]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2011 New Dream Network | |
7 | * Copyright (C) 2013,2014 Cloudwatt <libre.licensing@cloudwatt.com> | |
8 | * | |
9 | * Author: Loic Dachary <loic@dachary.org> | |
10 | * | |
11 | * This is free software; you can redistribute it and/or | |
12 | * modify it under the terms of the GNU Lesser General Public | |
13 | * License version 2.1, as published by the Free Software | |
14 | * Foundation. See file COPYING. | |
15 | * | |
16 | */ | |
17 | ||
1e59de90 | 18 | #include <algorithm> |
9f95a23c TL |
19 | #include <list> |
20 | #include <map> | |
21 | #include <ostream> | |
22 | #include <sstream> | |
23 | #include <set> | |
24 | #include <string> | |
25 | #include <utility> | |
26 | #include <vector> | |
27 | ||
28 | ||
7c673cae FG |
29 | #include <boost/assign/list_of.hpp> |
30 | ||
7c673cae | 31 | #include "include/ceph_features.h" |
9f95a23c | 32 | #include "include/encoding.h" |
11fdf7f2 | 33 | #include "include/stringify.h" |
7c673cae FG |
34 | extern "C" { |
35 | #include "crush/hash.h" | |
36 | } | |
9f95a23c TL |
37 | |
38 | #include "common/Formatter.h" | |
f67539c2 | 39 | #include "common/StackStringStream.h" |
20effc67 | 40 | #include "include/utime_fmt.h" |
7c673cae | 41 | #include "OSDMap.h" |
9f95a23c | 42 | #include "osd_types.h" |
20effc67 | 43 | #include "osd_types_fmt.h" |
9f95a23c TL |
44 | #include "os/Transaction.h" |
45 | ||
46 | using std::list; | |
47 | using std::make_pair; | |
48 | using std::map; | |
49 | using std::ostream; | |
9f95a23c TL |
50 | using std::pair; |
51 | using std::set; | |
20effc67 | 52 | using std::shared_ptr; |
9f95a23c | 53 | using std::string; |
9f95a23c TL |
54 | using std::unique_ptr; |
55 | using std::vector; | |
56 | ||
f67539c2 | 57 | using ceph::bufferlist; |
9f95a23c TL |
58 | using ceph::decode; |
59 | using ceph::decode_nohead; | |
60 | using ceph::encode; | |
61 | using ceph::encode_nohead; | |
62 | using ceph::Formatter; | |
f67539c2 TL |
63 | using ceph::make_timespan; |
64 | using ceph::JSONFormatter; | |
9f95a23c TL |
65 | |
66 | using namespace std::literals; | |
7c673cae FG |
67 | |
68 | const char *ceph_osd_flag_name(unsigned flag) | |
69 | { | |
70 | switch (flag) { | |
71 | case CEPH_OSD_FLAG_ACK: return "ack"; | |
72 | case CEPH_OSD_FLAG_ONNVRAM: return "onnvram"; | |
73 | case CEPH_OSD_FLAG_ONDISK: return "ondisk"; | |
74 | case CEPH_OSD_FLAG_RETRY: return "retry"; | |
75 | case CEPH_OSD_FLAG_READ: return "read"; | |
76 | case CEPH_OSD_FLAG_WRITE: return "write"; | |
77 | case CEPH_OSD_FLAG_ORDERSNAP: return "ordersnap"; | |
78 | case CEPH_OSD_FLAG_PEERSTAT_OLD: return "peerstat_old"; | |
79 | case CEPH_OSD_FLAG_BALANCE_READS: return "balance_reads"; | |
80 | case CEPH_OSD_FLAG_PARALLELEXEC: return "parallelexec"; | |
81 | case CEPH_OSD_FLAG_PGOP: return "pgop"; | |
82 | case CEPH_OSD_FLAG_EXEC: return "exec"; | |
83 | case CEPH_OSD_FLAG_EXEC_PUBLIC: return "exec_public"; | |
84 | case CEPH_OSD_FLAG_LOCALIZE_READS: return "localize_reads"; | |
85 | case CEPH_OSD_FLAG_RWORDERED: return "rwordered"; | |
86 | case CEPH_OSD_FLAG_IGNORE_CACHE: return "ignore_cache"; | |
87 | case CEPH_OSD_FLAG_SKIPRWLOCKS: return "skiprwlocks"; | |
88 | case CEPH_OSD_FLAG_IGNORE_OVERLAY: return "ignore_overlay"; | |
89 | case CEPH_OSD_FLAG_FLUSH: return "flush"; | |
90 | case CEPH_OSD_FLAG_MAP_SNAP_CLONE: return "map_snap_clone"; | |
91 | case CEPH_OSD_FLAG_ENFORCE_SNAPC: return "enforce_snapc"; | |
92 | case CEPH_OSD_FLAG_REDIRECTED: return "redirected"; | |
93 | case CEPH_OSD_FLAG_KNOWN_REDIR: return "known_if_redirected"; | |
94 | case CEPH_OSD_FLAG_FULL_TRY: return "full_try"; | |
95 | case CEPH_OSD_FLAG_FULL_FORCE: return "full_force"; | |
224ce89b | 96 | case CEPH_OSD_FLAG_IGNORE_REDIRECT: return "ignore_redirect"; |
9f95a23c | 97 | case CEPH_OSD_FLAG_RETURNVEC: return "returnvec"; |
20effc67 | 98 | case CEPH_OSD_FLAG_SUPPORTSPOOLEIO: return "supports_pool_eio"; |
7c673cae FG |
99 | default: return "???"; |
100 | } | |
101 | } | |
102 | ||
103 | string ceph_osd_flag_string(unsigned flags) | |
104 | { | |
105 | string s; | |
106 | for (unsigned i=0; i<32; ++i) { | |
107 | if (flags & (1u<<i)) { | |
108 | if (s.length()) | |
109 | s += "+"; | |
110 | s += ceph_osd_flag_name(1u << i); | |
111 | } | |
112 | } | |
113 | if (s.length()) | |
114 | return s; | |
115 | return string("-"); | |
116 | } | |
117 | ||
118 | const char * ceph_osd_op_flag_name(unsigned flag) | |
119 | { | |
120 | const char *name; | |
121 | ||
122 | switch(flag) { | |
123 | case CEPH_OSD_OP_FLAG_EXCL: | |
124 | name = "excl"; | |
125 | break; | |
126 | case CEPH_OSD_OP_FLAG_FAILOK: | |
127 | name = "failok"; | |
128 | break; | |
129 | case CEPH_OSD_OP_FLAG_FADVISE_RANDOM: | |
130 | name = "fadvise_random"; | |
131 | break; | |
132 | case CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL: | |
133 | name = "fadvise_sequential"; | |
134 | break; | |
135 | case CEPH_OSD_OP_FLAG_FADVISE_WILLNEED: | |
136 | name = "favise_willneed"; | |
137 | break; | |
138 | case CEPH_OSD_OP_FLAG_FADVISE_DONTNEED: | |
139 | name = "fadvise_dontneed"; | |
140 | break; | |
141 | case CEPH_OSD_OP_FLAG_FADVISE_NOCACHE: | |
142 | name = "fadvise_nocache"; | |
143 | break; | |
11fdf7f2 TL |
144 | case CEPH_OSD_OP_FLAG_WITH_REFERENCE: |
145 | name = "with_reference"; | |
146 | break; | |
91327a77 AA |
147 | case CEPH_OSD_OP_FLAG_BYPASS_CLEAN_CACHE: |
148 | name = "bypass_clean_cache"; | |
149 | break; | |
7c673cae FG |
150 | default: |
151 | name = "???"; | |
152 | }; | |
153 | ||
154 | return name; | |
155 | } | |
156 | ||
157 | string ceph_osd_op_flag_string(unsigned flags) | |
158 | { | |
159 | string s; | |
160 | for (unsigned i=0; i<32; ++i) { | |
161 | if (flags & (1u<<i)) { | |
162 | if (s.length()) | |
163 | s += "+"; | |
164 | s += ceph_osd_op_flag_name(1u << i); | |
165 | } | |
166 | } | |
167 | if (s.length()) | |
168 | return s; | |
169 | return string("-"); | |
170 | } | |
171 | ||
172 | string ceph_osd_alloc_hint_flag_string(unsigned flags) | |
173 | { | |
174 | string s; | |
175 | for (unsigned i=0; i<32; ++i) { | |
176 | if (flags & (1u<<i)) { | |
177 | if (s.length()) | |
178 | s += "+"; | |
179 | s += ceph_osd_alloc_hint_flag_name(1u << i); | |
180 | } | |
181 | } | |
182 | if (s.length()) | |
183 | return s; | |
184 | return string("-"); | |
185 | } | |
186 | ||
9f95a23c | 187 | void pg_shard_t::encode(ceph::buffer::list &bl) const |
7c673cae FG |
188 | { |
189 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
190 | encode(osd, bl); |
191 | encode(shard, bl); | |
7c673cae FG |
192 | ENCODE_FINISH(bl); |
193 | } | |
9f95a23c | 194 | void pg_shard_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
195 | { |
196 | DECODE_START(1, bl); | |
11fdf7f2 TL |
197 | decode(osd, bl); |
198 | decode(shard, bl); | |
7c673cae FG |
199 | DECODE_FINISH(bl); |
200 | } | |
201 | ||
202 | ostream &operator<<(ostream &lhs, const pg_shard_t &rhs) | |
203 | { | |
204 | if (rhs.is_undefined()) | |
205 | return lhs << "?"; | |
206 | if (rhs.shard == shard_id_t::NO_SHARD) | |
b32b8144 FG |
207 | return lhs << rhs.get_osd(); |
208 | return lhs << rhs.get_osd() << '(' << (unsigned)(rhs.shard) << ')'; | |
7c673cae FG |
209 | } |
210 | ||
11fdf7f2 TL |
211 | void dump(Formatter* f, const osd_alerts_t& alerts) |
212 | { | |
213 | for (auto& a : alerts) { | |
214 | string s0 = " osd: "; | |
215 | s0 += stringify(a.first); | |
216 | string s; | |
217 | for (auto& aa : a.second) { | |
218 | s = s0; | |
219 | s += " "; | |
220 | s += aa.first; | |
221 | s += ":"; | |
222 | s += aa.second; | |
223 | f->dump_string("alert", s); | |
224 | } | |
225 | } | |
226 | } | |
227 | ||
7c673cae FG |
228 | // -- osd_reqid_t -- |
229 | void osd_reqid_t::dump(Formatter *f) const | |
230 | { | |
231 | f->dump_stream("name") << name; | |
232 | f->dump_int("inc", inc); | |
233 | f->dump_unsigned("tid", tid); | |
234 | } | |
235 | ||
236 | void osd_reqid_t::generate_test_instances(list<osd_reqid_t*>& o) | |
237 | { | |
238 | o.push_back(new osd_reqid_t); | |
239 | o.push_back(new osd_reqid_t(entity_name_t::CLIENT(123), 1, 45678)); | |
240 | } | |
241 | ||
242 | // -- object_locator_t -- | |
243 | ||
9f95a23c | 244 | void object_locator_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
245 | { |
246 | // verify that nobody's corrupted the locator | |
11fdf7f2 | 247 | ceph_assert(hash == -1 || key.empty()); |
7c673cae FG |
248 | __u8 encode_compat = 3; |
249 | ENCODE_START(6, encode_compat, bl); | |
11fdf7f2 | 250 | encode(pool, bl); |
7c673cae | 251 | int32_t preferred = -1; // tell old code there is no preferred osd (-1). |
11fdf7f2 TL |
252 | encode(preferred, bl); |
253 | encode(key, bl); | |
254 | encode(nspace, bl); | |
255 | encode(hash, bl); | |
7c673cae | 256 | if (hash != -1) |
11fdf7f2 | 257 | encode_compat = std::max<std::uint8_t>(encode_compat, 6); // need to interpret the hash |
7c673cae FG |
258 | ENCODE_FINISH_NEW_COMPAT(bl, encode_compat); |
259 | } | |
260 | ||
9f95a23c | 261 | void object_locator_t::decode(ceph::buffer::list::const_iterator& p) |
7c673cae FG |
262 | { |
263 | DECODE_START_LEGACY_COMPAT_LEN(6, 3, 3, p); | |
264 | if (struct_v < 2) { | |
265 | int32_t op; | |
11fdf7f2 | 266 | decode(op, p); |
7c673cae FG |
267 | pool = op; |
268 | int16_t pref; | |
11fdf7f2 | 269 | decode(pref, p); |
7c673cae | 270 | } else { |
11fdf7f2 | 271 | decode(pool, p); |
7c673cae | 272 | int32_t preferred; |
11fdf7f2 | 273 | decode(preferred, p); |
7c673cae | 274 | } |
11fdf7f2 | 275 | decode(key, p); |
7c673cae | 276 | if (struct_v >= 5) |
11fdf7f2 | 277 | decode(nspace, p); |
7c673cae | 278 | if (struct_v >= 6) |
11fdf7f2 | 279 | decode(hash, p); |
7c673cae FG |
280 | else |
281 | hash = -1; | |
282 | DECODE_FINISH(p); | |
283 | // verify that nobody's corrupted the locator | |
11fdf7f2 | 284 | ceph_assert(hash == -1 || key.empty()); |
7c673cae FG |
285 | } |
286 | ||
287 | void object_locator_t::dump(Formatter *f) const | |
288 | { | |
289 | f->dump_int("pool", pool); | |
290 | f->dump_string("key", key); | |
291 | f->dump_string("namespace", nspace); | |
292 | f->dump_int("hash", hash); | |
293 | } | |
294 | ||
295 | void object_locator_t::generate_test_instances(list<object_locator_t*>& o) | |
296 | { | |
297 | o.push_back(new object_locator_t); | |
298 | o.push_back(new object_locator_t(123)); | |
299 | o.push_back(new object_locator_t(123, 876)); | |
300 | o.push_back(new object_locator_t(1, "n2")); | |
301 | o.push_back(new object_locator_t(1234, "", "key")); | |
302 | o.push_back(new object_locator_t(12, "n1", "key2")); | |
303 | } | |
304 | ||
305 | // -- request_redirect_t -- | |
9f95a23c | 306 | void request_redirect_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
307 | { |
308 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
309 | encode(redirect_locator, bl); |
310 | encode(redirect_object, bl); | |
311 | // legacy of the removed osd_instructions member | |
312 | encode((uint32_t)0, bl); | |
7c673cae FG |
313 | ENCODE_FINISH(bl); |
314 | } | |
315 | ||
9f95a23c | 316 | void request_redirect_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
317 | { |
318 | DECODE_START(1, bl); | |
11fdf7f2 TL |
319 | uint32_t legacy_osd_instructions_len; |
320 | decode(redirect_locator, bl); | |
321 | decode(redirect_object, bl); | |
322 | decode(legacy_osd_instructions_len, bl); | |
323 | if (legacy_osd_instructions_len) { | |
9f95a23c | 324 | bl += legacy_osd_instructions_len; |
11fdf7f2 | 325 | } |
7c673cae FG |
326 | DECODE_FINISH(bl); |
327 | } | |
328 | ||
329 | void request_redirect_t::dump(Formatter *f) const | |
330 | { | |
331 | f->dump_string("object", redirect_object); | |
332 | f->open_object_section("locator"); | |
333 | redirect_locator.dump(f); | |
334 | f->close_section(); // locator | |
335 | } | |
336 | ||
337 | void request_redirect_t::generate_test_instances(list<request_redirect_t*>& o) | |
338 | { | |
339 | object_locator_t loc(1, "redir_obj"); | |
340 | o.push_back(new request_redirect_t()); | |
341 | o.push_back(new request_redirect_t(loc, 0)); | |
342 | o.push_back(new request_redirect_t(loc, "redir_obj")); | |
343 | o.push_back(new request_redirect_t(loc)); | |
344 | } | |
345 | ||
346 | void objectstore_perf_stat_t::dump(Formatter *f) const | |
347 | { | |
11fdf7f2 TL |
348 | // *_ms values just for compatibility. |
349 | f->dump_float("commit_latency_ms", os_commit_latency_ns / 1000000.0); | |
350 | f->dump_float("apply_latency_ms", os_apply_latency_ns / 1000000.0); | |
351 | f->dump_unsigned("commit_latency_ns", os_commit_latency_ns); | |
352 | f->dump_unsigned("apply_latency_ns", os_apply_latency_ns); | |
7c673cae FG |
353 | } |
354 | ||
9f95a23c | 355 | void objectstore_perf_stat_t::encode(ceph::buffer::list &bl, uint64_t features) const |
7c673cae | 356 | { |
11fdf7f2 TL |
357 | uint8_t target_v = 2; |
358 | if (!HAVE_FEATURE(features, OS_PERF_STAT_NS)) { | |
359 | target_v = 1; | |
360 | } | |
361 | ENCODE_START(target_v, target_v, bl); | |
362 | if (target_v >= 2) { | |
363 | encode(os_commit_latency_ns, bl); | |
364 | encode(os_apply_latency_ns, bl); | |
365 | } else { | |
366 | constexpr auto NS_PER_MS = std::chrono::nanoseconds(1ms).count(); | |
367 | uint32_t commit_latency_ms = os_commit_latency_ns / NS_PER_MS; | |
368 | uint32_t apply_latency_ms = os_apply_latency_ns / NS_PER_MS; | |
369 | encode(commit_latency_ms, bl); // for compatibility with older monitor. | |
370 | encode(apply_latency_ms, bl); // for compatibility with older monitor. | |
371 | } | |
7c673cae FG |
372 | ENCODE_FINISH(bl); |
373 | } | |
374 | ||
9f95a23c | 375 | void objectstore_perf_stat_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 376 | { |
11fdf7f2 TL |
377 | DECODE_START(2, bl); |
378 | if (struct_v >= 2) { | |
379 | decode(os_commit_latency_ns, bl); | |
380 | decode(os_apply_latency_ns, bl); | |
381 | } else { | |
382 | uint32_t commit_latency_ms; | |
383 | uint32_t apply_latency_ms; | |
384 | decode(commit_latency_ms, bl); | |
385 | decode(apply_latency_ms, bl); | |
386 | constexpr auto NS_PER_MS = std::chrono::nanoseconds(1ms).count(); | |
387 | os_commit_latency_ns = commit_latency_ms * NS_PER_MS; | |
388 | os_apply_latency_ns = apply_latency_ms * NS_PER_MS; | |
389 | } | |
7c673cae FG |
390 | DECODE_FINISH(bl); |
391 | } | |
392 | ||
393 | void objectstore_perf_stat_t::generate_test_instances(std::list<objectstore_perf_stat_t*>& o) | |
394 | { | |
395 | o.push_back(new objectstore_perf_stat_t()); | |
396 | o.push_back(new objectstore_perf_stat_t()); | |
11fdf7f2 TL |
397 | o.back()->os_commit_latency_ns = 20000000; |
398 | o.back()->os_apply_latency_ns = 30000000; | |
7c673cae FG |
399 | } |
400 | ||
401 | // -- osd_stat_t -- | |
ded94939 | 402 | void osd_stat_t::dump(Formatter *f, bool with_net) const |
7c673cae | 403 | { |
31f18b77 FG |
404 | f->dump_unsigned("up_from", up_from); |
405 | f->dump_unsigned("seq", seq); | |
35e4c445 | 406 | f->dump_unsigned("num_pgs", num_pgs); |
81eedcae TL |
407 | f->dump_unsigned("num_osds", num_osds); |
408 | f->dump_unsigned("num_per_pool_osds", num_per_pool_osds); | |
9f95a23c | 409 | f->dump_unsigned("num_per_pool_omap_osds", num_per_pool_omap_osds); |
11fdf7f2 TL |
410 | |
411 | /// dump legacy stats fields to ensure backward compatibility. | |
412 | f->dump_unsigned("kb", statfs.kb()); | |
413 | f->dump_unsigned("kb_used", statfs.kb_used_raw()); | |
414 | f->dump_unsigned("kb_used_data", statfs.kb_used_data()); | |
415 | f->dump_unsigned("kb_used_omap", statfs.kb_used_omap()); | |
416 | f->dump_unsigned("kb_used_meta", statfs.kb_used_internal_metadata()); | |
417 | f->dump_unsigned("kb_avail", statfs.kb_avail()); | |
418 | //////////////////// | |
419 | ||
420 | f->open_object_section("statfs"); | |
421 | statfs.dump(f); | |
422 | f->close_section(); | |
7c673cae FG |
423 | f->open_array_section("hb_peers"); |
424 | for (auto p : hb_peers) | |
425 | f->dump_int("osd", p); | |
426 | f->close_section(); | |
427 | f->dump_int("snap_trim_queue_len", snap_trim_queue_len); | |
428 | f->dump_int("num_snap_trimming", num_snap_trimming); | |
11fdf7f2 | 429 | f->dump_int("num_shards_repaired", num_shards_repaired); |
7c673cae FG |
430 | f->open_object_section("op_queue_age_hist"); |
431 | op_queue_age_hist.dump(f); | |
432 | f->close_section(); | |
433 | f->open_object_section("perf_stat"); | |
434 | os_perf_stat.dump(f); | |
435 | f->close_section(); | |
11fdf7f2 TL |
436 | f->open_array_section("alerts"); |
437 | ::dump(f, os_alerts); | |
438 | f->close_section(); | |
ded94939 | 439 | if (with_net) { |
9f95a23c TL |
440 | dump_ping_time(f); |
441 | } | |
442 | } | |
443 | ||
444 | void osd_stat_t::dump_ping_time(Formatter *f) const | |
445 | { | |
eafe8130 TL |
446 | f->open_array_section("network_ping_times"); |
447 | for (auto &i : hb_pingtime) { | |
448 | f->open_object_section("entry"); | |
449 | f->dump_int("osd", i.first); | |
450 | const time_t lu(i.second.last_update); | |
451 | char buffer[26]; | |
452 | string lustr(ctime_r(&lu, buffer)); | |
453 | lustr.pop_back(); // Remove trailing \n | |
454 | f->dump_string("last update", lustr); | |
455 | f->open_array_section("interfaces"); | |
456 | f->open_object_section("interface"); | |
457 | f->dump_string("interface", "back"); | |
458 | f->open_object_section("average"); | |
9f95a23c TL |
459 | f->dump_float("1min", i.second.back_pingtime[0]/1000.0); |
460 | f->dump_float("5min", i.second.back_pingtime[1]/1000.0); | |
461 | f->dump_float("15min", i.second.back_pingtime[2]/1000.0); | |
eafe8130 TL |
462 | f->close_section(); // average |
463 | f->open_object_section("min"); | |
9f95a23c TL |
464 | f->dump_float("1min", i.second.back_min[0]/1000.0); |
465 | f->dump_float("5min", i.second.back_min[1]/1000.0); | |
466 | f->dump_float("15min", i.second.back_min[2]/1000.0); | |
eafe8130 TL |
467 | f->close_section(); // min |
468 | f->open_object_section("max"); | |
9f95a23c TL |
469 | f->dump_float("1min", i.second.back_max[0]/1000.0); |
470 | f->dump_float("5min", i.second.back_max[1]/1000.0); | |
471 | f->dump_float("15min", i.second.back_max[2]/1000.0); | |
eafe8130 | 472 | f->close_section(); // max |
9f95a23c | 473 | f->dump_float("last", i.second.back_last/1000.0); |
eafe8130 TL |
474 | f->close_section(); // interface |
475 | ||
476 | if (i.second.front_pingtime[0] != 0) { | |
477 | f->open_object_section("interface"); | |
478 | f->dump_string("interface", "front"); | |
479 | f->open_object_section("average"); | |
9f95a23c TL |
480 | f->dump_float("1min", i.second.front_pingtime[0]/1000.0); |
481 | f->dump_float("5min", i.second.front_pingtime[1]/1000.0); | |
482 | f->dump_float("15min", i.second.front_pingtime[2]/1000.0); | |
eafe8130 TL |
483 | f->close_section(); // average |
484 | f->open_object_section("min"); | |
9f95a23c TL |
485 | f->dump_float("1min", i.second.front_min[0]/1000.0); |
486 | f->dump_float("5min", i.second.front_min[1]/1000.0); | |
487 | f->dump_float("15min", i.second.front_min[2]/1000.0); | |
eafe8130 TL |
488 | f->close_section(); // min |
489 | f->open_object_section("max"); | |
9f95a23c TL |
490 | f->dump_float("1min", i.second.front_max[0]/1000.0); |
491 | f->dump_float("5min", i.second.front_max[1]/1000.0); | |
492 | f->dump_float("15min", i.second.front_max[2]/1000.0); | |
eafe8130 | 493 | f->close_section(); // max |
9f95a23c | 494 | f->dump_float("last", i.second.front_last/1000.0); |
eafe8130 TL |
495 | f->close_section(); // interface |
496 | } | |
497 | f->close_section(); // interfaces | |
498 | f->close_section(); // entry | |
499 | } | |
500 | f->close_section(); // network_ping_time | |
7c673cae FG |
501 | } |
502 | ||
9f95a23c | 503 | void osd_stat_t::encode(ceph::buffer::list &bl, uint64_t features) const |
11fdf7f2 | 504 | { |
eafe8130 | 505 | ENCODE_START(14, 2, bl); |
11fdf7f2 TL |
506 | |
507 | //////// for compatibility //////// | |
508 | int64_t kb = statfs.kb(); | |
509 | int64_t kb_used = statfs.kb_used_raw(); | |
510 | int64_t kb_avail = statfs.kb_avail(); | |
511 | encode(kb, bl); | |
512 | encode(kb_used, bl); | |
513 | encode(kb_avail, bl); | |
514 | /////////////////////////////////// | |
515 | ||
516 | encode(snap_trim_queue_len, bl); | |
517 | encode(num_snap_trimming, bl); | |
518 | encode(hb_peers, bl); | |
519 | encode((uint32_t)0, bl); | |
520 | encode(op_queue_age_hist, bl); | |
521 | encode(os_perf_stat, bl, features); | |
522 | encode(up_from, bl); | |
523 | encode(seq, bl); | |
524 | encode(num_pgs, bl); | |
525 | ||
526 | //////// for compatibility //////// | |
527 | int64_t kb_used_data = statfs.kb_used_data(); | |
528 | int64_t kb_used_omap = statfs.kb_used_omap(); | |
529 | int64_t kb_used_meta = statfs.kb_used_internal_metadata(); | |
530 | encode(kb_used_data, bl); | |
531 | encode(kb_used_omap, bl); | |
532 | encode(kb_used_meta, bl); | |
533 | encode(statfs, bl); | |
534 | /////////////////////////////////// | |
535 | encode(os_alerts, bl); | |
536 | encode(num_shards_repaired, bl); | |
81eedcae TL |
537 | encode(num_osds, bl); |
538 | encode(num_per_pool_osds, bl); | |
9f95a23c | 539 | encode(num_per_pool_omap_osds, bl); |
eafe8130 TL |
540 | |
541 | // hb_pingtime map | |
542 | encode((int)hb_pingtime.size(), bl); | |
543 | for (auto i : hb_pingtime) { | |
544 | encode(i.first, bl); // osd | |
545 | encode(i.second.last_update, bl); | |
546 | encode(i.second.back_pingtime[0], bl); | |
547 | encode(i.second.back_pingtime[1], bl); | |
548 | encode(i.second.back_pingtime[2], bl); | |
549 | encode(i.second.back_min[0], bl); | |
550 | encode(i.second.back_min[1], bl); | |
551 | encode(i.second.back_min[2], bl); | |
552 | encode(i.second.back_max[0], bl); | |
553 | encode(i.second.back_max[1], bl); | |
554 | encode(i.second.back_max[2], bl); | |
555 | encode(i.second.back_last, bl); | |
556 | encode(i.second.front_pingtime[0], bl); | |
557 | encode(i.second.front_pingtime[1], bl); | |
558 | encode(i.second.front_pingtime[2], bl); | |
559 | encode(i.second.front_min[0], bl); | |
560 | encode(i.second.front_min[1], bl); | |
561 | encode(i.second.front_min[2], bl); | |
562 | encode(i.second.front_max[0], bl); | |
563 | encode(i.second.front_max[1], bl); | |
564 | encode(i.second.front_max[2], bl); | |
565 | encode(i.second.front_last, bl); | |
566 | } | |
7c673cae FG |
567 | ENCODE_FINISH(bl); |
568 | } | |
569 | ||
9f95a23c | 570 | void osd_stat_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 571 | { |
11fdf7f2 TL |
572 | int64_t kb, kb_used,kb_avail; |
573 | int64_t kb_used_data, kb_used_omap, kb_used_meta; | |
eafe8130 | 574 | DECODE_START_LEGACY_COMPAT_LEN(14, 2, 2, bl); |
11fdf7f2 TL |
575 | decode(kb, bl); |
576 | decode(kb_used, bl); | |
577 | decode(kb_avail, bl); | |
578 | decode(snap_trim_queue_len, bl); | |
579 | decode(num_snap_trimming, bl); | |
580 | decode(hb_peers, bl); | |
7c673cae | 581 | vector<int> num_hb_out; |
11fdf7f2 | 582 | decode(num_hb_out, bl); |
7c673cae | 583 | if (struct_v >= 3) |
11fdf7f2 | 584 | decode(op_queue_age_hist, bl); |
7c673cae | 585 | if (struct_v >= 4) |
11fdf7f2 | 586 | decode(os_perf_stat, bl); |
31f18b77 | 587 | if (struct_v >= 6) { |
11fdf7f2 TL |
588 | decode(up_from, bl); |
589 | decode(seq, bl); | |
31f18b77 | 590 | } |
35e4c445 | 591 | if (struct_v >= 7) { |
11fdf7f2 TL |
592 | decode(num_pgs, bl); |
593 | } | |
594 | if (struct_v >= 8) { | |
595 | decode(kb_used_data, bl); | |
596 | decode(kb_used_omap, bl); | |
597 | decode(kb_used_meta, bl); | |
598 | } else { | |
599 | kb_used_data = kb_used; | |
600 | kb_used_omap = 0; | |
601 | kb_used_meta = 0; | |
602 | } | |
603 | if (struct_v >= 9) { | |
604 | decode(statfs, bl); | |
605 | } else { | |
606 | statfs.reset(); | |
607 | statfs.total = kb << 10; | |
608 | statfs.available = kb_avail << 10; | |
609 | // actually it's totally unexpected to have ststfs.total < statfs.available | |
610 | // here but unfortunately legacy generate_test_instances produced such a | |
611 | // case hence inserting some handling rather than assert | |
612 | statfs.internally_reserved = | |
613 | statfs.total > statfs.available ? statfs.total - statfs.available : 0; | |
614 | kb_used <<= 10; | |
615 | if ((int64_t)statfs.internally_reserved > kb_used) { | |
616 | statfs.internally_reserved -= kb_used; | |
617 | } else { | |
618 | statfs.internally_reserved = 0; | |
619 | } | |
620 | statfs.allocated = kb_used_data << 10; | |
621 | statfs.omap_allocated = kb_used_omap << 10; | |
622 | statfs.internal_metadata = kb_used_meta << 10; | |
623 | } | |
624 | if (struct_v >= 10) { | |
625 | decode(os_alerts, bl); | |
626 | } else { | |
627 | os_alerts.clear(); | |
628 | } | |
629 | if (struct_v >= 11) { | |
630 | decode(num_shards_repaired, bl); | |
631 | } else { | |
632 | num_shards_repaired = 0; | |
35e4c445 | 633 | } |
81eedcae TL |
634 | if (struct_v >= 12) { |
635 | decode(num_osds, bl); | |
636 | decode(num_per_pool_osds, bl); | |
637 | } else { | |
638 | num_osds = 0; | |
639 | num_per_pool_osds = 0; | |
640 | } | |
eafe8130 | 641 | if (struct_v >= 13) { |
9f95a23c TL |
642 | decode(num_per_pool_omap_osds, bl); |
643 | } else { | |
644 | num_per_pool_omap_osds = 0; | |
eafe8130 TL |
645 | } |
646 | hb_pingtime.clear(); | |
647 | if (struct_v >= 14) { | |
648 | int count; | |
649 | decode(count, bl); | |
650 | for (int i = 0 ; i < count ; i++) { | |
651 | int osd; | |
652 | decode(osd, bl); | |
653 | struct Interfaces ifs; | |
654 | decode(ifs.last_update, bl); | |
655 | decode(ifs.back_pingtime[0],bl); | |
656 | decode(ifs.back_pingtime[1], bl); | |
657 | decode(ifs.back_pingtime[2], bl); | |
658 | decode(ifs.back_min[0],bl); | |
659 | decode(ifs.back_min[1], bl); | |
660 | decode(ifs.back_min[2], bl); | |
661 | decode(ifs.back_max[0],bl); | |
662 | decode(ifs.back_max[1], bl); | |
663 | decode(ifs.back_max[2], bl); | |
664 | decode(ifs.back_last, bl); | |
665 | decode(ifs.front_pingtime[0], bl); | |
666 | decode(ifs.front_pingtime[1], bl); | |
667 | decode(ifs.front_pingtime[2], bl); | |
668 | decode(ifs.front_min[0], bl); | |
669 | decode(ifs.front_min[1], bl); | |
670 | decode(ifs.front_min[2], bl); | |
671 | decode(ifs.front_max[0], bl); | |
672 | decode(ifs.front_max[1], bl); | |
673 | decode(ifs.front_max[2], bl); | |
674 | decode(ifs.front_last, bl); | |
675 | hb_pingtime[osd] = ifs; | |
676 | } | |
677 | } | |
7c673cae FG |
678 | DECODE_FINISH(bl); |
679 | } | |
680 | ||
681 | void osd_stat_t::generate_test_instances(std::list<osd_stat_t*>& o) | |
682 | { | |
683 | o.push_back(new osd_stat_t); | |
684 | ||
685 | o.push_back(new osd_stat_t); | |
11fdf7f2 TL |
686 | list<store_statfs_t*> ll; |
687 | store_statfs_t::generate_test_instances(ll); | |
688 | o.back()->statfs = *ll.back(); | |
7c673cae FG |
689 | o.back()->hb_peers.push_back(7); |
690 | o.back()->snap_trim_queue_len = 8; | |
691 | o.back()->num_snap_trimming = 99; | |
11fdf7f2 TL |
692 | o.back()->num_shards_repaired = 101; |
693 | o.back()->os_alerts[0].emplace( | |
694 | "some alert", "some alert details"); | |
695 | o.back()->os_alerts[1].emplace( | |
696 | "some alert2", "some alert2 details"); | |
eafe8130 TL |
697 | struct Interfaces gen_interfaces = { |
698 | 123456789, { 1000, 900, 800 }, { 990, 890, 790 }, { 1010, 910, 810 }, 1001, | |
699 | { 1100, 1000, 900 }, { 1090, 990, 890 }, { 1110, 1010, 910 }, 1101 }; | |
700 | o.back()->hb_pingtime[20] = gen_interfaces; | |
701 | gen_interfaces = { | |
702 | 987654321, { 100, 200, 300 }, { 90, 190, 290 }, { 110, 210, 310 }, 101 }; | |
703 | o.back()->hb_pingtime[30] = gen_interfaces; | |
7c673cae FG |
704 | } |
705 | ||
706 | // -- pg_t -- | |
707 | ||
708 | int pg_t::print(char *o, int maxlen) const | |
709 | { | |
11fdf7f2 | 710 | return snprintf(o, maxlen, "%llu.%x", (unsigned long long)pool(), ps()); |
7c673cae FG |
711 | } |
712 | ||
713 | bool pg_t::parse(const char *s) | |
714 | { | |
715 | uint64_t ppool; | |
716 | uint32_t pseed; | |
11fdf7f2 | 717 | int r = sscanf(s, "%llu.%x", (long long unsigned *)&ppool, &pseed); |
7c673cae FG |
718 | if (r < 2) |
719 | return false; | |
720 | m_pool = ppool; | |
721 | m_seed = pseed; | |
7c673cae FG |
722 | return true; |
723 | } | |
724 | ||
725 | bool spg_t::parse(const char *s) | |
726 | { | |
7c673cae FG |
727 | shard = shard_id_t::NO_SHARD; |
728 | uint64_t ppool; | |
729 | uint32_t pseed; | |
7c673cae FG |
730 | uint32_t pshard; |
731 | int r = sscanf(s, "%llu.%x", (long long unsigned *)&ppool, &pseed); | |
732 | if (r < 2) | |
733 | return false; | |
734 | pgid.set_pool(ppool); | |
735 | pgid.set_ps(pseed); | |
736 | ||
11fdf7f2 | 737 | const char *p = strchr(s, 's'); |
7c673cae | 738 | if (p) { |
11fdf7f2 | 739 | r = sscanf(p, "s%u", &pshard); |
7c673cae FG |
740 | if (r == 1) { |
741 | shard = shard_id_t(pshard); | |
742 | } else { | |
743 | return false; | |
744 | } | |
745 | } | |
746 | return true; | |
747 | } | |
748 | ||
749 | char *spg_t::calc_name(char *buf, const char *suffix_backwords) const | |
750 | { | |
751 | while (*suffix_backwords) | |
752 | *--buf = *suffix_backwords++; | |
753 | ||
754 | if (!is_no_shard()) { | |
755 | buf = ritoa<uint8_t, 10>((uint8_t)shard.id, buf); | |
756 | *--buf = 's'; | |
757 | } | |
758 | ||
759 | return pgid.calc_name(buf, ""); | |
760 | } | |
761 | ||
1e59de90 TL |
762 | std::string spg_t::calc_name_sring() const |
763 | { | |
764 | char buf[spg_t::calc_name_buf_size]; | |
765 | buf[spg_t::calc_name_buf_size - 1] = '\0'; | |
766 | return string{calc_name(buf + spg_t::calc_name_buf_size - 1, "")}; | |
767 | } | |
768 | ||
7c673cae FG |
769 | ostream& operator<<(ostream& out, const spg_t &pg) |
770 | { | |
771 | char buf[spg_t::calc_name_buf_size]; | |
772 | buf[spg_t::calc_name_buf_size - 1] = '\0'; | |
773 | out << pg.calc_name(buf + spg_t::calc_name_buf_size - 1, ""); | |
774 | return out; | |
775 | } | |
776 | ||
777 | pg_t pg_t::get_ancestor(unsigned old_pg_num) const | |
778 | { | |
779 | int old_bits = cbits(old_pg_num); | |
780 | int old_mask = (1 << old_bits) - 1; | |
781 | pg_t ret = *this; | |
782 | ret.m_seed = ceph_stable_mod(m_seed, old_pg_num, old_mask); | |
783 | return ret; | |
784 | } | |
785 | ||
786 | bool pg_t::is_split(unsigned old_pg_num, unsigned new_pg_num, set<pg_t> *children) const | |
787 | { | |
11fdf7f2 TL |
788 | //ceph_assert(m_seed < old_pg_num); |
789 | if (m_seed >= old_pg_num) { | |
790 | // degenerate case | |
791 | return false; | |
792 | } | |
7c673cae FG |
793 | if (new_pg_num <= old_pg_num) |
794 | return false; | |
795 | ||
796 | bool split = false; | |
797 | if (true) { | |
798 | unsigned old_bits = cbits(old_pg_num); | |
799 | unsigned old_mask = (1 << old_bits) - 1; | |
800 | for (unsigned n = 1; ; n++) { | |
801 | unsigned next_bit = (n << (old_bits-1)); | |
802 | unsigned s = next_bit | m_seed; | |
803 | ||
804 | if (s < old_pg_num || s == m_seed) | |
805 | continue; | |
806 | if (s >= new_pg_num) | |
807 | break; | |
808 | if ((unsigned)ceph_stable_mod(s, old_pg_num, old_mask) == m_seed) { | |
809 | split = true; | |
810 | if (children) | |
11fdf7f2 | 811 | children->insert(pg_t(s, m_pool)); |
7c673cae FG |
812 | } |
813 | } | |
814 | } | |
815 | if (false) { | |
816 | // brute force | |
817 | int old_bits = cbits(old_pg_num); | |
818 | int old_mask = (1 << old_bits) - 1; | |
819 | for (unsigned x = old_pg_num; x < new_pg_num; ++x) { | |
820 | unsigned o = ceph_stable_mod(x, old_pg_num, old_mask); | |
821 | if (o == m_seed) { | |
822 | split = true; | |
11fdf7f2 | 823 | children->insert(pg_t(x, m_pool)); |
7c673cae FG |
824 | } |
825 | } | |
826 | } | |
827 | return split; | |
828 | } | |
829 | ||
830 | unsigned pg_t::get_split_bits(unsigned pg_num) const { | |
831 | if (pg_num == 1) | |
832 | return 0; | |
11fdf7f2 | 833 | ceph_assert(pg_num > 1); |
7c673cae FG |
834 | |
835 | // Find unique p such that pg_num \in [2^(p-1), 2^p) | |
836 | unsigned p = cbits(pg_num); | |
11fdf7f2 | 837 | ceph_assert(p); // silence coverity #751330 |
7c673cae FG |
838 | |
839 | if ((m_seed % (1<<(p-1))) < (pg_num % (1<<(p-1)))) | |
840 | return p; | |
841 | else | |
842 | return p - 1; | |
843 | } | |
844 | ||
11fdf7f2 TL |
845 | bool pg_t::is_merge_source( |
846 | unsigned old_pg_num, | |
847 | unsigned new_pg_num, | |
848 | pg_t *parent) const | |
849 | { | |
850 | if (m_seed < old_pg_num && | |
851 | m_seed >= new_pg_num) { | |
852 | if (parent) { | |
853 | pg_t t = *this; | |
854 | while (t.m_seed >= new_pg_num) { | |
855 | t = t.get_parent(); | |
856 | } | |
857 | *parent = t; | |
858 | } | |
859 | return true; | |
860 | } | |
861 | return false; | |
862 | } | |
863 | ||
7c673cae FG |
864 | pg_t pg_t::get_parent() const |
865 | { | |
866 | unsigned bits = cbits(m_seed); | |
11fdf7f2 | 867 | ceph_assert(bits); |
7c673cae FG |
868 | pg_t retval = *this; |
869 | retval.m_seed &= ~((~0)<<(bits - 1)); | |
870 | return retval; | |
871 | } | |
872 | ||
873 | hobject_t pg_t::get_hobj_start() const | |
874 | { | |
11fdf7f2 | 875 | return hobject_t(object_t(), string(), 0, m_seed, m_pool, |
7c673cae FG |
876 | string()); |
877 | } | |
878 | ||
879 | hobject_t pg_t::get_hobj_end(unsigned pg_num) const | |
880 | { | |
881 | // note: this assumes a bitwise sort; with the legacy nibblewise | |
882 | // sort a PG did not always cover a single contiguous range of the | |
883 | // (bit-reversed) hash range. | |
884 | unsigned bits = get_split_bits(pg_num); | |
885 | uint64_t rev_start = hobject_t::_reverse_bits(m_seed); | |
886 | uint64_t rev_end = (rev_start | (0xffffffff >> bits)) + 1; | |
887 | if (rev_end >= 0x100000000) { | |
11fdf7f2 | 888 | ceph_assert(rev_end == 0x100000000); |
7c673cae FG |
889 | return hobject_t::get_max(); |
890 | } else { | |
891 | return hobject_t(object_t(), string(), CEPH_NOSNAP, | |
892 | hobject_t::_reverse_bits(rev_end), m_pool, | |
893 | string()); | |
894 | } | |
895 | } | |
896 | ||
897 | void pg_t::dump(Formatter *f) const | |
898 | { | |
899 | f->dump_unsigned("pool", m_pool); | |
900 | f->dump_unsigned("seed", m_seed); | |
7c673cae FG |
901 | } |
902 | ||
903 | void pg_t::generate_test_instances(list<pg_t*>& o) | |
904 | { | |
905 | o.push_back(new pg_t); | |
11fdf7f2 TL |
906 | o.push_back(new pg_t(1, 2)); |
907 | o.push_back(new pg_t(13123, 3)); | |
908 | o.push_back(new pg_t(131223, 4)); | |
7c673cae FG |
909 | } |
910 | ||
911 | char *pg_t::calc_name(char *buf, const char *suffix_backwords) const | |
912 | { | |
913 | while (*suffix_backwords) | |
914 | *--buf = *suffix_backwords++; | |
915 | ||
7c673cae FG |
916 | buf = ritoa<uint32_t, 16>(m_seed, buf); |
917 | ||
918 | *--buf = '.'; | |
919 | ||
920 | return ritoa<uint64_t, 10>(m_pool, buf); | |
921 | } | |
922 | ||
923 | ostream& operator<<(ostream& out, const pg_t &pg) | |
924 | { | |
925 | char buf[pg_t::calc_name_buf_size]; | |
926 | buf[pg_t::calc_name_buf_size - 1] = '\0'; | |
927 | out << pg.calc_name(buf + pg_t::calc_name_buf_size - 1, ""); | |
928 | return out; | |
929 | } | |
930 | ||
931 | ||
932 | // -- coll_t -- | |
933 | ||
934 | void coll_t::calc_str() | |
935 | { | |
936 | switch (type) { | |
937 | case TYPE_META: | |
938 | strcpy(_str_buff, "meta"); | |
939 | _str = _str_buff; | |
940 | break; | |
941 | case TYPE_PG: | |
942 | _str_buff[spg_t::calc_name_buf_size - 1] = '\0'; | |
943 | _str = pgid.calc_name(_str_buff + spg_t::calc_name_buf_size - 1, "daeh_"); | |
944 | break; | |
945 | case TYPE_PG_TEMP: | |
946 | _str_buff[spg_t::calc_name_buf_size - 1] = '\0'; | |
947 | _str = pgid.calc_name(_str_buff + spg_t::calc_name_buf_size - 1, "PMET_"); | |
948 | break; | |
949 | default: | |
11fdf7f2 | 950 | ceph_abort_msg("unknown collection type"); |
7c673cae FG |
951 | } |
952 | } | |
953 | ||
954 | bool coll_t::parse(const std::string& s) | |
955 | { | |
956 | if (s == "meta") { | |
957 | type = TYPE_META; | |
958 | pgid = spg_t(); | |
959 | removal_seq = 0; | |
960 | calc_str(); | |
11fdf7f2 | 961 | ceph_assert(s == _str); |
7c673cae FG |
962 | return true; |
963 | } | |
964 | if (s.find("_head") == s.length() - 5 && | |
965 | pgid.parse(s.substr(0, s.length() - 5))) { | |
966 | type = TYPE_PG; | |
967 | removal_seq = 0; | |
968 | calc_str(); | |
11fdf7f2 | 969 | ceph_assert(s == _str); |
7c673cae FG |
970 | return true; |
971 | } | |
972 | if (s.find("_TEMP") == s.length() - 5 && | |
973 | pgid.parse(s.substr(0, s.length() - 5))) { | |
974 | type = TYPE_PG_TEMP; | |
975 | removal_seq = 0; | |
976 | calc_str(); | |
11fdf7f2 | 977 | ceph_assert(s == _str); |
7c673cae FG |
978 | return true; |
979 | } | |
980 | return false; | |
981 | } | |
982 | ||
9f95a23c | 983 | void coll_t::encode(ceph::buffer::list& bl) const |
7c673cae | 984 | { |
11fdf7f2 | 985 | using ceph::encode; |
7c673cae FG |
986 | // when changing this, remember to update encoded_size() too. |
987 | if (is_temp()) { | |
988 | // can't express this as v2... | |
989 | __u8 struct_v = 3; | |
11fdf7f2 TL |
990 | encode(struct_v, bl); |
991 | encode(to_str(), bl); | |
7c673cae FG |
992 | } else { |
993 | __u8 struct_v = 2; | |
11fdf7f2 TL |
994 | encode(struct_v, bl); |
995 | encode((__u8)type, bl); | |
996 | encode(pgid, bl); | |
7c673cae | 997 | snapid_t snap = CEPH_NOSNAP; |
11fdf7f2 | 998 | encode(snap, bl); |
7c673cae FG |
999 | } |
1000 | } | |
1001 | ||
1002 | size_t coll_t::encoded_size() const | |
1003 | { | |
1004 | size_t r = sizeof(__u8); | |
1005 | if (is_temp()) { | |
1006 | // v3 | |
1007 | r += sizeof(__u32); | |
1008 | if (_str) { | |
1009 | r += strlen(_str); | |
1010 | } | |
1011 | } else { | |
1012 | // v2 | |
1013 | // 1. type | |
1014 | r += sizeof(__u8); | |
1015 | // 2. pgid | |
1016 | // - encoding header | |
1017 | r += sizeof(ceph_le32) + 2 * sizeof(__u8); | |
1018 | // - pg_t | |
1019 | r += sizeof(__u8) + sizeof(uint64_t) + 2 * sizeof(uint32_t); | |
1020 | // - shard_id_t | |
1021 | r += sizeof(int8_t); | |
1022 | // 3. snapid_t | |
1023 | r += sizeof(uint64_t); | |
1024 | } | |
1025 | ||
1026 | return r; | |
1027 | } | |
1028 | ||
9f95a23c | 1029 | void coll_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 1030 | { |
11fdf7f2 | 1031 | using ceph::decode; |
7c673cae | 1032 | __u8 struct_v; |
11fdf7f2 | 1033 | decode(struct_v, bl); |
7c673cae FG |
1034 | switch (struct_v) { |
1035 | case 1: | |
1036 | { | |
1037 | snapid_t snap; | |
11fdf7f2 TL |
1038 | decode(pgid, bl); |
1039 | decode(snap, bl); | |
7c673cae FG |
1040 | |
1041 | // infer the type | |
1042 | if (pgid == spg_t() && snap == 0) { | |
1043 | type = TYPE_META; | |
1044 | } else { | |
1045 | type = TYPE_PG; | |
1046 | } | |
1047 | removal_seq = 0; | |
1048 | } | |
1049 | break; | |
1050 | ||
1051 | case 2: | |
1052 | { | |
1053 | __u8 _type; | |
1054 | snapid_t snap; | |
11fdf7f2 TL |
1055 | decode(_type, bl); |
1056 | decode(pgid, bl); | |
1057 | decode(snap, bl); | |
7c673cae FG |
1058 | type = (type_t)_type; |
1059 | removal_seq = 0; | |
1060 | } | |
1061 | break; | |
1062 | ||
1063 | case 3: | |
1064 | { | |
1065 | string str; | |
11fdf7f2 | 1066 | decode(str, bl); |
7c673cae FG |
1067 | bool ok = parse(str); |
1068 | if (!ok) | |
1069 | throw std::domain_error(std::string("unable to parse pg ") + str); | |
1070 | } | |
1071 | break; | |
1072 | ||
1073 | default: | |
1074 | { | |
f67539c2 TL |
1075 | CachedStackStringStream css; |
1076 | *css << "coll_t::decode(): don't know how to decode version " | |
1077 | << struct_v; | |
1078 | throw std::domain_error(css->str()); | |
7c673cae FG |
1079 | } |
1080 | } | |
1081 | } | |
1082 | ||
1083 | void coll_t::dump(Formatter *f) const | |
1084 | { | |
1085 | f->dump_unsigned("type_id", (unsigned)type); | |
1086 | if (type != TYPE_META) | |
1087 | f->dump_stream("pgid") << pgid; | |
1088 | f->dump_string("name", to_str()); | |
1089 | } | |
1090 | ||
1091 | void coll_t::generate_test_instances(list<coll_t*>& o) | |
1092 | { | |
1093 | o.push_back(new coll_t()); | |
1094 | o.push_back(new coll_t(spg_t(pg_t(1, 0), shard_id_t::NO_SHARD))); | |
1095 | o.push_back(new coll_t(o.back()->get_temp())); | |
1096 | o.push_back(new coll_t(spg_t(pg_t(3, 2), shard_id_t(12)))); | |
1097 | o.push_back(new coll_t(o.back()->get_temp())); | |
1098 | o.push_back(new coll_t()); | |
1099 | } | |
1100 | ||
1101 | // --- | |
1102 | ||
1103 | std::string pg_vector_string(const vector<int32_t> &a) | |
1104 | { | |
f67539c2 TL |
1105 | CachedStackStringStream css; |
1106 | *css << "["; | |
9f95a23c TL |
1107 | for (auto i = a.cbegin(); i != a.cend(); ++i) { |
1108 | if (i != a.begin()) | |
f67539c2 | 1109 | *css << ","; |
9f95a23c | 1110 | if (*i != CRUSH_ITEM_NONE) |
f67539c2 | 1111 | *css << *i; |
9f95a23c | 1112 | else |
f67539c2 | 1113 | *css << "NONE"; |
7c673cae | 1114 | } |
f67539c2 TL |
1115 | *css << "]"; |
1116 | return css->str(); | |
7c673cae FG |
1117 | } |
1118 | ||
11fdf7f2 | 1119 | std::string pg_state_string(uint64_t state) |
7c673cae | 1120 | { |
f67539c2 | 1121 | CachedStackStringStream css; |
7c673cae | 1122 | if (state & PG_STATE_STALE) |
f67539c2 | 1123 | *css << "stale+"; |
7c673cae | 1124 | if (state & PG_STATE_CREATING) |
f67539c2 | 1125 | *css << "creating+"; |
7c673cae | 1126 | if (state & PG_STATE_ACTIVE) |
f67539c2 | 1127 | *css << "active+"; |
7c673cae | 1128 | if (state & PG_STATE_ACTIVATING) |
f67539c2 | 1129 | *css << "activating+"; |
7c673cae | 1130 | if (state & PG_STATE_CLEAN) |
f67539c2 | 1131 | *css << "clean+"; |
7c673cae | 1132 | if (state & PG_STATE_RECOVERY_WAIT) |
f67539c2 | 1133 | *css << "recovery_wait+"; |
7c673cae | 1134 | if (state & PG_STATE_RECOVERY_TOOFULL) |
f67539c2 | 1135 | *css << "recovery_toofull+"; |
7c673cae | 1136 | if (state & PG_STATE_RECOVERING) |
f67539c2 | 1137 | *css << "recovering+"; |
c07f9fc5 | 1138 | if (state & PG_STATE_FORCED_RECOVERY) |
f67539c2 | 1139 | *css << "forced_recovery+"; |
7c673cae | 1140 | if (state & PG_STATE_DOWN) |
f67539c2 | 1141 | *css << "down+"; |
b32b8144 | 1142 | if (state & PG_STATE_RECOVERY_UNFOUND) |
f67539c2 | 1143 | *css << "recovery_unfound+"; |
b32b8144 | 1144 | if (state & PG_STATE_BACKFILL_UNFOUND) |
f67539c2 | 1145 | *css << "backfill_unfound+"; |
7c673cae | 1146 | if (state & PG_STATE_UNDERSIZED) |
f67539c2 | 1147 | *css << "undersized+"; |
7c673cae | 1148 | if (state & PG_STATE_DEGRADED) |
f67539c2 | 1149 | *css << "degraded+"; |
7c673cae | 1150 | if (state & PG_STATE_REMAPPED) |
f67539c2 | 1151 | *css << "remapped+"; |
11fdf7f2 | 1152 | if (state & PG_STATE_PREMERGE) |
f67539c2 | 1153 | *css << "premerge+"; |
7c673cae | 1154 | if (state & PG_STATE_SCRUBBING) |
f67539c2 | 1155 | *css << "scrubbing+"; |
7c673cae | 1156 | if (state & PG_STATE_DEEP_SCRUB) |
f67539c2 | 1157 | *css << "deep+"; |
7c673cae | 1158 | if (state & PG_STATE_INCONSISTENT) |
f67539c2 | 1159 | *css << "inconsistent+"; |
7c673cae | 1160 | if (state & PG_STATE_PEERING) |
f67539c2 | 1161 | *css << "peering+"; |
7c673cae | 1162 | if (state & PG_STATE_REPAIR) |
f67539c2 | 1163 | *css << "repair+"; |
3efd9988 | 1164 | if (state & PG_STATE_BACKFILL_WAIT) |
f67539c2 | 1165 | *css << "backfill_wait+"; |
3efd9988 | 1166 | if (state & PG_STATE_BACKFILLING) |
f67539c2 | 1167 | *css << "backfilling+"; |
c07f9fc5 | 1168 | if (state & PG_STATE_FORCED_BACKFILL) |
f67539c2 | 1169 | *css << "forced_backfill+"; |
7c673cae | 1170 | if (state & PG_STATE_BACKFILL_TOOFULL) |
f67539c2 | 1171 | *css << "backfill_toofull+"; |
7c673cae | 1172 | if (state & PG_STATE_INCOMPLETE) |
f67539c2 | 1173 | *css << "incomplete+"; |
7c673cae | 1174 | if (state & PG_STATE_PEERED) |
f67539c2 | 1175 | *css << "peered+"; |
7c673cae | 1176 | if (state & PG_STATE_SNAPTRIM) |
f67539c2 | 1177 | *css << "snaptrim+"; |
7c673cae | 1178 | if (state & PG_STATE_SNAPTRIM_WAIT) |
f67539c2 | 1179 | *css << "snaptrim_wait+"; |
224ce89b | 1180 | if (state & PG_STATE_SNAPTRIM_ERROR) |
f67539c2 | 1181 | *css << "snaptrim_error+"; |
11fdf7f2 | 1182 | if (state & PG_STATE_FAILED_REPAIR) |
f67539c2 | 1183 | *css << "failed_repair+"; |
9f95a23c | 1184 | if (state & PG_STATE_LAGGY) |
f67539c2 | 1185 | *css << "laggy+"; |
9f95a23c | 1186 | if (state & PG_STATE_WAIT) |
f67539c2 TL |
1187 | *css << "wait+"; |
1188 | auto ret = css->str(); | |
7c673cae FG |
1189 | if (ret.length() > 0) |
1190 | ret.resize(ret.length() - 1); | |
1191 | else | |
31f18b77 | 1192 | ret = "unknown"; |
7c673cae FG |
1193 | return ret; |
1194 | } | |
1195 | ||
9f95a23c | 1196 | std::optional<uint64_t> pg_string_state(const std::string& state) |
7c673cae | 1197 | { |
9f95a23c | 1198 | std::optional<uint64_t> type; |
7c673cae FG |
1199 | if (state == "active") |
1200 | type = PG_STATE_ACTIVE; | |
1201 | else if (state == "clean") | |
1202 | type = PG_STATE_CLEAN; | |
1203 | else if (state == "down") | |
1204 | type = PG_STATE_DOWN; | |
b32b8144 FG |
1205 | else if (state == "recovery_unfound") |
1206 | type = PG_STATE_RECOVERY_UNFOUND; | |
1207 | else if (state == "backfill_unfound") | |
1208 | type = PG_STATE_BACKFILL_UNFOUND; | |
11fdf7f2 TL |
1209 | else if (state == "premerge") |
1210 | type = PG_STATE_PREMERGE; | |
7c673cae FG |
1211 | else if (state == "scrubbing") |
1212 | type = PG_STATE_SCRUBBING; | |
1213 | else if (state == "degraded") | |
1214 | type = PG_STATE_DEGRADED; | |
1215 | else if (state == "inconsistent") | |
1216 | type = PG_STATE_INCONSISTENT; | |
1217 | else if (state == "peering") | |
1218 | type = PG_STATE_PEERING; | |
1219 | else if (state == "repair") | |
1220 | type = PG_STATE_REPAIR; | |
1221 | else if (state == "recovering") | |
1222 | type = PG_STATE_RECOVERING; | |
c07f9fc5 FG |
1223 | else if (state == "forced_recovery") |
1224 | type = PG_STATE_FORCED_RECOVERY; | |
7c673cae FG |
1225 | else if (state == "backfill_wait") |
1226 | type = PG_STATE_BACKFILL_WAIT; | |
1227 | else if (state == "incomplete") | |
1228 | type = PG_STATE_INCOMPLETE; | |
1229 | else if (state == "stale") | |
1230 | type = PG_STATE_STALE; | |
1231 | else if (state == "remapped") | |
1232 | type = PG_STATE_REMAPPED; | |
94b18763 | 1233 | else if (state == "deep") |
7c673cae | 1234 | type = PG_STATE_DEEP_SCRUB; |
3efd9988 FG |
1235 | else if (state == "backfilling") |
1236 | type = PG_STATE_BACKFILLING; | |
c07f9fc5 FG |
1237 | else if (state == "forced_backfill") |
1238 | type = PG_STATE_FORCED_BACKFILL; | |
7c673cae FG |
1239 | else if (state == "backfill_toofull") |
1240 | type = PG_STATE_BACKFILL_TOOFULL; | |
1241 | else if (state == "recovery_wait") | |
1242 | type = PG_STATE_RECOVERY_WAIT; | |
1243 | else if (state == "recovery_toofull") | |
1244 | type = PG_STATE_RECOVERY_TOOFULL; | |
1245 | else if (state == "undersized") | |
1246 | type = PG_STATE_UNDERSIZED; | |
1247 | else if (state == "activating") | |
1248 | type = PG_STATE_ACTIVATING; | |
1249 | else if (state == "peered") | |
1250 | type = PG_STATE_PEERED; | |
1251 | else if (state == "snaptrim") | |
1252 | type = PG_STATE_SNAPTRIM; | |
1253 | else if (state == "snaptrim_wait") | |
1254 | type = PG_STATE_SNAPTRIM_WAIT; | |
224ce89b WB |
1255 | else if (state == "snaptrim_error") |
1256 | type = PG_STATE_SNAPTRIM_ERROR; | |
91327a77 AA |
1257 | else if (state == "creating") |
1258 | type = PG_STATE_CREATING; | |
11fdf7f2 TL |
1259 | else if (state == "failed_repair") |
1260 | type = PG_STATE_FAILED_REPAIR; | |
9f95a23c TL |
1261 | else if (state == "laggy") |
1262 | type = PG_STATE_LAGGY; | |
1263 | else if (state == "wait") | |
1264 | type = PG_STATE_WAIT; | |
11fdf7f2 TL |
1265 | else if (state == "unknown") |
1266 | type = 0; | |
7c673cae | 1267 | else |
9f95a23c | 1268 | type = std::nullopt; |
7c673cae FG |
1269 | return type; |
1270 | } | |
1271 | ||
1272 | // -- eversion_t -- | |
1273 | string eversion_t::get_key_name() const | |
1274 | { | |
11fdf7f2 TL |
1275 | std::string key(32, ' '); |
1276 | get_key_name(&key[0]); | |
1277 | key.resize(31); // remove the null terminator | |
1278 | return key; | |
7c673cae FG |
1279 | } |
1280 | ||
7c673cae FG |
1281 | // -- pool_snap_info_t -- |
1282 | void pool_snap_info_t::dump(Formatter *f) const | |
1283 | { | |
1284 | f->dump_unsigned("snapid", snapid); | |
1285 | f->dump_stream("stamp") << stamp; | |
1286 | f->dump_string("name", name); | |
1287 | } | |
1288 | ||
9f95a23c | 1289 | void pool_snap_info_t::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 1290 | { |
11fdf7f2 | 1291 | using ceph::encode; |
7c673cae FG |
1292 | if ((features & CEPH_FEATURE_PGPOOL3) == 0) { |
1293 | __u8 struct_v = 1; | |
11fdf7f2 TL |
1294 | encode(struct_v, bl); |
1295 | encode(snapid, bl); | |
1296 | encode(stamp, bl); | |
1297 | encode(name, bl); | |
7c673cae FG |
1298 | return; |
1299 | } | |
1300 | ENCODE_START(2, 2, bl); | |
11fdf7f2 TL |
1301 | encode(snapid, bl); |
1302 | encode(stamp, bl); | |
1303 | encode(name, bl); | |
7c673cae FG |
1304 | ENCODE_FINISH(bl); |
1305 | } | |
1306 | ||
9f95a23c | 1307 | void pool_snap_info_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
1308 | { |
1309 | DECODE_START_LEGACY_COMPAT_LEN(2, 2, 2, bl); | |
11fdf7f2 TL |
1310 | decode(snapid, bl); |
1311 | decode(stamp, bl); | |
1312 | decode(name, bl); | |
7c673cae FG |
1313 | DECODE_FINISH(bl); |
1314 | } | |
1315 | ||
1316 | void pool_snap_info_t::generate_test_instances(list<pool_snap_info_t*>& o) | |
1317 | { | |
1318 | o.push_back(new pool_snap_info_t); | |
1319 | o.push_back(new pool_snap_info_t); | |
1320 | o.back()->snapid = 1; | |
1321 | o.back()->stamp = utime_t(1, 2); | |
1322 | o.back()->name = "foo"; | |
1323 | } | |
1324 | ||
1325 | // -- pool_opts_t -- | |
1326 | ||
1d09f67e TL |
1327 | // The order of items in the list is important, therefore, |
1328 | // you should always add to the end of the list when adding new options. | |
1329 | ||
7c673cae FG |
1330 | typedef std::map<std::string, pool_opts_t::opt_desc_t> opt_mapping_t; |
1331 | static opt_mapping_t opt_mapping = boost::assign::map_list_of | |
1332 | ("scrub_min_interval", pool_opts_t::opt_desc_t( | |
1333 | pool_opts_t::SCRUB_MIN_INTERVAL, pool_opts_t::DOUBLE)) | |
1334 | ("scrub_max_interval", pool_opts_t::opt_desc_t( | |
1335 | pool_opts_t::SCRUB_MAX_INTERVAL, pool_opts_t::DOUBLE)) | |
1336 | ("deep_scrub_interval", pool_opts_t::opt_desc_t( | |
1337 | pool_opts_t::DEEP_SCRUB_INTERVAL, pool_opts_t::DOUBLE)) | |
1338 | ("recovery_priority", pool_opts_t::opt_desc_t( | |
1339 | pool_opts_t::RECOVERY_PRIORITY, pool_opts_t::INT)) | |
1340 | ("recovery_op_priority", pool_opts_t::opt_desc_t( | |
1341 | pool_opts_t::RECOVERY_OP_PRIORITY, pool_opts_t::INT)) | |
1342 | ("scrub_priority", pool_opts_t::opt_desc_t( | |
1343 | pool_opts_t::SCRUB_PRIORITY, pool_opts_t::INT)) | |
1344 | ("compression_mode", pool_opts_t::opt_desc_t( | |
1345 | pool_opts_t::COMPRESSION_MODE, pool_opts_t::STR)) | |
1346 | ("compression_algorithm", pool_opts_t::opt_desc_t( | |
1347 | pool_opts_t::COMPRESSION_ALGORITHM, pool_opts_t::STR)) | |
1348 | ("compression_required_ratio", pool_opts_t::opt_desc_t( | |
1349 | pool_opts_t::COMPRESSION_REQUIRED_RATIO, pool_opts_t::DOUBLE)) | |
1350 | ("compression_max_blob_size", pool_opts_t::opt_desc_t( | |
1351 | pool_opts_t::COMPRESSION_MAX_BLOB_SIZE, pool_opts_t::INT)) | |
1352 | ("compression_min_blob_size", pool_opts_t::opt_desc_t( | |
1353 | pool_opts_t::COMPRESSION_MIN_BLOB_SIZE, pool_opts_t::INT)) | |
1354 | ("csum_type", pool_opts_t::opt_desc_t( | |
1355 | pool_opts_t::CSUM_TYPE, pool_opts_t::INT)) | |
1356 | ("csum_max_block", pool_opts_t::opt_desc_t( | |
1357 | pool_opts_t::CSUM_MAX_BLOCK, pool_opts_t::INT)) | |
1358 | ("csum_min_block", pool_opts_t::opt_desc_t( | |
11fdf7f2 TL |
1359 | pool_opts_t::CSUM_MIN_BLOCK, pool_opts_t::INT)) |
1360 | ("fingerprint_algorithm", pool_opts_t::opt_desc_t( | |
1361 | pool_opts_t::FINGERPRINT_ALGORITHM, pool_opts_t::STR)) | |
1362 | ("pg_num_min", pool_opts_t::opt_desc_t( | |
1363 | pool_opts_t::PG_NUM_MIN, pool_opts_t::INT)) | |
1364 | ("target_size_bytes", pool_opts_t::opt_desc_t( | |
1365 | pool_opts_t::TARGET_SIZE_BYTES, pool_opts_t::INT)) | |
1366 | ("target_size_ratio", pool_opts_t::opt_desc_t( | |
1367 | pool_opts_t::TARGET_SIZE_RATIO, pool_opts_t::DOUBLE)) | |
1368 | ("pg_autoscale_bias", pool_opts_t::opt_desc_t( | |
9f95a23c TL |
1369 | pool_opts_t::PG_AUTOSCALE_BIAS, pool_opts_t::DOUBLE)) |
1370 | ("read_lease_interval", pool_opts_t::opt_desc_t( | |
f67539c2 TL |
1371 | pool_opts_t::READ_LEASE_INTERVAL, pool_opts_t::DOUBLE)) |
1372 | ("dedup_tier", pool_opts_t::opt_desc_t( | |
1373 | pool_opts_t::DEDUP_TIER, pool_opts_t::INT)) | |
1374 | ("dedup_chunk_algorithm", pool_opts_t::opt_desc_t( | |
1375 | pool_opts_t::DEDUP_CHUNK_ALGORITHM, pool_opts_t::STR)) | |
1376 | ("dedup_cdc_chunk_size", pool_opts_t::opt_desc_t( | |
1d09f67e TL |
1377 | pool_opts_t::DEDUP_CDC_CHUNK_SIZE, pool_opts_t::INT)) |
1378 | ("pg_num_max", pool_opts_t::opt_desc_t( | |
1379 | pool_opts_t::PG_NUM_MAX, pool_opts_t::INT)); | |
7c673cae | 1380 | |
11fdf7f2 TL |
1381 | bool pool_opts_t::is_opt_name(const std::string& name) |
1382 | { | |
1383 | return opt_mapping.count(name); | |
7c673cae FG |
1384 | } |
1385 | ||
11fdf7f2 TL |
1386 | pool_opts_t::opt_desc_t pool_opts_t::get_opt_desc(const std::string& name) |
1387 | { | |
9f95a23c | 1388 | auto i = opt_mapping.find(name); |
11fdf7f2 TL |
1389 | ceph_assert(i != opt_mapping.end()); |
1390 | return i->second; | |
7c673cae FG |
1391 | } |
1392 | ||
11fdf7f2 TL |
1393 | bool pool_opts_t::is_set(pool_opts_t::key_t key) const |
1394 | { | |
1395 | return opts.count(key); | |
7c673cae FG |
1396 | } |
1397 | ||
11fdf7f2 TL |
1398 | const pool_opts_t::value_t& pool_opts_t::get(pool_opts_t::key_t key) const |
1399 | { | |
9f95a23c | 1400 | auto i = opts.find(key); |
11fdf7f2 | 1401 | ceph_assert(i != opts.end()); |
7c673cae FG |
1402 | return i->second; |
1403 | } | |
1404 | ||
1405 | bool pool_opts_t::unset(pool_opts_t::key_t key) { | |
1406 | return opts.erase(key) > 0; | |
1407 | } | |
1408 | ||
11fdf7f2 | 1409 | class pool_opts_dumper_t : public boost::static_visitor<> { |
7c673cae FG |
1410 | public: |
1411 | pool_opts_dumper_t(const std::string& name_, Formatter* f_) : | |
1412 | name(name_.c_str()), f(f_) {} | |
1413 | ||
1414 | void operator()(std::string s) const { | |
1415 | f->dump_string(name, s); | |
1416 | } | |
11fdf7f2 | 1417 | void operator()(int64_t i) const { |
7c673cae FG |
1418 | f->dump_int(name, i); |
1419 | } | |
1420 | void operator()(double d) const { | |
1421 | f->dump_float(name, d); | |
1422 | } | |
1423 | ||
1424 | private: | |
1425 | const char* name; | |
1426 | Formatter* f; | |
1427 | }; | |
1428 | ||
1429 | void pool_opts_t::dump(const std::string& name, Formatter* f) const | |
1430 | { | |
1431 | const opt_desc_t& desc = get_opt_desc(name); | |
9f95a23c | 1432 | auto i = opts.find(desc.key); |
7c673cae FG |
1433 | if (i == opts.end()) { |
1434 | return; | |
1435 | } | |
1436 | boost::apply_visitor(pool_opts_dumper_t(name, f), i->second); | |
1437 | } | |
1438 | ||
1439 | void pool_opts_t::dump(Formatter* f) const | |
1440 | { | |
9f95a23c | 1441 | for (auto i = opt_mapping.cbegin(); i != opt_mapping.cend(); ++i) { |
7c673cae FG |
1442 | const std::string& name = i->first; |
1443 | const opt_desc_t& desc = i->second; | |
9f95a23c | 1444 | auto j = opts.find(desc.key); |
7c673cae FG |
1445 | if (j == opts.end()) { |
1446 | continue; | |
1447 | } | |
1448 | boost::apply_visitor(pool_opts_dumper_t(name, f), j->second); | |
1449 | } | |
1450 | } | |
1451 | ||
11fdf7f2 | 1452 | class pool_opts_encoder_t : public boost::static_visitor<> { |
7c673cae | 1453 | public: |
9f95a23c | 1454 | explicit pool_opts_encoder_t(ceph::buffer::list& bl_, uint64_t features) |
11fdf7f2 TL |
1455 | : bl(bl_), |
1456 | features(features) {} | |
1457 | ||
1458 | void operator()(const std::string &s) const { | |
1459 | encode(static_cast<int32_t>(pool_opts_t::STR), bl); | |
1460 | encode(s, bl); | |
1461 | } | |
1462 | void operator()(int64_t i) const { | |
1463 | encode(static_cast<int32_t>(pool_opts_t::INT), bl); | |
1464 | if (HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
1465 | encode(i, bl); | |
1466 | } else { | |
1467 | encode(static_cast<int32_t>(i), bl); | |
1468 | } | |
7c673cae FG |
1469 | } |
1470 | void operator()(double d) const { | |
11fdf7f2 TL |
1471 | encode(static_cast<int32_t>(pool_opts_t::DOUBLE), bl); |
1472 | encode(d, bl); | |
7c673cae FG |
1473 | } |
1474 | ||
1475 | private: | |
9f95a23c | 1476 | ceph::buffer::list& bl; |
11fdf7f2 | 1477 | uint64_t features; |
7c673cae FG |
1478 | }; |
1479 | ||
9f95a23c | 1480 | void pool_opts_t::encode(ceph::buffer::list& bl, uint64_t features) const |
11fdf7f2 TL |
1481 | { |
1482 | unsigned v = 2; | |
1483 | if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
1484 | v = 1; | |
1485 | } | |
1486 | ENCODE_START(v, 1, bl); | |
7c673cae | 1487 | uint32_t n = static_cast<uint32_t>(opts.size()); |
11fdf7f2 | 1488 | encode(n, bl); |
9f95a23c | 1489 | for (auto i = opts.cbegin(); i != opts.cend(); ++i) { |
11fdf7f2 TL |
1490 | encode(static_cast<int32_t>(i->first), bl); |
1491 | boost::apply_visitor(pool_opts_encoder_t(bl, features), i->second); | |
7c673cae FG |
1492 | } |
1493 | ENCODE_FINISH(bl); | |
1494 | } | |
1495 | ||
9f95a23c | 1496 | void pool_opts_t::decode(ceph::buffer::list::const_iterator& bl) |
11fdf7f2 | 1497 | { |
7c673cae FG |
1498 | DECODE_START(1, bl); |
1499 | __u32 n; | |
11fdf7f2 | 1500 | decode(n, bl); |
7c673cae FG |
1501 | opts.clear(); |
1502 | while (n--) { | |
1503 | int32_t k, t; | |
11fdf7f2 TL |
1504 | decode(k, bl); |
1505 | decode(t, bl); | |
7c673cae FG |
1506 | if (t == STR) { |
1507 | std::string s; | |
11fdf7f2 | 1508 | decode(s, bl); |
7c673cae FG |
1509 | opts[static_cast<key_t>(k)] = s; |
1510 | } else if (t == INT) { | |
11fdf7f2 TL |
1511 | int64_t i; |
1512 | if (struct_v >= 2) { | |
1513 | decode(i, bl); | |
1514 | } else { | |
1515 | int ii; | |
1516 | decode(ii, bl); | |
1517 | i = ii; | |
1518 | } | |
7c673cae FG |
1519 | opts[static_cast<key_t>(k)] = i; |
1520 | } else if (t == DOUBLE) { | |
1521 | double d; | |
11fdf7f2 | 1522 | decode(d, bl); |
7c673cae FG |
1523 | opts[static_cast<key_t>(k)] = d; |
1524 | } else { | |
11fdf7f2 | 1525 | ceph_assert(!"invalid type"); |
7c673cae FG |
1526 | } |
1527 | } | |
1528 | DECODE_FINISH(bl); | |
1529 | } | |
1530 | ||
1531 | ostream& operator<<(ostream& out, const pool_opts_t& opts) | |
1532 | { | |
9f95a23c | 1533 | for (auto i = opt_mapping.begin(); i != opt_mapping.end(); ++i) { |
7c673cae FG |
1534 | const std::string& name = i->first; |
1535 | const pool_opts_t::opt_desc_t& desc = i->second; | |
9f95a23c | 1536 | auto j = opts.opts.find(desc.key); |
7c673cae FG |
1537 | if (j == opts.opts.end()) { |
1538 | continue; | |
1539 | } | |
1540 | out << " " << name << " " << j->second; | |
1541 | } | |
1542 | return out; | |
1543 | } | |
1544 | ||
1545 | // -- pg_pool_t -- | |
1546 | ||
c07f9fc5 FG |
1547 | const char *pg_pool_t::APPLICATION_NAME_CEPHFS("cephfs"); |
1548 | const char *pg_pool_t::APPLICATION_NAME_RBD("rbd"); | |
1549 | const char *pg_pool_t::APPLICATION_NAME_RGW("rgw"); | |
1550 | ||
7c673cae FG |
1551 | void pg_pool_t::dump(Formatter *f) const |
1552 | { | |
11fdf7f2 | 1553 | f->dump_stream("create_time") << get_create_time(); |
7c673cae FG |
1554 | f->dump_unsigned("flags", get_flags()); |
1555 | f->dump_string("flags_names", get_flags_string()); | |
1556 | f->dump_int("type", get_type()); | |
1557 | f->dump_int("size", get_size()); | |
1558 | f->dump_int("min_size", get_min_size()); | |
31f18b77 | 1559 | f->dump_int("crush_rule", get_crush_rule()); |
f67539c2 TL |
1560 | f->dump_int("peering_crush_bucket_count", peering_crush_bucket_count); |
1561 | f->dump_int("peering_crush_bucket_target", peering_crush_bucket_target); | |
1562 | f->dump_int("peering_crush_bucket_barrier", peering_crush_bucket_barrier); | |
1563 | f->dump_int("peering_crush_bucket_mandatory_member", peering_crush_mandatory_member); | |
7c673cae | 1564 | f->dump_int("object_hash", get_object_hash()); |
11fdf7f2 TL |
1565 | f->dump_string("pg_autoscale_mode", |
1566 | get_pg_autoscale_mode_name(pg_autoscale_mode)); | |
7c673cae FG |
1567 | f->dump_unsigned("pg_num", get_pg_num()); |
1568 | f->dump_unsigned("pg_placement_num", get_pgp_num()); | |
11fdf7f2 TL |
1569 | f->dump_unsigned("pg_placement_num_target", get_pgp_num_target()); |
1570 | f->dump_unsigned("pg_num_target", get_pg_num_target()); | |
1571 | f->dump_unsigned("pg_num_pending", get_pg_num_pending()); | |
1572 | f->dump_object("last_pg_merge_meta", last_pg_merge_meta); | |
7c673cae FG |
1573 | f->dump_stream("last_change") << get_last_change(); |
1574 | f->dump_stream("last_force_op_resend") << get_last_force_op_resend(); | |
11fdf7f2 TL |
1575 | f->dump_stream("last_force_op_resend_prenautilus") |
1576 | << get_last_force_op_resend_prenautilus(); | |
7c673cae FG |
1577 | f->dump_stream("last_force_op_resend_preluminous") |
1578 | << get_last_force_op_resend_preluminous(); | |
1579 | f->dump_unsigned("auid", get_auid()); | |
1580 | f->dump_string("snap_mode", is_pool_snaps_mode() ? "pool" : "selfmanaged"); | |
1581 | f->dump_unsigned("snap_seq", get_snap_seq()); | |
1582 | f->dump_unsigned("snap_epoch", get_snap_epoch()); | |
1583 | f->open_array_section("pool_snaps"); | |
9f95a23c | 1584 | for (auto p = snaps.cbegin(); p != snaps.cend(); ++p) { |
7c673cae FG |
1585 | f->open_object_section("pool_snap_info"); |
1586 | p->second.dump(f); | |
1587 | f->close_section(); | |
1588 | } | |
1589 | f->close_section(); | |
1590 | f->dump_stream("removed_snaps") << removed_snaps; | |
1591 | f->dump_unsigned("quota_max_bytes", quota_max_bytes); | |
1592 | f->dump_unsigned("quota_max_objects", quota_max_objects); | |
1593 | f->open_array_section("tiers"); | |
9f95a23c | 1594 | for (auto p = tiers.cbegin(); p != tiers.cend(); ++p) |
7c673cae FG |
1595 | f->dump_unsigned("pool_id", *p); |
1596 | f->close_section(); | |
1597 | f->dump_int("tier_of", tier_of); | |
1598 | f->dump_int("read_tier", read_tier); | |
1599 | f->dump_int("write_tier", write_tier); | |
1600 | f->dump_string("cache_mode", get_cache_mode_name()); | |
1601 | f->dump_unsigned("target_max_bytes", target_max_bytes); | |
1602 | f->dump_unsigned("target_max_objects", target_max_objects); | |
1603 | f->dump_unsigned("cache_target_dirty_ratio_micro", | |
1604 | cache_target_dirty_ratio_micro); | |
1605 | f->dump_unsigned("cache_target_dirty_high_ratio_micro", | |
1606 | cache_target_dirty_high_ratio_micro); | |
1607 | f->dump_unsigned("cache_target_full_ratio_micro", | |
1608 | cache_target_full_ratio_micro); | |
1609 | f->dump_unsigned("cache_min_flush_age", cache_min_flush_age); | |
1610 | f->dump_unsigned("cache_min_evict_age", cache_min_evict_age); | |
1611 | f->dump_string("erasure_code_profile", erasure_code_profile); | |
1612 | f->open_object_section("hit_set_params"); | |
1613 | hit_set_params.dump(f); | |
1614 | f->close_section(); // hit_set_params | |
1615 | f->dump_unsigned("hit_set_period", hit_set_period); | |
1616 | f->dump_unsigned("hit_set_count", hit_set_count); | |
1617 | f->dump_bool("use_gmt_hitset", use_gmt_hitset); | |
1618 | f->dump_unsigned("min_read_recency_for_promote", min_read_recency_for_promote); | |
1619 | f->dump_unsigned("min_write_recency_for_promote", min_write_recency_for_promote); | |
1620 | f->dump_unsigned("hit_set_grade_decay_rate", hit_set_grade_decay_rate); | |
1621 | f->dump_unsigned("hit_set_search_last_n", hit_set_search_last_n); | |
1622 | f->open_array_section("grade_table"); | |
1623 | for (unsigned i = 0; i < hit_set_count; ++i) | |
1624 | f->dump_unsigned("value", get_grade(i)); | |
1625 | f->close_section(); | |
1626 | f->dump_unsigned("stripe_width", get_stripe_width()); | |
1627 | f->dump_unsigned("expected_num_objects", expected_num_objects); | |
1628 | f->dump_bool("fast_read", fast_read); | |
1629 | f->open_object_section("options"); | |
1630 | opts.dump(f); | |
1631 | f->close_section(); // options | |
c07f9fc5 FG |
1632 | f->open_object_section("application_metadata"); |
1633 | for (auto &app_pair : application_metadata) { | |
1634 | f->open_object_section(app_pair.first.c_str()); | |
1635 | for (auto &kv_pair : app_pair.second) { | |
1636 | f->dump_string(kv_pair.first.c_str(), kv_pair.second); | |
1637 | } | |
1638 | f->close_section(); // application | |
1639 | } | |
1640 | f->close_section(); // application_metadata | |
7c673cae FG |
1641 | } |
1642 | ||
1643 | void pg_pool_t::convert_to_pg_shards(const vector<int> &from, set<pg_shard_t>* to) const { | |
1644 | for (size_t i = 0; i < from.size(); ++i) { | |
1645 | if (from[i] != CRUSH_ITEM_NONE) { | |
1646 | to->insert( | |
1647 | pg_shard_t( | |
1648 | from[i], | |
11fdf7f2 | 1649 | is_erasure() ? shard_id_t(i) : shard_id_t::NO_SHARD)); |
7c673cae FG |
1650 | } |
1651 | } | |
1652 | } | |
1653 | ||
1654 | void pg_pool_t::calc_pg_masks() | |
1655 | { | |
1656 | pg_num_mask = (1 << cbits(pg_num-1)) - 1; | |
1657 | pgp_num_mask = (1 << cbits(pgp_num-1)) - 1; | |
1658 | } | |
1659 | ||
1660 | unsigned pg_pool_t::get_pg_num_divisor(pg_t pgid) const | |
1661 | { | |
1662 | if (pg_num == pg_num_mask + 1) | |
1663 | return pg_num; // power-of-2 split | |
1664 | unsigned mask = pg_num_mask >> 1; | |
1665 | if ((pgid.ps() & mask) < (pg_num & mask)) | |
1666 | return pg_num_mask + 1; // smaller bin size (already split) | |
1667 | else | |
1668 | return (pg_num_mask + 1) >> 1; // bigger bin (not yet split) | |
1669 | } | |
1670 | ||
11fdf7f2 TL |
1671 | bool pg_pool_t::is_pending_merge(pg_t pgid, bool *target) const |
1672 | { | |
1673 | if (pg_num_pending >= pg_num) { | |
1674 | return false; | |
1675 | } | |
1676 | if (pgid.ps() >= pg_num_pending && pgid.ps() < pg_num) { | |
1677 | if (target) { | |
1678 | *target = false; | |
1679 | } | |
1680 | return true; | |
1681 | } | |
1682 | for (unsigned ps = pg_num_pending; ps < pg_num; ++ps) { | |
1683 | if (pg_t(ps, pgid.pool()).get_parent() == pgid) { | |
1684 | if (target) { | |
1685 | *target = true; | |
1686 | } | |
1687 | return true; | |
1688 | } | |
1689 | } | |
1690 | return false; | |
1691 | } | |
1692 | ||
7c673cae FG |
1693 | /* |
1694 | * we have two snap modes: | |
11fdf7f2 | 1695 | * - pool snaps |
7c673cae FG |
1696 | * - snap existence/non-existence defined by snaps[] and snap_seq |
1697 | * - user managed snaps | |
11fdf7f2 | 1698 | * - existence tracked by librados user |
7c673cae FG |
1699 | */ |
1700 | bool pg_pool_t::is_pool_snaps_mode() const | |
1701 | { | |
11fdf7f2 | 1702 | return has_flag(FLAG_POOL_SNAPS); |
7c673cae FG |
1703 | } |
1704 | ||
1705 | bool pg_pool_t::is_unmanaged_snaps_mode() const | |
1706 | { | |
11fdf7f2 | 1707 | return has_flag(FLAG_SELFMANAGED_SNAPS); |
7c673cae FG |
1708 | } |
1709 | ||
1710 | bool pg_pool_t::is_removed_snap(snapid_t s) const | |
1711 | { | |
1712 | if (is_pool_snaps_mode()) | |
1713 | return s <= get_snap_seq() && snaps.count(s) == 0; | |
1714 | else | |
1715 | return removed_snaps.contains(s); | |
1716 | } | |
1717 | ||
f67539c2 | 1718 | snapid_t pg_pool_t::snap_exists(std::string_view s) const |
7c673cae | 1719 | { |
9f95a23c | 1720 | for (auto p = snaps.cbegin(); p != snaps.cend(); ++p) |
7c673cae FG |
1721 | if (p->second.name == s) |
1722 | return p->second.snapid; | |
1723 | return 0; | |
1724 | } | |
1725 | ||
1726 | void pg_pool_t::add_snap(const char *n, utime_t stamp) | |
1727 | { | |
11fdf7f2 TL |
1728 | ceph_assert(!is_unmanaged_snaps_mode()); |
1729 | flags |= FLAG_POOL_SNAPS; | |
7c673cae FG |
1730 | snapid_t s = get_snap_seq() + 1; |
1731 | snap_seq = s; | |
1732 | snaps[s].snapid = s; | |
1733 | snaps[s].name = n; | |
1734 | snaps[s].stamp = stamp; | |
1735 | } | |
1736 | ||
9f95a23c | 1737 | uint64_t pg_pool_t::add_unmanaged_snap(bool preoctopus_compat) |
7c673cae | 1738 | { |
11fdf7f2 TL |
1739 | ceph_assert(!is_pool_snaps_mode()); |
1740 | if (snap_seq == 0) { | |
9f95a23c TL |
1741 | if (preoctopus_compat) { |
1742 | // kludge for pre-mimic tracking of pool vs selfmanaged snaps. after | |
1743 | // mimic this field is not decoded but our flag is set; pre-mimic, we | |
1744 | // have a non-empty removed_snaps to signifiy a non-pool-snaps pool. | |
1745 | removed_snaps.insert(snapid_t(1)); | |
1746 | } | |
7c673cae FG |
1747 | snap_seq = 1; |
1748 | } | |
11fdf7f2 | 1749 | flags |= FLAG_SELFMANAGED_SNAPS; |
9f95a23c TL |
1750 | snap_seq = snap_seq + 1; |
1751 | return snap_seq; | |
7c673cae FG |
1752 | } |
1753 | ||
1754 | void pg_pool_t::remove_snap(snapid_t s) | |
1755 | { | |
11fdf7f2 | 1756 | ceph_assert(snaps.count(s)); |
7c673cae FG |
1757 | snaps.erase(s); |
1758 | snap_seq = snap_seq + 1; | |
1759 | } | |
1760 | ||
9f95a23c | 1761 | void pg_pool_t::remove_unmanaged_snap(snapid_t s, bool preoctopus_compat) |
7c673cae | 1762 | { |
11fdf7f2 | 1763 | ceph_assert(is_unmanaged_snaps_mode()); |
9f95a23c TL |
1764 | ++snap_seq; |
1765 | if (preoctopus_compat) { | |
1766 | removed_snaps.insert(s); | |
1767 | // try to add in the new seq, just to try to keep the interval_set contiguous | |
1768 | if (!removed_snaps.contains(get_snap_seq())) { | |
1769 | removed_snaps.insert(get_snap_seq()); | |
1770 | } | |
28e407b8 | 1771 | } |
7c673cae FG |
1772 | } |
1773 | ||
1774 | SnapContext pg_pool_t::get_snap_context() const | |
1775 | { | |
1776 | vector<snapid_t> s(snaps.size()); | |
1777 | unsigned i = 0; | |
9f95a23c | 1778 | for (auto p = snaps.crbegin(); p != snaps.crend(); ++p) |
7c673cae FG |
1779 | s[i++] = p->first; |
1780 | return SnapContext(get_snap_seq(), s); | |
1781 | } | |
1782 | ||
1783 | uint32_t pg_pool_t::hash_key(const string& key, const string& ns) const | |
1784 | { | |
1785 | if (ns.empty()) | |
1786 | return ceph_str_hash(object_hash, key.data(), key.length()); | |
1787 | int nsl = ns.length(); | |
1788 | int len = key.length() + nsl + 1; | |
1789 | char buf[len]; | |
1790 | memcpy(&buf[0], ns.data(), nsl); | |
1791 | buf[nsl] = '\037'; | |
1792 | memcpy(&buf[nsl+1], key.data(), key.length()); | |
1793 | return ceph_str_hash(object_hash, &buf[0], len); | |
1794 | } | |
1795 | ||
1796 | uint32_t pg_pool_t::raw_hash_to_pg(uint32_t v) const | |
1797 | { | |
1798 | return ceph_stable_mod(v, pg_num, pg_num_mask); | |
1799 | } | |
1800 | ||
1801 | /* | |
1802 | * map a raw pg (with full precision ps) into an actual pg, for storage | |
1803 | */ | |
1804 | pg_t pg_pool_t::raw_pg_to_pg(pg_t pg) const | |
1805 | { | |
1806 | pg.set_ps(ceph_stable_mod(pg.ps(), pg_num, pg_num_mask)); | |
1807 | return pg; | |
1808 | } | |
1809 | ||
1810 | /* | |
1811 | * map raw pg (full precision ps) into a placement seed. include | |
1812 | * pool id in that value so that different pools don't use the same | |
1813 | * seeds. | |
1814 | */ | |
1815 | ps_t pg_pool_t::raw_pg_to_pps(pg_t pg) const | |
1816 | { | |
1817 | if (flags & FLAG_HASHPSPOOL) { | |
1818 | // Hash the pool id so that pool PGs do not overlap. | |
1819 | return | |
1820 | crush_hash32_2(CRUSH_HASH_RJENKINS1, | |
1821 | ceph_stable_mod(pg.ps(), pgp_num, pgp_num_mask), | |
1822 | pg.pool()); | |
1823 | } else { | |
1824 | // Legacy behavior; add ps and pool together. This is not a great | |
1825 | // idea because the PGs from each pool will essentially overlap on | |
1826 | // top of each other: 0.5 == 1.4 == 2.3 == ... | |
1827 | return | |
1828 | ceph_stable_mod(pg.ps(), pgp_num, pgp_num_mask) + | |
1829 | pg.pool(); | |
1830 | } | |
1831 | } | |
1832 | ||
1833 | uint32_t pg_pool_t::get_random_pg_position(pg_t pg, uint32_t seed) const | |
1834 | { | |
1835 | uint32_t r = crush_hash32_2(CRUSH_HASH_RJENKINS1, seed, 123); | |
1836 | if (pg_num == pg_num_mask + 1) { | |
1837 | r &= ~pg_num_mask; | |
1838 | } else { | |
1839 | unsigned smaller_mask = pg_num_mask >> 1; | |
1840 | if ((pg.ps() & smaller_mask) < (pg_num & smaller_mask)) { | |
1841 | r &= ~pg_num_mask; | |
1842 | } else { | |
1843 | r &= ~smaller_mask; | |
1844 | } | |
1845 | } | |
1846 | r |= pg.ps(); | |
1847 | return r; | |
1848 | } | |
1849 | ||
9f95a23c | 1850 | void pg_pool_t::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 1851 | { |
11fdf7f2 | 1852 | using ceph::encode; |
7c673cae FG |
1853 | if ((features & CEPH_FEATURE_PGPOOL3) == 0) { |
1854 | // this encoding matches the old struct ceph_pg_pool | |
1855 | __u8 struct_v = 2; | |
11fdf7f2 TL |
1856 | encode(struct_v, bl); |
1857 | encode(type, bl); | |
1858 | encode(size, bl); | |
1859 | encode(crush_rule, bl); | |
1860 | encode(object_hash, bl); | |
1861 | encode(pg_num, bl); | |
1862 | encode(pgp_num, bl); | |
7c673cae | 1863 | __u32 lpg_num = 0, lpgp_num = 0; // tell old code that there are no localized pgs. |
11fdf7f2 TL |
1864 | encode(lpg_num, bl); |
1865 | encode(lpgp_num, bl); | |
1866 | encode(last_change, bl); | |
1867 | encode(snap_seq, bl); | |
1868 | encode(snap_epoch, bl); | |
7c673cae FG |
1869 | |
1870 | __u32 n = snaps.size(); | |
11fdf7f2 | 1871 | encode(n, bl); |
7c673cae | 1872 | n = removed_snaps.num_intervals(); |
11fdf7f2 | 1873 | encode(n, bl); |
7c673cae | 1874 | |
11fdf7f2 | 1875 | encode(auid, bl); |
7c673cae | 1876 | |
11fdf7f2 TL |
1877 | encode_nohead(snaps, bl, features); |
1878 | encode_nohead(removed_snaps, bl); | |
7c673cae FG |
1879 | return; |
1880 | } | |
1881 | ||
1882 | if ((features & CEPH_FEATURE_OSDENC) == 0) { | |
1883 | __u8 struct_v = 4; | |
11fdf7f2 TL |
1884 | encode(struct_v, bl); |
1885 | encode(type, bl); | |
1886 | encode(size, bl); | |
1887 | encode(crush_rule, bl); | |
1888 | encode(object_hash, bl); | |
1889 | encode(pg_num, bl); | |
1890 | encode(pgp_num, bl); | |
7c673cae | 1891 | __u32 lpg_num = 0, lpgp_num = 0; // tell old code that there are no localized pgs. |
11fdf7f2 TL |
1892 | encode(lpg_num, bl); |
1893 | encode(lpgp_num, bl); | |
1894 | encode(last_change, bl); | |
1895 | encode(snap_seq, bl); | |
1896 | encode(snap_epoch, bl); | |
1897 | encode(snaps, bl, features); | |
1898 | encode(removed_snaps, bl); | |
1899 | encode(auid, bl); | |
1900 | encode(flags, bl); | |
1901 | encode((uint32_t)0, bl); // crash_replay_interval | |
7c673cae FG |
1902 | return; |
1903 | } | |
1904 | ||
1905 | if ((features & CEPH_FEATURE_OSD_POOLRESEND) == 0) { | |
1906 | // we simply added last_force_op_resend here, which is a fully | |
1907 | // backward compatible change. however, encoding the same map | |
1908 | // differently between monitors triggers scrub noise (even though | |
1909 | // they are decodable without the feature), so let's be pendantic | |
1910 | // about it. | |
1911 | ENCODE_START(14, 5, bl); | |
11fdf7f2 TL |
1912 | encode(type, bl); |
1913 | encode(size, bl); | |
1914 | encode(crush_rule, bl); | |
1915 | encode(object_hash, bl); | |
1916 | encode(pg_num, bl); | |
1917 | encode(pgp_num, bl); | |
7c673cae | 1918 | __u32 lpg_num = 0, lpgp_num = 0; // tell old code that there are no localized pgs. |
11fdf7f2 TL |
1919 | encode(lpg_num, bl); |
1920 | encode(lpgp_num, bl); | |
1921 | encode(last_change, bl); | |
1922 | encode(snap_seq, bl); | |
1923 | encode(snap_epoch, bl); | |
1924 | encode(snaps, bl, features); | |
1925 | encode(removed_snaps, bl); | |
1926 | encode(auid, bl); | |
1927 | encode(flags, bl); | |
1928 | encode((uint32_t)0, bl); // crash_replay_interval | |
1929 | encode(min_size, bl); | |
1930 | encode(quota_max_bytes, bl); | |
1931 | encode(quota_max_objects, bl); | |
1932 | encode(tiers, bl); | |
1933 | encode(tier_of, bl); | |
7c673cae | 1934 | __u8 c = cache_mode; |
11fdf7f2 TL |
1935 | encode(c, bl); |
1936 | encode(read_tier, bl); | |
1937 | encode(write_tier, bl); | |
1938 | encode(properties, bl); | |
1939 | encode(hit_set_params, bl); | |
1940 | encode(hit_set_period, bl); | |
1941 | encode(hit_set_count, bl); | |
1942 | encode(stripe_width, bl); | |
1943 | encode(target_max_bytes, bl); | |
1944 | encode(target_max_objects, bl); | |
1945 | encode(cache_target_dirty_ratio_micro, bl); | |
1946 | encode(cache_target_full_ratio_micro, bl); | |
1947 | encode(cache_min_flush_age, bl); | |
1948 | encode(cache_min_evict_age, bl); | |
1949 | encode(erasure_code_profile, bl); | |
7c673cae FG |
1950 | ENCODE_FINISH(bl); |
1951 | return; | |
1952 | } | |
1953 | ||
f67539c2 | 1954 | uint8_t v = 30; |
28e407b8 AA |
1955 | // NOTE: any new encoding dependencies must be reflected by |
1956 | // SIGNIFICANT_FEATURES | |
7c673cae FG |
1957 | if (!(features & CEPH_FEATURE_NEW_OSDOP_ENCODING)) { |
1958 | // this was the first post-hammer thing we added; if it's missing, encode | |
1959 | // like hammer. | |
1960 | v = 21; | |
94b18763 | 1961 | } else if (!HAVE_FEATURE(features, SERVER_LUMINOUS)) { |
7c673cae | 1962 | v = 24; |
11fdf7f2 TL |
1963 | } else if (!HAVE_FEATURE(features, SERVER_MIMIC)) { |
1964 | v = 26; | |
1965 | } else if (!HAVE_FEATURE(features, SERVER_NAUTILUS)) { | |
1966 | v = 27; | |
f67539c2 TL |
1967 | } else if (!is_stretch_pool()) { |
1968 | v = 29; | |
7c673cae FG |
1969 | } |
1970 | ||
1971 | ENCODE_START(v, 5, bl); | |
11fdf7f2 TL |
1972 | encode(type, bl); |
1973 | encode(size, bl); | |
1974 | encode(crush_rule, bl); | |
1975 | encode(object_hash, bl); | |
1976 | encode(pg_num, bl); | |
1977 | encode(pgp_num, bl); | |
7c673cae | 1978 | __u32 lpg_num = 0, lpgp_num = 0; // tell old code that there are no localized pgs. |
11fdf7f2 TL |
1979 | encode(lpg_num, bl); |
1980 | encode(lpgp_num, bl); | |
1981 | encode(last_change, bl); | |
1982 | encode(snap_seq, bl); | |
1983 | encode(snap_epoch, bl); | |
1984 | encode(snaps, bl, features); | |
1985 | encode(removed_snaps, bl); | |
1986 | encode(auid, bl); | |
1987 | if (v >= 27) { | |
1988 | encode(flags, bl); | |
1989 | } else { | |
1990 | auto tmp = flags; | |
1991 | tmp &= ~(FLAG_SELFMANAGED_SNAPS | FLAG_POOL_SNAPS | FLAG_CREATING); | |
1992 | encode(tmp, bl); | |
1993 | } | |
1994 | encode((uint32_t)0, bl); // crash_replay_interval | |
1995 | encode(min_size, bl); | |
1996 | encode(quota_max_bytes, bl); | |
1997 | encode(quota_max_objects, bl); | |
1998 | encode(tiers, bl); | |
1999 | encode(tier_of, bl); | |
7c673cae | 2000 | __u8 c = cache_mode; |
11fdf7f2 TL |
2001 | encode(c, bl); |
2002 | encode(read_tier, bl); | |
2003 | encode(write_tier, bl); | |
2004 | encode(properties, bl); | |
2005 | encode(hit_set_params, bl); | |
2006 | encode(hit_set_period, bl); | |
2007 | encode(hit_set_count, bl); | |
2008 | encode(stripe_width, bl); | |
2009 | encode(target_max_bytes, bl); | |
2010 | encode(target_max_objects, bl); | |
2011 | encode(cache_target_dirty_ratio_micro, bl); | |
2012 | encode(cache_target_full_ratio_micro, bl); | |
2013 | encode(cache_min_flush_age, bl); | |
2014 | encode(cache_min_evict_age, bl); | |
2015 | encode(erasure_code_profile, bl); | |
2016 | encode(last_force_op_resend_preluminous, bl); | |
2017 | encode(min_read_recency_for_promote, bl); | |
2018 | encode(expected_num_objects, bl); | |
7c673cae | 2019 | if (v >= 19) { |
11fdf7f2 | 2020 | encode(cache_target_dirty_high_ratio_micro, bl); |
7c673cae FG |
2021 | } |
2022 | if (v >= 20) { | |
11fdf7f2 | 2023 | encode(min_write_recency_for_promote, bl); |
7c673cae FG |
2024 | } |
2025 | if (v >= 21) { | |
11fdf7f2 | 2026 | encode(use_gmt_hitset, bl); |
7c673cae FG |
2027 | } |
2028 | if (v >= 22) { | |
11fdf7f2 | 2029 | encode(fast_read, bl); |
7c673cae FG |
2030 | } |
2031 | if (v >= 23) { | |
11fdf7f2 TL |
2032 | encode(hit_set_grade_decay_rate, bl); |
2033 | encode(hit_set_search_last_n, bl); | |
7c673cae FG |
2034 | } |
2035 | if (v >= 24) { | |
11fdf7f2 | 2036 | encode(opts, bl, features); |
7c673cae FG |
2037 | } |
2038 | if (v >= 25) { | |
11fdf7f2 | 2039 | encode(last_force_op_resend_prenautilus, bl); |
7c673cae | 2040 | } |
c07f9fc5 | 2041 | if (v >= 26) { |
11fdf7f2 TL |
2042 | encode(application_metadata, bl); |
2043 | } | |
2044 | if (v >= 27) { | |
2045 | encode(create_time, bl); | |
2046 | } | |
2047 | if (v >= 28) { | |
2048 | encode(pg_num_target, bl); | |
2049 | encode(pgp_num_target, bl); | |
2050 | encode(pg_num_pending, bl); | |
2051 | encode((epoch_t)0, bl); // pg_num_dec_last_epoch_started from 14.1.[01] | |
2052 | encode((epoch_t)0, bl); // pg_num_dec_last_epoch_clean from 14.1.[01] | |
2053 | encode(last_force_op_resend, bl); | |
2054 | encode(pg_autoscale_mode, bl); | |
2055 | } | |
2056 | if (v >= 29) { | |
2057 | encode(last_pg_merge_meta, bl); | |
c07f9fc5 | 2058 | } |
f67539c2 TL |
2059 | if (v >= 30) { |
2060 | encode(peering_crush_bucket_count, bl); | |
2061 | encode(peering_crush_bucket_target, bl); | |
2062 | encode(peering_crush_bucket_barrier, bl); | |
2063 | encode(peering_crush_mandatory_member, bl); | |
2064 | } | |
7c673cae FG |
2065 | ENCODE_FINISH(bl); |
2066 | } | |
2067 | ||
9f95a23c | 2068 | void pg_pool_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 2069 | { |
f67539c2 | 2070 | DECODE_START_LEGACY_COMPAT_LEN(30, 5, 5, bl); |
11fdf7f2 TL |
2071 | decode(type, bl); |
2072 | decode(size, bl); | |
2073 | decode(crush_rule, bl); | |
2074 | decode(object_hash, bl); | |
2075 | decode(pg_num, bl); | |
2076 | decode(pgp_num, bl); | |
7c673cae FG |
2077 | { |
2078 | __u32 lpg_num, lpgp_num; | |
11fdf7f2 TL |
2079 | decode(lpg_num, bl); |
2080 | decode(lpgp_num, bl); | |
7c673cae | 2081 | } |
11fdf7f2 TL |
2082 | decode(last_change, bl); |
2083 | decode(snap_seq, bl); | |
2084 | decode(snap_epoch, bl); | |
7c673cae FG |
2085 | |
2086 | if (struct_v >= 3) { | |
11fdf7f2 TL |
2087 | decode(snaps, bl); |
2088 | decode(removed_snaps, bl); | |
2089 | decode(auid, bl); | |
7c673cae FG |
2090 | } else { |
2091 | __u32 n, m; | |
11fdf7f2 TL |
2092 | decode(n, bl); |
2093 | decode(m, bl); | |
2094 | decode(auid, bl); | |
2095 | decode_nohead(n, snaps, bl); | |
2096 | decode_nohead(m, removed_snaps, bl); | |
7c673cae FG |
2097 | } |
2098 | ||
2099 | if (struct_v >= 4) { | |
11fdf7f2 TL |
2100 | decode(flags, bl); |
2101 | uint32_t crash_replay_interval; | |
2102 | decode(crash_replay_interval, bl); | |
7c673cae FG |
2103 | } else { |
2104 | flags = 0; | |
11fdf7f2 TL |
2105 | } |
2106 | // upgrade path for selfmanaged vs pool snaps | |
2107 | if (snap_seq > 0 && (flags & (FLAG_SELFMANAGED_SNAPS|FLAG_POOL_SNAPS)) == 0) { | |
2108 | if (!removed_snaps.empty()) { | |
2109 | flags |= FLAG_SELFMANAGED_SNAPS; | |
2110 | } else { | |
2111 | flags |= FLAG_POOL_SNAPS; | |
2112 | } | |
7c673cae FG |
2113 | } |
2114 | if (struct_v >= 7) { | |
11fdf7f2 | 2115 | decode(min_size, bl); |
7c673cae FG |
2116 | } else { |
2117 | min_size = size - size/2; | |
2118 | } | |
2119 | if (struct_v >= 8) { | |
11fdf7f2 TL |
2120 | decode(quota_max_bytes, bl); |
2121 | decode(quota_max_objects, bl); | |
7c673cae FG |
2122 | } |
2123 | if (struct_v >= 9) { | |
11fdf7f2 TL |
2124 | decode(tiers, bl); |
2125 | decode(tier_of, bl); | |
7c673cae | 2126 | __u8 v; |
11fdf7f2 | 2127 | decode(v, bl); |
7c673cae | 2128 | cache_mode = (cache_mode_t)v; |
11fdf7f2 TL |
2129 | decode(read_tier, bl); |
2130 | decode(write_tier, bl); | |
7c673cae FG |
2131 | } |
2132 | if (struct_v >= 10) { | |
11fdf7f2 | 2133 | decode(properties, bl); |
7c673cae FG |
2134 | } |
2135 | if (struct_v >= 11) { | |
11fdf7f2 TL |
2136 | decode(hit_set_params, bl); |
2137 | decode(hit_set_period, bl); | |
2138 | decode(hit_set_count, bl); | |
7c673cae FG |
2139 | } else { |
2140 | pg_pool_t def; | |
2141 | hit_set_period = def.hit_set_period; | |
2142 | hit_set_count = def.hit_set_count; | |
2143 | } | |
2144 | if (struct_v >= 12) { | |
11fdf7f2 | 2145 | decode(stripe_width, bl); |
7c673cae FG |
2146 | } else { |
2147 | set_stripe_width(0); | |
2148 | } | |
2149 | if (struct_v >= 13) { | |
11fdf7f2 TL |
2150 | decode(target_max_bytes, bl); |
2151 | decode(target_max_objects, bl); | |
2152 | decode(cache_target_dirty_ratio_micro, bl); | |
2153 | decode(cache_target_full_ratio_micro, bl); | |
2154 | decode(cache_min_flush_age, bl); | |
2155 | decode(cache_min_evict_age, bl); | |
7c673cae FG |
2156 | } else { |
2157 | target_max_bytes = 0; | |
2158 | target_max_objects = 0; | |
2159 | cache_target_dirty_ratio_micro = 0; | |
2160 | cache_target_full_ratio_micro = 0; | |
2161 | cache_min_flush_age = 0; | |
2162 | cache_min_evict_age = 0; | |
2163 | } | |
2164 | if (struct_v >= 14) { | |
11fdf7f2 | 2165 | decode(erasure_code_profile, bl); |
7c673cae FG |
2166 | } |
2167 | if (struct_v >= 15) { | |
11fdf7f2 | 2168 | decode(last_force_op_resend_preluminous, bl); |
7c673cae FG |
2169 | } else { |
2170 | last_force_op_resend_preluminous = 0; | |
2171 | } | |
2172 | if (struct_v >= 16) { | |
11fdf7f2 | 2173 | decode(min_read_recency_for_promote, bl); |
7c673cae FG |
2174 | } else { |
2175 | min_read_recency_for_promote = 1; | |
2176 | } | |
2177 | if (struct_v >= 17) { | |
11fdf7f2 | 2178 | decode(expected_num_objects, bl); |
7c673cae FG |
2179 | } else { |
2180 | expected_num_objects = 0; | |
2181 | } | |
2182 | if (struct_v >= 19) { | |
11fdf7f2 | 2183 | decode(cache_target_dirty_high_ratio_micro, bl); |
7c673cae FG |
2184 | } else { |
2185 | cache_target_dirty_high_ratio_micro = cache_target_dirty_ratio_micro; | |
2186 | } | |
2187 | if (struct_v >= 20) { | |
11fdf7f2 | 2188 | decode(min_write_recency_for_promote, bl); |
7c673cae FG |
2189 | } else { |
2190 | min_write_recency_for_promote = 1; | |
2191 | } | |
2192 | if (struct_v >= 21) { | |
11fdf7f2 | 2193 | decode(use_gmt_hitset, bl); |
7c673cae FG |
2194 | } else { |
2195 | use_gmt_hitset = false; | |
2196 | } | |
2197 | if (struct_v >= 22) { | |
11fdf7f2 | 2198 | decode(fast_read, bl); |
7c673cae FG |
2199 | } else { |
2200 | fast_read = false; | |
2201 | } | |
2202 | if (struct_v >= 23) { | |
11fdf7f2 TL |
2203 | decode(hit_set_grade_decay_rate, bl); |
2204 | decode(hit_set_search_last_n, bl); | |
7c673cae FG |
2205 | } else { |
2206 | hit_set_grade_decay_rate = 0; | |
2207 | hit_set_search_last_n = 1; | |
2208 | } | |
2209 | if (struct_v >= 24) { | |
11fdf7f2 | 2210 | decode(opts, bl); |
7c673cae FG |
2211 | } |
2212 | if (struct_v >= 25) { | |
11fdf7f2 | 2213 | decode(last_force_op_resend_prenautilus, bl); |
7c673cae | 2214 | } else { |
11fdf7f2 | 2215 | last_force_op_resend_prenautilus = last_force_op_resend_preluminous; |
7c673cae | 2216 | } |
c07f9fc5 | 2217 | if (struct_v >= 26) { |
11fdf7f2 TL |
2218 | decode(application_metadata, bl); |
2219 | } | |
2220 | if (struct_v >= 27) { | |
2221 | decode(create_time, bl); | |
2222 | } | |
2223 | if (struct_v >= 28) { | |
2224 | decode(pg_num_target, bl); | |
2225 | decode(pgp_num_target, bl); | |
2226 | decode(pg_num_pending, bl); | |
2227 | epoch_t old_merge_last_epoch_clean, old_merge_last_epoch_started; | |
2228 | decode(old_merge_last_epoch_started, bl); | |
2229 | decode(old_merge_last_epoch_clean, bl); | |
2230 | decode(last_force_op_resend, bl); | |
2231 | decode(pg_autoscale_mode, bl); | |
2232 | if (struct_v >= 29) { | |
2233 | decode(last_pg_merge_meta, bl); | |
2234 | } else { | |
2235 | last_pg_merge_meta.last_epoch_clean = old_merge_last_epoch_clean; | |
2236 | last_pg_merge_meta.last_epoch_started = old_merge_last_epoch_started; | |
2237 | } | |
2238 | } else { | |
2239 | pg_num_target = pg_num; | |
2240 | pgp_num_target = pgp_num; | |
2241 | pg_num_pending = pg_num; | |
2242 | last_force_op_resend = last_force_op_resend_prenautilus; | |
9f95a23c | 2243 | pg_autoscale_mode = pg_autoscale_mode_t::WARN; // default to warn on upgrade |
c07f9fc5 | 2244 | } |
f67539c2 TL |
2245 | if (struct_v >= 30) { |
2246 | decode(peering_crush_bucket_count, bl); | |
2247 | decode(peering_crush_bucket_target, bl); | |
2248 | decode(peering_crush_bucket_barrier, bl); | |
2249 | decode(peering_crush_mandatory_member, bl); | |
2250 | } | |
7c673cae FG |
2251 | DECODE_FINISH(bl); |
2252 | calc_pg_masks(); | |
2253 | calc_grade_table(); | |
2254 | } | |
2255 | ||
f67539c2 TL |
2256 | bool pg_pool_t::stretch_set_can_peer(const set<int>& want, const OSDMap& osdmap, |
2257 | std::ostream * out) const | |
2258 | { | |
2259 | if (!is_stretch_pool()) return true; | |
2260 | const uint32_t barrier_id = peering_crush_bucket_barrier; | |
2261 | const uint32_t barrier_count = peering_crush_bucket_count; | |
2262 | set<int> ancestors; | |
2263 | const shared_ptr<CrushWrapper>& crush = osdmap.crush; | |
2264 | for (int osdid : want) { | |
2265 | int ancestor = crush->get_parent_of_type(osdid, barrier_id, | |
2266 | crush_rule); | |
2267 | ancestors.insert(ancestor); | |
2268 | } | |
2269 | if (ancestors.size() < barrier_count) { | |
2270 | if (out) { | |
2271 | *out << __func__ << ": not enough crush buckets with OSDs in want set " | |
2272 | << want; | |
2273 | } | |
2274 | return false; | |
2275 | } else if (peering_crush_mandatory_member != CRUSH_ITEM_NONE && | |
2276 | !ancestors.count(peering_crush_mandatory_member)) { | |
2277 | if (out) { | |
2278 | *out << __func__ << ": missing mandatory crush bucket member " | |
2279 | << peering_crush_mandatory_member; | |
2280 | } | |
2281 | return false; | |
2282 | } | |
2283 | return true; | |
2284 | } | |
2285 | ||
7c673cae FG |
2286 | void pg_pool_t::generate_test_instances(list<pg_pool_t*>& o) |
2287 | { | |
2288 | pg_pool_t a; | |
2289 | o.push_back(new pg_pool_t(a)); | |
2290 | ||
11fdf7f2 | 2291 | a.create_time = utime_t(4,5); |
7c673cae FG |
2292 | a.type = TYPE_REPLICATED; |
2293 | a.size = 2; | |
31f18b77 | 2294 | a.crush_rule = 3; |
7c673cae FG |
2295 | a.object_hash = 4; |
2296 | a.pg_num = 6; | |
11fdf7f2 TL |
2297 | a.pgp_num = 4; |
2298 | a.pgp_num_target = 4; | |
2299 | a.pg_num_target = 5; | |
2300 | a.pg_num_pending = 5; | |
2301 | a.last_pg_merge_meta.last_epoch_started = 2; | |
2302 | a.last_pg_merge_meta.last_epoch_clean = 2; | |
7c673cae FG |
2303 | a.last_change = 9; |
2304 | a.last_force_op_resend = 123823; | |
2305 | a.last_force_op_resend_preluminous = 123824; | |
2306 | a.snap_seq = 10; | |
2307 | a.snap_epoch = 11; | |
11fdf7f2 | 2308 | a.flags = FLAG_POOL_SNAPS; |
7c673cae | 2309 | a.auid = 12; |
7c673cae FG |
2310 | a.quota_max_bytes = 473; |
2311 | a.quota_max_objects = 474; | |
2312 | o.push_back(new pg_pool_t(a)); | |
2313 | ||
2314 | a.snaps[3].name = "asdf"; | |
2315 | a.snaps[3].snapid = 3; | |
2316 | a.snaps[3].stamp = utime_t(123, 4); | |
2317 | a.snaps[6].name = "qwer"; | |
2318 | a.snaps[6].snapid = 6; | |
2319 | a.snaps[6].stamp = utime_t(23423, 4); | |
2320 | o.push_back(new pg_pool_t(a)); | |
2321 | ||
11fdf7f2 TL |
2322 | a.flags = FLAG_SELFMANAGED_SNAPS; |
2323 | a.snaps.clear(); | |
2324 | a.removed_snaps.insert(2); | |
7c673cae FG |
2325 | a.quota_max_bytes = 2473; |
2326 | a.quota_max_objects = 4374; | |
2327 | a.tiers.insert(0); | |
2328 | a.tiers.insert(1); | |
2329 | a.tier_of = 2; | |
2330 | a.cache_mode = CACHEMODE_WRITEBACK; | |
2331 | a.read_tier = 1; | |
2332 | a.write_tier = 1; | |
2333 | a.hit_set_params = HitSet::Params(new BloomHitSet::Params); | |
2334 | a.hit_set_period = 3600; | |
2335 | a.hit_set_count = 8; | |
2336 | a.min_read_recency_for_promote = 1; | |
2337 | a.min_write_recency_for_promote = 1; | |
2338 | a.hit_set_grade_decay_rate = 50; | |
2339 | a.hit_set_search_last_n = 1; | |
2340 | a.calc_grade_table(); | |
2341 | a.set_stripe_width(12345); | |
2342 | a.target_max_bytes = 1238132132; | |
2343 | a.target_max_objects = 1232132; | |
2344 | a.cache_target_dirty_ratio_micro = 187232; | |
2345 | a.cache_target_dirty_high_ratio_micro = 309856; | |
2346 | a.cache_target_full_ratio_micro = 987222; | |
2347 | a.cache_min_flush_age = 231; | |
2348 | a.cache_min_evict_age = 2321; | |
2349 | a.erasure_code_profile = "profile in osdmap"; | |
2350 | a.expected_num_objects = 123456; | |
2351 | a.fast_read = false; | |
c07f9fc5 | 2352 | a.application_metadata = {{"rbd", {{"key", "value"}}}}; |
7c673cae FG |
2353 | o.push_back(new pg_pool_t(a)); |
2354 | } | |
2355 | ||
2356 | ostream& operator<<(ostream& out, const pg_pool_t& p) | |
2357 | { | |
9f95a23c TL |
2358 | out << p.get_type_name(); |
2359 | if (p.get_type_name() == "erasure") { | |
2360 | out << " profile " << p.erasure_code_profile; | |
2361 | } | |
2362 | out << " size " << p.get_size() | |
7c673cae | 2363 | << " min_size " << p.get_min_size() |
31f18b77 | 2364 | << " crush_rule " << p.get_crush_rule() |
7c673cae FG |
2365 | << " object_hash " << p.get_object_hash_name() |
2366 | << " pg_num " << p.get_pg_num() | |
11fdf7f2 TL |
2367 | << " pgp_num " << p.get_pgp_num(); |
2368 | if (p.get_pg_num_target() != p.get_pg_num()) { | |
2369 | out << " pg_num_target " << p.get_pg_num_target(); | |
2370 | } | |
2371 | if (p.get_pgp_num_target() != p.get_pgp_num()) { | |
2372 | out << " pgp_num_target " << p.get_pgp_num_target(); | |
2373 | } | |
2374 | if (p.get_pg_num_pending() != p.get_pg_num()) { | |
2375 | out << " pg_num_pending " << p.get_pg_num_pending(); | |
2376 | } | |
9f95a23c | 2377 | if (p.pg_autoscale_mode != pg_pool_t::pg_autoscale_mode_t::UNKNOWN) { |
11fdf7f2 TL |
2378 | out << " autoscale_mode " << p.get_pg_autoscale_mode_name(p.pg_autoscale_mode); |
2379 | } | |
2380 | out << " last_change " << p.get_last_change(); | |
7c673cae | 2381 | if (p.get_last_force_op_resend() || |
11fdf7f2 | 2382 | p.get_last_force_op_resend_prenautilus() || |
7c673cae FG |
2383 | p.get_last_force_op_resend_preluminous()) |
2384 | out << " lfor " << p.get_last_force_op_resend() << "/" | |
11fdf7f2 | 2385 | << p.get_last_force_op_resend_prenautilus() << "/" |
7c673cae FG |
2386 | << p.get_last_force_op_resend_preluminous(); |
2387 | if (p.get_auid()) | |
2388 | out << " owner " << p.get_auid(); | |
2389 | if (p.flags) | |
2390 | out << " flags " << p.get_flags_string(); | |
7c673cae FG |
2391 | if (p.quota_max_bytes) |
2392 | out << " max_bytes " << p.quota_max_bytes; | |
2393 | if (p.quota_max_objects) | |
2394 | out << " max_objects " << p.quota_max_objects; | |
2395 | if (!p.tiers.empty()) | |
2396 | out << " tiers " << p.tiers; | |
2397 | if (p.is_tier()) | |
2398 | out << " tier_of " << p.tier_of; | |
2399 | if (p.has_read_tier()) | |
2400 | out << " read_tier " << p.read_tier; | |
2401 | if (p.has_write_tier()) | |
2402 | out << " write_tier " << p.write_tier; | |
2403 | if (p.cache_mode) | |
2404 | out << " cache_mode " << p.get_cache_mode_name(); | |
2405 | if (p.target_max_bytes) | |
2406 | out << " target_bytes " << p.target_max_bytes; | |
2407 | if (p.target_max_objects) | |
2408 | out << " target_objects " << p.target_max_objects; | |
2409 | if (p.hit_set_params.get_type() != HitSet::TYPE_NONE) { | |
2410 | out << " hit_set " << p.hit_set_params | |
2411 | << " " << p.hit_set_period << "s" | |
2412 | << " x" << p.hit_set_count << " decay_rate " | |
2413 | << p.hit_set_grade_decay_rate | |
2414 | << " search_last_n " << p.hit_set_search_last_n; | |
2415 | } | |
2416 | if (p.min_read_recency_for_promote) | |
2417 | out << " min_read_recency_for_promote " << p.min_read_recency_for_promote; | |
2418 | if (p.min_write_recency_for_promote) | |
2419 | out << " min_write_recency_for_promote " << p.min_write_recency_for_promote; | |
2420 | out << " stripe_width " << p.get_stripe_width(); | |
2421 | if (p.expected_num_objects) | |
2422 | out << " expected_num_objects " << p.expected_num_objects; | |
2423 | if (p.fast_read) | |
2424 | out << " fast_read " << p.fast_read; | |
2425 | out << p.opts; | |
c07f9fc5 FG |
2426 | if (!p.application_metadata.empty()) { |
2427 | out << " application "; | |
2428 | for (auto it = p.application_metadata.begin(); | |
2429 | it != p.application_metadata.end(); ++it) { | |
2430 | if (it != p.application_metadata.begin()) | |
2431 | out << ","; | |
2432 | out << it->first; | |
2433 | } | |
2434 | } | |
7c673cae FG |
2435 | return out; |
2436 | } | |
2437 | ||
2438 | ||
2439 | // -- object_stat_sum_t -- | |
2440 | ||
2441 | void object_stat_sum_t::dump(Formatter *f) const | |
2442 | { | |
2443 | f->dump_int("num_bytes", num_bytes); | |
2444 | f->dump_int("num_objects", num_objects); | |
2445 | f->dump_int("num_object_clones", num_object_clones); | |
2446 | f->dump_int("num_object_copies", num_object_copies); | |
2447 | f->dump_int("num_objects_missing_on_primary", num_objects_missing_on_primary); | |
2448 | f->dump_int("num_objects_missing", num_objects_missing); | |
2449 | f->dump_int("num_objects_degraded", num_objects_degraded); | |
2450 | f->dump_int("num_objects_misplaced", num_objects_misplaced); | |
2451 | f->dump_int("num_objects_unfound", num_objects_unfound); | |
2452 | f->dump_int("num_objects_dirty", num_objects_dirty); | |
2453 | f->dump_int("num_whiteouts", num_whiteouts); | |
2454 | f->dump_int("num_read", num_rd); | |
2455 | f->dump_int("num_read_kb", num_rd_kb); | |
2456 | f->dump_int("num_write", num_wr); | |
2457 | f->dump_int("num_write_kb", num_wr_kb); | |
2458 | f->dump_int("num_scrub_errors", num_scrub_errors); | |
2459 | f->dump_int("num_shallow_scrub_errors", num_shallow_scrub_errors); | |
2460 | f->dump_int("num_deep_scrub_errors", num_deep_scrub_errors); | |
2461 | f->dump_int("num_objects_recovered", num_objects_recovered); | |
2462 | f->dump_int("num_bytes_recovered", num_bytes_recovered); | |
2463 | f->dump_int("num_keys_recovered", num_keys_recovered); | |
2464 | f->dump_int("num_objects_omap", num_objects_omap); | |
2465 | f->dump_int("num_objects_hit_set_archive", num_objects_hit_set_archive); | |
2466 | f->dump_int("num_bytes_hit_set_archive", num_bytes_hit_set_archive); | |
2467 | f->dump_int("num_flush", num_flush); | |
2468 | f->dump_int("num_flush_kb", num_flush_kb); | |
2469 | f->dump_int("num_evict", num_evict); | |
2470 | f->dump_int("num_evict_kb", num_evict_kb); | |
2471 | f->dump_int("num_promote", num_promote); | |
2472 | f->dump_int("num_flush_mode_high", num_flush_mode_high); | |
2473 | f->dump_int("num_flush_mode_low", num_flush_mode_low); | |
2474 | f->dump_int("num_evict_mode_some", num_evict_mode_some); | |
2475 | f->dump_int("num_evict_mode_full", num_evict_mode_full); | |
2476 | f->dump_int("num_objects_pinned", num_objects_pinned); | |
2477 | f->dump_int("num_legacy_snapsets", num_legacy_snapsets); | |
28e407b8 | 2478 | f->dump_int("num_large_omap_objects", num_large_omap_objects); |
11fdf7f2 TL |
2479 | f->dump_int("num_objects_manifest", num_objects_manifest); |
2480 | f->dump_int("num_omap_bytes", num_omap_bytes); | |
2481 | f->dump_int("num_omap_keys", num_omap_keys); | |
2482 | f->dump_int("num_objects_repaired", num_objects_repaired); | |
7c673cae FG |
2483 | } |
2484 | ||
9f95a23c | 2485 | void object_stat_sum_t::encode(ceph::buffer::list& bl) const |
7c673cae | 2486 | { |
11fdf7f2 | 2487 | ENCODE_START(20, 14, bl); |
7c673cae FG |
2488 | #if defined(CEPH_LITTLE_ENDIAN) |
2489 | bl.append((char *)(&num_bytes), sizeof(object_stat_sum_t)); | |
2490 | #else | |
11fdf7f2 TL |
2491 | encode(num_bytes, bl); |
2492 | encode(num_objects, bl); | |
2493 | encode(num_object_clones, bl); | |
2494 | encode(num_object_copies, bl); | |
2495 | encode(num_objects_missing_on_primary, bl); | |
2496 | encode(num_objects_degraded, bl); | |
2497 | encode(num_objects_unfound, bl); | |
2498 | encode(num_rd, bl); | |
2499 | encode(num_rd_kb, bl); | |
2500 | encode(num_wr, bl); | |
2501 | encode(num_wr_kb, bl); | |
2502 | encode(num_scrub_errors, bl); | |
2503 | encode(num_objects_recovered, bl); | |
2504 | encode(num_bytes_recovered, bl); | |
2505 | encode(num_keys_recovered, bl); | |
2506 | encode(num_shallow_scrub_errors, bl); | |
2507 | encode(num_deep_scrub_errors, bl); | |
2508 | encode(num_objects_dirty, bl); | |
2509 | encode(num_whiteouts, bl); | |
2510 | encode(num_objects_omap, bl); | |
2511 | encode(num_objects_hit_set_archive, bl); | |
2512 | encode(num_objects_misplaced, bl); | |
2513 | encode(num_bytes_hit_set_archive, bl); | |
2514 | encode(num_flush, bl); | |
2515 | encode(num_flush_kb, bl); | |
2516 | encode(num_evict, bl); | |
2517 | encode(num_evict_kb, bl); | |
2518 | encode(num_promote, bl); | |
2519 | encode(num_flush_mode_high, bl); | |
2520 | encode(num_flush_mode_low, bl); | |
2521 | encode(num_evict_mode_some, bl); | |
2522 | encode(num_evict_mode_full, bl); | |
2523 | encode(num_objects_pinned, bl); | |
2524 | encode(num_objects_missing, bl); | |
2525 | encode(num_legacy_snapsets, bl); | |
2526 | encode(num_large_omap_objects, bl); | |
2527 | encode(num_objects_manifest, bl); | |
2528 | encode(num_omap_bytes, bl); | |
2529 | encode(num_omap_keys, bl); | |
2530 | encode(num_objects_repaired, bl); | |
7c673cae FG |
2531 | #endif |
2532 | ENCODE_FINISH(bl); | |
2533 | } | |
2534 | ||
9f95a23c | 2535 | void object_stat_sum_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
2536 | { |
2537 | bool decode_finish = false; | |
11fdf7f2 TL |
2538 | static const int STAT_SUM_DECODE_VERSION = 20; |
2539 | DECODE_START(STAT_SUM_DECODE_VERSION, bl); | |
7c673cae | 2540 | #if defined(CEPH_LITTLE_ENDIAN) |
11fdf7f2 | 2541 | if (struct_v == STAT_SUM_DECODE_VERSION) { |
7c673cae FG |
2542 | bl.copy(sizeof(object_stat_sum_t), (char*)(&num_bytes)); |
2543 | decode_finish = true; | |
2544 | } | |
2545 | #endif | |
2546 | if (!decode_finish) { | |
11fdf7f2 TL |
2547 | decode(num_bytes, bl); |
2548 | decode(num_objects, bl); | |
2549 | decode(num_object_clones, bl); | |
2550 | decode(num_object_copies, bl); | |
2551 | decode(num_objects_missing_on_primary, bl); | |
2552 | decode(num_objects_degraded, bl); | |
2553 | decode(num_objects_unfound, bl); | |
2554 | decode(num_rd, bl); | |
2555 | decode(num_rd_kb, bl); | |
2556 | decode(num_wr, bl); | |
2557 | decode(num_wr_kb, bl); | |
2558 | decode(num_scrub_errors, bl); | |
2559 | decode(num_objects_recovered, bl); | |
2560 | decode(num_bytes_recovered, bl); | |
2561 | decode(num_keys_recovered, bl); | |
2562 | decode(num_shallow_scrub_errors, bl); | |
2563 | decode(num_deep_scrub_errors, bl); | |
2564 | decode(num_objects_dirty, bl); | |
2565 | decode(num_whiteouts, bl); | |
2566 | decode(num_objects_omap, bl); | |
2567 | decode(num_objects_hit_set_archive, bl); | |
2568 | decode(num_objects_misplaced, bl); | |
2569 | decode(num_bytes_hit_set_archive, bl); | |
2570 | decode(num_flush, bl); | |
2571 | decode(num_flush_kb, bl); | |
2572 | decode(num_evict, bl); | |
2573 | decode(num_evict_kb, bl); | |
2574 | decode(num_promote, bl); | |
2575 | decode(num_flush_mode_high, bl); | |
2576 | decode(num_flush_mode_low, bl); | |
2577 | decode(num_evict_mode_some, bl); | |
2578 | decode(num_evict_mode_full, bl); | |
2579 | decode(num_objects_pinned, bl); | |
2580 | decode(num_objects_missing, bl); | |
7c673cae | 2581 | if (struct_v >= 16) { |
11fdf7f2 | 2582 | decode(num_legacy_snapsets, bl); |
7c673cae FG |
2583 | } else { |
2584 | num_legacy_snapsets = num_object_clones; // upper bound | |
2585 | } | |
28e407b8 | 2586 | if (struct_v >= 17) { |
11fdf7f2 TL |
2587 | decode(num_large_omap_objects, bl); |
2588 | } | |
2589 | if (struct_v >= 18) { | |
2590 | decode(num_objects_manifest, bl); | |
2591 | } | |
2592 | if (struct_v >= 19) { | |
2593 | decode(num_omap_bytes, bl); | |
2594 | decode(num_omap_keys, bl); | |
2595 | } | |
2596 | if (struct_v >= 20) { | |
2597 | decode(num_objects_repaired, bl); | |
28e407b8 | 2598 | } |
7c673cae FG |
2599 | } |
2600 | DECODE_FINISH(bl); | |
2601 | } | |
2602 | ||
2603 | void object_stat_sum_t::generate_test_instances(list<object_stat_sum_t*>& o) | |
2604 | { | |
2605 | object_stat_sum_t a; | |
2606 | ||
2607 | a.num_bytes = 1; | |
2608 | a.num_objects = 3; | |
2609 | a.num_object_clones = 4; | |
2610 | a.num_object_copies = 5; | |
2611 | a.num_objects_missing_on_primary = 6; | |
2612 | a.num_objects_missing = 123; | |
2613 | a.num_objects_degraded = 7; | |
2614 | a.num_objects_unfound = 8; | |
2615 | a.num_rd = 9; a.num_rd_kb = 10; | |
2616 | a.num_wr = 11; a.num_wr_kb = 12; | |
2617 | a.num_objects_recovered = 14; | |
2618 | a.num_bytes_recovered = 15; | |
2619 | a.num_keys_recovered = 16; | |
2620 | a.num_deep_scrub_errors = 17; | |
2621 | a.num_shallow_scrub_errors = 18; | |
2622 | a.num_scrub_errors = a.num_deep_scrub_errors + a.num_shallow_scrub_errors; | |
2623 | a.num_objects_dirty = 21; | |
2624 | a.num_whiteouts = 22; | |
2625 | a.num_objects_misplaced = 1232; | |
2626 | a.num_objects_hit_set_archive = 2; | |
2627 | a.num_bytes_hit_set_archive = 27; | |
2628 | a.num_flush = 5; | |
2629 | a.num_flush_kb = 6; | |
2630 | a.num_evict = 7; | |
2631 | a.num_evict_kb = 8; | |
2632 | a.num_promote = 9; | |
2633 | a.num_flush_mode_high = 0; | |
2634 | a.num_flush_mode_low = 1; | |
2635 | a.num_evict_mode_some = 1; | |
2636 | a.num_evict_mode_full = 0; | |
2637 | a.num_objects_pinned = 20; | |
28e407b8 | 2638 | a.num_large_omap_objects = 5; |
11fdf7f2 TL |
2639 | a.num_objects_manifest = 2; |
2640 | a.num_omap_bytes = 20000; | |
2641 | a.num_omap_keys = 200; | |
2642 | a.num_objects_repaired = 300; | |
7c673cae FG |
2643 | o.push_back(new object_stat_sum_t(a)); |
2644 | } | |
2645 | ||
2646 | void object_stat_sum_t::add(const object_stat_sum_t& o) | |
2647 | { | |
2648 | num_bytes += o.num_bytes; | |
2649 | num_objects += o.num_objects; | |
2650 | num_object_clones += o.num_object_clones; | |
2651 | num_object_copies += o.num_object_copies; | |
2652 | num_objects_missing_on_primary += o.num_objects_missing_on_primary; | |
2653 | num_objects_missing += o.num_objects_missing; | |
2654 | num_objects_degraded += o.num_objects_degraded; | |
2655 | num_objects_misplaced += o.num_objects_misplaced; | |
2656 | num_rd += o.num_rd; | |
2657 | num_rd_kb += o.num_rd_kb; | |
2658 | num_wr += o.num_wr; | |
2659 | num_wr_kb += o.num_wr_kb; | |
2660 | num_objects_unfound += o.num_objects_unfound; | |
2661 | num_scrub_errors += o.num_scrub_errors; | |
2662 | num_shallow_scrub_errors += o.num_shallow_scrub_errors; | |
2663 | num_deep_scrub_errors += o.num_deep_scrub_errors; | |
2664 | num_objects_recovered += o.num_objects_recovered; | |
2665 | num_bytes_recovered += o.num_bytes_recovered; | |
2666 | num_keys_recovered += o.num_keys_recovered; | |
2667 | num_objects_dirty += o.num_objects_dirty; | |
2668 | num_whiteouts += o.num_whiteouts; | |
2669 | num_objects_omap += o.num_objects_omap; | |
2670 | num_objects_hit_set_archive += o.num_objects_hit_set_archive; | |
2671 | num_bytes_hit_set_archive += o.num_bytes_hit_set_archive; | |
2672 | num_flush += o.num_flush; | |
2673 | num_flush_kb += o.num_flush_kb; | |
2674 | num_evict += o.num_evict; | |
2675 | num_evict_kb += o.num_evict_kb; | |
2676 | num_promote += o.num_promote; | |
2677 | num_flush_mode_high += o.num_flush_mode_high; | |
2678 | num_flush_mode_low += o.num_flush_mode_low; | |
2679 | num_evict_mode_some += o.num_evict_mode_some; | |
2680 | num_evict_mode_full += o.num_evict_mode_full; | |
2681 | num_objects_pinned += o.num_objects_pinned; | |
2682 | num_legacy_snapsets += o.num_legacy_snapsets; | |
28e407b8 | 2683 | num_large_omap_objects += o.num_large_omap_objects; |
11fdf7f2 TL |
2684 | num_objects_manifest += o.num_objects_manifest; |
2685 | num_omap_bytes += o.num_omap_bytes; | |
2686 | num_omap_keys += o.num_omap_keys; | |
2687 | num_objects_repaired += o.num_objects_repaired; | |
7c673cae FG |
2688 | } |
2689 | ||
2690 | void object_stat_sum_t::sub(const object_stat_sum_t& o) | |
2691 | { | |
2692 | num_bytes -= o.num_bytes; | |
2693 | num_objects -= o.num_objects; | |
2694 | num_object_clones -= o.num_object_clones; | |
2695 | num_object_copies -= o.num_object_copies; | |
2696 | num_objects_missing_on_primary -= o.num_objects_missing_on_primary; | |
2697 | num_objects_missing -= o.num_objects_missing; | |
2698 | num_objects_degraded -= o.num_objects_degraded; | |
2699 | num_objects_misplaced -= o.num_objects_misplaced; | |
2700 | num_rd -= o.num_rd; | |
2701 | num_rd_kb -= o.num_rd_kb; | |
2702 | num_wr -= o.num_wr; | |
2703 | num_wr_kb -= o.num_wr_kb; | |
2704 | num_objects_unfound -= o.num_objects_unfound; | |
2705 | num_scrub_errors -= o.num_scrub_errors; | |
2706 | num_shallow_scrub_errors -= o.num_shallow_scrub_errors; | |
2707 | num_deep_scrub_errors -= o.num_deep_scrub_errors; | |
2708 | num_objects_recovered -= o.num_objects_recovered; | |
2709 | num_bytes_recovered -= o.num_bytes_recovered; | |
2710 | num_keys_recovered -= o.num_keys_recovered; | |
2711 | num_objects_dirty -= o.num_objects_dirty; | |
2712 | num_whiteouts -= o.num_whiteouts; | |
2713 | num_objects_omap -= o.num_objects_omap; | |
2714 | num_objects_hit_set_archive -= o.num_objects_hit_set_archive; | |
2715 | num_bytes_hit_set_archive -= o.num_bytes_hit_set_archive; | |
2716 | num_flush -= o.num_flush; | |
2717 | num_flush_kb -= o.num_flush_kb; | |
2718 | num_evict -= o.num_evict; | |
2719 | num_evict_kb -= o.num_evict_kb; | |
2720 | num_promote -= o.num_promote; | |
2721 | num_flush_mode_high -= o.num_flush_mode_high; | |
2722 | num_flush_mode_low -= o.num_flush_mode_low; | |
2723 | num_evict_mode_some -= o.num_evict_mode_some; | |
2724 | num_evict_mode_full -= o.num_evict_mode_full; | |
2725 | num_objects_pinned -= o.num_objects_pinned; | |
2726 | num_legacy_snapsets -= o.num_legacy_snapsets; | |
28e407b8 | 2727 | num_large_omap_objects -= o.num_large_omap_objects; |
11fdf7f2 TL |
2728 | num_objects_manifest -= o.num_objects_manifest; |
2729 | num_omap_bytes -= o.num_omap_bytes; | |
2730 | num_omap_keys -= o.num_omap_keys; | |
2731 | num_objects_repaired -= o.num_objects_repaired; | |
7c673cae FG |
2732 | } |
2733 | ||
2734 | bool operator==(const object_stat_sum_t& l, const object_stat_sum_t& r) | |
2735 | { | |
2736 | return | |
2737 | l.num_bytes == r.num_bytes && | |
2738 | l.num_objects == r.num_objects && | |
2739 | l.num_object_clones == r.num_object_clones && | |
2740 | l.num_object_copies == r.num_object_copies && | |
2741 | l.num_objects_missing_on_primary == r.num_objects_missing_on_primary && | |
2742 | l.num_objects_missing == r.num_objects_missing && | |
2743 | l.num_objects_degraded == r.num_objects_degraded && | |
2744 | l.num_objects_misplaced == r.num_objects_misplaced && | |
2745 | l.num_objects_unfound == r.num_objects_unfound && | |
2746 | l.num_rd == r.num_rd && | |
2747 | l.num_rd_kb == r.num_rd_kb && | |
2748 | l.num_wr == r.num_wr && | |
2749 | l.num_wr_kb == r.num_wr_kb && | |
2750 | l.num_scrub_errors == r.num_scrub_errors && | |
2751 | l.num_shallow_scrub_errors == r.num_shallow_scrub_errors && | |
2752 | l.num_deep_scrub_errors == r.num_deep_scrub_errors && | |
2753 | l.num_objects_recovered == r.num_objects_recovered && | |
2754 | l.num_bytes_recovered == r.num_bytes_recovered && | |
2755 | l.num_keys_recovered == r.num_keys_recovered && | |
2756 | l.num_objects_dirty == r.num_objects_dirty && | |
2757 | l.num_whiteouts == r.num_whiteouts && | |
2758 | l.num_objects_omap == r.num_objects_omap && | |
2759 | l.num_objects_hit_set_archive == r.num_objects_hit_set_archive && | |
2760 | l.num_bytes_hit_set_archive == r.num_bytes_hit_set_archive && | |
2761 | l.num_flush == r.num_flush && | |
2762 | l.num_flush_kb == r.num_flush_kb && | |
2763 | l.num_evict == r.num_evict && | |
2764 | l.num_evict_kb == r.num_evict_kb && | |
2765 | l.num_promote == r.num_promote && | |
2766 | l.num_flush_mode_high == r.num_flush_mode_high && | |
2767 | l.num_flush_mode_low == r.num_flush_mode_low && | |
2768 | l.num_evict_mode_some == r.num_evict_mode_some && | |
2769 | l.num_evict_mode_full == r.num_evict_mode_full && | |
2770 | l.num_objects_pinned == r.num_objects_pinned && | |
28e407b8 | 2771 | l.num_legacy_snapsets == r.num_legacy_snapsets && |
11fdf7f2 TL |
2772 | l.num_large_omap_objects == r.num_large_omap_objects && |
2773 | l.num_objects_manifest == r.num_objects_manifest && | |
2774 | l.num_omap_bytes == r.num_omap_bytes && | |
2775 | l.num_omap_keys == r.num_omap_keys && | |
2776 | l.num_objects_repaired == r.num_objects_repaired; | |
7c673cae FG |
2777 | } |
2778 | ||
2779 | // -- object_stat_collection_t -- | |
2780 | ||
2781 | void object_stat_collection_t::dump(Formatter *f) const | |
2782 | { | |
2783 | f->open_object_section("stat_sum"); | |
2784 | sum.dump(f); | |
2785 | f->close_section(); | |
2786 | } | |
2787 | ||
9f95a23c | 2788 | void object_stat_collection_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
2789 | { |
2790 | ENCODE_START(2, 2, bl); | |
11fdf7f2 TL |
2791 | encode(sum, bl); |
2792 | encode((__u32)0, bl); | |
7c673cae FG |
2793 | ENCODE_FINISH(bl); |
2794 | } | |
2795 | ||
9f95a23c | 2796 | void object_stat_collection_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
2797 | { |
2798 | DECODE_START_LEGACY_COMPAT_LEN(2, 2, 2, bl); | |
11fdf7f2 | 2799 | decode(sum, bl); |
7c673cae FG |
2800 | { |
2801 | map<string,object_stat_sum_t> cat_sum; | |
11fdf7f2 | 2802 | decode(cat_sum, bl); |
7c673cae FG |
2803 | } |
2804 | DECODE_FINISH(bl); | |
2805 | } | |
2806 | ||
2807 | void object_stat_collection_t::generate_test_instances(list<object_stat_collection_t*>& o) | |
2808 | { | |
2809 | object_stat_collection_t a; | |
2810 | o.push_back(new object_stat_collection_t(a)); | |
2811 | list<object_stat_sum_t*> l; | |
2812 | object_stat_sum_t::generate_test_instances(l); | |
9f95a23c | 2813 | for (auto p = l.begin(); p != l.end(); ++p) { |
7c673cae FG |
2814 | a.add(**p); |
2815 | o.push_back(new object_stat_collection_t(a)); | |
2816 | } | |
2817 | } | |
2818 | ||
2819 | ||
2820 | // -- pg_stat_t -- | |
2821 | ||
2822 | bool pg_stat_t::is_acting_osd(int32_t osd, bool primary) const | |
2823 | { | |
2824 | if (primary && osd == acting_primary) { | |
2825 | return true; | |
2826 | } else if (!primary) { | |
9f95a23c | 2827 | for(auto it = acting.cbegin(); it != acting.cend(); ++it) |
7c673cae FG |
2828 | { |
2829 | if (*it == osd) | |
2830 | return true; | |
2831 | } | |
2832 | } | |
2833 | return false; | |
2834 | } | |
2835 | ||
2836 | void pg_stat_t::dump(Formatter *f) const | |
2837 | { | |
2838 | f->dump_stream("version") << version; | |
b3b6e05e TL |
2839 | f->dump_unsigned("reported_seq", reported_seq); |
2840 | f->dump_unsigned("reported_epoch", reported_epoch); | |
7c673cae FG |
2841 | f->dump_string("state", pg_state_string(state)); |
2842 | f->dump_stream("last_fresh") << last_fresh; | |
2843 | f->dump_stream("last_change") << last_change; | |
2844 | f->dump_stream("last_active") << last_active; | |
2845 | f->dump_stream("last_peered") << last_peered; | |
2846 | f->dump_stream("last_clean") << last_clean; | |
2847 | f->dump_stream("last_became_active") << last_became_active; | |
2848 | f->dump_stream("last_became_peered") << last_became_peered; | |
2849 | f->dump_stream("last_unstale") << last_unstale; | |
2850 | f->dump_stream("last_undegraded") << last_undegraded; | |
2851 | f->dump_stream("last_fullsized") << last_fullsized; | |
2852 | f->dump_unsigned("mapping_epoch", mapping_epoch); | |
2853 | f->dump_stream("log_start") << log_start; | |
2854 | f->dump_stream("ondisk_log_start") << ondisk_log_start; | |
2855 | f->dump_unsigned("created", created); | |
2856 | f->dump_unsigned("last_epoch_clean", last_epoch_clean); | |
2857 | f->dump_stream("parent") << parent; | |
2858 | f->dump_unsigned("parent_split_bits", parent_split_bits); | |
2859 | f->dump_stream("last_scrub") << last_scrub; | |
2860 | f->dump_stream("last_scrub_stamp") << last_scrub_stamp; | |
2861 | f->dump_stream("last_deep_scrub") << last_deep_scrub; | |
2862 | f->dump_stream("last_deep_scrub_stamp") << last_deep_scrub_stamp; | |
2863 | f->dump_stream("last_clean_scrub_stamp") << last_clean_scrub_stamp; | |
20effc67 | 2864 | f->dump_int("objects_scrubbed", objects_scrubbed); |
7c673cae | 2865 | f->dump_int("log_size", log_size); |
1e59de90 | 2866 | f->dump_int("log_dups_size", log_dups_size); |
7c673cae FG |
2867 | f->dump_int("ondisk_log_size", ondisk_log_size); |
2868 | f->dump_bool("stats_invalid", stats_invalid); | |
2869 | f->dump_bool("dirty_stats_invalid", dirty_stats_invalid); | |
2870 | f->dump_bool("omap_stats_invalid", omap_stats_invalid); | |
2871 | f->dump_bool("hitset_stats_invalid", hitset_stats_invalid); | |
2872 | f->dump_bool("hitset_bytes_stats_invalid", hitset_bytes_stats_invalid); | |
2873 | f->dump_bool("pin_stats_invalid", pin_stats_invalid); | |
11fdf7f2 | 2874 | f->dump_bool("manifest_stats_invalid", manifest_stats_invalid); |
b32b8144 | 2875 | f->dump_unsigned("snaptrimq_len", snaptrimq_len); |
20effc67 TL |
2876 | f->dump_int("last_scrub_duration", last_scrub_duration); |
2877 | f->dump_string("scrub_schedule", dump_scrub_schedule()); | |
1d09f67e TL |
2878 | f->dump_float("scrub_duration", scrub_duration); |
2879 | f->dump_int("objects_trimmed", objects_trimmed); | |
2880 | f->dump_float("snaptrim_duration", snaptrim_duration); | |
7c673cae FG |
2881 | stats.dump(f); |
2882 | f->open_array_section("up"); | |
9f95a23c | 2883 | for (auto p = up.cbegin(); p != up.cend(); ++p) |
7c673cae FG |
2884 | f->dump_int("osd", *p); |
2885 | f->close_section(); | |
2886 | f->open_array_section("acting"); | |
9f95a23c | 2887 | for (auto p = acting.cbegin(); p != acting.cend(); ++p) |
7c673cae FG |
2888 | f->dump_int("osd", *p); |
2889 | f->close_section(); | |
81eedcae TL |
2890 | f->open_array_section("avail_no_missing"); |
2891 | for (auto p = avail_no_missing.cbegin(); p != avail_no_missing.cend(); ++p) | |
2892 | f->dump_stream("shard") << *p; | |
2893 | f->close_section(); | |
2894 | f->open_array_section("object_location_counts"); | |
2895 | for (auto p = object_location_counts.cbegin(); p != object_location_counts.cend(); ++p) { | |
2896 | f->open_object_section("entry"); | |
2897 | f->dump_stream("shards") << p->first; | |
2898 | f->dump_int("objects", p->second); | |
2899 | f->close_section(); | |
2900 | } | |
2901 | f->close_section(); | |
7c673cae | 2902 | f->open_array_section("blocked_by"); |
9f95a23c | 2903 | for (auto p = blocked_by.cbegin(); p != blocked_by.cend(); ++p) |
7c673cae FG |
2904 | f->dump_int("osd", *p); |
2905 | f->close_section(); | |
2906 | f->dump_int("up_primary", up_primary); | |
2907 | f->dump_int("acting_primary", acting_primary); | |
11fdf7f2 | 2908 | f->open_array_section("purged_snaps"); |
9f95a23c | 2909 | for (auto i = purged_snaps.begin(); i != purged_snaps.end(); ++i) { |
11fdf7f2 TL |
2910 | f->open_object_section("interval"); |
2911 | f->dump_stream("start") << i.get_start(); | |
2912 | f->dump_stream("length") << i.get_len(); | |
2913 | f->close_section(); | |
2914 | } | |
2915 | f->close_section(); | |
7c673cae FG |
2916 | } |
2917 | ||
2918 | void pg_stat_t::dump_brief(Formatter *f) const | |
2919 | { | |
2920 | f->dump_string("state", pg_state_string(state)); | |
2921 | f->open_array_section("up"); | |
9f95a23c | 2922 | for (auto p = up.cbegin(); p != up.cend(); ++p) |
7c673cae FG |
2923 | f->dump_int("osd", *p); |
2924 | f->close_section(); | |
2925 | f->open_array_section("acting"); | |
9f95a23c | 2926 | for (auto p = acting.cbegin(); p != acting.cend(); ++p) |
7c673cae FG |
2927 | f->dump_int("osd", *p); |
2928 | f->close_section(); | |
2929 | f->dump_int("up_primary", up_primary); | |
2930 | f->dump_int("acting_primary", acting_primary); | |
2931 | } | |
2932 | ||
20effc67 TL |
2933 | std::string pg_stat_t::dump_scrub_schedule() const |
2934 | { | |
2935 | if (scrub_sched_status.m_is_active) { | |
1e59de90 TL |
2936 | // are we blocked (in fact, stuck) on some locked object? |
2937 | if (scrub_sched_status.m_sched_status == pg_scrub_sched_status_t::blocked) { | |
2938 | return fmt::format( | |
2939 | "Blocked! locked objects (for {}s)", | |
2940 | scrub_sched_status.m_duration_seconds); | |
2941 | } else { | |
2942 | return fmt::format( | |
2943 | "{}scrubbing for {}s", | |
2944 | ((scrub_sched_status.m_is_deep == scrub_level_t::deep) ? "deep " : ""), | |
2945 | scrub_sched_status.m_duration_seconds); | |
2946 | } | |
20effc67 TL |
2947 | } |
2948 | switch (scrub_sched_status.m_sched_status) { | |
2949 | case pg_scrub_sched_status_t::unknown: | |
2950 | // no reported scrub schedule yet | |
2951 | return "--"s; | |
2952 | case pg_scrub_sched_status_t::not_queued: | |
2953 | return "no scrub is scheduled"s; | |
2954 | case pg_scrub_sched_status_t::scheduled: | |
2955 | return fmt::format( | |
2956 | "{} {}scrub scheduled @ {}", | |
2957 | (scrub_sched_status.m_is_periodic ? "periodic" : "user requested"), | |
2958 | ((scrub_sched_status.m_is_deep == scrub_level_t::deep) ? "deep " : ""), | |
2959 | scrub_sched_status.m_scheduled_at); | |
2960 | case pg_scrub_sched_status_t::queued: | |
2961 | return fmt::format( | |
2962 | "queued for {}scrub", | |
2963 | ((scrub_sched_status.m_is_deep == scrub_level_t::deep) ? "deep " : "")); | |
2964 | default: | |
2965 | // a bug! | |
2966 | return "SCRUB STATE MISMATCH!"s; | |
2967 | } | |
2968 | } | |
2969 | ||
2970 | bool operator==(const pg_scrubbing_status_t& l, const pg_scrubbing_status_t& r) | |
2971 | { | |
2972 | return | |
2973 | l.m_sched_status == r.m_sched_status && | |
2974 | l.m_scheduled_at == r.m_scheduled_at && | |
2975 | l.m_duration_seconds == r.m_duration_seconds && | |
2976 | l.m_is_active == r.m_is_active && | |
2977 | l.m_is_deep == r.m_is_deep && | |
2978 | l.m_is_periodic == r.m_is_periodic; | |
2979 | } | |
2980 | ||
9f95a23c | 2981 | void pg_stat_t::encode(ceph::buffer::list &bl) const |
7c673cae | 2982 | { |
1e59de90 | 2983 | ENCODE_START(29, 22, bl); |
11fdf7f2 TL |
2984 | encode(version, bl); |
2985 | encode(reported_seq, bl); | |
2986 | encode(reported_epoch, bl); | |
2987 | encode((__u32)state, bl); // for older peers | |
2988 | encode(log_start, bl); | |
2989 | encode(ondisk_log_start, bl); | |
2990 | encode(created, bl); | |
2991 | encode(last_epoch_clean, bl); | |
2992 | encode(parent, bl); | |
2993 | encode(parent_split_bits, bl); | |
2994 | encode(last_scrub, bl); | |
2995 | encode(last_scrub_stamp, bl); | |
2996 | encode(stats, bl); | |
2997 | encode(log_size, bl); | |
2998 | encode(ondisk_log_size, bl); | |
2999 | encode(up, bl); | |
3000 | encode(acting, bl); | |
3001 | encode(last_fresh, bl); | |
3002 | encode(last_change, bl); | |
3003 | encode(last_active, bl); | |
3004 | encode(last_clean, bl); | |
3005 | encode(last_unstale, bl); | |
3006 | encode(mapping_epoch, bl); | |
3007 | encode(last_deep_scrub, bl); | |
3008 | encode(last_deep_scrub_stamp, bl); | |
3009 | encode(stats_invalid, bl); | |
3010 | encode(last_clean_scrub_stamp, bl); | |
3011 | encode(last_became_active, bl); | |
3012 | encode(dirty_stats_invalid, bl); | |
3013 | encode(up_primary, bl); | |
3014 | encode(acting_primary, bl); | |
3015 | encode(omap_stats_invalid, bl); | |
3016 | encode(hitset_stats_invalid, bl); | |
3017 | encode(blocked_by, bl); | |
3018 | encode(last_undegraded, bl); | |
3019 | encode(last_fullsized, bl); | |
3020 | encode(hitset_bytes_stats_invalid, bl); | |
3021 | encode(last_peered, bl); | |
3022 | encode(last_became_peered, bl); | |
3023 | encode(pin_stats_invalid, bl); | |
3024 | encode(snaptrimq_len, bl); | |
3025 | __u32 top_state = (state >> 32); | |
3026 | encode(top_state, bl); | |
3027 | encode(purged_snaps, bl); | |
3028 | encode(manifest_stats_invalid, bl); | |
81eedcae TL |
3029 | encode(avail_no_missing, bl); |
3030 | encode(object_location_counts, bl); | |
20effc67 TL |
3031 | encode(last_scrub_duration, bl); |
3032 | encode(scrub_sched_status.m_scheduled_at, bl); | |
3033 | encode(scrub_sched_status.m_duration_seconds, bl); | |
3034 | encode((__u16)scrub_sched_status.m_sched_status, bl); | |
3035 | encode(scrub_sched_status.m_is_active, bl); | |
3036 | encode((scrub_sched_status.m_is_deep==scrub_level_t::deep), bl); | |
3037 | encode(scrub_sched_status.m_is_periodic, bl); | |
3038 | encode(objects_scrubbed, bl); | |
1d09f67e TL |
3039 | encode(scrub_duration, bl); |
3040 | encode(objects_trimmed, bl); | |
3041 | encode(snaptrim_duration, bl); | |
1e59de90 | 3042 | encode(log_dups_size, bl); |
20effc67 | 3043 | |
7c673cae FG |
3044 | ENCODE_FINISH(bl); |
3045 | } | |
3046 | ||
9f95a23c | 3047 | void pg_stat_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
3048 | { |
3049 | bool tmp; | |
11fdf7f2 | 3050 | uint32_t old_state; |
1e59de90 | 3051 | DECODE_START(29, bl); |
11fdf7f2 TL |
3052 | decode(version, bl); |
3053 | decode(reported_seq, bl); | |
3054 | decode(reported_epoch, bl); | |
3055 | decode(old_state, bl); | |
3056 | decode(log_start, bl); | |
3057 | decode(ondisk_log_start, bl); | |
3058 | decode(created, bl); | |
3059 | decode(last_epoch_clean, bl); | |
3060 | decode(parent, bl); | |
3061 | decode(parent_split_bits, bl); | |
3062 | decode(last_scrub, bl); | |
3063 | decode(last_scrub_stamp, bl); | |
3064 | decode(stats, bl); | |
3065 | decode(log_size, bl); | |
3066 | decode(ondisk_log_size, bl); | |
3067 | decode(up, bl); | |
3068 | decode(acting, bl); | |
3069 | decode(last_fresh, bl); | |
3070 | decode(last_change, bl); | |
3071 | decode(last_active, bl); | |
3072 | decode(last_clean, bl); | |
3073 | decode(last_unstale, bl); | |
3074 | decode(mapping_epoch, bl); | |
3075 | decode(last_deep_scrub, bl); | |
3076 | decode(last_deep_scrub_stamp, bl); | |
3077 | decode(tmp, bl); | |
7c673cae | 3078 | stats_invalid = tmp; |
11fdf7f2 TL |
3079 | decode(last_clean_scrub_stamp, bl); |
3080 | decode(last_became_active, bl); | |
3081 | decode(tmp, bl); | |
7c673cae | 3082 | dirty_stats_invalid = tmp; |
11fdf7f2 TL |
3083 | decode(up_primary, bl); |
3084 | decode(acting_primary, bl); | |
3085 | decode(tmp, bl); | |
7c673cae | 3086 | omap_stats_invalid = tmp; |
11fdf7f2 | 3087 | decode(tmp, bl); |
7c673cae | 3088 | hitset_stats_invalid = tmp; |
11fdf7f2 TL |
3089 | decode(blocked_by, bl); |
3090 | decode(last_undegraded, bl); | |
3091 | decode(last_fullsized, bl); | |
3092 | decode(tmp, bl); | |
7c673cae | 3093 | hitset_bytes_stats_invalid = tmp; |
11fdf7f2 TL |
3094 | decode(last_peered, bl); |
3095 | decode(last_became_peered, bl); | |
3096 | decode(tmp, bl); | |
7c673cae | 3097 | pin_stats_invalid = tmp; |
b32b8144 | 3098 | if (struct_v >= 23) { |
11fdf7f2 TL |
3099 | decode(snaptrimq_len, bl); |
3100 | if (struct_v >= 24) { | |
3101 | __u32 top_state; | |
3102 | decode(top_state, bl); | |
3103 | state = (uint64_t)old_state | ((uint64_t)top_state << 32); | |
3104 | decode(purged_snaps, bl); | |
3105 | } else { | |
3106 | state = old_state; | |
3107 | } | |
3108 | if (struct_v >= 25) { | |
3109 | decode(tmp, bl); | |
3110 | manifest_stats_invalid = tmp; | |
3111 | } else { | |
3112 | manifest_stats_invalid = true; | |
3113 | } | |
81eedcae TL |
3114 | if (struct_v >= 26) { |
3115 | decode(avail_no_missing, bl); | |
3116 | decode(object_location_counts, bl); | |
3117 | } | |
20effc67 TL |
3118 | if (struct_v >= 27) { |
3119 | decode(last_scrub_duration, bl); | |
3120 | decode(scrub_sched_status.m_scheduled_at, bl); | |
3121 | decode(scrub_sched_status.m_duration_seconds, bl); | |
3122 | __u16 scrub_sched_as_u16; | |
3123 | decode(scrub_sched_as_u16, bl); | |
3124 | scrub_sched_status.m_sched_status = (pg_scrub_sched_status_t)(scrub_sched_as_u16); | |
3125 | decode(tmp, bl); | |
3126 | scrub_sched_status.m_is_active = tmp; | |
3127 | decode(tmp, bl); | |
3128 | scrub_sched_status.m_is_deep = tmp ? scrub_level_t::deep : scrub_level_t::shallow; | |
3129 | decode(tmp, bl); | |
3130 | scrub_sched_status.m_is_periodic = tmp; | |
3131 | decode(objects_scrubbed, bl); | |
3132 | } | |
1d09f67e TL |
3133 | if (struct_v >= 28) { |
3134 | decode(scrub_duration, bl); | |
3135 | decode(objects_trimmed, bl); | |
3136 | decode(snaptrim_duration, bl); | |
3137 | } | |
1e59de90 TL |
3138 | if (struct_v >= 29) { |
3139 | decode(log_dups_size, bl); | |
3140 | } | |
b32b8144 | 3141 | } |
7c673cae FG |
3142 | DECODE_FINISH(bl); |
3143 | } | |
3144 | ||
3145 | void pg_stat_t::generate_test_instances(list<pg_stat_t*>& o) | |
3146 | { | |
3147 | pg_stat_t a; | |
3148 | o.push_back(new pg_stat_t(a)); | |
3149 | ||
3150 | a.version = eversion_t(1, 3); | |
3151 | a.reported_epoch = 1; | |
3152 | a.reported_seq = 2; | |
3153 | a.state = 123; | |
3154 | a.mapping_epoch = 998; | |
3155 | a.last_fresh = utime_t(1002, 1); | |
3156 | a.last_change = utime_t(1002, 2); | |
3157 | a.last_active = utime_t(1002, 3); | |
3158 | a.last_clean = utime_t(1002, 4); | |
3159 | a.last_unstale = utime_t(1002, 5); | |
3160 | a.last_undegraded = utime_t(1002, 7); | |
3161 | a.last_fullsized = utime_t(1002, 8); | |
3162 | a.log_start = eversion_t(1, 4); | |
3163 | a.ondisk_log_start = eversion_t(1, 5); | |
3164 | a.created = 6; | |
3165 | a.last_epoch_clean = 7; | |
11fdf7f2 | 3166 | a.parent = pg_t(1, 2); |
7c673cae FG |
3167 | a.parent_split_bits = 12; |
3168 | a.last_scrub = eversion_t(9, 10); | |
3169 | a.last_scrub_stamp = utime_t(11, 12); | |
3170 | a.last_deep_scrub = eversion_t(13, 14); | |
3171 | a.last_deep_scrub_stamp = utime_t(15, 16); | |
3172 | a.last_clean_scrub_stamp = utime_t(17, 18); | |
20effc67 | 3173 | a.last_scrub_duration = 3617; |
1d09f67e | 3174 | a.scrub_duration = 0.003; |
b32b8144 | 3175 | a.snaptrimq_len = 1048576; |
20effc67 | 3176 | a.objects_scrubbed = 0; |
1d09f67e TL |
3177 | a.objects_trimmed = 0; |
3178 | a.snaptrim_duration = 0.123; | |
7c673cae FG |
3179 | list<object_stat_collection_t*> l; |
3180 | object_stat_collection_t::generate_test_instances(l); | |
3181 | a.stats = *l.back(); | |
3182 | a.log_size = 99; | |
3183 | a.ondisk_log_size = 88; | |
3184 | a.up.push_back(123); | |
3185 | a.up_primary = 123; | |
3186 | a.acting.push_back(456); | |
81eedcae TL |
3187 | a.avail_no_missing.push_back(pg_shard_t(456, shard_id_t::NO_SHARD)); |
3188 | set<pg_shard_t> sset = { pg_shard_t(0), pg_shard_t(1) }; | |
3189 | a.object_location_counts.insert(make_pair(sset, 10)); | |
3190 | sset.insert(pg_shard_t(2)); | |
3191 | a.object_location_counts.insert(make_pair(sset, 5)); | |
7c673cae FG |
3192 | a.acting_primary = 456; |
3193 | o.push_back(new pg_stat_t(a)); | |
3194 | ||
3195 | a.up.push_back(124); | |
3196 | a.up_primary = 124; | |
3197 | a.acting.push_back(124); | |
3198 | a.acting_primary = 124; | |
3199 | a.blocked_by.push_back(155); | |
3200 | a.blocked_by.push_back(156); | |
3201 | o.push_back(new pg_stat_t(a)); | |
3202 | } | |
3203 | ||
3204 | bool operator==(const pg_stat_t& l, const pg_stat_t& r) | |
3205 | { | |
3206 | return | |
3207 | l.version == r.version && | |
3208 | l.reported_seq == r.reported_seq && | |
3209 | l.reported_epoch == r.reported_epoch && | |
3210 | l.state == r.state && | |
3211 | l.last_fresh == r.last_fresh && | |
3212 | l.last_change == r.last_change && | |
3213 | l.last_active == r.last_active && | |
3214 | l.last_peered == r.last_peered && | |
3215 | l.last_clean == r.last_clean && | |
3216 | l.last_unstale == r.last_unstale && | |
3217 | l.last_undegraded == r.last_undegraded && | |
3218 | l.last_fullsized == r.last_fullsized && | |
3219 | l.log_start == r.log_start && | |
3220 | l.ondisk_log_start == r.ondisk_log_start && | |
3221 | l.created == r.created && | |
3222 | l.last_epoch_clean == r.last_epoch_clean && | |
3223 | l.parent == r.parent && | |
3224 | l.parent_split_bits == r.parent_split_bits && | |
3225 | l.last_scrub == r.last_scrub && | |
3226 | l.last_deep_scrub == r.last_deep_scrub && | |
3227 | l.last_scrub_stamp == r.last_scrub_stamp && | |
3228 | l.last_deep_scrub_stamp == r.last_deep_scrub_stamp && | |
3229 | l.last_clean_scrub_stamp == r.last_clean_scrub_stamp && | |
3230 | l.stats == r.stats && | |
3231 | l.stats_invalid == r.stats_invalid && | |
3232 | l.log_size == r.log_size && | |
1e59de90 | 3233 | l.log_dups_size == r.log_dups_size && |
7c673cae FG |
3234 | l.ondisk_log_size == r.ondisk_log_size && |
3235 | l.up == r.up && | |
3236 | l.acting == r.acting && | |
81eedcae TL |
3237 | l.avail_no_missing == r.avail_no_missing && |
3238 | l.object_location_counts == r.object_location_counts && | |
7c673cae FG |
3239 | l.mapping_epoch == r.mapping_epoch && |
3240 | l.blocked_by == r.blocked_by && | |
3241 | l.last_became_active == r.last_became_active && | |
3242 | l.last_became_peered == r.last_became_peered && | |
3243 | l.dirty_stats_invalid == r.dirty_stats_invalid && | |
3244 | l.omap_stats_invalid == r.omap_stats_invalid && | |
3245 | l.hitset_stats_invalid == r.hitset_stats_invalid && | |
3246 | l.hitset_bytes_stats_invalid == r.hitset_bytes_stats_invalid && | |
3247 | l.up_primary == r.up_primary && | |
3248 | l.acting_primary == r.acting_primary && | |
b32b8144 | 3249 | l.pin_stats_invalid == r.pin_stats_invalid && |
11fdf7f2 TL |
3250 | l.manifest_stats_invalid == r.manifest_stats_invalid && |
3251 | l.purged_snaps == r.purged_snaps && | |
20effc67 TL |
3252 | l.snaptrimq_len == r.snaptrimq_len && |
3253 | l.last_scrub_duration == r.last_scrub_duration && | |
3254 | l.scrub_sched_status == r.scrub_sched_status && | |
1d09f67e TL |
3255 | l.objects_scrubbed == r.objects_scrubbed && |
3256 | l.scrub_duration == r.scrub_duration && | |
3257 | l.objects_trimmed == r.objects_trimmed && | |
3258 | l.snaptrim_duration == r.snaptrim_duration; | |
7c673cae FG |
3259 | } |
3260 | ||
11fdf7f2 TL |
3261 | // -- store_statfs_t -- |
3262 | ||
3263 | bool store_statfs_t::operator==(const store_statfs_t& other) const | |
3264 | { | |
3265 | return total == other.total | |
3266 | && available == other.available | |
3267 | && allocated == other.allocated | |
3268 | && internally_reserved == other.internally_reserved | |
3269 | && data_stored == other.data_stored | |
3270 | && data_compressed == other.data_compressed | |
3271 | && data_compressed_allocated == other.data_compressed_allocated | |
3272 | && data_compressed_original == other.data_compressed_original | |
3273 | && omap_allocated == other.omap_allocated | |
3274 | && internal_metadata == other.internal_metadata; | |
3275 | } | |
3276 | ||
3277 | void store_statfs_t::dump(Formatter *f) const | |
3278 | { | |
3279 | f->dump_int("total", total); | |
3280 | f->dump_int("available", available); | |
3281 | f->dump_int("internally_reserved", internally_reserved); | |
3282 | f->dump_int("allocated", allocated); | |
3283 | f->dump_int("data_stored", data_stored); | |
3284 | f->dump_int("data_compressed", data_compressed); | |
3285 | f->dump_int("data_compressed_allocated", data_compressed_allocated); | |
3286 | f->dump_int("data_compressed_original", data_compressed_original); | |
3287 | f->dump_int("omap_allocated", omap_allocated); | |
3288 | f->dump_int("internal_metadata", internal_metadata); | |
3289 | } | |
3290 | ||
3291 | ostream& operator<<(ostream& out, const store_statfs_t &s) | |
3292 | { | |
3293 | out << std::hex | |
3294 | << "store_statfs(0x" << s.available | |
3295 | << "/0x" << s.internally_reserved | |
3296 | << "/0x" << s.total | |
3297 | << ", data 0x" << s.data_stored | |
3298 | << "/0x" << s.allocated | |
3299 | << ", compress 0x" << s.data_compressed | |
3300 | << "/0x" << s.data_compressed_allocated | |
3301 | << "/0x" << s.data_compressed_original | |
3302 | << ", omap 0x" << s.omap_allocated | |
3303 | << ", meta 0x" << s.internal_metadata | |
3304 | << std::dec | |
3305 | << ")"; | |
3306 | return out; | |
3307 | } | |
3308 | ||
3309 | void store_statfs_t::generate_test_instances(list<store_statfs_t*>& o) | |
3310 | { | |
3311 | store_statfs_t a; | |
3312 | o.push_back(new store_statfs_t(a)); | |
3313 | a.total = 234; | |
3314 | a.available = 123; | |
3315 | a.internally_reserved = 33; | |
3316 | a.allocated = 32; | |
3317 | a.data_stored = 44; | |
3318 | a.data_compressed = 21; | |
3319 | a.data_compressed_allocated = 12; | |
3320 | a.data_compressed_original = 13; | |
3321 | a.omap_allocated = 14; | |
3322 | a.internal_metadata = 15; | |
3323 | o.push_back(new store_statfs_t(a)); | |
3324 | } | |
3325 | ||
7c673cae FG |
3326 | // -- pool_stat_t -- |
3327 | ||
3328 | void pool_stat_t::dump(Formatter *f) const | |
3329 | { | |
3330 | stats.dump(f); | |
11fdf7f2 TL |
3331 | f->open_object_section("store_stats"); |
3332 | store_stats.dump(f); | |
3333 | f->close_section(); | |
7c673cae FG |
3334 | f->dump_int("log_size", log_size); |
3335 | f->dump_int("ondisk_log_size", ondisk_log_size); | |
3336 | f->dump_int("up", up); | |
3337 | f->dump_int("acting", acting); | |
eafe8130 | 3338 | f->dump_int("num_store_stats", num_store_stats); |
7c673cae FG |
3339 | } |
3340 | ||
9f95a23c | 3341 | void pool_stat_t::encode(ceph::buffer::list &bl, uint64_t features) const |
7c673cae | 3342 | { |
11fdf7f2 | 3343 | using ceph::encode; |
7c673cae FG |
3344 | if ((features & CEPH_FEATURE_OSDENC) == 0) { |
3345 | __u8 v = 4; | |
11fdf7f2 TL |
3346 | encode(v, bl); |
3347 | encode(stats, bl); | |
3348 | encode(log_size, bl); | |
3349 | encode(ondisk_log_size, bl); | |
7c673cae FG |
3350 | return; |
3351 | } | |
3352 | ||
11fdf7f2 TL |
3353 | ENCODE_START(7, 5, bl); |
3354 | encode(stats, bl); | |
3355 | encode(log_size, bl); | |
3356 | encode(ondisk_log_size, bl); | |
3357 | encode(up, bl); | |
3358 | encode(acting, bl); | |
3359 | encode(store_stats, bl); | |
3360 | encode(num_store_stats, bl); | |
7c673cae FG |
3361 | ENCODE_FINISH(bl); |
3362 | } | |
3363 | ||
9f95a23c | 3364 | void pool_stat_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 3365 | { |
11fdf7f2 | 3366 | DECODE_START_LEGACY_COMPAT_LEN(7, 5, 5, bl); |
7c673cae | 3367 | if (struct_v >= 4) { |
11fdf7f2 TL |
3368 | decode(stats, bl); |
3369 | decode(log_size, bl); | |
3370 | decode(ondisk_log_size, bl); | |
7c673cae | 3371 | if (struct_v >= 6) { |
11fdf7f2 TL |
3372 | decode(up, bl); |
3373 | decode(acting, bl); | |
7c673cae FG |
3374 | } else { |
3375 | up = 0; | |
3376 | acting = 0; | |
3377 | } | |
11fdf7f2 TL |
3378 | if (struct_v >= 7) { |
3379 | decode(store_stats, bl); | |
3380 | decode(num_store_stats, bl); | |
3381 | } else { | |
3382 | store_stats.reset(); | |
3383 | num_store_stats = 0; | |
3384 | } | |
3385 | ||
7c673cae | 3386 | } else { |
11fdf7f2 | 3387 | decode(stats.sum.num_bytes, bl); |
7c673cae | 3388 | uint64_t num_kb; |
11fdf7f2 TL |
3389 | decode(num_kb, bl); |
3390 | decode(stats.sum.num_objects, bl); | |
3391 | decode(stats.sum.num_object_clones, bl); | |
3392 | decode(stats.sum.num_object_copies, bl); | |
3393 | decode(stats.sum.num_objects_missing_on_primary, bl); | |
3394 | decode(stats.sum.num_objects_degraded, bl); | |
3395 | decode(log_size, bl); | |
3396 | decode(ondisk_log_size, bl); | |
7c673cae | 3397 | if (struct_v >= 2) { |
11fdf7f2 TL |
3398 | decode(stats.sum.num_rd, bl); |
3399 | decode(stats.sum.num_rd_kb, bl); | |
3400 | decode(stats.sum.num_wr, bl); | |
3401 | decode(stats.sum.num_wr_kb, bl); | |
7c673cae FG |
3402 | } |
3403 | if (struct_v >= 3) { | |
11fdf7f2 | 3404 | decode(stats.sum.num_objects_unfound, bl); |
7c673cae FG |
3405 | } |
3406 | } | |
3407 | DECODE_FINISH(bl); | |
3408 | } | |
3409 | ||
3410 | void pool_stat_t::generate_test_instances(list<pool_stat_t*>& o) | |
3411 | { | |
3412 | pool_stat_t a; | |
3413 | o.push_back(new pool_stat_t(a)); | |
3414 | ||
3415 | list<object_stat_collection_t*> l; | |
3416 | object_stat_collection_t::generate_test_instances(l); | |
11fdf7f2 TL |
3417 | list<store_statfs_t*> ll; |
3418 | store_statfs_t::generate_test_instances(ll); | |
7c673cae | 3419 | a.stats = *l.back(); |
11fdf7f2 | 3420 | a.store_stats = *ll.back(); |
7c673cae FG |
3421 | a.log_size = 123; |
3422 | a.ondisk_log_size = 456; | |
3423 | a.acting = 3; | |
3424 | a.up = 4; | |
11fdf7f2 | 3425 | a.num_store_stats = 1; |
7c673cae FG |
3426 | o.push_back(new pool_stat_t(a)); |
3427 | } | |
3428 | ||
3429 | ||
3430 | // -- pg_history_t -- | |
3431 | ||
9f95a23c | 3432 | void pg_history_t::encode(ceph::buffer::list &bl) const |
7c673cae | 3433 | { |
9f95a23c | 3434 | ENCODE_START(10, 4, bl); |
11fdf7f2 TL |
3435 | encode(epoch_created, bl); |
3436 | encode(last_epoch_started, bl); | |
3437 | encode(last_epoch_clean, bl); | |
3438 | encode(last_epoch_split, bl); | |
3439 | encode(same_interval_since, bl); | |
3440 | encode(same_up_since, bl); | |
3441 | encode(same_primary_since, bl); | |
3442 | encode(last_scrub, bl); | |
3443 | encode(last_scrub_stamp, bl); | |
3444 | encode(last_deep_scrub, bl); | |
3445 | encode(last_deep_scrub_stamp, bl); | |
3446 | encode(last_clean_scrub_stamp, bl); | |
3447 | encode(last_epoch_marked_full, bl); | |
3448 | encode(last_interval_started, bl); | |
3449 | encode(last_interval_clean, bl); | |
3450 | encode(epoch_pool_created, bl); | |
9f95a23c | 3451 | encode(prior_readable_until_ub, bl); |
7c673cae FG |
3452 | ENCODE_FINISH(bl); |
3453 | } | |
3454 | ||
9f95a23c | 3455 | void pg_history_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 3456 | { |
9f95a23c | 3457 | DECODE_START_LEGACY_COMPAT_LEN(10, 4, 4, bl); |
11fdf7f2 TL |
3458 | decode(epoch_created, bl); |
3459 | decode(last_epoch_started, bl); | |
7c673cae | 3460 | if (struct_v >= 3) |
11fdf7f2 | 3461 | decode(last_epoch_clean, bl); |
7c673cae FG |
3462 | else |
3463 | last_epoch_clean = last_epoch_started; // careful, it's a lie! | |
11fdf7f2 TL |
3464 | decode(last_epoch_split, bl); |
3465 | decode(same_interval_since, bl); | |
3466 | decode(same_up_since, bl); | |
3467 | decode(same_primary_since, bl); | |
7c673cae | 3468 | if (struct_v >= 2) { |
11fdf7f2 TL |
3469 | decode(last_scrub, bl); |
3470 | decode(last_scrub_stamp, bl); | |
7c673cae FG |
3471 | } |
3472 | if (struct_v >= 5) { | |
11fdf7f2 TL |
3473 | decode(last_deep_scrub, bl); |
3474 | decode(last_deep_scrub_stamp, bl); | |
7c673cae FG |
3475 | } |
3476 | if (struct_v >= 6) { | |
11fdf7f2 | 3477 | decode(last_clean_scrub_stamp, bl); |
7c673cae FG |
3478 | } |
3479 | if (struct_v >= 7) { | |
11fdf7f2 | 3480 | decode(last_epoch_marked_full, bl); |
7c673cae FG |
3481 | } |
3482 | if (struct_v >= 8) { | |
11fdf7f2 TL |
3483 | decode(last_interval_started, bl); |
3484 | decode(last_interval_clean, bl); | |
7c673cae FG |
3485 | } else { |
3486 | if (last_epoch_started >= same_interval_since) { | |
3487 | last_interval_started = same_interval_since; | |
3488 | } else { | |
3489 | last_interval_started = last_epoch_started; // best guess | |
3490 | } | |
3491 | if (last_epoch_clean >= same_interval_since) { | |
3492 | last_interval_clean = same_interval_since; | |
3493 | } else { | |
3494 | last_interval_clean = last_epoch_clean; // best guess | |
3495 | } | |
3496 | } | |
31f18b77 | 3497 | if (struct_v >= 9) { |
11fdf7f2 | 3498 | decode(epoch_pool_created, bl); |
31f18b77 FG |
3499 | } else { |
3500 | epoch_pool_created = epoch_created; | |
3501 | } | |
9f95a23c TL |
3502 | if (struct_v >= 10) { |
3503 | decode(prior_readable_until_ub, bl); | |
3504 | } | |
7c673cae FG |
3505 | DECODE_FINISH(bl); |
3506 | } | |
3507 | ||
3508 | void pg_history_t::dump(Formatter *f) const | |
3509 | { | |
3510 | f->dump_int("epoch_created", epoch_created); | |
31f18b77 | 3511 | f->dump_int("epoch_pool_created", epoch_pool_created); |
7c673cae FG |
3512 | f->dump_int("last_epoch_started", last_epoch_started); |
3513 | f->dump_int("last_interval_started", last_interval_started); | |
3514 | f->dump_int("last_epoch_clean", last_epoch_clean); | |
3515 | f->dump_int("last_interval_clean", last_interval_clean); | |
3516 | f->dump_int("last_epoch_split", last_epoch_split); | |
3517 | f->dump_int("last_epoch_marked_full", last_epoch_marked_full); | |
3518 | f->dump_int("same_up_since", same_up_since); | |
3519 | f->dump_int("same_interval_since", same_interval_since); | |
3520 | f->dump_int("same_primary_since", same_primary_since); | |
3521 | f->dump_stream("last_scrub") << last_scrub; | |
3522 | f->dump_stream("last_scrub_stamp") << last_scrub_stamp; | |
3523 | f->dump_stream("last_deep_scrub") << last_deep_scrub; | |
3524 | f->dump_stream("last_deep_scrub_stamp") << last_deep_scrub_stamp; | |
3525 | f->dump_stream("last_clean_scrub_stamp") << last_clean_scrub_stamp; | |
9f95a23c TL |
3526 | f->dump_float( |
3527 | "prior_readable_until_ub", | |
3528 | std::chrono::duration<double>(prior_readable_until_ub).count()); | |
7c673cae FG |
3529 | } |
3530 | ||
3531 | void pg_history_t::generate_test_instances(list<pg_history_t*>& o) | |
3532 | { | |
3533 | o.push_back(new pg_history_t); | |
3534 | o.push_back(new pg_history_t); | |
3535 | o.back()->epoch_created = 1; | |
31f18b77 | 3536 | o.back()->epoch_pool_created = 1; |
7c673cae FG |
3537 | o.back()->last_epoch_started = 2; |
3538 | o.back()->last_interval_started = 2; | |
3539 | o.back()->last_epoch_clean = 3; | |
3540 | o.back()->last_interval_clean = 2; | |
3541 | o.back()->last_epoch_split = 4; | |
9f95a23c | 3542 | o.back()->prior_readable_until_ub = make_timespan(3.1415); |
7c673cae FG |
3543 | o.back()->same_up_since = 5; |
3544 | o.back()->same_interval_since = 6; | |
3545 | o.back()->same_primary_since = 7; | |
3546 | o.back()->last_scrub = eversion_t(8, 9); | |
3547 | o.back()->last_scrub_stamp = utime_t(10, 11); | |
3548 | o.back()->last_deep_scrub = eversion_t(12, 13); | |
3549 | o.back()->last_deep_scrub_stamp = utime_t(14, 15); | |
3550 | o.back()->last_clean_scrub_stamp = utime_t(16, 17); | |
3551 | o.back()->last_epoch_marked_full = 18; | |
3552 | } | |
3553 | ||
3554 | ||
3555 | // -- pg_info_t -- | |
3556 | ||
9f95a23c | 3557 | void pg_info_t::encode(ceph::buffer::list &bl) const |
7c673cae FG |
3558 | { |
3559 | ENCODE_START(32, 26, bl); | |
11fdf7f2 TL |
3560 | encode(pgid.pgid, bl); |
3561 | encode(last_update, bl); | |
3562 | encode(last_complete, bl); | |
3563 | encode(log_tail, bl); | |
9f95a23c | 3564 | encode(hobject_t(), bl); // old (nibblewise) last_backfill |
11fdf7f2 | 3565 | encode(stats, bl); |
7c673cae | 3566 | history.encode(bl); |
11fdf7f2 TL |
3567 | encode(purged_snaps, bl); |
3568 | encode(last_epoch_started, bl); | |
3569 | encode(last_user_version, bl); | |
3570 | encode(hit_set, bl); | |
3571 | encode(pgid.shard, bl); | |
3572 | encode(last_backfill, bl); | |
9f95a23c | 3573 | encode(true, bl); // was last_backfill_bitwise |
11fdf7f2 | 3574 | encode(last_interval_started, bl); |
7c673cae FG |
3575 | ENCODE_FINISH(bl); |
3576 | } | |
3577 | ||
9f95a23c | 3578 | void pg_info_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
3579 | { |
3580 | DECODE_START(32, bl); | |
11fdf7f2 TL |
3581 | decode(pgid.pgid, bl); |
3582 | decode(last_update, bl); | |
3583 | decode(last_complete, bl); | |
3584 | decode(log_tail, bl); | |
7c673cae FG |
3585 | { |
3586 | hobject_t old_last_backfill; | |
11fdf7f2 | 3587 | decode(old_last_backfill, bl); |
7c673cae | 3588 | } |
11fdf7f2 | 3589 | decode(stats, bl); |
7c673cae | 3590 | history.decode(bl); |
11fdf7f2 TL |
3591 | decode(purged_snaps, bl); |
3592 | decode(last_epoch_started, bl); | |
3593 | decode(last_user_version, bl); | |
3594 | decode(hit_set, bl); | |
3595 | decode(pgid.shard, bl); | |
3596 | decode(last_backfill, bl); | |
9f95a23c TL |
3597 | { |
3598 | bool last_backfill_bitwise; | |
3599 | decode(last_backfill_bitwise, bl); | |
3600 | // note: we may see a false value here since the default value for | |
3601 | // the member was false, so it often didn't get set to true until | |
3602 | // peering progressed. | |
3603 | } | |
7c673cae | 3604 | if (struct_v >= 32) { |
11fdf7f2 | 3605 | decode(last_interval_started, bl); |
7c673cae FG |
3606 | } else { |
3607 | last_interval_started = last_epoch_started; | |
3608 | } | |
3609 | DECODE_FINISH(bl); | |
3610 | } | |
3611 | ||
3612 | // -- pg_info_t -- | |
3613 | ||
3614 | void pg_info_t::dump(Formatter *f) const | |
3615 | { | |
3616 | f->dump_stream("pgid") << pgid; | |
3617 | f->dump_stream("last_update") << last_update; | |
3618 | f->dump_stream("last_complete") << last_complete; | |
3619 | f->dump_stream("log_tail") << log_tail; | |
3620 | f->dump_int("last_user_version", last_user_version); | |
3621 | f->dump_stream("last_backfill") << last_backfill; | |
7c673cae FG |
3622 | f->open_array_section("purged_snaps"); |
3623 | for (interval_set<snapid_t>::const_iterator i=purged_snaps.begin(); | |
3624 | i != purged_snaps.end(); | |
3625 | ++i) { | |
3626 | f->open_object_section("purged_snap_interval"); | |
3627 | f->dump_stream("start") << i.get_start(); | |
3628 | f->dump_stream("length") << i.get_len(); | |
3629 | f->close_section(); | |
3630 | } | |
3631 | f->close_section(); | |
3632 | f->open_object_section("history"); | |
3633 | history.dump(f); | |
3634 | f->close_section(); | |
3635 | f->open_object_section("stats"); | |
3636 | stats.dump(f); | |
3637 | f->close_section(); | |
3638 | ||
3639 | f->dump_int("empty", is_empty()); | |
3640 | f->dump_int("dne", dne()); | |
3641 | f->dump_int("incomplete", is_incomplete()); | |
3642 | f->dump_int("last_epoch_started", last_epoch_started); | |
3643 | ||
3644 | f->open_object_section("hit_set_history"); | |
3645 | hit_set.dump(f); | |
3646 | f->close_section(); | |
3647 | } | |
3648 | ||
3649 | void pg_info_t::generate_test_instances(list<pg_info_t*>& o) | |
3650 | { | |
3651 | o.push_back(new pg_info_t); | |
3652 | o.push_back(new pg_info_t); | |
3653 | list<pg_history_t*> h; | |
3654 | pg_history_t::generate_test_instances(h); | |
3655 | o.back()->history = *h.back(); | |
11fdf7f2 | 3656 | o.back()->pgid = spg_t(pg_t(1, 2), shard_id_t::NO_SHARD); |
7c673cae FG |
3657 | o.back()->last_update = eversion_t(3, 4); |
3658 | o.back()->last_complete = eversion_t(5, 6); | |
3659 | o.back()->last_user_version = 2; | |
3660 | o.back()->log_tail = eversion_t(7, 8); | |
3661 | o.back()->last_backfill = hobject_t(object_t("objname"), "key", 123, 456, -1, ""); | |
7c673cae FG |
3662 | { |
3663 | list<pg_stat_t*> s; | |
3664 | pg_stat_t::generate_test_instances(s); | |
3665 | o.back()->stats = *s.back(); | |
3666 | } | |
3667 | { | |
3668 | list<pg_hit_set_history_t*> s; | |
3669 | pg_hit_set_history_t::generate_test_instances(s); | |
3670 | o.back()->hit_set = *s.back(); | |
3671 | } | |
3672 | } | |
3673 | ||
3674 | // -- pg_notify_t -- | |
9f95a23c | 3675 | void pg_notify_t::encode(ceph::buffer::list &bl) const |
7c673cae | 3676 | { |
9f95a23c | 3677 | ENCODE_START(3, 2, bl); |
11fdf7f2 TL |
3678 | encode(query_epoch, bl); |
3679 | encode(epoch_sent, bl); | |
3680 | encode(info, bl); | |
3681 | encode(to, bl); | |
3682 | encode(from, bl); | |
9f95a23c | 3683 | encode(past_intervals, bl); |
7c673cae FG |
3684 | ENCODE_FINISH(bl); |
3685 | } | |
3686 | ||
9f95a23c | 3687 | void pg_notify_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 3688 | { |
9f95a23c | 3689 | DECODE_START(3, bl); |
11fdf7f2 TL |
3690 | decode(query_epoch, bl); |
3691 | decode(epoch_sent, bl); | |
3692 | decode(info, bl); | |
3693 | decode(to, bl); | |
3694 | decode(from, bl); | |
9f95a23c TL |
3695 | if (struct_v >= 3) { |
3696 | decode(past_intervals, bl); | |
3697 | } | |
7c673cae FG |
3698 | DECODE_FINISH(bl); |
3699 | } | |
3700 | ||
3701 | void pg_notify_t::dump(Formatter *f) const | |
3702 | { | |
3703 | f->dump_int("from", from); | |
3704 | f->dump_int("to", to); | |
3705 | f->dump_unsigned("query_epoch", query_epoch); | |
3706 | f->dump_unsigned("epoch_sent", epoch_sent); | |
3707 | { | |
3708 | f->open_object_section("info"); | |
3709 | info.dump(f); | |
3710 | f->close_section(); | |
3711 | } | |
9f95a23c | 3712 | f->dump_object("past_intervals", past_intervals); |
7c673cae FG |
3713 | } |
3714 | ||
3715 | void pg_notify_t::generate_test_instances(list<pg_notify_t*>& o) | |
3716 | { | |
9f95a23c TL |
3717 | o.push_back(new pg_notify_t(shard_id_t(3), shard_id_t::NO_SHARD, 1, 1, |
3718 | pg_info_t(), PastIntervals())); | |
3719 | o.push_back(new pg_notify_t(shard_id_t(0), shard_id_t(0), 3, 10, | |
3720 | pg_info_t(), PastIntervals())); | |
7c673cae FG |
3721 | } |
3722 | ||
3723 | ostream &operator<<(ostream &lhs, const pg_notify_t ¬ify) | |
3724 | { | |
3725 | lhs << "(query:" << notify.query_epoch | |
3726 | << " sent:" << notify.epoch_sent | |
3727 | << " " << notify.info; | |
3728 | if (notify.from != shard_id_t::NO_SHARD || | |
3729 | notify.to != shard_id_t::NO_SHARD) | |
3730 | lhs << " " << (unsigned)notify.from | |
3731 | << "->" << (unsigned)notify.to; | |
9f95a23c | 3732 | lhs << " " << notify.past_intervals; |
7c673cae FG |
3733 | return lhs << ")"; |
3734 | } | |
3735 | ||
3736 | // -- pg_interval_t -- | |
3737 | ||
9f95a23c | 3738 | void PastIntervals::pg_interval_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
3739 | { |
3740 | ENCODE_START(4, 2, bl); | |
11fdf7f2 TL |
3741 | encode(first, bl); |
3742 | encode(last, bl); | |
3743 | encode(up, bl); | |
3744 | encode(acting, bl); | |
3745 | encode(maybe_went_rw, bl); | |
3746 | encode(primary, bl); | |
3747 | encode(up_primary, bl); | |
7c673cae FG |
3748 | ENCODE_FINISH(bl); |
3749 | } | |
3750 | ||
9f95a23c | 3751 | void PastIntervals::pg_interval_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
3752 | { |
3753 | DECODE_START_LEGACY_COMPAT_LEN(4, 2, 2, bl); | |
11fdf7f2 TL |
3754 | decode(first, bl); |
3755 | decode(last, bl); | |
3756 | decode(up, bl); | |
3757 | decode(acting, bl); | |
3758 | decode(maybe_went_rw, bl); | |
7c673cae | 3759 | if (struct_v >= 3) { |
11fdf7f2 | 3760 | decode(primary, bl); |
7c673cae FG |
3761 | } else { |
3762 | if (acting.size()) | |
3763 | primary = acting[0]; | |
3764 | } | |
3765 | if (struct_v >= 4) { | |
11fdf7f2 | 3766 | decode(up_primary, bl); |
7c673cae FG |
3767 | } else { |
3768 | if (up.size()) | |
3769 | up_primary = up[0]; | |
3770 | } | |
3771 | DECODE_FINISH(bl); | |
3772 | } | |
3773 | ||
3774 | void PastIntervals::pg_interval_t::dump(Formatter *f) const | |
3775 | { | |
3776 | f->dump_unsigned("first", first); | |
3777 | f->dump_unsigned("last", last); | |
3778 | f->dump_int("maybe_went_rw", maybe_went_rw ? 1 : 0); | |
3779 | f->open_array_section("up"); | |
9f95a23c | 3780 | for (auto p = up.cbegin(); p != up.cend(); ++p) |
7c673cae FG |
3781 | f->dump_int("osd", *p); |
3782 | f->close_section(); | |
3783 | f->open_array_section("acting"); | |
9f95a23c | 3784 | for (auto p = acting.cbegin(); p != acting.cend(); ++p) |
7c673cae FG |
3785 | f->dump_int("osd", *p); |
3786 | f->close_section(); | |
3787 | f->dump_int("primary", primary); | |
3788 | f->dump_int("up_primary", up_primary); | |
3789 | } | |
3790 | ||
3791 | void PastIntervals::pg_interval_t::generate_test_instances(list<pg_interval_t*>& o) | |
3792 | { | |
3793 | o.push_back(new pg_interval_t); | |
3794 | o.push_back(new pg_interval_t); | |
3795 | o.back()->up.push_back(1); | |
3796 | o.back()->acting.push_back(2); | |
3797 | o.back()->acting.push_back(3); | |
3798 | o.back()->first = 4; | |
3799 | o.back()->last = 5; | |
3800 | o.back()->maybe_went_rw = true; | |
3801 | } | |
3802 | ||
3803 | WRITE_CLASS_ENCODER(PastIntervals::pg_interval_t) | |
3804 | ||
7c673cae FG |
3805 | |
3806 | /** | |
3807 | * pi_compact_rep | |
3808 | * | |
3809 | * PastIntervals only needs to be able to answer two questions: | |
3810 | * 1) Where should the primary look for unfound objects? | |
3811 | * 2) List a set of subsets of the OSDs such that contacting at least | |
11fdf7f2 | 3812 | * one from each subset guarantees we speak to at least one witness |
7c673cae FG |
3813 | * of any completed write. |
3814 | * | |
3815 | * Crucially, 2) does not require keeping *all* past intervals. Certainly, | |
3816 | * we don't need to keep any where maybe_went_rw would be false. We also | |
3817 | * needn't keep two intervals where the actingset in one is a subset | |
3818 | * of the other (only need to keep the smaller of the two sets). In order | |
3819 | * to accurately trim the set of intervals as last_epoch_started changes | |
3820 | * without rebuilding the set from scratch, we'll retain the larger set | |
3821 | * if it in an older interval. | |
3822 | */ | |
3823 | struct compact_interval_t { | |
3824 | epoch_t first; | |
3825 | epoch_t last; | |
3826 | set<pg_shard_t> acting; | |
3827 | bool supersedes(const compact_interval_t &other) { | |
3828 | for (auto &&i: acting) { | |
3829 | if (!other.acting.count(i)) | |
3830 | return false; | |
3831 | } | |
3832 | return true; | |
3833 | } | |
3834 | void dump(Formatter *f) const { | |
3835 | f->open_object_section("compact_interval_t"); | |
3836 | f->dump_stream("first") << first; | |
3837 | f->dump_stream("last") << last; | |
3838 | f->dump_stream("acting") << acting; | |
3839 | f->close_section(); | |
3840 | } | |
9f95a23c | 3841 | void encode(ceph::buffer::list &bl) const { |
7c673cae | 3842 | ENCODE_START(1, 1, bl); |
11fdf7f2 TL |
3843 | encode(first, bl); |
3844 | encode(last, bl); | |
3845 | encode(acting, bl); | |
7c673cae FG |
3846 | ENCODE_FINISH(bl); |
3847 | } | |
9f95a23c | 3848 | void decode(ceph::buffer::list::const_iterator &bl) { |
7c673cae | 3849 | DECODE_START(1, bl); |
11fdf7f2 TL |
3850 | decode(first, bl); |
3851 | decode(last, bl); | |
3852 | decode(acting, bl); | |
7c673cae FG |
3853 | DECODE_FINISH(bl); |
3854 | } | |
3855 | static void generate_test_instances(list<compact_interval_t*> & o) { | |
3856 | /* Not going to be used, we'll generate pi_compact_rep directly */ | |
3857 | } | |
3858 | }; | |
3859 | ostream &operator<<(ostream &o, const compact_interval_t &rhs) | |
3860 | { | |
3861 | return o << "([" << rhs.first << "," << rhs.last | |
3862 | << "] acting " << rhs.acting << ")"; | |
3863 | } | |
3864 | WRITE_CLASS_ENCODER(compact_interval_t) | |
3865 | ||
3866 | class pi_compact_rep : public PastIntervals::interval_rep { | |
3867 | epoch_t first = 0; | |
3868 | epoch_t last = 0; // inclusive | |
3869 | set<pg_shard_t> all_participants; | |
3870 | list<compact_interval_t> intervals; | |
3871 | pi_compact_rep( | |
3872 | bool ec_pool, | |
3873 | std::list<PastIntervals::pg_interval_t> &&intervals) { | |
3874 | for (auto &&i: intervals) | |
3875 | add_interval(ec_pool, i); | |
3876 | } | |
3877 | public: | |
3878 | pi_compact_rep() = default; | |
3879 | pi_compact_rep(const pi_compact_rep &) = default; | |
3880 | pi_compact_rep(pi_compact_rep &&) = default; | |
3881 | pi_compact_rep &operator=(const pi_compact_rep &) = default; | |
3882 | pi_compact_rep &operator=(pi_compact_rep &&) = default; | |
3883 | ||
3884 | size_t size() const override { return intervals.size(); } | |
3885 | bool empty() const override { | |
3886 | return first > last || (first == 0 && last == 0); | |
3887 | } | |
3888 | void clear() override { | |
3889 | *this = pi_compact_rep(); | |
3890 | } | |
3891 | pair<epoch_t, epoch_t> get_bounds() const override { | |
3892 | return make_pair(first, last + 1); | |
3893 | } | |
f67539c2 | 3894 | void adjust_start_backwards(epoch_t last_epoch_clean) override { |
11fdf7f2 TL |
3895 | first = last_epoch_clean; |
3896 | } | |
3897 | ||
7c673cae FG |
3898 | set<pg_shard_t> get_all_participants( |
3899 | bool ec_pool) const override { | |
3900 | return all_participants; | |
3901 | } | |
3902 | void add_interval( | |
3903 | bool ec_pool, const PastIntervals::pg_interval_t &interval) override { | |
3904 | if (first == 0) | |
3905 | first = interval.first; | |
11fdf7f2 | 3906 | ceph_assert(interval.last > last); |
7c673cae FG |
3907 | last = interval.last; |
3908 | set<pg_shard_t> acting; | |
3909 | for (unsigned i = 0; i < interval.acting.size(); ++i) { | |
3910 | if (interval.acting[i] == CRUSH_ITEM_NONE) | |
3911 | continue; | |
3912 | acting.insert( | |
3913 | pg_shard_t( | |
3914 | interval.acting[i], | |
3915 | ec_pool ? shard_id_t(i) : shard_id_t::NO_SHARD)); | |
3916 | } | |
3917 | all_participants.insert(acting.begin(), acting.end()); | |
3918 | if (!interval.maybe_went_rw) | |
3919 | return; | |
3920 | intervals.push_back( | |
3921 | compact_interval_t{interval.first, interval.last, acting}); | |
3922 | auto plast = intervals.end(); | |
3923 | --plast; | |
3924 | for (auto cur = intervals.begin(); cur != plast; ) { | |
3925 | if (plast->supersedes(*cur)) { | |
3926 | intervals.erase(cur++); | |
3927 | } else { | |
3928 | ++cur; | |
3929 | } | |
3930 | } | |
3931 | } | |
3932 | unique_ptr<PastIntervals::interval_rep> clone() const override { | |
3933 | return unique_ptr<PastIntervals::interval_rep>(new pi_compact_rep(*this)); | |
3934 | } | |
3935 | ostream &print(ostream &out) const override { | |
3936 | return out << "([" << first << "," << last | |
9f95a23c TL |
3937 | << "] all_participants=" << all_participants |
3938 | << " intervals=" << intervals << ")"; | |
7c673cae | 3939 | } |
9f95a23c | 3940 | void encode(ceph::buffer::list &bl) const override { |
7c673cae | 3941 | ENCODE_START(1, 1, bl); |
11fdf7f2 TL |
3942 | encode(first, bl); |
3943 | encode(last, bl); | |
3944 | encode(all_participants, bl); | |
3945 | encode(intervals, bl); | |
7c673cae FG |
3946 | ENCODE_FINISH(bl); |
3947 | } | |
9f95a23c | 3948 | void decode(ceph::buffer::list::const_iterator &bl) override { |
7c673cae | 3949 | DECODE_START(1, bl); |
11fdf7f2 TL |
3950 | decode(first, bl); |
3951 | decode(last, bl); | |
3952 | decode(all_participants, bl); | |
3953 | decode(intervals, bl); | |
7c673cae FG |
3954 | DECODE_FINISH(bl); |
3955 | } | |
3956 | void dump(Formatter *f) const override { | |
3957 | f->open_object_section("PastIntervals::compact_rep"); | |
3958 | f->dump_stream("first") << first; | |
3959 | f->dump_stream("last") << last; | |
3960 | f->open_array_section("all_participants"); | |
3961 | for (auto& i : all_participants) { | |
3962 | f->dump_object("pg_shard", i); | |
3963 | } | |
3964 | f->close_section(); | |
3965 | f->open_array_section("intervals"); | |
3966 | for (auto &&i: intervals) { | |
3967 | i.dump(f); | |
3968 | } | |
3969 | f->close_section(); | |
3970 | f->close_section(); | |
3971 | } | |
7c673cae FG |
3972 | static void generate_test_instances(list<pi_compact_rep*> &o) { |
3973 | using ival = PastIntervals::pg_interval_t; | |
3974 | using ivallst = std::list<ival>; | |
3975 | o.push_back( | |
3976 | new pi_compact_rep( | |
3977 | true, ivallst | |
3978 | { ival{{0, 1, 2}, {0, 1, 2}, 10, 20, true, 0, 0} | |
3979 | , ival{{ 1, 2}, { 1, 2}, 21, 30, true, 1, 1} | |
3980 | , ival{{ 2}, { 2}, 31, 35, false, 2, 2} | |
3981 | , ival{{0, 2}, {0, 2}, 36, 50, true, 0, 0} | |
3982 | })); | |
3983 | o.push_back( | |
3984 | new pi_compact_rep( | |
3985 | false, ivallst | |
3986 | { ival{{0, 1, 2}, {0, 1, 2}, 10, 20, true, 0, 0} | |
3987 | , ival{{ 1, 2}, { 1, 2}, 21, 30, true, 1, 1} | |
3988 | , ival{{ 2}, { 2}, 31, 35, false, 2, 2} | |
3989 | , ival{{0, 2}, {0, 2}, 36, 50, true, 0, 0} | |
3990 | })); | |
3991 | o.push_back( | |
3992 | new pi_compact_rep( | |
3993 | true, ivallst | |
3994 | { ival{{2, 1, 0}, {2, 1, 0}, 10, 20, true, 1, 1} | |
3995 | , ival{{ 0, 2}, { 0, 2}, 21, 30, true, 0, 0} | |
3996 | , ival{{ 0, 2}, {2, 0}, 31, 35, true, 2, 2} | |
3997 | , ival{{ 0, 2}, { 0, 2}, 36, 50, true, 0, 0} | |
3998 | })); | |
3999 | } | |
4000 | void iterate_mayberw_back_to( | |
7c673cae FG |
4001 | epoch_t les, |
4002 | std::function<void(epoch_t, const set<pg_shard_t> &)> &&f) const override { | |
4003 | for (auto i = intervals.rbegin(); i != intervals.rend(); ++i) { | |
4004 | if (i->last < les) | |
4005 | break; | |
4006 | f(i->first, i->acting); | |
4007 | } | |
4008 | } | |
4009 | virtual ~pi_compact_rep() override {} | |
4010 | }; | |
4011 | WRITE_CLASS_ENCODER(pi_compact_rep) | |
4012 | ||
11fdf7f2 TL |
4013 | PastIntervals::PastIntervals() |
4014 | { | |
4015 | past_intervals.reset(new pi_compact_rep); | |
4016 | } | |
4017 | ||
7c673cae FG |
4018 | PastIntervals::PastIntervals(const PastIntervals &rhs) |
4019 | : past_intervals(rhs.past_intervals ? | |
4020 | rhs.past_intervals->clone() : | |
4021 | nullptr) {} | |
4022 | ||
4023 | PastIntervals &PastIntervals::operator=(const PastIntervals &rhs) | |
4024 | { | |
4025 | PastIntervals other(rhs); | |
31f18b77 | 4026 | swap(other); |
7c673cae FG |
4027 | return *this; |
4028 | } | |
4029 | ||
4030 | ostream& operator<<(ostream& out, const PastIntervals &i) | |
4031 | { | |
4032 | if (i.past_intervals) { | |
4033 | return i.past_intervals->print(out); | |
4034 | } else { | |
4035 | return out << "(empty)"; | |
4036 | } | |
4037 | } | |
4038 | ||
4039 | ostream& operator<<(ostream& out, const PastIntervals::PriorSet &i) | |
4040 | { | |
4041 | return out << "PriorSet(" | |
4042 | << "ec_pool: " << i.ec_pool | |
4043 | << ", probe: " << i.probe | |
4044 | << ", down: " << i.down | |
4045 | << ", blocked_by: " << i.blocked_by | |
4046 | << ", pg_down: " << i.pg_down | |
4047 | << ")"; | |
4048 | } | |
4049 | ||
9f95a23c | 4050 | void PastIntervals::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
4051 | { |
4052 | DECODE_START(1, bl); | |
4053 | __u8 type = 0; | |
11fdf7f2 | 4054 | decode(type, bl); |
7c673cae FG |
4055 | switch (type) { |
4056 | case 0: | |
4057 | break; | |
4058 | case 1: | |
11fdf7f2 | 4059 | ceph_abort_msg("pi_simple_rep support removed post-luminous"); |
7c673cae FG |
4060 | break; |
4061 | case 2: | |
4062 | past_intervals.reset(new pi_compact_rep); | |
4063 | past_intervals->decode(bl); | |
4064 | break; | |
4065 | } | |
4066 | DECODE_FINISH(bl); | |
4067 | } | |
4068 | ||
7c673cae FG |
4069 | void PastIntervals::generate_test_instances(list<PastIntervals*> &o) |
4070 | { | |
7c673cae FG |
4071 | { |
4072 | list<pi_compact_rep *> compact; | |
4073 | pi_compact_rep::generate_test_instances(compact); | |
4074 | for (auto &&i: compact) { | |
4075 | // takes ownership of contents | |
4076 | o.push_back(new PastIntervals(i)); | |
4077 | } | |
4078 | } | |
4079 | return; | |
4080 | } | |
4081 | ||
7c673cae FG |
4082 | bool PastIntervals::is_new_interval( |
4083 | int old_acting_primary, | |
4084 | int new_acting_primary, | |
4085 | const vector<int> &old_acting, | |
4086 | const vector<int> &new_acting, | |
4087 | int old_up_primary, | |
4088 | int new_up_primary, | |
4089 | const vector<int> &old_up, | |
4090 | const vector<int> &new_up, | |
4091 | int old_size, | |
4092 | int new_size, | |
4093 | int old_min_size, | |
4094 | int new_min_size, | |
4095 | unsigned old_pg_num, | |
4096 | unsigned new_pg_num, | |
11fdf7f2 TL |
4097 | unsigned old_pg_num_pending, |
4098 | unsigned new_pg_num_pending, | |
7c673cae FG |
4099 | bool old_sort_bitwise, |
4100 | bool new_sort_bitwise, | |
c07f9fc5 FG |
4101 | bool old_recovery_deletes, |
4102 | bool new_recovery_deletes, | |
f67539c2 TL |
4103 | uint32_t old_crush_count, |
4104 | uint32_t new_crush_count, | |
4105 | uint32_t old_crush_target, | |
4106 | uint32_t new_crush_target, | |
4107 | uint32_t old_crush_barrier, | |
4108 | uint32_t new_crush_barrier, | |
4109 | int32_t old_crush_member, | |
4110 | int32_t new_crush_member, | |
7c673cae FG |
4111 | pg_t pgid) { |
4112 | return old_acting_primary != new_acting_primary || | |
4113 | new_acting != old_acting || | |
4114 | old_up_primary != new_up_primary || | |
4115 | new_up != old_up || | |
4116 | old_min_size != new_min_size || | |
4117 | old_size != new_size || | |
4118 | pgid.is_split(old_pg_num, new_pg_num, 0) || | |
11fdf7f2 TL |
4119 | // (is or was) pre-merge source |
4120 | pgid.is_merge_source(old_pg_num_pending, new_pg_num_pending, 0) || | |
4121 | pgid.is_merge_source(new_pg_num_pending, old_pg_num_pending, 0) || | |
4122 | // merge source | |
4123 | pgid.is_merge_source(old_pg_num, new_pg_num, 0) || | |
4124 | // (is or was) pre-merge target | |
4125 | pgid.is_merge_target(old_pg_num_pending, new_pg_num_pending) || | |
4126 | pgid.is_merge_target(new_pg_num_pending, old_pg_num_pending) || | |
4127 | // merge target | |
4128 | pgid.is_merge_target(old_pg_num, new_pg_num) || | |
c07f9fc5 | 4129 | old_sort_bitwise != new_sort_bitwise || |
f67539c2 TL |
4130 | old_recovery_deletes != new_recovery_deletes || |
4131 | old_crush_count != new_crush_count || | |
4132 | old_crush_target != new_crush_target || | |
4133 | old_crush_barrier != new_crush_barrier || | |
4134 | old_crush_member != new_crush_member; | |
7c673cae FG |
4135 | } |
4136 | ||
4137 | bool PastIntervals::is_new_interval( | |
4138 | int old_acting_primary, | |
4139 | int new_acting_primary, | |
4140 | const vector<int> &old_acting, | |
4141 | const vector<int> &new_acting, | |
4142 | int old_up_primary, | |
4143 | int new_up_primary, | |
4144 | const vector<int> &old_up, | |
4145 | const vector<int> &new_up, | |
9f95a23c TL |
4146 | const OSDMap *osdmap, |
4147 | const OSDMap *lastmap, | |
11fdf7f2 TL |
4148 | pg_t pgid) |
4149 | { | |
4150 | const pg_pool_t *plast = lastmap->get_pg_pool(pgid.pool()); | |
4151 | if (!plast) { | |
4152 | return false; // after pool is deleted there are no more interval changes | |
4153 | } | |
4154 | const pg_pool_t *pi = osdmap->get_pg_pool(pgid.pool()); | |
4155 | if (!pi) { | |
4156 | return true; // pool was deleted this epoch -> (final!) interval change | |
4157 | } | |
4158 | return | |
7c673cae FG |
4159 | is_new_interval(old_acting_primary, |
4160 | new_acting_primary, | |
4161 | old_acting, | |
4162 | new_acting, | |
4163 | old_up_primary, | |
4164 | new_up_primary, | |
4165 | old_up, | |
4166 | new_up, | |
11fdf7f2 TL |
4167 | plast->size, |
4168 | pi->size, | |
4169 | plast->min_size, | |
4170 | pi->min_size, | |
4171 | plast->get_pg_num(), | |
4172 | pi->get_pg_num(), | |
4173 | plast->get_pg_num_pending(), | |
4174 | pi->get_pg_num_pending(), | |
7c673cae FG |
4175 | lastmap->test_flag(CEPH_OSDMAP_SORTBITWISE), |
4176 | osdmap->test_flag(CEPH_OSDMAP_SORTBITWISE), | |
c07f9fc5 FG |
4177 | lastmap->test_flag(CEPH_OSDMAP_RECOVERY_DELETES), |
4178 | osdmap->test_flag(CEPH_OSDMAP_RECOVERY_DELETES), | |
f67539c2 TL |
4179 | plast->peering_crush_bucket_count, pi->peering_crush_bucket_count, |
4180 | plast->peering_crush_bucket_target, pi->peering_crush_bucket_target, | |
4181 | plast->peering_crush_bucket_barrier, pi->peering_crush_bucket_barrier, | |
4182 | plast->peering_crush_mandatory_member, pi->peering_crush_mandatory_member, | |
7c673cae FG |
4183 | pgid); |
4184 | } | |
4185 | ||
4186 | bool PastIntervals::check_new_interval( | |
4187 | int old_acting_primary, | |
4188 | int new_acting_primary, | |
4189 | const vector<int> &old_acting, | |
4190 | const vector<int> &new_acting, | |
4191 | int old_up_primary, | |
4192 | int new_up_primary, | |
4193 | const vector<int> &old_up, | |
4194 | const vector<int> &new_up, | |
4195 | epoch_t same_interval_since, | |
4196 | epoch_t last_epoch_clean, | |
9f95a23c TL |
4197 | const OSDMap *osdmap, |
4198 | const OSDMap *lastmap, | |
7c673cae | 4199 | pg_t pgid, |
9f95a23c | 4200 | const IsPGRecoverablePredicate &could_have_gone_active, |
7c673cae FG |
4201 | PastIntervals *past_intervals, |
4202 | std::ostream *out) | |
4203 | { | |
4204 | /* | |
4205 | * We have to be careful to gracefully deal with situations like | |
4206 | * so. Say we have a power outage or something that takes out both | |
4207 | * OSDs, but the monitor doesn't mark them down in the same epoch. | |
4208 | * The history may look like | |
4209 | * | |
4210 | * 1: A B | |
4211 | * 2: B | |
4212 | * 3: let's say B dies for good, too (say, from the power spike) | |
4213 | * 4: A | |
4214 | * | |
4215 | * which makes it look like B may have applied updates to the PG | |
4216 | * that we need in order to proceed. This sucks... | |
4217 | * | |
4218 | * To minimize the risk of this happening, we CANNOT go active if | |
4219 | * _any_ OSDs in the prior set are down until we send an MOSDAlive | |
4220 | * to the monitor such that the OSDMap sets osd_up_thru to an epoch. | |
4221 | * Then, we have something like | |
4222 | * | |
4223 | * 1: A B | |
4224 | * 2: B up_thru[B]=0 | |
4225 | * 3: | |
4226 | * 4: A | |
4227 | * | |
4228 | * -> we can ignore B, bc it couldn't have gone active (up_thru still 0). | |
4229 | * | |
4230 | * or, | |
4231 | * | |
4232 | * 1: A B | |
4233 | * 2: B up_thru[B]=0 | |
4234 | * 3: B up_thru[B]=2 | |
4235 | * 4: | |
4236 | * 5: A | |
4237 | * | |
4238 | * -> we must wait for B, bc it was alive through 2, and could have | |
4239 | * written to the pg. | |
4240 | * | |
4241 | * If B is really dead, then an administrator will need to manually | |
4242 | * intervene by marking the OSD as "lost." | |
4243 | */ | |
4244 | ||
4245 | // remember past interval | |
4246 | // NOTE: a change in the up set primary triggers an interval | |
4247 | // change, even though the interval members in the pg_interval_t | |
4248 | // do not change. | |
11fdf7f2 TL |
4249 | ceph_assert(past_intervals); |
4250 | ceph_assert(past_intervals->past_intervals); | |
7c673cae FG |
4251 | if (is_new_interval( |
4252 | old_acting_primary, | |
4253 | new_acting_primary, | |
4254 | old_acting, | |
4255 | new_acting, | |
4256 | old_up_primary, | |
4257 | new_up_primary, | |
4258 | old_up, | |
4259 | new_up, | |
4260 | osdmap, | |
4261 | lastmap, | |
4262 | pgid)) { | |
4263 | pg_interval_t i; | |
4264 | i.first = same_interval_since; | |
4265 | i.last = osdmap->get_epoch() - 1; | |
11fdf7f2 | 4266 | ceph_assert(i.first <= i.last); |
7c673cae FG |
4267 | i.acting = old_acting; |
4268 | i.up = old_up; | |
4269 | i.primary = old_acting_primary; | |
4270 | i.up_primary = old_up_primary; | |
4271 | ||
4272 | unsigned num_acting = 0; | |
9f95a23c | 4273 | for (auto p = i.acting.cbegin(); p != i.acting.cend(); ++p) |
7c673cae FG |
4274 | if (*p != CRUSH_ITEM_NONE) |
4275 | ++num_acting; | |
4276 | ||
11fdf7f2 | 4277 | ceph_assert(lastmap->get_pools().count(pgid.pool())); |
7c673cae FG |
4278 | const pg_pool_t& old_pg_pool = lastmap->get_pools().find(pgid.pool())->second; |
4279 | set<pg_shard_t> old_acting_shards; | |
4280 | old_pg_pool.convert_to_pg_shards(old_acting, &old_acting_shards); | |
4281 | ||
4282 | if (num_acting && | |
4283 | i.primary != -1 && | |
4284 | num_acting >= old_pg_pool.min_size && | |
f67539c2 TL |
4285 | (!old_pg_pool.is_stretch_pool() || |
4286 | old_pg_pool.stretch_set_can_peer(old_acting, *lastmap, out)) && | |
9f95a23c | 4287 | could_have_gone_active(old_acting_shards)) { |
7c673cae FG |
4288 | if (out) |
4289 | *out << __func__ << " " << i | |
7c673cae FG |
4290 | << " up_thru " << lastmap->get_up_thru(i.primary) |
4291 | << " up_from " << lastmap->get_up_from(i.primary) | |
11fdf7f2 | 4292 | << " last_epoch_clean " << last_epoch_clean; |
7c673cae FG |
4293 | if (lastmap->get_up_thru(i.primary) >= i.first && |
4294 | lastmap->get_up_from(i.primary) <= i.first) { | |
4295 | i.maybe_went_rw = true; | |
4296 | if (out) | |
11fdf7f2 | 4297 | *out << " " << i |
7c673cae FG |
4298 | << " : primary up " << lastmap->get_up_from(i.primary) |
4299 | << "-" << lastmap->get_up_thru(i.primary) | |
4300 | << " includes interval" | |
11fdf7f2 | 4301 | << std::endl; |
7c673cae FG |
4302 | } else if (last_epoch_clean >= i.first && |
4303 | last_epoch_clean <= i.last) { | |
4304 | // If the last_epoch_clean is included in this interval, then | |
4305 | // the pg must have been rw (for recovery to have completed). | |
4306 | // This is important because we won't know the _real_ | |
4307 | // first_epoch because we stop at last_epoch_clean, and we | |
4308 | // don't want the oldest interval to randomly have | |
4309 | // maybe_went_rw false depending on the relative up_thru vs | |
4310 | // last_epoch_clean timing. | |
4311 | i.maybe_went_rw = true; | |
4312 | if (out) | |
11fdf7f2 | 4313 | *out << " " << i |
7c673cae FG |
4314 | << " : includes last_epoch_clean " << last_epoch_clean |
4315 | << " and presumed to have been rw" | |
4316 | << std::endl; | |
4317 | } else { | |
4318 | i.maybe_went_rw = false; | |
4319 | if (out) | |
11fdf7f2 | 4320 | *out << " " << i |
7c673cae FG |
4321 | << " : primary up " << lastmap->get_up_from(i.primary) |
4322 | << "-" << lastmap->get_up_thru(i.primary) | |
4323 | << " does not include interval" | |
11fdf7f2 | 4324 | << std::endl; |
7c673cae FG |
4325 | } |
4326 | } else { | |
4327 | i.maybe_went_rw = false; | |
4328 | if (out) | |
4329 | *out << __func__ << " " << i << " : acting set is too small" << std::endl; | |
4330 | } | |
11fdf7f2 | 4331 | past_intervals->past_intervals->add_interval(old_pg_pool.is_erasure(), i); |
7c673cae FG |
4332 | return true; |
4333 | } else { | |
4334 | return false; | |
4335 | } | |
4336 | } | |
4337 | ||
7c673cae FG |
4338 | // true if the given map affects the prior set |
4339 | bool PastIntervals::PriorSet::affected_by_map( | |
4340 | const OSDMap &osdmap, | |
4341 | const DoutPrefixProvider *dpp) const | |
4342 | { | |
9f95a23c | 4343 | for (auto p = probe.begin(); p != probe.end(); ++p) { |
7c673cae FG |
4344 | int o = p->osd; |
4345 | ||
4346 | // did someone in the prior set go down? | |
4347 | if (osdmap.is_down(o) && down.count(o) == 0) { | |
4348 | ldpp_dout(dpp, 10) << "affected_by_map osd." << o << " now down" << dendl; | |
4349 | return true; | |
4350 | } | |
4351 | ||
4352 | // did a down osd in cur get (re)marked as lost? | |
9f95a23c | 4353 | auto r = blocked_by.find(o); |
7c673cae FG |
4354 | if (r != blocked_by.end()) { |
4355 | if (!osdmap.exists(o)) { | |
4356 | ldpp_dout(dpp, 10) << "affected_by_map osd." << o << " no longer exists" << dendl; | |
4357 | return true; | |
4358 | } | |
4359 | if (osdmap.get_info(o).lost_at != r->second) { | |
4360 | ldpp_dout(dpp, 10) << "affected_by_map osd." << o << " (re)marked as lost" << dendl; | |
4361 | return true; | |
4362 | } | |
4363 | } | |
4364 | } | |
4365 | ||
4366 | // did someone in the prior down set go up? | |
9f95a23c | 4367 | for (auto p = down.cbegin(); p != down.cend(); ++p) { |
7c673cae FG |
4368 | int o = *p; |
4369 | ||
4370 | if (osdmap.is_up(o)) { | |
4371 | ldpp_dout(dpp, 10) << "affected_by_map osd." << o << " now up" << dendl; | |
4372 | return true; | |
4373 | } | |
4374 | ||
4375 | // did someone in the prior set get lost or destroyed? | |
4376 | if (!osdmap.exists(o)) { | |
4377 | ldpp_dout(dpp, 10) << "affected_by_map osd." << o << " no longer exists" << dendl; | |
4378 | return true; | |
4379 | } | |
4380 | // did a down osd in down get (re)marked as lost? | |
9f95a23c | 4381 | auto r = blocked_by.find(o); |
7c673cae FG |
4382 | if (r != blocked_by.end()) { |
4383 | if (osdmap.get_info(o).lost_at != r->second) { | |
4384 | ldpp_dout(dpp, 10) << "affected_by_map osd." << o << " (re)marked as lost" << dendl; | |
4385 | return true; | |
4386 | } | |
4387 | } | |
4388 | } | |
4389 | ||
4390 | return false; | |
4391 | } | |
4392 | ||
4393 | ostream& operator<<(ostream& out, const PastIntervals::pg_interval_t& i) | |
4394 | { | |
4395 | out << "interval(" << i.first << "-" << i.last | |
4396 | << " up " << i.up << "(" << i.up_primary << ")" | |
4397 | << " acting " << i.acting << "(" << i.primary << ")"; | |
4398 | if (i.maybe_went_rw) | |
4399 | out << " maybe_went_rw"; | |
4400 | out << ")"; | |
4401 | return out; | |
4402 | } | |
4403 | ||
4404 | ||
4405 | ||
4406 | // -- pg_query_t -- | |
4407 | ||
9f95a23c | 4408 | void pg_query_t::encode(ceph::buffer::list &bl, uint64_t features) const { |
7c673cae | 4409 | ENCODE_START(3, 3, bl); |
11fdf7f2 TL |
4410 | encode(type, bl); |
4411 | encode(since, bl); | |
7c673cae | 4412 | history.encode(bl); |
11fdf7f2 TL |
4413 | encode(epoch_sent, bl); |
4414 | encode(to, bl); | |
4415 | encode(from, bl); | |
7c673cae FG |
4416 | ENCODE_FINISH(bl); |
4417 | } | |
4418 | ||
9f95a23c | 4419 | void pg_query_t::decode(ceph::buffer::list::const_iterator &bl) { |
7c673cae | 4420 | DECODE_START(3, bl); |
11fdf7f2 TL |
4421 | decode(type, bl); |
4422 | decode(since, bl); | |
7c673cae | 4423 | history.decode(bl); |
11fdf7f2 TL |
4424 | decode(epoch_sent, bl); |
4425 | decode(to, bl); | |
4426 | decode(from, bl); | |
7c673cae FG |
4427 | DECODE_FINISH(bl); |
4428 | } | |
4429 | ||
4430 | void pg_query_t::dump(Formatter *f) const | |
4431 | { | |
4432 | f->dump_int("from", from); | |
4433 | f->dump_int("to", to); | |
4434 | f->dump_string("type", get_type_name()); | |
4435 | f->dump_stream("since") << since; | |
4436 | f->dump_stream("epoch_sent") << epoch_sent; | |
4437 | f->open_object_section("history"); | |
4438 | history.dump(f); | |
4439 | f->close_section(); | |
4440 | } | |
4441 | void pg_query_t::generate_test_instances(list<pg_query_t*>& o) | |
4442 | { | |
4443 | o.push_back(new pg_query_t()); | |
4444 | list<pg_history_t*> h; | |
4445 | pg_history_t::generate_test_instances(h); | |
4446 | o.push_back(new pg_query_t(pg_query_t::INFO, shard_id_t(1), shard_id_t(2), *h.back(), 4)); | |
4447 | o.push_back(new pg_query_t(pg_query_t::MISSING, shard_id_t(2), shard_id_t(3), *h.back(), 4)); | |
4448 | o.push_back(new pg_query_t(pg_query_t::LOG, shard_id_t(0), shard_id_t(0), | |
4449 | eversion_t(4, 5), *h.back(), 4)); | |
4450 | o.push_back(new pg_query_t(pg_query_t::FULLLOG, | |
4451 | shard_id_t::NO_SHARD, shard_id_t::NO_SHARD, | |
4452 | *h.back(), 5)); | |
4453 | } | |
4454 | ||
9f95a23c TL |
4455 | // -- pg_lease_t -- |
4456 | ||
4457 | void pg_lease_t::encode(bufferlist& bl) const | |
4458 | { | |
4459 | ENCODE_START(1, 1, bl); | |
4460 | encode(readable_until, bl); | |
4461 | encode(readable_until_ub, bl); | |
4462 | encode(interval, bl); | |
4463 | ENCODE_FINISH(bl); | |
4464 | } | |
4465 | ||
4466 | void pg_lease_t::decode(bufferlist::const_iterator& p) | |
4467 | { | |
4468 | DECODE_START(1, p); | |
4469 | decode(readable_until, p); | |
4470 | decode(readable_until_ub, p); | |
4471 | decode(interval, p); | |
4472 | DECODE_FINISH(p); | |
4473 | } | |
4474 | ||
4475 | void pg_lease_t::dump(Formatter *f) const | |
4476 | { | |
4477 | f->dump_stream("readable_until") << readable_until; | |
4478 | f->dump_stream("readable_until_ub") << readable_until_ub; | |
4479 | f->dump_stream("interval") << interval; | |
4480 | } | |
4481 | ||
4482 | void pg_lease_t::generate_test_instances(std::list<pg_lease_t*>& o) | |
4483 | { | |
4484 | o.push_back(new pg_lease_t()); | |
4485 | o.push_back(new pg_lease_t()); | |
4486 | o.back()->readable_until = make_timespan(1.5); | |
4487 | o.back()->readable_until_ub = make_timespan(3.4); | |
4488 | o.back()->interval = make_timespan(1.0); | |
4489 | } | |
4490 | ||
4491 | // -- pg_lease_ack_t -- | |
4492 | ||
4493 | void pg_lease_ack_t::encode(bufferlist& bl) const | |
4494 | { | |
4495 | ENCODE_START(1, 1, bl); | |
4496 | encode(readable_until_ub, bl); | |
4497 | ENCODE_FINISH(bl); | |
4498 | } | |
4499 | ||
4500 | void pg_lease_ack_t::decode(bufferlist::const_iterator& p) | |
4501 | { | |
4502 | DECODE_START(1, p); | |
4503 | decode(readable_until_ub, p); | |
4504 | DECODE_FINISH(p); | |
4505 | } | |
4506 | ||
4507 | void pg_lease_ack_t::dump(Formatter *f) const | |
4508 | { | |
4509 | f->dump_stream("readable_until_ub") << readable_until_ub; | |
4510 | } | |
4511 | ||
4512 | void pg_lease_ack_t::generate_test_instances(std::list<pg_lease_ack_t*>& o) | |
4513 | { | |
4514 | o.push_back(new pg_lease_ack_t()); | |
4515 | o.push_back(new pg_lease_ack_t()); | |
4516 | o.back()->readable_until_ub = make_timespan(3.4); | |
4517 | } | |
4518 | ||
4519 | ||
7c673cae FG |
4520 | // -- ObjectModDesc -- |
4521 | void ObjectModDesc::visit(Visitor *visitor) const | |
4522 | { | |
11fdf7f2 | 4523 | auto bp = bl.cbegin(); |
7c673cae FG |
4524 | try { |
4525 | while (!bp.end()) { | |
4526 | DECODE_START(max_required_version, bp); | |
4527 | uint8_t code; | |
11fdf7f2 | 4528 | decode(code, bp); |
7c673cae FG |
4529 | switch (code) { |
4530 | case APPEND: { | |
4531 | uint64_t size; | |
11fdf7f2 | 4532 | decode(size, bp); |
7c673cae FG |
4533 | visitor->append(size); |
4534 | break; | |
4535 | } | |
4536 | case SETATTRS: { | |
9f95a23c | 4537 | map<string, std::optional<ceph::buffer::list> > attrs; |
11fdf7f2 | 4538 | decode(attrs, bp); |
7c673cae FG |
4539 | visitor->setattrs(attrs); |
4540 | break; | |
4541 | } | |
4542 | case DELETE: { | |
4543 | version_t old_version; | |
11fdf7f2 | 4544 | decode(old_version, bp); |
7c673cae FG |
4545 | visitor->rmobject(old_version); |
4546 | break; | |
4547 | } | |
4548 | case CREATE: { | |
4549 | visitor->create(); | |
4550 | break; | |
4551 | } | |
4552 | case UPDATE_SNAPS: { | |
4553 | set<snapid_t> snaps; | |
11fdf7f2 | 4554 | decode(snaps, bp); |
7c673cae FG |
4555 | visitor->update_snaps(snaps); |
4556 | break; | |
4557 | } | |
4558 | case TRY_DELETE: { | |
4559 | version_t old_version; | |
11fdf7f2 | 4560 | decode(old_version, bp); |
7c673cae FG |
4561 | visitor->try_rmobject(old_version); |
4562 | break; | |
4563 | } | |
4564 | case ROLLBACK_EXTENTS: { | |
4565 | vector<pair<uint64_t, uint64_t> > extents; | |
4566 | version_t gen; | |
11fdf7f2 TL |
4567 | decode(gen, bp); |
4568 | decode(extents, bp); | |
7c673cae FG |
4569 | visitor->rollback_extents(gen,extents); |
4570 | break; | |
4571 | } | |
4572 | default: | |
11fdf7f2 | 4573 | ceph_abort_msg("Invalid rollback code"); |
7c673cae FG |
4574 | } |
4575 | DECODE_FINISH(bp); | |
4576 | } | |
4577 | } catch (...) { | |
11fdf7f2 | 4578 | ceph_abort_msg("Invalid encoding"); |
7c673cae FG |
4579 | } |
4580 | } | |
4581 | ||
4582 | struct DumpVisitor : public ObjectModDesc::Visitor { | |
4583 | Formatter *f; | |
4584 | explicit DumpVisitor(Formatter *f) : f(f) {} | |
4585 | void append(uint64_t old_size) override { | |
4586 | f->open_object_section("op"); | |
4587 | f->dump_string("code", "APPEND"); | |
4588 | f->dump_unsigned("old_size", old_size); | |
4589 | f->close_section(); | |
4590 | } | |
9f95a23c | 4591 | void setattrs(map<string, std::optional<ceph::buffer::list> > &attrs) override { |
7c673cae FG |
4592 | f->open_object_section("op"); |
4593 | f->dump_string("code", "SETATTRS"); | |
4594 | f->open_array_section("attrs"); | |
9f95a23c | 4595 | for (auto i = attrs.begin(); i != attrs.end(); ++i) { |
7c673cae FG |
4596 | f->dump_string("attr_name", i->first); |
4597 | } | |
4598 | f->close_section(); | |
4599 | f->close_section(); | |
4600 | } | |
4601 | void rmobject(version_t old_version) override { | |
4602 | f->open_object_section("op"); | |
4603 | f->dump_string("code", "RMOBJECT"); | |
4604 | f->dump_unsigned("old_version", old_version); | |
4605 | f->close_section(); | |
4606 | } | |
4607 | void try_rmobject(version_t old_version) override { | |
4608 | f->open_object_section("op"); | |
4609 | f->dump_string("code", "TRY_RMOBJECT"); | |
4610 | f->dump_unsigned("old_version", old_version); | |
4611 | f->close_section(); | |
4612 | } | |
4613 | void create() override { | |
4614 | f->open_object_section("op"); | |
4615 | f->dump_string("code", "CREATE"); | |
4616 | f->close_section(); | |
4617 | } | |
4618 | void update_snaps(const set<snapid_t> &snaps) override { | |
4619 | f->open_object_section("op"); | |
4620 | f->dump_string("code", "UPDATE_SNAPS"); | |
4621 | f->dump_stream("snaps") << snaps; | |
4622 | f->close_section(); | |
4623 | } | |
4624 | void rollback_extents( | |
4625 | version_t gen, | |
4626 | const vector<pair<uint64_t, uint64_t> > &extents) override { | |
4627 | f->open_object_section("op"); | |
4628 | f->dump_string("code", "ROLLBACK_EXTENTS"); | |
4629 | f->dump_unsigned("gen", gen); | |
4630 | f->dump_stream("snaps") << extents; | |
4631 | f->close_section(); | |
4632 | } | |
4633 | }; | |
4634 | ||
4635 | void ObjectModDesc::dump(Formatter *f) const | |
4636 | { | |
4637 | f->open_object_section("object_mod_desc"); | |
4638 | f->dump_bool("can_local_rollback", can_local_rollback); | |
4639 | f->dump_bool("rollback_info_completed", rollback_info_completed); | |
4640 | { | |
4641 | f->open_array_section("ops"); | |
4642 | DumpVisitor vis(f); | |
4643 | visit(&vis); | |
4644 | f->close_section(); | |
4645 | } | |
4646 | f->close_section(); | |
4647 | } | |
4648 | ||
4649 | void ObjectModDesc::generate_test_instances(list<ObjectModDesc*>& o) | |
4650 | { | |
9f95a23c | 4651 | map<string, std::optional<ceph::buffer::list> > attrs; |
7c673cae FG |
4652 | attrs[OI_ATTR]; |
4653 | attrs[SS_ATTR]; | |
4654 | attrs["asdf"]; | |
4655 | o.push_back(new ObjectModDesc()); | |
4656 | o.back()->append(100); | |
4657 | o.back()->setattrs(attrs); | |
4658 | o.push_back(new ObjectModDesc()); | |
4659 | o.back()->rmobject(1001); | |
4660 | o.push_back(new ObjectModDesc()); | |
4661 | o.back()->create(); | |
4662 | o.back()->setattrs(attrs); | |
4663 | o.push_back(new ObjectModDesc()); | |
4664 | o.back()->create(); | |
4665 | o.back()->setattrs(attrs); | |
4666 | o.back()->mark_unrollbackable(); | |
4667 | o.back()->append(1000); | |
4668 | } | |
4669 | ||
9f95a23c | 4670 | void ObjectModDesc::encode(ceph::buffer::list &_bl) const |
7c673cae FG |
4671 | { |
4672 | ENCODE_START(max_required_version, max_required_version, _bl); | |
11fdf7f2 TL |
4673 | encode(can_local_rollback, _bl); |
4674 | encode(rollback_info_completed, _bl); | |
4675 | encode(bl, _bl); | |
7c673cae FG |
4676 | ENCODE_FINISH(_bl); |
4677 | } | |
9f95a23c | 4678 | void ObjectModDesc::decode(ceph::buffer::list::const_iterator &_bl) |
7c673cae FG |
4679 | { |
4680 | DECODE_START(2, _bl); | |
4681 | max_required_version = struct_v; | |
11fdf7f2 TL |
4682 | decode(can_local_rollback, _bl); |
4683 | decode(rollback_info_completed, _bl); | |
4684 | decode(bl, _bl); | |
9f95a23c | 4685 | // ensure bl does not pin a larger ceph::buffer in memory |
7c673cae | 4686 | bl.rebuild(); |
31f18b77 | 4687 | bl.reassign_to_mempool(mempool::mempool_osd_pglog); |
7c673cae FG |
4688 | DECODE_FINISH(_bl); |
4689 | } | |
4690 | ||
9f95a23c TL |
4691 | std::atomic<uint32_t> ObjectCleanRegions::max_num_intervals = {10}; |
4692 | ||
4693 | void ObjectCleanRegions::set_max_num_intervals(uint32_t num) | |
4694 | { | |
4695 | max_num_intervals = num; | |
4696 | } | |
4697 | ||
4698 | void ObjectCleanRegions::trim() | |
4699 | { | |
4700 | while(clean_offsets.num_intervals() > max_num_intervals) { | |
4701 | typename interval_set<uint64_t>::iterator shortest_interval = clean_offsets.begin(); | |
4702 | if (shortest_interval == clean_offsets.end()) | |
4703 | break; | |
4704 | for (typename interval_set<uint64_t>::iterator it = clean_offsets.begin(); | |
4705 | it != clean_offsets.end(); | |
4706 | ++it) { | |
4707 | if (it.get_len() < shortest_interval.get_len()) | |
4708 | shortest_interval = it; | |
4709 | } | |
4710 | clean_offsets.erase(shortest_interval); | |
4711 | } | |
4712 | } | |
4713 | ||
4714 | void ObjectCleanRegions::merge(const ObjectCleanRegions &other) | |
4715 | { | |
4716 | clean_offsets.intersection_of(other.clean_offsets); | |
4717 | clean_omap = clean_omap && other.clean_omap; | |
4718 | trim(); | |
4719 | } | |
4720 | ||
4721 | void ObjectCleanRegions::mark_data_region_dirty(uint64_t offset, uint64_t len) | |
4722 | { | |
4723 | interval_set<uint64_t> clean_region; | |
4724 | clean_region.insert(0, (uint64_t)-1); | |
4725 | clean_region.erase(offset, len); | |
4726 | clean_offsets.intersection_of(clean_region); | |
4727 | trim(); | |
4728 | } | |
4729 | ||
f67539c2 TL |
4730 | bool ObjectCleanRegions::is_clean_region(uint64_t offset, uint64_t len) const |
4731 | { | |
4732 | return clean_offsets.contains(offset, len); | |
4733 | } | |
4734 | ||
9f95a23c TL |
4735 | void ObjectCleanRegions::mark_omap_dirty() |
4736 | { | |
4737 | clean_omap = false; | |
4738 | } | |
4739 | ||
4740 | void ObjectCleanRegions::mark_object_new() | |
4741 | { | |
4742 | new_object = true; | |
4743 | } | |
4744 | ||
4745 | void ObjectCleanRegions::mark_fully_dirty() | |
4746 | { | |
4747 | mark_data_region_dirty(0, (uint64_t)-1); | |
4748 | mark_omap_dirty(); | |
4749 | mark_object_new(); | |
4750 | } | |
4751 | ||
4752 | interval_set<uint64_t> ObjectCleanRegions::get_dirty_regions() const | |
4753 | { | |
4754 | interval_set<uint64_t> dirty_region; | |
4755 | dirty_region.insert(0, (uint64_t)-1); | |
4756 | dirty_region.subtract(clean_offsets); | |
4757 | return dirty_region; | |
4758 | } | |
4759 | ||
4760 | bool ObjectCleanRegions::omap_is_dirty() const | |
4761 | { | |
4762 | return !clean_omap; | |
4763 | } | |
4764 | ||
4765 | bool ObjectCleanRegions::object_is_exist() const | |
4766 | { | |
4767 | return !new_object; | |
4768 | } | |
4769 | ||
4770 | void ObjectCleanRegions::encode(bufferlist &bl) const | |
4771 | { | |
4772 | ENCODE_START(1, 1, bl); | |
4773 | using ceph::encode; | |
4774 | encode(clean_offsets, bl); | |
4775 | encode(clean_omap, bl); | |
4776 | encode(new_object, bl); | |
4777 | ENCODE_FINISH(bl); | |
4778 | } | |
4779 | ||
4780 | void ObjectCleanRegions::decode(bufferlist::const_iterator &bl) | |
4781 | { | |
4782 | DECODE_START(1, bl); | |
4783 | using ceph::decode; | |
4784 | decode(clean_offsets, bl); | |
4785 | decode(clean_omap, bl); | |
4786 | decode(new_object, bl); | |
4787 | DECODE_FINISH(bl); | |
4788 | } | |
4789 | ||
4790 | void ObjectCleanRegions::dump(Formatter *f) const | |
4791 | { | |
4792 | f->open_object_section("object_clean_regions"); | |
4793 | f->dump_stream("clean_offsets") << clean_offsets; | |
4794 | f->dump_bool("clean_omap", clean_omap); | |
4795 | f->dump_bool("new_object", new_object); | |
4796 | f->close_section(); | |
4797 | } | |
4798 | ||
4799 | void ObjectCleanRegions::generate_test_instances(list<ObjectCleanRegions*>& o) | |
4800 | { | |
4801 | o.push_back(new ObjectCleanRegions()); | |
4802 | o.push_back(new ObjectCleanRegions()); | |
4803 | o.back()->mark_data_region_dirty(4096, 40960); | |
4804 | o.back()->mark_omap_dirty(); | |
4805 | o.back()->mark_object_new(); | |
4806 | } | |
4807 | ||
4808 | ostream& operator<<(ostream& out, const ObjectCleanRegions& ocr) | |
4809 | { | |
4810 | return out << "clean_offsets: " << ocr.clean_offsets | |
4811 | << ", clean_omap: " << ocr.clean_omap | |
4812 | << ", new_object: " << ocr.new_object; | |
4813 | } | |
4814 | ||
7c673cae FG |
4815 | // -- pg_log_entry_t -- |
4816 | ||
4817 | string pg_log_entry_t::get_key_name() const | |
4818 | { | |
4819 | return version.get_key_name(); | |
4820 | } | |
4821 | ||
9f95a23c | 4822 | void pg_log_entry_t::encode_with_checksum(ceph::buffer::list& bl) const |
7c673cae | 4823 | { |
11fdf7f2 | 4824 | using ceph::encode; |
9f95a23c | 4825 | ceph::buffer::list ebl(sizeof(*this)*2); |
11fdf7f2 | 4826 | this->encode(ebl); |
7c673cae | 4827 | __u32 crc = ebl.crc32c(0); |
11fdf7f2 TL |
4828 | encode(ebl, bl); |
4829 | encode(crc, bl); | |
7c673cae FG |
4830 | } |
4831 | ||
9f95a23c | 4832 | void pg_log_entry_t::decode_with_checksum(ceph::buffer::list::const_iterator& p) |
7c673cae | 4833 | { |
11fdf7f2 | 4834 | using ceph::decode; |
9f95a23c | 4835 | ceph::buffer::list bl; |
11fdf7f2 | 4836 | decode(bl, p); |
7c673cae | 4837 | __u32 crc; |
11fdf7f2 | 4838 | decode(crc, p); |
7c673cae | 4839 | if (crc != bl.crc32c(0)) |
9f95a23c | 4840 | throw ceph::buffer::malformed_input("bad checksum on pg_log_entry_t"); |
11fdf7f2 TL |
4841 | auto q = bl.cbegin(); |
4842 | this->decode(q); | |
7c673cae FG |
4843 | } |
4844 | ||
9f95a23c | 4845 | void pg_log_entry_t::encode(ceph::buffer::list &bl) const |
7c673cae | 4846 | { |
9f95a23c | 4847 | ENCODE_START(14, 4, bl); |
11fdf7f2 TL |
4848 | encode(op, bl); |
4849 | encode(soid, bl); | |
4850 | encode(version, bl); | |
7c673cae FG |
4851 | |
4852 | /** | |
4853 | * Added with reverting_to: | |
4854 | * Previous code used prior_version to encode | |
4855 | * what we now call reverting_to. This will | |
4856 | * allow older code to decode reverting_to | |
4857 | * into prior_version as expected. | |
4858 | */ | |
4859 | if (op == LOST_REVERT) | |
11fdf7f2 | 4860 | encode(reverting_to, bl); |
7c673cae | 4861 | else |
11fdf7f2 | 4862 | encode(prior_version, bl); |
7c673cae | 4863 | |
11fdf7f2 TL |
4864 | encode(reqid, bl); |
4865 | encode(mtime, bl); | |
7c673cae | 4866 | if (op == LOST_REVERT) |
11fdf7f2 TL |
4867 | encode(prior_version, bl); |
4868 | encode(snaps, bl); | |
4869 | encode(user_version, bl); | |
4870 | encode(mod_desc, bl); | |
4871 | encode(extra_reqids, bl); | |
7c673cae | 4872 | if (op == ERROR) |
11fdf7f2 TL |
4873 | encode(return_code, bl); |
4874 | if (!extra_reqids.empty()) | |
4875 | encode(extra_reqid_return_codes, bl); | |
9f95a23c TL |
4876 | encode(clean_regions, bl); |
4877 | if (op != ERROR) | |
4878 | encode(return_code, bl); | |
4879 | encode(op_returns, bl); | |
7c673cae FG |
4880 | ENCODE_FINISH(bl); |
4881 | } | |
4882 | ||
9f95a23c | 4883 | void pg_log_entry_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 4884 | { |
9f95a23c | 4885 | DECODE_START_LEGACY_COMPAT_LEN(14, 4, 4, bl); |
11fdf7f2 | 4886 | decode(op, bl); |
7c673cae FG |
4887 | if (struct_v < 2) { |
4888 | sobject_t old_soid; | |
11fdf7f2 | 4889 | decode(old_soid, bl); |
7c673cae FG |
4890 | soid.oid = old_soid.oid; |
4891 | soid.snap = old_soid.snap; | |
4892 | invalid_hash = true; | |
4893 | } else { | |
11fdf7f2 | 4894 | decode(soid, bl); |
7c673cae FG |
4895 | } |
4896 | if (struct_v < 3) | |
4897 | invalid_hash = true; | |
11fdf7f2 | 4898 | decode(version, bl); |
7c673cae FG |
4899 | |
4900 | if (struct_v >= 6 && op == LOST_REVERT) | |
11fdf7f2 | 4901 | decode(reverting_to, bl); |
7c673cae | 4902 | else |
11fdf7f2 | 4903 | decode(prior_version, bl); |
7c673cae | 4904 | |
11fdf7f2 | 4905 | decode(reqid, bl); |
7c673cae | 4906 | |
11fdf7f2 | 4907 | decode(mtime, bl); |
7c673cae FG |
4908 | if (struct_v < 5) |
4909 | invalid_pool = true; | |
4910 | ||
4911 | if (op == LOST_REVERT) { | |
4912 | if (struct_v >= 6) { | |
11fdf7f2 | 4913 | decode(prior_version, bl); |
7c673cae FG |
4914 | } else { |
4915 | reverting_to = prior_version; | |
4916 | } | |
4917 | } | |
4918 | if (struct_v >= 7 || // for v >= 7, this is for all ops. | |
4919 | op == CLONE) { // for v < 7, it's only present for CLONE. | |
11fdf7f2 | 4920 | decode(snaps, bl); |
9f95a23c | 4921 | // ensure snaps does not pin a larger ceph::buffer in memory |
7c673cae | 4922 | snaps.rebuild(); |
31f18b77 | 4923 | snaps.reassign_to_mempool(mempool::mempool_osd_pglog); |
7c673cae FG |
4924 | } |
4925 | ||
4926 | if (struct_v >= 8) | |
11fdf7f2 | 4927 | decode(user_version, bl); |
7c673cae FG |
4928 | else |
4929 | user_version = version.version; | |
4930 | ||
4931 | if (struct_v >= 9) | |
11fdf7f2 | 4932 | decode(mod_desc, bl); |
7c673cae FG |
4933 | else |
4934 | mod_desc.mark_unrollbackable(); | |
4935 | if (struct_v >= 10) | |
11fdf7f2 | 4936 | decode(extra_reqids, bl); |
7c673cae | 4937 | if (struct_v >= 11 && op == ERROR) |
11fdf7f2 TL |
4938 | decode(return_code, bl); |
4939 | if (struct_v >= 12 && !extra_reqids.empty()) | |
4940 | decode(extra_reqid_return_codes, bl); | |
9f95a23c TL |
4941 | if (struct_v >= 13) |
4942 | decode(clean_regions, bl); | |
4943 | else | |
4944 | clean_regions.mark_fully_dirty(); | |
4945 | if (struct_v >= 14) { | |
4946 | if (op != ERROR) { | |
4947 | decode(return_code, bl); | |
4948 | } | |
4949 | decode(op_returns, bl); | |
4950 | } | |
7c673cae FG |
4951 | DECODE_FINISH(bl); |
4952 | } | |
4953 | ||
4954 | void pg_log_entry_t::dump(Formatter *f) const | |
4955 | { | |
4956 | f->dump_string("op", get_op_name()); | |
4957 | f->dump_stream("object") << soid; | |
4958 | f->dump_stream("version") << version; | |
4959 | f->dump_stream("prior_version") << prior_version; | |
4960 | f->dump_stream("reqid") << reqid; | |
4961 | f->open_array_section("extra_reqids"); | |
11fdf7f2 | 4962 | uint32_t idx = 0; |
31f18b77 | 4963 | for (auto p = extra_reqids.begin(); |
7c673cae | 4964 | p != extra_reqids.end(); |
11fdf7f2 | 4965 | ++idx, ++p) { |
7c673cae FG |
4966 | f->open_object_section("extra_reqid"); |
4967 | f->dump_stream("reqid") << p->first; | |
4968 | f->dump_stream("user_version") << p->second; | |
11fdf7f2 TL |
4969 | auto it = extra_reqid_return_codes.find(idx); |
4970 | if (it != extra_reqid_return_codes.end()) { | |
4971 | f->dump_int("return_code", it->second); | |
4972 | } | |
7c673cae FG |
4973 | f->close_section(); |
4974 | } | |
4975 | f->close_section(); | |
4976 | f->dump_stream("mtime") << mtime; | |
4977 | f->dump_int("return_code", return_code); | |
9f95a23c TL |
4978 | if (!op_returns.empty()) { |
4979 | f->open_array_section("op_returns"); | |
4980 | for (auto& i : op_returns) { | |
4981 | f->dump_object("op", i); | |
4982 | } | |
4983 | f->close_section(); | |
4984 | } | |
7c673cae FG |
4985 | if (snaps.length() > 0) { |
4986 | vector<snapid_t> v; | |
9f95a23c | 4987 | ceph::buffer::list c = snaps; |
11fdf7f2 | 4988 | auto p = c.cbegin(); |
7c673cae | 4989 | try { |
11fdf7f2 TL |
4990 | using ceph::decode; |
4991 | decode(v, p); | |
7c673cae FG |
4992 | } catch (...) { |
4993 | v.clear(); | |
4994 | } | |
4995 | f->open_object_section("snaps"); | |
9f95a23c | 4996 | for (auto p = v.begin(); p != v.end(); ++p) |
7c673cae FG |
4997 | f->dump_unsigned("snap", *p); |
4998 | f->close_section(); | |
4999 | } | |
5000 | { | |
5001 | f->open_object_section("mod_desc"); | |
5002 | mod_desc.dump(f); | |
5003 | f->close_section(); | |
5004 | } | |
9f95a23c TL |
5005 | { |
5006 | f->open_object_section("clean_regions"); | |
5007 | clean_regions.dump(f); | |
5008 | f->close_section(); | |
5009 | } | |
7c673cae FG |
5010 | } |
5011 | ||
5012 | void pg_log_entry_t::generate_test_instances(list<pg_log_entry_t*>& o) | |
5013 | { | |
5014 | o.push_back(new pg_log_entry_t()); | |
5015 | hobject_t oid(object_t("objname"), "key", 123, 456, 0, ""); | |
5016 | o.push_back(new pg_log_entry_t(MODIFY, oid, eversion_t(1,2), eversion_t(3,4), | |
5017 | 1, osd_reqid_t(entity_name_t::CLIENT(777), 8, 999), | |
5018 | utime_t(8,9), 0)); | |
5019 | o.push_back(new pg_log_entry_t(ERROR, oid, eversion_t(1,2), eversion_t(3,4), | |
5020 | 1, osd_reqid_t(entity_name_t::CLIENT(777), 8, 999), | |
5021 | utime_t(8,9), -ENOENT)); | |
5022 | } | |
5023 | ||
5024 | ostream& operator<<(ostream& out, const pg_log_entry_t& e) | |
5025 | { | |
5026 | out << e.version << " (" << e.prior_version << ") " | |
5027 | << std::left << std::setw(8) << e.get_op_name() << ' ' | |
5028 | << e.soid << " by " << e.reqid << " " << e.mtime | |
5029 | << " " << e.return_code; | |
9f95a23c TL |
5030 | if (!e.op_returns.empty()) { |
5031 | out << " " << e.op_returns; | |
5032 | } | |
7c673cae FG |
5033 | if (e.snaps.length()) { |
5034 | vector<snapid_t> snaps; | |
9f95a23c | 5035 | ceph::buffer::list c = e.snaps; |
11fdf7f2 | 5036 | auto p = c.cbegin(); |
7c673cae | 5037 | try { |
11fdf7f2 | 5038 | decode(snaps, p); |
7c673cae FG |
5039 | } catch (...) { |
5040 | snaps.clear(); | |
5041 | } | |
5042 | out << " snaps " << snaps; | |
5043 | } | |
9f95a23c | 5044 | out << " ObjectCleanRegions " << e.clean_regions; |
7c673cae FG |
5045 | return out; |
5046 | } | |
5047 | ||
c07f9fc5 FG |
5048 | // -- pg_log_dup_t -- |
5049 | ||
11fdf7f2 | 5050 | std::string pg_log_dup_t::get_key_name() const |
c07f9fc5 | 5051 | { |
11fdf7f2 TL |
5052 | static const char prefix[] = "dup_"; |
5053 | std::string key(36, ' '); | |
5054 | memcpy(&key[0], prefix, 4); | |
5055 | version.get_key_name(&key[4]); | |
5056 | key.resize(35); // remove the null terminator | |
5057 | return key; | |
c07f9fc5 FG |
5058 | } |
5059 | ||
9f95a23c | 5060 | void pg_log_dup_t::encode(ceph::buffer::list &bl) const |
c07f9fc5 | 5061 | { |
9f95a23c | 5062 | ENCODE_START(2, 1, bl); |
11fdf7f2 TL |
5063 | encode(reqid, bl); |
5064 | encode(version, bl); | |
5065 | encode(user_version, bl); | |
5066 | encode(return_code, bl); | |
9f95a23c | 5067 | encode(op_returns, bl); |
c07f9fc5 FG |
5068 | ENCODE_FINISH(bl); |
5069 | } | |
5070 | ||
9f95a23c | 5071 | void pg_log_dup_t::decode(ceph::buffer::list::const_iterator &bl) |
c07f9fc5 | 5072 | { |
9f95a23c | 5073 | DECODE_START(2, bl); |
11fdf7f2 TL |
5074 | decode(reqid, bl); |
5075 | decode(version, bl); | |
5076 | decode(user_version, bl); | |
5077 | decode(return_code, bl); | |
9f95a23c TL |
5078 | if (struct_v >= 2) { |
5079 | decode(op_returns, bl); | |
5080 | } | |
c07f9fc5 FG |
5081 | DECODE_FINISH(bl); |
5082 | } | |
5083 | ||
5084 | void pg_log_dup_t::dump(Formatter *f) const | |
5085 | { | |
5086 | f->dump_stream("reqid") << reqid; | |
5087 | f->dump_stream("version") << version; | |
5088 | f->dump_stream("user_version") << user_version; | |
5089 | f->dump_stream("return_code") << return_code; | |
9f95a23c TL |
5090 | if (!op_returns.empty()) { |
5091 | f->open_array_section("op_returns"); | |
5092 | for (auto& i : op_returns) { | |
5093 | f->dump_object("op", i); | |
5094 | } | |
5095 | f->close_section(); | |
5096 | } | |
c07f9fc5 FG |
5097 | } |
5098 | ||
5099 | void pg_log_dup_t::generate_test_instances(list<pg_log_dup_t*>& o) | |
5100 | { | |
5101 | o.push_back(new pg_log_dup_t()); | |
5102 | o.push_back(new pg_log_dup_t(eversion_t(1,2), | |
5103 | 1, | |
5104 | osd_reqid_t(entity_name_t::CLIENT(777), 8, 999), | |
5105 | 0)); | |
5106 | o.push_back(new pg_log_dup_t(eversion_t(1,2), | |
5107 | 2, | |
5108 | osd_reqid_t(entity_name_t::CLIENT(777), 8, 999), | |
5109 | -ENOENT)); | |
5110 | } | |
5111 | ||
5112 | ||
5113 | std::ostream& operator<<(std::ostream& out, const pg_log_dup_t& e) { | |
9f95a23c | 5114 | out << "log_dup(reqid=" << e.reqid << |
c07f9fc5 | 5115 | " v=" << e.version << " uv=" << e.user_version << |
9f95a23c TL |
5116 | " rc=" << e.return_code; |
5117 | if (!e.op_returns.empty()) { | |
5118 | out << " " << e.op_returns; | |
5119 | } | |
5120 | return out << ")"; | |
c07f9fc5 FG |
5121 | } |
5122 | ||
7c673cae FG |
5123 | |
5124 | // -- pg_log_t -- | |
5125 | ||
5126 | // out: pg_log_t that only has entries that apply to import_pgid using curmap | |
5127 | // reject: Entries rejected from "in" are in the reject.log. Other fields not set. | |
5128 | void pg_log_t::filter_log(spg_t import_pgid, const OSDMap &curmap, | |
5129 | const string &hit_set_namespace, const pg_log_t &in, | |
5130 | pg_log_t &out, pg_log_t &reject) | |
5131 | { | |
5132 | out = in; | |
5133 | out.log.clear(); | |
5134 | reject.log.clear(); | |
5135 | ||
9f95a23c | 5136 | for (auto i = in.log.cbegin(); i != in.log.cend(); ++i) { |
7c673cae FG |
5137 | |
5138 | // Reject pg log entries for temporary objects | |
5139 | if (i->soid.is_temp()) { | |
5140 | reject.log.push_back(*i); | |
5141 | continue; | |
5142 | } | |
5143 | ||
5144 | if (i->soid.nspace != hit_set_namespace) { | |
5145 | object_t oid = i->soid.oid; | |
5146 | object_locator_t loc(i->soid); | |
5147 | pg_t raw_pgid = curmap.object_locator_to_pg(oid, loc); | |
5148 | pg_t pgid = curmap.raw_pg_to_pg(raw_pgid); | |
5149 | ||
5150 | if (import_pgid.pgid == pgid) { | |
5151 | out.log.push_back(*i); | |
5152 | } else { | |
5153 | reject.log.push_back(*i); | |
5154 | } | |
5155 | } else { | |
5156 | out.log.push_back(*i); | |
5157 | } | |
5158 | } | |
5159 | } | |
5160 | ||
9f95a23c | 5161 | void pg_log_t::encode(ceph::buffer::list& bl) const |
7c673cae | 5162 | { |
c07f9fc5 | 5163 | ENCODE_START(7, 3, bl); |
11fdf7f2 TL |
5164 | encode(head, bl); |
5165 | encode(tail, bl); | |
5166 | encode(log, bl); | |
5167 | encode(can_rollback_to, bl); | |
5168 | encode(rollback_info_trimmed_to, bl); | |
5169 | encode(dups, bl); | |
7c673cae FG |
5170 | ENCODE_FINISH(bl); |
5171 | } | |
5172 | ||
9f95a23c | 5173 | void pg_log_t::decode(ceph::buffer::list::const_iterator &bl, int64_t pool) |
7c673cae | 5174 | { |
c07f9fc5 | 5175 | DECODE_START_LEGACY_COMPAT_LEN(7, 3, 3, bl); |
11fdf7f2 TL |
5176 | decode(head, bl); |
5177 | decode(tail, bl); | |
7c673cae FG |
5178 | if (struct_v < 2) { |
5179 | bool backlog; | |
11fdf7f2 | 5180 | decode(backlog, bl); |
7c673cae | 5181 | } |
11fdf7f2 | 5182 | decode(log, bl); |
7c673cae | 5183 | if (struct_v >= 5) |
11fdf7f2 | 5184 | decode(can_rollback_to, bl); |
7c673cae FG |
5185 | |
5186 | if (struct_v >= 6) | |
11fdf7f2 | 5187 | decode(rollback_info_trimmed_to, bl); |
7c673cae FG |
5188 | else |
5189 | rollback_info_trimmed_to = tail; | |
c07f9fc5 FG |
5190 | |
5191 | if (struct_v >= 7) | |
11fdf7f2 | 5192 | decode(dups, bl); |
c07f9fc5 | 5193 | |
7c673cae FG |
5194 | DECODE_FINISH(bl); |
5195 | ||
5196 | // handle hobject_t format change | |
5197 | if (struct_v < 4) { | |
9f95a23c | 5198 | for (auto i = log.begin(); i != log.end(); ++i) { |
7c673cae FG |
5199 | if (!i->soid.is_max() && i->soid.pool == -1) |
5200 | i->soid.pool = pool; | |
5201 | } | |
5202 | } | |
5203 | } | |
5204 | ||
5205 | void pg_log_t::dump(Formatter *f) const | |
5206 | { | |
5207 | f->dump_stream("head") << head; | |
5208 | f->dump_stream("tail") << tail; | |
5209 | f->open_array_section("log"); | |
9f95a23c | 5210 | for (auto p = log.cbegin(); p != log.cend(); ++p) { |
7c673cae FG |
5211 | f->open_object_section("entry"); |
5212 | p->dump(f); | |
5213 | f->close_section(); | |
5214 | } | |
5215 | f->close_section(); | |
c07f9fc5 FG |
5216 | f->open_array_section("dups"); |
5217 | for (const auto& entry : dups) { | |
5218 | f->open_object_section("entry"); | |
5219 | entry.dump(f); | |
5220 | f->close_section(); | |
5221 | } | |
5222 | f->close_section(); | |
7c673cae FG |
5223 | } |
5224 | ||
5225 | void pg_log_t::generate_test_instances(list<pg_log_t*>& o) | |
5226 | { | |
5227 | o.push_back(new pg_log_t); | |
5228 | ||
5229 | // this is nonsensical: | |
5230 | o.push_back(new pg_log_t); | |
5231 | o.back()->head = eversion_t(1,2); | |
5232 | o.back()->tail = eversion_t(3,4); | |
5233 | list<pg_log_entry_t*> e; | |
5234 | pg_log_entry_t::generate_test_instances(e); | |
9f95a23c | 5235 | for (auto p = e.begin(); p != e.end(); ++p) |
7c673cae FG |
5236 | o.back()->log.push_back(**p); |
5237 | } | |
5238 | ||
81eedcae TL |
5239 | static void _handle_dups(CephContext* cct, pg_log_t &target, const pg_log_t &other, unsigned maxdups) |
5240 | { | |
5241 | auto earliest_dup_version = | |
5242 | target.head.version < maxdups ? 0u : target.head.version - maxdups + 1; | |
2a845540 TL |
5243 | lgeneric_subdout(cct, osd, 20) << __func__ << " earliest_dup_version " |
5244 | << earliest_dup_version << dendl; | |
81eedcae TL |
5245 | |
5246 | for (auto d = other.dups.cbegin(); d != other.dups.cend(); ++d) { | |
5247 | if (d->version.version >= earliest_dup_version) { | |
5248 | lgeneric_subdout(cct, osd, 20) | |
5249 | << "copy_up_to/copy_after copy dup version " | |
5250 | << d->version << dendl; | |
5251 | target.dups.push_back(pg_log_dup_t(*d)); | |
5252 | } | |
5253 | } | |
5254 | ||
5255 | for (auto i = other.log.cbegin(); i != other.log.cend(); ++i) { | |
5256 | ceph_assert(i->version > other.tail); | |
5257 | if (i->version > target.tail) | |
5258 | break; | |
5259 | if (i->version.version >= earliest_dup_version) { | |
5260 | lgeneric_subdout(cct, osd, 20) | |
5261 | << "copy_up_to/copy_after copy dup from log version " | |
5262 | << i->version << dendl; | |
5263 | target.dups.push_back(pg_log_dup_t(*i)); | |
5264 | } | |
5265 | } | |
5266 | } | |
5267 | ||
5268 | ||
5269 | void pg_log_t::copy_after(CephContext* cct, const pg_log_t &other, eversion_t v) | |
7c673cae FG |
5270 | { |
5271 | can_rollback_to = other.can_rollback_to; | |
5272 | head = other.head; | |
5273 | tail = other.tail; | |
2a845540 TL |
5274 | lgeneric_subdout(cct, osd, 20) << __func__ << " v " << v |
5275 | << " dups.size()=" << dups.size() | |
5276 | << " other.dups.size()=" << other.dups.size() << dendl; | |
9f95a23c | 5277 | for (auto i = other.log.crbegin(); i != other.log.crend(); ++i) { |
11fdf7f2 | 5278 | ceph_assert(i->version > other.tail); |
7c673cae FG |
5279 | if (i->version <= v) { |
5280 | // make tail accurate. | |
5281 | tail = i->version; | |
5282 | break; | |
5283 | } | |
81eedcae | 5284 | lgeneric_subdout(cct, osd, 20) << __func__ << " copy log version " << i->version << dendl; |
7c673cae FG |
5285 | log.push_front(*i); |
5286 | } | |
81eedcae | 5287 | _handle_dups(cct, *this, other, cct->_conf->osd_pg_log_dups_tracked); |
2a845540 TL |
5288 | lgeneric_subdout(cct, osd, 20) << __func__ << " END v " << v |
5289 | << " dups.size()=" << dups.size() | |
5290 | << " other.dups.size()=" << other.dups.size() << dendl; | |
7c673cae FG |
5291 | } |
5292 | ||
81eedcae | 5293 | void pg_log_t::copy_up_to(CephContext* cct, const pg_log_t &other, int max) |
7c673cae FG |
5294 | { |
5295 | can_rollback_to = other.can_rollback_to; | |
5296 | int n = 0; | |
5297 | head = other.head; | |
5298 | tail = other.tail; | |
2a845540 TL |
5299 | lgeneric_subdout(cct, osd, 20) << __func__ << " max " << max |
5300 | << " dups.size()=" << dups.size() | |
5301 | << " other.dups.size()=" << other.dups.size() << dendl; | |
9f95a23c | 5302 | for (auto i = other.log.crbegin(); i != other.log.crend(); ++i) { |
81eedcae | 5303 | ceph_assert(i->version > other.tail); |
7c673cae FG |
5304 | if (n++ >= max) { |
5305 | tail = i->version; | |
5306 | break; | |
5307 | } | |
81eedcae | 5308 | lgeneric_subdout(cct, osd, 20) << __func__ << " copy log version " << i->version << dendl; |
7c673cae FG |
5309 | log.push_front(*i); |
5310 | } | |
81eedcae | 5311 | _handle_dups(cct, *this, other, cct->_conf->osd_pg_log_dups_tracked); |
2a845540 TL |
5312 | lgeneric_subdout(cct, osd, 20) << __func__ << " END max " << max |
5313 | << " dups.size()=" << dups.size() | |
5314 | << " other.dups.size()=" << other.dups.size() << dendl; | |
7c673cae FG |
5315 | } |
5316 | ||
c07f9fc5 | 5317 | ostream& pg_log_t::print(ostream& out) const |
7c673cae FG |
5318 | { |
5319 | out << *this << std::endl; | |
9f95a23c | 5320 | for (auto p = log.cbegin(); p != log.cend(); ++p) |
7c673cae | 5321 | out << *p << std::endl; |
c07f9fc5 FG |
5322 | for (const auto& entry : dups) { |
5323 | out << " dup entry: " << entry << std::endl; | |
5324 | } | |
7c673cae FG |
5325 | return out; |
5326 | } | |
5327 | ||
5328 | // -- pg_missing_t -- | |
5329 | ||
5330 | ostream& operator<<(ostream& out, const pg_missing_item& i) | |
5331 | { | |
5332 | out << i.need; | |
5333 | if (i.have != eversion_t()) | |
5334 | out << "(" << i.have << ")"; | |
9f95a23c TL |
5335 | out << " flags = " << i.flag_str() |
5336 | << " " << i.clean_regions; | |
7c673cae FG |
5337 | return out; |
5338 | } | |
5339 | ||
5340 | // -- object_copy_cursor_t -- | |
5341 | ||
9f95a23c | 5342 | void object_copy_cursor_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
5343 | { |
5344 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
5345 | encode(attr_complete, bl); |
5346 | encode(data_offset, bl); | |
5347 | encode(data_complete, bl); | |
5348 | encode(omap_offset, bl); | |
5349 | encode(omap_complete, bl); | |
7c673cae FG |
5350 | ENCODE_FINISH(bl); |
5351 | } | |
5352 | ||
9f95a23c | 5353 | void object_copy_cursor_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
5354 | { |
5355 | DECODE_START(1, bl); | |
11fdf7f2 TL |
5356 | decode(attr_complete, bl); |
5357 | decode(data_offset, bl); | |
5358 | decode(data_complete, bl); | |
5359 | decode(omap_offset, bl); | |
5360 | decode(omap_complete, bl); | |
7c673cae FG |
5361 | DECODE_FINISH(bl); |
5362 | } | |
5363 | ||
5364 | void object_copy_cursor_t::dump(Formatter *f) const | |
5365 | { | |
5366 | f->dump_unsigned("attr_complete", (int)attr_complete); | |
5367 | f->dump_unsigned("data_offset", data_offset); | |
5368 | f->dump_unsigned("data_complete", (int)data_complete); | |
5369 | f->dump_string("omap_offset", omap_offset); | |
5370 | f->dump_unsigned("omap_complete", (int)omap_complete); | |
5371 | } | |
5372 | ||
5373 | void object_copy_cursor_t::generate_test_instances(list<object_copy_cursor_t*>& o) | |
5374 | { | |
5375 | o.push_back(new object_copy_cursor_t); | |
5376 | o.push_back(new object_copy_cursor_t); | |
5377 | o.back()->attr_complete = true; | |
5378 | o.back()->data_offset = 123; | |
5379 | o.push_back(new object_copy_cursor_t); | |
5380 | o.back()->attr_complete = true; | |
5381 | o.back()->data_complete = true; | |
5382 | o.back()->omap_offset = "foo"; | |
5383 | o.push_back(new object_copy_cursor_t); | |
5384 | o.back()->attr_complete = true; | |
5385 | o.back()->data_complete = true; | |
5386 | o.back()->omap_complete = true; | |
5387 | } | |
5388 | ||
5389 | // -- object_copy_data_t -- | |
5390 | ||
9f95a23c | 5391 | void object_copy_data_t::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae | 5392 | { |
11fdf7f2 TL |
5393 | ENCODE_START(8, 5, bl); |
5394 | encode(size, bl); | |
5395 | encode(mtime, bl); | |
5396 | encode(attrs, bl); | |
5397 | encode(data, bl); | |
5398 | encode(omap_data, bl); | |
5399 | encode(cursor, bl); | |
5400 | encode(omap_header, bl); | |
5401 | encode(snaps, bl); | |
5402 | encode(snap_seq, bl); | |
5403 | encode(flags, bl); | |
5404 | encode(data_digest, bl); | |
5405 | encode(omap_digest, bl); | |
5406 | encode(reqids, bl); | |
5407 | encode(truncate_seq, bl); | |
5408 | encode(truncate_size, bl); | |
5409 | encode(reqid_return_codes, bl); | |
7c673cae FG |
5410 | ENCODE_FINISH(bl); |
5411 | } | |
5412 | ||
9f95a23c | 5413 | void object_copy_data_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 5414 | { |
9f95a23c | 5415 | DECODE_START(8, bl); |
7c673cae FG |
5416 | if (struct_v < 5) { |
5417 | // old | |
11fdf7f2 TL |
5418 | decode(size, bl); |
5419 | decode(mtime, bl); | |
7c673cae FG |
5420 | { |
5421 | string category; | |
11fdf7f2 | 5422 | decode(category, bl); // no longer used |
7c673cae | 5423 | } |
11fdf7f2 TL |
5424 | decode(attrs, bl); |
5425 | decode(data, bl); | |
7c673cae | 5426 | { |
9f95a23c | 5427 | map<string,ceph::buffer::list> omap; |
11fdf7f2 | 5428 | decode(omap, bl); |
7c673cae | 5429 | omap_data.clear(); |
11fdf7f2 TL |
5430 | if (!omap.empty()) { |
5431 | using ceph::encode; | |
5432 | encode(omap, omap_data); | |
5433 | } | |
7c673cae | 5434 | } |
11fdf7f2 | 5435 | decode(cursor, bl); |
7c673cae | 5436 | if (struct_v >= 2) |
11fdf7f2 | 5437 | decode(omap_header, bl); |
7c673cae | 5438 | if (struct_v >= 3) { |
11fdf7f2 TL |
5439 | decode(snaps, bl); |
5440 | decode(snap_seq, bl); | |
7c673cae FG |
5441 | } else { |
5442 | snaps.clear(); | |
5443 | snap_seq = 0; | |
5444 | } | |
5445 | if (struct_v >= 4) { | |
11fdf7f2 TL |
5446 | decode(flags, bl); |
5447 | decode(data_digest, bl); | |
5448 | decode(omap_digest, bl); | |
7c673cae FG |
5449 | } |
5450 | } else { | |
5451 | // current | |
11fdf7f2 TL |
5452 | decode(size, bl); |
5453 | decode(mtime, bl); | |
5454 | decode(attrs, bl); | |
5455 | decode(data, bl); | |
5456 | decode(omap_data, bl); | |
5457 | decode(cursor, bl); | |
5458 | decode(omap_header, bl); | |
5459 | decode(snaps, bl); | |
5460 | decode(snap_seq, bl); | |
7c673cae | 5461 | if (struct_v >= 4) { |
11fdf7f2 TL |
5462 | decode(flags, bl); |
5463 | decode(data_digest, bl); | |
5464 | decode(omap_digest, bl); | |
7c673cae FG |
5465 | } |
5466 | if (struct_v >= 6) { | |
11fdf7f2 | 5467 | decode(reqids, bl); |
7c673cae FG |
5468 | } |
5469 | if (struct_v >= 7) { | |
11fdf7f2 TL |
5470 | decode(truncate_seq, bl); |
5471 | decode(truncate_size, bl); | |
5472 | } | |
5473 | if (struct_v >= 8) { | |
5474 | decode(reqid_return_codes, bl); | |
7c673cae FG |
5475 | } |
5476 | } | |
5477 | DECODE_FINISH(bl); | |
5478 | } | |
5479 | ||
5480 | void object_copy_data_t::generate_test_instances(list<object_copy_data_t*>& o) | |
5481 | { | |
5482 | o.push_back(new object_copy_data_t()); | |
5483 | ||
5484 | list<object_copy_cursor_t*> cursors; | |
5485 | object_copy_cursor_t::generate_test_instances(cursors); | |
9f95a23c | 5486 | auto ci = cursors.begin(); |
7c673cae FG |
5487 | o.back()->cursor = **(ci++); |
5488 | ||
5489 | o.push_back(new object_copy_data_t()); | |
5490 | o.back()->cursor = **(ci++); | |
5491 | ||
5492 | o.push_back(new object_copy_data_t()); | |
5493 | o.back()->size = 1234; | |
5494 | o.back()->mtime.set_from_double(1234); | |
9f95a23c TL |
5495 | ceph::buffer::ptr bp("there", 5); |
5496 | ceph::buffer::list bl; | |
7c673cae FG |
5497 | bl.push_back(bp); |
5498 | o.back()->attrs["hello"] = bl; | |
9f95a23c TL |
5499 | ceph::buffer::ptr bp2("not", 3); |
5500 | ceph::buffer::list bl2; | |
7c673cae | 5501 | bl2.push_back(bp2); |
9f95a23c | 5502 | map<string,ceph::buffer::list> omap; |
7c673cae | 5503 | omap["why"] = bl2; |
11fdf7f2 TL |
5504 | using ceph::encode; |
5505 | encode(omap, o.back()->omap_data); | |
9f95a23c | 5506 | ceph::buffer::ptr databp("iamsomedatatocontain", 20); |
7c673cae FG |
5507 | o.back()->data.push_back(databp); |
5508 | o.back()->omap_header.append("this is an omap header"); | |
5509 | o.back()->snaps.push_back(123); | |
5510 | o.back()->reqids.push_back(make_pair(osd_reqid_t(), version_t())); | |
5511 | } | |
5512 | ||
5513 | void object_copy_data_t::dump(Formatter *f) const | |
5514 | { | |
5515 | f->open_object_section("cursor"); | |
5516 | cursor.dump(f); | |
5517 | f->close_section(); // cursor | |
5518 | f->dump_int("size", size); | |
5519 | f->dump_stream("mtime") << mtime; | |
9f95a23c | 5520 | /* we should really print out the attrs here, but ceph::buffer::list |
7c673cae FG |
5521 | const-correctness prevents that */ |
5522 | f->dump_int("attrs_size", attrs.size()); | |
5523 | f->dump_int("flags", flags); | |
5524 | f->dump_unsigned("data_digest", data_digest); | |
5525 | f->dump_unsigned("omap_digest", omap_digest); | |
5526 | f->dump_int("omap_data_length", omap_data.length()); | |
5527 | f->dump_int("omap_header_length", omap_header.length()); | |
5528 | f->dump_int("data_length", data.length()); | |
5529 | f->open_array_section("snaps"); | |
9f95a23c | 5530 | for (auto p = snaps.cbegin(); p != snaps.cend(); ++p) |
7c673cae FG |
5531 | f->dump_unsigned("snap", *p); |
5532 | f->close_section(); | |
5533 | f->open_array_section("reqids"); | |
11fdf7f2 | 5534 | uint32_t idx = 0; |
31f18b77 | 5535 | for (auto p = reqids.begin(); |
7c673cae | 5536 | p != reqids.end(); |
11fdf7f2 | 5537 | ++idx, ++p) { |
7c673cae FG |
5538 | f->open_object_section("extra_reqid"); |
5539 | f->dump_stream("reqid") << p->first; | |
5540 | f->dump_stream("user_version") << p->second; | |
11fdf7f2 TL |
5541 | auto it = reqid_return_codes.find(idx); |
5542 | if (it != reqid_return_codes.end()) { | |
5543 | f->dump_int("return_code", it->second); | |
5544 | } | |
7c673cae FG |
5545 | f->close_section(); |
5546 | } | |
5547 | f->close_section(); | |
5548 | } | |
5549 | ||
5550 | // -- pg_create_t -- | |
5551 | ||
9f95a23c | 5552 | void pg_create_t::encode(ceph::buffer::list &bl) const |
7c673cae FG |
5553 | { |
5554 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
5555 | encode(created, bl); |
5556 | encode(parent, bl); | |
5557 | encode(split_bits, bl); | |
7c673cae FG |
5558 | ENCODE_FINISH(bl); |
5559 | } | |
5560 | ||
9f95a23c | 5561 | void pg_create_t::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
5562 | { |
5563 | DECODE_START(1, bl); | |
11fdf7f2 TL |
5564 | decode(created, bl); |
5565 | decode(parent, bl); | |
5566 | decode(split_bits, bl); | |
7c673cae FG |
5567 | DECODE_FINISH(bl); |
5568 | } | |
5569 | ||
5570 | void pg_create_t::dump(Formatter *f) const | |
5571 | { | |
5572 | f->dump_unsigned("created", created); | |
5573 | f->dump_stream("parent") << parent; | |
5574 | f->dump_int("split_bits", split_bits); | |
5575 | } | |
5576 | ||
5577 | void pg_create_t::generate_test_instances(list<pg_create_t*>& o) | |
5578 | { | |
5579 | o.push_back(new pg_create_t); | |
11fdf7f2 | 5580 | o.push_back(new pg_create_t(1, pg_t(3, 4), 2)); |
7c673cae FG |
5581 | } |
5582 | ||
5583 | ||
5584 | // -- pg_hit_set_info_t -- | |
5585 | ||
9f95a23c | 5586 | void pg_hit_set_info_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
5587 | { |
5588 | ENCODE_START(2, 1, bl); | |
11fdf7f2 TL |
5589 | encode(begin, bl); |
5590 | encode(end, bl); | |
5591 | encode(version, bl); | |
5592 | encode(using_gmt, bl); | |
7c673cae FG |
5593 | ENCODE_FINISH(bl); |
5594 | } | |
5595 | ||
9f95a23c | 5596 | void pg_hit_set_info_t::decode(ceph::buffer::list::const_iterator& p) |
7c673cae FG |
5597 | { |
5598 | DECODE_START(2, p); | |
11fdf7f2 TL |
5599 | decode(begin, p); |
5600 | decode(end, p); | |
5601 | decode(version, p); | |
7c673cae | 5602 | if (struct_v >= 2) { |
11fdf7f2 | 5603 | decode(using_gmt, p); |
7c673cae FG |
5604 | } else { |
5605 | using_gmt = false; | |
5606 | } | |
5607 | DECODE_FINISH(p); | |
5608 | } | |
5609 | ||
5610 | void pg_hit_set_info_t::dump(Formatter *f) const | |
5611 | { | |
5612 | f->dump_stream("begin") << begin; | |
5613 | f->dump_stream("end") << end; | |
5614 | f->dump_stream("version") << version; | |
5615 | f->dump_stream("using_gmt") << using_gmt; | |
5616 | } | |
5617 | ||
5618 | void pg_hit_set_info_t::generate_test_instances(list<pg_hit_set_info_t*>& ls) | |
5619 | { | |
5620 | ls.push_back(new pg_hit_set_info_t); | |
5621 | ls.push_back(new pg_hit_set_info_t); | |
5622 | ls.back()->begin = utime_t(1, 2); | |
5623 | ls.back()->end = utime_t(3, 4); | |
5624 | } | |
5625 | ||
5626 | ||
5627 | // -- pg_hit_set_history_t -- | |
5628 | ||
9f95a23c | 5629 | void pg_hit_set_history_t::encode(ceph::buffer::list& bl) const |
7c673cae FG |
5630 | { |
5631 | ENCODE_START(1, 1, bl); | |
11fdf7f2 | 5632 | encode(current_last_update, bl); |
7c673cae FG |
5633 | { |
5634 | utime_t dummy_stamp; | |
11fdf7f2 | 5635 | encode(dummy_stamp, bl); |
7c673cae FG |
5636 | } |
5637 | { | |
5638 | pg_hit_set_info_t dummy_info; | |
11fdf7f2 | 5639 | encode(dummy_info, bl); |
7c673cae | 5640 | } |
11fdf7f2 | 5641 | encode(history, bl); |
7c673cae FG |
5642 | ENCODE_FINISH(bl); |
5643 | } | |
5644 | ||
9f95a23c | 5645 | void pg_hit_set_history_t::decode(ceph::buffer::list::const_iterator& p) |
7c673cae FG |
5646 | { |
5647 | DECODE_START(1, p); | |
11fdf7f2 | 5648 | decode(current_last_update, p); |
7c673cae FG |
5649 | { |
5650 | utime_t dummy_stamp; | |
11fdf7f2 | 5651 | decode(dummy_stamp, p); |
7c673cae FG |
5652 | } |
5653 | { | |
5654 | pg_hit_set_info_t dummy_info; | |
11fdf7f2 | 5655 | decode(dummy_info, p); |
7c673cae | 5656 | } |
11fdf7f2 | 5657 | decode(history, p); |
7c673cae FG |
5658 | DECODE_FINISH(p); |
5659 | } | |
5660 | ||
5661 | void pg_hit_set_history_t::dump(Formatter *f) const | |
5662 | { | |
5663 | f->dump_stream("current_last_update") << current_last_update; | |
5664 | f->open_array_section("history"); | |
9f95a23c | 5665 | for (auto p = history.cbegin(); p != history.cend(); ++p) { |
7c673cae FG |
5666 | f->open_object_section("info"); |
5667 | p->dump(f); | |
5668 | f->close_section(); | |
5669 | } | |
5670 | f->close_section(); | |
5671 | } | |
5672 | ||
5673 | void pg_hit_set_history_t::generate_test_instances(list<pg_hit_set_history_t*>& ls) | |
5674 | { | |
5675 | ls.push_back(new pg_hit_set_history_t); | |
5676 | ls.push_back(new pg_hit_set_history_t); | |
5677 | ls.back()->current_last_update = eversion_t(1, 2); | |
5678 | ls.back()->history.push_back(pg_hit_set_info_t()); | |
5679 | } | |
5680 | ||
7c673cae FG |
5681 | // -- OSDSuperblock -- |
5682 | ||
9f95a23c | 5683 | void OSDSuperblock::encode(ceph::buffer::list &bl) const |
7c673cae | 5684 | { |
1e59de90 | 5685 | ENCODE_START(10, 5, bl); |
11fdf7f2 TL |
5686 | encode(cluster_fsid, bl); |
5687 | encode(whoami, bl); | |
5688 | encode(current_epoch, bl); | |
5689 | encode(oldest_map, bl); | |
5690 | encode(newest_map, bl); | |
5691 | encode(weight, bl); | |
7c673cae | 5692 | compat_features.encode(bl); |
11fdf7f2 TL |
5693 | encode(clean_thru, bl); |
5694 | encode(mounted, bl); | |
5695 | encode(osd_fsid, bl); | |
5696 | encode((epoch_t)0, bl); // epoch_t last_epoch_marked_full | |
5697 | encode((uint32_t)0, bl); // map<int64_t,epoch_t> pool_last_epoch_marked_full | |
9f95a23c TL |
5698 | encode(purged_snaps_last, bl); |
5699 | encode(last_purged_snaps_scrub, bl); | |
1e59de90 | 5700 | encode(cluster_osdmap_trim_lower_bound, bl); |
7c673cae FG |
5701 | ENCODE_FINISH(bl); |
5702 | } | |
5703 | ||
9f95a23c | 5704 | void OSDSuperblock::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae | 5705 | { |
1e59de90 | 5706 | DECODE_START_LEGACY_COMPAT_LEN(10, 5, 5, bl); |
7c673cae FG |
5707 | if (struct_v < 3) { |
5708 | string magic; | |
11fdf7f2 TL |
5709 | decode(magic, bl); |
5710 | } | |
5711 | decode(cluster_fsid, bl); | |
5712 | decode(whoami, bl); | |
5713 | decode(current_epoch, bl); | |
5714 | decode(oldest_map, bl); | |
5715 | decode(newest_map, bl); | |
5716 | decode(weight, bl); | |
7c673cae FG |
5717 | if (struct_v >= 2) { |
5718 | compat_features.decode(bl); | |
5719 | } else { //upgrade it! | |
5720 | compat_features.incompat.insert(CEPH_OSD_FEATURE_INCOMPAT_BASE); | |
5721 | } | |
11fdf7f2 TL |
5722 | decode(clean_thru, bl); |
5723 | decode(mounted, bl); | |
7c673cae | 5724 | if (struct_v >= 4) |
11fdf7f2 | 5725 | decode(osd_fsid, bl); |
7c673cae FG |
5726 | if (struct_v >= 6) { |
5727 | epoch_t last_map_marked_full; | |
11fdf7f2 | 5728 | decode(last_map_marked_full, bl); |
7c673cae FG |
5729 | } |
5730 | if (struct_v >= 7) { | |
5731 | map<int64_t,epoch_t> pool_last_map_marked_full; | |
11fdf7f2 | 5732 | decode(pool_last_map_marked_full, bl); |
7c673cae | 5733 | } |
9f95a23c TL |
5734 | if (struct_v >= 9) { |
5735 | decode(purged_snaps_last, bl); | |
5736 | decode(last_purged_snaps_scrub, bl); | |
5737 | } else { | |
5738 | purged_snaps_last = 0; | |
5739 | } | |
1e59de90 TL |
5740 | if (struct_v >= 10) { |
5741 | decode(cluster_osdmap_trim_lower_bound, bl); | |
5742 | } else { | |
5743 | cluster_osdmap_trim_lower_bound = 0; | |
5744 | } | |
7c673cae FG |
5745 | DECODE_FINISH(bl); |
5746 | } | |
5747 | ||
5748 | void OSDSuperblock::dump(Formatter *f) const | |
5749 | { | |
5750 | f->dump_stream("cluster_fsid") << cluster_fsid; | |
5751 | f->dump_stream("osd_fsid") << osd_fsid; | |
5752 | f->dump_int("whoami", whoami); | |
5753 | f->dump_int("current_epoch", current_epoch); | |
5754 | f->dump_int("oldest_map", oldest_map); | |
5755 | f->dump_int("newest_map", newest_map); | |
5756 | f->dump_float("weight", weight); | |
5757 | f->open_object_section("compat"); | |
5758 | compat_features.dump(f); | |
5759 | f->close_section(); | |
5760 | f->dump_int("clean_thru", clean_thru); | |
5761 | f->dump_int("last_epoch_mounted", mounted); | |
9f95a23c TL |
5762 | f->dump_unsigned("purged_snaps_last", purged_snaps_last); |
5763 | f->dump_stream("last_purged_snaps_scrub") << last_purged_snaps_scrub; | |
1e59de90 TL |
5764 | f->dump_int("cluster_osdmap_trim_lower_bound", |
5765 | cluster_osdmap_trim_lower_bound); | |
7c673cae FG |
5766 | } |
5767 | ||
5768 | void OSDSuperblock::generate_test_instances(list<OSDSuperblock*>& o) | |
5769 | { | |
5770 | OSDSuperblock z; | |
5771 | o.push_back(new OSDSuperblock(z)); | |
11fdf7f2 TL |
5772 | z.cluster_fsid.parse("01010101-0101-0101-0101-010101010101"); |
5773 | z.osd_fsid.parse("02020202-0202-0202-0202-020202020202"); | |
7c673cae FG |
5774 | z.whoami = 3; |
5775 | z.current_epoch = 4; | |
5776 | z.oldest_map = 5; | |
5777 | z.newest_map = 9; | |
5778 | z.mounted = 8; | |
5779 | z.clean_thru = 7; | |
5780 | o.push_back(new OSDSuperblock(z)); | |
5781 | o.push_back(new OSDSuperblock(z)); | |
5782 | } | |
5783 | ||
5784 | // -- SnapSet -- | |
5785 | ||
9f95a23c | 5786 | void SnapSet::encode(ceph::buffer::list& bl) const |
7c673cae FG |
5787 | { |
5788 | ENCODE_START(3, 2, bl); | |
11fdf7f2 TL |
5789 | encode(seq, bl); |
5790 | encode(true, bl); // head_exists | |
5791 | encode(snaps, bl); | |
5792 | encode(clones, bl); | |
5793 | encode(clone_overlap, bl); | |
5794 | encode(clone_size, bl); | |
5795 | encode(clone_snaps, bl); | |
7c673cae FG |
5796 | ENCODE_FINISH(bl); |
5797 | } | |
5798 | ||
9f95a23c | 5799 | void SnapSet::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
5800 | { |
5801 | DECODE_START_LEGACY_COMPAT_LEN(3, 2, 2, bl); | |
11fdf7f2 | 5802 | decode(seq, bl); |
9f95a23c | 5803 | bl += 1u; // skip legacy head_exists (always true) |
11fdf7f2 TL |
5804 | decode(snaps, bl); |
5805 | decode(clones, bl); | |
5806 | decode(clone_overlap, bl); | |
5807 | decode(clone_size, bl); | |
7c673cae | 5808 | if (struct_v >= 3) { |
11fdf7f2 | 5809 | decode(clone_snaps, bl); |
7c673cae FG |
5810 | } else { |
5811 | clone_snaps.clear(); | |
5812 | } | |
5813 | DECODE_FINISH(bl); | |
5814 | } | |
5815 | ||
5816 | void SnapSet::dump(Formatter *f) const | |
5817 | { | |
9f95a23c | 5818 | f->dump_unsigned("seq", seq); |
7c673cae | 5819 | f->open_array_section("clones"); |
9f95a23c | 5820 | for (auto p = clones.cbegin(); p != clones.cend(); ++p) { |
7c673cae FG |
5821 | f->open_object_section("clone"); |
5822 | f->dump_unsigned("snap", *p); | |
94b18763 FG |
5823 | auto cs = clone_size.find(*p); |
5824 | if (cs != clone_size.end()) | |
5825 | f->dump_unsigned("size", cs->second); | |
5826 | else | |
5827 | f->dump_string("size", "????"); | |
5828 | auto co = clone_overlap.find(*p); | |
5829 | if (co != clone_overlap.end()) | |
5830 | f->dump_stream("overlap") << co->second; | |
5831 | else | |
5832 | f->dump_stream("overlap") << "????"; | |
7c673cae FG |
5833 | auto q = clone_snaps.find(*p); |
5834 | if (q != clone_snaps.end()) { | |
5835 | f->open_array_section("snaps"); | |
5836 | for (auto s : q->second) { | |
5837 | f->dump_unsigned("snap", s); | |
5838 | } | |
5839 | f->close_section(); | |
5840 | } | |
5841 | f->close_section(); | |
5842 | } | |
5843 | f->close_section(); | |
5844 | } | |
5845 | ||
5846 | void SnapSet::generate_test_instances(list<SnapSet*>& o) | |
5847 | { | |
5848 | o.push_back(new SnapSet); | |
5849 | o.push_back(new SnapSet); | |
7c673cae FG |
5850 | o.back()->seq = 123; |
5851 | o.back()->snaps.push_back(123); | |
5852 | o.back()->snaps.push_back(12); | |
5853 | o.push_back(new SnapSet); | |
7c673cae FG |
5854 | o.back()->seq = 123; |
5855 | o.back()->snaps.push_back(123); | |
5856 | o.back()->snaps.push_back(12); | |
5857 | o.back()->clones.push_back(12); | |
5858 | o.back()->clone_size[12] = 12345; | |
5859 | o.back()->clone_overlap[12]; | |
5860 | o.back()->clone_snaps[12] = {12, 10, 8}; | |
5861 | } | |
5862 | ||
5863 | ostream& operator<<(ostream& out, const SnapSet& cs) | |
5864 | { | |
11fdf7f2 TL |
5865 | return out << cs.seq << "=" << cs.snaps << ":" |
5866 | << cs.clone_snaps; | |
7c673cae FG |
5867 | } |
5868 | ||
5869 | void SnapSet::from_snap_set(const librados::snap_set_t& ss, bool legacy) | |
5870 | { | |
5871 | // NOTE: our reconstruction of snaps (and the snapc) is not strictly | |
5872 | // correct: it will not include snaps that still logically exist | |
5873 | // but for which there was no clone that is defined. For all | |
5874 | // practical purposes this doesn't matter, since we only use that | |
5875 | // information to clone on the OSD, and we have already moved | |
5876 | // forward past that part of the object history. | |
5877 | ||
5878 | seq = ss.seq; | |
5879 | set<snapid_t> _snaps; | |
5880 | set<snapid_t> _clones; | |
9f95a23c | 5881 | for (auto p = ss.clones.cbegin(); p != ss.clones.cend(); ++p) { |
11fdf7f2 | 5882 | if (p->cloneid != librados::SNAP_HEAD) { |
7c673cae FG |
5883 | _clones.insert(p->cloneid); |
5884 | _snaps.insert(p->snaps.begin(), p->snaps.end()); | |
5885 | clone_size[p->cloneid] = p->size; | |
5886 | clone_overlap[p->cloneid]; // the entry must exist, even if it's empty. | |
9f95a23c | 5887 | for (auto q = p->overlap.cbegin(); q != p->overlap.cend(); ++q) |
7c673cae FG |
5888 | clone_overlap[p->cloneid].insert(q->first, q->second); |
5889 | if (!legacy) { | |
5890 | // p->snaps is ascending; clone_snaps is descending | |
5891 | vector<snapid_t>& v = clone_snaps[p->cloneid]; | |
5892 | for (auto q = p->snaps.rbegin(); q != p->snaps.rend(); ++q) { | |
5893 | v.push_back(*q); | |
5894 | } | |
5895 | } | |
5896 | } | |
5897 | } | |
5898 | ||
5899 | // ascending | |
5900 | clones.clear(); | |
5901 | clones.reserve(_clones.size()); | |
9f95a23c | 5902 | for (auto p = _clones.begin(); p != _clones.end(); ++p) |
7c673cae FG |
5903 | clones.push_back(*p); |
5904 | ||
5905 | // descending | |
5906 | snaps.clear(); | |
5907 | snaps.reserve(_snaps.size()); | |
9f95a23c | 5908 | for (auto p = _snaps.rbegin(); |
7c673cae FG |
5909 | p != _snaps.rend(); ++p) |
5910 | snaps.push_back(*p); | |
5911 | } | |
5912 | ||
5913 | uint64_t SnapSet::get_clone_bytes(snapid_t clone) const | |
5914 | { | |
11fdf7f2 | 5915 | ceph_assert(clone_size.count(clone)); |
7c673cae | 5916 | uint64_t size = clone_size.find(clone)->second; |
11fdf7f2 | 5917 | ceph_assert(clone_overlap.count(clone)); |
7c673cae | 5918 | const interval_set<uint64_t> &overlap = clone_overlap.find(clone)->second; |
11fdf7f2 TL |
5919 | ceph_assert(size >= (uint64_t)overlap.size()); |
5920 | return size - overlap.size(); | |
7c673cae FG |
5921 | } |
5922 | ||
5923 | void SnapSet::filter(const pg_pool_t &pinfo) | |
5924 | { | |
5925 | vector<snapid_t> oldsnaps; | |
5926 | oldsnaps.swap(snaps); | |
9f95a23c | 5927 | for (auto i = oldsnaps.cbegin(); i != oldsnaps.cend(); ++i) { |
7c673cae FG |
5928 | if (!pinfo.is_removed_snap(*i)) |
5929 | snaps.push_back(*i); | |
5930 | } | |
5931 | } | |
5932 | ||
5933 | SnapSet SnapSet::get_filtered(const pg_pool_t &pinfo) const | |
5934 | { | |
5935 | SnapSet ss = *this; | |
5936 | ss.filter(pinfo); | |
5937 | return ss; | |
5938 | } | |
5939 | ||
5940 | // -- watch_info_t -- | |
5941 | ||
9f95a23c | 5942 | void watch_info_t::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae FG |
5943 | { |
5944 | ENCODE_START(4, 3, bl); | |
11fdf7f2 TL |
5945 | encode(cookie, bl); |
5946 | encode(timeout_seconds, bl); | |
5947 | encode(addr, bl, features); | |
7c673cae FG |
5948 | ENCODE_FINISH(bl); |
5949 | } | |
5950 | ||
9f95a23c | 5951 | void watch_info_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
5952 | { |
5953 | DECODE_START_LEGACY_COMPAT_LEN(4, 3, 3, bl); | |
11fdf7f2 | 5954 | decode(cookie, bl); |
7c673cae FG |
5955 | if (struct_v < 2) { |
5956 | uint64_t ver; | |
11fdf7f2 | 5957 | decode(ver, bl); |
7c673cae | 5958 | } |
11fdf7f2 | 5959 | decode(timeout_seconds, bl); |
7c673cae | 5960 | if (struct_v >= 4) { |
11fdf7f2 | 5961 | decode(addr, bl); |
7c673cae FG |
5962 | } |
5963 | DECODE_FINISH(bl); | |
5964 | } | |
5965 | ||
5966 | void watch_info_t::dump(Formatter *f) const | |
5967 | { | |
5968 | f->dump_unsigned("cookie", cookie); | |
5969 | f->dump_unsigned("timeout_seconds", timeout_seconds); | |
5970 | f->open_object_section("addr"); | |
5971 | addr.dump(f); | |
5972 | f->close_section(); | |
5973 | } | |
5974 | ||
5975 | void watch_info_t::generate_test_instances(list<watch_info_t*>& o) | |
5976 | { | |
5977 | o.push_back(new watch_info_t); | |
5978 | o.push_back(new watch_info_t); | |
5979 | o.back()->cookie = 123; | |
5980 | o.back()->timeout_seconds = 99; | |
5981 | entity_addr_t ea; | |
5982 | ea.set_type(entity_addr_t::TYPE_LEGACY); | |
5983 | ea.set_nonce(1); | |
5984 | ea.set_family(AF_INET); | |
5985 | ea.set_in4_quad(0, 127); | |
5986 | ea.set_in4_quad(1, 0); | |
5987 | ea.set_in4_quad(2, 1); | |
5988 | ea.set_in4_quad(3, 2); | |
5989 | ea.set_port(2); | |
5990 | o.back()->addr = ea; | |
5991 | } | |
5992 | ||
11fdf7f2 TL |
5993 | // -- chunk_info_t -- |
5994 | ||
9f95a23c | 5995 | void chunk_info_t::encode(ceph::buffer::list& bl) const |
11fdf7f2 TL |
5996 | { |
5997 | ENCODE_START(1, 1, bl); | |
5998 | encode(offset, bl); | |
5999 | encode(length, bl); | |
6000 | encode(oid, bl); | |
6001 | __u32 _flags = flags; | |
6002 | encode(_flags, bl); | |
6003 | ENCODE_FINISH(bl); | |
6004 | } | |
6005 | ||
9f95a23c | 6006 | void chunk_info_t::decode(ceph::buffer::list::const_iterator& bl) |
11fdf7f2 TL |
6007 | { |
6008 | DECODE_START(1, bl); | |
6009 | decode(offset, bl); | |
6010 | decode(length, bl); | |
6011 | decode(oid, bl); | |
6012 | __u32 _flags; | |
6013 | decode(_flags, bl); | |
6014 | flags = (cflag_t)_flags; | |
6015 | DECODE_FINISH(bl); | |
6016 | } | |
6017 | ||
6018 | void chunk_info_t::dump(Formatter *f) const | |
6019 | { | |
6020 | f->dump_unsigned("length", length); | |
6021 | f->open_object_section("oid"); | |
6022 | oid.dump(f); | |
6023 | f->close_section(); | |
6024 | f->dump_unsigned("flags", flags); | |
6025 | } | |
6026 | ||
f67539c2 TL |
6027 | |
6028 | bool chunk_info_t::operator==(const chunk_info_t& cit) const | |
6029 | { | |
6030 | if (has_fingerprint()) { | |
6031 | if (oid.oid.name == cit.oid.oid.name) { | |
6032 | return true; | |
6033 | } | |
6034 | } else { | |
6035 | if (offset == cit.offset && length == cit.length && | |
6036 | oid.oid.name == cit.oid.oid.name) { | |
6037 | return true; | |
6038 | } | |
6039 | ||
6040 | } | |
6041 | return false; | |
6042 | } | |
6043 | ||
6044 | bool operator==(const std::pair<const long unsigned int, chunk_info_t> & l, | |
6045 | const std::pair<const long unsigned int, chunk_info_t> & r) | |
6046 | { | |
6047 | return l.first == r.first && | |
6048 | l.second == r.second; | |
6049 | } | |
6050 | ||
11fdf7f2 TL |
6051 | ostream& operator<<(ostream& out, const chunk_info_t& ci) |
6052 | { | |
6053 | return out << "(len: " << ci.length << " oid: " << ci.oid | |
6054 | << " offset: " << ci.offset | |
6055 | << " flags: " << ci.get_flag_string(ci.flags) << ")"; | |
6056 | } | |
6057 | ||
31f18b77 FG |
6058 | // -- object_manifest_t -- |
6059 | ||
f67539c2 TL |
6060 | std::ostream& operator<<(std::ostream& out, const object_ref_delta_t & ci) |
6061 | { | |
6062 | return out << ci.ref_delta << std::endl; | |
6063 | } | |
6064 | ||
6065 | void object_manifest_t::calc_refs_to_inc_on_set( | |
6066 | const object_manifest_t* _g, | |
6067 | const object_manifest_t* _l, | |
6068 | object_ref_delta_t &refs) const | |
6069 | { | |
6070 | /* avoid to increment the same reference on adjacent clones */ | |
6071 | auto iter = chunk_map.begin(); | |
6072 | auto find_chunk = [](decltype(iter) &i, const object_manifest_t* cur) | |
6073 | -> bool { | |
6074 | if (cur) { | |
6075 | auto c = cur->chunk_map.find(i->first); | |
6076 | if (c != cur->chunk_map.end() && c->second == i->second) { | |
6077 | return true; | |
6078 | ||
6079 | } | |
6080 | } | |
6081 | return false; | |
6082 | }; | |
6083 | ||
6084 | /* If at least a same chunk exists on either _g or _l, do not increment | |
6085 | * the reference | |
6086 | * | |
6087 | * head: [0, 2) ccc, [6, 2) bbb, [8, 2) ccc | |
6088 | * 20: [0, 2) aaa, <- set_chunk | |
6089 | * 30: [0, 2) abc, [6, 2) bbb, [8, 2) ccc | |
6090 | * --> incremnt the reference | |
6091 | * | |
6092 | * head: [0, 2) ccc, [6, 2) bbb, [8, 2) ccc | |
6093 | * 20: [0, 2) ccc, <- set_chunk | |
6094 | * 30: [0, 2) abc, [6, 2) bbb, [8, 2) ccc | |
6095 | * --> do not need to increment | |
6096 | * | |
6097 | * head: [0, 2) ccc, [6, 2) bbb, [8, 2) ccc | |
6098 | * 20: [0, 2) ccc, <- set_chunk | |
6099 | * 30: [0, 2) ccc, [6, 2) bbb, [8, 2) ccc | |
6100 | * --> decrement the reference of ccc | |
6101 | * | |
6102 | */ | |
6103 | for (; iter != chunk_map.end(); ++iter) { | |
6104 | auto found_g = find_chunk(iter, _g); | |
6105 | auto found_l = find_chunk(iter, _l); | |
6106 | if (!found_g && !found_l) { | |
6107 | refs.inc_ref(iter->second.oid); | |
6108 | } else if (found_g && found_l) { | |
6109 | refs.dec_ref(iter->second.oid); | |
6110 | } | |
6111 | } | |
6112 | } | |
6113 | ||
6114 | void object_manifest_t::calc_refs_to_drop_on_modify( | |
6115 | const object_manifest_t* _l, | |
6116 | const ObjectCleanRegions& clean_regions, | |
6117 | object_ref_delta_t &refs) const | |
6118 | { | |
6119 | for (auto &p : chunk_map) { | |
6120 | if (!clean_regions.is_clean_region(p.first, p.second.length)) { | |
6121 | // has previous snapshot | |
6122 | if (_l) { | |
6123 | /* | |
6124 | * Let's assume that there is a manifest snapshotted object which has three chunks | |
6125 | * head: [0, 2) aaa, [6, 2) bbb, [8, 2) ccc | |
6126 | * 20: [0, 2) aaa, [6, 2) bbb, [8, 2) ccc | |
6127 | * | |
6128 | * If we modify [6, 2) at head, we shouldn't decrement bbb's refcount because | |
6129 | * 20 has the reference for bbb. Therefore, we only drop the reference if two chunks | |
6130 | * (head: [6, 2) and 20: [6, 2)) are different. | |
6131 | * | |
6132 | */ | |
6133 | auto c = _l->chunk_map.find(p.first); | |
6134 | if (c != _l->chunk_map.end()) { | |
6135 | if (p.second == c->second) { | |
6136 | continue; | |
6137 | } | |
6138 | } | |
6139 | refs.dec_ref(p.second.oid); | |
6140 | } else { | |
6141 | // decrement the reference of the updated chunks if the manifest object has no snapshot | |
6142 | refs.dec_ref(p.second.oid); | |
6143 | } | |
6144 | } | |
6145 | } | |
6146 | } | |
6147 | ||
6148 | void object_manifest_t::calc_refs_to_drop_on_removal( | |
6149 | const object_manifest_t* _g, | |
6150 | const object_manifest_t* _l, | |
6151 | object_ref_delta_t &refs) const | |
6152 | { | |
6153 | /* At a high level, the rule is that consecutive clones with the same reference | |
6154 | * at the same offset share a reference. As such, removing *this may result | |
6155 | * in removing references in two cases: | |
6156 | * 1) *this has a reference which it shares with neither _g nor _l | |
6157 | * 2) _g and _l have a reference which they share with each other but not | |
6158 | * *this. | |
6159 | * | |
6160 | * For a particular offset, both 1 and 2 can happen. | |
6161 | * | |
6162 | * Notably, this means that to evaluate the reference change from removing | |
6163 | * the object with *this, we only need to look at the two adjacent clones. | |
6164 | */ | |
6165 | ||
6166 | // Paper over possibly missing _g or _l -- nullopt is semantically the same | |
6167 | // as an empty chunk_map | |
6168 | static const object_manifest_t empty; | |
6169 | const object_manifest_t &g = _g ? *_g : empty; | |
6170 | const object_manifest_t &l = _l ? *_l : empty; | |
6171 | ||
6172 | auto giter = g.chunk_map.begin(); | |
6173 | auto iter = chunk_map.begin(); | |
6174 | auto liter = l.chunk_map.begin(); | |
6175 | ||
6176 | // Translate iter, map pair to the current offset, end() -> max | |
6177 | auto get_offset = [](decltype(iter) &i, const object_manifest_t &manifest) | |
6178 | -> uint64_t { | |
6179 | return i == manifest.chunk_map.end() ? | |
6180 | std::numeric_limits<uint64_t>::max() : i->first; | |
6181 | }; | |
6182 | ||
6183 | /* If current matches the offset at iter, returns the chunk at *iter | |
6184 | * and increments iter. Otherwise, returns nullptr. | |
6185 | * | |
6186 | * current will always be derived from the min of *giter, *iter, and | |
6187 | * *liter on each cycle, so the result will be that each loop iteration | |
6188 | * will pick up all chunks at the offest being considered, each offset | |
6189 | * will be considered once, and all offsets will be considered. | |
6190 | */ | |
6191 | auto get_chunk = []( | |
6192 | uint64_t current, decltype(iter) &i, const object_manifest_t &manifest) | |
6193 | -> const chunk_info_t * { | |
6194 | if (i == manifest.chunk_map.end() || current != i->first) { | |
6195 | return nullptr; | |
6196 | } else { | |
6197 | return &(i++)->second; | |
6198 | } | |
6199 | }; | |
6200 | ||
6201 | while (giter != g.chunk_map.end() || | |
6202 | iter != chunk_map.end() || | |
6203 | liter != l.chunk_map.end()) { | |
6204 | auto current = std::min( | |
6205 | std::min(get_offset(giter, g), get_offset(iter, *this)), | |
6206 | get_offset(liter, l)); | |
6207 | ||
6208 | auto gchunk = get_chunk(current, giter, g); | |
6209 | auto chunk = get_chunk(current, iter, *this); | |
6210 | auto lchunk = get_chunk(current, liter, l); | |
6211 | ||
6212 | if (gchunk && lchunk && *gchunk == *lchunk && | |
6213 | (!chunk || *gchunk != *chunk)) { | |
6214 | // case 1 from above: l and g match, chunk does not | |
6215 | refs.dec_ref(gchunk->oid); | |
6216 | } | |
6217 | ||
6218 | if (chunk && | |
6219 | (!gchunk || chunk->oid != gchunk->oid) && | |
6220 | (!lchunk || chunk->oid != lchunk->oid)) { | |
6221 | // case 2 from above: *this matches neither | |
6222 | refs.dec_ref(chunk->oid); | |
6223 | } | |
6224 | } | |
6225 | } | |
6226 | ||
9f95a23c | 6227 | void object_manifest_t::encode(ceph::buffer::list& bl) const |
31f18b77 FG |
6228 | { |
6229 | ENCODE_START(1, 1, bl); | |
11fdf7f2 | 6230 | encode(type, bl); |
31f18b77 FG |
6231 | switch (type) { |
6232 | case TYPE_NONE: break; | |
6233 | case TYPE_REDIRECT: | |
11fdf7f2 TL |
6234 | encode(redirect_target, bl); |
6235 | break; | |
6236 | case TYPE_CHUNKED: | |
9f95a23c | 6237 | encode(chunk_map, bl); |
31f18b77 FG |
6238 | break; |
6239 | default: | |
6240 | ceph_abort(); | |
6241 | } | |
6242 | ENCODE_FINISH(bl); | |
6243 | } | |
6244 | ||
9f95a23c | 6245 | void object_manifest_t::decode(ceph::buffer::list::const_iterator& bl) |
31f18b77 FG |
6246 | { |
6247 | DECODE_START(1, bl); | |
11fdf7f2 | 6248 | decode(type, bl); |
31f18b77 FG |
6249 | switch (type) { |
6250 | case TYPE_NONE: break; | |
6251 | case TYPE_REDIRECT: | |
11fdf7f2 TL |
6252 | decode(redirect_target, bl); |
6253 | break; | |
6254 | case TYPE_CHUNKED: | |
6255 | decode(chunk_map, bl); | |
31f18b77 FG |
6256 | break; |
6257 | default: | |
6258 | ceph_abort(); | |
6259 | } | |
6260 | DECODE_FINISH(bl); | |
6261 | } | |
6262 | ||
6263 | void object_manifest_t::dump(Formatter *f) const | |
6264 | { | |
6265 | f->dump_unsigned("type", type); | |
11fdf7f2 TL |
6266 | if (type == TYPE_REDIRECT) { |
6267 | f->open_object_section("redirect_target"); | |
6268 | redirect_target.dump(f); | |
6269 | f->close_section(); | |
6270 | } else if (type == TYPE_CHUNKED) { | |
6271 | f->open_array_section("chunk_map"); | |
6272 | for (auto& p : chunk_map) { | |
6273 | f->open_object_section("chunk"); | |
6274 | f->dump_unsigned("offset", p.first); | |
6275 | p.second.dump(f); | |
6276 | f->close_section(); | |
6277 | } | |
6278 | f->close_section(); | |
6279 | } | |
31f18b77 FG |
6280 | } |
6281 | ||
6282 | void object_manifest_t::generate_test_instances(list<object_manifest_t*>& o) | |
6283 | { | |
6284 | o.push_back(new object_manifest_t()); | |
6285 | o.back()->type = TYPE_REDIRECT; | |
6286 | } | |
6287 | ||
6288 | ostream& operator<<(ostream& out, const object_manifest_t& om) | |
6289 | { | |
11fdf7f2 TL |
6290 | out << "manifest(" << om.get_type_name(); |
6291 | if (om.is_redirect()) { | |
6292 | out << " " << om.redirect_target; | |
6293 | } else if (om.is_chunked()) { | |
6294 | out << " " << om.chunk_map; | |
6295 | } | |
6296 | out << ")"; | |
6297 | return out; | |
31f18b77 | 6298 | } |
7c673cae FG |
6299 | |
6300 | // -- object_info_t -- | |
6301 | ||
6302 | void object_info_t::copy_user_bits(const object_info_t& other) | |
6303 | { | |
6304 | // these bits are copied from head->clone. | |
6305 | size = other.size; | |
6306 | mtime = other.mtime; | |
6307 | local_mtime = other.local_mtime; | |
6308 | last_reqid = other.last_reqid; | |
6309 | truncate_seq = other.truncate_seq; | |
6310 | truncate_size = other.truncate_size; | |
6311 | flags = other.flags; | |
6312 | user_version = other.user_version; | |
6313 | data_digest = other.data_digest; | |
6314 | omap_digest = other.omap_digest; | |
6315 | } | |
6316 | ||
9f95a23c | 6317 | void object_info_t::encode(ceph::buffer::list& bl, uint64_t features) const |
7c673cae FG |
6318 | { |
6319 | object_locator_t myoloc(soid); | |
6320 | map<entity_name_t, watch_info_t> old_watchers; | |
9f95a23c | 6321 | for (auto i = watchers.cbegin(); i != watchers.cend(); ++i) { |
7c673cae FG |
6322 | old_watchers.insert(make_pair(i->first.second, i->second)); |
6323 | } | |
31f18b77 | 6324 | ENCODE_START(17, 8, bl); |
11fdf7f2 TL |
6325 | encode(soid, bl); |
6326 | encode(myoloc, bl); //Retained for compatibility | |
6327 | encode((__u32)0, bl); // was category, no longer used | |
6328 | encode(version, bl); | |
6329 | encode(prior_version, bl); | |
6330 | encode(last_reqid, bl); | |
6331 | encode(size, bl); | |
6332 | encode(mtime, bl); | |
7c673cae | 6333 | if (soid.snap == CEPH_NOSNAP) |
11fdf7f2 | 6334 | encode(osd_reqid_t(), bl); // used to be wrlock_by |
7c673cae | 6335 | else |
11fdf7f2 TL |
6336 | encode((uint32_t)0, bl); // was legacy_snaps |
6337 | encode(truncate_seq, bl); | |
6338 | encode(truncate_size, bl); | |
6339 | encode(is_lost(), bl); | |
6340 | encode(old_watchers, bl, features); | |
7c673cae FG |
6341 | /* shenanigans to avoid breaking backwards compatibility in the disk format. |
6342 | * When we can, switch this out for simply putting the version_t on disk. */ | |
6343 | eversion_t user_eversion(0, user_version); | |
11fdf7f2 TL |
6344 | encode(user_eversion, bl); |
6345 | encode(test_flag(FLAG_USES_TMAP), bl); | |
6346 | encode(watchers, bl, features); | |
7c673cae | 6347 | __u32 _flags = flags; |
11fdf7f2 TL |
6348 | encode(_flags, bl); |
6349 | encode(local_mtime, bl); | |
6350 | encode(data_digest, bl); | |
6351 | encode(omap_digest, bl); | |
6352 | encode(expected_object_size, bl); | |
6353 | encode(expected_write_size, bl); | |
6354 | encode(alloc_hint_flags, bl); | |
31f18b77 | 6355 | if (has_manifest()) { |
11fdf7f2 | 6356 | encode(manifest, bl); |
31f18b77 | 6357 | } |
7c673cae FG |
6358 | ENCODE_FINISH(bl); |
6359 | } | |
6360 | ||
9f95a23c | 6361 | void object_info_t::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae FG |
6362 | { |
6363 | object_locator_t myoloc; | |
31f18b77 | 6364 | DECODE_START_LEGACY_COMPAT_LEN(17, 8, 8, bl); |
7c673cae | 6365 | map<entity_name_t, watch_info_t> old_watchers; |
11fdf7f2 TL |
6366 | decode(soid, bl); |
6367 | decode(myoloc, bl); | |
7c673cae FG |
6368 | { |
6369 | string category; | |
11fdf7f2 | 6370 | decode(category, bl); // no longer used |
7c673cae | 6371 | } |
11fdf7f2 TL |
6372 | decode(version, bl); |
6373 | decode(prior_version, bl); | |
6374 | decode(last_reqid, bl); | |
6375 | decode(size, bl); | |
6376 | decode(mtime, bl); | |
7c673cae FG |
6377 | if (soid.snap == CEPH_NOSNAP) { |
6378 | osd_reqid_t wrlock_by; | |
11fdf7f2 | 6379 | decode(wrlock_by, bl); |
7c673cae | 6380 | } else { |
11fdf7f2 TL |
6381 | vector<snapid_t> legacy_snaps; |
6382 | decode(legacy_snaps, bl); | |
7c673cae | 6383 | } |
11fdf7f2 TL |
6384 | decode(truncate_seq, bl); |
6385 | decode(truncate_size, bl); | |
7c673cae FG |
6386 | |
6387 | // if this is struct_v >= 13, we will overwrite this | |
6388 | // below since this field is just here for backwards | |
6389 | // compatibility | |
6390 | __u8 lo; | |
11fdf7f2 | 6391 | decode(lo, bl); |
7c673cae FG |
6392 | flags = (flag_t)lo; |
6393 | ||
11fdf7f2 | 6394 | decode(old_watchers, bl); |
7c673cae | 6395 | eversion_t user_eversion; |
11fdf7f2 | 6396 | decode(user_eversion, bl); |
7c673cae FG |
6397 | user_version = user_eversion.version; |
6398 | ||
6399 | if (struct_v >= 9) { | |
6400 | bool uses_tmap = false; | |
11fdf7f2 | 6401 | decode(uses_tmap, bl); |
7c673cae FG |
6402 | if (uses_tmap) |
6403 | set_flag(FLAG_USES_TMAP); | |
6404 | } else { | |
6405 | set_flag(FLAG_USES_TMAP); | |
6406 | } | |
6407 | if (struct_v < 10) | |
6408 | soid.pool = myoloc.pool; | |
6409 | if (struct_v >= 11) { | |
11fdf7f2 | 6410 | decode(watchers, bl); |
7c673cae | 6411 | } else { |
9f95a23c | 6412 | for (auto i = old_watchers.begin(); i != old_watchers.end(); ++i) { |
7c673cae FG |
6413 | watchers.insert( |
6414 | make_pair( | |
6415 | make_pair(i->second.cookie, i->first), i->second)); | |
6416 | } | |
6417 | } | |
6418 | if (struct_v >= 13) { | |
6419 | __u32 _flags; | |
11fdf7f2 | 6420 | decode(_flags, bl); |
7c673cae FG |
6421 | flags = (flag_t)_flags; |
6422 | } | |
6423 | if (struct_v >= 14) { | |
11fdf7f2 | 6424 | decode(local_mtime, bl); |
7c673cae FG |
6425 | } else { |
6426 | local_mtime = utime_t(); | |
6427 | } | |
6428 | if (struct_v >= 15) { | |
11fdf7f2 TL |
6429 | decode(data_digest, bl); |
6430 | decode(omap_digest, bl); | |
7c673cae FG |
6431 | } else { |
6432 | data_digest = omap_digest = -1; | |
6433 | clear_flag(FLAG_DATA_DIGEST); | |
6434 | clear_flag(FLAG_OMAP_DIGEST); | |
6435 | } | |
6436 | if (struct_v >= 16) { | |
11fdf7f2 TL |
6437 | decode(expected_object_size, bl); |
6438 | decode(expected_write_size, bl); | |
6439 | decode(alloc_hint_flags, bl); | |
7c673cae FG |
6440 | } else { |
6441 | expected_object_size = 0; | |
6442 | expected_write_size = 0; | |
6443 | alloc_hint_flags = 0; | |
6444 | } | |
31f18b77 FG |
6445 | if (struct_v >= 17) { |
6446 | if (has_manifest()) { | |
11fdf7f2 | 6447 | decode(manifest, bl); |
31f18b77 FG |
6448 | } |
6449 | } | |
7c673cae FG |
6450 | DECODE_FINISH(bl); |
6451 | } | |
6452 | ||
6453 | void object_info_t::dump(Formatter *f) const | |
6454 | { | |
6455 | f->open_object_section("oid"); | |
6456 | soid.dump(f); | |
6457 | f->close_section(); | |
6458 | f->dump_stream("version") << version; | |
6459 | f->dump_stream("prior_version") << prior_version; | |
6460 | f->dump_stream("last_reqid") << last_reqid; | |
6461 | f->dump_unsigned("user_version", user_version); | |
6462 | f->dump_unsigned("size", size); | |
6463 | f->dump_stream("mtime") << mtime; | |
6464 | f->dump_stream("local_mtime") << local_mtime; | |
6465 | f->dump_unsigned("lost", (int)is_lost()); | |
94b18763 FG |
6466 | vector<string> sv = get_flag_vector(flags); |
6467 | f->open_array_section("flags"); | |
20effc67 | 6468 | for (const auto& str: sv) { |
94b18763 | 6469 | f->dump_string("flags", str); |
20effc67 | 6470 | } |
94b18763 | 6471 | f->close_section(); |
7c673cae FG |
6472 | f->dump_unsigned("truncate_seq", truncate_seq); |
6473 | f->dump_unsigned("truncate_size", truncate_size); | |
94b18763 FG |
6474 | f->dump_format("data_digest", "0x%08x", data_digest); |
6475 | f->dump_format("omap_digest", "0x%08x", omap_digest); | |
7c673cae FG |
6476 | f->dump_unsigned("expected_object_size", expected_object_size); |
6477 | f->dump_unsigned("expected_write_size", expected_write_size); | |
6478 | f->dump_unsigned("alloc_hint_flags", alloc_hint_flags); | |
31f18b77 | 6479 | f->dump_object("manifest", manifest); |
7c673cae | 6480 | f->open_object_section("watchers"); |
9f95a23c | 6481 | for (auto p = watchers.cbegin(); p != watchers.cend(); ++p) { |
f67539c2 TL |
6482 | CachedStackStringStream css; |
6483 | *css << p->first.second; | |
6484 | f->open_object_section(css->strv()); | |
7c673cae FG |
6485 | p->second.dump(f); |
6486 | f->close_section(); | |
6487 | } | |
6488 | f->close_section(); | |
6489 | } | |
6490 | ||
6491 | void object_info_t::generate_test_instances(list<object_info_t*>& o) | |
6492 | { | |
6493 | o.push_back(new object_info_t()); | |
6494 | ||
6495 | // fixme | |
6496 | } | |
6497 | ||
6498 | ||
6499 | ostream& operator<<(ostream& out, const object_info_t& oi) | |
6500 | { | |
6501 | out << oi.soid << "(" << oi.version | |
6502 | << " " << oi.last_reqid; | |
7c673cae FG |
6503 | if (oi.flags) |
6504 | out << " " << oi.get_flag_string(); | |
6505 | out << " s " << oi.size; | |
6506 | out << " uv " << oi.user_version; | |
6507 | if (oi.is_data_digest()) | |
6508 | out << " dd " << std::hex << oi.data_digest << std::dec; | |
6509 | if (oi.is_omap_digest()) | |
6510 | out << " od " << std::hex << oi.omap_digest << std::dec; | |
6511 | out << " alloc_hint [" << oi.expected_object_size | |
6512 | << " " << oi.expected_write_size | |
6513 | << " " << oi.alloc_hint_flags << "]"; | |
31f18b77 FG |
6514 | if (oi.has_manifest()) |
6515 | out << " " << oi.manifest; | |
7c673cae FG |
6516 | out << ")"; |
6517 | return out; | |
6518 | } | |
6519 | ||
6520 | // -- ObjectRecovery -- | |
9f95a23c | 6521 | void ObjectRecoveryProgress::encode(ceph::buffer::list &bl) const |
7c673cae FG |
6522 | { |
6523 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
6524 | encode(first, bl); |
6525 | encode(data_complete, bl); | |
6526 | encode(data_recovered_to, bl); | |
6527 | encode(omap_recovered_to, bl); | |
6528 | encode(omap_complete, bl); | |
7c673cae FG |
6529 | ENCODE_FINISH(bl); |
6530 | } | |
6531 | ||
9f95a23c | 6532 | void ObjectRecoveryProgress::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
6533 | { |
6534 | DECODE_START(1, bl); | |
11fdf7f2 TL |
6535 | decode(first, bl); |
6536 | decode(data_complete, bl); | |
6537 | decode(data_recovered_to, bl); | |
6538 | decode(omap_recovered_to, bl); | |
6539 | decode(omap_complete, bl); | |
7c673cae FG |
6540 | DECODE_FINISH(bl); |
6541 | } | |
6542 | ||
6543 | ostream &operator<<(ostream &out, const ObjectRecoveryProgress &prog) | |
6544 | { | |
6545 | return prog.print(out); | |
6546 | } | |
6547 | ||
6548 | void ObjectRecoveryProgress::generate_test_instances( | |
6549 | list<ObjectRecoveryProgress*>& o) | |
6550 | { | |
6551 | o.push_back(new ObjectRecoveryProgress); | |
6552 | o.back()->first = false; | |
6553 | o.back()->data_complete = true; | |
6554 | o.back()->omap_complete = true; | |
6555 | o.back()->data_recovered_to = 100; | |
6556 | ||
6557 | o.push_back(new ObjectRecoveryProgress); | |
6558 | o.back()->first = true; | |
6559 | o.back()->data_complete = false; | |
6560 | o.back()->omap_complete = false; | |
6561 | o.back()->data_recovered_to = 0; | |
6562 | } | |
6563 | ||
6564 | ostream &ObjectRecoveryProgress::print(ostream &out) const | |
6565 | { | |
6566 | return out << "ObjectRecoveryProgress(" | |
6567 | << ( first ? "" : "!" ) << "first, " | |
6568 | << "data_recovered_to:" << data_recovered_to | |
6569 | << ", data_complete:" << ( data_complete ? "true" : "false" ) | |
6570 | << ", omap_recovered_to:" << omap_recovered_to | |
6571 | << ", omap_complete:" << ( omap_complete ? "true" : "false" ) | |
224ce89b | 6572 | << ", error:" << ( error ? "true" : "false" ) |
7c673cae FG |
6573 | << ")"; |
6574 | } | |
6575 | ||
6576 | void ObjectRecoveryProgress::dump(Formatter *f) const | |
6577 | { | |
6578 | f->dump_int("first?", first); | |
6579 | f->dump_int("data_complete?", data_complete); | |
6580 | f->dump_unsigned("data_recovered_to", data_recovered_to); | |
6581 | f->dump_int("omap_complete?", omap_complete); | |
6582 | f->dump_string("omap_recovered_to", omap_recovered_to); | |
6583 | } | |
6584 | ||
9f95a23c | 6585 | void ObjectRecoveryInfo::encode(ceph::buffer::list &bl, uint64_t features) const |
7c673cae | 6586 | { |
9f95a23c | 6587 | ENCODE_START(3, 1, bl); |
11fdf7f2 TL |
6588 | encode(soid, bl); |
6589 | encode(version, bl); | |
6590 | encode(size, bl); | |
6591 | encode(oi, bl, features); | |
6592 | encode(ss, bl); | |
6593 | encode(copy_subset, bl); | |
6594 | encode(clone_subset, bl); | |
9f95a23c | 6595 | encode(object_exist, bl); |
7c673cae FG |
6596 | ENCODE_FINISH(bl); |
6597 | } | |
6598 | ||
9f95a23c | 6599 | void ObjectRecoveryInfo::decode(ceph::buffer::list::const_iterator &bl, |
7c673cae FG |
6600 | int64_t pool) |
6601 | { | |
9f95a23c | 6602 | DECODE_START(3, bl); |
11fdf7f2 TL |
6603 | decode(soid, bl); |
6604 | decode(version, bl); | |
6605 | decode(size, bl); | |
6606 | decode(oi, bl); | |
6607 | decode(ss, bl); | |
6608 | decode(copy_subset, bl); | |
6609 | decode(clone_subset, bl); | |
9f95a23c TL |
6610 | if (struct_v > 2) |
6611 | decode(object_exist, bl); | |
6612 | else | |
6613 | object_exist = false; | |
7c673cae | 6614 | DECODE_FINISH(bl); |
7c673cae FG |
6615 | if (struct_v < 2) { |
6616 | if (!soid.is_max() && soid.pool == -1) | |
6617 | soid.pool = pool; | |
6618 | map<hobject_t, interval_set<uint64_t>> tmp; | |
6619 | tmp.swap(clone_subset); | |
9f95a23c | 6620 | for (auto i = tmp.begin(); i != tmp.end(); ++i) { |
7c673cae FG |
6621 | hobject_t first(i->first); |
6622 | if (!first.is_max() && first.pool == -1) | |
6623 | first.pool = pool; | |
6624 | clone_subset[first].swap(i->second); | |
6625 | } | |
6626 | } | |
6627 | } | |
6628 | ||
6629 | void ObjectRecoveryInfo::generate_test_instances( | |
6630 | list<ObjectRecoveryInfo*>& o) | |
6631 | { | |
6632 | o.push_back(new ObjectRecoveryInfo); | |
6633 | o.back()->soid = hobject_t(sobject_t("key", CEPH_NOSNAP)); | |
6634 | o.back()->version = eversion_t(0,0); | |
6635 | o.back()->size = 100; | |
9f95a23c | 6636 | o.back()->object_exist = false; |
7c673cae FG |
6637 | } |
6638 | ||
6639 | ||
6640 | void ObjectRecoveryInfo::dump(Formatter *f) const | |
6641 | { | |
6642 | f->dump_stream("object") << soid; | |
6643 | f->dump_stream("at_version") << version; | |
6644 | f->dump_stream("size") << size; | |
6645 | { | |
6646 | f->open_object_section("object_info"); | |
6647 | oi.dump(f); | |
6648 | f->close_section(); | |
6649 | } | |
6650 | { | |
6651 | f->open_object_section("snapset"); | |
6652 | ss.dump(f); | |
6653 | f->close_section(); | |
6654 | } | |
6655 | f->dump_stream("copy_subset") << copy_subset; | |
6656 | f->dump_stream("clone_subset") << clone_subset; | |
9f95a23c | 6657 | f->dump_stream("object_exist") << object_exist; |
7c673cae FG |
6658 | } |
6659 | ||
6660 | ostream& operator<<(ostream& out, const ObjectRecoveryInfo &inf) | |
6661 | { | |
6662 | return inf.print(out); | |
6663 | } | |
6664 | ||
6665 | ostream &ObjectRecoveryInfo::print(ostream &out) const | |
6666 | { | |
6667 | return out << "ObjectRecoveryInfo(" | |
6668 | << soid << "@" << version | |
6669 | << ", size: " << size | |
6670 | << ", copy_subset: " << copy_subset | |
6671 | << ", clone_subset: " << clone_subset | |
6672 | << ", snapset: " << ss | |
9f95a23c | 6673 | << ", object_exist: " << object_exist |
7c673cae FG |
6674 | << ")"; |
6675 | } | |
6676 | ||
6677 | // -- PushReplyOp -- | |
6678 | void PushReplyOp::generate_test_instances(list<PushReplyOp*> &o) | |
6679 | { | |
6680 | o.push_back(new PushReplyOp); | |
6681 | o.push_back(new PushReplyOp); | |
6682 | o.back()->soid = hobject_t(sobject_t("asdf", 2)); | |
6683 | o.push_back(new PushReplyOp); | |
6684 | o.back()->soid = hobject_t(sobject_t("asdf", CEPH_NOSNAP)); | |
6685 | } | |
6686 | ||
9f95a23c | 6687 | void PushReplyOp::encode(ceph::buffer::list &bl) const |
7c673cae FG |
6688 | { |
6689 | ENCODE_START(1, 1, bl); | |
11fdf7f2 | 6690 | encode(soid, bl); |
7c673cae FG |
6691 | ENCODE_FINISH(bl); |
6692 | } | |
6693 | ||
9f95a23c | 6694 | void PushReplyOp::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
6695 | { |
6696 | DECODE_START(1, bl); | |
11fdf7f2 | 6697 | decode(soid, bl); |
7c673cae FG |
6698 | DECODE_FINISH(bl); |
6699 | } | |
6700 | ||
6701 | void PushReplyOp::dump(Formatter *f) const | |
6702 | { | |
6703 | f->dump_stream("soid") << soid; | |
6704 | } | |
6705 | ||
6706 | ostream &PushReplyOp::print(ostream &out) const | |
6707 | { | |
6708 | return out | |
6709 | << "PushReplyOp(" << soid | |
6710 | << ")"; | |
6711 | } | |
6712 | ||
6713 | ostream& operator<<(ostream& out, const PushReplyOp &op) | |
6714 | { | |
6715 | return op.print(out); | |
6716 | } | |
6717 | ||
6718 | uint64_t PushReplyOp::cost(CephContext *cct) const | |
6719 | { | |
1e59de90 TL |
6720 | if (cct->_conf->osd_op_queue == "mclock_scheduler") { |
6721 | /* In general, we really never want to throttle PushReplyOp messages. | |
6722 | * As long as the object is smaller than osd_recovery_max_chunk (8M at | |
6723 | * time of writing this comment, so this is basically always true), | |
6724 | * processing the PushReplyOp does not cost any further IO and simply | |
6725 | * permits the object once more to be written to. | |
6726 | * | |
6727 | * In the unlikely event that the object is larger than | |
6728 | * osd_recovery_max_chunk (again, 8M at the moment, so never for common | |
6729 | * configurations of rbd and virtually never for cephfs and rgw), | |
6730 | * we *still* want to push out the next portion immediately so that we can | |
6731 | * release the object for IO. | |
6732 | * | |
6733 | * The throttling for this operation on the primary occurs at the point | |
6734 | * where we queue the PGRecoveryContext which calls into recover_missing | |
6735 | * and recover_backfill to initiate pushes. | |
6736 | * See OSD::queue_recovery_context. | |
6737 | */ | |
6738 | return 1; | |
6739 | } else { | |
6740 | /* We retain this legacy behavior for WeightedPriorityQueue. It seems to | |
6741 | * require very large costs for several messages in order to do any | |
6742 | * meaningful amount of throttling. This branch should be removed after | |
6743 | * Reef. | |
6744 | */ | |
6745 | return cct->_conf->osd_push_per_object_cost + | |
6746 | cct->_conf->osd_recovery_max_chunk; | |
6747 | } | |
7c673cae FG |
6748 | } |
6749 | ||
6750 | // -- PullOp -- | |
6751 | void PullOp::generate_test_instances(list<PullOp*> &o) | |
6752 | { | |
6753 | o.push_back(new PullOp); | |
6754 | o.push_back(new PullOp); | |
6755 | o.back()->soid = hobject_t(sobject_t("asdf", 2)); | |
6756 | o.back()->recovery_info.version = eversion_t(3, 10); | |
6757 | o.push_back(new PullOp); | |
6758 | o.back()->soid = hobject_t(sobject_t("asdf", CEPH_NOSNAP)); | |
6759 | o.back()->recovery_info.version = eversion_t(0, 0); | |
6760 | } | |
6761 | ||
9f95a23c | 6762 | void PullOp::encode(ceph::buffer::list &bl, uint64_t features) const |
7c673cae FG |
6763 | { |
6764 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
6765 | encode(soid, bl); |
6766 | encode(recovery_info, bl, features); | |
6767 | encode(recovery_progress, bl); | |
7c673cae FG |
6768 | ENCODE_FINISH(bl); |
6769 | } | |
6770 | ||
9f95a23c | 6771 | void PullOp::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
6772 | { |
6773 | DECODE_START(1, bl); | |
11fdf7f2 TL |
6774 | decode(soid, bl); |
6775 | decode(recovery_info, bl); | |
6776 | decode(recovery_progress, bl); | |
7c673cae FG |
6777 | DECODE_FINISH(bl); |
6778 | } | |
6779 | ||
6780 | void PullOp::dump(Formatter *f) const | |
6781 | { | |
6782 | f->dump_stream("soid") << soid; | |
6783 | { | |
6784 | f->open_object_section("recovery_info"); | |
6785 | recovery_info.dump(f); | |
6786 | f->close_section(); | |
6787 | } | |
6788 | { | |
6789 | f->open_object_section("recovery_progress"); | |
6790 | recovery_progress.dump(f); | |
6791 | f->close_section(); | |
6792 | } | |
6793 | } | |
6794 | ||
6795 | ostream &PullOp::print(ostream &out) const | |
6796 | { | |
6797 | return out | |
6798 | << "PullOp(" << soid | |
6799 | << ", recovery_info: " << recovery_info | |
6800 | << ", recovery_progress: " << recovery_progress | |
6801 | << ")"; | |
6802 | } | |
6803 | ||
6804 | ostream& operator<<(ostream& out, const PullOp &op) | |
6805 | { | |
6806 | return op.print(out); | |
6807 | } | |
6808 | ||
6809 | uint64_t PullOp::cost(CephContext *cct) const | |
6810 | { | |
1e59de90 TL |
6811 | if (cct->_conf->osd_op_queue == "mclock_scheduler") { |
6812 | return std::clamp<uint64_t>( | |
6813 | recovery_progress.estimate_remaining_data_to_recover(recovery_info), | |
6814 | 1, | |
6815 | cct->_conf->osd_recovery_max_chunk); | |
6816 | } else { | |
6817 | /* We retain this legacy behavior for WeightedPriorityQueue. It seems to | |
6818 | * require very large costs for several messages in order to do any | |
6819 | * meaningful amount of throttling. This branch should be removed after | |
6820 | * Reef. | |
6821 | */ | |
6822 | return cct->_conf->osd_push_per_object_cost + | |
6823 | cct->_conf->osd_recovery_max_chunk; | |
6824 | } | |
7c673cae FG |
6825 | } |
6826 | ||
6827 | // -- PushOp -- | |
6828 | void PushOp::generate_test_instances(list<PushOp*> &o) | |
6829 | { | |
6830 | o.push_back(new PushOp); | |
6831 | o.push_back(new PushOp); | |
6832 | o.back()->soid = hobject_t(sobject_t("asdf", 2)); | |
6833 | o.back()->version = eversion_t(3, 10); | |
6834 | o.push_back(new PushOp); | |
6835 | o.back()->soid = hobject_t(sobject_t("asdf", CEPH_NOSNAP)); | |
6836 | o.back()->version = eversion_t(0, 0); | |
6837 | } | |
6838 | ||
9f95a23c | 6839 | void PushOp::encode(ceph::buffer::list &bl, uint64_t features) const |
7c673cae FG |
6840 | { |
6841 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
6842 | encode(soid, bl); |
6843 | encode(version, bl); | |
6844 | encode(data, bl); | |
6845 | encode(data_included, bl); | |
6846 | encode(omap_header, bl); | |
6847 | encode(omap_entries, bl); | |
6848 | encode(attrset, bl); | |
6849 | encode(recovery_info, bl, features); | |
6850 | encode(after_progress, bl); | |
6851 | encode(before_progress, bl); | |
7c673cae FG |
6852 | ENCODE_FINISH(bl); |
6853 | } | |
6854 | ||
9f95a23c | 6855 | void PushOp::decode(ceph::buffer::list::const_iterator &bl) |
7c673cae FG |
6856 | { |
6857 | DECODE_START(1, bl); | |
11fdf7f2 TL |
6858 | decode(soid, bl); |
6859 | decode(version, bl); | |
6860 | decode(data, bl); | |
6861 | decode(data_included, bl); | |
6862 | decode(omap_header, bl); | |
6863 | decode(omap_entries, bl); | |
6864 | decode(attrset, bl); | |
6865 | decode(recovery_info, bl); | |
6866 | decode(after_progress, bl); | |
6867 | decode(before_progress, bl); | |
7c673cae FG |
6868 | DECODE_FINISH(bl); |
6869 | } | |
6870 | ||
6871 | void PushOp::dump(Formatter *f) const | |
6872 | { | |
6873 | f->dump_stream("soid") << soid; | |
6874 | f->dump_stream("version") << version; | |
6875 | f->dump_int("data_len", data.length()); | |
6876 | f->dump_stream("data_included") << data_included; | |
6877 | f->dump_int("omap_header_len", omap_header.length()); | |
6878 | f->dump_int("omap_entries_len", omap_entries.size()); | |
6879 | f->dump_int("attrset_len", attrset.size()); | |
6880 | { | |
6881 | f->open_object_section("recovery_info"); | |
6882 | recovery_info.dump(f); | |
6883 | f->close_section(); | |
6884 | } | |
6885 | { | |
6886 | f->open_object_section("after_progress"); | |
6887 | after_progress.dump(f); | |
6888 | f->close_section(); | |
6889 | } | |
6890 | { | |
6891 | f->open_object_section("before_progress"); | |
6892 | before_progress.dump(f); | |
6893 | f->close_section(); | |
6894 | } | |
6895 | } | |
6896 | ||
6897 | ostream &PushOp::print(ostream &out) const | |
6898 | { | |
6899 | return out | |
6900 | << "PushOp(" << soid | |
6901 | << ", version: " << version | |
6902 | << ", data_included: " << data_included | |
6903 | << ", data_size: " << data.length() | |
6904 | << ", omap_header_size: " << omap_header.length() | |
6905 | << ", omap_entries_size: " << omap_entries.size() | |
6906 | << ", attrset_size: " << attrset.size() | |
6907 | << ", recovery_info: " << recovery_info | |
6908 | << ", after_progress: " << after_progress | |
6909 | << ", before_progress: " << before_progress | |
6910 | << ")"; | |
6911 | } | |
6912 | ||
6913 | ostream& operator<<(ostream& out, const PushOp &op) | |
6914 | { | |
6915 | return op.print(out); | |
6916 | } | |
6917 | ||
6918 | uint64_t PushOp::cost(CephContext *cct) const | |
6919 | { | |
6920 | uint64_t cost = data_included.size(); | |
9f95a23c | 6921 | for (auto i = omap_entries.cbegin(); i != omap_entries.cend(); ++i) { |
7c673cae FG |
6922 | cost += i->second.length(); |
6923 | } | |
6924 | cost += cct->_conf->osd_push_per_object_cost; | |
6925 | return cost; | |
6926 | } | |
6927 | ||
6928 | // -- ScrubMap -- | |
6929 | ||
6930 | void ScrubMap::merge_incr(const ScrubMap &l) | |
6931 | { | |
11fdf7f2 | 6932 | ceph_assert(valid_through == l.incr_since); |
7c673cae FG |
6933 | valid_through = l.valid_through; |
6934 | ||
9f95a23c | 6935 | for (auto p = l.objects.cbegin(); p != l.objects.cend(); ++p){ |
7c673cae | 6936 | if (p->second.negative) { |
9f95a23c | 6937 | auto q = objects.find(p->first); |
7c673cae FG |
6938 | if (q != objects.end()) { |
6939 | objects.erase(q); | |
6940 | } | |
6941 | } else { | |
6942 | objects[p->first] = p->second; | |
6943 | } | |
6944 | } | |
6945 | } | |
6946 | ||
9f95a23c | 6947 | void ScrubMap::encode(ceph::buffer::list& bl) const |
7c673cae FG |
6948 | { |
6949 | ENCODE_START(3, 2, bl); | |
11fdf7f2 TL |
6950 | encode(objects, bl); |
6951 | encode((__u32)0, bl); // used to be attrs; now deprecated | |
9f95a23c | 6952 | ceph::buffer::list old_logbl; // not used |
11fdf7f2 TL |
6953 | encode(old_logbl, bl); |
6954 | encode(valid_through, bl); | |
6955 | encode(incr_since, bl); | |
7c673cae FG |
6956 | ENCODE_FINISH(bl); |
6957 | } | |
6958 | ||
9f95a23c | 6959 | void ScrubMap::decode(ceph::buffer::list::const_iterator& bl, int64_t pool) |
7c673cae FG |
6960 | { |
6961 | DECODE_START_LEGACY_COMPAT_LEN(3, 2, 2, bl); | |
11fdf7f2 | 6962 | decode(objects, bl); |
7c673cae FG |
6963 | { |
6964 | map<string,string> attrs; // deprecated | |
11fdf7f2 | 6965 | decode(attrs, bl); |
7c673cae | 6966 | } |
9f95a23c | 6967 | ceph::buffer::list old_logbl; // not used |
11fdf7f2 TL |
6968 | decode(old_logbl, bl); |
6969 | decode(valid_through, bl); | |
6970 | decode(incr_since, bl); | |
7c673cae FG |
6971 | DECODE_FINISH(bl); |
6972 | ||
6973 | // handle hobject_t upgrade | |
6974 | if (struct_v < 3) { | |
6975 | map<hobject_t, object> tmp; | |
6976 | tmp.swap(objects); | |
9f95a23c | 6977 | for (auto i = tmp.begin(); i != tmp.end(); ++i) { |
7c673cae FG |
6978 | hobject_t first(i->first); |
6979 | if (!first.is_max() && first.pool == -1) | |
6980 | first.pool = pool; | |
6981 | objects[first] = i->second; | |
6982 | } | |
6983 | } | |
6984 | } | |
6985 | ||
6986 | void ScrubMap::dump(Formatter *f) const | |
6987 | { | |
6988 | f->dump_stream("valid_through") << valid_through; | |
6989 | f->dump_stream("incremental_since") << incr_since; | |
6990 | f->open_array_section("objects"); | |
9f95a23c | 6991 | for (auto p = objects.cbegin(); p != objects.cend(); ++p) { |
7c673cae FG |
6992 | f->open_object_section("object"); |
6993 | f->dump_string("name", p->first.oid.name); | |
6994 | f->dump_unsigned("hash", p->first.get_hash()); | |
6995 | f->dump_string("key", p->first.get_key()); | |
6996 | f->dump_int("snapid", p->first.snap); | |
6997 | p->second.dump(f); | |
6998 | f->close_section(); | |
6999 | } | |
7000 | f->close_section(); | |
7001 | } | |
7002 | ||
7003 | void ScrubMap::generate_test_instances(list<ScrubMap*>& o) | |
7004 | { | |
7005 | o.push_back(new ScrubMap); | |
7006 | o.push_back(new ScrubMap); | |
7007 | o.back()->valid_through = eversion_t(1, 2); | |
7008 | o.back()->incr_since = eversion_t(3, 4); | |
7009 | list<object*> obj; | |
7010 | object::generate_test_instances(obj); | |
7011 | o.back()->objects[hobject_t(object_t("foo"), "fookey", 123, 456, 0, "")] = *obj.back(); | |
7012 | obj.pop_back(); | |
7013 | o.back()->objects[hobject_t(object_t("bar"), string(), 123, 456, 0, "")] = *obj.back(); | |
7014 | } | |
7015 | ||
7016 | // -- ScrubMap::object -- | |
7017 | ||
9f95a23c | 7018 | void ScrubMap::object::encode(ceph::buffer::list& bl) const |
7c673cae FG |
7019 | { |
7020 | bool compat_read_error = read_error || ec_hash_mismatch || ec_size_mismatch; | |
11fdf7f2 TL |
7021 | ENCODE_START(10, 7, bl); |
7022 | encode(size, bl); | |
7023 | encode(negative, bl); | |
7024 | encode(attrs, bl); | |
7025 | encode(digest, bl); | |
7026 | encode(digest_present, bl); | |
7027 | encode((uint32_t)0, bl); // obsolete nlinks | |
7028 | encode((uint32_t)0, bl); // snapcolls | |
7029 | encode(omap_digest, bl); | |
7030 | encode(omap_digest_present, bl); | |
7031 | encode(compat_read_error, bl); | |
7032 | encode(stat_error, bl); | |
7033 | encode(read_error, bl); | |
7034 | encode(ec_hash_mismatch, bl); | |
7035 | encode(ec_size_mismatch, bl); | |
7036 | encode(large_omap_object_found, bl); | |
7037 | encode(large_omap_object_key_count, bl); | |
7038 | encode(large_omap_object_value_size, bl); | |
7039 | encode(object_omap_bytes, bl); | |
7040 | encode(object_omap_keys, bl); | |
7c673cae FG |
7041 | ENCODE_FINISH(bl); |
7042 | } | |
7043 | ||
9f95a23c | 7044 | void ScrubMap::object::decode(ceph::buffer::list::const_iterator& bl) |
7c673cae | 7045 | { |
11fdf7f2 TL |
7046 | DECODE_START(10, bl); |
7047 | decode(size, bl); | |
7c673cae | 7048 | bool tmp, compat_read_error = false; |
11fdf7f2 | 7049 | decode(tmp, bl); |
7c673cae | 7050 | negative = tmp; |
11fdf7f2 TL |
7051 | decode(attrs, bl); |
7052 | decode(digest, bl); | |
7053 | decode(tmp, bl); | |
7c673cae FG |
7054 | digest_present = tmp; |
7055 | { | |
7056 | uint32_t nlinks; | |
11fdf7f2 | 7057 | decode(nlinks, bl); |
7c673cae | 7058 | set<snapid_t> snapcolls; |
11fdf7f2 | 7059 | decode(snapcolls, bl); |
7c673cae | 7060 | } |
11fdf7f2 TL |
7061 | decode(omap_digest, bl); |
7062 | decode(tmp, bl); | |
7c673cae | 7063 | omap_digest_present = tmp; |
11fdf7f2 TL |
7064 | decode(compat_read_error, bl); |
7065 | decode(tmp, bl); | |
7c673cae FG |
7066 | stat_error = tmp; |
7067 | if (struct_v >= 8) { | |
11fdf7f2 | 7068 | decode(tmp, bl); |
7c673cae | 7069 | read_error = tmp; |
11fdf7f2 | 7070 | decode(tmp, bl); |
7c673cae | 7071 | ec_hash_mismatch = tmp; |
11fdf7f2 | 7072 | decode(tmp, bl); |
7c673cae FG |
7073 | ec_size_mismatch = tmp; |
7074 | } | |
7075 | // If older encoder found a read_error, set read_error | |
7076 | if (compat_read_error && !read_error && !ec_hash_mismatch && !ec_size_mismatch) | |
7077 | read_error = true; | |
28e407b8 | 7078 | if (struct_v >= 9) { |
11fdf7f2 | 7079 | decode(tmp, bl); |
28e407b8 | 7080 | large_omap_object_found = tmp; |
11fdf7f2 TL |
7081 | decode(large_omap_object_key_count, bl); |
7082 | decode(large_omap_object_value_size, bl); | |
7083 | } | |
7084 | if (struct_v >= 10) { | |
7085 | decode(object_omap_bytes, bl); | |
7086 | decode(object_omap_keys, bl); | |
28e407b8 | 7087 | } |
7c673cae FG |
7088 | DECODE_FINISH(bl); |
7089 | } | |
7090 | ||
7091 | void ScrubMap::object::dump(Formatter *f) const | |
7092 | { | |
7093 | f->dump_int("size", size); | |
7094 | f->dump_int("negative", negative); | |
7095 | f->open_array_section("attrs"); | |
9f95a23c | 7096 | for (auto p = attrs.cbegin(); p != attrs.cend(); ++p) { |
7c673cae FG |
7097 | f->open_object_section("attr"); |
7098 | f->dump_string("name", p->first); | |
7099 | f->dump_int("length", p->second.length()); | |
7100 | f->close_section(); | |
7101 | } | |
7102 | f->close_section(); | |
7103 | } | |
7104 | ||
7105 | void ScrubMap::object::generate_test_instances(list<object*>& o) | |
7106 | { | |
7107 | o.push_back(new object); | |
7108 | o.push_back(new object); | |
7109 | o.back()->negative = true; | |
7110 | o.push_back(new object); | |
7111 | o.back()->size = 123; | |
9f95a23c TL |
7112 | o.back()->attrs["foo"] = ceph::buffer::copy("foo", 3); |
7113 | o.back()->attrs["bar"] = ceph::buffer::copy("barval", 6); | |
7c673cae FG |
7114 | } |
7115 | ||
7116 | // -- OSDOp -- | |
7117 | ||
7118 | ostream& operator<<(ostream& out, const OSDOp& op) | |
7119 | { | |
7120 | out << ceph_osd_op_name(op.op.op); | |
7121 | if (ceph_osd_op_type_data(op.op.op)) { | |
7122 | // data extent | |
7123 | switch (op.op.op) { | |
7124 | case CEPH_OSD_OP_ASSERT_VER: | |
7125 | out << " v" << op.op.assert_ver.ver; | |
7126 | break; | |
7127 | case CEPH_OSD_OP_TRUNCATE: | |
7128 | out << " " << op.op.extent.offset; | |
7129 | break; | |
7130 | case CEPH_OSD_OP_MASKTRUNC: | |
7131 | case CEPH_OSD_OP_TRIMTRUNC: | |
7132 | out << " " << op.op.extent.truncate_seq << "@" | |
7133 | << (int64_t)op.op.extent.truncate_size; | |
7134 | break; | |
7135 | case CEPH_OSD_OP_ROLLBACK: | |
7136 | out << " " << snapid_t(op.op.snap.snapid); | |
7137 | break; | |
7138 | case CEPH_OSD_OP_WATCH: | |
7139 | out << " " << ceph_osd_watch_op_name(op.op.watch.op) | |
7140 | << " cookie " << op.op.watch.cookie; | |
7141 | if (op.op.watch.gen) | |
7142 | out << " gen " << op.op.watch.gen; | |
7143 | break; | |
7144 | case CEPH_OSD_OP_NOTIFY: | |
7c673cae FG |
7145 | out << " cookie " << op.op.notify.cookie; |
7146 | break; | |
7147 | case CEPH_OSD_OP_COPY_GET: | |
7148 | out << " max " << op.op.copy_get.max; | |
7149 | break; | |
7150 | case CEPH_OSD_OP_COPY_FROM: | |
7151 | out << " ver " << op.op.copy_from.src_version; | |
7152 | break; | |
7153 | case CEPH_OSD_OP_SETALLOCHINT: | |
7154 | out << " object_size " << op.op.alloc_hint.expected_object_size | |
7155 | << " write_size " << op.op.alloc_hint.expected_write_size; | |
7156 | break; | |
7157 | case CEPH_OSD_OP_READ: | |
7158 | case CEPH_OSD_OP_SPARSE_READ: | |
7159 | case CEPH_OSD_OP_SYNC_READ: | |
7160 | case CEPH_OSD_OP_WRITE: | |
7161 | case CEPH_OSD_OP_WRITEFULL: | |
7162 | case CEPH_OSD_OP_ZERO: | |
7163 | case CEPH_OSD_OP_APPEND: | |
7164 | case CEPH_OSD_OP_MAPEXT: | |
11fdf7f2 | 7165 | case CEPH_OSD_OP_CMPEXT: |
7c673cae FG |
7166 | out << " " << op.op.extent.offset << "~" << op.op.extent.length; |
7167 | if (op.op.extent.truncate_seq) | |
7168 | out << " [" << op.op.extent.truncate_seq << "@" | |
7169 | << (int64_t)op.op.extent.truncate_size << "]"; | |
7170 | if (op.op.flags) | |
7171 | out << " [" << ceph_osd_op_flag_string(op.op.flags) << "]"; | |
7172 | default: | |
7173 | // don't show any arg info | |
7174 | break; | |
7175 | } | |
7176 | } else if (ceph_osd_op_type_attr(op.op.op)) { | |
7177 | // xattr name | |
7178 | if (op.op.xattr.name_len && op.indata.length()) { | |
7179 | out << " "; | |
7180 | op.indata.write(0, op.op.xattr.name_len, out); | |
7181 | } | |
7182 | if (op.op.xattr.value_len) | |
7183 | out << " (" << op.op.xattr.value_len << ")"; | |
7184 | if (op.op.op == CEPH_OSD_OP_CMPXATTR) | |
7185 | out << " op " << (int)op.op.xattr.cmp_op | |
7186 | << " mode " << (int)op.op.xattr.cmp_mode; | |
7187 | } else if (ceph_osd_op_type_exec(op.op.op)) { | |
7188 | // class.method | |
7189 | if (op.op.cls.class_len && op.indata.length()) { | |
7190 | out << " "; | |
7191 | op.indata.write(0, op.op.cls.class_len, out); | |
7192 | out << "."; | |
7193 | op.indata.write(op.op.cls.class_len, op.op.cls.method_len, out); | |
7194 | } | |
7195 | } else if (ceph_osd_op_type_pg(op.op.op)) { | |
7196 | switch (op.op.op) { | |
7197 | case CEPH_OSD_OP_PGLS: | |
7198 | case CEPH_OSD_OP_PGLS_FILTER: | |
7199 | case CEPH_OSD_OP_PGNLS: | |
7200 | case CEPH_OSD_OP_PGNLS_FILTER: | |
7201 | out << " start_epoch " << op.op.pgls.start_epoch; | |
7202 | break; | |
7203 | case CEPH_OSD_OP_PG_HITSET_LS: | |
7204 | break; | |
7205 | case CEPH_OSD_OP_PG_HITSET_GET: | |
7206 | out << " " << utime_t(op.op.hit_set_get.stamp); | |
7207 | break; | |
7208 | case CEPH_OSD_OP_SCRUBLS: | |
7209 | break; | |
7210 | } | |
7211 | } | |
9f95a23c TL |
7212 | if (op.indata.length()) { |
7213 | out << " in=" << op.indata.length() << "b"; | |
7214 | } | |
7215 | if (op.outdata.length()) { | |
7216 | out << " out=" << op.outdata.length() << "b"; | |
7217 | } | |
7c673cae FG |
7218 | return out; |
7219 | } | |
7220 | ||
7221 | ||
9f95a23c | 7222 | void OSDOp::split_osd_op_vector_out_data(vector<OSDOp>& ops, ceph::buffer::list& in) |
7c673cae | 7223 | { |
9f95a23c | 7224 | auto datap = in.begin(); |
7c673cae FG |
7225 | for (unsigned i = 0; i < ops.size(); i++) { |
7226 | if (ops[i].op.payload_len) { | |
7227 | datap.copy(ops[i].op.payload_len, ops[i].outdata); | |
7228 | } | |
7229 | } | |
7230 | } | |
7231 | ||
9f95a23c | 7232 | void OSDOp::merge_osd_op_vector_out_data(vector<OSDOp>& ops, ceph::buffer::list& out) |
7c673cae FG |
7233 | { |
7234 | for (unsigned i = 0; i < ops.size(); i++) { | |
9f95a23c | 7235 | ops[i].op.payload_len = ops[i].outdata.length(); |
7c673cae | 7236 | if (ops[i].outdata.length()) { |
7c673cae FG |
7237 | out.append(ops[i].outdata); |
7238 | } | |
7239 | } | |
7240 | } | |
7241 | ||
9f95a23c TL |
7242 | int prepare_info_keymap( |
7243 | CephContext* cct, | |
7244 | map<string,bufferlist> *km, | |
7245 | string *key_to_remove, | |
7246 | epoch_t epoch, | |
7247 | pg_info_t &info, | |
7248 | pg_info_t &last_written_info, | |
7249 | PastIntervals &past_intervals, | |
7250 | bool dirty_big_info, | |
7251 | bool dirty_epoch, | |
7252 | bool try_fast_info, | |
7253 | PerfCounters *logger, | |
7254 | DoutPrefixProvider *dpp) | |
7255 | { | |
7256 | if (dirty_epoch) { | |
7257 | encode(epoch, (*km)[string(epoch_key)]); | |
7258 | } | |
7259 | ||
7260 | if (logger) | |
7261 | logger->inc(l_osd_pg_info); | |
7262 | ||
7263 | // try to do info efficiently? | |
7264 | if (!dirty_big_info && try_fast_info && | |
7265 | info.last_update > last_written_info.last_update) { | |
7266 | pg_fast_info_t fast; | |
7267 | fast.populate_from(info); | |
7268 | bool did = fast.try_apply_to(&last_written_info); | |
7269 | ceph_assert(did); // we verified last_update increased above | |
7270 | if (info == last_written_info) { | |
7271 | encode(fast, (*km)[string(fastinfo_key)]); | |
7272 | if (logger) | |
7273 | logger->inc(l_osd_pg_fastinfo); | |
7274 | return 0; | |
7275 | } | |
7276 | if (dpp) { | |
7277 | ldpp_dout(dpp, 30) << __func__ << " fastinfo failed, info:\n"; | |
7278 | { | |
7279 | JSONFormatter jf(true); | |
7280 | jf.dump_object("info", info); | |
7281 | jf.flush(*_dout); | |
7282 | } | |
7283 | { | |
7284 | *_dout << "\nlast_written_info:\n"; | |
7285 | JSONFormatter jf(true); | |
7286 | jf.dump_object("last_written_info", last_written_info); | |
7287 | jf.flush(*_dout); | |
7288 | } | |
7289 | *_dout << dendl; | |
7290 | } | |
7291 | } else if (info.last_update <= last_written_info.last_update) { | |
7292 | // clean up any potentially stale fastinfo key resulting from last_update | |
7293 | // not moving forwards (e.g., a backwards jump during peering) | |
7294 | *key_to_remove = fastinfo_key; | |
7295 | } | |
7296 | ||
7297 | last_written_info = info; | |
7298 | ||
7299 | // info. store purged_snaps separately. | |
7300 | interval_set<snapid_t> purged_snaps; | |
7301 | purged_snaps.swap(info.purged_snaps); | |
7302 | encode(info, (*km)[string(info_key)]); | |
7303 | purged_snaps.swap(info.purged_snaps); | |
7304 | ||
7305 | if (dirty_big_info) { | |
7306 | // potentially big stuff | |
7307 | bufferlist& bigbl = (*km)[string(biginfo_key)]; | |
7308 | encode(past_intervals, bigbl); | |
7309 | encode(info.purged_snaps, bigbl); | |
7310 | //dout(20) << "write_info bigbl " << bigbl.length() << dendl; | |
7311 | if (logger) | |
7312 | logger->inc(l_osd_pg_biginfo); | |
7313 | } | |
7314 | ||
7315 | return 0; | |
7316 | } | |
7317 | ||
7318 | void create_pg_collection( | |
7319 | ceph::os::Transaction& t, spg_t pgid, int bits) | |
7320 | { | |
7321 | coll_t coll(pgid); | |
7322 | t.create_collection(coll, bits); | |
7323 | } | |
7324 | ||
7325 | void init_pg_ondisk( | |
7326 | ceph::os::Transaction& t, | |
7327 | spg_t pgid, | |
7328 | const pg_pool_t *pool) | |
7329 | { | |
7330 | coll_t coll(pgid); | |
7331 | if (pool) { | |
7332 | // Give a hint to the PG collection | |
7333 | bufferlist hint; | |
7334 | uint32_t pg_num = pool->get_pg_num(); | |
7335 | uint64_t expected_num_objects_pg = pool->expected_num_objects / pg_num; | |
7336 | encode(pg_num, hint); | |
7337 | encode(expected_num_objects_pg, hint); | |
7338 | uint32_t hint_type = ceph::os::Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS; | |
7339 | t.collection_hint(coll, hint_type, hint); | |
7340 | } | |
7341 | ||
7342 | ghobject_t pgmeta_oid(pgid.make_pgmeta_oid()); | |
7343 | t.touch(coll, pgmeta_oid); | |
7344 | map<string,bufferlist> values; | |
7345 | __u8 struct_v = pg_latest_struct_v; | |
7346 | encode(struct_v, values[string(infover_key)]); | |
7347 | t.omap_setkeys(coll, pgmeta_oid, values); | |
7348 | } | |
7349 | ||
7350 | PGLSFilter::PGLSFilter() : cct(nullptr) | |
7351 | { | |
7352 | } | |
7353 | ||
7354 | PGLSFilter::~PGLSFilter() | |
7355 | { | |
7356 | } | |
7357 | ||
7358 | int PGLSPlainFilter::init(ceph::bufferlist::const_iterator ¶ms) | |
7359 | { | |
7360 | try { | |
7361 | decode(xattr, params); | |
7362 | decode(val, params); | |
f67539c2 | 7363 | } catch (ceph::buffer::error &e) { |
9f95a23c TL |
7364 | return -EINVAL; |
7365 | } | |
7366 | return 0; | |
7367 | } | |
7368 | ||
7369 | bool PGLSPlainFilter::filter(const hobject_t& obj, | |
7370 | const ceph::bufferlist& xattr_data) const | |
7371 | { | |
7372 | return xattr_data.contents_equal(val.c_str(), val.size()); | |
7373 | } |