( | |
10345 | TransContext *txc, | |
10346 | CollectionRef& c, | |
10347 | OnodeRef o, | |
10348 | uint64_t offset, | |
10349 | uint64_t length, | |
10350 | bufferlist& bl, | |
10351 | uint32_t fadvise_flags) | |
10352 | { | |
10353 | int r = 0; | |
10354 | ||
10355 | dout(20) << __func__ | |
10356 | << " " << o->oid | |
10357 | << " 0x" << std::hex << offset << "~" << length | |
10358 | << " - have 0x" << o->onode.size | |
10359 | << " (" << std::dec << o->onode.size << ")" | |
10360 | << " bytes" | |
10361 | << " fadvise_flags 0x" << std::hex << fadvise_flags << std::dec | |
10362 | << dendl; | |
10363 | _dump_onode(o); | |
10364 | ||
10365 | if (length == 0) { | |
10366 | return 0; | |
10367 | } | |
10368 | ||
10369 | uint64_t end = offset + length; | |
10370 | ||
10371 | GarbageCollector gc(c->store->cct); | |
10372 | int64_t benefit; | |
10373 | auto dirty_start = offset; | |
10374 | auto dirty_end = end; | |
10375 | ||
10376 | WriteContext wctx; | |
10377 | _choose_write_options(c, o, fadvise_flags, &wctx); | |
7c673cae FG |
10378 | o->extent_map.fault_range(db, offset, length); |
10379 | _do_write_data(txc, c, o, offset, length, bl, &wctx); | |
7c673cae FG |
10380 | r = _do_alloc_write(txc, c, o, &wctx); |
10381 | if (r < 0) { | |
10382 | derr << __func__ << " _do_alloc_write failed with " << cpp_strerror(r) | |
10383 | << dendl; | |
10384 | goto out; | |
10385 | } | |
10386 | ||
31f18b77 FG |
10387 | // NB: _wctx_finish() will empty old_extents |
10388 | // so we must do gc estimation before that | |
7c673cae | 10389 | benefit = gc.estimate(offset, |
31f18b77 FG |
10390 | length, |
10391 | o->extent_map, | |
10392 | wctx.old_extents, | |
10393 | min_alloc_size); | |
7c673cae FG |
10394 | |
10395 | _wctx_finish(txc, c, o, &wctx); | |
10396 | if (end > o->onode.size) { | |
10397 | dout(20) << __func__ << " extending size to 0x" << std::hex << end | |
31f18b77 | 10398 | << std::dec << dendl; |
7c673cae FG |
10399 | o->onode.size = end; |
10400 | } | |
10401 | ||
10402 | if (benefit >= g_conf->bluestore_gc_enable_total_threshold) { | |
31f18b77 FG |
10403 | if (!gc.get_extents_to_collect().empty()) { |
10404 | dout(20) << __func__ << " perform garbage collection, " | |
10405 | << "expected benefit = " << benefit << " AUs" << dendl; | |
10406 | r = _do_gc(txc, c, o, gc, wctx, &dirty_start, &dirty_end); | |
10407 | if (r < 0) { | |
10408 | derr << __func__ << " _do_gc failed with " << cpp_strerror(r) | |
10409 | << dendl; | |
10410 | goto out; | |
7c673cae FG |
10411 | } |
10412 | } | |
10413 | } | |
7c673cae FG |
10414 | |
10415 | o->extent_map.compress_extent_map(dirty_start, dirty_end - dirty_start); | |
31f18b77 FG |
10416 | o->extent_map.dirty_range(dirty_start, dirty_end - dirty_start); |
10417 | ||
7c673cae FG |
10418 | r = 0; |
10419 | ||
10420 | out: | |
10421 | return r; | |
10422 | } | |
10423 | ||
10424 | int BlueStore::_write(TransContext *txc, | |
10425 | CollectionRef& c, | |
10426 | OnodeRef& o, | |
31f18b77 FG |
10427 | uint64_t offset, size_t length, |
10428 | bufferlist& bl, | |
10429 | uint32_t fadvise_flags) | |
7c673cae FG |
10430 | { |
10431 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10432 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
10433 | << dendl; | |
35e4c445 FG |
10434 | int r = 0; |
10435 | if (offset + length >= OBJECT_MAX_SIZE) { | |
10436 | r = -E2BIG; | |
10437 | } else { | |
10438 | _assign_nid(txc, o); | |
10439 | r = _do_write(txc, c, o, offset, length, bl, fadvise_flags); | |
10440 | txc->write_onode(o); | |
10441 | } | |
7c673cae FG |
10442 | dout(10) << __func__ << " " << c->cid << " " << o->oid |
10443 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
10444 | << " = " << r << dendl; | |
10445 | return r; | |
10446 | } | |
10447 | ||
10448 | int BlueStore::_zero(TransContext *txc, | |
10449 | CollectionRef& c, | |
10450 | OnodeRef& o, | |
10451 | uint64_t offset, size_t length) | |
10452 | { | |
10453 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10454 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
10455 | << dendl; | |
35e4c445 FG |
10456 | int r = 0; |
10457 | if (offset + length >= OBJECT_MAX_SIZE) { | |
10458 | r = -E2BIG; | |
10459 | } else { | |
10460 | _assign_nid(txc, o); | |
10461 | r = _do_zero(txc, c, o, offset, length); | |
10462 | } | |
7c673cae FG |
10463 | dout(10) << __func__ << " " << c->cid << " " << o->oid |
10464 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
10465 | << " = " << r << dendl; | |
10466 | return r; | |
10467 | } | |
10468 | ||
10469 | int BlueStore::_do_zero(TransContext *txc, | |
10470 | CollectionRef& c, | |
10471 | OnodeRef& o, | |
10472 | uint64_t offset, size_t length) | |
10473 | { | |
10474 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10475 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
10476 | << dendl; | |
10477 | int r = 0; | |
10478 | ||
10479 | _dump_onode(o); | |
10480 | ||
10481 | WriteContext wctx; | |
10482 | o->extent_map.fault_range(db, offset, length); | |
10483 | o->extent_map.punch_hole(c, offset, length, &wctx.old_extents); | |
31f18b77 | 10484 | o->extent_map.dirty_range(offset, length); |
7c673cae FG |
10485 | _wctx_finish(txc, c, o, &wctx); |
10486 | ||
10487 | if (offset + length > o->onode.size) { | |
10488 | o->onode.size = offset + length; | |
10489 | dout(20) << __func__ << " extending size to " << offset + length | |
10490 | << dendl; | |
10491 | } | |
10492 | txc->write_onode(o); | |
10493 | ||
10494 | dout(10) << __func__ << " " << c->cid << " " << o->oid | |
10495 | << " 0x" << std::hex << offset << "~" << length << std::dec | |
10496 | << " = " << r << dendl; | |
10497 | return r; | |
10498 | } | |
10499 | ||
10500 | void BlueStore::_do_truncate( | |
31f18b77 FG |
10501 | TransContext *txc, CollectionRef& c, OnodeRef o, uint64_t offset, |
10502 | set<SharedBlob*> *maybe_unshared_blobs) | |
7c673cae FG |
10503 | { |
10504 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10505 | << " 0x" << std::hex << offset << std::dec << dendl; | |
10506 | ||
10507 | _dump_onode(o, 30); | |
10508 | ||
10509 | if (offset == o->onode.size) | |
31f18b77 | 10510 | return; |
7c673cae FG |
10511 | |
10512 | if (offset < o->onode.size) { | |
10513 | WriteContext wctx; | |
10514 | uint64_t length = o->onode.size - offset; | |
10515 | o->extent_map.fault_range(db, offset, length); | |
10516 | o->extent_map.punch_hole(c, offset, length, &wctx.old_extents); | |
31f18b77 FG |
10517 | o->extent_map.dirty_range(offset, length); |
10518 | _wctx_finish(txc, c, o, &wctx, maybe_unshared_blobs); | |
7c673cae FG |
10519 | |
10520 | // if we have shards past EOF, ask for a reshard | |
10521 | if (!o->onode.extent_map_shards.empty() && | |
10522 | o->onode.extent_map_shards.back().offset >= offset) { | |
10523 | dout(10) << __func__ << " request reshard past EOF" << dendl; | |
10524 | if (offset) { | |
10525 | o->extent_map.request_reshard(offset - 1, offset + length); | |
10526 | } else { | |
10527 | o->extent_map.request_reshard(0, length); | |
10528 | } | |
10529 | } | |
10530 | } | |
10531 | ||
10532 | o->onode.size = offset; | |
10533 | ||
10534 | txc->write_onode(o); | |
10535 | } | |
10536 | ||
35e4c445 | 10537 | int BlueStore::_truncate(TransContext *txc, |
7c673cae FG |
10538 | CollectionRef& c, |
10539 | OnodeRef& o, | |
10540 | uint64_t offset) | |
10541 | { | |
10542 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10543 | << " 0x" << std::hex << offset << std::dec | |
10544 | << dendl; | |
35e4c445 FG |
10545 | int r = 0; |
10546 | if (offset >= OBJECT_MAX_SIZE) { | |
10547 | r = -E2BIG; | |
10548 | } else { | |
10549 | _do_truncate(txc, c, o, offset); | |
10550 | } | |
10551 | dout(10) << __func__ << " " << c->cid << " " << o->oid | |
10552 | << " 0x" << std::hex << offset << std::dec | |
10553 | << " = " << r << dendl; | |
10554 | return r; | |
7c673cae FG |
10555 | } |
10556 | ||
10557 | int BlueStore::_do_remove( | |
10558 | TransContext *txc, | |
10559 | CollectionRef& c, | |
10560 | OnodeRef o) | |
10561 | { | |
31f18b77 | 10562 | set<SharedBlob*> maybe_unshared_blobs; |
224ce89b WB |
10563 | bool is_gen = !o->oid.is_no_gen(); |
10564 | _do_truncate(txc, c, o, 0, is_gen ? &maybe_unshared_blobs : nullptr); | |
7c673cae FG |
10565 | if (o->onode.has_omap()) { |
10566 | o->flush(); | |
10567 | _do_omap_clear(txc, o->onode.nid); | |
10568 | } | |
10569 | o->exists = false; | |
10570 | string key; | |
10571 | for (auto &s : o->extent_map.shards) { | |
10572 | dout(20) << __func__ << " removing shard 0x" << std::hex | |
10573 | << s.shard_info->offset << std::dec << dendl; | |
10574 | generate_extent_shard_key_and_apply(o->key, s.shard_info->offset, &key, | |
10575 | [&](const string& final_key) { | |
10576 | txc->t->rmkey(PREFIX_OBJ, final_key); | |
10577 | } | |
10578 | ); | |
10579 | } | |
10580 | txc->t->rmkey(PREFIX_OBJ, o->key.c_str(), o->key.size()); | |
10581 | txc->removed(o); | |
10582 | o->extent_map.clear(); | |
10583 | o->onode = bluestore_onode_t(); | |
10584 | _debug_obj_on_delete(o->oid); | |
31f18b77 | 10585 | |
224ce89b WB |
10586 | if (!is_gen || maybe_unshared_blobs.empty()) { |
10587 | return 0; | |
10588 | } | |
31f18b77 | 10589 | |
224ce89b WB |
10590 | // see if we can unshare blobs still referenced by the head |
10591 | dout(10) << __func__ << " gen and maybe_unshared_blobs " | |
10592 | << maybe_unshared_blobs << dendl; | |
10593 | ghobject_t nogen = o->oid; | |
10594 | nogen.generation = ghobject_t::NO_GEN; | |
10595 | OnodeRef h = c->onode_map.lookup(nogen); | |
10596 | ||
10597 | if (!h || !h->exists) { | |
10598 | return 0; | |
10599 | } | |
10600 | ||
10601 | dout(20) << __func__ << " checking for unshareable blobs on " << h | |
10602 | << " " << h->oid << dendl; | |
10603 | map<SharedBlob*,bluestore_extent_ref_map_t> expect; | |
10604 | for (auto& e : h->extent_map.extent_map) { | |
10605 | const bluestore_blob_t& b = e.blob->get_blob(); | |
10606 | SharedBlob *sb = e.blob->shared_blob.get(); | |
10607 | if (b.is_shared() && | |
10608 | sb->loaded && | |
10609 | maybe_unshared_blobs.count(sb)) { | |
10610 | b.map(e.blob_offset, e.length, [&](uint64_t off, uint64_t len) { | |
10611 | expect[sb].get(off, len); | |
10612 | return 0; | |
10613 | }); | |
10614 | } | |
10615 | } | |
31f18b77 | 10616 | |
224ce89b WB |
10617 | vector<SharedBlob*> unshared_blobs; |
10618 | unshared_blobs.reserve(maybe_unshared_blobs.size()); | |
10619 | for (auto& p : expect) { | |
10620 | dout(20) << " ? " << *p.first << " vs " << p.second << dendl; | |
10621 | if (p.first->persistent->ref_map == p.second) { | |
10622 | SharedBlob *sb = p.first; | |
10623 | dout(20) << __func__ << " unsharing " << *sb << dendl; | |
10624 | unshared_blobs.push_back(sb); | |
10625 | txc->unshare_blob(sb); | |
10626 | uint64_t sbid = c->make_blob_unshared(sb); | |
10627 | string key; | |
10628 | get_shared_blob_key(sbid, &key); | |
10629 | txc->t->rmkey(PREFIX_SHARED_BLOB, key); | |
10630 | } | |
10631 | } | |
10632 | ||
10633 | if (unshared_blobs.empty()) { | |
10634 | return 0; | |
10635 | } | |
10636 | ||
224ce89b WB |
10637 | for (auto& e : h->extent_map.extent_map) { |
10638 | const bluestore_blob_t& b = e.blob->get_blob(); | |
10639 | SharedBlob *sb = e.blob->shared_blob.get(); | |
10640 | if (b.is_shared() && | |
10641 | std::find(unshared_blobs.begin(), unshared_blobs.end(), | |
10642 | sb) != unshared_blobs.end()) { | |
10643 | dout(20) << __func__ << " unsharing " << e << dendl; | |
10644 | bluestore_blob_t& blob = e.blob->dirty_blob(); | |
10645 | blob.clear_flag(bluestore_blob_t::FLAG_SHARED); | |
d2e6a577 | 10646 | h->extent_map.dirty_range(e.logical_offset, 1); |
31f18b77 FG |
10647 | } |
10648 | } | |
224ce89b WB |
10649 | txc->write_onode(h); |
10650 | ||
7c673cae FG |
10651 | return 0; |
10652 | } | |
10653 | ||
10654 | int BlueStore::_remove(TransContext *txc, | |
10655 | CollectionRef& c, | |
10656 | OnodeRef &o) | |
10657 | { | |
10658 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10659 | int r = _do_remove(txc, c, o); | |
10660 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10661 | return r; | |
10662 | } | |
10663 | ||
10664 | int BlueStore::_setattr(TransContext *txc, | |
10665 | CollectionRef& c, | |
10666 | OnodeRef& o, | |
10667 | const string& name, | |
10668 | bufferptr& val) | |
10669 | { | |
10670 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10671 | << " " << name << " (" << val.length() << " bytes)" | |
10672 | << dendl; | |
10673 | int r = 0; | |
10674 | if (val.is_partial()) | |
10675 | o->onode.attrs[name.c_str()] = bufferptr(val.c_str(), val.length()); | |
10676 | else | |
10677 | o->onode.attrs[name.c_str()] = val; | |
10678 | txc->write_onode(o); | |
10679 | dout(10) << __func__ << " " << c->cid << " " << o->oid | |
10680 | << " " << name << " (" << val.length() << " bytes)" | |
10681 | << " = " << r << dendl; | |
10682 | return r; | |
10683 | } | |
10684 | ||
10685 | int BlueStore::_setattrs(TransContext *txc, | |
10686 | CollectionRef& c, | |
10687 | OnodeRef& o, | |
10688 | const map<string,bufferptr>& aset) | |
10689 | { | |
10690 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10691 | << " " << aset.size() << " keys" | |
10692 | << dendl; | |
10693 | int r = 0; | |
10694 | for (map<string,bufferptr>::const_iterator p = aset.begin(); | |
10695 | p != aset.end(); ++p) { | |
10696 | if (p->second.is_partial()) | |
10697 | o->onode.attrs[p->first.c_str()] = | |
10698 | bufferptr(p->second.c_str(), p->second.length()); | |
10699 | else | |
10700 | o->onode.attrs[p->first.c_str()] = p->second; | |
10701 | } | |
10702 | txc->write_onode(o); | |
10703 | dout(10) << __func__ << " " << c->cid << " " << o->oid | |
10704 | << " " << aset.size() << " keys" | |
10705 | << " = " << r << dendl; | |
10706 | return r; | |
10707 | } | |
10708 | ||
10709 | ||
10710 | int BlueStore::_rmattr(TransContext *txc, | |
10711 | CollectionRef& c, | |
10712 | OnodeRef& o, | |
10713 | const string& name) | |
10714 | { | |
10715 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10716 | << " " << name << dendl; | |
10717 | int r = 0; | |
10718 | auto it = o->onode.attrs.find(name.c_str()); | |
10719 | if (it == o->onode.attrs.end()) | |
10720 | goto out; | |
10721 | ||
10722 | o->onode.attrs.erase(it); | |
10723 | txc->write_onode(o); | |
10724 | ||
10725 | out: | |
10726 | dout(10) << __func__ << " " << c->cid << " " << o->oid | |
10727 | << " " << name << " = " << r << dendl; | |
10728 | return r; | |
10729 | } | |
10730 | ||
10731 | int BlueStore::_rmattrs(TransContext *txc, | |
10732 | CollectionRef& c, | |
10733 | OnodeRef& o) | |
10734 | { | |
10735 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10736 | int r = 0; | |
10737 | ||
10738 | if (o->onode.attrs.empty()) | |
10739 | goto out; | |
10740 | ||
10741 | o->onode.attrs.clear(); | |
10742 | txc->write_onode(o); | |
10743 | ||
10744 | out: | |
10745 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10746 | return r; | |
10747 | } | |
10748 | ||
10749 | void BlueStore::_do_omap_clear(TransContext *txc, uint64_t id) | |
10750 | { | |
10751 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_OMAP); | |
10752 | string prefix, tail; | |
10753 | get_omap_header(id, &prefix); | |
10754 | get_omap_tail(id, &tail); | |
10755 | it->lower_bound(prefix); | |
10756 | while (it->valid()) { | |
10757 | if (it->key() >= tail) { | |
10758 | dout(30) << __func__ << " stop at " << pretty_binary_string(tail) | |
10759 | << dendl; | |
10760 | break; | |
10761 | } | |
10762 | txc->t->rmkey(PREFIX_OMAP, it->key()); | |
10763 | dout(30) << __func__ << " rm " << pretty_binary_string(it->key()) << dendl; | |
10764 | it->next(); | |
10765 | } | |
10766 | } | |
10767 | ||
10768 | int BlueStore::_omap_clear(TransContext *txc, | |
10769 | CollectionRef& c, | |
10770 | OnodeRef& o) | |
10771 | { | |
10772 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10773 | int r = 0; | |
10774 | if (o->onode.has_omap()) { | |
10775 | o->flush(); | |
10776 | _do_omap_clear(txc, o->onode.nid); | |
10777 | o->onode.clear_omap_flag(); | |
10778 | txc->write_onode(o); | |
10779 | } | |
10780 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10781 | return r; | |
10782 | } | |
10783 | ||
10784 | int BlueStore::_omap_setkeys(TransContext *txc, | |
10785 | CollectionRef& c, | |
10786 | OnodeRef& o, | |
10787 | bufferlist &bl) | |
10788 | { | |
10789 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10790 | int r; | |
10791 | bufferlist::iterator p = bl.begin(); | |
10792 | __u32 num; | |
10793 | if (!o->onode.has_omap()) { | |
10794 | o->onode.set_omap_flag(); | |
10795 | txc->write_onode(o); | |
10796 | } else { | |
10797 | txc->note_modified_object(o); | |
10798 | } | |
10799 | string final_key; | |
10800 | _key_encode_u64(o->onode.nid, &final_key); | |
10801 | final_key.push_back('.'); | |
10802 | ::decode(num, p); | |
10803 | while (num--) { | |
10804 | string key; | |
10805 | bufferlist value; | |
10806 | ::decode(key, p); | |
10807 | ::decode(value, p); | |
10808 | final_key.resize(9); // keep prefix | |
10809 | final_key += key; | |
10810 | dout(30) << __func__ << " " << pretty_binary_string(final_key) | |
10811 | << " <- " << key << dendl; | |
10812 | txc->t->set(PREFIX_OMAP, final_key, value); | |
10813 | } | |
10814 | r = 0; | |
10815 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10816 | return r; | |
10817 | } | |
10818 | ||
10819 | int BlueStore::_omap_setheader(TransContext *txc, | |
10820 | CollectionRef& c, | |
10821 | OnodeRef &o, | |
10822 | bufferlist& bl) | |
10823 | { | |
10824 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10825 | int r; | |
10826 | string key; | |
10827 | if (!o->onode.has_omap()) { | |
10828 | o->onode.set_omap_flag(); | |
10829 | txc->write_onode(o); | |
10830 | } else { | |
10831 | txc->note_modified_object(o); | |
10832 | } | |
10833 | get_omap_header(o->onode.nid, &key); | |
10834 | txc->t->set(PREFIX_OMAP, key, bl); | |
10835 | r = 0; | |
10836 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10837 | return r; | |
10838 | } | |
10839 | ||
10840 | int BlueStore::_omap_rmkeys(TransContext *txc, | |
10841 | CollectionRef& c, | |
10842 | OnodeRef& o, | |
10843 | bufferlist& bl) | |
10844 | { | |
10845 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10846 | int r = 0; | |
10847 | bufferlist::iterator p = bl.begin(); | |
10848 | __u32 num; | |
10849 | string final_key; | |
10850 | ||
10851 | if (!o->onode.has_omap()) { | |
10852 | goto out; | |
10853 | } | |
10854 | _key_encode_u64(o->onode.nid, &final_key); | |
10855 | final_key.push_back('.'); | |
10856 | ::decode(num, p); | |
10857 | while (num--) { | |
10858 | string key; | |
10859 | ::decode(key, p); | |
10860 | final_key.resize(9); // keep prefix | |
10861 | final_key += key; | |
10862 | dout(30) << __func__ << " rm " << pretty_binary_string(final_key) | |
10863 | << " <- " << key << dendl; | |
10864 | txc->t->rmkey(PREFIX_OMAP, final_key); | |
10865 | } | |
10866 | txc->note_modified_object(o); | |
10867 | ||
10868 | out: | |
10869 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10870 | return r; | |
10871 | } | |
10872 | ||
10873 | int BlueStore::_omap_rmkey_range(TransContext *txc, | |
10874 | CollectionRef& c, | |
10875 | OnodeRef& o, | |
10876 | const string& first, const string& last) | |
10877 | { | |
10878 | dout(15) << __func__ << " " << c->cid << " " << o->oid << dendl; | |
10879 | KeyValueDB::Iterator it; | |
10880 | string key_first, key_last; | |
10881 | int r = 0; | |
10882 | if (!o->onode.has_omap()) { | |
10883 | goto out; | |
10884 | } | |
10885 | o->flush(); | |
10886 | it = db->get_iterator(PREFIX_OMAP); | |
10887 | get_omap_key(o->onode.nid, first, &key_first); | |
10888 | get_omap_key(o->onode.nid, last, &key_last); | |
10889 | it->lower_bound(key_first); | |
10890 | while (it->valid()) { | |
10891 | if (it->key() >= key_last) { | |
10892 | dout(30) << __func__ << " stop at " << pretty_binary_string(key_last) | |
10893 | << dendl; | |
10894 | break; | |
10895 | } | |
10896 | txc->t->rmkey(PREFIX_OMAP, it->key()); | |
10897 | dout(30) << __func__ << " rm " << pretty_binary_string(it->key()) << dendl; | |
10898 | it->next(); | |
10899 | } | |
10900 | txc->note_modified_object(o); | |
10901 | ||
10902 | out: | |
10903 | dout(10) << __func__ << " " << c->cid << " " << o->oid << " = " << r << dendl; | |
10904 | return r; | |
10905 | } | |
10906 | ||
10907 | int BlueStore::_set_alloc_hint( | |
10908 | TransContext *txc, | |
10909 | CollectionRef& c, | |
10910 | OnodeRef& o, | |
10911 | uint64_t expected_object_size, | |
10912 | uint64_t expected_write_size, | |
10913 | uint32_t flags) | |
10914 | { | |
10915 | dout(15) << __func__ << " " << c->cid << " " << o->oid | |
10916 | << " object_size " << expected_object_size | |
10917 | << " write_size " << expected_write_size | |
10918 | << " flags " << ceph_osd_alloc_hint_flag_string(flags) | |
10919 | << dendl; | |
10920 | int r = 0; | |
10921 | o->onode.expected_object_size = expected_object_size; | |
10922 | o->onode.expected_write_size = expected_write_size; | |
10923 | o->onode.alloc_hint_flags = flags; | |
10924 | txc->write_onode(o); | |
10925 | dout(10) << __func__ << " " << c->cid << " " << o->oid | |
10926 | << " object_size " << expected_object_size | |
10927 | << " write_size " << expected_write_size | |
10928 | << " flags " << ceph_osd_alloc_hint_flag_string(flags) | |
10929 | << " = " << r << dendl; | |
10930 | return r; | |
10931 | } | |
10932 | ||
10933 | int BlueStore::_clone(TransContext *txc, | |
10934 | CollectionRef& c, | |
10935 | OnodeRef& oldo, | |
10936 | OnodeRef& newo) | |
10937 | { | |
10938 | dout(15) << __func__ << " " << c->cid << " " << oldo->oid << " -> " | |
10939 | << newo->oid << dendl; | |
10940 | int r = 0; | |
10941 | if (oldo->oid.hobj.get_hash() != newo->oid.hobj.get_hash()) { | |
10942 | derr << __func__ << " mismatched hash on " << oldo->oid | |
10943 | << " and " << newo->oid << dendl; | |
10944 | return -EINVAL; | |
10945 | } | |
10946 | ||
7c673cae FG |
10947 | _assign_nid(txc, newo); |
10948 | ||
10949 | // clone data | |
10950 | oldo->flush(); | |
10951 | _do_truncate(txc, c, newo, 0); | |
10952 | if (cct->_conf->bluestore_clone_cow) { | |
10953 | _do_clone_range(txc, c, oldo, newo, 0, oldo->onode.size, 0); | |
10954 | } else { | |
10955 | bufferlist bl; | |
10956 | r = _do_read(c.get(), oldo, 0, oldo->onode.size, bl, 0); | |
10957 | if (r < 0) | |
10958 | goto out; | |
10959 | r = _do_write(txc, c, newo, 0, oldo->onode.size, bl, 0); | |
10960 | if (r < 0) | |
10961 | goto out; | |
10962 | } | |
10963 | ||
10964 | // clone attrs | |
10965 | newo->onode.attrs = oldo->onode.attrs; | |
10966 | ||
10967 | // clone omap | |
10968 | if (newo->onode.has_omap()) { | |
10969 | dout(20) << __func__ << " clearing old omap data" << dendl; | |
10970 | newo->flush(); | |
10971 | _do_omap_clear(txc, newo->onode.nid); | |
10972 | } | |
10973 | if (oldo->onode.has_omap()) { | |
10974 | dout(20) << __func__ << " copying omap data" << dendl; | |
10975 | if (!newo->onode.has_omap()) { | |
10976 | newo->onode.set_omap_flag(); | |
10977 | } | |
10978 | KeyValueDB::Iterator it = db->get_iterator(PREFIX_OMAP); | |
10979 | string head, tail; | |
10980 | get_omap_header(oldo->onode.nid, &head); | |
10981 | get_omap_tail(oldo->onode.nid, &tail); | |
10982 | it->lower_bound(head); | |
10983 | while (it->valid()) { | |
10984 | if (it->key() >= tail) { | |
10985 | dout(30) << __func__ << " reached tail" << dendl; | |
10986 | break; | |
10987 | } else { | |
10988 | dout(30) << __func__ << " got header/data " | |
10989 | << pretty_binary_string(it->key()) << dendl; | |
10990 | string key; | |
10991 | rewrite_omap_key(newo->onode.nid, it->key(), &key); | |
10992 | txc->t->set(PREFIX_OMAP, key, it->value()); | |
10993 | } | |
10994 | it->next(); | |
10995 | } | |
10996 | } else { | |
10997 | newo->onode.clear_omap_flag(); | |
10998 | } | |
10999 | ||
11000 | txc->write_onode(newo); | |
11001 | r = 0; | |
11002 | ||
11003 | out: | |
11004 | dout(10) << __func__ << " " << c->cid << " " << oldo->oid << " -> " | |
11005 | << newo->oid << " = " << r << dendl; | |
11006 | return r; | |
11007 | } | |
11008 | ||
11009 | int BlueStore::_do_clone_range( | |
11010 | TransContext *txc, | |
11011 | CollectionRef& c, | |
11012 | OnodeRef& oldo, | |
11013 | OnodeRef& newo, | |
224ce89b WB |
11014 | uint64_t srcoff, |
11015 | uint64_t length, | |
11016 | uint64_t dstoff) | |
7c673cae FG |
11017 | { |
11018 | dout(15) << __func__ << " " << c->cid << " " << oldo->oid << " -> " | |
11019 | << newo->oid | |
11020 | << " 0x" << std::hex << srcoff << "~" << length << " -> " | |
11021 | << " 0x" << dstoff << "~" << length << std::dec << dendl; | |
11022 | oldo->extent_map.fault_range(db, srcoff, length); | |
11023 | newo->extent_map.fault_range(db, dstoff, length); | |
11024 | _dump_onode(oldo); | |
11025 | _dump_onode(newo); | |
11026 | ||
11027 | // hmm, this could go into an ExtentMap::dup() method. | |
11028 | vector<BlobRef> id_to_blob(oldo->extent_map.extent_map.size()); | |
11029 | for (auto &e : oldo->extent_map.extent_map) { | |
11030 | e.blob->last_encoded_id = -1; | |
11031 | } | |
11032 | int n = 0; | |
7c673cae | 11033 | uint64_t end = srcoff + length; |
224ce89b WB |
11034 | uint32_t dirty_range_begin = 0; |
11035 | uint32_t dirty_range_end = 0; | |
35e4c445 | 11036 | bool src_dirty = false; |
7c673cae FG |
11037 | for (auto ep = oldo->extent_map.seek_lextent(srcoff); |
11038 | ep != oldo->extent_map.extent_map.end(); | |
11039 | ++ep) { | |
11040 | auto& e = *ep; | |
11041 | if (e.logical_offset >= end) { | |
11042 | break; | |
11043 | } | |
11044 | dout(20) << __func__ << " src " << e << dendl; | |
11045 | BlobRef cb; | |
11046 | bool blob_duped = true; | |
11047 | if (e.blob->last_encoded_id >= 0) { | |
11048 | // blob is already duped | |
11049 | cb = id_to_blob[e.blob->last_encoded_id]; | |
11050 | blob_duped = false; | |
11051 | } else { | |
11052 | // dup the blob | |
11053 | const bluestore_blob_t& blob = e.blob->get_blob(); | |
11054 | // make sure it is shared | |
11055 | if (!blob.is_shared()) { | |
11056 | c->make_blob_shared(_assign_blobid(txc), e.blob); | |
35e4c445 FG |
11057 | if (!src_dirty) { |
11058 | src_dirty = true; | |
224ce89b WB |
11059 | dirty_range_begin = e.logical_offset; |
11060 | } | |
11061 | assert(e.logical_end() > 0); | |
11062 | // -1 to exclude next potential shard | |
11063 | dirty_range_end = e.logical_end() - 1; | |
7c673cae FG |
11064 | } else { |
11065 | c->load_shared_blob(e.blob->shared_blob); | |
11066 | } | |
11067 | cb = new Blob(); | |
11068 | e.blob->last_encoded_id = n; | |
11069 | id_to_blob[n] = cb; | |
11070 | e.blob->dup(*cb); | |
11071 | // bump the extent refs on the copied blob's extents | |
11072 | for (auto p : blob.get_extents()) { | |
11073 | if (p.is_valid()) { | |
11074 | e.blob->shared_blob->get_ref(p.offset, p.length); | |
11075 | } | |
11076 | } | |
11077 | txc->write_shared_blob(e.blob->shared_blob); | |
11078 | dout(20) << __func__ << " new " << *cb << dendl; | |
11079 | } | |
11080 | // dup extent | |
11081 | int skip_front, skip_back; | |
11082 | if (e.logical_offset < srcoff) { | |
11083 | skip_front = srcoff - e.logical_offset; | |
11084 | } else { | |
11085 | skip_front = 0; | |
11086 | } | |
11087 | if (e.logical_end() > end) { | |
11088 | skip_back = e.logical_end() - end; | |
11089 | } else { | |
11090 | skip_back = 0; | |
11091 | } | |
11092 | Extent *ne = new Extent(e.logical_offset + skip_front + dstoff - srcoff, | |
11093 | e.blob_offset + skip_front, | |
11094 | e.length - skip_front - skip_back, cb); | |
11095 | newo->extent_map.extent_map.insert(*ne); | |
11096 | ne->blob->get_ref(c.get(), ne->blob_offset, ne->length); | |
11097 | // fixme: we may leave parts of new blob unreferenced that could | |
11098 | // be freed (relative to the shared_blob). | |
11099 | txc->statfs_delta.stored() += ne->length; | |
11100 | if (e.blob->get_blob().is_compressed()) { | |
11101 | txc->statfs_delta.compressed_original() += ne->length; | |
11102 | if (blob_duped){ | |
11103 | txc->statfs_delta.compressed() += | |
11104 | cb->get_blob().get_compressed_payload_length(); | |
11105 | } | |
11106 | } | |
11107 | dout(20) << __func__ << " dst " << *ne << dendl; | |
11108 | ++n; | |
11109 | } | |
35e4c445 | 11110 | if (src_dirty) { |
224ce89b WB |
11111 | oldo->extent_map.dirty_range(dirty_range_begin, |
11112 | dirty_range_end - dirty_range_begin); | |
7c673cae FG |
11113 | txc->write_onode(oldo); |
11114 | } | |
11115 | txc->write_onode(newo); | |
11116 | ||
11117 | if (dstoff + length > newo->onode.size) { | |
11118 | newo->onode.size = dstoff + length; | |
11119 | } | |
31f18b77 | 11120 | newo->extent_map.dirty_range(dstoff, length); |
7c673cae FG |
11121 | _dump_onode(oldo); |
11122 | _dump_onode(newo); | |
11123 | return 0; | |
11124 | } | |
11125 | ||
11126 | int BlueStore::_clone_range(TransContext *txc, | |
11127 | CollectionRef& c, | |
11128 | OnodeRef& oldo, | |
11129 | OnodeRef& newo, | |
11130 | uint64_t srcoff, uint64_t length, uint64_t dstoff) | |
11131 | { | |
11132 | dout(15) << __func__ << " " << c->cid << " " << oldo->oid << " -> " | |
11133 | << newo->oid << " from 0x" << std::hex << srcoff << "~" << length | |
11134 | << " to offset 0x" << dstoff << std::dec << dendl; | |
11135 | int r = 0; | |
11136 | ||
35e4c445 FG |
11137 | if (srcoff + length >= OBJECT_MAX_SIZE || |
11138 | dstoff + length >= OBJECT_MAX_SIZE) { | |
11139 | r = -E2BIG; | |
11140 | goto out; | |
11141 | } | |
7c673cae FG |
11142 | if (srcoff + length > oldo->onode.size) { |
11143 | r = -EINVAL; | |
11144 | goto out; | |
11145 | } | |
11146 | ||
7c673cae FG |
11147 | _assign_nid(txc, newo); |
11148 | ||
11149 | if (length > 0) { | |
11150 | if (cct->_conf->bluestore_clone_cow) { | |
11151 | _do_zero(txc, c, newo, dstoff, length); | |
11152 | _do_clone_range(txc, c, oldo, newo, srcoff, length, dstoff); | |
11153 | } else { | |
11154 | bufferlist bl; | |
11155 | r = _do_read(c.get(), oldo, srcoff, length, bl, 0); | |
11156 | if (r < 0) | |
11157 | goto out; | |
11158 | r = _do_write(txc, c, newo, dstoff, bl.length(), bl, 0); | |
11159 | if (r < 0) | |
11160 | goto out; | |
11161 | } | |
11162 | } | |
11163 | ||
11164 | txc->write_onode(newo); | |
11165 | r = 0; | |
11166 | ||
11167 | out: | |
11168 | dout(10) << __func__ << " " << c->cid << " " << oldo->oid << " -> " | |
11169 | << newo->oid << " from 0x" << std::hex << srcoff << "~" << length | |
11170 | << " to offset 0x" << dstoff << std::dec | |
11171 | << " = " << r << dendl; | |
11172 | return r; | |
11173 | } | |
11174 | ||
11175 | int BlueStore::_rename(TransContext *txc, | |
11176 | CollectionRef& c, | |
11177 | OnodeRef& oldo, | |
11178 | OnodeRef& newo, | |
11179 | const ghobject_t& new_oid) | |
11180 | { | |
11181 | dout(15) << __func__ << " " << c->cid << " " << oldo->oid << " -> " | |
11182 | << new_oid << dendl; | |
11183 | int r; | |
11184 | ghobject_t old_oid = oldo->oid; | |
31f18b77 | 11185 | mempool::bluestore_cache_other::string new_okey; |
7c673cae FG |
11186 | |
11187 | if (newo) { | |
11188 | if (newo->exists) { | |
11189 | r = -EEXIST; | |
11190 | goto out; | |
11191 | } | |
11192 | assert(txc->onodes.count(newo) == 0); | |
11193 | } | |
11194 | ||
11195 | txc->t->rmkey(PREFIX_OBJ, oldo->key.c_str(), oldo->key.size()); | |
11196 | ||
11197 | // rewrite shards | |
11198 | { | |
11199 | oldo->extent_map.fault_range(db, 0, oldo->onode.size); | |
11200 | get_object_key(cct, new_oid, &new_okey); | |
11201 | string key; | |
11202 | for (auto &s : oldo->extent_map.shards) { | |
11203 | generate_extent_shard_key_and_apply(oldo->key, s.shard_info->offset, &key, | |
11204 | [&](const string& final_key) { | |
11205 | txc->t->rmkey(PREFIX_OBJ, final_key); | |
11206 | } | |
11207 | ); | |
11208 | s.dirty = true; | |
11209 | } | |
11210 | } | |
11211 | ||
11212 | newo = oldo; | |
11213 | txc->write_onode(newo); | |
11214 | ||
11215 | // this adjusts oldo->{oid,key}, and reset oldo to a fresh empty | |
11216 | // Onode in the old slot | |
11217 | c->onode_map.rename(oldo, old_oid, new_oid, new_okey); | |
11218 | r = 0; | |
11219 | ||
11220 | out: | |
11221 | dout(10) << __func__ << " " << c->cid << " " << old_oid << " -> " | |
11222 | << new_oid << " = " << r << dendl; | |
11223 | return r; | |
11224 | } | |
11225 | ||
11226 | // collections | |
11227 | ||
11228 | int BlueStore::_create_collection( | |
11229 | TransContext *txc, | |
11230 | const coll_t &cid, | |
11231 | unsigned bits, | |
11232 | CollectionRef *c) | |
11233 | { | |
11234 | dout(15) << __func__ << " " << cid << " bits " << bits << dendl; | |
11235 | int r; | |
11236 | bufferlist bl; | |
11237 | ||
11238 | { | |
11239 | RWLock::WLocker l(coll_lock); | |
11240 | if (*c) { | |
11241 | r = -EEXIST; | |
11242 | goto out; | |
11243 | } | |
11244 | c->reset( | |
11245 | new Collection( | |
11246 | this, | |
11247 | cache_shards[cid.hash_to_shard(cache_shards.size())], | |
11248 | cid)); | |
11249 | (*c)->cnode.bits = bits; | |
11250 | coll_map[cid] = *c; | |
11251 | } | |
11252 | ::encode((*c)->cnode, bl); | |
11253 | txc->t->set(PREFIX_COLL, stringify(cid), bl); | |
11254 | r = 0; | |
11255 | ||
11256 | out: | |
11257 | dout(10) << __func__ << " " << cid << " bits " << bits << " = " << r << dendl; | |
11258 | return r; | |
11259 | } | |
11260 | ||
11261 | int BlueStore::_remove_collection(TransContext *txc, const coll_t &cid, | |
11262 | CollectionRef *c) | |
11263 | { | |
11264 | dout(15) << __func__ << " " << cid << dendl; | |
11265 | int r; | |
11266 | ||
11267 | { | |
11268 | RWLock::WLocker l(coll_lock); | |
11269 | if (!*c) { | |
11270 | r = -ENOENT; | |
11271 | goto out; | |
11272 | } | |
11273 | size_t nonexistent_count = 0; | |
11274 | assert((*c)->exists); | |
11275 | if ((*c)->onode_map.map_any([&](OnodeRef o) { | |
11276 | if (o->exists) { | |
11277 | dout(10) << __func__ << " " << o->oid << " " << o | |
11278 | << " exists in onode_map" << dendl; | |
11279 | return true; | |
11280 | } | |
11281 | ++nonexistent_count; | |
11282 | return false; | |
11283 | })) { | |
11284 | r = -ENOTEMPTY; | |
11285 | goto out; | |
11286 | } | |
11287 | ||
11288 | vector<ghobject_t> ls; | |
11289 | ghobject_t next; | |
11290 | // Enumerate onodes in db, up to nonexistent_count + 1 | |
11291 | // then check if all of them are marked as non-existent. | |
11292 | // Bypass the check if returned number is greater than nonexistent_count | |
11293 | r = _collection_list(c->get(), ghobject_t(), ghobject_t::get_max(), | |
11294 | nonexistent_count + 1, &ls, &next); | |
11295 | if (r >= 0) { | |
11296 | bool exists = false; //ls.size() > nonexistent_count; | |
11297 | for (auto it = ls.begin(); !exists && it < ls.end(); ++it) { | |
11298 | dout(10) << __func__ << " oid " << *it << dendl; | |
11299 | auto onode = (*c)->onode_map.lookup(*it); | |
11300 | exists = !onode || onode->exists; | |
11301 | if (exists) { | |
11302 | dout(10) << __func__ << " " << *it | |
11303 | << " exists in db" << dendl; | |
11304 | } | |
11305 | } | |
11306 | if (!exists) { | |
11307 | coll_map.erase(cid); | |
11308 | txc->removed_collections.push_back(*c); | |
11309 | (*c)->exists = false; | |
11310 | c->reset(); | |
11311 | txc->t->rmkey(PREFIX_COLL, stringify(cid)); | |
11312 | r = 0; | |
11313 | } else { | |
11314 | dout(10) << __func__ << " " << cid | |
11315 | << " is non-empty" << dendl; | |
11316 | r = -ENOTEMPTY; | |
11317 | } | |
11318 | } | |
11319 | } | |
11320 | ||
11321 | out: | |
11322 | dout(10) << __func__ << " " << cid << " = " << r << dendl; | |
11323 | return r; | |
11324 | } | |
11325 | ||
11326 | int BlueStore::_split_collection(TransContext *txc, | |
11327 | CollectionRef& c, | |
11328 | CollectionRef& d, | |
11329 | unsigned bits, int rem) | |
11330 | { | |
11331 | dout(15) << __func__ << " " << c->cid << " to " << d->cid << " " | |
11332 | << " bits " << bits << dendl; | |
11333 | RWLock::WLocker l(c->lock); | |
11334 | RWLock::WLocker l2(d->lock); | |
11335 | int r; | |
11336 | ||
11337 | // flush all previous deferred writes on this sequencer. this is a bit | |
11338 | // heavyweight, but we need to make sure all deferred writes complete | |
11339 | // before we split as the new collection's sequencer may need to order | |
11340 | // this after those writes, and we don't bother with the complexity of | |
11341 | // moving those TransContexts over to the new osr. | |
11342 | _osr_drain_preceding(txc); | |
11343 | ||
11344 | // move any cached items (onodes and referenced shared blobs) that will | |
11345 | // belong to the child collection post-split. leave everything else behind. | |
11346 | // this may include things that don't strictly belong to the now-smaller | |
11347 | // parent split, but the OSD will always send us a split for every new | |
11348 | // child. | |
11349 | ||
11350 | spg_t pgid, dest_pgid; | |
11351 | bool is_pg = c->cid.is_pg(&pgid); | |
11352 | assert(is_pg); | |
11353 | is_pg = d->cid.is_pg(&dest_pgid); | |
11354 | assert(is_pg); | |
11355 | ||
11356 | // the destination should initially be empty. | |
11357 | assert(d->onode_map.empty()); | |
11358 | assert(d->shared_blob_set.empty()); | |
11359 | assert(d->cnode.bits == bits); | |
11360 | ||
11361 | c->split_cache(d.get()); | |
11362 | ||
11363 | // adjust bits. note that this will be redundant for all but the first | |
11364 | // split call for this parent (first child). | |
11365 | c->cnode.bits = bits; | |
11366 | assert(d->cnode.bits == bits); | |
11367 | r = 0; | |
11368 | ||
11369 | bufferlist bl; | |
11370 | ::encode(c->cnode, bl); | |
11371 | txc->t->set(PREFIX_COLL, stringify(c->cid), bl); | |
11372 | ||
11373 | dout(10) << __func__ << " " << c->cid << " to " << d->cid << " " | |
11374 | << " bits " << bits << " = " << r << dendl; | |
11375 | return r; | |
11376 | } | |
11377 | ||
11378 | // DB key value Histogram | |
11379 | #define KEY_SLAB 32 | |
11380 | #define VALUE_SLAB 64 | |
11381 | ||
11382 | const string prefix_onode = "o"; | |
11383 | const string prefix_onode_shard = "x"; | |
11384 | const string prefix_other = "Z"; | |
11385 | ||
11386 | int BlueStore::DBHistogram::get_key_slab(size_t sz) | |
11387 | { | |
11388 | return (sz/KEY_SLAB); | |
11389 | } | |
11390 | ||
11391 | string BlueStore::DBHistogram::get_key_slab_to_range(int slab) | |
11392 | { | |
11393 | int lower_bound = slab * KEY_SLAB; | |
11394 | int upper_bound = (slab + 1) * KEY_SLAB; | |
11395 | string ret = "[" + stringify(lower_bound) + "," + stringify(upper_bound) + ")"; | |
11396 | return ret; | |
11397 | } | |
11398 | ||
11399 | int BlueStore::DBHistogram::get_value_slab(size_t sz) | |
11400 | { | |
11401 | return (sz/VALUE_SLAB); | |
11402 | } | |
11403 | ||
11404 | string BlueStore::DBHistogram::get_value_slab_to_range(int slab) | |
11405 | { | |
11406 | int lower_bound = slab * VALUE_SLAB; | |
11407 | int upper_bound = (slab + 1) * VALUE_SLAB; | |
11408 | string ret = "[" + stringify(lower_bound) + "," + stringify(upper_bound) + ")"; | |
11409 | return ret; | |
11410 | } | |
11411 | ||
11412 | void BlueStore::DBHistogram::update_hist_entry(map<string, map<int, struct key_dist> > &key_hist, | |
11413 | const string &prefix, size_t key_size, size_t value_size) | |
11414 | { | |
11415 | uint32_t key_slab = get_key_slab(key_size); | |
11416 | uint32_t value_slab = get_value_slab(value_size); | |
11417 | key_hist[prefix][key_slab].count++; | |
11418 | key_hist[prefix][key_slab].max_len = MAX(key_size, key_hist[prefix][key_slab].max_len); | |
11419 | key_hist[prefix][key_slab].val_map[value_slab].count++; | |
11420 | key_hist[prefix][key_slab].val_map[value_slab].max_len = | |
11421 | MAX(value_size, key_hist[prefix][key_slab].val_map[value_slab].max_len); | |
11422 | } | |
11423 | ||
11424 | void BlueStore::DBHistogram::dump(Formatter *f) | |
11425 | { | |
11426 | f->open_object_section("rocksdb_value_distribution"); | |
11427 | for (auto i : value_hist) { | |
11428 | f->dump_unsigned(get_value_slab_to_range(i.first).data(), i.second); | |
11429 | } | |
11430 | f->close_section(); | |
11431 | ||
11432 | f->open_object_section("rocksdb_key_value_histogram"); | |
11433 | for (auto i : key_hist) { | |
11434 | f->dump_string("prefix", i.first); | |
11435 | f->open_object_section("key_hist"); | |
11436 | for ( auto k : i.second) { | |
11437 | f->dump_unsigned(get_key_slab_to_range(k.first).data(), k.second.count); | |
11438 | f->dump_unsigned("max_len", k.second.max_len); | |
11439 | f->open_object_section("value_hist"); | |
11440 | for ( auto j : k.second.val_map) { | |
11441 | f->dump_unsigned(get_value_slab_to_range(j.first).data(), j.second.count); | |
11442 | f->dump_unsigned("max_len", j.second.max_len); | |
11443 | } | |
11444 | f->close_section(); | |
11445 | } | |
11446 | f->close_section(); | |
11447 | } | |
11448 | f->close_section(); | |
11449 | } | |
11450 | ||
11451 | //Itrerates through the db and collects the stats | |
11452 | void BlueStore::generate_db_histogram(Formatter *f) | |
11453 | { | |
11454 | //globals | |
11455 | uint64_t num_onodes = 0; | |
11456 | uint64_t num_shards = 0; | |
11457 | uint64_t num_super = 0; | |
11458 | uint64_t num_coll = 0; | |
11459 | uint64_t num_omap = 0; | |
11460 | uint64_t num_deferred = 0; | |
11461 | uint64_t num_alloc = 0; | |
11462 | uint64_t num_stat = 0; | |
11463 | uint64_t num_others = 0; | |
11464 | uint64_t num_shared_shards = 0; | |
11465 | size_t max_key_size =0, max_value_size = 0; | |
11466 | uint64_t total_key_size = 0, total_value_size = 0; | |
11467 | size_t key_size = 0, value_size = 0; | |
11468 | DBHistogram hist; | |
11469 | ||
11470 | utime_t start = ceph_clock_now(); | |
11471 | ||
11472 | KeyValueDB::WholeSpaceIterator iter = db->get_iterator(); | |
11473 | iter->seek_to_first(); | |
11474 | while (iter->valid()) { | |
11475 | dout(30) << __func__ << " Key: " << iter->key() << dendl; | |
11476 | key_size = iter->key_size(); | |
11477 | value_size = iter->value_size(); | |
11478 | hist.value_hist[hist.get_value_slab(value_size)]++; | |
11479 | max_key_size = MAX(max_key_size, key_size); | |
11480 | max_value_size = MAX(max_value_size, value_size); | |
11481 | total_key_size += key_size; | |
11482 | total_value_size += value_size; | |
11483 | ||
11484 | pair<string,string> key(iter->raw_key()); | |
11485 | ||
11486 | if (key.first == PREFIX_SUPER) { | |
11487 | hist.update_hist_entry(hist.key_hist, PREFIX_SUPER, key_size, value_size); | |
11488 | num_super++; | |
11489 | } else if (key.first == PREFIX_STAT) { | |
11490 | hist.update_hist_entry(hist.key_hist, PREFIX_STAT, key_size, value_size); | |
11491 | num_stat++; | |
11492 | } else if (key.first == PREFIX_COLL) { | |
11493 | hist.update_hist_entry(hist.key_hist, PREFIX_COLL, key_size, value_size); | |
11494 | num_coll++; | |
11495 | } else if (key.first == PREFIX_OBJ) { | |
11496 | if (key.second.back() == ONODE_KEY_SUFFIX) { | |
11497 | hist.update_hist_entry(hist.key_hist, prefix_onode, key_size, value_size); | |
11498 | num_onodes++; | |
11499 | } else { | |
11500 | hist.update_hist_entry(hist.key_hist, prefix_onode_shard, key_size, value_size); | |
11501 | num_shards++; | |
11502 | } | |
11503 | } else if (key.first == PREFIX_OMAP) { | |
11504 | hist.update_hist_entry(hist.key_hist, PREFIX_OMAP, key_size, value_size); | |
11505 | num_omap++; | |
11506 | } else if (key.first == PREFIX_DEFERRED) { | |
11507 | hist.update_hist_entry(hist.key_hist, PREFIX_DEFERRED, key_size, value_size); | |
11508 | num_deferred++; | |
11509 | } else if (key.first == PREFIX_ALLOC || key.first == "b" ) { | |
11510 | hist.update_hist_entry(hist.key_hist, PREFIX_ALLOC, key_size, value_size); | |
11511 | num_alloc++; | |
11512 | } else if (key.first == PREFIX_SHARED_BLOB) { | |
11513 | hist.update_hist_entry(hist.key_hist, PREFIX_SHARED_BLOB, key_size, value_size); | |
11514 | num_shared_shards++; | |
11515 | } else { | |
11516 | hist.update_hist_entry(hist.key_hist, prefix_other, key_size, value_size); | |
11517 | num_others++; | |
11518 | } | |
11519 | iter->next(); | |
11520 | } | |
11521 | ||
11522 | utime_t duration = ceph_clock_now() - start; | |
11523 | f->open_object_section("rocksdb_key_value_stats"); | |
11524 | f->dump_unsigned("num_onodes", num_onodes); | |
11525 | f->dump_unsigned("num_shards", num_shards); | |
11526 | f->dump_unsigned("num_super", num_super); | |
11527 | f->dump_unsigned("num_coll", num_coll); | |
11528 | f->dump_unsigned("num_omap", num_omap); | |
11529 | f->dump_unsigned("num_deferred", num_deferred); | |
11530 | f->dump_unsigned("num_alloc", num_alloc); | |
11531 | f->dump_unsigned("num_stat", num_stat); | |
11532 | f->dump_unsigned("num_shared_shards", num_shared_shards); | |
11533 | f->dump_unsigned("num_others", num_others); | |
11534 | f->dump_unsigned("max_key_size", max_key_size); | |
11535 | f->dump_unsigned("max_value_size", max_value_size); | |
11536 | f->dump_unsigned("total_key_size", total_key_size); | |
11537 | f->dump_unsigned("total_value_size", total_value_size); | |
11538 | f->close_section(); | |
11539 | ||
11540 | hist.dump(f); | |
11541 | ||
11542 | dout(20) << __func__ << " finished in " << duration << " seconds" << dendl; | |
11543 | ||
11544 | } | |
11545 | ||
31f18b77 | 11546 | void BlueStore::_flush_cache() |
7c673cae FG |
11547 | { |
11548 | dout(10) << __func__ << dendl; | |
11549 | for (auto i : cache_shards) { | |
11550 | i->trim_all(); | |
31f18b77 | 11551 | assert(i->empty()); |
7c673cae FG |
11552 | } |
11553 | for (auto& p : coll_map) { | |
11554 | assert(p.second->onode_map.empty()); | |
11555 | assert(p.second->shared_blob_set.empty()); | |
11556 | } | |
11557 | coll_map.clear(); | |
11558 | } | |
11559 | ||
31f18b77 FG |
11560 | // For external caller. |
11561 | // We use a best-effort policy instead, e.g., | |
11562 | // we don't care if there are still some pinned onodes/data in the cache | |
11563 | // after this command is completed. | |
11564 | void BlueStore::flush_cache() | |
11565 | { | |
11566 | dout(10) << __func__ << dendl; | |
11567 | for (auto i : cache_shards) { | |
11568 | i->trim_all(); | |
11569 | } | |
11570 | } | |
11571 | ||
7c673cae FG |
11572 | void BlueStore::_apply_padding(uint64_t head_pad, |
11573 | uint64_t tail_pad, | |
7c673cae FG |
11574 | bufferlist& padded) |
11575 | { | |
7c673cae | 11576 | if (head_pad) { |
224ce89b | 11577 | padded.prepend_zero(head_pad); |
7c673cae FG |
11578 | } |
11579 | if (tail_pad) { | |
11580 | padded.append_zero(tail_pad); | |
11581 | } | |
11582 | if (head_pad || tail_pad) { | |
11583 | dout(20) << __func__ << " can pad head 0x" << std::hex << head_pad | |
11584 | << " tail 0x" << tail_pad << std::dec << dendl; | |
11585 | logger->inc(l_bluestore_write_pad_bytes, head_pad + tail_pad); | |
11586 | } | |
11587 | } | |
11588 | ||
11589 | // =========================================== |