]>
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 | ||
7c673cae FG |
4 | #include "include/types.h" |
5 | ||
7c673cae FG |
6 | #include <errno.h> |
7 | ||
7c673cae FG |
8 | #include "objclass/objclass.h" |
9 | #include "cls/rgw/cls_rgw_ops.h" | |
10 | #include "cls/rgw/cls_rgw_const.h" | |
11 | #include "common/Clock.h" | |
12 | #include "common/strtol.h" | |
13 | #include "common/escape.h" | |
14 | ||
7c673cae FG |
15 | #include "include/compat.h" |
16 | ||
17 | CLS_VER(1,0) | |
18 | CLS_NAME(rgw) | |
19 | ||
20 | ||
21 | #define BI_PREFIX_CHAR 0x80 | |
22 | ||
23 | #define BI_BUCKET_OBJS_INDEX 0 | |
24 | #define BI_BUCKET_LOG_INDEX 1 | |
25 | #define BI_BUCKET_OBJ_INSTANCE_INDEX 2 | |
26 | #define BI_BUCKET_OLH_DATA_INDEX 3 | |
27 | ||
28 | #define BI_BUCKET_LAST_INDEX 4 | |
29 | ||
30 | static string bucket_index_prefixes[] = { "", /* special handling for the objs list index */ | |
31 | "0_", /* bucket log index */ | |
32 | "1000_", /* obj instance index */ | |
33 | "1001_", /* olh data index */ | |
34 | ||
35 | /* this must be the last index */ | |
36 | "9999_",}; | |
37 | ||
38 | static bool bi_is_objs_index(const string& s) { | |
39 | return ((unsigned char)s[0] != BI_PREFIX_CHAR); | |
40 | } | |
41 | ||
42 | int bi_entry_type(const string& s) | |
43 | { | |
44 | if (bi_is_objs_index(s)) { | |
45 | return BI_BUCKET_OBJS_INDEX; | |
46 | } | |
47 | ||
48 | for (size_t i = 1; | |
49 | i < sizeof(bucket_index_prefixes) / sizeof(bucket_index_prefixes[0]); | |
50 | ++i) { | |
51 | const string& t = bucket_index_prefixes[i]; | |
52 | ||
53 | if (s.compare(1, t.size(), t) == 0) { | |
54 | return i; | |
55 | } | |
56 | } | |
57 | ||
58 | return -EINVAL; | |
59 | } | |
60 | ||
61 | static bool bi_entry_gt(const string& first, const string& second) | |
62 | { | |
63 | int fi = bi_entry_type(first); | |
64 | int si = bi_entry_type(second); | |
65 | ||
66 | if (fi > si) { | |
67 | return true; | |
68 | } else if (fi < si) { | |
69 | return false; | |
70 | } | |
71 | ||
72 | return first > second; | |
73 | } | |
74 | ||
75 | static void get_time_key(real_time& ut, string *key) | |
76 | { | |
77 | char buf[32]; | |
78 | ceph_timespec ts = ceph::real_clock::to_ceph_timespec(ut); | |
79 | snprintf(buf, 32, "%011llu.%09u", (unsigned long long)ts.tv_sec, (unsigned int)ts.tv_nsec); | |
80 | *key = buf; | |
81 | } | |
82 | ||
83 | static void get_index_ver_key(cls_method_context_t hctx, uint64_t index_ver, string *key) | |
84 | { | |
85 | char buf[48]; | |
86 | snprintf(buf, sizeof(buf), "%011llu.%llu.%d", (unsigned long long)index_ver, | |
87 | (unsigned long long)cls_current_version(hctx), | |
88 | cls_current_subop_num(hctx)); | |
89 | *key = buf; | |
90 | } | |
91 | ||
92 | static void bi_log_prefix(string& key) | |
93 | { | |
94 | key = BI_PREFIX_CHAR; | |
95 | key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]); | |
96 | } | |
97 | ||
98 | static void bi_log_index_key(cls_method_context_t hctx, string& key, string& id, uint64_t index_ver) | |
99 | { | |
100 | bi_log_prefix(key); | |
101 | get_index_ver_key(hctx, index_ver, &id); | |
102 | key.append(id); | |
103 | } | |
104 | ||
105 | static int log_index_operation(cls_method_context_t hctx, cls_rgw_obj_key& obj_key, RGWModifyOp op, | |
106 | string& tag, real_time& timestamp, | |
107 | rgw_bucket_entry_ver& ver, RGWPendingState state, uint64_t index_ver, | |
31f18b77 | 108 | string& max_marker, uint16_t bilog_flags, string *owner, string *owner_display_name, rgw_zone_set *zones_trace) |
7c673cae FG |
109 | { |
110 | bufferlist bl; | |
111 | ||
112 | struct rgw_bi_log_entry entry; | |
113 | ||
114 | entry.object = obj_key.name; | |
115 | entry.instance = obj_key.instance; | |
116 | entry.timestamp = timestamp; | |
117 | entry.op = op; | |
118 | entry.ver = ver; | |
119 | entry.state = state; | |
120 | entry.index_ver = index_ver; | |
121 | entry.tag = tag; | |
122 | entry.bilog_flags = bilog_flags; | |
123 | if (owner) { | |
124 | entry.owner = *owner; | |
125 | } | |
126 | if (owner_display_name) { | |
127 | entry.owner_display_name = *owner_display_name; | |
128 | } | |
31f18b77 FG |
129 | if (zones_trace) { |
130 | entry.zones_trace = std::move(*zones_trace); | |
131 | } | |
7c673cae FG |
132 | |
133 | string key; | |
134 | bi_log_index_key(hctx, key, entry.id, index_ver); | |
135 | ||
136 | ::encode(entry, bl); | |
137 | ||
138 | if (entry.id > max_marker) | |
139 | max_marker = entry.id; | |
140 | ||
141 | return cls_cxx_map_set_val(hctx, key, &bl); | |
142 | } | |
143 | ||
144 | /* | |
145 | * read list of objects, skips objects in the ugly namespace | |
146 | */ | |
147 | static int get_obj_vals(cls_method_context_t hctx, const string& start, const string& filter_prefix, | |
c07f9fc5 | 148 | int num_entries, map<string, bufferlist> *pkeys, bool *pmore) |
7c673cae | 149 | { |
c07f9fc5 | 150 | int ret = cls_cxx_map_get_vals(hctx, start, filter_prefix, num_entries, pkeys, pmore); |
7c673cae FG |
151 | if (ret < 0) |
152 | return ret; | |
153 | ||
154 | if (pkeys->empty()) | |
155 | return 0; | |
156 | ||
1adf2230 | 157 | auto last_element = pkeys->rbegin(); |
7c673cae FG |
158 | if ((unsigned char)last_element->first[0] < BI_PREFIX_CHAR) { |
159 | /* nothing to see here, move along */ | |
160 | return 0; | |
161 | } | |
162 | ||
1adf2230 | 163 | auto first_element = pkeys->begin(); |
7c673cae FG |
164 | if ((unsigned char)first_element->first[0] > BI_PREFIX_CHAR) { |
165 | return 0; | |
166 | } | |
167 | ||
168 | /* let's rebuild the list, only keep entries we're interested in */ | |
1adf2230 AA |
169 | auto comp = [](const pair<string, bufferlist>& l, const string &r) { return l.first < r; }; |
170 | string new_start = {static_cast<char>(BI_PREFIX_CHAR + 1)}; | |
7c673cae | 171 | |
1adf2230 AA |
172 | auto lower = pkeys->lower_bound(string{static_cast<char>(BI_PREFIX_CHAR)}); |
173 | auto upper = std::lower_bound(lower, pkeys->end(), new_start, comp); | |
174 | pkeys->erase(lower, upper); | |
7c673cae FG |
175 | |
176 | if (num_entries == (int)pkeys->size()) | |
177 | return 0; | |
178 | ||
179 | map<string, bufferlist> new_keys; | |
7c673cae FG |
180 | |
181 | /* now get some more keys */ | |
c07f9fc5 | 182 | ret = cls_cxx_map_get_vals(hctx, new_start, filter_prefix, num_entries - pkeys->size(), &new_keys, pmore); |
7c673cae FG |
183 | if (ret < 0) |
184 | return ret; | |
185 | ||
1adf2230 AA |
186 | pkeys->insert(std::make_move_iterator(new_keys.begin()), |
187 | std::make_move_iterator(new_keys.end())); | |
7c673cae FG |
188 | return 0; |
189 | } | |
190 | ||
191 | /* | |
192 | * get a monotonically decreasing string representation. | |
193 | * For num = x, num = y, where x > y, str(x) < str(y) | |
194 | * Another property is that string size starts short and grows as num increases | |
195 | */ | |
196 | static void decreasing_str(uint64_t num, string *str) | |
197 | { | |
198 | char buf[32]; | |
199 | if (num < 0x10) { /* 16 */ | |
200 | snprintf(buf, sizeof(buf), "9%02lld", 15 - (long long)num); | |
201 | } else if (num < 0x100) { /* 256 */ | |
202 | snprintf(buf, sizeof(buf), "8%03lld", 255 - (long long)num); | |
203 | } else if (num < 0x1000) /* 4096 */ { | |
204 | snprintf(buf, sizeof(buf), "7%04lld", 4095 - (long long)num); | |
205 | } else if (num < 0x10000) /* 65536 */ { | |
206 | snprintf(buf, sizeof(buf), "6%05lld", 65535 - (long long)num); | |
207 | } else if (num < 0x100000000) /* 4G */ { | |
208 | snprintf(buf, sizeof(buf), "5%010lld", 0xFFFFFFFF - (long long)num); | |
209 | } else { | |
210 | snprintf(buf, sizeof(buf), "4%020lld", (long long)-num); | |
211 | } | |
212 | ||
213 | *str = buf; | |
214 | } | |
215 | ||
216 | /* | |
217 | * we now hold two different indexes for objects. The first one holds the list of objects in the | |
218 | * order that we want them to be listed. The second one only holds the objects instances (for | |
219 | * versioned objects), and they're not arranged in any particular order. | |
220 | * When listing objects we'll use the first index, when doing operations on the objects themselves | |
221 | * we'll use the second index. Note that regular objects only map to the first index anyway | |
222 | */ | |
223 | ||
224 | static void get_list_index_key(struct rgw_bucket_dir_entry& entry, string *index_key) | |
225 | { | |
226 | *index_key = entry.key.name; | |
227 | ||
228 | string ver_str; | |
229 | decreasing_str(entry.versioned_epoch, &ver_str); | |
230 | string instance_delim("\0i", 2); | |
231 | string ver_delim("\0v", 2); | |
232 | ||
233 | index_key->append(ver_delim); | |
234 | index_key->append(ver_str); | |
235 | index_key->append(instance_delim); | |
236 | index_key->append(entry.key.instance); | |
237 | } | |
238 | ||
239 | static void encode_obj_versioned_data_key(const cls_rgw_obj_key& key, string *index_key, bool append_delete_marker_suffix = false) | |
240 | { | |
241 | *index_key = BI_PREFIX_CHAR; | |
242 | index_key->append(bucket_index_prefixes[BI_BUCKET_OBJ_INSTANCE_INDEX]); | |
243 | index_key->append(key.name); | |
244 | string delim("\0i", 2); | |
245 | index_key->append(delim); | |
246 | index_key->append(key.instance); | |
247 | if (append_delete_marker_suffix) { | |
248 | string dm("\0d", 2); | |
249 | index_key->append(dm); | |
250 | } | |
251 | } | |
252 | ||
253 | static void encode_obj_index_key(const cls_rgw_obj_key& key, string *index_key) | |
254 | { | |
255 | if (key.instance.empty()) { | |
256 | *index_key = key.name; | |
257 | } else { | |
258 | encode_obj_versioned_data_key(key, index_key); | |
259 | } | |
260 | } | |
261 | ||
262 | static void encode_olh_data_key(const cls_rgw_obj_key& key, string *index_key) | |
263 | { | |
264 | *index_key = BI_PREFIX_CHAR; | |
265 | index_key->append(bucket_index_prefixes[BI_BUCKET_OLH_DATA_INDEX]); | |
266 | index_key->append(key.name); | |
267 | } | |
268 | ||
269 | template <class T> | |
270 | static int read_index_entry(cls_method_context_t hctx, string& name, T *entry); | |
271 | ||
272 | static int encode_list_index_key(cls_method_context_t hctx, const cls_rgw_obj_key& key, string *index_key) | |
273 | { | |
274 | if (key.instance.empty()) { | |
275 | *index_key = key.name; | |
276 | return 0; | |
277 | } | |
278 | ||
279 | string obj_index_key; | |
280 | encode_obj_index_key(key, &obj_index_key); | |
281 | ||
282 | rgw_bucket_dir_entry entry; | |
283 | ||
284 | int ret = read_index_entry(hctx, obj_index_key, &entry); | |
285 | if (ret == -ENOENT) { | |
286 | /* couldn't find the entry, set key value after the current object */ | |
287 | char buf[2] = { 0x1, 0 }; | |
288 | string s(buf); | |
289 | *index_key = key.name + s; | |
290 | return 0; | |
291 | } | |
292 | if (ret < 0) { | |
293 | CLS_LOG(1, "ERROR: encode_list_index_key(): cls_cxx_map_get_val returned %d\n", ret); | |
294 | return ret; | |
295 | } | |
296 | ||
297 | get_list_index_key(entry, index_key); | |
298 | ||
299 | return 0; | |
300 | } | |
301 | ||
302 | static void split_key(const string& key, list<string>& vals) | |
303 | { | |
304 | size_t pos = 0; | |
305 | const char *p = key.c_str(); | |
306 | while (pos < key.size()) { | |
307 | size_t len = strlen(p); | |
308 | vals.push_back(p); | |
309 | pos += len + 1; | |
310 | p += len + 1; | |
311 | } | |
312 | } | |
313 | ||
91327a77 AA |
314 | static string escape_str(const string& s) |
315 | { | |
316 | int len = escape_json_attr_len(s.c_str(), s.size()); | |
317 | char escaped[len]; | |
318 | escape_json_attr(s.c_str(), s.size(), escaped); | |
319 | return string(escaped); | |
320 | } | |
321 | ||
7c673cae FG |
322 | /* |
323 | * list index key structure: | |
324 | * | |
325 | * <obj name>\0[v<ver>\0i<instance id>] | |
326 | */ | |
91327a77 | 327 | static int decode_list_index_key(const string& index_key, cls_rgw_obj_key *key, uint64_t *ver) |
7c673cae FG |
328 | { |
329 | size_t len = strlen(index_key.c_str()); | |
330 | ||
331 | key->instance.clear(); | |
332 | *ver = 0; | |
333 | ||
334 | if (len == index_key.size()) { | |
335 | key->name = index_key; | |
91327a77 | 336 | return 0; |
7c673cae FG |
337 | } |
338 | ||
339 | list<string> vals; | |
340 | split_key(index_key, vals); | |
341 | ||
91327a77 AA |
342 | if (vals.empty()) { |
343 | CLS_LOG(0, "ERROR: %s(): bad index_key (%s): split_key() returned empty vals", __func__, escape_str(index_key).c_str()); | |
344 | return -EIO; | |
345 | } | |
7c673cae FG |
346 | |
347 | list<string>::iterator iter = vals.begin(); | |
348 | key->name = *iter; | |
349 | ++iter; | |
350 | ||
91327a77 AA |
351 | if (iter == vals.end()) { |
352 | CLS_LOG(0, "ERROR: %s(): bad index_key (%s): no vals", __func__, escape_str(index_key).c_str()); | |
353 | return -EIO; | |
354 | } | |
7c673cae FG |
355 | |
356 | for (; iter != vals.end(); ++iter) { | |
357 | string& val = *iter; | |
358 | if (val[0] == 'i') { | |
359 | key->instance = val.substr(1); | |
360 | } else if (val[0] == 'v') { | |
361 | string err; | |
362 | const char *s = val.c_str() + 1; | |
363 | *ver = strict_strtoll(s, 10, &err); | |
91327a77 AA |
364 | if (!err.empty()) { |
365 | CLS_LOG(0, "ERROR: %s(): bad index_key (%s): could not parse val (v=%s)", __func__, escape_str(index_key).c_str(), s); | |
366 | return -EIO; | |
367 | } | |
7c673cae FG |
368 | } |
369 | } | |
91327a77 AA |
370 | |
371 | return 0; | |
7c673cae FG |
372 | } |
373 | ||
374 | static int read_bucket_header(cls_method_context_t hctx, struct rgw_bucket_dir_header *header) | |
375 | { | |
376 | bufferlist bl; | |
377 | int rc = cls_cxx_map_read_header(hctx, &bl); | |
378 | if (rc < 0) | |
379 | return rc; | |
31f18b77 FG |
380 | |
381 | if (bl.length() == 0) { | |
382 | *header = rgw_bucket_dir_header(); | |
383 | return 0; | |
384 | } | |
7c673cae FG |
385 | bufferlist::iterator iter = bl.begin(); |
386 | try { | |
387 | ::decode(*header, iter); | |
388 | } catch (buffer::error& err) { | |
389 | CLS_LOG(1, "ERROR: read_bucket_header(): failed to decode header\n"); | |
390 | return -EIO; | |
391 | } | |
392 | ||
393 | return 0; | |
394 | } | |
395 | ||
396 | int rgw_bucket_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
397 | { | |
398 | bufferlist::iterator iter = in->begin(); | |
399 | ||
400 | struct rgw_cls_list_op op; | |
401 | try { | |
402 | ::decode(op, iter); | |
403 | } catch (buffer::error& err) { | |
404 | CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode request\n"); | |
405 | return -EINVAL; | |
406 | } | |
407 | ||
408 | struct rgw_cls_list_ret ret; | |
409 | struct rgw_bucket_dir& new_dir = ret.dir; | |
410 | int rc = read_bucket_header(hctx, &new_dir.header); | |
411 | if (rc < 0) { | |
412 | CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to read header\n"); | |
413 | return rc; | |
414 | } | |
415 | ||
416 | map<string, bufferlist> keys; | |
417 | std::map<string, bufferlist>::iterator kiter; | |
418 | string start_key; | |
419 | encode_list_index_key(hctx, op.start_obj, &start_key); | |
420 | bool done = false; | |
c07f9fc5 FG |
421 | uint32_t left_to_read = op.num_entries; |
422 | bool more; | |
7c673cae FG |
423 | |
424 | do { | |
c07f9fc5 | 425 | rc = get_obj_vals(hctx, start_key, op.filter_prefix, left_to_read, &keys, &more); |
7c673cae FG |
426 | if (rc < 0) |
427 | return rc; | |
428 | ||
429 | std::map<string, struct rgw_bucket_dir_entry>& m = new_dir.m; | |
430 | ||
431 | done = keys.empty(); | |
432 | ||
433 | for (kiter = keys.begin(); kiter != keys.end(); ++kiter) { | |
434 | struct rgw_bucket_dir_entry entry; | |
435 | ||
436 | if (!bi_is_objs_index(kiter->first)) { | |
437 | done = true; | |
438 | break; | |
439 | } | |
440 | ||
441 | bufferlist& entrybl = kiter->second; | |
442 | bufferlist::iterator eiter = entrybl.begin(); | |
443 | try { | |
444 | ::decode(entry, eiter); | |
445 | } catch (buffer::error& err) { | |
446 | CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode entry, key=%s\n", kiter->first.c_str()); | |
447 | return -EINVAL; | |
448 | } | |
449 | ||
450 | cls_rgw_obj_key key; | |
451 | uint64_t ver; | |
7c673cae FG |
452 | |
453 | start_key = kiter->first; | |
454 | CLS_LOG(20, "start_key=%s len=%zu", start_key.c_str(), start_key.size()); | |
455 | ||
91327a77 AA |
456 | int ret = decode_list_index_key(kiter->first, &key, &ver); |
457 | if (ret < 0) { | |
458 | CLS_LOG(0, "ERROR: failed to decode list index key (%s)\n", escape_str(kiter->first).c_str()); | |
459 | continue; | |
460 | } | |
461 | ||
7c673cae FG |
462 | if (!entry.is_valid()) { |
463 | CLS_LOG(20, "entry %s[%s] is not valid\n", key.name.c_str(), key.instance.c_str()); | |
464 | continue; | |
465 | } | |
3efd9988 FG |
466 | |
467 | // filter out noncurrent versions, delete markers, and initial marker | |
468 | if (!op.list_versions && (!entry.is_visible() || op.start_obj.name == key.name)) { | |
7c673cae FG |
469 | CLS_LOG(20, "entry %s[%s] is not visible\n", key.name.c_str(), key.instance.c_str()); |
470 | continue; | |
471 | } | |
472 | if (m.size() < op.num_entries) { | |
473 | m[kiter->first] = entry; | |
474 | } | |
475 | left_to_read--; | |
476 | ||
477 | CLS_LOG(20, "got entry %s[%s] m.size()=%d\n", key.name.c_str(), key.instance.c_str(), (int)m.size()); | |
478 | } | |
479 | } while (left_to_read > 0 && !done); | |
480 | ||
c07f9fc5 | 481 | ret.is_truncated = more && !done; |
7c673cae FG |
482 | |
483 | ::encode(ret, *out); | |
484 | return 0; | |
485 | } | |
486 | ||
487 | static int check_index(cls_method_context_t hctx, struct rgw_bucket_dir_header *existing_header, struct rgw_bucket_dir_header *calc_header) | |
488 | { | |
489 | int rc = read_bucket_header(hctx, existing_header); | |
490 | if (rc < 0) { | |
491 | CLS_LOG(1, "ERROR: check_index(): failed to read header\n"); | |
492 | return rc; | |
493 | } | |
494 | ||
495 | calc_header->tag_timeout = existing_header->tag_timeout; | |
496 | calc_header->ver = existing_header->ver; | |
497 | ||
498 | map<string, bufferlist> keys; | |
499 | string start_obj; | |
500 | string filter_prefix; | |
501 | ||
502 | #define CHECK_CHUNK_SIZE 1000 | |
503 | bool done = false; | |
c07f9fc5 | 504 | bool more; |
7c673cae FG |
505 | |
506 | do { | |
c07f9fc5 | 507 | rc = get_obj_vals(hctx, start_obj, filter_prefix, CHECK_CHUNK_SIZE, &keys, &more); |
7c673cae FG |
508 | if (rc < 0) |
509 | return rc; | |
510 | ||
511 | std::map<string, bufferlist>::iterator kiter = keys.begin(); | |
512 | for (; kiter != keys.end(); ++kiter) { | |
513 | if (!bi_is_objs_index(kiter->first)) { | |
514 | done = true; | |
515 | break; | |
516 | } | |
517 | ||
518 | struct rgw_bucket_dir_entry entry; | |
519 | bufferlist::iterator eiter = kiter->second.begin(); | |
520 | try { | |
521 | ::decode(entry, eiter); | |
522 | } catch (buffer::error& err) { | |
523 | CLS_LOG(1, "ERROR: rgw_bucket_list(): failed to decode entry, key=%s\n", kiter->first.c_str()); | |
524 | return -EIO; | |
525 | } | |
526 | struct rgw_bucket_category_stats& stats = calc_header->stats[entry.meta.category]; | |
527 | stats.num_entries++; | |
528 | stats.total_size += entry.meta.accounted_size; | |
529 | stats.total_size_rounded += cls_rgw_get_rounded_size(entry.meta.accounted_size); | |
530 | stats.actual_size += entry.meta.size; | |
531 | ||
532 | start_obj = kiter->first; | |
533 | } | |
534 | } while (keys.size() == CHECK_CHUNK_SIZE && !done); | |
535 | ||
536 | return 0; | |
537 | } | |
538 | ||
539 | int rgw_bucket_check_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
540 | { | |
541 | struct rgw_cls_check_index_ret ret; | |
542 | ||
543 | int rc = check_index(hctx, &ret.existing_header, &ret.calculated_header); | |
544 | if (rc < 0) | |
545 | return rc; | |
546 | ||
547 | ::encode(ret, *out); | |
548 | ||
549 | return 0; | |
550 | } | |
551 | ||
552 | static int write_bucket_header(cls_method_context_t hctx, struct rgw_bucket_dir_header *header) | |
553 | { | |
554 | header->ver++; | |
555 | ||
556 | bufferlist header_bl; | |
557 | ::encode(*header, header_bl); | |
558 | return cls_cxx_map_write_header(hctx, &header_bl); | |
559 | } | |
560 | ||
561 | ||
562 | int rgw_bucket_rebuild_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
563 | { | |
564 | struct rgw_bucket_dir_header existing_header; | |
565 | struct rgw_bucket_dir_header calc_header; | |
566 | int rc = check_index(hctx, &existing_header, &calc_header); | |
567 | if (rc < 0) | |
568 | return rc; | |
569 | ||
570 | return write_bucket_header(hctx, &calc_header); | |
571 | } | |
572 | ||
573 | int rgw_bucket_update_stats(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
574 | { | |
575 | // decode request | |
576 | rgw_cls_bucket_update_stats_op op; | |
577 | auto iter = in->begin(); | |
578 | try { | |
579 | ::decode(op, iter); | |
580 | } catch (buffer::error& err) { | |
581 | CLS_LOG(1, "ERROR: %s(): failed to decode request\n", __func__); | |
582 | return -EINVAL; | |
583 | } | |
584 | ||
585 | struct rgw_bucket_dir_header header; | |
586 | int rc = read_bucket_header(hctx, &header); | |
587 | if (rc < 0) { | |
588 | CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__); | |
589 | return rc; | |
590 | } | |
591 | ||
592 | for (auto& s : op.stats) { | |
593 | auto& dest = header.stats[s.first]; | |
594 | if (op.absolute) { | |
595 | dest = s.second; | |
596 | } else { | |
597 | dest.total_size += s.second.total_size; | |
598 | dest.total_size_rounded += s.second.total_size_rounded; | |
599 | dest.num_entries += s.second.num_entries; | |
91327a77 | 600 | dest.actual_size += s.second.actual_size; |
7c673cae FG |
601 | } |
602 | } | |
603 | ||
604 | return write_bucket_header(hctx, &header); | |
605 | } | |
606 | ||
607 | int rgw_bucket_init_index(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
608 | { | |
609 | bufferlist::iterator iter; | |
610 | ||
611 | bufferlist header_bl; | |
612 | int rc = cls_cxx_map_read_header(hctx, &header_bl); | |
613 | if (rc < 0) { | |
614 | switch (rc) { | |
615 | case -ENODATA: | |
616 | case -ENOENT: | |
617 | break; | |
618 | default: | |
619 | return rc; | |
620 | } | |
621 | } | |
622 | ||
623 | if (header_bl.length() != 0) { | |
624 | CLS_LOG(1, "ERROR: index already initialized\n"); | |
625 | return -EINVAL; | |
626 | } | |
627 | ||
628 | rgw_bucket_dir dir; | |
629 | ||
630 | return write_bucket_header(hctx, &dir.header); | |
631 | } | |
632 | ||
633 | int rgw_bucket_set_tag_timeout(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
634 | { | |
635 | // decode request | |
636 | rgw_cls_tag_timeout_op op; | |
637 | bufferlist::iterator iter = in->begin(); | |
638 | try { | |
639 | ::decode(op, iter); | |
640 | } catch (buffer::error& err) { | |
641 | CLS_LOG(1, "ERROR: rgw_bucket_set_tag_timeout(): failed to decode request\n"); | |
642 | return -EINVAL; | |
643 | } | |
644 | ||
645 | struct rgw_bucket_dir_header header; | |
646 | int rc = read_bucket_header(hctx, &header); | |
647 | if (rc < 0) { | |
648 | CLS_LOG(1, "ERROR: rgw_bucket_set_tag_timeout(): failed to read header\n"); | |
649 | return rc; | |
650 | } | |
651 | ||
652 | header.tag_timeout = op.tag_timeout; | |
653 | ||
654 | return write_bucket_header(hctx, &header); | |
655 | } | |
656 | ||
657 | static int read_key_entry(cls_method_context_t hctx, cls_rgw_obj_key& key, string *idx, struct rgw_bucket_dir_entry *entry, | |
658 | bool special_delete_marker_name = false); | |
659 | ||
660 | int rgw_bucket_prepare_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
661 | { | |
662 | // decode request | |
663 | rgw_cls_obj_prepare_op op; | |
664 | bufferlist::iterator iter = in->begin(); | |
665 | try { | |
666 | ::decode(op, iter); | |
667 | } catch (buffer::error& err) { | |
668 | CLS_LOG(1, "ERROR: rgw_bucket_prepare_op(): failed to decode request\n"); | |
669 | return -EINVAL; | |
670 | } | |
671 | ||
672 | if (op.tag.empty()) { | |
673 | CLS_LOG(1, "ERROR: tag is empty\n"); | |
674 | return -EINVAL; | |
675 | } | |
676 | ||
677 | CLS_LOG(1, "rgw_bucket_prepare_op(): request: op=%d name=%s instance=%s tag=%s\n", | |
678 | op.op, op.key.name.c_str(), op.key.instance.c_str(), op.tag.c_str()); | |
679 | ||
680 | // get on-disk state | |
681 | string idx; | |
682 | ||
683 | struct rgw_bucket_dir_entry entry; | |
684 | int rc = read_key_entry(hctx, op.key, &idx, &entry); | |
685 | if (rc < 0 && rc != -ENOENT) | |
686 | return rc; | |
687 | ||
688 | bool noent = (rc == -ENOENT); | |
689 | ||
690 | rc = 0; | |
691 | ||
692 | if (noent) { // no entry, initialize fields | |
693 | entry.key = op.key; | |
694 | entry.ver = rgw_bucket_entry_ver(); | |
695 | entry.exists = false; | |
696 | entry.locator = op.locator; | |
697 | } | |
698 | ||
699 | // fill in proper state | |
700 | struct rgw_bucket_pending_info info; | |
701 | info.timestamp = real_clock::now(); | |
702 | info.state = CLS_RGW_STATE_PENDING_MODIFY; | |
703 | info.op = op.op; | |
704 | entry.pending_map.insert(pair<string, rgw_bucket_pending_info>(op.tag, info)); | |
705 | ||
706 | struct rgw_bucket_dir_header header; | |
707 | rc = read_bucket_header(hctx, &header); | |
708 | if (rc < 0) { | |
709 | CLS_LOG(1, "ERROR: rgw_bucket_prepare_op(): failed to read header\n"); | |
710 | return rc; | |
711 | } | |
712 | ||
c07f9fc5 | 713 | if (op.log_op && !header.syncstopped) { |
7c673cae | 714 | rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, |
31f18b77 | 715 | entry.ver, info.state, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace); |
7c673cae FG |
716 | if (rc < 0) |
717 | return rc; | |
718 | } | |
719 | ||
720 | // write out new key to disk | |
721 | bufferlist info_bl; | |
722 | ::encode(entry, info_bl); | |
723 | rc = cls_cxx_map_set_val(hctx, idx, &info_bl); | |
724 | if (rc < 0) | |
725 | return rc; | |
726 | ||
727 | return write_bucket_header(hctx, &header); | |
728 | } | |
729 | ||
730 | static void unaccount_entry(struct rgw_bucket_dir_header& header, struct rgw_bucket_dir_entry& entry) | |
731 | { | |
732 | struct rgw_bucket_category_stats& stats = header.stats[entry.meta.category]; | |
733 | stats.num_entries--; | |
734 | stats.total_size -= entry.meta.accounted_size; | |
735 | stats.total_size_rounded -= cls_rgw_get_rounded_size(entry.meta.accounted_size); | |
736 | stats.actual_size -= entry.meta.size; | |
737 | } | |
738 | ||
739 | static void log_entry(const char *func, const char *str, struct rgw_bucket_dir_entry *entry) | |
740 | { | |
741 | CLS_LOG(1, "%s(): %s: ver=%ld:%llu name=%s instance=%s locator=%s\n", func, str, | |
742 | (long)entry->ver.pool, (unsigned long long)entry->ver.epoch, | |
743 | entry->key.name.c_str(), entry->key.instance.c_str(), entry->locator.c_str()); | |
744 | } | |
745 | ||
746 | static void log_entry(const char *func, const char *str, struct rgw_bucket_olh_entry *entry) | |
747 | { | |
748 | CLS_LOG(1, "%s(): %s: epoch=%llu name=%s instance=%s tag=%s\n", func, str, | |
749 | (unsigned long long)entry->epoch, entry->key.name.c_str(), entry->key.instance.c_str(), | |
750 | entry->tag.c_str()); | |
751 | } | |
752 | ||
753 | template <class T> | |
754 | static int read_index_entry(cls_method_context_t hctx, string& name, T *entry) | |
755 | { | |
756 | bufferlist current_entry; | |
757 | int rc = cls_cxx_map_get_val(hctx, name, ¤t_entry); | |
758 | if (rc < 0) { | |
759 | return rc; | |
760 | } | |
761 | ||
762 | bufferlist::iterator cur_iter = current_entry.begin(); | |
763 | try { | |
764 | ::decode(*entry, cur_iter); | |
765 | } catch (buffer::error& err) { | |
766 | CLS_LOG(1, "ERROR: read_index_entry(): failed to decode entry\n"); | |
767 | return -EIO; | |
768 | } | |
769 | ||
770 | log_entry(__func__, "existing entry", entry); | |
771 | return 0; | |
772 | } | |
773 | ||
774 | static int read_key_entry(cls_method_context_t hctx, cls_rgw_obj_key& key, string *idx, struct rgw_bucket_dir_entry *entry, | |
775 | bool special_delete_marker_name) | |
776 | { | |
777 | encode_obj_index_key(key, idx); | |
778 | int rc = read_index_entry(hctx, *idx, entry); | |
779 | if (rc < 0) { | |
780 | return rc; | |
781 | } | |
782 | ||
783 | if (key.instance.empty() && | |
784 | entry->flags & RGW_BUCKET_DIRENT_FLAG_VER_MARKER) { | |
785 | /* we only do it where key.instance is empty. In this case the delete marker will have a | |
786 | * separate entry in the index to avoid collisions with the actual object, as it's mutable | |
787 | */ | |
788 | if (special_delete_marker_name) { | |
789 | encode_obj_versioned_data_key(key, idx, true); | |
790 | rc = read_index_entry(hctx, *idx, entry); | |
791 | if (rc == 0) { | |
792 | return 0; | |
793 | } | |
794 | } | |
795 | encode_obj_versioned_data_key(key, idx); | |
796 | rc = read_index_entry(hctx, *idx, entry); | |
797 | if (rc < 0) { | |
798 | *entry = rgw_bucket_dir_entry(); /* need to reset entry because we initialized it earlier */ | |
799 | return rc; | |
800 | } | |
801 | } | |
802 | ||
803 | return 0; | |
804 | } | |
805 | ||
806 | int rgw_bucket_complete_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
807 | { | |
808 | // decode request | |
809 | rgw_cls_obj_complete_op op; | |
810 | bufferlist::iterator iter = in->begin(); | |
811 | try { | |
812 | ::decode(op, iter); | |
813 | } catch (buffer::error& err) { | |
814 | CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to decode request\n"); | |
815 | return -EINVAL; | |
816 | } | |
817 | CLS_LOG(1, "rgw_bucket_complete_op(): request: op=%d name=%s instance=%s ver=%lu:%llu tag=%s\n", | |
818 | op.op, op.key.name.c_str(), op.key.instance.c_str(), | |
819 | (unsigned long)op.ver.pool, (unsigned long long)op.ver.epoch, | |
820 | op.tag.c_str()); | |
821 | ||
822 | struct rgw_bucket_dir_header header; | |
823 | int rc = read_bucket_header(hctx, &header); | |
824 | if (rc < 0) { | |
825 | CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n"); | |
826 | return -EINVAL; | |
827 | } | |
828 | ||
829 | struct rgw_bucket_dir_entry entry; | |
830 | bool ondisk = true; | |
831 | ||
832 | string idx; | |
833 | rc = read_key_entry(hctx, op.key, &idx, &entry); | |
834 | if (rc == -ENOENT) { | |
835 | entry.key = op.key; | |
836 | entry.ver = op.ver; | |
837 | entry.meta = op.meta; | |
838 | entry.locator = op.locator; | |
839 | ondisk = false; | |
840 | } else if (rc < 0) { | |
841 | return rc; | |
842 | } | |
843 | ||
844 | entry.index_ver = header.ver; | |
845 | entry.flags = (entry.key.instance.empty() ? 0 : RGW_BUCKET_DIRENT_FLAG_VER); /* resetting entry flags, entry might have been previously a delete marker */ | |
846 | ||
847 | if (op.tag.size()) { | |
848 | map<string, struct rgw_bucket_pending_info>::iterator pinter = entry.pending_map.find(op.tag); | |
849 | if (pinter == entry.pending_map.end()) { | |
850 | CLS_LOG(1, "ERROR: couldn't find tag for pending operation\n"); | |
851 | return -EINVAL; | |
852 | } | |
853 | entry.pending_map.erase(pinter); | |
854 | } | |
855 | ||
856 | bool cancel = false; | |
857 | bufferlist update_bl; | |
858 | ||
859 | if (op.tag.size() && op.op == CLS_RGW_OP_CANCEL) { | |
860 | CLS_LOG(1, "rgw_bucket_complete_op(): cancel requested\n"); | |
861 | cancel = true; | |
862 | } else if (op.ver.pool == entry.ver.pool && | |
863 | op.ver.epoch && op.ver.epoch <= entry.ver.epoch) { | |
864 | CLS_LOG(1, "rgw_bucket_complete_op(): skipping request, old epoch\n"); | |
865 | cancel = true; | |
866 | } | |
867 | ||
868 | bufferlist op_bl; | |
869 | if (cancel) { | |
c07f9fc5 | 870 | if (op.log_op && !header.syncstopped) { |
7c673cae | 871 | rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, entry.ver, |
31f18b77 | 872 | CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace); |
7c673cae FG |
873 | if (rc < 0) |
874 | return rc; | |
875 | } | |
876 | ||
877 | if (op.tag.size()) { | |
878 | bufferlist new_key_bl; | |
879 | ::encode(entry, new_key_bl); | |
880 | return cls_cxx_map_set_val(hctx, idx, &new_key_bl); | |
881 | } else { | |
882 | return 0; | |
883 | } | |
884 | } | |
885 | ||
886 | if (entry.exists) { | |
887 | unaccount_entry(header, entry); | |
888 | } | |
889 | ||
890 | entry.ver = op.ver; | |
891 | switch ((int)op.op) { | |
892 | case CLS_RGW_OP_DEL: | |
893 | entry.meta = op.meta; | |
894 | if (ondisk) { | |
895 | if (!entry.pending_map.size()) { | |
896 | int ret = cls_cxx_map_remove_key(hctx, idx); | |
897 | if (ret < 0) | |
898 | return ret; | |
899 | } else { | |
900 | entry.exists = false; | |
901 | bufferlist new_key_bl; | |
902 | ::encode(entry, new_key_bl); | |
903 | int ret = cls_cxx_map_set_val(hctx, idx, &new_key_bl); | |
904 | if (ret < 0) | |
905 | return ret; | |
906 | } | |
907 | } else { | |
908 | return -ENOENT; | |
909 | } | |
910 | break; | |
911 | case CLS_RGW_OP_ADD: | |
912 | { | |
913 | struct rgw_bucket_dir_entry_meta& meta = op.meta; | |
914 | struct rgw_bucket_category_stats& stats = header.stats[meta.category]; | |
915 | entry.meta = meta; | |
916 | entry.key = op.key; | |
917 | entry.exists = true; | |
918 | entry.tag = op.tag; | |
919 | stats.num_entries++; | |
920 | stats.total_size += meta.accounted_size; | |
921 | stats.total_size_rounded += cls_rgw_get_rounded_size(meta.accounted_size); | |
922 | stats.actual_size += meta.size; | |
923 | bufferlist new_key_bl; | |
924 | ::encode(entry, new_key_bl); | |
925 | int ret = cls_cxx_map_set_val(hctx, idx, &new_key_bl); | |
926 | if (ret < 0) | |
927 | return ret; | |
928 | } | |
929 | break; | |
930 | } | |
931 | ||
c07f9fc5 | 932 | if (op.log_op && !header.syncstopped) { |
7c673cae | 933 | rc = log_index_operation(hctx, op.key, op.op, op.tag, entry.meta.mtime, entry.ver, |
31f18b77 | 934 | CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace); |
7c673cae FG |
935 | if (rc < 0) |
936 | return rc; | |
937 | } | |
938 | ||
939 | list<cls_rgw_obj_key>::iterator remove_iter; | |
940 | CLS_LOG(20, "rgw_bucket_complete_op(): remove_objs.size()=%d\n", (int)op.remove_objs.size()); | |
941 | for (remove_iter = op.remove_objs.begin(); remove_iter != op.remove_objs.end(); ++remove_iter) { | |
942 | cls_rgw_obj_key& remove_key = *remove_iter; | |
943 | CLS_LOG(1, "rgw_bucket_complete_op(): removing entries, read_index_entry name=%s instance=%s\n", | |
944 | remove_key.name.c_str(), remove_key.instance.c_str()); | |
945 | struct rgw_bucket_dir_entry remove_entry; | |
946 | string k; | |
947 | int ret = read_key_entry(hctx, remove_key, &k, &remove_entry); | |
948 | if (ret < 0) { | |
949 | CLS_LOG(1, "rgw_bucket_complete_op(): removing entries, read_index_entry name=%s instance=%s ret=%d\n", | |
950 | remove_key.name.c_str(), remove_key.instance.c_str(), ret); | |
951 | continue; | |
952 | } | |
953 | CLS_LOG(0, "rgw_bucket_complete_op(): entry.name=%s entry.instance=%s entry.meta.category=%d\n", | |
954 | remove_entry.key.name.c_str(), remove_entry.key.instance.c_str(), remove_entry.meta.category); | |
955 | unaccount_entry(header, remove_entry); | |
956 | ||
c07f9fc5 | 957 | if (op.log_op && !header.syncstopped) { |
3efd9988 | 958 | ++header.ver; // increment index version, or we'll overwrite keys previously written |
7c673cae | 959 | rc = log_index_operation(hctx, remove_key, CLS_RGW_OP_DEL, op.tag, remove_entry.meta.mtime, |
31f18b77 | 960 | remove_entry.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags, NULL, NULL, &op.zones_trace); |
7c673cae FG |
961 | if (rc < 0) |
962 | continue; | |
963 | } | |
964 | ||
965 | ret = cls_cxx_map_remove_key(hctx, k); | |
966 | if (ret < 0) { | |
967 | CLS_LOG(1, "rgw_bucket_complete_op(): cls_cxx_map_remove_key, failed to remove entry, name=%s instance=%s read_index_entry ret=%d\n", remove_key.name.c_str(), remove_key.instance.c_str(), rc); | |
968 | continue; | |
969 | } | |
970 | } | |
971 | ||
972 | return write_bucket_header(hctx, &header); | |
973 | } | |
974 | ||
975 | template <class T> | |
976 | static int write_entry(cls_method_context_t hctx, T& entry, const string& key) | |
977 | { | |
978 | bufferlist bl; | |
979 | ::encode(entry, bl); | |
980 | return cls_cxx_map_set_val(hctx, key, &bl); | |
981 | } | |
982 | ||
983 | static int read_olh(cls_method_context_t hctx,cls_rgw_obj_key& obj_key, struct rgw_bucket_olh_entry *olh_data_entry, string *index_key, bool *found) | |
984 | { | |
985 | cls_rgw_obj_key olh_key; | |
986 | olh_key.name = obj_key.name; | |
987 | ||
988 | encode_olh_data_key(olh_key, index_key); | |
989 | int ret = read_index_entry(hctx, *index_key, olh_data_entry); | |
990 | if (ret < 0 && ret != -ENOENT) { | |
991 | CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_key.name.c_str(), ret); | |
992 | return ret; | |
993 | } | |
994 | if (found) { | |
995 | *found = (ret != -ENOENT); | |
996 | } | |
997 | return 0; | |
998 | } | |
999 | ||
1000 | static void update_olh_log(struct rgw_bucket_olh_entry& olh_data_entry, OLHLogOp op, const string& op_tag, | |
1001 | cls_rgw_obj_key& key, bool delete_marker, uint64_t epoch) | |
1002 | { | |
1003 | vector<rgw_bucket_olh_log_entry>& log = olh_data_entry.pending_log[olh_data_entry.epoch]; | |
1004 | rgw_bucket_olh_log_entry log_entry; | |
1005 | log_entry.epoch = epoch; | |
1006 | log_entry.op = op; | |
1007 | log_entry.op_tag = op_tag; | |
1008 | log_entry.key = key; | |
1009 | log_entry.delete_marker = delete_marker; | |
1010 | log.push_back(log_entry); | |
1011 | } | |
1012 | ||
7c673cae FG |
1013 | static int write_obj_instance_entry(cls_method_context_t hctx, struct rgw_bucket_dir_entry& instance_entry, const string& instance_idx) |
1014 | { | |
1015 | CLS_LOG(20, "write_entry() instance=%s idx=%s flags=%d", escape_str(instance_entry.key.instance).c_str(), instance_idx.c_str(), instance_entry.flags); | |
1016 | /* write the instance entry */ | |
1017 | int ret = write_entry(hctx, instance_entry, instance_idx); | |
1018 | if (ret < 0) { | |
1019 | CLS_LOG(0, "ERROR: write_entry() instance_key=%s ret=%d", escape_str(instance_idx).c_str(), ret); | |
1020 | return ret; | |
1021 | } | |
1022 | return 0; | |
1023 | } | |
1024 | ||
1025 | /* | |
1026 | * write object instance entry, and if needed also the list entry | |
1027 | */ | |
1028 | static int write_obj_entries(cls_method_context_t hctx, struct rgw_bucket_dir_entry& instance_entry, const string& instance_idx) | |
1029 | { | |
1030 | int ret = write_obj_instance_entry(hctx, instance_entry, instance_idx); | |
1031 | if (ret < 0) { | |
1032 | return ret; | |
1033 | } | |
1034 | string instance_list_idx; | |
1035 | get_list_index_key(instance_entry, &instance_list_idx); | |
1036 | ||
1037 | if (instance_idx != instance_list_idx) { | |
1038 | CLS_LOG(20, "write_entry() idx=%s flags=%d", escape_str(instance_list_idx).c_str(), instance_entry.flags); | |
1039 | /* write a new list entry for the object instance */ | |
1040 | ret = write_entry(hctx, instance_entry, instance_list_idx); | |
1041 | if (ret < 0) { | |
1042 | CLS_LOG(0, "ERROR: write_entry() instance=%s instance_list_idx=%s ret=%d", instance_entry.key.instance.c_str(), instance_list_idx.c_str(), ret); | |
1043 | return ret; | |
1044 | } | |
1045 | } | |
1046 | return 0; | |
1047 | } | |
1048 | ||
1049 | ||
1050 | class BIVerObjEntry { | |
1051 | cls_method_context_t hctx; | |
1052 | cls_rgw_obj_key key; | |
1053 | string instance_idx; | |
1054 | ||
1055 | struct rgw_bucket_dir_entry instance_entry; | |
1056 | ||
1057 | bool initialized; | |
1058 | ||
1059 | public: | |
1060 | BIVerObjEntry(cls_method_context_t& _hctx, const cls_rgw_obj_key& _key) : hctx(_hctx), key(_key), initialized(false) { | |
1061 | } | |
1062 | ||
1063 | int init(bool check_delete_marker = true) { | |
1064 | int ret = read_key_entry(hctx, key, &instance_idx, &instance_entry, | |
1065 | check_delete_marker && key.instance.empty()); /* this is potentially a delete marker, for null objects we | |
1066 | keep separate instance entry for the delete markers */ | |
1067 | ||
1068 | if (ret < 0) { | |
1069 | CLS_LOG(0, "ERROR: read_key_entry() idx=%s ret=%d", instance_idx.c_str(), ret); | |
1070 | return ret; | |
1071 | } | |
1072 | initialized = true; | |
1073 | CLS_LOG(20, "read instance_entry key.name=%s key.instance=%s flags=%d", instance_entry.key.name.c_str(), instance_entry.key.instance.c_str(), instance_entry.flags); | |
1074 | return 0; | |
1075 | } | |
1076 | ||
1077 | rgw_bucket_dir_entry& get_dir_entry() { | |
1078 | return instance_entry; | |
1079 | } | |
1080 | ||
1081 | void init_as_delete_marker(rgw_bucket_dir_entry_meta& meta) { | |
1082 | /* a deletion marker, need to initialize it, there's no instance entry for it yet */ | |
1083 | instance_entry.key = key; | |
1084 | instance_entry.flags = RGW_BUCKET_DIRENT_FLAG_DELETE_MARKER; | |
1085 | instance_entry.meta = meta; | |
1086 | instance_entry.tag = "delete-marker"; | |
1087 | ||
1088 | initialized = true; | |
1089 | } | |
1090 | ||
1091 | void set_epoch(uint64_t epoch) { | |
1092 | instance_entry.versioned_epoch = epoch; | |
1093 | } | |
1094 | ||
1095 | int unlink_list_entry() { | |
1096 | string list_idx; | |
1097 | /* this instance has a previous list entry, remove that entry */ | |
1098 | get_list_index_key(instance_entry, &list_idx); | |
1099 | CLS_LOG(20, "unlink_list_entry() list_idx=%s", escape_str(list_idx).c_str()); | |
1100 | int ret = cls_cxx_map_remove_key(hctx, list_idx); | |
1101 | if (ret < 0) { | |
1102 | CLS_LOG(0, "ERROR: cls_cxx_map_remove_key() list_idx=%s ret=%d", list_idx.c_str(), ret); | |
1103 | return ret; | |
1104 | } | |
1105 | return 0; | |
1106 | } | |
1107 | ||
1108 | int unlink() { | |
1109 | /* remove the instance entry */ | |
1110 | CLS_LOG(20, "unlink() idx=%s", escape_str(instance_idx).c_str()); | |
1111 | int ret = cls_cxx_map_remove_key(hctx, instance_idx); | |
1112 | if (ret < 0) { | |
1113 | CLS_LOG(0, "ERROR: cls_cxx_map_remove_key() instance_idx=%s ret=%d", instance_idx.c_str(), ret); | |
1114 | return ret; | |
1115 | } | |
1116 | return 0; | |
1117 | } | |
1118 | ||
1119 | int write_entries(uint64_t flags_set, uint64_t flags_reset) { | |
1120 | if (!initialized) { | |
1121 | int ret = init(); | |
1122 | if (ret < 0) { | |
1123 | return ret; | |
1124 | } | |
1125 | } | |
1126 | instance_entry.flags &= ~flags_reset; | |
1127 | instance_entry.flags |= flags_set; | |
1128 | ||
1129 | /* write the instance and list entries */ | |
1130 | bool special_delete_marker_key = (instance_entry.is_delete_marker() && instance_entry.key.instance.empty()); | |
1131 | encode_obj_versioned_data_key(key, &instance_idx, special_delete_marker_key); | |
1132 | int ret = write_obj_entries(hctx, instance_entry, instance_idx); | |
1133 | if (ret < 0) { | |
1134 | CLS_LOG(0, "ERROR: write_obj_entries() instance_idx=%s ret=%d", instance_idx.c_str(), ret); | |
1135 | return ret; | |
1136 | } | |
1137 | ||
1138 | return 0; | |
1139 | } | |
1140 | ||
1141 | int write(uint64_t epoch, bool current) { | |
1142 | if (instance_entry.versioned_epoch > 0) { | |
1143 | CLS_LOG(20, "%s(): instance_entry.versioned_epoch=%d epoch=%d", __func__, (int)instance_entry.versioned_epoch, (int)epoch); | |
1144 | /* this instance has a previous list entry, remove that entry */ | |
1145 | int ret = unlink_list_entry(); | |
1146 | if (ret < 0) { | |
1147 | return ret; | |
1148 | } | |
1149 | } | |
1150 | ||
1151 | uint64_t flags = RGW_BUCKET_DIRENT_FLAG_VER; | |
1152 | if (current) { | |
1153 | flags |= RGW_BUCKET_DIRENT_FLAG_CURRENT; | |
1154 | } | |
1155 | ||
1156 | instance_entry.versioned_epoch = epoch; | |
1157 | return write_entries(flags, 0); | |
1158 | } | |
1159 | ||
1160 | int demote_current() { | |
1161 | return write_entries(0, RGW_BUCKET_DIRENT_FLAG_CURRENT); | |
1162 | } | |
1163 | ||
1164 | bool is_delete_marker() { | |
1165 | return instance_entry.is_delete_marker(); | |
1166 | } | |
1167 | ||
1168 | int find_next_key(cls_rgw_obj_key *next_key, bool *found) { | |
1169 | string list_idx; | |
1170 | /* this instance has a previous list entry, remove that entry */ | |
1171 | get_list_index_key(instance_entry, &list_idx); | |
1172 | /* this is the current head, need to update! */ | |
1173 | map<string, bufferlist> keys; | |
c07f9fc5 | 1174 | bool more; |
7c673cae FG |
1175 | string filter = key.name; /* list key starts with key name, filter it to avoid a case where we cross to |
1176 | different namespace */ | |
c07f9fc5 | 1177 | int ret = cls_cxx_map_get_vals(hctx, list_idx, filter, 1, &keys, &more); |
7c673cae FG |
1178 | if (ret < 0) { |
1179 | return ret; | |
1180 | } | |
1181 | ||
1182 | if (keys.size() < 1) { | |
1183 | *found = false; | |
1184 | return 0; | |
1185 | } | |
1186 | ||
1187 | rgw_bucket_dir_entry next_entry; | |
1188 | ||
1189 | map<string, bufferlist>::reverse_iterator last = keys.rbegin(); | |
1190 | try { | |
1191 | bufferlist::iterator iter = last->second.begin(); | |
1192 | ::decode(next_entry, iter); | |
1193 | } catch (buffer::error& err) { | |
1194 | CLS_LOG(0, "ERROR; failed to decode entry: %s", last->first.c_str()); | |
1195 | return -EIO; | |
1196 | } | |
1197 | ||
1198 | *found = (key.name == next_entry.key.name); | |
1199 | if (*found) { | |
1200 | *next_key = next_entry.key; | |
1201 | } | |
1202 | ||
1203 | return 0; | |
1204 | } | |
1205 | ||
1206 | real_time mtime() { | |
1207 | return instance_entry.meta.mtime; | |
1208 | } | |
1209 | }; | |
1210 | ||
1211 | ||
1212 | class BIOLHEntry { | |
1213 | cls_method_context_t hctx; | |
1214 | cls_rgw_obj_key key; | |
1215 | ||
1216 | string olh_data_idx; | |
1217 | struct rgw_bucket_olh_entry olh_data_entry; | |
1218 | ||
1219 | bool initialized; | |
1220 | public: | |
1221 | BIOLHEntry(cls_method_context_t& _hctx, const cls_rgw_obj_key& _key) : hctx(_hctx), key(_key), initialized(false) { } | |
1222 | ||
1223 | int init(bool *exists) { | |
1224 | /* read olh */ | |
1225 | int ret = read_olh(hctx, key, &olh_data_entry, &olh_data_idx, exists); | |
1226 | if (ret < 0) { | |
1227 | return ret; | |
1228 | } | |
1229 | ||
1230 | initialized = true; | |
1231 | return 0; | |
1232 | } | |
1233 | ||
1234 | bool apply_epoch(uint64_t candidate_epoch) { | |
1235 | if (candidate_epoch < olh_data_entry.epoch) { | |
1236 | return false; | |
1237 | } | |
1238 | ||
1239 | olh_data_entry.epoch = candidate_epoch; | |
1240 | return true; | |
1241 | } | |
1242 | ||
1243 | bool start_modify(uint64_t candidate_epoch) { | |
1244 | if (candidate_epoch) { | |
1245 | if (candidate_epoch < olh_data_entry.epoch) { | |
1246 | return false; /* olh cannot be modified, old epoch */ | |
1247 | } | |
1248 | olh_data_entry.epoch = candidate_epoch; | |
1249 | } else { | |
1250 | if (olh_data_entry.epoch == 0) { | |
1251 | olh_data_entry.epoch = 2; /* versioned epoch should start with 2, 1 is reserved to converted plain entries */ | |
1252 | } else { | |
1253 | olh_data_entry.epoch++; | |
1254 | } | |
1255 | } | |
1256 | return true; | |
1257 | } | |
1258 | ||
1259 | uint64_t get_epoch() { | |
1260 | return olh_data_entry.epoch; | |
1261 | } | |
1262 | ||
1263 | rgw_bucket_olh_entry& get_entry() { | |
1264 | return olh_data_entry; | |
1265 | } | |
1266 | ||
1267 | void update(cls_rgw_obj_key& key, bool delete_marker) { | |
1268 | olh_data_entry.delete_marker = delete_marker; | |
1269 | olh_data_entry.key = key; | |
1270 | } | |
1271 | ||
1272 | int write() { | |
1273 | /* write the olh data entry */ | |
1274 | int ret = write_entry(hctx, olh_data_entry, olh_data_idx); | |
1275 | if (ret < 0) { | |
1276 | CLS_LOG(0, "ERROR: write_entry() olh_key=%s ret=%d", olh_data_idx.c_str(), ret); | |
1277 | return ret; | |
1278 | } | |
1279 | ||
1280 | return 0; | |
1281 | } | |
1282 | ||
1283 | void update_log(OLHLogOp op, const string& op_tag, cls_rgw_obj_key& key, bool delete_marker, uint64_t epoch = 0) { | |
1284 | if (epoch == 0) { | |
1285 | epoch = olh_data_entry.epoch; | |
1286 | } | |
1287 | update_olh_log(olh_data_entry, op, op_tag, key, delete_marker, epoch); | |
1288 | } | |
1289 | ||
1290 | bool exists() { return olh_data_entry.exists; } | |
1291 | ||
1292 | void set_exists(bool exists) { | |
1293 | olh_data_entry.exists = exists; | |
1294 | } | |
1295 | ||
1296 | bool pending_removal() { return olh_data_entry.pending_removal; } | |
1297 | ||
1298 | void set_pending_removal(bool pending_removal) { | |
1299 | olh_data_entry.pending_removal = pending_removal; | |
1300 | } | |
1301 | ||
1302 | const string& get_tag() { return olh_data_entry.tag; } | |
1303 | void set_tag(const string& tag) { | |
1304 | olh_data_entry.tag = tag; | |
1305 | } | |
1306 | }; | |
1307 | ||
1308 | static int write_version_marker(cls_method_context_t hctx, cls_rgw_obj_key& key) | |
1309 | { | |
1310 | struct rgw_bucket_dir_entry entry; | |
1311 | entry.key = key; | |
1312 | entry.flags = RGW_BUCKET_DIRENT_FLAG_VER_MARKER; | |
1313 | int ret = write_entry(hctx, entry, key.name); | |
1314 | if (ret < 0) { | |
1315 | CLS_LOG(0, "ERROR: write_entry returned ret=%d", ret); | |
1316 | return ret; | |
1317 | } | |
1318 | return 0; | |
1319 | } | |
1320 | ||
1321 | /* | |
1322 | * plain entries are the ones who were created when bucket was not versioned, | |
1323 | * if we override these objects, we need to convert these to versioned entries -- ones that have | |
1324 | * both data entry, and listing key. Their version is going to be empty though | |
1325 | */ | |
1326 | static int convert_plain_entry_to_versioned(cls_method_context_t hctx, cls_rgw_obj_key& key, bool demote_current, bool instance_only) | |
1327 | { | |
1328 | if (!key.instance.empty()) { | |
1329 | return -EINVAL; | |
1330 | } | |
1331 | ||
1332 | struct rgw_bucket_dir_entry entry; | |
1333 | ||
1334 | string orig_idx; | |
1335 | int ret = read_key_entry(hctx, key, &orig_idx, &entry); | |
1336 | if (ret != -ENOENT) { | |
1337 | if (ret < 0) { | |
1338 | CLS_LOG(0, "ERROR: read_key_entry() returned ret=%d", ret); | |
1339 | return ret; | |
1340 | } | |
1341 | ||
1342 | entry.versioned_epoch = 1; /* converted entries are always 1 */ | |
1343 | entry.flags |= RGW_BUCKET_DIRENT_FLAG_VER; | |
1344 | ||
1345 | if (demote_current) { | |
1346 | entry.flags &= ~RGW_BUCKET_DIRENT_FLAG_CURRENT; | |
1347 | } | |
1348 | ||
1349 | string new_idx; | |
1350 | encode_obj_versioned_data_key(key, &new_idx); | |
1351 | ||
1352 | if (instance_only) { | |
1353 | ret = write_obj_instance_entry(hctx, entry, new_idx); | |
1354 | } else { | |
1355 | ret = write_obj_entries(hctx, entry, new_idx); | |
1356 | } | |
1357 | if (ret < 0) { | |
1358 | CLS_LOG(0, "ERROR: write_obj_entries new_idx=%s returned %d", new_idx.c_str(), ret); | |
1359 | return ret; | |
1360 | } | |
1361 | } | |
1362 | ||
1363 | ret = write_version_marker(hctx, key); | |
1364 | if (ret < 0) { | |
1365 | return ret; | |
1366 | } | |
1367 | ||
1368 | return 0; | |
1369 | } | |
1370 | ||
1371 | /* | |
1372 | * link an object version to an olh, update the relevant index entries. It will also handle the | |
1373 | * deletion marker case. We have a few entries that we need to take care of. For object 'foo', | |
1374 | * instance BAR, we'd update the following (not actual encoding): | |
1375 | * - olh data: [BI_BUCKET_OLH_DATA_INDEX]foo | |
1376 | * - object instance data: [BI_BUCKET_OBJ_INSTANCE_INDEX]foo,BAR | |
1377 | * - object instance list entry: foo,123,BAR | |
1378 | * | |
1379 | * The instance list entry needs to be ordered by newer to older, so we generate an appropriate | |
1380 | * number string that follows the name. | |
1381 | * The top instance for each object is marked appropriately. | |
1382 | * We generate instance entry for deletion markers here, as they are not created prior. | |
1383 | */ | |
1384 | static int rgw_bucket_link_olh(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
1385 | { | |
1386 | string olh_data_idx; | |
1387 | string instance_idx; | |
1388 | ||
1389 | // decode request | |
1390 | rgw_cls_link_olh_op op; | |
1391 | bufferlist::iterator iter = in->begin(); | |
1392 | try { | |
1393 | ::decode(op, iter); | |
1394 | } catch (buffer::error& err) { | |
1395 | CLS_LOG(0, "ERROR: rgw_bucket_link_olh_op(): failed to decode request\n"); | |
1396 | return -EINVAL; | |
1397 | } | |
1398 | ||
1399 | BIVerObjEntry obj(hctx, op.key); | |
1400 | BIOLHEntry olh(hctx, op.key); | |
1401 | ||
1402 | /* read instance entry */ | |
1403 | int ret = obj.init(op.delete_marker); | |
1404 | bool existed = (ret == 0); | |
1405 | if (ret == -ENOENT && op.delete_marker) { | |
1406 | ret = 0; | |
1407 | } | |
1408 | if (ret < 0) { | |
1409 | return ret; | |
1410 | } | |
1411 | ||
1412 | if (existed && !real_clock::is_zero(op.unmod_since)) { | |
1413 | struct timespec mtime = ceph::real_clock::to_timespec(obj.mtime()); | |
1414 | struct timespec unmod = ceph::real_clock::to_timespec(op.unmod_since); | |
1415 | if (!op.high_precision_time) { | |
1416 | mtime.tv_nsec = 0; | |
1417 | unmod.tv_nsec = 0; | |
1418 | } | |
1419 | if (mtime >= unmod) { | |
1420 | return 0; /* no need to set error, we just return 0 and avoid writing to the bi log */ | |
1421 | } | |
1422 | } | |
1423 | ||
1424 | bool removing; | |
1425 | ||
1426 | /* | |
1427 | * Special handling for null instance object / delete-marker. For these objects we're going to | |
1428 | * have separate instances for a data object vs. delete-marker to avoid collisions. We now check | |
1429 | * if we got to overwrite a previous entry, and in that case we'll remove its list entry. | |
1430 | */ | |
1431 | if (op.key.instance.empty()) { | |
1432 | BIVerObjEntry other_obj(hctx, op.key); | |
1433 | ret = other_obj.init(!op.delete_marker); /* try reading the other null versioned entry */ | |
1434 | existed = (ret >= 0 && !other_obj.is_delete_marker()); | |
1435 | if (ret >= 0 && other_obj.is_delete_marker() != op.delete_marker) { | |
1436 | ret = other_obj.unlink_list_entry(); | |
1437 | if (ret < 0) { | |
1438 | return ret; | |
1439 | } | |
1440 | ret = other_obj.unlink(); | |
1441 | if (ret < 0) { | |
1442 | return ret; | |
1443 | } | |
1444 | } | |
1445 | ||
1446 | removing = existed && op.delete_marker; | |
1447 | } else { | |
1448 | removing = (existed && !obj.is_delete_marker() && op.delete_marker); | |
1449 | } | |
1450 | ||
1451 | if (op.delete_marker) { | |
1452 | /* a deletion marker, need to initialize entry as such */ | |
1453 | obj.init_as_delete_marker(op.meta); | |
1454 | } | |
1455 | ||
1456 | /* read olh */ | |
1457 | bool olh_found; | |
1458 | ret = olh.init(&olh_found); | |
1459 | if (ret < 0) { | |
1460 | return ret; | |
1461 | } | |
1462 | ||
1463 | if (!olh.start_modify(op.olh_epoch)) { | |
1464 | ret = obj.write(op.olh_epoch, false); | |
1465 | if (ret < 0) { | |
1466 | return ret; | |
1467 | } | |
1468 | if (removing) { | |
1469 | olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false, op.olh_epoch); | |
1470 | } | |
1471 | return 0; | |
1472 | } | |
1473 | ||
1474 | if (olh_found) { | |
1475 | const string& olh_tag = olh.get_tag(); | |
1476 | if (op.olh_tag != olh_tag) { | |
1477 | if (!olh.pending_removal()) { | |
1478 | CLS_LOG(5, "NOTICE: op.olh_tag (%s) != olh.tag (%s)", op.olh_tag.c_str(), olh_tag.c_str()); | |
1479 | return -ECANCELED; | |
1480 | } | |
1481 | /* if pending removal, this is a new olh instance */ | |
1482 | olh.set_tag(op.olh_tag); | |
1483 | } | |
1484 | if (olh.exists()) { | |
1485 | rgw_bucket_olh_entry& olh_entry = olh.get_entry(); | |
1486 | /* found olh, previous instance is no longer the latest, need to update */ | |
1487 | if (!(olh_entry.key == op.key)) { | |
1488 | BIVerObjEntry old_obj(hctx, olh_entry.key); | |
1489 | ||
1490 | ret = old_obj.demote_current(); | |
1491 | if (ret < 0) { | |
1492 | CLS_LOG(0, "ERROR: could not demote current on previous key ret=%d", ret); | |
1493 | return ret; | |
1494 | } | |
1495 | } | |
1496 | } | |
1497 | olh.set_pending_removal(false); | |
1498 | } else { | |
1499 | bool instance_only = (op.key.instance.empty() && op.delete_marker); | |
1500 | cls_rgw_obj_key key(op.key.name); | |
1501 | ret = convert_plain_entry_to_versioned(hctx, key, true, instance_only); | |
1502 | if (ret < 0) { | |
1503 | CLS_LOG(0, "ERROR: convert_plain_entry_to_versioned ret=%d", ret); | |
1504 | return ret; | |
1505 | } | |
1506 | olh.set_tag(op.olh_tag); | |
1507 | } | |
1508 | ||
1509 | /* update the olh log */ | |
1510 | olh.update_log(CLS_RGW_OLH_OP_LINK_OLH, op.op_tag, op.key, op.delete_marker); | |
1511 | if (removing) { | |
1512 | olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false); | |
1513 | } | |
1514 | ||
1515 | olh.update(op.key, op.delete_marker); | |
1516 | ||
1517 | olh.set_exists(true); | |
1518 | ||
1519 | ret = olh.write(); | |
1520 | if (ret < 0) { | |
1521 | CLS_LOG(0, "ERROR: failed to update olh ret=%d", ret); | |
1522 | return ret; | |
1523 | } | |
1524 | ||
1525 | /* write the instance and list entries */ | |
1526 | ret = obj.write(olh.get_epoch(), true); | |
1527 | if (ret < 0) { | |
1528 | return ret; | |
1529 | } | |
1530 | ||
1531 | struct rgw_bucket_dir_header header; | |
1532 | ret = read_bucket_header(hctx, &header); | |
1533 | if (ret < 0) { | |
1534 | CLS_LOG(1, "ERROR: rgw_bucket_unlink_instance(): failed to read header\n"); | |
1535 | return ret; | |
1536 | } | |
1537 | ||
c07f9fc5 | 1538 | if (op.log_op && !header.syncstopped) { |
7c673cae FG |
1539 | rgw_bucket_dir_entry& entry = obj.get_dir_entry(); |
1540 | ||
1541 | rgw_bucket_entry_ver ver; | |
1542 | ver.epoch = (op.olh_epoch ? op.olh_epoch : olh.get_epoch()); | |
1543 | ||
1544 | string *powner = NULL; | |
1545 | string *powner_display_name = NULL; | |
1546 | ||
1547 | if (op.delete_marker) { | |
1548 | powner = &entry.meta.owner; | |
1549 | powner_display_name = &entry.meta.owner_display_name; | |
1550 | } | |
1551 | ||
1552 | RGWModifyOp operation = (op.delete_marker ? CLS_RGW_OP_LINK_OLH_DM : CLS_RGW_OP_LINK_OLH); | |
1553 | ret = log_index_operation(hctx, op.key, operation, op.op_tag, | |
1554 | entry.meta.mtime, ver, | |
1555 | CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, op.bilog_flags | RGW_BILOG_FLAG_VERSIONED_OP, | |
31f18b77 | 1556 | powner, powner_display_name, &op.zones_trace); |
7c673cae FG |
1557 | if (ret < 0) |
1558 | return ret; | |
1559 | } | |
1560 | ||
1561 | return write_bucket_header(hctx, &header); /* updates header version */ | |
1562 | } | |
1563 | ||
1564 | static int rgw_bucket_unlink_instance(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
1565 | { | |
1566 | string olh_data_idx; | |
1567 | string instance_idx; | |
1568 | ||
1569 | // decode request | |
1570 | rgw_cls_unlink_instance_op op; | |
1571 | bufferlist::iterator iter = in->begin(); | |
1572 | try { | |
1573 | ::decode(op, iter); | |
1574 | } catch (buffer::error& err) { | |
1575 | CLS_LOG(0, "ERROR: rgw_bucket_rm_obj_instance_op(): failed to decode request\n"); | |
1576 | return -EINVAL; | |
1577 | } | |
1578 | ||
1579 | cls_rgw_obj_key dest_key = op.key; | |
1580 | if (dest_key.instance == "null") { | |
1581 | dest_key.instance.clear(); | |
1582 | } | |
1583 | ||
1584 | BIVerObjEntry obj(hctx, dest_key); | |
1585 | BIOLHEntry olh(hctx, dest_key); | |
1586 | ||
1587 | int ret = obj.init(); | |
1588 | if (ret == -ENOENT) { | |
1589 | return 0; /* already removed */ | |
1590 | } | |
1591 | if (ret < 0) { | |
1592 | CLS_LOG(0, "ERROR: obj.init() returned ret=%d", ret); | |
1593 | return ret; | |
1594 | } | |
1595 | ||
1596 | bool olh_found; | |
1597 | ret = olh.init(&olh_found); | |
1598 | if (ret < 0) { | |
1599 | CLS_LOG(0, "ERROR: olh.init() returned ret=%d", ret); | |
1600 | return ret; | |
1601 | } | |
1602 | ||
1603 | if (!olh_found) { | |
1604 | bool instance_only = false; | |
1605 | cls_rgw_obj_key key(dest_key.name); | |
1606 | ret = convert_plain_entry_to_versioned(hctx, key, true, instance_only); | |
1607 | if (ret < 0) { | |
1608 | CLS_LOG(0, "ERROR: convert_plain_entry_to_versioned ret=%d", ret); | |
1609 | return ret; | |
1610 | } | |
1611 | olh.update(dest_key, false); | |
1612 | olh.set_tag(op.olh_tag); | |
1613 | ||
1614 | obj.set_epoch(1); | |
1615 | } | |
1616 | ||
1617 | if (!olh.start_modify(op.olh_epoch)) { | |
1618 | ret = obj.unlink_list_entry(); | |
1619 | if (ret < 0) { | |
1620 | return ret; | |
1621 | } | |
1622 | ||
1623 | if (!obj.is_delete_marker()) { | |
1624 | olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false, op.olh_epoch); | |
1625 | } | |
1626 | ||
1627 | return 0; | |
1628 | } | |
1629 | ||
1630 | rgw_bucket_olh_entry& olh_entry = olh.get_entry(); | |
1631 | cls_rgw_obj_key& olh_key = olh_entry.key; | |
1632 | CLS_LOG(20, "%s(): updating olh log: existing olh entry: %s[%s] (delete_marker=%d)", __func__, | |
1633 | olh_key.name.c_str(), olh_key.instance.c_str(), olh_entry.delete_marker); | |
1634 | ||
1635 | if (olh_key == dest_key) { | |
1636 | /* this is the current head, need to update! */ | |
1637 | cls_rgw_obj_key next_key; | |
1638 | bool found; | |
1639 | ret = obj.find_next_key(&next_key, &found); | |
1640 | if (ret < 0) { | |
1641 | CLS_LOG(0, "ERROR: obj.find_next_key() returned ret=%d", ret); | |
1642 | return ret; | |
1643 | } | |
1644 | ||
1645 | if (found) { | |
1646 | BIVerObjEntry next(hctx, next_key); | |
1647 | ret = next.write(olh.get_epoch(), true); | |
1648 | if (ret < 0) { | |
1649 | CLS_LOG(0, "ERROR: next.write() returned ret=%d", ret); | |
1650 | return ret; | |
1651 | } | |
1652 | ||
1653 | CLS_LOG(20, "%s(): updating olh log: link olh -> %s[%s] (is_delete=%d)", __func__, | |
1654 | next_key.name.c_str(), next_key.instance.c_str(), (int)next.is_delete_marker()); | |
1655 | ||
1656 | olh.update(next_key, next.is_delete_marker()); | |
1657 | olh.update_log(CLS_RGW_OLH_OP_LINK_OLH, op.op_tag, next_key, next.is_delete_marker()); | |
1658 | } else { | |
1659 | /* next_key is empty */ | |
1660 | olh.update(next_key, false); | |
1661 | olh.update_log(CLS_RGW_OLH_OP_UNLINK_OLH, op.op_tag, next_key, false); | |
1662 | olh.set_exists(false); | |
1663 | olh.set_pending_removal(true); | |
1664 | } | |
1665 | } | |
1666 | ||
1667 | if (!obj.is_delete_marker()) { | |
1668 | olh.update_log(CLS_RGW_OLH_OP_REMOVE_INSTANCE, op.op_tag, op.key, false); | |
1669 | } else { | |
1670 | /* this is a delete marker, it's our responsibility to remove its instance entry */ | |
1671 | ret = obj.unlink(); | |
1672 | if (ret < 0) { | |
1673 | return ret; | |
1674 | } | |
1675 | } | |
1676 | ||
1677 | ret = obj.unlink_list_entry(); | |
1678 | if (ret < 0) { | |
1679 | return ret; | |
1680 | } | |
1681 | ||
1682 | ret = olh.write(); | |
1683 | if (ret < 0) { | |
1684 | return ret; | |
1685 | } | |
1686 | ||
1687 | struct rgw_bucket_dir_header header; | |
1688 | ret = read_bucket_header(hctx, &header); | |
1689 | if (ret < 0) { | |
1690 | CLS_LOG(1, "ERROR: rgw_bucket_unlink_instance(): failed to read header\n"); | |
1691 | return ret; | |
1692 | } | |
1693 | ||
c07f9fc5 | 1694 | if (op.log_op && !header.syncstopped) { |
7c673cae FG |
1695 | rgw_bucket_entry_ver ver; |
1696 | ver.epoch = (op.olh_epoch ? op.olh_epoch : olh.get_epoch()); | |
1697 | ||
94b18763 | 1698 | real_time mtime = obj.mtime(); /* mtime has no real meaning in instance removal context */ |
7c673cae FG |
1699 | ret = log_index_operation(hctx, op.key, CLS_RGW_OP_UNLINK_INSTANCE, op.op_tag, |
1700 | mtime, ver, | |
1701 | CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, | |
31f18b77 | 1702 | op.bilog_flags | RGW_BILOG_FLAG_VERSIONED_OP, NULL, NULL, &op.zones_trace); |
7c673cae FG |
1703 | if (ret < 0) |
1704 | return ret; | |
1705 | } | |
1706 | ||
1707 | return write_bucket_header(hctx, &header); /* updates header version */ | |
1708 | } | |
1709 | ||
1710 | static int rgw_bucket_read_olh_log(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
1711 | { | |
1712 | // decode request | |
1713 | rgw_cls_read_olh_log_op op; | |
1714 | bufferlist::iterator iter = in->begin(); | |
1715 | try { | |
1716 | ::decode(op, iter); | |
1717 | } catch (buffer::error& err) { | |
1718 | CLS_LOG(0, "ERROR: rgw_bucket_read_olh_log(): failed to decode request\n"); | |
1719 | return -EINVAL; | |
1720 | } | |
1721 | ||
1722 | if (!op.olh.instance.empty()) { | |
1723 | CLS_LOG(1, "bad key passed in (non empty instance)"); | |
1724 | return -EINVAL; | |
1725 | } | |
1726 | ||
1727 | struct rgw_bucket_olh_entry olh_data_entry; | |
1728 | string olh_data_key; | |
1729 | encode_olh_data_key(op.olh, &olh_data_key); | |
1730 | int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry); | |
1731 | if (ret < 0 && ret != -ENOENT) { | |
1732 | CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret); | |
1733 | return ret; | |
1734 | } | |
1735 | ||
1736 | if (olh_data_entry.tag != op.olh_tag) { | |
1737 | CLS_LOG(1, "NOTICE: %s(): olh_tag_mismatch olh_data_entry.tag=%s op.olh_tag=%s", __func__, olh_data_entry.tag.c_str(), op.olh_tag.c_str()); | |
1738 | return -ECANCELED; | |
1739 | } | |
1740 | ||
1741 | rgw_cls_read_olh_log_ret op_ret; | |
1742 | ||
1743 | #define MAX_OLH_LOG_ENTRIES 1000 | |
1744 | map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log = olh_data_entry.pending_log; | |
1745 | ||
1746 | if (log.begin()->first > op.ver_marker && log.size() <= MAX_OLH_LOG_ENTRIES) { | |
1747 | op_ret.log = log; | |
1748 | op_ret.is_truncated = false; | |
1749 | } else { | |
1750 | map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator iter = log.upper_bound(op.ver_marker); | |
1751 | ||
1752 | for (int i = 0; i < MAX_OLH_LOG_ENTRIES && iter != log.end(); ++i, ++iter) { | |
1753 | op_ret.log[iter->first] = iter->second; | |
1754 | } | |
1755 | op_ret.is_truncated = (iter != log.end()); | |
1756 | } | |
1757 | ||
1758 | ::encode(op_ret, *out); | |
1759 | ||
1760 | return 0; | |
1761 | } | |
1762 | ||
1763 | static int rgw_bucket_trim_olh_log(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
1764 | { | |
1765 | // decode request | |
1766 | rgw_cls_trim_olh_log_op op; | |
1767 | bufferlist::iterator iter = in->begin(); | |
1768 | try { | |
1769 | ::decode(op, iter); | |
1770 | } catch (buffer::error& err) { | |
1771 | CLS_LOG(0, "ERROR: rgw_bucket_trim_olh_log(): failed to decode request\n"); | |
1772 | return -EINVAL; | |
1773 | } | |
1774 | ||
1775 | if (!op.olh.instance.empty()) { | |
1776 | CLS_LOG(1, "bad key passed in (non empty instance)"); | |
1777 | return -EINVAL; | |
1778 | } | |
1779 | ||
1780 | /* read olh entry */ | |
1781 | struct rgw_bucket_olh_entry olh_data_entry; | |
1782 | string olh_data_key; | |
1783 | encode_olh_data_key(op.olh, &olh_data_key); | |
1784 | int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry); | |
1785 | if (ret < 0 && ret != -ENOENT) { | |
1786 | CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret); | |
1787 | return ret; | |
1788 | } | |
1789 | ||
1790 | if (olh_data_entry.tag != op.olh_tag) { | |
1791 | CLS_LOG(1, "NOTICE: %s(): olh_tag_mismatch olh_data_entry.tag=%s op.olh_tag=%s", __func__, olh_data_entry.tag.c_str(), op.olh_tag.c_str()); | |
1792 | return -ECANCELED; | |
1793 | } | |
1794 | ||
1795 | /* remove all versions up to and including ver from the pending map */ | |
1796 | map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log = olh_data_entry.pending_log; | |
1797 | map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator liter = log.begin(); | |
1798 | while (liter != log.end() && liter->first <= op.ver) { | |
1799 | map<uint64_t, vector<rgw_bucket_olh_log_entry> >::iterator rm_iter = liter; | |
1800 | ++liter; | |
1801 | log.erase(rm_iter); | |
1802 | } | |
1803 | ||
1804 | /* write the olh data entry */ | |
1805 | ret = write_entry(hctx, olh_data_entry, olh_data_key); | |
1806 | if (ret < 0) { | |
1807 | CLS_LOG(0, "ERROR: write_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret); | |
1808 | return ret; | |
1809 | } | |
1810 | ||
1811 | return 0; | |
1812 | } | |
1813 | ||
1814 | static int rgw_bucket_clear_olh(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
1815 | { | |
1816 | // decode request | |
1817 | rgw_cls_bucket_clear_olh_op op; | |
1818 | bufferlist::iterator iter = in->begin(); | |
1819 | try { | |
1820 | ::decode(op, iter); | |
1821 | } catch (buffer::error& err) { | |
1822 | CLS_LOG(0, "ERROR: rgw_bucket_clear_olh(): failed to decode request\n"); | |
1823 | return -EINVAL; | |
1824 | } | |
1825 | ||
1826 | if (!op.key.instance.empty()) { | |
1827 | CLS_LOG(1, "bad key passed in (non empty instance)"); | |
1828 | return -EINVAL; | |
1829 | } | |
1830 | ||
1831 | /* read olh entry */ | |
1832 | struct rgw_bucket_olh_entry olh_data_entry; | |
1833 | string olh_data_key; | |
1834 | encode_olh_data_key(op.key, &olh_data_key); | |
1835 | int ret = read_index_entry(hctx, olh_data_key, &olh_data_entry); | |
1836 | if (ret < 0 && ret != -ENOENT) { | |
1837 | CLS_LOG(0, "ERROR: read_index_entry() olh_key=%s ret=%d", olh_data_key.c_str(), ret); | |
1838 | return ret; | |
1839 | } | |
1840 | ||
1841 | if (olh_data_entry.tag != op.olh_tag) { | |
1842 | CLS_LOG(1, "NOTICE: %s(): olh_tag_mismatch olh_data_entry.tag=%s op.olh_tag=%s", __func__, olh_data_entry.tag.c_str(), op.olh_tag.c_str()); | |
1843 | return -ECANCELED; | |
1844 | } | |
1845 | ||
1846 | ret = cls_cxx_map_remove_key(hctx, olh_data_key); | |
1847 | if (ret < 0) { | |
1848 | CLS_LOG(1, "NOTICE: %s(): can't remove key %s ret=%d", __func__, olh_data_key.c_str(), ret); | |
1849 | return ret; | |
1850 | } | |
1851 | ||
1852 | rgw_bucket_dir_entry plain_entry; | |
1853 | ||
1854 | /* read plain entry, make sure it's a versioned place holder */ | |
1855 | ret = read_index_entry(hctx, op.key.name, &plain_entry); | |
1856 | if (ret == -ENOENT) { | |
1857 | /* we're done, no entry existing */ | |
1858 | return 0; | |
1859 | } | |
1860 | if (ret < 0) { | |
1861 | CLS_LOG(0, "ERROR: read_index_entry key=%s ret=%d", op.key.name.c_str(), ret); | |
1862 | return ret; | |
1863 | } | |
1864 | ||
1865 | if ((plain_entry.flags & RGW_BUCKET_DIRENT_FLAG_VER_MARKER) == 0) { | |
1866 | /* it's not a version marker, don't remove it */ | |
1867 | return 0; | |
1868 | } | |
1869 | ||
1870 | ret = cls_cxx_map_remove_key(hctx, op.key.name); | |
1871 | if (ret < 0) { | |
1872 | CLS_LOG(1, "NOTICE: %s(): can't remove key %s ret=%d", __func__, op.key.name.c_str(), ret); | |
1873 | return ret; | |
1874 | } | |
1875 | ||
1876 | return 0; | |
1877 | } | |
1878 | ||
3efd9988 FG |
1879 | int rgw_dir_suggest_changes(cls_method_context_t hctx, |
1880 | bufferlist *in, bufferlist *out) | |
7c673cae FG |
1881 | { |
1882 | CLS_LOG(1, "rgw_dir_suggest_changes()"); | |
1883 | ||
1884 | bufferlist header_bl; | |
1885 | struct rgw_bucket_dir_header header; | |
1886 | bool header_changed = false; | |
1887 | ||
1888 | int rc = read_bucket_header(hctx, &header); | |
1889 | if (rc < 0) { | |
1890 | CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to read header\n"); | |
1891 | return rc; | |
1892 | } | |
1893 | ||
224ce89b WB |
1894 | timespan tag_timeout( |
1895 | std::chrono::seconds( | |
1896 | header.tag_timeout ? header.tag_timeout : CEPH_RGW_TAG_TIMEOUT)); | |
7c673cae FG |
1897 | |
1898 | bufferlist::iterator in_iter = in->begin(); | |
1899 | ||
1900 | while (!in_iter.end()) { | |
1901 | __u8 op; | |
1902 | rgw_bucket_dir_entry cur_change; | |
1903 | rgw_bucket_dir_entry cur_disk; | |
1904 | try { | |
1905 | ::decode(op, in_iter); | |
1906 | ::decode(cur_change, in_iter); | |
1907 | } catch (buffer::error& err) { | |
1908 | CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to decode request\n"); | |
1909 | return -EINVAL; | |
1910 | } | |
1911 | ||
1912 | bufferlist cur_disk_bl; | |
1913 | string cur_change_key; | |
1914 | encode_obj_index_key(cur_change.key, &cur_change_key); | |
1915 | int ret = cls_cxx_map_get_val(hctx, cur_change_key, &cur_disk_bl); | |
1916 | if (ret < 0 && ret != -ENOENT) | |
1917 | return -EINVAL; | |
1918 | ||
1adf2230 AA |
1919 | if (ret == -ENOENT) { |
1920 | continue; | |
1921 | } | |
1922 | ||
7c673cae FG |
1923 | if (cur_disk_bl.length()) { |
1924 | bufferlist::iterator cur_disk_iter = cur_disk_bl.begin(); | |
1925 | try { | |
1926 | ::decode(cur_disk, cur_disk_iter); | |
1927 | } catch (buffer::error& error) { | |
1928 | CLS_LOG(1, "ERROR: rgw_dir_suggest_changes(): failed to decode cur_disk\n"); | |
1929 | return -EINVAL; | |
1930 | } | |
1931 | ||
1932 | real_time cur_time = real_clock::now(); | |
1933 | map<string, struct rgw_bucket_pending_info>::iterator iter = | |
1934 | cur_disk.pending_map.begin(); | |
1935 | while(iter != cur_disk.pending_map.end()) { | |
1936 | map<string, struct rgw_bucket_pending_info>::iterator cur_iter=iter++; | |
1937 | if (cur_time > (cur_iter->second.timestamp + timespan(tag_timeout))) { | |
1938 | cur_disk.pending_map.erase(cur_iter); | |
1939 | } | |
1940 | } | |
1941 | } | |
1942 | ||
1943 | CLS_LOG(20, "cur_disk.pending_map.empty()=%d op=%d cur_disk.exists=%d cur_change.pending_map.size()=%d cur_change.exists=%d\n", | |
1944 | cur_disk.pending_map.empty(), (int)op, cur_disk.exists, | |
1945 | (int)cur_change.pending_map.size(), cur_change.exists); | |
1946 | ||
1947 | if (cur_disk.pending_map.empty()) { | |
1948 | if (cur_disk.exists) { | |
1949 | struct rgw_bucket_category_stats& old_stats = header.stats[cur_disk.meta.category]; | |
1950 | CLS_LOG(10, "total_entries: %" PRId64 " -> %" PRId64 "\n", old_stats.num_entries, old_stats.num_entries - 1); | |
1951 | old_stats.num_entries--; | |
1952 | old_stats.total_size -= cur_disk.meta.accounted_size; | |
1953 | old_stats.total_size_rounded -= cls_rgw_get_rounded_size(cur_disk.meta.accounted_size); | |
1954 | old_stats.actual_size -= cur_disk.meta.size; | |
1955 | header_changed = true; | |
1956 | } | |
1957 | struct rgw_bucket_category_stats& stats = | |
1958 | header.stats[cur_change.meta.category]; | |
1959 | bool log_op = (op & CEPH_RGW_DIR_SUGGEST_LOG_OP) != 0; | |
1960 | op &= CEPH_RGW_DIR_SUGGEST_OP_MASK; | |
1961 | switch(op) { | |
1962 | case CEPH_RGW_REMOVE: | |
1963 | CLS_LOG(10, "CEPH_RGW_REMOVE name=%s instance=%s\n", cur_change.key.name.c_str(), cur_change.key.instance.c_str()); | |
1964 | ret = cls_cxx_map_remove_key(hctx, cur_change_key); | |
1965 | if (ret < 0) | |
1966 | return ret; | |
c07f9fc5 | 1967 | if (log_op && cur_disk.exists && !header.syncstopped) { |
7c673cae | 1968 | ret = log_index_operation(hctx, cur_disk.key, CLS_RGW_OP_DEL, cur_disk.tag, cur_disk.meta.mtime, |
31f18b77 | 1969 | cur_disk.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, 0, NULL, NULL, NULL); |
7c673cae FG |
1970 | if (ret < 0) { |
1971 | CLS_LOG(0, "ERROR: %s(): failed to log operation ret=%d", __func__, ret); | |
1972 | return ret; | |
1973 | } | |
1974 | } | |
1975 | break; | |
1976 | case CEPH_RGW_UPDATE: | |
1977 | CLS_LOG(10, "CEPH_RGW_UPDATE name=%s instance=%s total_entries: %" PRId64 " -> %" PRId64 "\n", | |
1978 | cur_change.key.name.c_str(), cur_change.key.instance.c_str(), stats.num_entries, stats.num_entries + 1); | |
3efd9988 | 1979 | |
7c673cae FG |
1980 | stats.num_entries++; |
1981 | stats.total_size += cur_change.meta.accounted_size; | |
1982 | stats.total_size_rounded += cls_rgw_get_rounded_size(cur_change.meta.accounted_size); | |
1983 | stats.actual_size += cur_change.meta.size; | |
1984 | header_changed = true; | |
1985 | cur_change.index_ver = header.ver; | |
1986 | bufferlist cur_state_bl; | |
1987 | ::encode(cur_change, cur_state_bl); | |
1988 | ret = cls_cxx_map_set_val(hctx, cur_change_key, &cur_state_bl); | |
1989 | if (ret < 0) | |
1990 | return ret; | |
c07f9fc5 | 1991 | if (log_op && !header.syncstopped) { |
7c673cae | 1992 | ret = log_index_operation(hctx, cur_change.key, CLS_RGW_OP_ADD, cur_change.tag, cur_change.meta.mtime, |
31f18b77 | 1993 | cur_change.ver, CLS_RGW_STATE_COMPLETE, header.ver, header.max_marker, 0, NULL, NULL, NULL); |
7c673cae FG |
1994 | if (ret < 0) { |
1995 | CLS_LOG(0, "ERROR: %s(): failed to log operation ret=%d", __func__, ret); | |
1996 | return ret; | |
1997 | } | |
1998 | } | |
1999 | break; | |
3efd9988 FG |
2000 | } // switch(op) |
2001 | } // if (cur_disk.pending_map.empty()) | |
2002 | } // while (!in_iter.end()) | |
7c673cae FG |
2003 | |
2004 | if (header_changed) { | |
2005 | return write_bucket_header(hctx, &header); | |
2006 | } | |
2007 | return 0; | |
2008 | } | |
2009 | ||
2010 | static int rgw_obj_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2011 | { | |
2012 | // decode request | |
2013 | rgw_cls_obj_remove_op op; | |
2014 | bufferlist::iterator iter = in->begin(); | |
2015 | try { | |
2016 | ::decode(op, iter); | |
2017 | } catch (buffer::error& err) { | |
2018 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2019 | return -EINVAL; | |
2020 | } | |
2021 | ||
2022 | if (op.keep_attr_prefixes.empty()) { | |
2023 | return cls_cxx_remove(hctx); | |
2024 | } | |
2025 | ||
2026 | map<string, bufferlist> attrset; | |
2027 | int ret = cls_cxx_getxattrs(hctx, &attrset); | |
2028 | if (ret < 0 && ret != -ENOENT) { | |
2029 | CLS_LOG(0, "ERROR: %s(): cls_cxx_getxattrs() returned %d", __func__, ret); | |
2030 | return ret; | |
2031 | } | |
2032 | ||
2033 | map<string, bufferlist> new_attrs; | |
2034 | for (list<string>::iterator iter = op.keep_attr_prefixes.begin(); | |
2035 | iter != op.keep_attr_prefixes.end(); ++iter) { | |
2036 | string& check_prefix = *iter; | |
2037 | ||
2038 | for (map<string, bufferlist>::iterator aiter = attrset.lower_bound(check_prefix); | |
2039 | aiter != attrset.end(); ++aiter) { | |
2040 | const string& attr = aiter->first; | |
2041 | ||
2042 | if (attr.substr(0, check_prefix.size()) > check_prefix) { | |
2043 | break; | |
2044 | } | |
2045 | ||
2046 | new_attrs[attr] = aiter->second; | |
2047 | } | |
2048 | } | |
2049 | ||
2050 | CLS_LOG(20, "%s(): removing object", __func__); | |
2051 | ret = cls_cxx_remove(hctx); | |
2052 | if (ret < 0) { | |
2053 | CLS_LOG(0, "ERROR: %s(): cls_cxx_remove returned %d", __func__, ret); | |
2054 | return ret; | |
2055 | } | |
2056 | ||
2057 | if (new_attrs.empty()) { | |
2058 | /* no data to keep */ | |
2059 | return 0; | |
2060 | } | |
2061 | ||
2062 | ret = cls_cxx_create(hctx, false); | |
2063 | if (ret < 0) { | |
2064 | CLS_LOG(0, "ERROR: %s(): cls_cxx_create returned %d", __func__, ret); | |
2065 | return ret; | |
2066 | } | |
2067 | ||
2068 | for (map<string, bufferlist>::iterator aiter = new_attrs.begin(); | |
2069 | aiter != new_attrs.end(); ++aiter) { | |
2070 | const string& attr = aiter->first; | |
2071 | ||
2072 | ret = cls_cxx_setxattr(hctx, attr.c_str(), &aiter->second); | |
2073 | CLS_LOG(20, "%s(): setting attr: %s", __func__, attr.c_str()); | |
2074 | if (ret < 0) { | |
2075 | CLS_LOG(0, "ERROR: %s(): cls_cxx_setxattr (attr=%s) returned %d", __func__, attr.c_str(), ret); | |
2076 | return ret; | |
2077 | } | |
2078 | } | |
2079 | ||
2080 | return 0; | |
2081 | } | |
2082 | ||
2083 | static int rgw_obj_store_pg_ver(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2084 | { | |
2085 | // decode request | |
2086 | rgw_cls_obj_store_pg_ver_op op; | |
2087 | bufferlist::iterator iter = in->begin(); | |
2088 | try { | |
2089 | ::decode(op, iter); | |
2090 | } catch (buffer::error& err) { | |
2091 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2092 | return -EINVAL; | |
2093 | } | |
2094 | ||
2095 | bufferlist bl; | |
2096 | uint64_t ver = cls_current_version(hctx); | |
2097 | ::encode(ver, bl); | |
2098 | int ret = cls_cxx_setxattr(hctx, op.attr.c_str(), &bl); | |
2099 | if (ret < 0) { | |
2100 | CLS_LOG(0, "ERROR: %s(): cls_cxx_setxattr (attr=%s) returned %d", __func__, op.attr.c_str(), ret); | |
2101 | return ret; | |
2102 | } | |
2103 | ||
2104 | return 0; | |
2105 | } | |
2106 | ||
2107 | static int rgw_obj_check_attrs_prefix(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2108 | { | |
2109 | // decode request | |
2110 | rgw_cls_obj_check_attrs_prefix op; | |
2111 | bufferlist::iterator iter = in->begin(); | |
2112 | try { | |
2113 | ::decode(op, iter); | |
2114 | } catch (buffer::error& err) { | |
2115 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2116 | return -EINVAL; | |
2117 | } | |
2118 | ||
2119 | if (op.check_prefix.empty()) { | |
2120 | return -EINVAL; | |
2121 | } | |
2122 | ||
2123 | map<string, bufferlist> attrset; | |
2124 | int ret = cls_cxx_getxattrs(hctx, &attrset); | |
2125 | if (ret < 0 && ret != -ENOENT) { | |
2126 | CLS_LOG(0, "ERROR: %s(): cls_cxx_getxattrs() returned %d", __func__, ret); | |
2127 | return ret; | |
2128 | } | |
2129 | ||
2130 | bool exist = false; | |
2131 | ||
2132 | for (map<string, bufferlist>::iterator aiter = attrset.lower_bound(op.check_prefix); | |
2133 | aiter != attrset.end(); ++aiter) { | |
2134 | const string& attr = aiter->first; | |
2135 | ||
2136 | if (attr.substr(0, op.check_prefix.size()) > op.check_prefix) { | |
2137 | break; | |
2138 | } | |
2139 | ||
2140 | exist = true; | |
2141 | } | |
2142 | ||
2143 | if (exist == op.fail_if_exist) { | |
2144 | return -ECANCELED; | |
2145 | } | |
2146 | ||
2147 | return 0; | |
2148 | } | |
2149 | ||
2150 | static int rgw_obj_check_mtime(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2151 | { | |
2152 | // decode request | |
2153 | rgw_cls_obj_check_mtime op; | |
2154 | bufferlist::iterator iter = in->begin(); | |
2155 | try { | |
2156 | ::decode(op, iter); | |
2157 | } catch (buffer::error& err) { | |
2158 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2159 | return -EINVAL; | |
2160 | } | |
2161 | ||
2162 | real_time obj_ut; | |
2163 | int ret = cls_cxx_stat2(hctx, NULL, &obj_ut); | |
2164 | if (ret < 0 && ret != -ENOENT) { | |
2165 | CLS_LOG(0, "ERROR: %s(): cls_cxx_stat() returned %d", __func__, ret); | |
2166 | return ret; | |
2167 | } | |
2168 | if (ret == -ENOENT) { | |
2169 | CLS_LOG(10, "object does not exist, skipping check"); | |
2170 | } | |
2171 | ||
2172 | ceph_timespec obj_ts = ceph::real_clock::to_ceph_timespec(obj_ut); | |
2173 | ceph_timespec op_ts = ceph::real_clock::to_ceph_timespec(op.mtime); | |
2174 | ||
2175 | if (!op.high_precision_time) { | |
2176 | obj_ts.tv_nsec = 0; | |
2177 | op_ts.tv_nsec = 0; | |
2178 | } | |
2179 | ||
2180 | CLS_LOG(10, "%s: obj_ut=%lld.%06lld op.mtime=%lld.%06lld", __func__, | |
2181 | (long long)obj_ts.tv_sec, (long long)obj_ts.tv_nsec, | |
2182 | (long long)op_ts.tv_sec, (long long)op_ts.tv_nsec); | |
2183 | ||
2184 | bool check; | |
2185 | ||
2186 | switch (op.type) { | |
2187 | case CLS_RGW_CHECK_TIME_MTIME_EQ: | |
2188 | check = (obj_ts == op_ts); | |
2189 | break; | |
2190 | case CLS_RGW_CHECK_TIME_MTIME_LT: | |
2191 | check = (obj_ts < op_ts); | |
2192 | break; | |
2193 | case CLS_RGW_CHECK_TIME_MTIME_LE: | |
2194 | check = (obj_ts <= op_ts); | |
2195 | break; | |
2196 | case CLS_RGW_CHECK_TIME_MTIME_GT: | |
2197 | check = (obj_ts > op_ts); | |
2198 | break; | |
2199 | case CLS_RGW_CHECK_TIME_MTIME_GE: | |
2200 | check = (obj_ts >= op_ts); | |
2201 | break; | |
2202 | default: | |
2203 | return -EINVAL; | |
2204 | }; | |
2205 | ||
2206 | if (!check) { | |
2207 | return -ECANCELED; | |
2208 | } | |
2209 | ||
2210 | return 0; | |
2211 | } | |
2212 | ||
2213 | static int rgw_bi_get_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2214 | { | |
2215 | // decode request | |
2216 | rgw_cls_bi_get_op op; | |
2217 | bufferlist::iterator iter = in->begin(); | |
2218 | try { | |
2219 | ::decode(op, iter); | |
2220 | } catch (buffer::error& err) { | |
2221 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2222 | return -EINVAL; | |
2223 | } | |
2224 | ||
2225 | string idx; | |
2226 | ||
2227 | switch (op.type) { | |
2228 | case PlainIdx: | |
2229 | idx = op.key.name; | |
2230 | break; | |
2231 | case InstanceIdx: | |
2232 | encode_obj_index_key(op.key, &idx); | |
2233 | break; | |
2234 | case OLHIdx: | |
2235 | encode_olh_data_key(op.key, &idx); | |
2236 | break; | |
2237 | default: | |
2238 | CLS_LOG(10, "%s(): invalid key type encoding: %d", __func__, op.type); | |
2239 | return -EINVAL; | |
2240 | } | |
2241 | ||
2242 | rgw_cls_bi_get_ret op_ret; | |
2243 | ||
2244 | rgw_cls_bi_entry& entry = op_ret.entry; | |
2245 | ||
2246 | entry.type = op.type; | |
2247 | entry.idx = idx; | |
2248 | ||
2249 | int r = cls_cxx_map_get_val(hctx, idx, &entry.data); | |
2250 | if (r < 0) { | |
2251 | CLS_LOG(10, "%s(): cls_cxx_map_get_val() returned %d", __func__, r); | |
2252 | return r; | |
2253 | } | |
2254 | ||
2255 | ::encode(op_ret, *out); | |
2256 | ||
2257 | return 0; | |
2258 | } | |
2259 | ||
2260 | static int rgw_bi_put_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2261 | { | |
2262 | // decode request | |
2263 | rgw_cls_bi_put_op op; | |
2264 | bufferlist::iterator iter = in->begin(); | |
2265 | try { | |
2266 | ::decode(op, iter); | |
2267 | } catch (buffer::error& err) { | |
2268 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2269 | return -EINVAL; | |
2270 | } | |
2271 | ||
2272 | rgw_cls_bi_entry& entry = op.entry; | |
2273 | ||
2274 | int r = cls_cxx_map_set_val(hctx, entry.idx, &entry.data); | |
2275 | if (r < 0) { | |
2276 | CLS_LOG(0, "ERROR: %s(): cls_cxx_map_set_val() returned r=%d", __func__, r); | |
2277 | } | |
2278 | ||
2279 | return 0; | |
2280 | } | |
2281 | ||
2282 | static int list_plain_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max, | |
c07f9fc5 | 2283 | list<rgw_cls_bi_entry> *entries, bool *pmore) |
7c673cae FG |
2284 | { |
2285 | string filter = name; | |
2286 | string start_key = marker; | |
2287 | ||
2288 | string end_key; // stop listing at bi_log_prefix | |
2289 | bi_log_prefix(end_key); | |
2290 | ||
2291 | int count = 0; | |
2292 | map<string, bufferlist> keys; | |
c07f9fc5 FG |
2293 | int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max, &keys, pmore); |
2294 | if (ret < 0) { | |
2295 | return ret; | |
2296 | } | |
2297 | ||
2298 | map<string, bufferlist>::iterator iter; | |
2299 | for (iter = keys.begin(); iter != keys.end(); ++iter) { | |
2300 | if (iter->first >= end_key) { | |
2301 | /* past the end of plain namespace */ | |
28e407b8 AA |
2302 | if (pmore) { |
2303 | *pmore = false; | |
2304 | } | |
7c673cae FG |
2305 | return count; |
2306 | } | |
7c673cae | 2307 | |
c07f9fc5 FG |
2308 | rgw_cls_bi_entry entry; |
2309 | entry.type = PlainIdx; | |
2310 | entry.idx = iter->first; | |
2311 | entry.data = iter->second; | |
7c673cae | 2312 | |
c07f9fc5 | 2313 | bufferlist::iterator biter = entry.data.begin(); |
7c673cae | 2314 | |
c07f9fc5 FG |
2315 | rgw_bucket_dir_entry e; |
2316 | try { | |
2317 | ::decode(e, biter); | |
2318 | } catch (buffer::error& err) { | |
2319 | CLS_LOG(0, "ERROR: %s(): failed to decode buffer", __func__); | |
2320 | return -EIO; | |
2321 | } | |
7c673cae | 2322 | |
c07f9fc5 | 2323 | CLS_LOG(20, "%s(): entry.idx=%s e.key.name=%s", __func__, escape_str(entry.idx).c_str(), escape_str(e.key.name).c_str()); |
7c673cae | 2324 | |
c07f9fc5 | 2325 | if (!name.empty() && e.key.name != name) { |
28e407b8 AA |
2326 | /* we are skipping the rest of the entries */ |
2327 | if (pmore) { | |
2328 | *pmore = false; | |
2329 | } | |
c07f9fc5 FG |
2330 | return count; |
2331 | } | |
7c673cae | 2332 | |
c07f9fc5 FG |
2333 | entries->push_back(entry); |
2334 | count++; | |
2335 | if (count >= (int)max) { | |
2336 | return count; | |
7c673cae | 2337 | } |
c07f9fc5 FG |
2338 | start_key = entry.idx; |
2339 | } | |
7c673cae FG |
2340 | |
2341 | return count; | |
2342 | } | |
2343 | ||
2344 | static int list_instance_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max, | |
c07f9fc5 | 2345 | list<rgw_cls_bi_entry> *entries, bool *pmore) |
7c673cae FG |
2346 | { |
2347 | cls_rgw_obj_key key(name); | |
2348 | string first_instance_idx; | |
2349 | encode_obj_versioned_data_key(key, &first_instance_idx); | |
2350 | string start_key; | |
2351 | ||
2352 | if (!name.empty()) { | |
2353 | start_key = first_instance_idx; | |
2354 | } else { | |
2355 | start_key = BI_PREFIX_CHAR; | |
2356 | start_key.append(bucket_index_prefixes[BI_BUCKET_OBJ_INSTANCE_INDEX]); | |
2357 | } | |
2358 | string filter = start_key; | |
2359 | if (bi_entry_gt(marker, start_key)) { | |
2360 | start_key = marker; | |
2361 | } | |
2362 | int count = 0; | |
2363 | map<string, bufferlist> keys; | |
c07f9fc5 FG |
2364 | bufferlist k; |
2365 | int ret = cls_cxx_map_get_val(hctx, start_key, &k); | |
2366 | if (ret < 0 && ret != -ENOENT) { | |
2367 | return ret; | |
2368 | } | |
2369 | bool found_first = (ret == 0); | |
2370 | if (found_first) { | |
2371 | --max; | |
2372 | } | |
2373 | if (max > 0) { | |
2374 | ret = cls_cxx_map_get_vals(hctx, start_key, string(), max, &keys, pmore); | |
7c673cae FG |
2375 | CLS_LOG(20, "%s(): start_key=%s first_instance_idx=%s keys.size()=%d", __func__, escape_str(start_key).c_str(), escape_str(first_instance_idx).c_str(), (int)keys.size()); |
2376 | if (ret < 0) { | |
2377 | return ret; | |
2378 | } | |
c07f9fc5 FG |
2379 | } |
2380 | if (found_first) { | |
2381 | keys[start_key].claim(k); | |
2382 | } | |
7c673cae | 2383 | |
c07f9fc5 FG |
2384 | map<string, bufferlist>::iterator iter; |
2385 | for (iter = keys.begin(); iter != keys.end(); ++iter) { | |
2386 | rgw_cls_bi_entry entry; | |
2387 | entry.type = InstanceIdx; | |
2388 | entry.idx = iter->first; | |
2389 | entry.data = iter->second; | |
7c673cae | 2390 | |
c07f9fc5 | 2391 | if (!filter.empty() && entry.idx.compare(0, filter.size(), filter) != 0) { |
28e407b8 AA |
2392 | /* we are skipping the rest of the entries */ |
2393 | if (pmore) { | |
2394 | *pmore = false; | |
2395 | } | |
c07f9fc5 FG |
2396 | return count; |
2397 | } | |
7c673cae | 2398 | |
c07f9fc5 | 2399 | CLS_LOG(20, "%s(): entry.idx=%s", __func__, escape_str(entry.idx).c_str()); |
7c673cae | 2400 | |
c07f9fc5 | 2401 | bufferlist::iterator biter = entry.data.begin(); |
7c673cae | 2402 | |
c07f9fc5 FG |
2403 | rgw_bucket_dir_entry e; |
2404 | try { | |
2405 | ::decode(e, biter); | |
2406 | } catch (buffer::error& err) { | |
2407 | CLS_LOG(0, "ERROR: %s(): failed to decode buffer (size=%d)", __func__, entry.data.length()); | |
2408 | return -EIO; | |
2409 | } | |
7c673cae | 2410 | |
c07f9fc5 | 2411 | if (!name.empty() && e.key.name != name) { |
28e407b8 AA |
2412 | /* we are skipping the rest of the entries */ |
2413 | if (pmore) { | |
2414 | *pmore = false; | |
2415 | } | |
c07f9fc5 | 2416 | return count; |
7c673cae | 2417 | } |
c07f9fc5 FG |
2418 | |
2419 | entries->push_back(entry); | |
2420 | count++; | |
2421 | start_key = entry.idx; | |
2422 | } | |
7c673cae FG |
2423 | |
2424 | return count; | |
2425 | } | |
2426 | ||
2427 | static int list_olh_entries(cls_method_context_t hctx, const string& name, const string& marker, uint32_t max, | |
c07f9fc5 | 2428 | list<rgw_cls_bi_entry> *entries, bool *pmore) |
7c673cae FG |
2429 | { |
2430 | cls_rgw_obj_key key(name); | |
2431 | string first_instance_idx; | |
2432 | encode_olh_data_key(key, &first_instance_idx); | |
2433 | string start_key; | |
2434 | ||
2435 | if (!name.empty()) { | |
2436 | start_key = first_instance_idx; | |
2437 | } else { | |
2438 | start_key = BI_PREFIX_CHAR; | |
2439 | start_key.append(bucket_index_prefixes[BI_BUCKET_OLH_DATA_INDEX]); | |
2440 | } | |
2441 | string filter = start_key; | |
2442 | if (bi_entry_gt(marker, start_key)) { | |
2443 | start_key = marker; | |
2444 | } | |
2445 | int count = 0; | |
2446 | map<string, bufferlist> keys; | |
c07f9fc5 FG |
2447 | int ret; |
2448 | bufferlist k; | |
2449 | ret = cls_cxx_map_get_val(hctx, start_key, &k); | |
2450 | if (ret < 0 && ret != -ENOENT) { | |
2451 | return ret; | |
2452 | } | |
2453 | bool found_first = (ret == 0); | |
2454 | if (found_first) { | |
2455 | --max; | |
2456 | } | |
2457 | if (max > 0) { | |
2458 | ret = cls_cxx_map_get_vals(hctx, start_key, string(), max, &keys, pmore); | |
7c673cae FG |
2459 | CLS_LOG(20, "%s(): start_key=%s first_instance_idx=%s keys.size()=%d", __func__, escape_str(start_key).c_str(), escape_str(first_instance_idx).c_str(), (int)keys.size()); |
2460 | if (ret < 0) { | |
2461 | return ret; | |
2462 | } | |
c07f9fc5 | 2463 | } |
7c673cae | 2464 | |
c07f9fc5 FG |
2465 | if (found_first) { |
2466 | keys[start_key].claim(k); | |
2467 | } | |
7c673cae | 2468 | |
c07f9fc5 FG |
2469 | map<string, bufferlist>::iterator iter; |
2470 | for (iter = keys.begin(); iter != keys.end(); ++iter) { | |
2471 | rgw_cls_bi_entry entry; | |
2472 | entry.type = OLHIdx; | |
2473 | entry.idx = iter->first; | |
2474 | entry.data = iter->second; | |
7c673cae | 2475 | |
c07f9fc5 | 2476 | if (!filter.empty() && entry.idx.compare(0, filter.size(), filter) != 0) { |
28e407b8 AA |
2477 | /* we are skipping the rest of the entries */ |
2478 | if (pmore) { | |
2479 | *pmore = false; | |
2480 | } | |
c07f9fc5 FG |
2481 | return count; |
2482 | } | |
7c673cae | 2483 | |
c07f9fc5 | 2484 | CLS_LOG(20, "%s(): entry.idx=%s", __func__, escape_str(entry.idx).c_str()); |
7c673cae | 2485 | |
c07f9fc5 | 2486 | bufferlist::iterator biter = entry.data.begin(); |
7c673cae | 2487 | |
c07f9fc5 FG |
2488 | rgw_bucket_olh_entry e; |
2489 | try { | |
2490 | ::decode(e, biter); | |
2491 | } catch (buffer::error& err) { | |
2492 | CLS_LOG(0, "ERROR: %s(): failed to decode buffer (size=%d)", __func__, entry.data.length()); | |
2493 | return -EIO; | |
2494 | } | |
7c673cae | 2495 | |
c07f9fc5 | 2496 | if (!name.empty() && e.key.name != name) { |
28e407b8 AA |
2497 | /* we are skipping the rest of the entries */ |
2498 | if (pmore) { | |
2499 | *pmore = false; | |
2500 | } | |
c07f9fc5 | 2501 | return count; |
7c673cae | 2502 | } |
c07f9fc5 FG |
2503 | |
2504 | entries->push_back(entry); | |
2505 | count++; | |
2506 | start_key = entry.idx; | |
2507 | } | |
7c673cae FG |
2508 | |
2509 | return count; | |
2510 | } | |
2511 | ||
2512 | static int rgw_bi_list_op(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2513 | { | |
2514 | // decode request | |
2515 | rgw_cls_bi_list_op op; | |
2516 | bufferlist::iterator iter = in->begin(); | |
2517 | try { | |
2518 | ::decode(op, iter); | |
2519 | } catch (buffer::error& err) { | |
2520 | CLS_LOG(0, "ERROR: %s(): failed to decode request", __func__); | |
2521 | return -EINVAL; | |
2522 | } | |
2523 | ||
2524 | rgw_cls_bi_list_ret op_ret; | |
2525 | ||
2526 | string filter = op.name; | |
2527 | #define MAX_BI_LIST_ENTRIES 1000 | |
c07f9fc5 | 2528 | int32_t max = (op.max < MAX_BI_LIST_ENTRIES ? op.max : MAX_BI_LIST_ENTRIES); |
7c673cae | 2529 | string start_key = op.marker; |
c07f9fc5 | 2530 | bool more; |
28e407b8 | 2531 | int ret = list_plain_entries(hctx, op.name, op.marker, max, &op_ret.entries, &more); |
7c673cae FG |
2532 | if (ret < 0) { |
2533 | CLS_LOG(0, "ERROR: %s(): list_plain_entries retured ret=%d", __func__, ret); | |
2534 | return ret; | |
2535 | } | |
2536 | int count = ret; | |
2537 | ||
2538 | CLS_LOG(20, "found %d plain entries", count); | |
2539 | ||
c07f9fc5 FG |
2540 | if (!more) { |
2541 | ret = list_instance_entries(hctx, op.name, op.marker, max - count, &op_ret.entries, &more); | |
2542 | if (ret < 0) { | |
2543 | CLS_LOG(0, "ERROR: %s(): list_instance_entries retured ret=%d", __func__, ret); | |
2544 | return ret; | |
2545 | } | |
2546 | ||
2547 | count += ret; | |
7c673cae FG |
2548 | } |
2549 | ||
c07f9fc5 FG |
2550 | if (!more) { |
2551 | ret = list_olh_entries(hctx, op.name, op.marker, max - count, &op_ret.entries, &more); | |
2552 | if (ret < 0) { | |
2553 | CLS_LOG(0, "ERROR: %s(): list_olh_entries retured ret=%d", __func__, ret); | |
2554 | return ret; | |
2555 | } | |
7c673cae | 2556 | |
c07f9fc5 | 2557 | count += ret; |
7c673cae FG |
2558 | } |
2559 | ||
c07f9fc5 | 2560 | op_ret.is_truncated = (count >= max) || more; |
1adf2230 | 2561 | while (count > max) { |
7c673cae FG |
2562 | op_ret.entries.pop_back(); |
2563 | count--; | |
2564 | } | |
2565 | ||
2566 | ::encode(op_ret, *out); | |
2567 | ||
2568 | return 0; | |
2569 | } | |
2570 | ||
2571 | int bi_log_record_decode(bufferlist& bl, rgw_bi_log_entry& e) | |
2572 | { | |
2573 | bufferlist::iterator iter = bl.begin(); | |
2574 | try { | |
2575 | ::decode(e, iter); | |
2576 | } catch (buffer::error& err) { | |
2577 | CLS_LOG(0, "ERROR: failed to decode rgw_bi_log_entry"); | |
2578 | return -EIO; | |
2579 | } | |
2580 | return 0; | |
2581 | } | |
2582 | ||
2583 | static int bi_log_iterate_entries(cls_method_context_t hctx, const string& marker, const string& end_marker, | |
2584 | string& key_iter, uint32_t max_entries, bool *truncated, | |
2585 | int (*cb)(cls_method_context_t, const string&, rgw_bi_log_entry&, void *), | |
2586 | void *param) | |
2587 | { | |
2588 | CLS_LOG(10, "bi_log_iterate_range"); | |
2589 | ||
2590 | map<string, bufferlist> keys; | |
2591 | string filter_prefix, end_key; | |
2592 | uint32_t i = 0; | |
2593 | string key; | |
2594 | ||
2595 | if (truncated) | |
2596 | *truncated = false; | |
2597 | ||
2598 | string start_key; | |
2599 | if (key_iter.empty()) { | |
2600 | key = BI_PREFIX_CHAR; | |
2601 | key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]); | |
2602 | key.append(marker); | |
2603 | ||
2604 | start_key = key; | |
2605 | } else { | |
2606 | start_key = key_iter; | |
2607 | } | |
2608 | ||
2609 | if (end_marker.empty()) { | |
2610 | end_key = BI_PREFIX_CHAR; | |
2611 | end_key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX + 1]); | |
2612 | } else { | |
2613 | end_key = BI_PREFIX_CHAR; | |
2614 | end_key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]); | |
2615 | end_key.append(end_marker); | |
2616 | } | |
2617 | ||
2618 | CLS_LOG(0, "bi_log_iterate_entries start_key=%s end_key=%s\n", start_key.c_str(), end_key.c_str()); | |
2619 | ||
2620 | string filter; | |
2621 | ||
c07f9fc5 FG |
2622 | int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max_entries, &keys, truncated); |
2623 | if (ret < 0) | |
2624 | return ret; | |
7c673cae | 2625 | |
c07f9fc5 FG |
2626 | map<string, bufferlist>::iterator iter = keys.begin(); |
2627 | if (iter == keys.end()) | |
2628 | return 0; | |
7c673cae | 2629 | |
c07f9fc5 | 2630 | uint32_t num_keys = keys.size(); |
7c673cae | 2631 | |
c07f9fc5 FG |
2632 | for (; iter != keys.end(); ++iter,++i) { |
2633 | const string& key = iter->first; | |
2634 | rgw_bi_log_entry e; | |
7c673cae | 2635 | |
c07f9fc5 | 2636 | CLS_LOG(0, "bi_log_iterate_entries key=%s bl.length=%d\n", key.c_str(), (int)iter->second.length()); |
7c673cae | 2637 | |
c07f9fc5 FG |
2638 | if (key.compare(end_key) > 0) { |
2639 | key_iter = key; | |
94b18763 FG |
2640 | if (truncated) { |
2641 | *truncated = false; | |
2642 | } | |
c07f9fc5 FG |
2643 | return 0; |
2644 | } | |
7c673cae | 2645 | |
c07f9fc5 FG |
2646 | ret = bi_log_record_decode(iter->second, e); |
2647 | if (ret < 0) | |
2648 | return ret; | |
7c673cae | 2649 | |
c07f9fc5 FG |
2650 | ret = cb(hctx, key, e, param); |
2651 | if (ret < 0) | |
2652 | return ret; | |
7c673cae | 2653 | |
c07f9fc5 FG |
2654 | if (i == num_keys - 1) { |
2655 | key_iter = key; | |
7c673cae | 2656 | } |
c07f9fc5 FG |
2657 | } |
2658 | ||
7c673cae FG |
2659 | return 0; |
2660 | } | |
2661 | ||
2662 | static int bi_log_list_cb(cls_method_context_t hctx, const string& key, rgw_bi_log_entry& info, void *param) | |
2663 | { | |
2664 | list<rgw_bi_log_entry> *l = (list<rgw_bi_log_entry> *)param; | |
2665 | l->push_back(info); | |
2666 | return 0; | |
2667 | } | |
2668 | ||
2669 | static int bi_log_list_entries(cls_method_context_t hctx, const string& marker, | |
2670 | uint32_t max, list<rgw_bi_log_entry>& entries, bool *truncated) | |
2671 | { | |
2672 | string key_iter; | |
2673 | string end_marker; | |
2674 | int ret = bi_log_iterate_entries(hctx, marker, end_marker, | |
2675 | key_iter, max, truncated, | |
2676 | bi_log_list_cb, &entries); | |
2677 | return ret; | |
2678 | } | |
2679 | ||
2680 | static int rgw_bi_log_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2681 | { | |
2682 | bufferlist::iterator in_iter = in->begin(); | |
2683 | ||
2684 | cls_rgw_bi_log_list_op op; | |
2685 | try { | |
2686 | ::decode(op, in_iter); | |
2687 | } catch (buffer::error& err) { | |
2688 | CLS_LOG(1, "ERROR: rgw_bi_log_list(): failed to decode entry\n"); | |
2689 | return -EINVAL; | |
2690 | } | |
2691 | ||
2692 | cls_rgw_bi_log_list_ret op_ret; | |
2693 | int ret = bi_log_list_entries(hctx, op.marker, op.max, op_ret.entries, &op_ret.truncated); | |
2694 | if (ret < 0) | |
2695 | return ret; | |
2696 | ||
2697 | ::encode(op_ret, *out); | |
2698 | ||
2699 | return 0; | |
2700 | } | |
2701 | ||
2702 | static int bi_log_list_trim_cb(cls_method_context_t hctx, const string& key, rgw_bi_log_entry& info, void *param) | |
2703 | { | |
2704 | list<rgw_bi_log_entry> *entries = (list<rgw_bi_log_entry> *)param; | |
2705 | ||
2706 | entries->push_back(info); | |
2707 | return 0; | |
2708 | } | |
2709 | ||
2710 | static int bi_log_remove_entry(cls_method_context_t hctx, rgw_bi_log_entry& entry) | |
2711 | { | |
2712 | string key; | |
2713 | key = BI_PREFIX_CHAR; | |
2714 | key.append(bucket_index_prefixes[BI_BUCKET_LOG_INDEX]); | |
2715 | key.append(entry.id); | |
2716 | return cls_cxx_map_remove_key(hctx, key); | |
2717 | } | |
2718 | ||
2719 | static int bi_log_list_trim_entries(cls_method_context_t hctx, | |
2720 | const string& start_marker, const string& end_marker, | |
2721 | list<rgw_bi_log_entry>& entries, bool *truncated) | |
2722 | { | |
2723 | string key_iter; | |
2724 | #define MAX_TRIM_ENTRIES 1000 /* max entries to trim in a single operation */ | |
2725 | int ret = bi_log_iterate_entries(hctx, start_marker, end_marker, | |
2726 | key_iter, MAX_TRIM_ENTRIES, truncated, | |
2727 | bi_log_list_trim_cb, &entries); | |
2728 | return ret; | |
2729 | } | |
2730 | ||
2731 | static int rgw_bi_log_trim(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2732 | { | |
2733 | bufferlist::iterator in_iter = in->begin(); | |
2734 | ||
2735 | cls_rgw_bi_log_trim_op op; | |
2736 | try { | |
2737 | ::decode(op, in_iter); | |
2738 | } catch (buffer::error& err) { | |
2739 | CLS_LOG(1, "ERROR: rgw_bi_log_list(): failed to decode entry\n"); | |
2740 | return -EINVAL; | |
2741 | } | |
2742 | ||
2743 | cls_rgw_bi_log_list_ret op_ret; | |
2744 | list<rgw_bi_log_entry> entries; | |
2745 | #define MAX_TRIM_ENTRIES 1000 /* don't do more than that in a single operation */ | |
2746 | bool truncated; | |
2747 | int ret = bi_log_list_trim_entries(hctx, op.start_marker, op.end_marker, entries, &truncated); | |
2748 | if (ret < 0) | |
2749 | return ret; | |
2750 | ||
2751 | if (entries.empty()) | |
2752 | return -ENODATA; | |
2753 | ||
2754 | list<rgw_bi_log_entry>::iterator iter; | |
2755 | for (iter = entries.begin(); iter != entries.end(); ++iter) { | |
2756 | rgw_bi_log_entry& entry = *iter; | |
2757 | ||
2758 | ret = bi_log_remove_entry(hctx, entry); | |
2759 | if (ret < 0) | |
2760 | return ret; | |
2761 | } | |
2762 | ||
2763 | return 0; | |
2764 | } | |
2765 | ||
c07f9fc5 FG |
2766 | static int rgw_bi_log_resync(cls_method_context_t hctx, bufferlist *in, bufferlist *out) |
2767 | { | |
2768 | struct rgw_bucket_dir_header header; | |
2769 | int rc = read_bucket_header(hctx, &header); | |
2770 | if (rc < 0) { | |
2771 | CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n"); | |
2772 | return rc; | |
2773 | } | |
2774 | ||
2775 | bufferlist bl; | |
2776 | ||
2777 | struct rgw_bi_log_entry entry; | |
2778 | ||
2779 | entry.timestamp = real_clock::now(); | |
2780 | entry.op = RGWModifyOp::CLS_RGW_OP_RESYNC; | |
2781 | entry.state = RGWPendingState::CLS_RGW_STATE_COMPLETE; | |
2782 | ||
2783 | string key; | |
2784 | bi_log_index_key(hctx, key, entry.id, header.ver); | |
2785 | ||
2786 | ::encode(entry, bl); | |
2787 | ||
2788 | if (entry.id > header.max_marker) | |
2789 | header.max_marker = entry.id; | |
2790 | ||
2791 | header.syncstopped = false; | |
2792 | ||
2793 | rc = cls_cxx_map_set_val(hctx, key, &bl); | |
2794 | if (rc < 0) | |
2795 | return rc; | |
2796 | ||
2797 | return write_bucket_header(hctx, &header); | |
2798 | } | |
2799 | ||
2800 | static int rgw_bi_log_stop(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2801 | { | |
2802 | struct rgw_bucket_dir_header header; | |
2803 | int rc = read_bucket_header(hctx, &header); | |
2804 | if (rc < 0) { | |
2805 | CLS_LOG(1, "ERROR: rgw_bucket_complete_op(): failed to read header\n"); | |
2806 | return rc; | |
2807 | } | |
2808 | ||
2809 | bufferlist bl; | |
2810 | ||
2811 | struct rgw_bi_log_entry entry; | |
2812 | ||
2813 | entry.timestamp = real_clock::now(); | |
2814 | entry.op = RGWModifyOp::CLS_RGW_OP_SYNCSTOP; | |
2815 | entry.state = RGWPendingState::CLS_RGW_STATE_COMPLETE; | |
2816 | ||
2817 | string key; | |
2818 | bi_log_index_key(hctx, key, entry.id, header.ver); | |
2819 | ||
2820 | ::encode(entry, bl); | |
2821 | ||
2822 | if (entry.id > header.max_marker) | |
2823 | header.max_marker = entry.id; | |
2824 | header.syncstopped = true; | |
2825 | ||
2826 | rc = cls_cxx_map_set_val(hctx, key, &bl); | |
2827 | if (rc < 0) | |
2828 | return rc; | |
2829 | ||
2830 | return write_bucket_header(hctx, &header); | |
2831 | } | |
2832 | ||
2833 | ||
7c673cae FG |
2834 | static void usage_record_prefix_by_time(uint64_t epoch, string& key) |
2835 | { | |
2836 | char buf[32]; | |
2837 | snprintf(buf, sizeof(buf), "%011llu", (long long unsigned)epoch); | |
2838 | key = buf; | |
2839 | } | |
2840 | ||
2841 | static void usage_record_prefix_by_user(string& user, uint64_t epoch, string& key) | |
2842 | { | |
2843 | char buf[user.size() + 32]; | |
2844 | snprintf(buf, sizeof(buf), "%s_%011llu_", user.c_str(), (long long unsigned)epoch); | |
2845 | key = buf; | |
2846 | } | |
2847 | ||
2848 | static void usage_record_name_by_time(uint64_t epoch, const string& user, string& bucket, string& key) | |
2849 | { | |
2850 | char buf[32 + user.size() + bucket.size()]; | |
2851 | snprintf(buf, sizeof(buf), "%011llu_%s_%s", (long long unsigned)epoch, user.c_str(), bucket.c_str()); | |
2852 | key = buf; | |
2853 | } | |
2854 | ||
2855 | static void usage_record_name_by_user(const string& user, uint64_t epoch, string& bucket, string& key) | |
2856 | { | |
2857 | char buf[32 + user.size() + bucket.size()]; | |
2858 | snprintf(buf, sizeof(buf), "%s_%011llu_%s", user.c_str(), (long long unsigned)epoch, bucket.c_str()); | |
2859 | key = buf; | |
2860 | } | |
2861 | ||
2862 | static int usage_record_decode(bufferlist& record_bl, rgw_usage_log_entry& e) | |
2863 | { | |
2864 | bufferlist::iterator kiter = record_bl.begin(); | |
2865 | try { | |
2866 | ::decode(e, kiter); | |
2867 | } catch (buffer::error& err) { | |
2868 | CLS_LOG(1, "ERROR: usage_record_decode(): failed to decode record_bl\n"); | |
2869 | return -EINVAL; | |
2870 | } | |
2871 | ||
2872 | return 0; | |
2873 | } | |
2874 | ||
2875 | int rgw_user_usage_log_add(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
2876 | { | |
2877 | CLS_LOG(10, "rgw_user_usage_log_add()"); | |
2878 | ||
2879 | bufferlist::iterator in_iter = in->begin(); | |
2880 | rgw_cls_usage_log_add_op op; | |
2881 | ||
2882 | try { | |
2883 | ::decode(op, in_iter); | |
2884 | } catch (buffer::error& err) { | |
2885 | CLS_LOG(1, "ERROR: rgw_user_usage_log_add(): failed to decode request\n"); | |
2886 | return -EINVAL; | |
2887 | } | |
2888 | ||
2889 | rgw_usage_log_info& info = op.info; | |
2890 | vector<rgw_usage_log_entry>::iterator iter; | |
2891 | ||
2892 | for (iter = info.entries.begin(); iter != info.entries.end(); ++iter) { | |
2893 | rgw_usage_log_entry& entry = *iter; | |
2894 | string key_by_time; | |
2895 | ||
2896 | rgw_user *puser = (entry.payer.empty() ? &entry.owner : &entry.payer); | |
2897 | ||
2898 | usage_record_name_by_time(entry.epoch, puser->to_str(), entry.bucket, key_by_time); | |
2899 | ||
2900 | CLS_LOG(10, "rgw_user_usage_log_add user=%s bucket=%s\n", puser->to_str().c_str(), entry.bucket.c_str()); | |
2901 | ||
2902 | bufferlist record_bl; | |
2903 | int ret = cls_cxx_map_get_val(hctx, key_by_time, &record_bl); | |
2904 | if (ret < 0 && ret != -ENOENT) { | |
2905 | CLS_LOG(1, "ERROR: rgw_user_usage_log_add(): cls_cxx_map_read_key returned %d\n", ret); | |
2906 | return -EINVAL; | |
2907 | } | |
2908 | if (ret >= 0) { | |
2909 | rgw_usage_log_entry e; | |
2910 | ret = usage_record_decode(record_bl, e); | |
2911 | if (ret < 0) | |
2912 | return ret; | |
2913 | CLS_LOG(10, "rgw_user_usage_log_add aggregating existing bucket\n"); | |
2914 | entry.aggregate(e); | |
2915 | } | |
2916 | ||
2917 | bufferlist new_record_bl; | |
2918 | ::encode(entry, new_record_bl); | |
2919 | ret = cls_cxx_map_set_val(hctx, key_by_time, &new_record_bl); | |
2920 | if (ret < 0) | |
2921 | return ret; | |
2922 | ||
2923 | string key_by_user; | |
2924 | usage_record_name_by_user(puser->to_str(), entry.epoch, entry.bucket, key_by_user); | |
2925 | ret = cls_cxx_map_set_val(hctx, key_by_user, &new_record_bl); | |
2926 | if (ret < 0) | |
2927 | return ret; | |
2928 | } | |
2929 | ||
2930 | return 0; | |
2931 | } | |
2932 | ||
2933 | static int usage_iterate_range(cls_method_context_t hctx, uint64_t start, uint64_t end, | |
2934 | string& user, string& key_iter, uint32_t max_entries, bool *truncated, | |
2935 | int (*cb)(cls_method_context_t, const string&, rgw_usage_log_entry&, void *), | |
2936 | void *param) | |
2937 | { | |
2938 | CLS_LOG(10, "usage_iterate_range"); | |
2939 | ||
2940 | map<string, bufferlist> keys; | |
2941 | #define NUM_KEYS 32 | |
2942 | string filter_prefix; | |
2943 | string start_key, end_key; | |
2944 | bool by_user = !user.empty(); | |
2945 | uint32_t i = 0; | |
2946 | string user_key; | |
3efd9988 | 2947 | bool truncated_status = false; |
7c673cae | 2948 | |
94b18763 FG |
2949 | assert(truncated != nullptr); |
2950 | ||
7c673cae FG |
2951 | if (!by_user) { |
2952 | usage_record_prefix_by_time(end, end_key); | |
2953 | } else { | |
2954 | user_key = user; | |
2955 | user_key.append("_"); | |
2956 | } | |
2957 | ||
2958 | if (key_iter.empty()) { | |
2959 | if (by_user) { | |
2960 | usage_record_prefix_by_user(user, start, start_key); | |
2961 | } else { | |
2962 | usage_record_prefix_by_time(start, start_key); | |
2963 | } | |
2964 | } else { | |
2965 | start_key = key_iter; | |
2966 | } | |
2967 | ||
c07f9fc5 | 2968 | CLS_LOG(20, "usage_iterate_range start_key=%s", start_key.c_str()); |
3efd9988 | 2969 | int ret = cls_cxx_map_get_vals(hctx, start_key, filter_prefix, max_entries, &keys, &truncated_status); |
c07f9fc5 FG |
2970 | if (ret < 0) |
2971 | return ret; | |
7c673cae | 2972 | |
94b18763 | 2973 | *truncated = truncated_status; |
3efd9988 | 2974 | |
c07f9fc5 FG |
2975 | map<string, bufferlist>::iterator iter = keys.begin(); |
2976 | if (iter == keys.end()) | |
2977 | return 0; | |
7c673cae | 2978 | |
c07f9fc5 | 2979 | uint32_t num_keys = keys.size(); |
7c673cae | 2980 | |
c07f9fc5 FG |
2981 | for (; iter != keys.end(); ++iter,++i) { |
2982 | const string& key = iter->first; | |
2983 | rgw_usage_log_entry e; | |
7c673cae | 2984 | |
c07f9fc5 FG |
2985 | if (!by_user && key.compare(end_key) >= 0) { |
2986 | CLS_LOG(20, "usage_iterate_range reached key=%s, done", key.c_str()); | |
94b18763 | 2987 | *truncated = false; |
b32b8144 | 2988 | key_iter = key; |
c07f9fc5 FG |
2989 | return 0; |
2990 | } | |
7c673cae | 2991 | |
c07f9fc5 FG |
2992 | if (by_user && key.compare(0, user_key.size(), user_key) != 0) { |
2993 | CLS_LOG(20, "usage_iterate_range reached key=%s, done", key.c_str()); | |
94b18763 | 2994 | *truncated = false; |
b32b8144 | 2995 | key_iter = key; |
c07f9fc5 FG |
2996 | return 0; |
2997 | } | |
7c673cae | 2998 | |
c07f9fc5 FG |
2999 | ret = usage_record_decode(iter->second, e); |
3000 | if (ret < 0) | |
3001 | return ret; | |
7c673cae | 3002 | |
c07f9fc5 FG |
3003 | if (e.epoch < start) |
3004 | continue; | |
7c673cae | 3005 | |
c07f9fc5 | 3006 | /* keys are sorted by epoch, so once we're past end we're done */ |
94b18763 FG |
3007 | if (e.epoch >= end) { |
3008 | *truncated = false; | |
c07f9fc5 | 3009 | return 0; |
94b18763 | 3010 | } |
7c673cae | 3011 | |
c07f9fc5 FG |
3012 | ret = cb(hctx, key, e, param); |
3013 | if (ret < 0) | |
3014 | return ret; | |
7c673cae | 3015 | |
c07f9fc5 FG |
3016 | |
3017 | if (i == num_keys - 1) { | |
3018 | key_iter = key; | |
3019 | return 0; | |
7c673cae | 3020 | } |
c07f9fc5 | 3021 | } |
7c673cae FG |
3022 | return 0; |
3023 | } | |
3024 | ||
3025 | static int usage_log_read_cb(cls_method_context_t hctx, const string& key, rgw_usage_log_entry& entry, void *param) | |
3026 | { | |
3027 | map<rgw_user_bucket, rgw_usage_log_entry> *usage = (map<rgw_user_bucket, rgw_usage_log_entry> *)param; | |
3028 | rgw_user *puser; | |
3029 | if (!entry.payer.empty()) { | |
3030 | puser = &entry.payer; | |
3031 | } else { | |
3032 | puser = &entry.owner; | |
3033 | } | |
3034 | rgw_user_bucket ub(puser->to_str(), entry.bucket); | |
3035 | rgw_usage_log_entry& le = (*usage)[ub]; | |
3036 | le.aggregate(entry); | |
3037 | ||
3038 | return 0; | |
3039 | } | |
3040 | ||
3041 | int rgw_user_usage_log_read(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3042 | { | |
3043 | CLS_LOG(10, "rgw_user_usage_log_read()"); | |
3044 | ||
3045 | bufferlist::iterator in_iter = in->begin(); | |
3046 | rgw_cls_usage_log_read_op op; | |
3047 | ||
3048 | try { | |
3049 | ::decode(op, in_iter); | |
3050 | } catch (buffer::error& err) { | |
3051 | CLS_LOG(1, "ERROR: rgw_user_usage_log_read(): failed to decode request\n"); | |
3052 | return -EINVAL; | |
3053 | } | |
3054 | ||
3055 | rgw_cls_usage_log_read_ret ret_info; | |
3056 | map<rgw_user_bucket, rgw_usage_log_entry> *usage = &ret_info.usage; | |
3057 | string iter = op.iter; | |
3058 | #define MAX_ENTRIES 1000 | |
3059 | uint32_t max_entries = (op.max_entries ? op.max_entries : MAX_ENTRIES); | |
3060 | int ret = usage_iterate_range(hctx, op.start_epoch, op.end_epoch, op.owner, iter, max_entries, &ret_info.truncated, usage_log_read_cb, (void *)usage); | |
3061 | if (ret < 0) | |
3062 | return ret; | |
3063 | ||
3064 | if (ret_info.truncated) | |
3065 | ret_info.next_iter = iter; | |
3066 | ||
3067 | ::encode(ret_info, *out); | |
3068 | return 0; | |
3069 | } | |
3070 | ||
3071 | static int usage_log_trim_cb(cls_method_context_t hctx, const string& key, rgw_usage_log_entry& entry, void *param) | |
3072 | { | |
94b18763 FG |
3073 | bool *found = (bool *)param; |
3074 | if (found) { | |
3075 | *found = true; | |
3076 | } | |
7c673cae FG |
3077 | string key_by_time; |
3078 | string key_by_user; | |
3079 | ||
3080 | string o = entry.owner.to_str(); | |
3081 | usage_record_name_by_time(entry.epoch, o, entry.bucket, key_by_time); | |
3082 | usage_record_name_by_user(o, entry.epoch, entry.bucket, key_by_user); | |
3083 | ||
3084 | int ret = cls_cxx_map_remove_key(hctx, key_by_time); | |
3085 | if (ret < 0) | |
3086 | return ret; | |
3087 | ||
3088 | return cls_cxx_map_remove_key(hctx, key_by_user); | |
3089 | } | |
3090 | ||
3091 | int rgw_user_usage_log_trim(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3092 | { | |
3093 | CLS_LOG(10, "rgw_user_usage_log_trim()"); | |
3094 | ||
3095 | /* only continue if object exists! */ | |
3096 | int ret = cls_cxx_stat(hctx, NULL, NULL); | |
3097 | if (ret < 0) | |
3098 | return ret; | |
3099 | ||
3100 | bufferlist::iterator in_iter = in->begin(); | |
3101 | rgw_cls_usage_log_trim_op op; | |
3102 | ||
3103 | try { | |
3104 | ::decode(op, in_iter); | |
3105 | } catch (buffer::error& err) { | |
3106 | CLS_LOG(1, "ERROR: rgw_user_log_usage_log_trim(): failed to decode request\n"); | |
3107 | return -EINVAL; | |
3108 | } | |
3109 | ||
3110 | string iter; | |
c07f9fc5 | 3111 | bool more; |
94b18763 | 3112 | bool found = false; |
c07f9fc5 | 3113 | #define MAX_USAGE_TRIM_ENTRIES 128 |
94b18763 | 3114 | ret = usage_iterate_range(hctx, op.start_epoch, op.end_epoch, op.user, iter, MAX_USAGE_TRIM_ENTRIES, &more, usage_log_trim_cb, (void *)&found); |
7c673cae FG |
3115 | if (ret < 0) |
3116 | return ret; | |
3117 | ||
94b18763 | 3118 | if (!more && !found) |
b32b8144 FG |
3119 | return -ENODATA; |
3120 | ||
7c673cae FG |
3121 | return 0; |
3122 | } | |
3123 | ||
3124 | /* | |
3125 | * We hold the garbage collection chain data under two different indexes: the first 'name' index | |
3126 | * keeps them under a unique tag that represents the chains, and a second 'time' index keeps | |
3127 | * them by their expiration timestamp | |
3128 | */ | |
3129 | #define GC_OBJ_NAME_INDEX 0 | |
3130 | #define GC_OBJ_TIME_INDEX 1 | |
3131 | ||
3132 | static string gc_index_prefixes[] = { "0_", | |
3133 | "1_" }; | |
3134 | ||
3135 | static void prepend_index_prefix(const string& src, int index, string *dest) | |
3136 | { | |
3137 | *dest = gc_index_prefixes[index]; | |
3138 | dest->append(src); | |
3139 | } | |
3140 | ||
3141 | static int gc_omap_get(cls_method_context_t hctx, int type, const string& key, cls_rgw_gc_obj_info *info) | |
3142 | { | |
3143 | string index; | |
3144 | prepend_index_prefix(key, type, &index); | |
3145 | ||
3146 | bufferlist bl; | |
3147 | int ret = cls_cxx_map_get_val(hctx, index, &bl); | |
3148 | if (ret < 0) | |
3149 | return ret; | |
3150 | ||
3151 | try { | |
3152 | bufferlist::iterator iter = bl.begin(); | |
3153 | ::decode(*info, iter); | |
3154 | } catch (buffer::error& err) { | |
3155 | CLS_LOG(0, "ERROR: rgw_cls_gc_omap_get(): failed to decode index=%s\n", index.c_str()); | |
3156 | } | |
3157 | ||
3158 | return 0; | |
3159 | } | |
3160 | ||
3161 | static int gc_omap_set(cls_method_context_t hctx, int type, const string& key, const cls_rgw_gc_obj_info *info) | |
3162 | { | |
3163 | bufferlist bl; | |
3164 | ::encode(*info, bl); | |
3165 | ||
3166 | string index = gc_index_prefixes[type]; | |
3167 | index.append(key); | |
3168 | ||
3169 | int ret = cls_cxx_map_set_val(hctx, index, &bl); | |
3170 | if (ret < 0) | |
3171 | return ret; | |
3172 | ||
3173 | return 0; | |
3174 | } | |
3175 | ||
3176 | static int gc_omap_remove(cls_method_context_t hctx, int type, const string& key) | |
3177 | { | |
3178 | string index = gc_index_prefixes[type]; | |
3179 | index.append(key); | |
3180 | ||
3181 | int ret = cls_cxx_map_remove_key(hctx, index); | |
3182 | if (ret < 0) | |
3183 | return ret; | |
3184 | ||
3185 | return 0; | |
3186 | } | |
3187 | ||
3188 | static bool key_in_index(const string& key, int index_type) | |
3189 | { | |
3190 | const string& prefix = gc_index_prefixes[index_type]; | |
3191 | return (key.compare(0, prefix.size(), prefix) == 0); | |
3192 | } | |
3193 | ||
3194 | ||
3195 | static int gc_update_entry(cls_method_context_t hctx, uint32_t expiration_secs, | |
3196 | cls_rgw_gc_obj_info& info) | |
3197 | { | |
3198 | cls_rgw_gc_obj_info old_info; | |
3199 | int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, info.tag, &old_info); | |
3200 | if (ret == 0) { | |
3201 | string key; | |
3202 | get_time_key(old_info.time, &key); | |
3203 | ret = gc_omap_remove(hctx, GC_OBJ_TIME_INDEX, key); | |
3204 | if (ret < 0 && ret != -ENOENT) { | |
3205 | CLS_LOG(0, "ERROR: failed to remove key=%s\n", key.c_str()); | |
3206 | return ret; | |
3207 | } | |
3208 | } | |
3209 | info.time = ceph::real_clock::now(); | |
3210 | info.time += make_timespan(expiration_secs); | |
3211 | ret = gc_omap_set(hctx, GC_OBJ_NAME_INDEX, info.tag, &info); | |
3212 | if (ret < 0) | |
3213 | return ret; | |
3214 | ||
3215 | string key; | |
3216 | get_time_key(info.time, &key); | |
3217 | ret = gc_omap_set(hctx, GC_OBJ_TIME_INDEX, key, &info); | |
3218 | if (ret < 0) | |
3219 | goto done_err; | |
3220 | ||
3221 | return 0; | |
3222 | ||
3223 | done_err: | |
3224 | CLS_LOG(0, "ERROR: gc_set_entry error info.tag=%s, ret=%d\n", info.tag.c_str(), ret); | |
3225 | gc_omap_remove(hctx, GC_OBJ_NAME_INDEX, info.tag); | |
3226 | return ret; | |
3227 | } | |
3228 | ||
3229 | static int gc_defer_entry(cls_method_context_t hctx, const string& tag, uint32_t expiration_secs) | |
3230 | { | |
3231 | cls_rgw_gc_obj_info info; | |
3232 | int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, tag, &info); | |
3233 | if (ret == -ENOENT) | |
3234 | return 0; | |
3235 | if (ret < 0) | |
3236 | return ret; | |
3237 | return gc_update_entry(hctx, expiration_secs, info); | |
3238 | } | |
3239 | ||
3240 | int gc_record_decode(bufferlist& bl, cls_rgw_gc_obj_info& e) | |
3241 | { | |
3242 | bufferlist::iterator iter = bl.begin(); | |
3243 | try { | |
3244 | ::decode(e, iter); | |
3245 | } catch (buffer::error& err) { | |
3246 | CLS_LOG(0, "ERROR: failed to decode cls_rgw_gc_obj_info"); | |
3247 | return -EIO; | |
3248 | } | |
3249 | return 0; | |
3250 | } | |
3251 | ||
3252 | static int rgw_cls_gc_set_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3253 | { | |
3254 | bufferlist::iterator in_iter = in->begin(); | |
3255 | ||
3256 | cls_rgw_gc_set_entry_op op; | |
3257 | try { | |
3258 | ::decode(op, in_iter); | |
3259 | } catch (buffer::error& err) { | |
3260 | CLS_LOG(1, "ERROR: rgw_cls_gc_set_entry(): failed to decode entry\n"); | |
3261 | return -EINVAL; | |
3262 | } | |
3263 | ||
3264 | return gc_update_entry(hctx, op.expiration_secs, op.info); | |
3265 | } | |
3266 | ||
3267 | static int rgw_cls_gc_defer_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3268 | { | |
3269 | bufferlist::iterator in_iter = in->begin(); | |
3270 | ||
3271 | cls_rgw_gc_defer_entry_op op; | |
3272 | try { | |
3273 | ::decode(op, in_iter); | |
3274 | } catch (buffer::error& err) { | |
3275 | CLS_LOG(1, "ERROR: rgw_cls_gc_defer_entry(): failed to decode entry\n"); | |
3276 | return -EINVAL; | |
3277 | } | |
3278 | ||
3279 | return gc_defer_entry(hctx, op.tag, op.expiration_secs); | |
3280 | } | |
3281 | ||
3282 | static int gc_iterate_entries(cls_method_context_t hctx, const string& marker, bool expired_only, | |
3283 | string& key_iter, uint32_t max_entries, bool *truncated, | |
3284 | int (*cb)(cls_method_context_t, const string&, cls_rgw_gc_obj_info&, void *), | |
3285 | void *param) | |
3286 | { | |
3287 | CLS_LOG(10, "gc_iterate_range"); | |
3288 | ||
3289 | map<string, bufferlist> keys; | |
3290 | string filter_prefix, end_key; | |
3291 | uint32_t i = 0; | |
3292 | string key; | |
3293 | ||
3294 | if (truncated) | |
3295 | *truncated = false; | |
3296 | ||
3297 | string start_key; | |
31f18b77 | 3298 | if (marker.empty()) { |
7c673cae FG |
3299 | prepend_index_prefix(marker, GC_OBJ_TIME_INDEX, &start_key); |
3300 | } else { | |
31f18b77 | 3301 | start_key = marker; |
7c673cae FG |
3302 | } |
3303 | ||
3304 | if (expired_only) { | |
3305 | real_time now = ceph::real_clock::now(); | |
3306 | string now_str; | |
3307 | get_time_key(now, &now_str); | |
3308 | prepend_index_prefix(now_str, GC_OBJ_TIME_INDEX, &end_key); | |
3309 | ||
91327a77 | 3310 | CLS_LOG(10, "gc_iterate_entries end_key=%s\n", end_key.c_str()); |
7c673cae FG |
3311 | } |
3312 | ||
3313 | string filter; | |
3314 | ||
c07f9fc5 FG |
3315 | int ret = cls_cxx_map_get_vals(hctx, start_key, filter, max_entries, &keys, truncated); |
3316 | if (ret < 0) | |
3317 | return ret; | |
7c673cae FG |
3318 | |
3319 | ||
c07f9fc5 FG |
3320 | map<string, bufferlist>::iterator iter = keys.begin(); |
3321 | if (iter == keys.end()) | |
3322 | return 0; | |
7c673cae | 3323 | |
c07f9fc5 | 3324 | uint32_t num_keys = keys.size(); |
7c673cae | 3325 | |
c07f9fc5 FG |
3326 | for (; iter != keys.end(); ++iter, ++i) { |
3327 | const string& key = iter->first; | |
3328 | cls_rgw_gc_obj_info e; | |
7c673cae | 3329 | |
c07f9fc5 | 3330 | CLS_LOG(10, "gc_iterate_entries key=%s\n", key.c_str()); |
7c673cae | 3331 | |
1adf2230 AA |
3332 | if (!end_key.empty() && key.compare(end_key) >= 0) { |
3333 | if (truncated) | |
3334 | *truncated = false; | |
c07f9fc5 | 3335 | return 0; |
1adf2230 | 3336 | } |
7c673cae | 3337 | |
c07f9fc5 FG |
3338 | if (!key_in_index(key, GC_OBJ_TIME_INDEX)) |
3339 | return 0; | |
7c673cae | 3340 | |
c07f9fc5 FG |
3341 | ret = gc_record_decode(iter->second, e); |
3342 | if (ret < 0) | |
3343 | return ret; | |
7c673cae | 3344 | |
c07f9fc5 FG |
3345 | ret = cb(hctx, key, e, param); |
3346 | if (ret < 0) | |
3347 | return ret; | |
7c673cae | 3348 | |
c07f9fc5 FG |
3349 | if (i == num_keys - 1) { |
3350 | key_iter = key; | |
7c673cae | 3351 | } |
c07f9fc5 FG |
3352 | } |
3353 | ||
7c673cae FG |
3354 | return 0; |
3355 | } | |
3356 | ||
3357 | static int gc_list_cb(cls_method_context_t hctx, const string& key, cls_rgw_gc_obj_info& info, void *param) | |
3358 | { | |
3359 | list<cls_rgw_gc_obj_info> *l = (list<cls_rgw_gc_obj_info> *)param; | |
3360 | l->push_back(info); | |
3361 | return 0; | |
3362 | } | |
3363 | ||
3364 | static int gc_list_entries(cls_method_context_t hctx, const string& marker, | |
3365 | uint32_t max, bool expired_only, | |
31f18b77 | 3366 | list<cls_rgw_gc_obj_info>& entries, bool *truncated, string& next_marker) |
7c673cae | 3367 | { |
7c673cae | 3368 | int ret = gc_iterate_entries(hctx, marker, expired_only, |
31f18b77 | 3369 | next_marker, max, truncated, |
7c673cae FG |
3370 | gc_list_cb, &entries); |
3371 | return ret; | |
3372 | } | |
3373 | ||
3374 | static int rgw_cls_gc_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3375 | { | |
3376 | bufferlist::iterator in_iter = in->begin(); | |
3377 | ||
3378 | cls_rgw_gc_list_op op; | |
3379 | try { | |
3380 | ::decode(op, in_iter); | |
3381 | } catch (buffer::error& err) { | |
3382 | CLS_LOG(1, "ERROR: rgw_cls_gc_list(): failed to decode entry\n"); | |
3383 | return -EINVAL; | |
3384 | } | |
3385 | ||
3386 | cls_rgw_gc_list_ret op_ret; | |
c07f9fc5 FG |
3387 | #define GC_LIST_ENTRIES_DEFAULT 128 |
3388 | int ret = gc_list_entries(hctx, op.marker, (op.max ? op.max : GC_LIST_ENTRIES_DEFAULT), op.expired_only, | |
31f18b77 | 3389 | op_ret.entries, &op_ret.truncated, op_ret.next_marker); |
7c673cae FG |
3390 | if (ret < 0) |
3391 | return ret; | |
3392 | ||
3393 | ::encode(op_ret, *out); | |
3394 | ||
3395 | return 0; | |
3396 | } | |
3397 | ||
3398 | static int gc_remove(cls_method_context_t hctx, list<string>& tags) | |
3399 | { | |
3400 | list<string>::iterator iter; | |
3401 | ||
3402 | for (iter = tags.begin(); iter != tags.end(); ++iter) { | |
3403 | string& tag = *iter; | |
3404 | cls_rgw_gc_obj_info info; | |
3405 | int ret = gc_omap_get(hctx, GC_OBJ_NAME_INDEX, tag, &info); | |
3406 | if (ret == -ENOENT) { | |
3407 | CLS_LOG(0, "couldn't find tag in name index tag=%s\n", tag.c_str()); | |
3408 | continue; | |
3409 | } | |
3410 | ||
3411 | if (ret < 0) | |
3412 | return ret; | |
3413 | ||
3414 | string time_key; | |
3415 | get_time_key(info.time, &time_key); | |
3416 | ret = gc_omap_remove(hctx, GC_OBJ_TIME_INDEX, time_key); | |
3417 | if (ret < 0 && ret != -ENOENT) | |
3418 | return ret; | |
3419 | if (ret == -ENOENT) { | |
3420 | CLS_LOG(0, "couldn't find key in time index key=%s\n", time_key.c_str()); | |
3421 | } | |
3422 | ||
3423 | ret = gc_omap_remove(hctx, GC_OBJ_NAME_INDEX, tag); | |
3424 | if (ret < 0 && ret != -ENOENT) | |
3425 | return ret; | |
3426 | } | |
3427 | ||
3428 | return 0; | |
3429 | } | |
3430 | ||
3431 | static int rgw_cls_gc_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3432 | { | |
3433 | bufferlist::iterator in_iter = in->begin(); | |
3434 | ||
3435 | cls_rgw_gc_remove_op op; | |
3436 | try { | |
3437 | ::decode(op, in_iter); | |
3438 | } catch (buffer::error& err) { | |
3439 | CLS_LOG(1, "ERROR: rgw_cls_gc_remove(): failed to decode entry\n"); | |
3440 | return -EINVAL; | |
3441 | } | |
3442 | ||
3443 | return gc_remove(hctx, op.tags); | |
3444 | } | |
3445 | ||
3446 | static int rgw_cls_lc_set_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3447 | { | |
3448 | bufferlist::iterator in_iter = in->begin(); | |
3449 | ||
3450 | cls_rgw_lc_set_entry_op op; | |
3451 | try { | |
3452 | ::decode(op, in_iter); | |
3453 | } catch (buffer::error& err) { | |
3454 | CLS_LOG(1, "ERROR: rgw_cls_lc_set_entry(): failed to decode entry\n"); | |
3455 | return -EINVAL; | |
3456 | } | |
3457 | ||
3458 | bufferlist bl; | |
3459 | ::encode(op.entry, bl); | |
3460 | ||
3461 | int ret = cls_cxx_map_set_val(hctx, op.entry.first, &bl); | |
3462 | return ret; | |
3463 | } | |
3464 | ||
3465 | static int rgw_cls_lc_rm_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3466 | { | |
3467 | bufferlist::iterator in_iter = in->begin(); | |
3468 | ||
3469 | cls_rgw_lc_rm_entry_op op; | |
3470 | try { | |
3471 | ::decode(op, in_iter); | |
3472 | } catch (buffer::error& err) { | |
3473 | CLS_LOG(1, "ERROR: rgw_cls_lc_rm_entry(): failed to decode entry\n"); | |
3474 | return -EINVAL; | |
3475 | } | |
3476 | ||
3477 | bufferlist bl; | |
3478 | ::encode(op.entry, bl); | |
3479 | ||
3480 | int ret = cls_cxx_map_remove_key(hctx, op.entry.first); | |
3481 | return ret; | |
3482 | } | |
3483 | ||
3484 | static int rgw_cls_lc_get_next_entry(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3485 | { | |
3486 | bufferlist::iterator in_iter = in->begin(); | |
3487 | cls_rgw_lc_get_next_entry_ret op_ret; | |
3488 | cls_rgw_lc_get_next_entry_op op; | |
3489 | try { | |
3490 | ::decode(op, in_iter); | |
3491 | } catch (buffer::error& err) { | |
3492 | CLS_LOG(1, "ERROR: rgw_cls_lc_get_next_entry: failed to decode op\n"); | |
3493 | return -EINVAL; | |
3494 | } | |
3495 | ||
3496 | map<string, bufferlist> vals; | |
3497 | string filter_prefix; | |
c07f9fc5 FG |
3498 | bool more; |
3499 | int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, 1, &vals, &more); | |
7c673cae FG |
3500 | if (ret < 0) |
3501 | return ret; | |
3502 | map<string, bufferlist>::iterator it; | |
3503 | pair<string, int> entry; | |
3504 | if (!vals.empty()) { | |
3505 | it=vals.begin(); | |
3506 | in_iter = it->second.begin(); | |
3507 | try { | |
3508 | ::decode(entry, in_iter); | |
3509 | } catch (buffer::error& err) { | |
3510 | CLS_LOG(1, "ERROR: rgw_cls_lc_get_next_entry(): failed to decode entry\n"); | |
3511 | return -EIO; | |
3512 | } | |
3513 | } | |
3514 | op_ret.entry = entry; | |
3515 | ::encode(op_ret, *out); | |
3516 | return 0; | |
3517 | } | |
3518 | ||
3519 | static int rgw_cls_lc_list_entries(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3520 | { | |
3521 | cls_rgw_lc_list_entries_op op; | |
3522 | bufferlist::iterator in_iter = in->begin(); | |
3523 | try { | |
3524 | ::decode(op, in_iter); | |
3525 | } catch (buffer::error& err) { | |
3526 | CLS_LOG(1, "ERROR: rgw_cls_lc_list_entries(): failed to decode op\n"); | |
3527 | return -EINVAL; | |
3528 | } | |
3529 | ||
3530 | cls_rgw_lc_list_entries_ret op_ret; | |
3531 | bufferlist::iterator iter; | |
3532 | map<string, bufferlist> vals; | |
3533 | string filter_prefix; | |
c07f9fc5 | 3534 | int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, op.max_entries, &vals, &op_ret.is_truncated); |
7c673cae FG |
3535 | if (ret < 0) |
3536 | return ret; | |
3537 | map<string, bufferlist>::iterator it; | |
3538 | pair<string, int> entry; | |
3539 | for (it = vals.begin(); it != vals.end(); ++it) { | |
3540 | iter = it->second.begin(); | |
3541 | try { | |
3542 | ::decode(entry, iter); | |
3543 | } catch (buffer::error& err) { | |
3544 | CLS_LOG(1, "ERROR: rgw_cls_lc_list_entries(): failed to decode entry\n"); | |
3545 | return -EIO; | |
3546 | } | |
3547 | op_ret.entries.insert(entry); | |
3548 | } | |
3549 | ::encode(op_ret, *out); | |
3550 | return 0; | |
3551 | } | |
3552 | ||
3553 | static int rgw_cls_lc_put_head(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3554 | { | |
3555 | bufferlist::iterator in_iter = in->begin(); | |
3556 | ||
3557 | cls_rgw_lc_put_head_op op; | |
3558 | try { | |
3559 | ::decode(op, in_iter); | |
3560 | } catch (buffer::error& err) { | |
3561 | CLS_LOG(1, "ERROR: rgw_cls_lc_put_head(): failed to decode entry\n"); | |
3562 | return -EINVAL; | |
3563 | } | |
3564 | ||
3565 | bufferlist bl; | |
3566 | ::encode(op.head, bl); | |
3567 | int ret = cls_cxx_map_write_header(hctx,&bl); | |
3568 | return ret; | |
3569 | } | |
3570 | ||
3571 | static int rgw_cls_lc_get_head(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3572 | { | |
3573 | bufferlist bl; | |
3574 | int ret = cls_cxx_map_read_header(hctx, &bl); | |
3575 | if (ret < 0) | |
3576 | return ret; | |
3577 | cls_rgw_lc_obj_head head; | |
3578 | if (bl.length() != 0) { | |
3579 | bufferlist::iterator iter = bl.begin(); | |
3580 | try { | |
3581 | ::decode(head, iter); | |
3582 | } catch (buffer::error& err) { | |
3583 | CLS_LOG(0, "ERROR: rgw_cls_lc_get_head(): failed to decode entry %s\n",err.what()); | |
3584 | return -EINVAL; | |
3585 | } | |
3586 | } else { | |
3587 | head.start_date = 0; | |
3588 | head.marker.clear(); | |
3589 | } | |
3590 | cls_rgw_lc_get_head_ret op_ret; | |
3591 | op_ret.head = head; | |
3592 | ::encode(op_ret, *out); | |
3593 | return 0; | |
3594 | } | |
3595 | ||
31f18b77 FG |
3596 | static int rgw_reshard_add(cls_method_context_t hctx, bufferlist *in, bufferlist *out) |
3597 | { | |
3598 | bufferlist::iterator in_iter = in->begin(); | |
3599 | ||
3600 | cls_rgw_reshard_add_op op; | |
3601 | try { | |
3602 | ::decode(op, in_iter); | |
3603 | } catch (buffer::error& err) { | |
3604 | CLS_LOG(1, "ERROR: rgw_reshard_add: failed to decode entry\n"); | |
3605 | return -EINVAL; | |
3606 | } | |
3607 | ||
3608 | ||
3609 | string key; | |
3610 | op.entry.get_key(&key); | |
3611 | ||
3612 | bufferlist bl; | |
3613 | ::encode(op.entry, bl); | |
3614 | int ret = cls_cxx_map_set_val(hctx, key, &bl); | |
3615 | if (ret < 0) { | |
3616 | CLS_ERR("error adding reshard job for bucket %s with key %s",op.entry.bucket_name.c_str(), key.c_str()); | |
3617 | return ret; | |
3618 | } | |
3619 | ||
3620 | return ret; | |
3621 | } | |
3622 | ||
3623 | static int rgw_reshard_list(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3624 | { | |
3625 | cls_rgw_reshard_list_op op; | |
3626 | bufferlist::iterator in_iter = in->begin(); | |
3627 | try { | |
3628 | ::decode(op, in_iter); | |
3629 | } catch (buffer::error& err) { | |
3630 | CLS_LOG(1, "ERROR: rgw_cls_rehard_list(): failed to decode entry\n"); | |
3631 | return -EINVAL; | |
3632 | } | |
3633 | cls_rgw_reshard_list_ret op_ret; | |
3634 | bufferlist::iterator iter; | |
3635 | map<string, bufferlist> vals; | |
3636 | string filter_prefix; | |
3637 | #define MAX_RESHARD_LIST_ENTRIES 1000 | |
3638 | /* one extra entry for identifying truncation */ | |
c07f9fc5 FG |
3639 | int32_t max = (op.max && (op.max < MAX_RESHARD_LIST_ENTRIES) ? op.max : MAX_RESHARD_LIST_ENTRIES); |
3640 | int ret = cls_cxx_map_get_vals(hctx, op.marker, filter_prefix, max, &vals, &op_ret.is_truncated); | |
31f18b77 FG |
3641 | if (ret < 0) |
3642 | return ret; | |
3643 | map<string, bufferlist>::iterator it; | |
3644 | cls_rgw_reshard_entry entry; | |
3645 | int i = 0; | |
3646 | for (it = vals.begin(); i < (int)op.max && it != vals.end(); ++it, ++i) { | |
3647 | iter = it->second.begin(); | |
3648 | try { | |
3649 | ::decode(entry, iter); | |
3650 | } catch (buffer::error& err) { | |
3651 | CLS_LOG(1, "ERROR: rgw_cls_rehard_list(): failed to decode entry\n"); | |
3652 | return -EIO; | |
3653 | } | |
3654 | op_ret.entries.push_back(entry); | |
3655 | } | |
31f18b77 FG |
3656 | ::encode(op_ret, *out); |
3657 | return 0; | |
3658 | } | |
3659 | ||
3660 | static int get_reshard_entry(cls_method_context_t hctx, const string& key, cls_rgw_reshard_entry *entry) | |
3661 | { | |
3662 | bufferlist bl; | |
3663 | int ret = cls_cxx_map_get_val(hctx, key, &bl); | |
3664 | if (ret < 0) | |
3665 | return ret; | |
3666 | bufferlist::iterator iter = bl.begin(); | |
3667 | try { | |
3668 | ::decode(*entry, iter); | |
3669 | } catch (buffer::error& err) { | |
3670 | CLS_LOG(0, "ERROR: %s : failed to decode entry %s\n", __func__, err.what()); | |
3671 | return -EIO; | |
3672 | } | |
3673 | return 0; | |
3674 | } | |
3675 | ||
3676 | static int rgw_reshard_get(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3677 | { | |
3678 | bufferlist::iterator in_iter = in->begin(); | |
3679 | ||
3680 | cls_rgw_reshard_get_op op; | |
3681 | try { | |
3682 | ::decode(op, in_iter); | |
3683 | } catch (buffer::error& err) { | |
3684 | CLS_LOG(1, "ERROR: rgw_reshard_get: failed to decode entry\n"); | |
3685 | return -EINVAL; | |
3686 | } | |
3687 | ||
3688 | string key; | |
3689 | cls_rgw_reshard_entry entry; | |
3690 | op.entry.get_key(&key); | |
3691 | int ret = get_reshard_entry(hctx, key, &entry); | |
3692 | if (ret < 0) { | |
3693 | return ret; | |
3694 | } | |
3695 | ||
3696 | cls_rgw_reshard_get_ret op_ret; | |
3697 | op_ret.entry = entry; | |
3698 | ::encode(op_ret, *out); | |
3699 | return 0; | |
3700 | } | |
3701 | ||
3702 | static int rgw_reshard_remove(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3703 | { | |
3704 | bufferlist::iterator in_iter = in->begin(); | |
3705 | ||
3706 | cls_rgw_reshard_remove_op op; | |
3707 | try { | |
3708 | ::decode(op, in_iter); | |
3709 | } catch (buffer::error& err) { | |
3710 | CLS_LOG(1, "ERROR: rgw_cls_rehard_remove: failed to decode entry\n"); | |
3711 | return -EINVAL; | |
3712 | } | |
3713 | ||
3714 | string key; | |
3715 | cls_rgw_reshard_entry entry; | |
3716 | cls_rgw_reshard_entry::generate_key(op.tenant, op.bucket_name, &key); | |
3717 | int ret = get_reshard_entry(hctx, key, &entry); | |
3718 | if (ret < 0) { | |
3719 | return ret; | |
3720 | } | |
3721 | ||
3722 | if (!op.bucket_id.empty() && | |
3723 | entry.bucket_id != op.bucket_id) { | |
3724 | return 0; | |
3725 | } | |
3726 | ||
3727 | ret = cls_cxx_map_remove_key(hctx, key); | |
3728 | if (ret < 0) { | |
3729 | CLS_LOG(0, "ERROR: failed to remove key: key=%s ret=%d", key.c_str(), ret); | |
3730 | return 0; | |
3731 | } | |
3732 | return ret; | |
3733 | } | |
3734 | ||
3735 | static int rgw_set_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3736 | { | |
3737 | cls_rgw_set_bucket_resharding_op op; | |
3738 | ||
3739 | bufferlist::iterator in_iter = in->begin(); | |
3740 | try { | |
3741 | ::decode(op, in_iter); | |
3742 | } catch (buffer::error& err) { | |
3743 | CLS_LOG(1, "ERROR: cls_rgw_set_bucket_resharding: failed to decode entry\n"); | |
3744 | return -EINVAL; | |
3745 | } | |
3746 | ||
3747 | struct rgw_bucket_dir_header header; | |
3748 | int rc = read_bucket_header(hctx, &header); | |
3749 | if (rc < 0) { | |
3750 | CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__); | |
3751 | return rc; | |
3752 | } | |
3753 | ||
3754 | header.new_instance.set_status(op.entry.new_bucket_instance_id, op.entry.num_shards, op.entry.reshard_status); | |
3755 | ||
3756 | return write_bucket_header(hctx, &header); | |
3757 | } | |
3758 | ||
3759 | static int rgw_clear_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3760 | { | |
3761 | cls_rgw_set_bucket_resharding_op op; | |
3762 | ||
3763 | bufferlist::iterator in_iter = in->begin(); | |
3764 | try { | |
3765 | ::decode(op, in_iter); | |
3766 | } catch (buffer::error& err) { | |
3767 | CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n"); | |
3768 | return -EINVAL; | |
3769 | } | |
3770 | ||
3771 | struct rgw_bucket_dir_header header; | |
3772 | int rc = read_bucket_header(hctx, &header); | |
3773 | if (rc < 0) { | |
3774 | CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__); | |
3775 | return rc; | |
3776 | } | |
3777 | header.new_instance.clear(); | |
3778 | ||
3779 | return write_bucket_header(hctx, &header); | |
3780 | } | |
3781 | ||
3782 | static int rgw_guard_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3783 | { | |
3784 | cls_rgw_guard_bucket_resharding_op op; | |
3785 | ||
3786 | bufferlist::iterator in_iter = in->begin(); | |
3787 | try { | |
3788 | ::decode(op, in_iter); | |
3789 | } catch (buffer::error& err) { | |
3790 | CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n"); | |
3791 | return -EINVAL; | |
3792 | } | |
3793 | ||
3794 | struct rgw_bucket_dir_header header; | |
3795 | int rc = read_bucket_header(hctx, &header); | |
3796 | if (rc < 0) { | |
3797 | CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__); | |
3798 | return rc; | |
3799 | } | |
3800 | ||
3801 | if (header.resharding()) { | |
3802 | return op.ret_err; | |
3803 | } | |
3804 | ||
3805 | return 0; | |
3806 | } | |
3807 | ||
3808 | static int rgw_get_bucket_resharding(cls_method_context_t hctx, bufferlist *in, bufferlist *out) | |
3809 | { | |
3810 | cls_rgw_get_bucket_resharding_op op; | |
3811 | ||
3812 | bufferlist::iterator in_iter = in->begin(); | |
3813 | try { | |
3814 | ::decode(op, in_iter); | |
3815 | } catch (buffer::error& err) { | |
3816 | CLS_LOG(1, "ERROR: cls_rgw_clear_bucket_resharding: failed to decode entry\n"); | |
3817 | return -EINVAL; | |
3818 | } | |
3819 | ||
3820 | struct rgw_bucket_dir_header header; | |
3821 | int rc = read_bucket_header(hctx, &header); | |
3822 | if (rc < 0) { | |
3823 | CLS_LOG(1, "ERROR: %s(): failed to read header\n", __func__); | |
3824 | return rc; | |
3825 | } | |
3826 | ||
3827 | cls_rgw_get_bucket_resharding_ret op_ret; | |
3828 | op_ret.new_instance = header.new_instance; | |
3829 | ||
3830 | ::encode(op_ret, *out); | |
3831 | ||
3832 | return 0; | |
3833 | } | |
3834 | ||
7c673cae FG |
3835 | CLS_INIT(rgw) |
3836 | { | |
3837 | CLS_LOG(1, "Loaded rgw class!"); | |
3838 | ||
3839 | cls_handle_t h_class; | |
3840 | cls_method_handle_t h_rgw_bucket_init_index; | |
3841 | cls_method_handle_t h_rgw_bucket_set_tag_timeout; | |
3842 | cls_method_handle_t h_rgw_bucket_list; | |
3843 | cls_method_handle_t h_rgw_bucket_check_index; | |
3844 | cls_method_handle_t h_rgw_bucket_rebuild_index; | |
3845 | cls_method_handle_t h_rgw_bucket_update_stats; | |
3846 | cls_method_handle_t h_rgw_bucket_prepare_op; | |
3847 | cls_method_handle_t h_rgw_bucket_complete_op; | |
3848 | cls_method_handle_t h_rgw_bucket_link_olh; | |
3849 | cls_method_handle_t h_rgw_bucket_unlink_instance_op; | |
3850 | cls_method_handle_t h_rgw_bucket_read_olh_log; | |
3851 | cls_method_handle_t h_rgw_bucket_trim_olh_log; | |
3852 | cls_method_handle_t h_rgw_bucket_clear_olh; | |
3853 | cls_method_handle_t h_rgw_obj_remove; | |
3854 | cls_method_handle_t h_rgw_obj_store_pg_ver; | |
3855 | cls_method_handle_t h_rgw_obj_check_attrs_prefix; | |
3856 | cls_method_handle_t h_rgw_obj_check_mtime; | |
3857 | cls_method_handle_t h_rgw_bi_get_op; | |
3858 | cls_method_handle_t h_rgw_bi_put_op; | |
3859 | cls_method_handle_t h_rgw_bi_list_op; | |
3860 | cls_method_handle_t h_rgw_bi_log_list_op; | |
c07f9fc5 FG |
3861 | cls_method_handle_t h_rgw_bi_log_resync_op; |
3862 | cls_method_handle_t h_rgw_bi_log_stop_op; | |
7c673cae FG |
3863 | cls_method_handle_t h_rgw_dir_suggest_changes; |
3864 | cls_method_handle_t h_rgw_user_usage_log_add; | |
3865 | cls_method_handle_t h_rgw_user_usage_log_read; | |
3866 | cls_method_handle_t h_rgw_user_usage_log_trim; | |
3867 | cls_method_handle_t h_rgw_gc_set_entry; | |
3868 | cls_method_handle_t h_rgw_gc_list; | |
3869 | cls_method_handle_t h_rgw_gc_remove; | |
3870 | cls_method_handle_t h_rgw_lc_set_entry; | |
3871 | cls_method_handle_t h_rgw_lc_rm_entry; | |
3872 | cls_method_handle_t h_rgw_lc_get_next_entry; | |
3873 | cls_method_handle_t h_rgw_lc_put_head; | |
3874 | cls_method_handle_t h_rgw_lc_get_head; | |
3875 | cls_method_handle_t h_rgw_lc_list_entries; | |
31f18b77 FG |
3876 | cls_method_handle_t h_rgw_reshard_add; |
3877 | cls_method_handle_t h_rgw_reshard_list; | |
3878 | cls_method_handle_t h_rgw_reshard_get; | |
3879 | cls_method_handle_t h_rgw_reshard_remove; | |
3880 | cls_method_handle_t h_rgw_set_bucket_resharding; | |
3881 | cls_method_handle_t h_rgw_clear_bucket_resharding; | |
3882 | cls_method_handle_t h_rgw_guard_bucket_resharding; | |
3883 | cls_method_handle_t h_rgw_get_bucket_resharding; | |
7c673cae FG |
3884 | |
3885 | ||
3886 | cls_register(RGW_CLASS, &h_class); | |
3887 | ||
3888 | /* bucket index */ | |
3889 | cls_register_cxx_method(h_class, RGW_BUCKET_INIT_INDEX, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_init_index, &h_rgw_bucket_init_index); | |
3890 | cls_register_cxx_method(h_class, RGW_BUCKET_SET_TAG_TIMEOUT, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_set_tag_timeout, &h_rgw_bucket_set_tag_timeout); | |
3891 | cls_register_cxx_method(h_class, RGW_BUCKET_LIST, CLS_METHOD_RD, rgw_bucket_list, &h_rgw_bucket_list); | |
3892 | cls_register_cxx_method(h_class, RGW_BUCKET_CHECK_INDEX, CLS_METHOD_RD, rgw_bucket_check_index, &h_rgw_bucket_check_index); | |
3893 | cls_register_cxx_method(h_class, RGW_BUCKET_REBUILD_INDEX, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_rebuild_index, &h_rgw_bucket_rebuild_index); | |
3894 | cls_register_cxx_method(h_class, RGW_BUCKET_UPDATE_STATS, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_update_stats, &h_rgw_bucket_update_stats); | |
3895 | cls_register_cxx_method(h_class, RGW_BUCKET_PREPARE_OP, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_prepare_op, &h_rgw_bucket_prepare_op); | |
3896 | cls_register_cxx_method(h_class, RGW_BUCKET_COMPLETE_OP, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_complete_op, &h_rgw_bucket_complete_op); | |
3897 | cls_register_cxx_method(h_class, RGW_BUCKET_LINK_OLH, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_link_olh, &h_rgw_bucket_link_olh); | |
3898 | cls_register_cxx_method(h_class, RGW_BUCKET_UNLINK_INSTANCE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_unlink_instance, &h_rgw_bucket_unlink_instance_op); | |
3899 | cls_register_cxx_method(h_class, RGW_BUCKET_READ_OLH_LOG, CLS_METHOD_RD, rgw_bucket_read_olh_log, &h_rgw_bucket_read_olh_log); | |
3900 | cls_register_cxx_method(h_class, RGW_BUCKET_TRIM_OLH_LOG, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_trim_olh_log, &h_rgw_bucket_trim_olh_log); | |
3901 | cls_register_cxx_method(h_class, RGW_BUCKET_CLEAR_OLH, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bucket_clear_olh, &h_rgw_bucket_clear_olh); | |
3902 | ||
3903 | cls_register_cxx_method(h_class, RGW_OBJ_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_obj_remove, &h_rgw_obj_remove); | |
3904 | cls_register_cxx_method(h_class, RGW_OBJ_STORE_PG_VER, CLS_METHOD_WR, rgw_obj_store_pg_ver, &h_rgw_obj_store_pg_ver); | |
3905 | cls_register_cxx_method(h_class, RGW_OBJ_CHECK_ATTRS_PREFIX, CLS_METHOD_RD, rgw_obj_check_attrs_prefix, &h_rgw_obj_check_attrs_prefix); | |
3906 | cls_register_cxx_method(h_class, RGW_OBJ_CHECK_MTIME, CLS_METHOD_RD, rgw_obj_check_mtime, &h_rgw_obj_check_mtime); | |
3907 | ||
3908 | cls_register_cxx_method(h_class, RGW_BI_GET, CLS_METHOD_RD, rgw_bi_get_op, &h_rgw_bi_get_op); | |
3909 | cls_register_cxx_method(h_class, RGW_BI_PUT, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_put_op, &h_rgw_bi_put_op); | |
3910 | cls_register_cxx_method(h_class, RGW_BI_LIST, CLS_METHOD_RD, rgw_bi_list_op, &h_rgw_bi_list_op); | |
3911 | ||
3912 | cls_register_cxx_method(h_class, RGW_BI_LOG_LIST, CLS_METHOD_RD, rgw_bi_log_list, &h_rgw_bi_log_list_op); | |
3913 | cls_register_cxx_method(h_class, RGW_BI_LOG_TRIM, CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_trim, &h_rgw_bi_log_list_op); | |
3914 | cls_register_cxx_method(h_class, RGW_DIR_SUGGEST_CHANGES, CLS_METHOD_RD | CLS_METHOD_WR, rgw_dir_suggest_changes, &h_rgw_dir_suggest_changes); | |
3915 | ||
c07f9fc5 FG |
3916 | cls_register_cxx_method(h_class, "bi_log_resync", CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_resync, &h_rgw_bi_log_resync_op); |
3917 | cls_register_cxx_method(h_class, "bi_log_stop", CLS_METHOD_RD | CLS_METHOD_WR, rgw_bi_log_stop, &h_rgw_bi_log_stop_op); | |
3918 | ||
7c673cae FG |
3919 | /* usage logging */ |
3920 | cls_register_cxx_method(h_class, RGW_USER_USAGE_LOG_ADD, CLS_METHOD_RD | CLS_METHOD_WR, rgw_user_usage_log_add, &h_rgw_user_usage_log_add); | |
3921 | cls_register_cxx_method(h_class, RGW_USER_USAGE_LOG_READ, CLS_METHOD_RD, rgw_user_usage_log_read, &h_rgw_user_usage_log_read); | |
3922 | cls_register_cxx_method(h_class, RGW_USER_USAGE_LOG_TRIM, CLS_METHOD_RD | CLS_METHOD_WR, rgw_user_usage_log_trim, &h_rgw_user_usage_log_trim); | |
3923 | ||
3924 | /* garbage collection */ | |
3925 | cls_register_cxx_method(h_class, RGW_GC_SET_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_set_entry, &h_rgw_gc_set_entry); | |
3926 | cls_register_cxx_method(h_class, RGW_GC_DEFER_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_defer_entry, &h_rgw_gc_set_entry); | |
3927 | cls_register_cxx_method(h_class, RGW_GC_LIST, CLS_METHOD_RD, rgw_cls_gc_list, &h_rgw_gc_list); | |
3928 | cls_register_cxx_method(h_class, RGW_GC_REMOVE, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_gc_remove, &h_rgw_gc_remove); | |
3929 | ||
3930 | /* lifecycle bucket list */ | |
3931 | cls_register_cxx_method(h_class, RGW_LC_SET_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_lc_set_entry, &h_rgw_lc_set_entry); | |
3932 | cls_register_cxx_method(h_class, RGW_LC_RM_ENTRY, CLS_METHOD_RD | CLS_METHOD_WR, rgw_cls_lc_rm_entry, &h_rgw_lc_rm_entry); | |
3933 | cls_register_cxx_method(h_class, RGW_LC_GET_NEXT_ENTRY, CLS_METHOD_RD, rgw_cls_lc_get_next_entry, &h_rgw_lc_get_next_entry); | |
3934 | cls_register_cxx_method(h_class, RGW_LC_PUT_HEAD, CLS_METHOD_RD| CLS_METHOD_WR, rgw_cls_lc_put_head, &h_rgw_lc_put_head); | |
3935 | cls_register_cxx_method(h_class, RGW_LC_GET_HEAD, CLS_METHOD_RD, rgw_cls_lc_get_head, &h_rgw_lc_get_head); | |
3936 | cls_register_cxx_method(h_class, RGW_LC_LIST_ENTRIES, CLS_METHOD_RD, rgw_cls_lc_list_entries, &h_rgw_lc_list_entries); | |
31f18b77 FG |
3937 | cls_register_cxx_method(h_class, "reshard_add", CLS_METHOD_RD | CLS_METHOD_WR, rgw_reshard_add, &h_rgw_reshard_add); |
3938 | cls_register_cxx_method(h_class, "reshard_list", CLS_METHOD_RD, rgw_reshard_list, &h_rgw_reshard_list); | |
3939 | cls_register_cxx_method(h_class, "reshard_get", CLS_METHOD_RD,rgw_reshard_get, &h_rgw_reshard_get); | |
3940 | cls_register_cxx_method(h_class, "reshard_remove", CLS_METHOD_RD | CLS_METHOD_WR, rgw_reshard_remove, &h_rgw_reshard_remove); | |
3941 | cls_register_cxx_method(h_class, "set_bucket_resharding", CLS_METHOD_RD | CLS_METHOD_WR, | |
3942 | rgw_set_bucket_resharding, &h_rgw_set_bucket_resharding); | |
3943 | cls_register_cxx_method(h_class, "clear_bucket_resharding", CLS_METHOD_RD | CLS_METHOD_WR, | |
3944 | rgw_clear_bucket_resharding, &h_rgw_clear_bucket_resharding); | |
3945 | cls_register_cxx_method(h_class, "guard_bucket_resharding", CLS_METHOD_RD , | |
3946 | rgw_guard_bucket_resharding, &h_rgw_guard_bucket_resharding); | |
3947 | cls_register_cxx_method(h_class, "get_bucket_resharding", CLS_METHOD_RD , | |
3948 | rgw_get_bucket_resharding, &h_rgw_get_bucket_resharding); | |
3949 | ||
7c673cae FG |
3950 | return; |
3951 | } | |
3952 |