]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | ||
4 | #ifndef CEPH_RGWRADOS_H | |
5 | #define CEPH_RGWRADOS_H | |
6 | ||
7 | #include <functional> | |
8 | ||
9 | #include "include/rados/librados.hpp" | |
10 | #include "include/Context.h" | |
3a9019d9 | 11 | #include "common/admin_socket.h" |
7c673cae FG |
12 | #include "common/RefCountedObj.h" |
13 | #include "common/RWLock.h" | |
14 | #include "common/ceph_time.h" | |
15 | #include "common/lru_map.h" | |
11fdf7f2 | 16 | #include "common/ceph_json.h" |
7c673cae FG |
17 | #include "rgw_common.h" |
18 | #include "cls/rgw/cls_rgw_types.h" | |
19 | #include "cls/version/cls_version_types.h" | |
20 | #include "cls/log/cls_log_types.h" | |
7c673cae | 21 | #include "cls/timeindex/cls_timeindex_types.h" |
11fdf7f2 | 22 | #include "cls/otp/cls_otp_types.h" |
7c673cae FG |
23 | #include "rgw_log.h" |
24 | #include "rgw_metadata.h" | |
25 | #include "rgw_meta_sync_status.h" | |
26 | #include "rgw_period_puller.h" | |
27 | #include "rgw_sync_module.h" | |
b32b8144 | 28 | #include "rgw_sync_log_trim.h" |
11fdf7f2 TL |
29 | #include "rgw_service.h" |
30 | ||
31 | #include "services/svc_rados.h" | |
32 | #include "services/svc_zone.h" | |
7c673cae FG |
33 | |
34 | class RGWWatcher; | |
35 | class SafeTimer; | |
36 | class ACLOwner; | |
37 | class RGWGC; | |
38 | class RGWMetaNotifier; | |
39 | class RGWDataNotifier; | |
40 | class RGWLC; | |
41 | class RGWObjectExpirer; | |
42 | class RGWMetaSyncProcessorThread; | |
43 | class RGWDataSyncProcessorThread; | |
44 | class RGWSyncLogTrimThread; | |
11fdf7f2 | 45 | class RGWSyncTraceManager; |
7c673cae FG |
46 | struct RGWZoneGroup; |
47 | struct RGWZoneParams; | |
31f18b77 FG |
48 | class RGWReshard; |
49 | class RGWReshardWait; | |
7c673cae | 50 | |
11fdf7f2 TL |
51 | class RGWSysObjectCtx; |
52 | ||
7c673cae FG |
53 | /* flags for put_obj_meta() */ |
54 | #define PUT_OBJ_CREATE 0x01 | |
55 | #define PUT_OBJ_EXCL 0x02 | |
56 | #define PUT_OBJ_CREATE_EXCL (PUT_OBJ_CREATE | PUT_OBJ_EXCL) | |
57 | ||
58 | #define RGW_OBJ_NS_MULTIPART "multipart" | |
59 | #define RGW_OBJ_NS_SHADOW "shadow" | |
60 | ||
61 | #define RGW_BUCKET_INSTANCE_MD_PREFIX ".bucket.meta." | |
62 | ||
63 | #define RGW_NO_SHARD -1 | |
64 | ||
31f18b77 FG |
65 | #define RGW_SHARDS_PRIME_0 7877 |
66 | #define RGW_SHARDS_PRIME_1 65521 | |
67 | ||
11fdf7f2 TL |
68 | extern const std::string MP_META_SUFFIX; |
69 | ||
1adf2230 | 70 | // only called by rgw_shard_id and rgw_bucket_shard_index |
31f18b77 FG |
71 | static inline int rgw_shards_mod(unsigned hval, int max_shards) |
72 | { | |
73 | if (max_shards <= RGW_SHARDS_PRIME_0) { | |
74 | return hval % RGW_SHARDS_PRIME_0 % max_shards; | |
75 | } | |
76 | return hval % RGW_SHARDS_PRIME_1 % max_shards; | |
77 | } | |
78 | ||
1adf2230 AA |
79 | // used for logging and tagging |
80 | static inline int rgw_shard_id(const string& key, int max_shards) | |
31f18b77 | 81 | { |
1adf2230 AA |
82 | return rgw_shards_mod(ceph_str_hash_linux(key.c_str(), key.size()), |
83 | max_shards); | |
84 | } | |
85 | ||
86 | // used for bucket indices | |
87 | static inline uint32_t rgw_bucket_shard_index(const std::string& key, | |
88 | int num_shards) { | |
89 | uint32_t sid = ceph_str_hash_linux(key.c_str(), key.size()); | |
90 | uint32_t sid2 = sid ^ ((sid & 0xFF) << 24); | |
91 | return rgw_shards_mod(sid2, num_shards); | |
31f18b77 FG |
92 | } |
93 | ||
94 | static inline int rgw_shards_max() | |
95 | { | |
96 | return RGW_SHARDS_PRIME_1; | |
97 | } | |
7c673cae FG |
98 | |
99 | static inline void prepend_bucket_marker(const rgw_bucket& bucket, const string& orig_oid, string& oid) | |
100 | { | |
101 | if (bucket.marker.empty() || orig_oid.empty()) { | |
102 | oid = orig_oid; | |
103 | } else { | |
104 | oid = bucket.marker; | |
105 | oid.append("_"); | |
106 | oid.append(orig_oid); | |
107 | } | |
108 | } | |
109 | ||
110 | static inline void get_obj_bucket_and_oid_loc(const rgw_obj& obj, string& oid, string& locator) | |
111 | { | |
112 | const rgw_bucket& bucket = obj.bucket; | |
113 | prepend_bucket_marker(bucket, obj.get_oid(), oid); | |
114 | const string& loc = obj.key.get_loc(); | |
115 | if (!loc.empty()) { | |
116 | prepend_bucket_marker(bucket, loc, locator); | |
117 | } else { | |
118 | locator.clear(); | |
119 | } | |
120 | } | |
121 | ||
7c673cae FG |
122 | int rgw_policy_from_attrset(CephContext *cct, map<string, bufferlist>& attrset, RGWAccessControlPolicy *policy); |
123 | ||
124 | static inline bool rgw_raw_obj_to_obj(const rgw_bucket& bucket, const rgw_raw_obj& raw_obj, rgw_obj *obj) | |
125 | { | |
126 | ssize_t pos = raw_obj.oid.find('_'); | |
127 | if (pos < 0) { | |
128 | return false; | |
129 | } | |
130 | ||
131 | if (!rgw_obj_key::parse_raw_oid(raw_obj.oid.substr(pos + 1), &obj->key)) { | |
132 | return false; | |
133 | } | |
134 | obj->bucket = bucket; | |
135 | ||
136 | return true; | |
137 | } | |
138 | ||
11fdf7f2 | 139 | |
7c673cae | 140 | struct rgw_bucket_placement { |
11fdf7f2 | 141 | rgw_placement_rule placement_rule; |
7c673cae FG |
142 | rgw_bucket bucket; |
143 | ||
144 | void dump(Formatter *f) const; | |
145 | }; | |
146 | ||
147 | class rgw_obj_select { | |
11fdf7f2 | 148 | rgw_placement_rule placement_rule; |
7c673cae FG |
149 | rgw_obj obj; |
150 | rgw_raw_obj raw_obj; | |
151 | bool is_raw; | |
152 | ||
153 | public: | |
154 | rgw_obj_select() : is_raw(false) {} | |
11fdf7f2 TL |
155 | explicit rgw_obj_select(const rgw_obj& _obj) : obj(_obj), is_raw(false) {} |
156 | explicit rgw_obj_select(const rgw_raw_obj& _raw_obj) : raw_obj(_raw_obj), is_raw(true) {} | |
7c673cae | 157 | rgw_obj_select(const rgw_obj_select& rhs) { |
c07f9fc5 | 158 | placement_rule = rhs.placement_rule; |
7c673cae FG |
159 | is_raw = rhs.is_raw; |
160 | if (is_raw) { | |
161 | raw_obj = rhs.raw_obj; | |
162 | } else { | |
163 | obj = rhs.obj; | |
164 | } | |
165 | } | |
166 | ||
167 | rgw_raw_obj get_raw_obj(const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params) const; | |
168 | rgw_raw_obj get_raw_obj(RGWRados *store) const; | |
169 | ||
170 | rgw_obj_select& operator=(const rgw_obj& rhs) { | |
171 | obj = rhs; | |
172 | is_raw = false; | |
173 | return *this; | |
174 | } | |
175 | ||
176 | rgw_obj_select& operator=(const rgw_raw_obj& rhs) { | |
177 | raw_obj = rhs; | |
178 | is_raw = true; | |
179 | return *this; | |
180 | } | |
181 | ||
11fdf7f2 | 182 | void set_placement_rule(const rgw_placement_rule& rule) { |
7c673cae FG |
183 | placement_rule = rule; |
184 | } | |
11fdf7f2 | 185 | void dump(Formatter *f) const; |
7c673cae FG |
186 | }; |
187 | ||
188 | struct compression_block { | |
189 | uint64_t old_ofs; | |
190 | uint64_t new_ofs; | |
191 | uint64_t len; | |
192 | ||
193 | void encode(bufferlist& bl) const { | |
194 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
195 | encode(old_ofs, bl); |
196 | encode(new_ofs, bl); | |
197 | encode(len, bl); | |
7c673cae FG |
198 | ENCODE_FINISH(bl); |
199 | } | |
200 | ||
11fdf7f2 | 201 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 202 | DECODE_START(1, bl); |
11fdf7f2 TL |
203 | decode(old_ofs, bl); |
204 | decode(new_ofs, bl); | |
205 | decode(len, bl); | |
7c673cae FG |
206 | DECODE_FINISH(bl); |
207 | } | |
11fdf7f2 | 208 | void dump(Formatter *f) const; |
7c673cae FG |
209 | }; |
210 | WRITE_CLASS_ENCODER(compression_block) | |
211 | ||
212 | struct RGWCompressionInfo { | |
213 | string compression_type; | |
214 | uint64_t orig_size; | |
215 | vector<compression_block> blocks; | |
216 | ||
217 | RGWCompressionInfo() : compression_type("none"), orig_size(0) {} | |
218 | RGWCompressionInfo(const RGWCompressionInfo& cs_info) : compression_type(cs_info.compression_type), | |
219 | orig_size(cs_info.orig_size), | |
220 | blocks(cs_info.blocks) {} | |
221 | ||
222 | void encode(bufferlist& bl) const { | |
223 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
224 | encode(compression_type, bl); |
225 | encode(orig_size, bl); | |
226 | encode(blocks, bl); | |
7c673cae FG |
227 | ENCODE_FINISH(bl); |
228 | } | |
229 | ||
11fdf7f2 | 230 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 231 | DECODE_START(1, bl); |
11fdf7f2 TL |
232 | decode(compression_type, bl); |
233 | decode(orig_size, bl); | |
234 | decode(blocks, bl); | |
7c673cae | 235 | DECODE_FINISH(bl); |
11fdf7f2 TL |
236 | } |
237 | void dump(Formatter *f) const; | |
7c673cae FG |
238 | }; |
239 | WRITE_CLASS_ENCODER(RGWCompressionInfo) | |
240 | ||
241 | int rgw_compression_info_from_attrset(map<string, bufferlist>& attrs, bool& need_decompress, RGWCompressionInfo& cs_info); | |
242 | ||
243 | struct RGWOLHInfo { | |
244 | rgw_obj target; | |
245 | bool removed; | |
246 | ||
247 | RGWOLHInfo() : removed(false) {} | |
248 | ||
249 | void encode(bufferlist& bl) const { | |
250 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
251 | encode(target, bl); |
252 | encode(removed, bl); | |
7c673cae FG |
253 | ENCODE_FINISH(bl); |
254 | } | |
255 | ||
11fdf7f2 | 256 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 257 | DECODE_START(1, bl); |
11fdf7f2 TL |
258 | decode(target, bl); |
259 | decode(removed, bl); | |
7c673cae FG |
260 | DECODE_FINISH(bl); |
261 | } | |
262 | static void generate_test_instances(list<RGWOLHInfo*>& o); | |
263 | void dump(Formatter *f) const; | |
264 | }; | |
265 | WRITE_CLASS_ENCODER(RGWOLHInfo) | |
266 | ||
267 | struct RGWOLHPendingInfo { | |
268 | ceph::real_time time; | |
269 | ||
270 | RGWOLHPendingInfo() {} | |
271 | ||
272 | void encode(bufferlist& bl) const { | |
273 | ENCODE_START(1, 1, bl); | |
11fdf7f2 | 274 | encode(time, bl); |
7c673cae FG |
275 | ENCODE_FINISH(bl); |
276 | } | |
277 | ||
11fdf7f2 | 278 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 279 | DECODE_START(1, bl); |
11fdf7f2 | 280 | decode(time, bl); |
7c673cae FG |
281 | DECODE_FINISH(bl); |
282 | } | |
283 | ||
284 | void dump(Formatter *f) const; | |
285 | }; | |
286 | WRITE_CLASS_ENCODER(RGWOLHPendingInfo) | |
287 | ||
288 | struct RGWUsageBatch { | |
289 | map<ceph::real_time, rgw_usage_log_entry> m; | |
290 | ||
291 | void insert(ceph::real_time& t, rgw_usage_log_entry& entry, bool *account) { | |
292 | bool exists = m.find(t) != m.end(); | |
293 | *account = !exists; | |
294 | m[t].aggregate(entry); | |
295 | } | |
296 | }; | |
297 | ||
298 | struct RGWUsageIter { | |
299 | string read_iter; | |
300 | uint32_t index; | |
301 | ||
302 | RGWUsageIter() : index(0) {} | |
303 | }; | |
304 | ||
305 | class RGWGetDataCB { | |
7c673cae FG |
306 | public: |
307 | virtual int handle_data(bufferlist& bl, off_t bl_ofs, off_t bl_len) = 0; | |
11fdf7f2 | 308 | RGWGetDataCB() {} |
7c673cae | 309 | virtual ~RGWGetDataCB() {} |
7c673cae FG |
310 | }; |
311 | ||
312 | struct RGWCloneRangeInfo { | |
313 | rgw_obj src; | |
314 | off_t src_ofs; | |
315 | off_t dst_ofs; | |
316 | uint64_t len; | |
317 | }; | |
318 | ||
319 | struct RGWObjManifestPart { | |
320 | rgw_obj loc; /* the object where the data is located */ | |
321 | uint64_t loc_ofs; /* the offset at that object where the data is located */ | |
322 | uint64_t size; /* the part size */ | |
323 | ||
324 | RGWObjManifestPart() : loc_ofs(0), size(0) {} | |
325 | ||
326 | void encode(bufferlist& bl) const { | |
327 | ENCODE_START(2, 2, bl); | |
11fdf7f2 TL |
328 | encode(loc, bl); |
329 | encode(loc_ofs, bl); | |
330 | encode(size, bl); | |
7c673cae FG |
331 | ENCODE_FINISH(bl); |
332 | } | |
333 | ||
11fdf7f2 | 334 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 335 | DECODE_START_LEGACY_COMPAT_LEN_32(2, 2, 2, bl); |
11fdf7f2 TL |
336 | decode(loc, bl); |
337 | decode(loc_ofs, bl); | |
338 | decode(size, bl); | |
7c673cae FG |
339 | DECODE_FINISH(bl); |
340 | } | |
341 | ||
342 | void dump(Formatter *f) const; | |
343 | static void generate_test_instances(list<RGWObjManifestPart*>& o); | |
344 | }; | |
345 | WRITE_CLASS_ENCODER(RGWObjManifestPart) | |
346 | ||
347 | /* | |
348 | The manifest defines a set of rules for structuring the object parts. | |
349 | There are a few terms to note: | |
350 | - head: the head part of the object, which is the part that contains | |
351 | the first chunk of data. An object might not have a head (as in the | |
352 | case of multipart-part objects). | |
353 | - stripe: data portion of a single rgw object that resides on a single | |
354 | rados object. | |
355 | - part: a collection of stripes that make a contiguous part of an | |
356 | object. A regular object will only have one part (although might have | |
357 | many stripes), a multipart object might have many parts. Each part | |
358 | has a fixed stripe size, although the last stripe of a part might | |
359 | be smaller than that. Consecutive parts may be merged if their stripe | |
360 | value is the same. | |
361 | */ | |
362 | ||
363 | struct RGWObjManifestRule { | |
364 | uint32_t start_part_num; | |
365 | uint64_t start_ofs; | |
366 | uint64_t part_size; /* each part size, 0 if there's no part size, meaning it's unlimited */ | |
367 | uint64_t stripe_max_size; /* underlying obj max size */ | |
368 | string override_prefix; | |
369 | ||
370 | RGWObjManifestRule() : start_part_num(0), start_ofs(0), part_size(0), stripe_max_size(0) {} | |
371 | RGWObjManifestRule(uint32_t _start_part_num, uint64_t _start_ofs, uint64_t _part_size, uint64_t _stripe_max_size) : | |
372 | start_part_num(_start_part_num), start_ofs(_start_ofs), part_size(_part_size), stripe_max_size(_stripe_max_size) {} | |
373 | ||
374 | void encode(bufferlist& bl) const { | |
375 | ENCODE_START(2, 1, bl); | |
11fdf7f2 TL |
376 | encode(start_part_num, bl); |
377 | encode(start_ofs, bl); | |
378 | encode(part_size, bl); | |
379 | encode(stripe_max_size, bl); | |
380 | encode(override_prefix, bl); | |
7c673cae FG |
381 | ENCODE_FINISH(bl); |
382 | } | |
383 | ||
11fdf7f2 | 384 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 385 | DECODE_START(2, bl); |
11fdf7f2 TL |
386 | decode(start_part_num, bl); |
387 | decode(start_ofs, bl); | |
388 | decode(part_size, bl); | |
389 | decode(stripe_max_size, bl); | |
7c673cae | 390 | if (struct_v >= 2) |
11fdf7f2 | 391 | decode(override_prefix, bl); |
7c673cae FG |
392 | DECODE_FINISH(bl); |
393 | } | |
394 | void dump(Formatter *f) const; | |
395 | }; | |
396 | WRITE_CLASS_ENCODER(RGWObjManifestRule) | |
397 | ||
398 | class RGWObjManifest { | |
399 | protected: | |
400 | bool explicit_objs; /* old manifest? */ | |
401 | map<uint64_t, RGWObjManifestPart> objs; | |
402 | ||
403 | uint64_t obj_size; | |
404 | ||
405 | rgw_obj obj; | |
406 | uint64_t head_size; | |
11fdf7f2 | 407 | rgw_placement_rule head_placement_rule; |
7c673cae FG |
408 | |
409 | uint64_t max_head_size; | |
410 | string prefix; | |
411 | rgw_bucket_placement tail_placement; /* might be different than the original bucket, | |
412 | as object might have been copied across pools */ | |
413 | map<uint64_t, RGWObjManifestRule> rules; | |
414 | ||
415 | string tail_instance; /* tail object's instance */ | |
416 | ||
417 | void convert_to_explicit(const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params); | |
418 | int append_explicit(RGWObjManifest& m, const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params); | |
419 | void append_rules(RGWObjManifest& m, map<uint64_t, RGWObjManifestRule>::iterator& iter, string *override_prefix); | |
420 | ||
421 | void update_iterators() { | |
422 | begin_iter.seek(0); | |
423 | end_iter.seek(obj_size); | |
424 | } | |
425 | public: | |
426 | ||
427 | RGWObjManifest() : explicit_objs(false), obj_size(0), head_size(0), max_head_size(0), | |
428 | begin_iter(this), end_iter(this) {} | |
429 | RGWObjManifest(const RGWObjManifest& rhs) { | |
430 | *this = rhs; | |
431 | } | |
432 | RGWObjManifest& operator=(const RGWObjManifest& rhs) { | |
433 | explicit_objs = rhs.explicit_objs; | |
434 | objs = rhs.objs; | |
435 | obj_size = rhs.obj_size; | |
436 | obj = rhs.obj; | |
437 | head_size = rhs.head_size; | |
438 | max_head_size = rhs.max_head_size; | |
439 | prefix = rhs.prefix; | |
440 | tail_placement = rhs.tail_placement; | |
441 | rules = rhs.rules; | |
442 | tail_instance = rhs.tail_instance; | |
443 | ||
444 | begin_iter.set_manifest(this); | |
445 | end_iter.set_manifest(this); | |
446 | ||
447 | begin_iter.seek(rhs.begin_iter.get_ofs()); | |
448 | end_iter.seek(rhs.end_iter.get_ofs()); | |
449 | ||
450 | return *this; | |
451 | } | |
452 | ||
453 | map<uint64_t, RGWObjManifestPart>& get_explicit_objs() { | |
454 | return objs; | |
455 | } | |
456 | ||
457 | ||
458 | void set_explicit(uint64_t _size, map<uint64_t, RGWObjManifestPart>& _objs) { | |
459 | explicit_objs = true; | |
460 | obj_size = _size; | |
461 | objs.swap(_objs); | |
462 | } | |
463 | ||
464 | void get_implicit_location(uint64_t cur_part_id, uint64_t cur_stripe, uint64_t ofs, string *override_prefix, rgw_obj_select *location); | |
465 | ||
466 | void set_trivial_rule(uint64_t tail_ofs, uint64_t stripe_max_size) { | |
467 | RGWObjManifestRule rule(0, tail_ofs, 0, stripe_max_size); | |
468 | rules[0] = rule; | |
469 | max_head_size = tail_ofs; | |
470 | } | |
471 | ||
472 | void set_multipart_part_rule(uint64_t stripe_max_size, uint64_t part_num) { | |
473 | RGWObjManifestRule rule(0, 0, 0, stripe_max_size); | |
474 | rule.start_part_num = part_num; | |
475 | rules[0] = rule; | |
476 | max_head_size = 0; | |
477 | } | |
478 | ||
479 | void encode(bufferlist& bl) const { | |
480 | ENCODE_START(7, 6, bl); | |
11fdf7f2 TL |
481 | encode(obj_size, bl); |
482 | encode(objs, bl); | |
483 | encode(explicit_objs, bl); | |
484 | encode(obj, bl); | |
485 | encode(head_size, bl); | |
486 | encode(max_head_size, bl); | |
487 | encode(prefix, bl); | |
488 | encode(rules, bl); | |
7c673cae | 489 | bool encode_tail_bucket = !(tail_placement.bucket == obj.bucket); |
11fdf7f2 | 490 | encode(encode_tail_bucket, bl); |
7c673cae | 491 | if (encode_tail_bucket) { |
11fdf7f2 | 492 | encode(tail_placement.bucket, bl); |
7c673cae FG |
493 | } |
494 | bool encode_tail_instance = (tail_instance != obj.key.instance); | |
11fdf7f2 | 495 | encode(encode_tail_instance, bl); |
7c673cae | 496 | if (encode_tail_instance) { |
11fdf7f2 | 497 | encode(tail_instance, bl); |
7c673cae | 498 | } |
11fdf7f2 TL |
499 | encode(head_placement_rule, bl); |
500 | encode(tail_placement.placement_rule, bl); | |
7c673cae FG |
501 | ENCODE_FINISH(bl); |
502 | } | |
503 | ||
11fdf7f2 | 504 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 505 | DECODE_START_LEGACY_COMPAT_LEN_32(7, 2, 2, bl); |
11fdf7f2 TL |
506 | decode(obj_size, bl); |
507 | decode(objs, bl); | |
7c673cae | 508 | if (struct_v >= 3) { |
11fdf7f2 TL |
509 | decode(explicit_objs, bl); |
510 | decode(obj, bl); | |
511 | decode(head_size, bl); | |
512 | decode(max_head_size, bl); | |
513 | decode(prefix, bl); | |
514 | decode(rules, bl); | |
7c673cae FG |
515 | } else { |
516 | explicit_objs = true; | |
517 | if (!objs.empty()) { | |
518 | map<uint64_t, RGWObjManifestPart>::iterator iter = objs.begin(); | |
519 | obj = iter->second.loc; | |
520 | head_size = iter->second.size; | |
521 | max_head_size = head_size; | |
522 | } | |
523 | } | |
524 | ||
525 | if (explicit_objs && head_size > 0 && !objs.empty()) { | |
526 | /* patch up manifest due to issue 16435: | |
527 | * the first object in the explicit objs list might not be the one we need to access, use the | |
528 | * head object instead if set. This would happen if we had an old object that was created | |
529 | * when the explicit objs manifest was around, and it got copied. | |
530 | */ | |
531 | rgw_obj& obj_0 = objs[0].loc; | |
532 | if (!obj_0.get_oid().empty() && obj_0.key.ns.empty()) { | |
533 | objs[0].loc = obj; | |
534 | objs[0].size = head_size; | |
535 | } | |
536 | } | |
537 | ||
538 | if (struct_v >= 4) { | |
539 | if (struct_v < 6) { | |
11fdf7f2 | 540 | decode(tail_placement.bucket, bl); |
7c673cae FG |
541 | } else { |
542 | bool need_to_decode; | |
11fdf7f2 | 543 | decode(need_to_decode, bl); |
7c673cae | 544 | if (need_to_decode) { |
11fdf7f2 | 545 | decode(tail_placement.bucket, bl); |
7c673cae FG |
546 | } else { |
547 | tail_placement.bucket = obj.bucket; | |
548 | } | |
549 | } | |
550 | } | |
551 | ||
552 | if (struct_v >= 5) { | |
553 | if (struct_v < 6) { | |
11fdf7f2 | 554 | decode(tail_instance, bl); |
7c673cae FG |
555 | } else { |
556 | bool need_to_decode; | |
11fdf7f2 | 557 | decode(need_to_decode, bl); |
7c673cae | 558 | if (need_to_decode) { |
11fdf7f2 | 559 | decode(tail_instance, bl); |
7c673cae FG |
560 | } else { |
561 | tail_instance = obj.key.instance; | |
562 | } | |
563 | } | |
564 | } else { // old object created before 'tail_instance' field added to manifest | |
565 | tail_instance = obj.key.instance; | |
566 | } | |
567 | ||
568 | if (struct_v >= 7) { | |
11fdf7f2 TL |
569 | decode(head_placement_rule, bl); |
570 | decode(tail_placement.placement_rule, bl); | |
7c673cae FG |
571 | } |
572 | ||
573 | update_iterators(); | |
574 | DECODE_FINISH(bl); | |
575 | } | |
576 | ||
577 | void dump(Formatter *f) const; | |
578 | static void generate_test_instances(list<RGWObjManifest*>& o); | |
579 | ||
11fdf7f2 TL |
580 | int append(RGWObjManifest& m, const RGWZoneGroup& zonegroup, |
581 | const RGWZoneParams& zone_params); | |
582 | int append(RGWObjManifest& m, RGWSI_Zone *zone_svc); | |
7c673cae FG |
583 | |
584 | bool get_rule(uint64_t ofs, RGWObjManifestRule *rule); | |
585 | ||
586 | bool empty() { | |
587 | if (explicit_objs) | |
588 | return objs.empty(); | |
589 | return rules.empty(); | |
590 | } | |
591 | ||
592 | bool has_explicit_objs() { | |
593 | return explicit_objs; | |
594 | } | |
595 | ||
596 | bool has_tail() { | |
597 | if (explicit_objs) { | |
598 | if (objs.size() == 1) { | |
599 | map<uint64_t, RGWObjManifestPart>::iterator iter = objs.begin(); | |
600 | rgw_obj& o = iter->second.loc; | |
601 | return !(obj == o); | |
602 | } | |
603 | return (objs.size() >= 2); | |
604 | } | |
605 | return (obj_size > head_size); | |
606 | } | |
607 | ||
11fdf7f2 | 608 | void set_head(const rgw_placement_rule& placement_rule, const rgw_obj& _o, uint64_t _s) { |
7c673cae FG |
609 | head_placement_rule = placement_rule; |
610 | obj = _o; | |
611 | head_size = _s; | |
612 | ||
613 | if (explicit_objs && head_size > 0) { | |
614 | objs[0].loc = obj; | |
615 | objs[0].size = head_size; | |
616 | } | |
617 | } | |
618 | ||
619 | const rgw_obj& get_obj() { | |
620 | return obj; | |
621 | } | |
622 | ||
11fdf7f2 | 623 | void set_tail_placement(const rgw_placement_rule& placement_rule, const rgw_bucket& _b) { |
7c673cae FG |
624 | tail_placement.placement_rule = placement_rule; |
625 | tail_placement.bucket = _b; | |
626 | } | |
627 | ||
628 | const rgw_bucket_placement& get_tail_placement() { | |
629 | return tail_placement; | |
630 | } | |
631 | ||
11fdf7f2 | 632 | const rgw_placement_rule& get_head_placement_rule() { |
7c673cae FG |
633 | return head_placement_rule; |
634 | } | |
635 | ||
636 | void set_prefix(const string& _p) { | |
637 | prefix = _p; | |
638 | } | |
639 | ||
640 | const string& get_prefix() { | |
641 | return prefix; | |
642 | } | |
643 | ||
644 | void set_tail_instance(const string& _ti) { | |
645 | tail_instance = _ti; | |
646 | } | |
647 | ||
648 | const string& get_tail_instance() { | |
649 | return tail_instance; | |
650 | } | |
651 | ||
652 | void set_head_size(uint64_t _s) { | |
653 | head_size = _s; | |
654 | } | |
655 | ||
656 | void set_obj_size(uint64_t s) { | |
657 | obj_size = s; | |
658 | ||
659 | update_iterators(); | |
660 | } | |
661 | ||
662 | uint64_t get_obj_size() { | |
663 | return obj_size; | |
664 | } | |
665 | ||
666 | uint64_t get_head_size() { | |
667 | return head_size; | |
668 | } | |
669 | ||
7c673cae FG |
670 | uint64_t get_max_head_size() { |
671 | return max_head_size; | |
672 | } | |
673 | ||
674 | class obj_iterator { | |
675 | RGWObjManifest *manifest; | |
676 | uint64_t part_ofs; /* where current part starts */ | |
677 | uint64_t stripe_ofs; /* where current stripe starts */ | |
678 | uint64_t ofs; /* current position within the object */ | |
679 | uint64_t stripe_size; /* current part size */ | |
680 | ||
681 | int cur_part_id; | |
682 | int cur_stripe; | |
683 | string cur_override_prefix; | |
684 | ||
685 | rgw_obj_select location; | |
686 | ||
687 | map<uint64_t, RGWObjManifestRule>::iterator rule_iter; | |
688 | map<uint64_t, RGWObjManifestRule>::iterator next_rule_iter; | |
689 | ||
690 | map<uint64_t, RGWObjManifestPart>::iterator explicit_iter; | |
691 | ||
692 | void init() { | |
693 | part_ofs = 0; | |
694 | stripe_ofs = 0; | |
695 | ofs = 0; | |
696 | stripe_size = 0; | |
697 | cur_part_id = 0; | |
698 | cur_stripe = 0; | |
699 | } | |
700 | ||
701 | void update_explicit_pos(); | |
702 | ||
703 | ||
704 | protected: | |
705 | ||
706 | void set_manifest(RGWObjManifest *m) { | |
707 | manifest = m; | |
708 | } | |
709 | ||
710 | public: | |
711 | obj_iterator() : manifest(NULL) { | |
712 | init(); | |
713 | } | |
714 | explicit obj_iterator(RGWObjManifest *_m) : manifest(_m) { | |
715 | init(); | |
716 | if (!manifest->empty()) { | |
717 | seek(0); | |
718 | } | |
719 | } | |
720 | obj_iterator(RGWObjManifest *_m, uint64_t _ofs) : manifest(_m) { | |
721 | init(); | |
722 | if (!manifest->empty()) { | |
723 | seek(_ofs); | |
724 | } | |
725 | } | |
726 | void seek(uint64_t ofs); | |
727 | ||
728 | void operator++(); | |
729 | bool operator==(const obj_iterator& rhs) { | |
730 | return (ofs == rhs.ofs); | |
731 | } | |
732 | bool operator!=(const obj_iterator& rhs) { | |
733 | return (ofs != rhs.ofs); | |
734 | } | |
735 | const rgw_obj_select& get_location() { | |
736 | return location; | |
737 | } | |
738 | ||
739 | /* start of current stripe */ | |
740 | uint64_t get_stripe_ofs() { | |
741 | if (manifest->explicit_objs) { | |
742 | return explicit_iter->first; | |
743 | } | |
744 | return stripe_ofs; | |
745 | } | |
746 | ||
747 | /* current ofs relative to start of rgw object */ | |
748 | uint64_t get_ofs() const { | |
749 | return ofs; | |
750 | } | |
751 | ||
752 | /* stripe number */ | |
753 | int get_cur_stripe() const { | |
754 | return cur_stripe; | |
755 | } | |
756 | ||
757 | /* current stripe size */ | |
758 | uint64_t get_stripe_size() { | |
759 | if (manifest->explicit_objs) { | |
760 | return explicit_iter->second.size; | |
761 | } | |
762 | return stripe_size; | |
763 | } | |
764 | ||
765 | /* offset where data starts within current stripe */ | |
766 | uint64_t location_ofs() { | |
767 | if (manifest->explicit_objs) { | |
768 | return explicit_iter->second.loc_ofs; | |
769 | } | |
770 | return 0; /* all stripes start at zero offset */ | |
771 | } | |
772 | ||
773 | void update_location(); | |
774 | ||
775 | friend class RGWObjManifest; | |
11fdf7f2 | 776 | void dump(Formatter *f) const; |
7c673cae FG |
777 | }; |
778 | ||
779 | const obj_iterator& obj_begin(); | |
780 | const obj_iterator& obj_end(); | |
781 | obj_iterator obj_find(uint64_t ofs); | |
782 | ||
783 | obj_iterator begin_iter; | |
784 | obj_iterator end_iter; | |
785 | ||
786 | /* | |
787 | * simple object generator. Using a simple single rule manifest. | |
788 | */ | |
789 | class generator { | |
790 | RGWObjManifest *manifest; | |
791 | uint64_t last_ofs; | |
792 | uint64_t cur_part_ofs; | |
793 | int cur_part_id; | |
794 | int cur_stripe; | |
795 | uint64_t cur_stripe_size; | |
796 | string cur_oid; | |
797 | ||
798 | string oid_prefix; | |
799 | ||
800 | rgw_obj_select cur_obj; | |
7c673cae FG |
801 | |
802 | RGWObjManifestRule rule; | |
803 | ||
804 | public: | |
805 | generator() : manifest(NULL), last_ofs(0), cur_part_ofs(0), cur_part_id(0), | |
806 | cur_stripe(0), cur_stripe_size(0) {} | |
11fdf7f2 TL |
807 | int create_begin(CephContext *cct, RGWObjManifest *manifest, |
808 | const rgw_placement_rule& head_placement_rule, | |
809 | const rgw_placement_rule *tail_placement_rule, | |
810 | const rgw_bucket& bucket, | |
811 | const rgw_obj& obj); | |
7c673cae FG |
812 | |
813 | int create_next(uint64_t ofs); | |
814 | ||
815 | rgw_raw_obj get_cur_obj(RGWZoneGroup& zonegroup, RGWZoneParams& zone_params) { return cur_obj.get_raw_obj(zonegroup, zone_params); } | |
11fdf7f2 | 816 | rgw_raw_obj get_cur_obj(RGWRados *store) const { return cur_obj.get_raw_obj(store); } |
7c673cae FG |
817 | |
818 | /* total max size of current stripe (including head obj) */ | |
11fdf7f2 | 819 | uint64_t cur_stripe_max_size() const { |
7c673cae FG |
820 | return cur_stripe_size; |
821 | } | |
822 | }; | |
823 | }; | |
824 | WRITE_CLASS_ENCODER(RGWObjManifest) | |
825 | ||
826 | struct RGWUploadPartInfo { | |
827 | uint32_t num; | |
828 | uint64_t size; | |
829 | uint64_t accounted_size{0}; | |
830 | string etag; | |
831 | ceph::real_time modified; | |
832 | RGWObjManifest manifest; | |
833 | RGWCompressionInfo cs_info; | |
834 | ||
835 | RGWUploadPartInfo() : num(0), size(0) {} | |
836 | ||
837 | void encode(bufferlist& bl) const { | |
838 | ENCODE_START(4, 2, bl); | |
11fdf7f2 TL |
839 | encode(num, bl); |
840 | encode(size, bl); | |
841 | encode(etag, bl); | |
842 | encode(modified, bl); | |
843 | encode(manifest, bl); | |
844 | encode(cs_info, bl); | |
845 | encode(accounted_size, bl); | |
7c673cae FG |
846 | ENCODE_FINISH(bl); |
847 | } | |
11fdf7f2 | 848 | void decode(bufferlist::const_iterator& bl) { |
7c673cae | 849 | DECODE_START_LEGACY_COMPAT_LEN(4, 2, 2, bl); |
11fdf7f2 TL |
850 | decode(num, bl); |
851 | decode(size, bl); | |
852 | decode(etag, bl); | |
853 | decode(modified, bl); | |
7c673cae | 854 | if (struct_v >= 3) |
11fdf7f2 | 855 | decode(manifest, bl); |
7c673cae | 856 | if (struct_v >= 4) { |
11fdf7f2 TL |
857 | decode(cs_info, bl); |
858 | decode(accounted_size, bl); | |
7c673cae FG |
859 | } else { |
860 | accounted_size = size; | |
861 | } | |
862 | DECODE_FINISH(bl); | |
863 | } | |
864 | void dump(Formatter *f) const; | |
865 | static void generate_test_instances(list<RGWUploadPartInfo*>& o); | |
866 | }; | |
867 | WRITE_CLASS_ENCODER(RGWUploadPartInfo) | |
868 | ||
869 | struct RGWObjState { | |
870 | rgw_obj obj; | |
871 | bool is_atomic; | |
872 | bool has_attrs; | |
873 | bool exists; | |
874 | uint64_t size; //< size of raw object | |
875 | uint64_t accounted_size{0}; //< size before compression, encryption | |
876 | ceph::real_time mtime; | |
877 | uint64_t epoch; | |
878 | bufferlist obj_tag; | |
181888fb | 879 | bufferlist tail_tag; |
7c673cae FG |
880 | string write_tag; |
881 | bool fake_tag; | |
882 | RGWObjManifest manifest; | |
883 | bool has_manifest; | |
884 | string shadow_obj; | |
885 | bool has_data; | |
886 | bufferlist data; | |
887 | bool prefetch_data; | |
888 | bool keep_tail; | |
889 | bool is_olh; | |
890 | bufferlist olh_tag; | |
891 | uint64_t pg_ver; | |
892 | uint32_t zone_short_id; | |
893 | ||
894 | /* important! don't forget to update copy constructor */ | |
895 | ||
896 | RGWObjVersionTracker objv_tracker; | |
897 | ||
898 | map<string, bufferlist> attrset; | |
899 | RGWObjState() : is_atomic(false), has_attrs(0), exists(false), | |
900 | size(0), epoch(0), fake_tag(false), has_manifest(false), | |
901 | has_data(false), prefetch_data(false), keep_tail(false), is_olh(false), | |
902 | pg_ver(0), zone_short_id(0) {} | |
903 | RGWObjState(const RGWObjState& rhs) : obj (rhs.obj) { | |
904 | is_atomic = rhs.is_atomic; | |
905 | has_attrs = rhs.has_attrs; | |
906 | exists = rhs.exists; | |
907 | size = rhs.size; | |
908 | accounted_size = rhs.accounted_size; | |
909 | mtime = rhs.mtime; | |
910 | epoch = rhs.epoch; | |
911 | if (rhs.obj_tag.length()) { | |
912 | obj_tag = rhs.obj_tag; | |
913 | } | |
181888fb FG |
914 | if (rhs.tail_tag.length()) { |
915 | tail_tag = rhs.tail_tag; | |
916 | } | |
7c673cae FG |
917 | write_tag = rhs.write_tag; |
918 | fake_tag = rhs.fake_tag; | |
919 | if (rhs.has_manifest) { | |
920 | manifest = rhs.manifest; | |
921 | } | |
922 | has_manifest = rhs.has_manifest; | |
923 | shadow_obj = rhs.shadow_obj; | |
924 | has_data = rhs.has_data; | |
925 | if (rhs.data.length()) { | |
926 | data = rhs.data; | |
927 | } | |
928 | prefetch_data = rhs.prefetch_data; | |
929 | keep_tail = rhs.keep_tail; | |
930 | is_olh = rhs.is_olh; | |
931 | objv_tracker = rhs.objv_tracker; | |
932 | pg_ver = rhs.pg_ver; | |
933 | } | |
934 | ||
935 | bool get_attr(string name, bufferlist& dest) { | |
936 | map<string, bufferlist>::iterator iter = attrset.find(name); | |
937 | if (iter != attrset.end()) { | |
938 | dest = iter->second; | |
939 | return true; | |
940 | } | |
941 | return false; | |
942 | } | |
943 | }; | |
944 | ||
945 | struct RGWRawObjState { | |
946 | rgw_raw_obj obj; | |
947 | bool has_attrs{false}; | |
948 | bool exists{false}; | |
949 | uint64_t size{0}; | |
950 | ceph::real_time mtime; | |
11fdf7f2 | 951 | uint64_t epoch{0}; |
7c673cae FG |
952 | bufferlist obj_tag; |
953 | bool has_data{false}; | |
954 | bufferlist data; | |
955 | bool prefetch_data{false}; | |
956 | uint64_t pg_ver{0}; | |
957 | ||
958 | /* important! don't forget to update copy constructor */ | |
959 | ||
960 | RGWObjVersionTracker objv_tracker; | |
961 | ||
962 | map<string, bufferlist> attrset; | |
963 | RGWRawObjState() {} | |
964 | RGWRawObjState(const RGWRawObjState& rhs) : obj (rhs.obj) { | |
965 | has_attrs = rhs.has_attrs; | |
966 | exists = rhs.exists; | |
967 | size = rhs.size; | |
968 | mtime = rhs.mtime; | |
969 | epoch = rhs.epoch; | |
970 | if (rhs.obj_tag.length()) { | |
971 | obj_tag = rhs.obj_tag; | |
972 | } | |
973 | has_data = rhs.has_data; | |
974 | if (rhs.data.length()) { | |
975 | data = rhs.data; | |
976 | } | |
977 | prefetch_data = rhs.prefetch_data; | |
978 | pg_ver = rhs.pg_ver; | |
979 | objv_tracker = rhs.objv_tracker; | |
980 | } | |
981 | }; | |
982 | ||
983 | struct RGWPoolIterCtx { | |
984 | librados::IoCtx io_ctx; | |
985 | librados::NObjectIterator iter; | |
986 | }; | |
987 | ||
988 | struct RGWListRawObjsCtx { | |
989 | bool initialized; | |
990 | RGWPoolIterCtx iter_ctx; | |
991 | ||
992 | RGWListRawObjsCtx() : initialized(false) {} | |
993 | }; | |
994 | ||
7c673cae FG |
995 | struct objexp_hint_entry { |
996 | string tenant; | |
997 | string bucket_name; | |
998 | string bucket_id; | |
999 | rgw_obj_key obj_key; | |
1000 | ceph::real_time exp_time; | |
1001 | ||
1002 | void encode(bufferlist& bl) const { | |
1003 | ENCODE_START(2, 1, bl); | |
11fdf7f2 TL |
1004 | encode(bucket_name, bl); |
1005 | encode(bucket_id, bl); | |
1006 | encode(obj_key, bl); | |
1007 | encode(exp_time, bl); | |
1008 | encode(tenant, bl); | |
7c673cae FG |
1009 | ENCODE_FINISH(bl); |
1010 | } | |
1011 | ||
11fdf7f2 | 1012 | void decode(bufferlist::const_iterator& bl) { |
7c673cae FG |
1013 | // XXX Do we want DECODE_START_LEGACY_COMPAT_LEN(2, 1, 1, bl); ? |
1014 | DECODE_START(2, bl); | |
11fdf7f2 TL |
1015 | decode(bucket_name, bl); |
1016 | decode(bucket_id, bl); | |
1017 | decode(obj_key, bl); | |
1018 | decode(exp_time, bl); | |
7c673cae | 1019 | if (struct_v >= 2) { |
11fdf7f2 | 1020 | decode(tenant, bl); |
7c673cae FG |
1021 | } else { |
1022 | tenant.clear(); | |
1023 | } | |
1024 | DECODE_FINISH(bl); | |
1025 | } | |
1026 | }; | |
1027 | WRITE_CLASS_ENCODER(objexp_hint_entry) | |
1028 | ||
7c673cae FG |
1029 | class RGWDataChangesLog; |
1030 | class RGWMetaSyncStatusManager; | |
1031 | class RGWDataSyncStatusManager; | |
7c673cae | 1032 | class RGWCoroutinesManagerRegistry; |
7c673cae FG |
1033 | |
1034 | class RGWGetBucketStats_CB : public RefCountedObject { | |
1035 | protected: | |
1036 | rgw_bucket bucket; | |
1037 | map<RGWObjCategory, RGWStorageStats> *stats; | |
1038 | public: | |
224ce89b | 1039 | explicit RGWGetBucketStats_CB(const rgw_bucket& _bucket) : bucket(_bucket), stats(NULL) {} |
7c673cae FG |
1040 | ~RGWGetBucketStats_CB() override {} |
1041 | virtual void handle_response(int r) = 0; | |
1042 | virtual void set_response(map<RGWObjCategory, RGWStorageStats> *_stats) { | |
1043 | stats = _stats; | |
1044 | } | |
1045 | }; | |
1046 | ||
1047 | class RGWGetUserStats_CB : public RefCountedObject { | |
1048 | protected: | |
1049 | rgw_user user; | |
1050 | RGWStorageStats stats; | |
1051 | public: | |
1052 | explicit RGWGetUserStats_CB(const rgw_user& _user) : user(_user) {} | |
1053 | ~RGWGetUserStats_CB() override {} | |
1054 | virtual void handle_response(int r) = 0; | |
1055 | virtual void set_response(RGWStorageStats& _stats) { | |
1056 | stats = _stats; | |
1057 | } | |
1058 | }; | |
1059 | ||
1060 | class RGWGetDirHeader_CB; | |
1061 | class RGWGetUserHeader_CB; | |
1062 | ||
11fdf7f2 TL |
1063 | class RGWObjectCtx { |
1064 | RGWRados *store; | |
1065 | RWLock lock{"RGWObjectCtx"}; | |
1066 | void *s{nullptr}; | |
7c673cae | 1067 | |
11fdf7f2 | 1068 | std::map<rgw_obj, RGWObjState> objs_state; |
7c673cae | 1069 | public: |
11fdf7f2 TL |
1070 | explicit RGWObjectCtx(RGWRados *_store) : store(_store) {} |
1071 | explicit RGWObjectCtx(RGWRados *_store, void *_s) : store(_store), s(_s) {} | |
7c673cae | 1072 | |
11fdf7f2 TL |
1073 | void *get_private() { |
1074 | return s; | |
1075 | } | |
7c673cae | 1076 | |
11fdf7f2 TL |
1077 | RGWRados *get_store() { |
1078 | return store; | |
1079 | } | |
7c673cae | 1080 | |
11fdf7f2 TL |
1081 | RGWObjState *get_state(const rgw_obj& obj) { |
1082 | RGWObjState *result; | |
1083 | typename std::map<rgw_obj, RGWObjState>::iterator iter; | |
7c673cae FG |
1084 | lock.get_read(); |
1085 | assert (!obj.empty()); | |
1086 | iter = objs_state.find(obj); | |
1087 | if (iter != objs_state.end()) { | |
1088 | result = &iter->second; | |
1089 | lock.unlock(); | |
1090 | } else { | |
1091 | lock.unlock(); | |
1092 | lock.get_write(); | |
1093 | result = &objs_state[obj]; | |
1094 | lock.unlock(); | |
1095 | } | |
1096 | return result; | |
1097 | } | |
1098 | ||
11fdf7f2 | 1099 | void set_atomic(rgw_obj& obj) { |
7c673cae FG |
1100 | RWLock::WLocker wl(lock); |
1101 | assert (!obj.empty()); | |
1102 | objs_state[obj].is_atomic = true; | |
1103 | } | |
11fdf7f2 | 1104 | void set_prefetch_data(const rgw_obj& obj) { |
7c673cae FG |
1105 | RWLock::WLocker wl(lock); |
1106 | assert (!obj.empty()); | |
1107 | objs_state[obj].prefetch_data = true; | |
1108 | } | |
11fdf7f2 TL |
1109 | |
1110 | void invalidate(const rgw_obj& obj) { | |
7c673cae FG |
1111 | RWLock::WLocker wl(lock); |
1112 | auto iter = objs_state.find(obj); | |
1113 | if (iter == objs_state.end()) { | |
1114 | return; | |
1115 | } | |
1116 | bool is_atomic = iter->second.is_atomic; | |
1117 | bool prefetch_data = iter->second.prefetch_data; | |
1118 | ||
1119 | objs_state.erase(iter); | |
1120 | ||
1121 | if (is_atomic || prefetch_data) { | |
11fdf7f2 TL |
1122 | auto& state = objs_state[obj]; |
1123 | state.is_atomic = is_atomic; | |
1124 | state.prefetch_data = prefetch_data; | |
7c673cae FG |
1125 | } |
1126 | } | |
1127 | }; | |
1128 | ||
7c673cae FG |
1129 | class RGWAsyncRadosProcessor; |
1130 | ||
1131 | template <class T> | |
1132 | class RGWChainedCacheImpl; | |
1133 | ||
1134 | struct bucket_info_entry { | |
1135 | RGWBucketInfo info; | |
1136 | real_time mtime; | |
1137 | map<string, bufferlist> attrs; | |
1138 | }; | |
1139 | ||
1140 | struct tombstone_entry { | |
1141 | ceph::real_time mtime; | |
1142 | uint32_t zone_short_id; | |
1143 | uint64_t pg_ver; | |
1144 | ||
1145 | tombstone_entry() = default; | |
11fdf7f2 | 1146 | explicit tombstone_entry(const RGWObjState& state) |
7c673cae FG |
1147 | : mtime(state.mtime), zone_short_id(state.zone_short_id), |
1148 | pg_ver(state.pg_ver) {} | |
1149 | }; | |
1150 | ||
31f18b77 FG |
1151 | class RGWIndexCompletionManager; |
1152 | ||
3a9019d9 | 1153 | class RGWRados : public AdminSocketHook |
7c673cae FG |
1154 | { |
1155 | friend class RGWGC; | |
1156 | friend class RGWMetaNotifier; | |
1157 | friend class RGWDataNotifier; | |
1158 | friend class RGWLC; | |
1159 | friend class RGWObjectExpirer; | |
1160 | friend class RGWMetaSyncProcessorThread; | |
1161 | friend class RGWDataSyncProcessorThread; | |
31f18b77 FG |
1162 | friend class RGWReshard; |
1163 | friend class RGWBucketReshard; | |
f64942e4 | 1164 | friend class RGWBucketReshardLock; |
31f18b77 | 1165 | friend class BucketIndexLockGuard; |
d2e6a577 | 1166 | friend class RGWCompleteMultipart; |
7c673cae | 1167 | |
11fdf7f2 TL |
1168 | static constexpr const char* admin_commands[4][3] = { |
1169 | { "cache list", | |
1170 | "cache list name=filter,type=CephString,req=false", | |
1171 | "cache list [filter_str]: list object cache, possibly matching substrings" }, | |
1172 | { "cache inspect", | |
1173 | "cache inspect name=target,type=CephString,req=true", | |
1174 | "cache inspect target: print cache element" }, | |
1175 | { "cache erase", | |
1176 | "cache erase name=target,type=CephString,req=true", | |
1177 | "cache erase target: erase element from cache" }, | |
1178 | { "cache zap", | |
1179 | "cache zap", | |
1180 | "cache zap: erase all elements from cache" } | |
1181 | }; | |
3a9019d9 | 1182 | |
7c673cae FG |
1183 | /** Open the pool used as root for this gateway */ |
1184 | int open_root_pool_ctx(); | |
1185 | int open_gc_pool_ctx(); | |
1186 | int open_lc_pool_ctx(); | |
1187 | int open_objexp_pool_ctx(); | |
31f18b77 | 1188 | int open_reshard_pool_ctx(); |
7c673cae | 1189 | |
494da23a TL |
1190 | int open_pool_ctx(const rgw_pool& pool, librados::IoCtx& io_ctx, |
1191 | bool mostly_omap); | |
7c673cae FG |
1192 | int open_bucket_index_ctx(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx); |
1193 | int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, string& bucket_oid); | |
1194 | int open_bucket_index_base(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
1195 | string& bucket_oid_base); | |
1196 | int open_bucket_index_shard(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
1197 | const string& obj_key, string *bucket_obj, int *shard_id); | |
1198 | int open_bucket_index_shard(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
1199 | int shard_id, string *bucket_obj); | |
1200 | int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
1201 | map<int, string>& bucket_objs, int shard_id = -1, map<int, string> *bucket_instance_ids = NULL); | |
1202 | template<typename T> | |
1203 | int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
1204 | map<int, string>& oids, map<int, T>& bucket_objs, | |
1205 | int shard_id = -1, map<int, string> *bucket_instance_ids = NULL); | |
1206 | void build_bucket_index_marker(const string& shard_id_str, const string& shard_marker, | |
1207 | string *marker); | |
1208 | ||
1209 | void get_bucket_instance_ids(const RGWBucketInfo& bucket_info, int shard_id, map<int, string> *result); | |
1210 | ||
1211 | std::atomic<int64_t> max_req_id = { 0 }; | |
1212 | Mutex lock; | |
7c673cae FG |
1213 | SafeTimer *timer; |
1214 | ||
1215 | RGWGC *gc; | |
1216 | RGWLC *lc; | |
1217 | RGWObjectExpirer *obj_expirer; | |
1218 | bool use_gc_thread; | |
1219 | bool use_lc_thread; | |
1220 | bool quota_threads; | |
1221 | bool run_sync_thread; | |
31f18b77 | 1222 | bool run_reshard_thread; |
7c673cae FG |
1223 | |
1224 | RGWAsyncRadosProcessor* async_rados; | |
1225 | ||
1226 | RGWMetaNotifier *meta_notifier; | |
1227 | RGWDataNotifier *data_notifier; | |
1228 | RGWMetaSyncProcessorThread *meta_sync_processor_thread; | |
11fdf7f2 | 1229 | RGWSyncTraceManager *sync_tracer = nullptr; |
7c673cae FG |
1230 | map<string, RGWDataSyncProcessorThread *> data_sync_processor_threads; |
1231 | ||
b32b8144 | 1232 | boost::optional<rgw::BucketTrimManager> bucket_trim; |
7c673cae FG |
1233 | RGWSyncLogTrimThread *sync_log_trimmer{nullptr}; |
1234 | ||
1235 | Mutex meta_sync_thread_lock; | |
1236 | Mutex data_sync_thread_lock; | |
1237 | ||
7c673cae | 1238 | librados::IoCtx root_pool_ctx; // .rgw |
11fdf7f2 TL |
1239 | |
1240 | double inject_notify_timeout_probability = 0; | |
1241 | unsigned max_notify_retries = 0; | |
7c673cae FG |
1242 | |
1243 | friend class RGWWatcher; | |
1244 | ||
1245 | Mutex bucket_id_lock; | |
1246 | ||
1247 | // This field represents the number of bucket index object shards | |
1248 | uint32_t bucket_index_max_shards; | |
1249 | ||
1250 | int get_obj_head_ioctx(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::IoCtx *ioctx); | |
1251 | int get_obj_head_ref(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_rados_ref *ref); | |
224ce89b | 1252 | int get_system_obj_ref(const rgw_raw_obj& obj, rgw_rados_ref *ref); |
7c673cae FG |
1253 | uint64_t max_bucket_id; |
1254 | ||
1255 | int get_olh_target_state(RGWObjectCtx& rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
1256 | RGWObjState *olh_state, RGWObjState **target_state); | |
7c673cae FG |
1257 | int get_obj_state_impl(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state, |
1258 | bool follow_olh, bool assume_noent = false); | |
1259 | int append_atomic_test(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
1260 | librados::ObjectOperation& op, RGWObjState **state); | |
11fdf7f2 | 1261 | int append_atomic_test(const RGWObjState* astate, librados::ObjectOperation& op); |
7c673cae FG |
1262 | |
1263 | int update_placement_map(); | |
1264 | int store_bucket_info(RGWBucketInfo& info, map<string, bufferlist> *pattrs, RGWObjVersionTracker *objv_tracker, bool exclusive); | |
1265 | ||
1266 | void remove_rgw_head_obj(librados::ObjectWriteOperation& op); | |
1267 | void cls_obj_check_prefix_exist(librados::ObjectOperation& op, const string& prefix, bool fail_if_exist); | |
1268 | void cls_obj_check_mtime(librados::ObjectOperation& op, const real_time& mtime, bool high_precision_time, RGWCheckMTimeType type); | |
1269 | protected: | |
1270 | CephContext *cct; | |
1271 | ||
494da23a | 1272 | librados::Rados rados; |
7c673cae FG |
1273 | |
1274 | using RGWChainedCacheImpl_bucket_info_entry = RGWChainedCacheImpl<bucket_info_entry>; | |
1275 | RGWChainedCacheImpl_bucket_info_entry *binfo_cache; | |
1276 | ||
1277 | using tombstone_cache_t = lru_map<rgw_obj, tombstone_entry>; | |
1278 | tombstone_cache_t *obj_tombstone_cache; | |
1279 | ||
1280 | librados::IoCtx gc_pool_ctx; // .rgw.gc | |
1281 | librados::IoCtx lc_pool_ctx; // .rgw.lc | |
1282 | librados::IoCtx objexp_pool_ctx; | |
31f18b77 | 1283 | librados::IoCtx reshard_pool_ctx; |
7c673cae | 1284 | |
11fdf7f2 | 1285 | bool pools_initialized; |
7c673cae | 1286 | |
11fdf7f2 | 1287 | RGWQuotaHandler *quota_handler; |
7c673cae | 1288 | |
11fdf7f2 | 1289 | RGWCoroutinesManagerRegistry *cr_registry; |
7c673cae | 1290 | |
11fdf7f2 TL |
1291 | RGWSyncModuleInstanceRef sync_module; |
1292 | bool writeable_zone{false}; | |
7c673cae | 1293 | |
11fdf7f2 | 1294 | RGWIndexCompletionManager *index_completion_manager{nullptr}; |
7c673cae | 1295 | |
11fdf7f2 TL |
1296 | bool use_cache{false}; |
1297 | public: | |
1298 | RGWRados(): lock("rados_timer_lock"), timer(NULL), | |
1299 | gc(NULL), lc(NULL), obj_expirer(NULL), use_gc_thread(false), use_lc_thread(false), quota_threads(false), | |
1300 | run_sync_thread(false), run_reshard_thread(false), async_rados(nullptr), meta_notifier(NULL), | |
1301 | data_notifier(NULL), meta_sync_processor_thread(NULL), | |
1302 | meta_sync_thread_lock("meta_sync_thread_lock"), data_sync_thread_lock("data_sync_thread_lock"), | |
1303 | bucket_id_lock("rados_bucket_id"), | |
1304 | bucket_index_max_shards(0), | |
1305 | max_bucket_id(0), cct(NULL), | |
11fdf7f2 TL |
1306 | binfo_cache(NULL), obj_tombstone_cache(nullptr), |
1307 | pools_initialized(false), | |
1308 | quota_handler(NULL), | |
1309 | cr_registry(NULL), | |
1310 | meta_mgr(NULL), data_log(NULL), reshard(NULL) {} | |
7c673cae | 1311 | |
11fdf7f2 TL |
1312 | RGWRados& set_use_cache(bool status) { |
1313 | use_cache = status; | |
1314 | return *this; | |
7c673cae FG |
1315 | } |
1316 | ||
11fdf7f2 TL |
1317 | RGWLC *get_lc() { |
1318 | return lc; | |
7c673cae FG |
1319 | } |
1320 | ||
11fdf7f2 TL |
1321 | RGWRados& set_run_gc_thread(bool _use_gc_thread) { |
1322 | use_gc_thread = _use_gc_thread; | |
1323 | return *this; | |
7c673cae FG |
1324 | } |
1325 | ||
11fdf7f2 TL |
1326 | RGWRados& set_run_lc_thread(bool _use_lc_thread) { |
1327 | use_lc_thread = _use_lc_thread; | |
1328 | return *this; | |
7c673cae FG |
1329 | } |
1330 | ||
11fdf7f2 TL |
1331 | RGWRados& set_run_quota_threads(bool _run_quota_threads) { |
1332 | quota_threads = _run_quota_threads; | |
1333 | return *this; | |
7c673cae FG |
1334 | } |
1335 | ||
11fdf7f2 TL |
1336 | RGWRados& set_run_sync_thread(bool _run_sync_thread) { |
1337 | run_sync_thread = _run_sync_thread; | |
1338 | return *this; | |
7c673cae FG |
1339 | } |
1340 | ||
11fdf7f2 TL |
1341 | RGWRados& set_run_reshard_thread(bool _run_reshard_thread) { |
1342 | run_reshard_thread = _run_reshard_thread; | |
1343 | return *this; | |
7c673cae FG |
1344 | } |
1345 | ||
11fdf7f2 TL |
1346 | uint64_t get_new_req_id() { |
1347 | return ++max_req_id; | |
7c673cae FG |
1348 | } |
1349 | ||
11fdf7f2 TL |
1350 | librados::IoCtx* get_lc_pool_ctx() { |
1351 | return &lc_pool_ctx; | |
7c673cae | 1352 | } |
11fdf7f2 TL |
1353 | void set_context(CephContext *_cct) { |
1354 | cct = _cct; | |
7c673cae | 1355 | } |
31f18b77 | 1356 | |
11fdf7f2 TL |
1357 | RGWServices svc; |
1358 | ||
1359 | /** | |
1360 | * AmazonS3 errors contain a HostId string, but is an opaque base64 blob; we | |
1361 | * try to be more transparent. This has a wrapper so we can update it when zonegroup/zone are changed. | |
1362 | */ | |
1363 | string host_id; | |
31f18b77 | 1364 | |
7c673cae FG |
1365 | // pulls missing periods for period_history |
1366 | std::unique_ptr<RGWPeriodPuller> period_puller; | |
1367 | // maintains a connected history of periods | |
1368 | std::unique_ptr<RGWPeriodHistory> period_history; | |
1369 | ||
1370 | RGWAsyncRadosProcessor* get_async_rados() const { return async_rados; }; | |
1371 | ||
1372 | RGWMetadataManager *meta_mgr; | |
1373 | ||
1374 | RGWDataChangesLog *data_log; | |
1375 | ||
31f18b77 FG |
1376 | RGWReshard *reshard; |
1377 | std::shared_ptr<RGWReshardWait> reshard_wait; | |
1378 | ||
7c673cae FG |
1379 | virtual ~RGWRados() = default; |
1380 | ||
1381 | tombstone_cache_t *get_tombstone_cache() { | |
1382 | return obj_tombstone_cache; | |
1383 | } | |
7c673cae FG |
1384 | const RGWSyncModuleInstanceRef& get_sync_module() { |
1385 | return sync_module; | |
1386 | } | |
11fdf7f2 TL |
1387 | RGWSyncTraceManager *get_sync_tracer() { |
1388 | return sync_tracer; | |
1389 | } | |
7c673cae FG |
1390 | |
1391 | int get_required_alignment(const rgw_pool& pool, uint64_t *alignment); | |
11fdf7f2 TL |
1392 | void get_max_aligned_size(uint64_t size, uint64_t alignment, uint64_t *max_size); |
1393 | int get_max_chunk_size(const rgw_pool& pool, uint64_t *max_chunk_size, uint64_t *palignment = nullptr); | |
1394 | int get_max_chunk_size(const rgw_placement_rule& placement_rule, const rgw_obj& obj, uint64_t *max_chunk_size, uint64_t *palignment = nullptr); | |
7c673cae FG |
1395 | |
1396 | uint32_t get_max_bucket_shards() { | |
31f18b77 | 1397 | return rgw_shards_max(); |
7c673cae FG |
1398 | } |
1399 | ||
181888fb | 1400 | |
224ce89b | 1401 | int get_raw_obj_ref(const rgw_raw_obj& obj, rgw_rados_ref *ref); |
7c673cae | 1402 | |
181888fb FG |
1403 | int list_raw_objects_init(const rgw_pool& pool, const string& marker, RGWListRawObjsCtx *ctx); |
1404 | int list_raw_objects_next(const string& prefix_filter, int max, | |
1405 | RGWListRawObjsCtx& ctx, list<string>& oids, | |
1406 | bool *is_truncated); | |
7c673cae FG |
1407 | int list_raw_objects(const rgw_pool& pool, const string& prefix_filter, int max, |
1408 | RGWListRawObjsCtx& ctx, list<string>& oids, | |
1409 | bool *is_truncated); | |
181888fb | 1410 | string list_raw_objs_get_cursor(RGWListRawObjsCtx& ctx); |
7c673cae | 1411 | |
7c673cae FG |
1412 | CephContext *ctx() { return cct; } |
1413 | /** do all necessary setup of the storage device */ | |
11fdf7f2 | 1414 | int initialize(CephContext *_cct) { |
7c673cae | 1415 | set_context(_cct); |
7c673cae FG |
1416 | return initialize(); |
1417 | } | |
1418 | /** Initialize the RADOS instance and prepare to do other ops */ | |
11fdf7f2 TL |
1419 | int init_svc(bool raw); |
1420 | int init_rados(); | |
7c673cae | 1421 | int init_complete(); |
7c673cae FG |
1422 | int initialize(); |
1423 | void finalize(); | |
1424 | ||
224ce89b | 1425 | int register_to_service_map(const string& daemon_type, const map<string, string>& meta); |
11fdf7f2 | 1426 | int update_service_map(std::map<std::string, std::string>&& status); |
7c673cae FG |
1427 | |
1428 | /// list logs | |
1429 | int log_list_init(const string& prefix, RGWAccessHandle *handle); | |
1430 | int log_list_next(RGWAccessHandle handle, string *name); | |
1431 | ||
1432 | /// remove log | |
1433 | int log_remove(const string& name); | |
1434 | ||
1435 | /// show log | |
1436 | int log_show_init(const string& name, RGWAccessHandle *handle); | |
1437 | int log_show_next(RGWAccessHandle handle, rgw_log_entry *entry); | |
1438 | ||
1439 | // log bandwidth info | |
1440 | int log_usage(map<rgw_user_bucket, RGWUsageBatch>& usage_info); | |
11fdf7f2 TL |
1441 | int read_usage(const rgw_user& user, const string& bucket_name, uint64_t start_epoch, uint64_t end_epoch, |
1442 | uint32_t max_entries, bool *is_truncated, RGWUsageIter& read_iter, map<rgw_user_bucket, | |
1443 | rgw_usage_log_entry>& usage); | |
1444 | int trim_usage(const rgw_user& user, const string& bucket_name, uint64_t start_epoch, uint64_t end_epoch); | |
1445 | int clear_usage(); | |
7c673cae FG |
1446 | |
1447 | int create_pool(const rgw_pool& pool); | |
1448 | ||
7c673cae | 1449 | int init_bucket_index(RGWBucketInfo& bucket_info, int num_shards); |
f64942e4 | 1450 | int clean_bucket_index(RGWBucketInfo& bucket_info, int num_shards); |
7c673cae FG |
1451 | void create_bucket_id(string *bucket_id); |
1452 | ||
11fdf7f2 TL |
1453 | bool get_obj_data_pool(const rgw_placement_rule& placement_rule, const rgw_obj& obj, rgw_pool *pool); |
1454 | bool obj_to_raw(const rgw_placement_rule& placement_rule, const rgw_obj& obj, rgw_raw_obj *raw_obj); | |
7c673cae | 1455 | |
11fdf7f2 | 1456 | int create_bucket(const RGWUserInfo& owner, rgw_bucket& bucket, |
7c673cae | 1457 | const string& zonegroup_id, |
11fdf7f2 | 1458 | const rgw_placement_rule& placement_rule, |
7c673cae FG |
1459 | const string& swift_ver_location, |
1460 | const RGWQuotaInfo * pquota_info, | |
1461 | map<std::string,bufferlist>& attrs, | |
1462 | RGWBucketInfo& bucket_info, | |
1463 | obj_version *pobjv, | |
1464 | obj_version *pep_objv, | |
1465 | ceph::real_time creation_time, | |
1466 | rgw_bucket *master_bucket, | |
1467 | uint32_t *master_num_shards, | |
1468 | bool exclusive = true); | |
7c673cae FG |
1469 | |
1470 | RGWCoroutinesManagerRegistry *get_cr_registry() { return cr_registry; } | |
1471 | ||
7c673cae FG |
1472 | struct BucketShard { |
1473 | RGWRados *store; | |
1474 | rgw_bucket bucket; | |
1475 | int shard_id; | |
1476 | librados::IoCtx index_ctx; | |
1477 | string bucket_obj; | |
1478 | ||
1479 | explicit BucketShard(RGWRados *_store) : store(_store), shard_id(-1) {} | |
f64942e4 AA |
1480 | int init(const rgw_bucket& _bucket, const rgw_obj& obj, RGWBucketInfo* out); |
1481 | int init(const rgw_bucket& _bucket, int sid, RGWBucketInfo* out); | |
a8e16298 | 1482 | int init(const RGWBucketInfo& bucket_info, const rgw_obj& obj); |
b32b8144 | 1483 | int init(const RGWBucketInfo& bucket_info, int sid); |
7c673cae FG |
1484 | }; |
1485 | ||
1486 | class Object { | |
1487 | RGWRados *store; | |
1488 | RGWBucketInfo bucket_info; | |
1489 | RGWObjectCtx& ctx; | |
1490 | rgw_obj obj; | |
1491 | ||
1492 | BucketShard bs; | |
1493 | ||
1494 | RGWObjState *state; | |
1495 | ||
1496 | bool versioning_disabled; | |
1497 | ||
1498 | bool bs_initialized; | |
1499 | ||
1500 | protected: | |
1501 | int get_state(RGWObjState **pstate, bool follow_olh, bool assume_noent = false); | |
1502 | void invalidate_state(); | |
1503 | ||
1504 | int prepare_atomic_modification(librados::ObjectWriteOperation& op, bool reset_obj, const string *ptag, | |
181888fb | 1505 | const char *ifmatch, const char *ifnomatch, bool removal_op, bool modify_tail); |
7c673cae FG |
1506 | int complete_atomic_modification(); |
1507 | ||
1508 | public: | |
1509 | Object(RGWRados *_store, const RGWBucketInfo& _bucket_info, RGWObjectCtx& _ctx, const rgw_obj& _obj) : store(_store), bucket_info(_bucket_info), | |
1510 | ctx(_ctx), obj(_obj), bs(store), | |
1511 | state(NULL), versioning_disabled(false), | |
1512 | bs_initialized(false) {} | |
1513 | ||
1514 | RGWRados *get_store() { return store; } | |
1515 | rgw_obj& get_obj() { return obj; } | |
1516 | RGWObjectCtx& get_ctx() { return ctx; } | |
1517 | RGWBucketInfo& get_bucket_info() { return bucket_info; } | |
1518 | int get_manifest(RGWObjManifest **pmanifest); | |
1519 | ||
1520 | int get_bucket_shard(BucketShard **pbs) { | |
1521 | if (!bs_initialized) { | |
f64942e4 AA |
1522 | int r = |
1523 | bs.init(bucket_info.bucket, obj, nullptr /* no RGWBucketInfo */); | |
7c673cae FG |
1524 | if (r < 0) { |
1525 | return r; | |
1526 | } | |
1527 | bs_initialized = true; | |
1528 | } | |
1529 | *pbs = &bs; | |
1530 | return 0; | |
1531 | } | |
1532 | ||
1533 | void set_versioning_disabled(bool status) { | |
1534 | versioning_disabled = status; | |
1535 | } | |
1536 | ||
1537 | bool versioning_enabled() { | |
1538 | return (!versioning_disabled && bucket_info.versioning_enabled()); | |
1539 | } | |
1540 | ||
1541 | struct Read { | |
1542 | RGWRados::Object *source; | |
1543 | ||
1544 | struct GetObjState { | |
11fdf7f2 TL |
1545 | map<rgw_pool, librados::IoCtx> io_ctxs; |
1546 | rgw_pool cur_pool; | |
1547 | librados::IoCtx *cur_ioctx{nullptr}; | |
7c673cae FG |
1548 | rgw_obj obj; |
1549 | rgw_raw_obj head_obj; | |
1550 | } state; | |
1551 | ||
1552 | struct ConditionParams { | |
1553 | const ceph::real_time *mod_ptr; | |
1554 | const ceph::real_time *unmod_ptr; | |
1555 | bool high_precision_time; | |
1556 | uint32_t mod_zone_id; | |
1557 | uint64_t mod_pg_ver; | |
1558 | const char *if_match; | |
1559 | const char *if_nomatch; | |
1560 | ||
1561 | ConditionParams() : | |
1562 | mod_ptr(NULL), unmod_ptr(NULL), high_precision_time(false), mod_zone_id(0), mod_pg_ver(0), | |
1563 | if_match(NULL), if_nomatch(NULL) {} | |
1564 | } conds; | |
1565 | ||
1566 | struct Params { | |
1567 | ceph::real_time *lastmod; | |
1568 | uint64_t *obj_size; | |
1569 | map<string, bufferlist> *attrs; | |
eafe8130 | 1570 | rgw_obj *target_obj; |
7c673cae | 1571 | |
eafe8130 TL |
1572 | Params() : lastmod(nullptr), obj_size(nullptr), attrs(nullptr), |
1573 | target_obj(nullptr) {} | |
7c673cae FG |
1574 | } params; |
1575 | ||
1576 | explicit Read(RGWRados::Object *_source) : source(_source) {} | |
1577 | ||
1578 | int prepare(); | |
1579 | static int range_to_ofs(uint64_t obj_size, int64_t &ofs, int64_t &end); | |
1580 | int read(int64_t ofs, int64_t end, bufferlist& bl); | |
1581 | int iterate(int64_t ofs, int64_t end, RGWGetDataCB *cb); | |
1582 | int get_attr(const char *name, bufferlist& dest); | |
1583 | }; | |
1584 | ||
1585 | struct Write { | |
1586 | RGWRados::Object *target; | |
1587 | ||
1588 | struct MetaParams { | |
1589 | ceph::real_time *mtime; | |
1590 | map<std::string, bufferlist>* rmattrs; | |
1591 | const bufferlist *data; | |
1592 | RGWObjManifest *manifest; | |
1593 | const string *ptag; | |
1594 | list<rgw_obj_index_key> *remove_objs; | |
1595 | ceph::real_time set_mtime; | |
1596 | rgw_user owner; | |
1597 | RGWObjCategory category; | |
1598 | int flags; | |
1599 | const char *if_match; | |
1600 | const char *if_nomatch; | |
11fdf7f2 | 1601 | std::optional<uint64_t> olh_epoch; |
7c673cae FG |
1602 | ceph::real_time delete_at; |
1603 | bool canceled; | |
1604 | const string *user_data; | |
31f18b77 | 1605 | rgw_zone_set *zones_trace; |
181888fb | 1606 | bool modify_tail; |
3efd9988 | 1607 | bool completeMultipart; |
11fdf7f2 | 1608 | bool appendable; |
7c673cae FG |
1609 | |
1610 | MetaParams() : mtime(NULL), rmattrs(NULL), data(NULL), manifest(NULL), ptag(NULL), | |
11fdf7f2 | 1611 | remove_objs(NULL), category(RGWObjCategory::Main), flags(0), |
91327a77 | 1612 | if_match(NULL), if_nomatch(NULL), canceled(false), user_data(nullptr), zones_trace(nullptr), |
11fdf7f2 | 1613 | modify_tail(false), completeMultipart(false), appendable(false) {} |
7c673cae FG |
1614 | } meta; |
1615 | ||
1616 | explicit Write(RGWRados::Object *_target) : target(_target) {} | |
1617 | ||
1618 | int _do_write_meta(uint64_t size, uint64_t accounted_size, | |
1619 | map<std::string, bufferlist>& attrs, | |
181888fb | 1620 | bool modify_tail, bool assume_noent, |
7c673cae FG |
1621 | void *index_op); |
1622 | int write_meta(uint64_t size, uint64_t accounted_size, | |
1623 | map<std::string, bufferlist>& attrs); | |
1624 | int write_data(const char *data, uint64_t ofs, uint64_t len, bool exclusive); | |
11fdf7f2 TL |
1625 | const req_state* get_req_state() { |
1626 | return (req_state *)target->get_ctx().get_private(); | |
1627 | } | |
7c673cae FG |
1628 | }; |
1629 | ||
1630 | struct Delete { | |
1631 | RGWRados::Object *target; | |
1632 | ||
1633 | struct DeleteParams { | |
1634 | rgw_user bucket_owner; | |
1635 | int versioning_status; | |
1636 | ACLOwner obj_owner; /* needed for creation of deletion marker */ | |
1637 | uint64_t olh_epoch; | |
1638 | string marker_version_id; | |
1639 | uint32_t bilog_flags; | |
1640 | list<rgw_obj_index_key> *remove_objs; | |
1641 | ceph::real_time expiration_time; | |
1642 | ceph::real_time unmod_since; | |
1643 | ceph::real_time mtime; /* for setting delete marker mtime */ | |
1644 | bool high_precision_time; | |
31f18b77 | 1645 | rgw_zone_set *zones_trace; |
7c673cae | 1646 | |
31f18b77 | 1647 | DeleteParams() : versioning_status(0), olh_epoch(0), bilog_flags(0), remove_objs(NULL), high_precision_time(false), zones_trace(nullptr) {} |
7c673cae FG |
1648 | } params; |
1649 | ||
1650 | struct DeleteResult { | |
1651 | bool delete_marker; | |
1652 | string version_id; | |
1653 | ||
1654 | DeleteResult() : delete_marker(false) {} | |
1655 | } result; | |
1656 | ||
1657 | explicit Delete(RGWRados::Object *_target) : target(_target) {} | |
1658 | ||
1659 | int delete_obj(); | |
1660 | }; | |
1661 | ||
1662 | struct Stat { | |
1663 | RGWRados::Object *source; | |
1664 | ||
1665 | struct Result { | |
1666 | rgw_obj obj; | |
1667 | RGWObjManifest manifest; | |
1668 | bool has_manifest; | |
1669 | uint64_t size; | |
1670 | struct timespec mtime; | |
1671 | map<string, bufferlist> attrs; | |
1672 | ||
1673 | Result() : has_manifest(false), size(0) {} | |
1674 | } result; | |
1675 | ||
1676 | struct State { | |
1677 | librados::IoCtx io_ctx; | |
1678 | librados::AioCompletion *completion; | |
1679 | int ret; | |
1680 | ||
1681 | State() : completion(NULL), ret(0) {} | |
1682 | } state; | |
1683 | ||
1684 | ||
1685 | explicit Stat(RGWRados::Object *_source) : source(_source) {} | |
1686 | ||
1687 | int stat_async(); | |
1688 | int wait(); | |
1689 | int stat(); | |
1690 | private: | |
1691 | int finish(); | |
1692 | }; | |
1693 | }; | |
1694 | ||
1695 | class Bucket { | |
1696 | RGWRados *store; | |
1697 | RGWBucketInfo bucket_info; | |
1698 | rgw_bucket& bucket; | |
1699 | int shard_id; | |
1700 | ||
1701 | public: | |
1702 | Bucket(RGWRados *_store, const RGWBucketInfo& _bucket_info) : store(_store), bucket_info(_bucket_info), bucket(bucket_info.bucket), | |
1703 | shard_id(RGW_NO_SHARD) {} | |
1704 | RGWRados *get_store() { return store; } | |
1705 | rgw_bucket& get_bucket() { return bucket; } | |
1706 | RGWBucketInfo& get_bucket_info() { return bucket_info; } | |
1707 | ||
31f18b77 FG |
1708 | int update_bucket_id(const string& new_bucket_id); |
1709 | ||
7c673cae FG |
1710 | int get_shard_id() { return shard_id; } |
1711 | void set_shard_id(int id) { | |
1712 | shard_id = id; | |
1713 | } | |
1714 | ||
1715 | class UpdateIndex { | |
1716 | RGWRados::Bucket *target; | |
1717 | string optag; | |
1718 | rgw_obj obj; | |
1719 | uint16_t bilog_flags{0}; | |
1720 | BucketShard bs; | |
1721 | bool bs_initialized{false}; | |
1722 | bool blind; | |
1723 | bool prepared{false}; | |
31f18b77 FG |
1724 | rgw_zone_set *zones_trace{nullptr}; |
1725 | ||
1726 | int init_bs() { | |
f64942e4 AA |
1727 | int r = |
1728 | bs.init(target->get_bucket(), obj, nullptr /* no RGWBucketInfo */); | |
31f18b77 FG |
1729 | if (r < 0) { |
1730 | return r; | |
1731 | } | |
1732 | bs_initialized = true; | |
1733 | return 0; | |
1734 | } | |
1735 | ||
1736 | void invalidate_bs() { | |
1737 | bs_initialized = false; | |
1738 | } | |
1739 | ||
1740 | int guard_reshard(BucketShard **pbs, std::function<int(BucketShard *)> call); | |
7c673cae FG |
1741 | public: |
1742 | ||
1743 | UpdateIndex(RGWRados::Bucket *_target, const rgw_obj& _obj) : target(_target), obj(_obj), | |
1744 | bs(target->get_store()) { | |
1745 | blind = (target->get_bucket_info().index_type == RGWBIType_Indexless); | |
1746 | } | |
1747 | ||
1748 | int get_bucket_shard(BucketShard **pbs) { | |
1749 | if (!bs_initialized) { | |
31f18b77 | 1750 | int r = init_bs(); |
7c673cae FG |
1751 | if (r < 0) { |
1752 | return r; | |
1753 | } | |
7c673cae FG |
1754 | } |
1755 | *pbs = &bs; | |
1756 | return 0; | |
1757 | } | |
1758 | ||
1759 | void set_bilog_flags(uint16_t flags) { | |
1760 | bilog_flags = flags; | |
1761 | } | |
31f18b77 FG |
1762 | |
1763 | void set_zones_trace(rgw_zone_set *_zones_trace) { | |
1764 | zones_trace = _zones_trace; | |
1765 | } | |
7c673cae FG |
1766 | |
1767 | int prepare(RGWModifyOp, const string *write_tag); | |
1768 | int complete(int64_t poolid, uint64_t epoch, uint64_t size, | |
1769 | uint64_t accounted_size, ceph::real_time& ut, | |
1770 | const string& etag, const string& content_type, | |
11fdf7f2 | 1771 | const string& storage_class, |
7c673cae | 1772 | bufferlist *acl_bl, RGWObjCategory category, |
11fdf7f2 | 1773 | list<rgw_obj_index_key> *remove_objs, const string *user_data = nullptr, bool appendable = false); |
7c673cae FG |
1774 | int complete_del(int64_t poolid, uint64_t epoch, |
1775 | ceph::real_time& removed_mtime, /* mtime of removed object */ | |
1776 | list<rgw_obj_index_key> *remove_objs); | |
1777 | int cancel(); | |
1778 | ||
1779 | const string *get_optag() { return &optag; } | |
1780 | ||
1781 | bool is_prepared() { return prepared; } | |
1adf2230 AA |
1782 | }; // class UpdateIndex |
1783 | ||
1784 | class List { | |
1785 | protected: | |
eafe8130 TL |
1786 | // absolute maximum number of objects that |
1787 | // list_objects_(un)ordered can return | |
1788 | static constexpr int64_t bucket_list_objects_absolute_max = 25000; | |
7c673cae | 1789 | |
7c673cae FG |
1790 | RGWRados::Bucket *target; |
1791 | rgw_obj_key next_marker; | |
1792 | ||
1adf2230 AA |
1793 | int list_objects_ordered(int64_t max, |
1794 | vector<rgw_bucket_dir_entry> *result, | |
1795 | map<string, bool> *common_prefixes, | |
1796 | bool *is_truncated); | |
1797 | int list_objects_unordered(int64_t max, | |
1798 | vector<rgw_bucket_dir_entry> *result, | |
1799 | map<string, bool> *common_prefixes, | |
1800 | bool *is_truncated); | |
1801 | ||
1802 | public: | |
1803 | ||
7c673cae FG |
1804 | struct Params { |
1805 | string prefix; | |
1806 | string delim; | |
1807 | rgw_obj_key marker; | |
1808 | rgw_obj_key end_marker; | |
1809 | string ns; | |
1810 | bool enforce_ns; | |
1811 | RGWAccessListFilter *filter; | |
1812 | bool list_versions; | |
1adf2230 AA |
1813 | bool allow_unordered; |
1814 | ||
1815 | Params() : | |
1816 | enforce_ns(true), | |
1817 | filter(NULL), | |
1818 | list_versions(false), | |
1819 | allow_unordered(false) | |
1820 | {} | |
7c673cae FG |
1821 | } params; |
1822 | ||
7c673cae FG |
1823 | explicit List(RGWRados::Bucket *_target) : target(_target) {} |
1824 | ||
1adf2230 AA |
1825 | int list_objects(int64_t max, |
1826 | vector<rgw_bucket_dir_entry> *result, | |
1827 | map<string, bool> *common_prefixes, | |
1828 | bool *is_truncated) { | |
1829 | if (params.allow_unordered) { | |
1830 | return list_objects_unordered(max, result, common_prefixes, | |
1831 | is_truncated); | |
1832 | } else { | |
1833 | return list_objects_ordered(max, result, common_prefixes, | |
1834 | is_truncated); | |
1835 | } | |
1836 | } | |
7c673cae FG |
1837 | rgw_obj_key& get_next_marker() { |
1838 | return next_marker; | |
1839 | } | |
1adf2230 AA |
1840 | }; // class List |
1841 | }; // class Bucket | |
7c673cae | 1842 | |
7c673cae FG |
1843 | int on_last_entry_in_listing(RGWBucketInfo& bucket_info, |
1844 | const std::string& obj_prefix, | |
1845 | const std::string& obj_delim, | |
1846 | std::function<int(const rgw_bucket_dir_entry&)> handler); | |
1847 | ||
1848 | bool swift_versioning_enabled(const RGWBucketInfo& bucket_info) const { | |
1849 | return bucket_info.has_swift_versioning() && | |
1850 | bucket_info.swift_ver_location.size(); | |
1851 | } | |
1852 | ||
1853 | int swift_versioning_copy(RGWObjectCtx& obj_ctx, /* in/out */ | |
1854 | const rgw_user& user, /* in */ | |
1855 | RGWBucketInfo& bucket_info, /* in */ | |
1856 | rgw_obj& obj); /* in */ | |
11fdf7f2 TL |
1857 | int swift_versioning_restore(RGWSysObjectCtx& sysobj_ctx, |
1858 | RGWObjectCtx& obj_ctx, /* in/out */ | |
7c673cae FG |
1859 | const rgw_user& user, /* in */ |
1860 | RGWBucketInfo& bucket_info, /* in */ | |
1861 | rgw_obj& obj, /* in */ | |
1862 | bool& restored); /* out */ | |
1863 | int copy_obj_to_remote_dest(RGWObjState *astate, | |
1864 | map<string, bufferlist>& src_attrs, | |
1865 | RGWRados::Object::Read& read_op, | |
1866 | const rgw_user& user_id, | |
1867 | rgw_obj& dest_obj, | |
1868 | ceph::real_time *mtime); | |
1869 | ||
1870 | enum AttrsMod { | |
1871 | ATTRSMOD_NONE = 0, | |
1872 | ATTRSMOD_REPLACE = 1, | |
1873 | ATTRSMOD_MERGE = 2 | |
1874 | }; | |
1875 | ||
11fdf7f2 | 1876 | int rewrite_obj(RGWBucketInfo& dest_bucket_info, const rgw_obj& obj); |
7c673cae FG |
1877 | |
1878 | int stat_remote_obj(RGWObjectCtx& obj_ctx, | |
1879 | const rgw_user& user_id, | |
7c673cae FG |
1880 | req_info *info, |
1881 | const string& source_zone, | |
1882 | rgw_obj& src_obj, | |
1883 | RGWBucketInfo& src_bucket_info, | |
1884 | real_time *src_mtime, | |
1885 | uint64_t *psize, | |
1886 | const real_time *mod_ptr, | |
1887 | const real_time *unmod_ptr, | |
1888 | bool high_precision_time, | |
1889 | const char *if_match, | |
1890 | const char *if_nomatch, | |
1891 | map<string, bufferlist> *pattrs, | |
11fdf7f2 | 1892 | map<string, string> *pheaders, |
7c673cae FG |
1893 | string *version_id, |
1894 | string *ptag, | |
1895 | string *petag); | |
1896 | ||
1897 | int fetch_remote_obj(RGWObjectCtx& obj_ctx, | |
1898 | const rgw_user& user_id, | |
7c673cae FG |
1899 | req_info *info, |
1900 | const string& source_zone, | |
11fdf7f2 TL |
1901 | const rgw_obj& dest_obj, |
1902 | const rgw_obj& src_obj, | |
7c673cae FG |
1903 | RGWBucketInfo& dest_bucket_info, |
1904 | RGWBucketInfo& src_bucket_info, | |
11fdf7f2 | 1905 | std::optional<rgw_placement_rule> dest_placement, |
7c673cae FG |
1906 | ceph::real_time *src_mtime, |
1907 | ceph::real_time *mtime, | |
1908 | const ceph::real_time *mod_ptr, | |
1909 | const ceph::real_time *unmod_ptr, | |
1910 | bool high_precision_time, | |
1911 | const char *if_match, | |
1912 | const char *if_nomatch, | |
1913 | AttrsMod attrs_mod, | |
1914 | bool copy_if_newer, | |
1915 | map<string, bufferlist>& attrs, | |
1916 | RGWObjCategory category, | |
11fdf7f2 | 1917 | std::optional<uint64_t> olh_epoch, |
7c673cae | 1918 | ceph::real_time delete_at, |
7c673cae | 1919 | string *ptag, |
11fdf7f2 | 1920 | string *petag, |
7c673cae | 1921 | void (*progress_cb)(off_t, void *), |
31f18b77 | 1922 | void *progress_data, |
81eedcae TL |
1923 | rgw_zone_set *zones_trace= nullptr, |
1924 | std::optional<uint64_t>* bytes_transferred = 0); | |
7c673cae FG |
1925 | /** |
1926 | * Copy an object. | |
1927 | * dest_obj: the object to copy into | |
1928 | * src_obj: the object to copy from | |
1929 | * attrs: usage depends on attrs_mod parameter | |
1930 | * attrs_mod: the modification mode of the attrs, may have the following values: | |
1931 | * ATTRSMOD_NONE - the attributes of the source object will be | |
1932 | * copied without modifications, attrs parameter is ignored; | |
1933 | * ATTRSMOD_REPLACE - new object will have the attributes provided by attrs | |
1934 | * parameter, source object attributes are not copied; | |
1935 | * ATTRSMOD_MERGE - any conflicting meta keys on the source object's attributes | |
1936 | * are overwritten by values contained in attrs parameter. | |
7c673cae FG |
1937 | * Returns: 0 on success, -ERR# otherwise. |
1938 | */ | |
1939 | int copy_obj(RGWObjectCtx& obj_ctx, | |
1940 | const rgw_user& user_id, | |
7c673cae FG |
1941 | req_info *info, |
1942 | const string& source_zone, | |
1943 | rgw_obj& dest_obj, | |
1944 | rgw_obj& src_obj, | |
1945 | RGWBucketInfo& dest_bucket_info, | |
1946 | RGWBucketInfo& src_bucket_info, | |
11fdf7f2 | 1947 | const rgw_placement_rule& dest_placement, |
7c673cae FG |
1948 | ceph::real_time *src_mtime, |
1949 | ceph::real_time *mtime, | |
1950 | const ceph::real_time *mod_ptr, | |
1951 | const ceph::real_time *unmod_ptr, | |
1952 | bool high_precision_time, | |
1953 | const char *if_match, | |
1954 | const char *if_nomatch, | |
1955 | AttrsMod attrs_mod, | |
1956 | bool copy_if_newer, | |
1957 | map<std::string, bufferlist>& attrs, | |
1958 | RGWObjCategory category, | |
1959 | uint64_t olh_epoch, | |
1960 | ceph::real_time delete_at, | |
1961 | string *version_id, | |
1962 | string *ptag, | |
11fdf7f2 | 1963 | string *petag, |
7c673cae FG |
1964 | void (*progress_cb)(off_t, void *), |
1965 | void *progress_data); | |
1966 | ||
1967 | int copy_obj_data(RGWObjectCtx& obj_ctx, | |
1968 | RGWBucketInfo& dest_bucket_info, | |
11fdf7f2 | 1969 | const rgw_placement_rule& dest_placement, |
7c673cae | 1970 | RGWRados::Object::Read& read_op, off_t end, |
11fdf7f2 | 1971 | const rgw_obj& dest_obj, |
7c673cae FG |
1972 | ceph::real_time *mtime, |
1973 | ceph::real_time set_mtime, | |
1974 | map<string, bufferlist>& attrs, | |
7c673cae FG |
1975 | uint64_t olh_epoch, |
1976 | ceph::real_time delete_at, | |
11fdf7f2 | 1977 | string *petag); |
7c673cae | 1978 | |
11fdf7f2 TL |
1979 | int transition_obj(RGWObjectCtx& obj_ctx, |
1980 | RGWBucketInfo& bucket_info, | |
1981 | rgw_obj& obj, | |
1982 | const rgw_placement_rule& placement_rule, | |
1983 | const real_time& mtime, | |
1984 | uint64_t olh_epoch); | |
1985 | ||
7c673cae FG |
1986 | int check_bucket_empty(RGWBucketInfo& bucket_info); |
1987 | ||
1988 | /** | |
1989 | * Delete a bucket. | |
1990 | * bucket: the name of the bucket to delete | |
1991 | * Returns 0 on success, -ERR# otherwise. | |
1992 | */ | |
1993 | int delete_bucket(RGWBucketInfo& bucket_info, RGWObjVersionTracker& objv_tracker, bool check_empty = true); | |
1994 | ||
7c673cae FG |
1995 | void wakeup_meta_sync_shards(set<int>& shard_ids); |
1996 | void wakeup_data_sync_shards(const string& source_zone, map<int, set<string> >& shard_ids); | |
1997 | ||
1998 | RGWMetaSyncStatusManager* get_meta_sync_manager(); | |
1999 | RGWDataSyncStatusManager* get_data_sync_manager(const std::string& source_zone); | |
2000 | ||
2001 | int set_bucket_owner(rgw_bucket& bucket, ACLOwner& owner); | |
2002 | int set_buckets_enabled(std::vector<rgw_bucket>& buckets, bool enabled); | |
2003 | int bucket_suspended(rgw_bucket& bucket, bool *suspended); | |
2004 | ||
2005 | /** Delete an object.*/ | |
2006 | int delete_obj(RGWObjectCtx& obj_ctx, | |
2007 | const RGWBucketInfo& bucket_owner, | |
2008 | const rgw_obj& src_obj, | |
2009 | int versioning_status, | |
2010 | uint16_t bilog_flags = 0, | |
31f18b77 FG |
2011 | const ceph::real_time& expiration_time = ceph::real_time(), |
2012 | rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 2013 | |
7c673cae FG |
2014 | int delete_raw_obj(const rgw_raw_obj& obj); |
2015 | ||
7c673cae | 2016 | /** Remove an object from the bucket index */ |
494da23a | 2017 | int delete_obj_index(const rgw_obj& obj, ceph::real_time mtime); |
7c673cae | 2018 | |
7c673cae FG |
2019 | /** |
2020 | * Set an attr on an object. | |
2021 | * bucket: name of the bucket holding the object | |
2022 | * obj: name of the object to set the attr on | |
2023 | * name: the attr to set | |
2024 | * bl: the contents of the attr | |
2025 | * Returns: 0 on success, -ERR# otherwise. | |
2026 | */ | |
2027 | int set_attr(void *ctx, const RGWBucketInfo& bucket_info, rgw_obj& obj, const char *name, bufferlist& bl); | |
2028 | ||
2029 | int set_attrs(void *ctx, const RGWBucketInfo& bucket_info, rgw_obj& obj, | |
2030 | map<string, bufferlist>& attrs, | |
2031 | map<string, bufferlist>* rmattrs); | |
2032 | ||
7c673cae FG |
2033 | int get_obj_state(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state, |
2034 | bool follow_olh, bool assume_noent = false); | |
2035 | int get_obj_state(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state) { | |
2036 | return get_obj_state(rctx, bucket_info, obj, state, true); | |
2037 | } | |
2038 | ||
11fdf7f2 TL |
2039 | using iterate_obj_cb = int (*)(const rgw_raw_obj&, off_t, off_t, |
2040 | off_t, bool, RGWObjState*, void*); | |
2041 | ||
2042 | int iterate_obj(RGWObjectCtx& ctx, const RGWBucketInfo& bucket_info, | |
2043 | const rgw_obj& obj, off_t ofs, off_t end, | |
2044 | uint64_t max_chunk_size, iterate_obj_cb cb, void *arg); | |
7c673cae FG |
2045 | |
2046 | int flush_read_list(struct get_obj_data *d); | |
2047 | ||
11fdf7f2 TL |
2048 | int get_obj_iterate_cb(const rgw_raw_obj& read_obj, off_t obj_ofs, |
2049 | off_t read_ofs, off_t len, bool is_head_obj, | |
2050 | RGWObjState *astate, void *arg); | |
7c673cae FG |
2051 | |
2052 | void get_obj_aio_completion_cb(librados::completion_t cb, void *arg); | |
2053 | ||
2054 | /** | |
2055 | * a simple object read without keeping state | |
2056 | */ | |
2057 | ||
11fdf7f2 TL |
2058 | int raw_obj_stat(rgw_raw_obj& obj, uint64_t *psize, ceph::real_time *pmtime, uint64_t *epoch, |
2059 | map<string, bufferlist> *attrs, bufferlist *first_chunk, | |
2060 | RGWObjVersionTracker *objv_tracker); | |
7c673cae FG |
2061 | |
2062 | int obj_operate(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::ObjectWriteOperation *op); | |
2063 | int obj_operate(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::ObjectReadOperation *op); | |
2064 | ||
f64942e4 AA |
2065 | int guard_reshard(BucketShard *bs, |
2066 | const rgw_obj& obj_instance, | |
2067 | const RGWBucketInfo& bucket_info, | |
2068 | std::function<int(BucketShard *)> call); | |
2069 | int block_while_resharding(RGWRados::BucketShard *bs, | |
2070 | string *new_bucket_id, | |
11fdf7f2 TL |
2071 | const RGWBucketInfo& bucket_info, |
2072 | optional_yield y); | |
31f18b77 | 2073 | |
7c673cae FG |
2074 | void bucket_index_guard_olh_op(RGWObjState& olh_state, librados::ObjectOperation& op); |
2075 | int olh_init_modification(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, string *op_tag); | |
2076 | int olh_init_modification_impl(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, string *op_tag); | |
2077 | int bucket_index_link_olh(const RGWBucketInfo& bucket_info, RGWObjState& olh_state, | |
2078 | const rgw_obj& obj_instance, bool delete_marker, | |
2079 | const string& op_tag, struct rgw_bucket_dir_entry_meta *meta, | |
2080 | uint64_t olh_epoch, | |
91327a77 AA |
2081 | ceph::real_time unmod_since, bool high_precision_time, |
2082 | rgw_zone_set *zones_trace = nullptr, | |
2083 | bool log_data_change = false); | |
31f18b77 | 2084 | int bucket_index_unlink_instance(const RGWBucketInfo& bucket_info, const rgw_obj& obj_instance, const string& op_tag, const string& olh_tag, uint64_t olh_epoch, rgw_zone_set *zones_trace = nullptr); |
7c673cae FG |
2085 | int bucket_index_read_olh_log(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& obj_instance, uint64_t ver_marker, |
2086 | map<uint64_t, vector<rgw_bucket_olh_log_entry> > *log, bool *is_truncated); | |
2087 | int bucket_index_trim_olh_log(const RGWBucketInfo& bucket_info, RGWObjState& obj_state, const rgw_obj& obj_instance, uint64_t ver); | |
2088 | int bucket_index_clear_olh(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& obj_instance); | |
2089 | int apply_olh_log(RGWObjectCtx& ctx, RGWObjState& obj_state, const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
2090 | bufferlist& obj_tag, map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log, | |
31f18b77 FG |
2091 | uint64_t *plast_ver, rgw_zone_set *zones_trace = nullptr); |
2092 | int update_olh(RGWObjectCtx& obj_ctx, RGWObjState *state, const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 2093 | int set_olh(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, const rgw_obj& target_obj, bool delete_marker, rgw_bucket_dir_entry_meta *meta, |
91327a77 AA |
2094 | uint64_t olh_epoch, ceph::real_time unmod_since, bool high_precision_time, |
2095 | rgw_zone_set *zones_trace = nullptr, bool log_data_change = false); | |
a8e16298 TL |
2096 | int repair_olh(RGWObjState* state, const RGWBucketInfo& bucket_info, |
2097 | const rgw_obj& obj); | |
7c673cae | 2098 | int unlink_obj_instance(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, const rgw_obj& target_obj, |
31f18b77 | 2099 | uint64_t olh_epoch, rgw_zone_set *zones_trace = nullptr); |
7c673cae FG |
2100 | |
2101 | void check_pending_olh_entries(map<string, bufferlist>& pending_entries, map<string, bufferlist> *rm_pending_entries); | |
2102 | int remove_olh_pending_entries(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, map<string, bufferlist>& pending_attrs); | |
2103 | int follow_olh(const RGWBucketInfo& bucket_info, RGWObjectCtx& ctx, RGWObjState *state, const rgw_obj& olh_obj, rgw_obj *target); | |
2104 | int get_olh(const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWOLHInfo *olh); | |
2105 | ||
11fdf7f2 | 2106 | void gen_rand_obj_instance_name(rgw_obj_key *target_key); |
7c673cae FG |
2107 | void gen_rand_obj_instance_name(rgw_obj *target); |
2108 | ||
7c673cae FG |
2109 | int update_containers_stats(map<string, RGWBucketEnt>& m); |
2110 | int append_async(rgw_raw_obj& obj, size_t size, bufferlist& bl); | |
2111 | ||
11fdf7f2 | 2112 | public: |
7c673cae FG |
2113 | void set_atomic(void *ctx, rgw_obj& obj) { |
2114 | RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx); | |
11fdf7f2 | 2115 | rctx->set_atomic(obj); |
7c673cae | 2116 | } |
11fdf7f2 | 2117 | void set_prefetch_data(void *ctx, const rgw_obj& obj) { |
7c673cae | 2118 | RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx); |
11fdf7f2 | 2119 | rctx->set_prefetch_data(obj); |
7c673cae | 2120 | } |
7c673cae FG |
2121 | int decode_policy(bufferlist& bl, ACLOwner *owner); |
2122 | int get_bucket_stats(RGWBucketInfo& bucket_info, int shard_id, string *bucket_ver, string *master_ver, | |
c07f9fc5 | 2123 | map<RGWObjCategory, RGWStorageStats>& stats, string *max_marker, bool* syncstopped = NULL); |
7c673cae FG |
2124 | int get_bucket_stats_async(RGWBucketInfo& bucket_info, int shard_id, RGWGetBucketStats_CB *cb); |
2125 | int get_user_stats(const rgw_user& user, RGWStorageStats& stats); | |
2126 | int get_user_stats_async(const rgw_user& user, RGWGetUserStats_CB *cb); | |
2127 | void get_bucket_instance_obj(const rgw_bucket& bucket, rgw_raw_obj& obj); | |
2128 | void get_bucket_meta_oid(const rgw_bucket& bucket, string& oid); | |
2129 | ||
2130 | int put_bucket_entrypoint_info(const string& tenant_name, const string& bucket_name, RGWBucketEntryPoint& entry_point, | |
2131 | bool exclusive, RGWObjVersionTracker& objv_tracker, ceph::real_time mtime, | |
2132 | map<string, bufferlist> *pattrs); | |
2133 | int put_bucket_instance_info(RGWBucketInfo& info, bool exclusive, ceph::real_time mtime, map<string, bufferlist> *pattrs); | |
11fdf7f2 | 2134 | int get_bucket_entrypoint_info(RGWSysObjectCtx& obj_ctx, const string& tenant_name, const string& bucket_name, |
7c673cae | 2135 | RGWBucketEntryPoint& entry_point, RGWObjVersionTracker *objv_tracker, |
b32b8144 FG |
2136 | ceph::real_time *pmtime, map<string, bufferlist> *pattrs, rgw_cache_entry_info *cache_info = NULL, |
2137 | boost::optional<obj_version> refresh_version = boost::none); | |
11fdf7f2 TL |
2138 | int get_bucket_instance_info(RGWSysObjectCtx& obj_ctx, const string& meta_key, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs); |
2139 | int get_bucket_instance_info(RGWSysObjectCtx& obj_ctx, const rgw_bucket& bucket, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs); | |
2140 | int get_bucket_instance_from_oid(RGWSysObjectCtx& obj_ctx, const string& oid, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs, | |
b32b8144 FG |
2141 | rgw_cache_entry_info *cache_info = NULL, |
2142 | boost::optional<obj_version> refresh_version = boost::none); | |
7c673cae | 2143 | |
11fdf7f2 | 2144 | int convert_old_bucket_info(RGWSysObjectCtx& obj_ctx, const string& tenant_name, const string& bucket_name); |
7c673cae | 2145 | static void make_bucket_entry_name(const string& tenant_name, const string& bucket_name, string& bucket_entry); |
b32b8144 FG |
2146 | |
2147 | ||
2148 | private: | |
11fdf7f2 | 2149 | int _get_bucket_info(RGWSysObjectCtx& obj_ctx, const string& tenant, |
b32b8144 FG |
2150 | const string& bucket_name, RGWBucketInfo& info, |
2151 | real_time *pmtime, | |
2152 | map<string, bufferlist> *pattrs, | |
2153 | boost::optional<obj_version> refresh_version); | |
2154 | public: | |
2155 | ||
11fdf7f2 TL |
2156 | bool call(std::string_view command, const cmdmap_t& cmdmap, |
2157 | std::string_view format, | |
3a9019d9 FG |
2158 | bufferlist& out) override final; |
2159 | ||
3a9019d9 | 2160 | protected: |
3a9019d9 FG |
2161 | // `call_list` must iterate over all cache entries and call |
2162 | // `cache_list_dump_helper` with the supplied Formatter on any that | |
2163 | // include `filter` as a substring. | |
2164 | // | |
11fdf7f2 | 2165 | void call_list(const std::optional<std::string>& filter, |
3a9019d9 FG |
2166 | Formatter* format); |
2167 | // `call_inspect` must look up the requested target and, if found, | |
2168 | // dump it to the supplied Formatter and return true. If not found, | |
2169 | // it must return false. | |
2170 | // | |
11fdf7f2 | 2171 | bool call_inspect(const std::string& target, Formatter* format); |
3a9019d9 FG |
2172 | |
2173 | // `call_erase` must erase the requested target and return true. If | |
2174 | // the requested target does not exist, it should return false. | |
11fdf7f2 | 2175 | bool call_erase(const std::string& target); |
3a9019d9 FG |
2176 | |
2177 | // `call_zap` must erase the cache. | |
11fdf7f2 | 2178 | void call_zap(); |
3a9019d9 | 2179 | public: |
b32b8144 | 2180 | |
11fdf7f2 | 2181 | int get_bucket_info(RGWSysObjectCtx& obj_ctx, |
b32b8144 FG |
2182 | const string& tenant_name, const string& bucket_name, |
2183 | RGWBucketInfo& info, | |
2184 | ceph::real_time *pmtime, map<string, bufferlist> *pattrs = NULL); | |
2185 | ||
81eedcae TL |
2186 | // Returns 0 on successful refresh. Returns error code if there was |
2187 | // an error or the version stored on the OSD is the same as that | |
b32b8144 FG |
2188 | // presented in the BucketInfo structure. |
2189 | // | |
2190 | int try_refresh_bucket_info(RGWBucketInfo& info, | |
2191 | ceph::real_time *pmtime, | |
2192 | map<string, bufferlist> *pattrs = nullptr); | |
2193 | ||
7c673cae | 2194 | int put_linked_bucket_info(RGWBucketInfo& info, bool exclusive, ceph::real_time mtime, obj_version *pep_objv, |
b32b8144 | 2195 | map<string, bufferlist> *pattrs, bool create_entry_point); |
7c673cae | 2196 | |
31f18b77 FG |
2197 | int cls_obj_prepare_op(BucketShard& bs, RGWModifyOp op, string& tag, rgw_obj& obj, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); |
2198 | int cls_obj_complete_op(BucketShard& bs, const rgw_obj& obj, RGWModifyOp op, string& tag, int64_t pool, uint64_t epoch, | |
2199 | rgw_bucket_dir_entry& ent, RGWObjCategory category, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); | |
2200 | int cls_obj_complete_add(BucketShard& bs, const rgw_obj& obj, string& tag, int64_t pool, uint64_t epoch, rgw_bucket_dir_entry& ent, | |
2201 | RGWObjCategory category, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 2202 | int cls_obj_complete_del(BucketShard& bs, string& tag, int64_t pool, uint64_t epoch, rgw_obj& obj, |
31f18b77 FG |
2203 | ceph::real_time& removed_mtime, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); |
2204 | int cls_obj_complete_cancel(BucketShard& bs, string& tag, rgw_obj& obj, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 2205 | int cls_obj_set_bucket_tag_timeout(RGWBucketInfo& bucket_info, uint64_t timeout); |
1adf2230 | 2206 | int cls_bucket_list_ordered(RGWBucketInfo& bucket_info, int shard_id, |
11fdf7f2 TL |
2207 | const rgw_obj_index_key& start, |
2208 | const string& prefix, | |
1adf2230 AA |
2209 | uint32_t num_entries, bool list_versions, |
2210 | map<string, rgw_bucket_dir_entry>& m, | |
2211 | bool *is_truncated, | |
2212 | rgw_obj_index_key *last_entry, | |
2213 | bool (*force_check_filter)(const string& name) = nullptr); | |
2214 | int cls_bucket_list_unordered(RGWBucketInfo& bucket_info, int shard_id, | |
11fdf7f2 TL |
2215 | const rgw_obj_index_key& start, |
2216 | const string& prefix, | |
1adf2230 AA |
2217 | uint32_t num_entries, bool list_versions, |
2218 | vector<rgw_bucket_dir_entry>& ent_list, | |
2219 | bool *is_truncated, rgw_obj_index_key *last_entry, | |
2220 | bool (*force_check_filter)(const string& name) = nullptr); | |
a8e16298 | 2221 | int cls_bucket_head(const RGWBucketInfo& bucket_info, int shard_id, vector<rgw_bucket_dir_header>& headers, map<int, string> *bucket_instance_ids = NULL); |
7c673cae FG |
2222 | int cls_bucket_head_async(const RGWBucketInfo& bucket_info, int shard_id, RGWGetDirHeader_CB *ctx, int *num_aio); |
2223 | int list_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id, string& marker, uint32_t max, std::list<rgw_bi_log_entry>& result, bool *truncated); | |
2224 | int trim_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id, string& marker, string& end_marker); | |
c07f9fc5 FG |
2225 | int resync_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id); |
2226 | int stop_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id); | |
7c673cae FG |
2227 | int get_bi_log_status(RGWBucketInfo& bucket_info, int shard_id, map<int, string>& max_marker); |
2228 | ||
a8e16298 TL |
2229 | int bi_get_instance(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_bucket_dir_entry *dirent); |
2230 | int bi_get_olh(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_bucket_olh_entry *olh); | |
2231 | int bi_get(const RGWBucketInfo& bucket_info, const rgw_obj& obj, BIIndexType index_type, rgw_cls_bi_entry *entry); | |
7c673cae FG |
2232 | void bi_put(librados::ObjectWriteOperation& op, BucketShard& bs, rgw_cls_bi_entry& entry); |
2233 | int bi_put(BucketShard& bs, rgw_cls_bi_entry& entry); | |
2234 | int bi_put(rgw_bucket& bucket, rgw_obj& obj, rgw_cls_bi_entry& entry); | |
2235 | int bi_list(rgw_bucket& bucket, int shard_id, const string& filter_obj, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated); | |
2236 | int bi_list(BucketShard& bs, const string& filter_obj, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated); | |
2237 | int bi_list(rgw_bucket& bucket, const string& obj_name, const string& marker, uint32_t max, | |
2238 | list<rgw_cls_bi_entry> *entries, bool *is_truncated); | |
2239 | int bi_remove(BucketShard& bs); | |
2240 | ||
2241 | int cls_obj_usage_log_add(const string& oid, rgw_usage_log_info& info); | |
11fdf7f2 TL |
2242 | int cls_obj_usage_log_read(const string& oid, const string& user, const string& bucket, uint64_t start_epoch, |
2243 | uint64_t end_epoch, uint32_t max_entries, string& read_iter, map<rgw_user_bucket, | |
2244 | rgw_usage_log_entry>& usage, bool *is_truncated); | |
2245 | int cls_obj_usage_log_trim(const string& oid, const string& user, const string& bucket, uint64_t start_epoch, | |
2246 | uint64_t end_epoch); | |
2247 | int cls_obj_usage_log_clear(string& oid); | |
7c673cae FG |
2248 | |
2249 | int key_to_shard_id(const string& key, int max_shards); | |
2250 | void shard_name(const string& prefix, unsigned max_shards, const string& key, string& name, int *shard_id); | |
2251 | void shard_name(const string& prefix, unsigned max_shards, const string& section, const string& key, string& name); | |
2252 | void shard_name(const string& prefix, unsigned shard_id, string& name); | |
2253 | int get_target_shard_id(const RGWBucketInfo& bucket_info, const string& obj_key, int *shard_id); | |
2254 | void time_log_prepare_entry(cls_log_entry& entry, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl); | |
2255 | int time_log_add_init(librados::IoCtx& io_ctx); | |
2256 | int time_log_add(const string& oid, list<cls_log_entry>& entries, | |
2257 | librados::AioCompletion *completion, bool monotonic_inc = true); | |
2258 | int time_log_add(const string& oid, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl); | |
2259 | int time_log_list(const string& oid, const ceph::real_time& start_time, const ceph::real_time& end_time, | |
2260 | int max_entries, list<cls_log_entry>& entries, | |
2261 | const string& marker, string *out_marker, bool *truncated); | |
2262 | int time_log_info(const string& oid, cls_log_header *header); | |
2263 | int time_log_info_async(librados::IoCtx& io_ctx, const string& oid, cls_log_header *header, librados::AioCompletion *completion); | |
2264 | int time_log_trim(const string& oid, const ceph::real_time& start_time, const ceph::real_time& end_time, | |
2265 | const string& from_marker, const string& to_marker, | |
2266 | librados::AioCompletion *completion = nullptr); | |
2267 | ||
2268 | string objexp_hint_get_shardname(int shard_num); | |
2269 | int objexp_key_shard(const rgw_obj_index_key& key); | |
2270 | void objexp_get_shard(int shard_num, | |
2271 | string& shard); /* out */ | |
2272 | int objexp_hint_add(const ceph::real_time& delete_at, | |
2273 | const string& tenant_name, | |
2274 | const string& bucket_name, | |
2275 | const string& bucket_id, | |
2276 | const rgw_obj_index_key& obj_key); | |
2277 | int objexp_hint_list(const string& oid, | |
2278 | const ceph::real_time& start_time, | |
2279 | const ceph::real_time& end_time, | |
2280 | const int max_entries, | |
2281 | const string& marker, | |
2282 | list<cls_timeindex_entry>& entries, /* out */ | |
2283 | string *out_marker, /* out */ | |
2284 | bool *truncated); /* out */ | |
2285 | int objexp_hint_parse(cls_timeindex_entry &ti_entry, | |
2286 | objexp_hint_entry& hint_entry); /* out */ | |
2287 | int objexp_hint_trim(const string& oid, | |
2288 | const ceph::real_time& start_time, | |
2289 | const ceph::real_time& end_time, | |
2290 | const string& from_marker = std::string(), | |
2291 | const string& to_marker = std::string()); | |
2292 | ||
11fdf7f2 TL |
2293 | int lock_exclusive(const rgw_pool& pool, const string& oid, ceph::timespan& duration, string& zone_id, string& owner_id); |
2294 | int unlock(const rgw_pool& pool, const string& oid, string& zone_id, string& owner_id); | |
7c673cae FG |
2295 | |
2296 | void update_gc_chain(rgw_obj& head_obj, RGWObjManifest& manifest, cls_rgw_obj_chain *chain); | |
2297 | int send_chain_to_gc(cls_rgw_obj_chain& chain, const string& tag, bool sync); | |
2298 | int gc_operate(string& oid, librados::ObjectWriteOperation *op); | |
11fdf7f2 | 2299 | int gc_aio_operate(string& oid, librados::ObjectWriteOperation *op, librados::AioCompletion **pc = nullptr); |
7c673cae FG |
2300 | int gc_operate(string& oid, librados::ObjectReadOperation *op, bufferlist *pbl); |
2301 | ||
2302 | int list_gc_objs(int *index, string& marker, uint32_t max, bool expired_only, std::list<cls_rgw_gc_obj_info>& result, bool *truncated); | |
11fdf7f2 | 2303 | int process_gc(bool expired_only); |
1adf2230 | 2304 | bool process_expire_objects(); |
7c673cae FG |
2305 | int defer_gc(void *ctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj); |
2306 | ||
2307 | int process_lc(); | |
2308 | int list_lc_progress(const string& marker, uint32_t max_entries, map<string, int> *progress_map); | |
2309 | ||
2310 | int bucket_check_index(RGWBucketInfo& bucket_info, | |
2311 | map<RGWObjCategory, RGWStorageStats> *existing_stats, | |
2312 | map<RGWObjCategory, RGWStorageStats> *calculated_stats); | |
2313 | int bucket_rebuild_index(RGWBucketInfo& bucket_info); | |
f64942e4 | 2314 | int bucket_set_reshard(const RGWBucketInfo& bucket_info, const cls_rgw_bucket_instance_entry& entry); |
7c673cae FG |
2315 | int remove_objs_from_index(RGWBucketInfo& bucket_info, list<rgw_obj_index_key>& oid_list); |
2316 | int move_rados_obj(librados::IoCtx& src_ioctx, | |
2317 | const string& src_oid, const string& src_locator, | |
2318 | librados::IoCtx& dst_ioctx, | |
2319 | const string& dst_oid, const string& dst_locator); | |
2320 | int fix_head_obj_locator(const RGWBucketInfo& bucket_info, bool copy_obj, bool remove_bad, rgw_obj_key& key); | |
2321 | int fix_tail_obj_locator(const RGWBucketInfo& bucket_info, rgw_obj_key& key, bool fix, bool *need_fix); | |
2322 | ||
2323 | int cls_user_get_header(const string& user_id, cls_user_header *header); | |
94b18763 | 2324 | int cls_user_reset_stats(const string& user_id); |
7c673cae FG |
2325 | int cls_user_get_header_async(const string& user_id, RGWGetUserHeader_CB *ctx); |
2326 | int cls_user_sync_bucket_stats(rgw_raw_obj& user_obj, const RGWBucketInfo& bucket_info); | |
2327 | int cls_user_list_buckets(rgw_raw_obj& obj, | |
2328 | const string& in_marker, | |
2329 | const string& end_marker, | |
2330 | int max_entries, | |
2331 | list<cls_user_bucket_entry>& entries, | |
2332 | string *out_marker, | |
2333 | bool *truncated); | |
2334 | int cls_user_add_bucket(rgw_raw_obj& obj, const cls_user_bucket_entry& entry); | |
2335 | int cls_user_update_buckets(rgw_raw_obj& obj, list<cls_user_bucket_entry>& entries, bool add); | |
2336 | int cls_user_complete_stats_sync(rgw_raw_obj& obj); | |
2337 | int complete_sync_user_stats(const rgw_user& user_id); | |
7c673cae | 2338 | int cls_user_remove_bucket(rgw_raw_obj& obj, const cls_user_bucket& bucket); |
c07f9fc5 | 2339 | int cls_user_get_bucket_stats(const rgw_bucket& bucket, cls_user_bucket_entry& entry); |
7c673cae FG |
2340 | |
2341 | int check_quota(const rgw_user& bucket_owner, rgw_bucket& bucket, | |
11fdf7f2 | 2342 | RGWQuotaInfo& user_quota, RGWQuotaInfo& bucket_quota, uint64_t obj_size, bool check_size_only = false); |
7c673cae | 2343 | |
224ce89b | 2344 | int check_bucket_shards(const RGWBucketInfo& bucket_info, const rgw_bucket& bucket, |
31f18b77 FG |
2345 | RGWQuotaInfo& bucket_quota); |
2346 | ||
2347 | int add_bucket_to_reshard(const RGWBucketInfo& bucket_info, uint32_t new_num_shards); | |
2348 | ||
7c673cae | 2349 | uint64_t instance_id(); |
3efd9988 | 2350 | |
7c673cae FG |
2351 | librados::Rados* get_rados_handle(); |
2352 | ||
2353 | int delete_raw_obj_aio(const rgw_raw_obj& obj, list<librados::AioCompletion *>& handles); | |
2354 | int delete_obj_aio(const rgw_obj& obj, RGWBucketInfo& info, RGWObjState *astate, | |
2355 | list<librados::AioCompletion *>& handles, bool keep_index_consistent); | |
11fdf7f2 TL |
2356 | |
2357 | /* mfa/totp stuff */ | |
2358 | private: | |
2359 | void prepare_mfa_write(librados::ObjectWriteOperation *op, | |
2360 | RGWObjVersionTracker *objv_tracker, | |
2361 | const ceph::real_time& mtime); | |
2362 | public: | |
2363 | string get_mfa_oid(const rgw_user& user); | |
2364 | int get_mfa_ref(const rgw_user& user, rgw_rados_ref *ref); | |
2365 | int check_mfa(const rgw_user& user, const string& otp_id, const string& pin); | |
2366 | int create_mfa(const rgw_user& user, const rados::cls::otp::otp_info_t& config, | |
2367 | RGWObjVersionTracker *objv_tracker, const ceph::real_time& mtime); | |
2368 | int remove_mfa(const rgw_user& user, const string& id, | |
2369 | RGWObjVersionTracker *objv_tracker, const ceph::real_time& mtime); | |
2370 | int get_mfa(const rgw_user& user, const string& id, rados::cls::otp::otp_info_t *result); | |
2371 | int list_mfa(const rgw_user& user, list<rados::cls::otp::otp_info_t> *result); | |
2372 | int otp_get_current_time(const rgw_user& user, ceph::real_time *result); | |
2373 | ||
2374 | /* mfa interfaces used by metadata engine */ | |
2375 | int set_mfa(const string& oid, const list<rados::cls::otp::otp_info_t>& entries, bool reset_obj, | |
2376 | RGWObjVersionTracker *objv_tracker, const ceph::real_time& mtime); | |
2377 | int list_mfa(const string& oid, list<rados::cls::otp::otp_info_t> *result, | |
2378 | RGWObjVersionTracker *objv_tracker, ceph::real_time *pmtime); | |
7c673cae FG |
2379 | private: |
2380 | /** | |
2381 | * This is a helper method, it generates a list of bucket index objects with the given | |
2382 | * bucket base oid and number of shards. | |
2383 | * | |
2384 | * bucket_oid_base [in] - base name of the bucket index object; | |
2385 | * num_shards [in] - number of bucket index object shards. | |
2386 | * bucket_objs [out] - filled by this method, a list of bucket index objects. | |
2387 | */ | |
2388 | void get_bucket_index_objects(const string& bucket_oid_base, uint32_t num_shards, | |
2389 | map<int, string>& bucket_objs, int shard_id = -1); | |
2390 | ||
2391 | /** | |
2392 | * Get the bucket index object with the given base bucket index object and object key, | |
2393 | * and the number of bucket index shards. | |
2394 | * | |
2395 | * bucket_oid_base [in] - bucket object base name. | |
2396 | * obj_key [in] - object key. | |
2397 | * num_shards [in] - number of bucket index shards. | |
2398 | * hash_type [in] - type of hash to find the shard ID. | |
2399 | * bucket_obj [out] - the bucket index object for the given object. | |
2400 | * | |
2401 | * Return 0 on success, a failure code otherwise. | |
2402 | */ | |
2403 | int get_bucket_index_object(const string& bucket_oid_base, const string& obj_key, | |
2404 | uint32_t num_shards, RGWBucketInfo::BIShardsHashType hash_type, string *bucket_obj, int *shard); | |
2405 | ||
2406 | void get_bucket_index_object(const string& bucket_oid_base, uint32_t num_shards, | |
2407 | int shard_id, string *bucket_obj); | |
2408 | ||
2409 | /** | |
2410 | * Check the actual on-disk state of the object specified | |
2411 | * by list_state, and fill in the time and size of object. | |
2412 | * Then append any changes to suggested_updates for | |
2413 | * the rgw class' dir_suggest_changes function. | |
2414 | * | |
2415 | * Note that this can maul list_state; don't use it afterwards. Also | |
2416 | * it expects object to already be filled in from list_state; it only | |
2417 | * sets the size and mtime. | |
2418 | * | |
2419 | * Returns 0 on success, -ENOENT if the object doesn't exist on disk, | |
2420 | * and -errno on other failures. (-ENOENT is not a failure, and it | |
2421 | * will encode that info as a suggested update.) | |
2422 | */ | |
2423 | int check_disk_state(librados::IoCtx io_ctx, | |
2424 | const RGWBucketInfo& bucket_info, | |
2425 | rgw_bucket_dir_entry& list_state, | |
2426 | rgw_bucket_dir_entry& object, | |
2427 | bufferlist& suggested_updates); | |
2428 | ||
2429 | /** | |
2430 | * Init pool iteration | |
31f18b77 | 2431 | * pool: pool to use for the ctx initialization |
7c673cae FG |
2432 | * ctx: context object to use for the iteration |
2433 | * Returns: 0 on success, -ERR# otherwise. | |
2434 | */ | |
2435 | int pool_iterate_begin(const rgw_pool& pool, RGWPoolIterCtx& ctx); | |
31f18b77 | 2436 | |
181888fb FG |
2437 | /** |
2438 | * Init pool iteration | |
2439 | * pool: pool to use | |
2440 | * cursor: position to start iteration | |
2441 | * ctx: context object to use for the iteration | |
2442 | * Returns: 0 on success, -ERR# otherwise. | |
2443 | */ | |
2444 | int pool_iterate_begin(const rgw_pool& pool, const string& cursor, RGWPoolIterCtx& ctx); | |
2445 | ||
2446 | /** | |
2447 | * Get pool iteration position | |
2448 | * ctx: context object to use for the iteration | |
2449 | * Returns: string representation of position | |
2450 | */ | |
2451 | string pool_iterate_get_cursor(RGWPoolIterCtx& ctx); | |
2452 | ||
7c673cae FG |
2453 | /** |
2454 | * Iterate over pool return object names, use optional filter | |
2455 | * ctx: iteration context, initialized with pool_iterate_begin() | |
2456 | * num: max number of objects to return | |
2457 | * objs: a vector that the results will append into | |
2458 | * is_truncated: if not NULL, will hold true iff iteration is complete | |
2459 | * filter: if not NULL, will be used to filter returned objects | |
2460 | * Returns: 0 on success, -ERR# otherwise. | |
2461 | */ | |
2462 | int pool_iterate(RGWPoolIterCtx& ctx, uint32_t num, vector<rgw_bucket_dir_entry>& objs, | |
2463 | bool *is_truncated, RGWAccessListFilter *filter); | |
2464 | ||
2465 | uint64_t next_bucket_id(); | |
2466 | }; | |
2467 | ||
2468 | class RGWStoreManager { | |
2469 | public: | |
2470 | RGWStoreManager() {} | |
28e407b8 AA |
2471 | static RGWRados *get_storage(CephContext *cct, bool use_gc_thread, bool use_lc_thread, bool quota_threads, |
2472 | bool run_sync_thread, bool run_reshard_thread, bool use_cache = true) { | |
31f18b77 | 2473 | RGWRados *store = init_storage_provider(cct, use_gc_thread, use_lc_thread, quota_threads, run_sync_thread, |
28e407b8 | 2474 | run_reshard_thread, use_cache); |
7c673cae FG |
2475 | return store; |
2476 | } | |
2477 | static RGWRados *get_raw_storage(CephContext *cct) { | |
2478 | RGWRados *store = init_raw_storage_provider(cct); | |
2479 | return store; | |
2480 | } | |
28e407b8 | 2481 | static RGWRados *init_storage_provider(CephContext *cct, bool use_gc_thread, bool use_lc_thread, bool quota_threads, bool run_sync_thread, bool run_reshard_thread, bool use_metadata_cache); |
7c673cae FG |
2482 | static RGWRados *init_raw_storage_provider(CephContext *cct); |
2483 | static void close_storage(RGWRados *store); | |
2484 | ||
2485 | }; | |
2486 | ||
7c673cae FG |
2487 | class RGWMPObj { |
2488 | string oid; | |
2489 | string prefix; | |
2490 | string meta; | |
2491 | string upload_id; | |
2492 | public: | |
2493 | RGWMPObj() {} | |
2494 | RGWMPObj(const string& _oid, const string& _upload_id) { | |
2495 | init(_oid, _upload_id, _upload_id); | |
2496 | } | |
2497 | void init(const string& _oid, const string& _upload_id) { | |
2498 | init(_oid, _upload_id, _upload_id); | |
2499 | } | |
2500 | void init(const string& _oid, const string& _upload_id, const string& part_unique_str) { | |
2501 | if (_oid.empty()) { | |
2502 | clear(); | |
2503 | return; | |
2504 | } | |
2505 | oid = _oid; | |
2506 | upload_id = _upload_id; | |
2507 | prefix = oid + "."; | |
2508 | meta = prefix + upload_id + MP_META_SUFFIX; | |
2509 | prefix.append(part_unique_str); | |
2510 | } | |
11fdf7f2 TL |
2511 | const string& get_meta() const { return meta; } |
2512 | string get_part(int num) const { | |
7c673cae FG |
2513 | char buf[16]; |
2514 | snprintf(buf, 16, ".%d", num); | |
2515 | string s = prefix; | |
2516 | s.append(buf); | |
2517 | return s; | |
2518 | } | |
11fdf7f2 | 2519 | string get_part(const string& part) const { |
7c673cae FG |
2520 | string s = prefix; |
2521 | s.append("."); | |
2522 | s.append(part); | |
2523 | return s; | |
2524 | } | |
11fdf7f2 | 2525 | const string& get_upload_id() const { |
7c673cae FG |
2526 | return upload_id; |
2527 | } | |
11fdf7f2 | 2528 | const string& get_key() const { |
7c673cae FG |
2529 | return oid; |
2530 | } | |
2531 | bool from_meta(string& meta) { | |
2532 | int end_pos = meta.rfind('.'); // search for ".meta" | |
2533 | if (end_pos < 0) | |
2534 | return false; | |
2535 | int mid_pos = meta.rfind('.', end_pos - 1); // <key>.<upload_id> | |
2536 | if (mid_pos < 0) | |
2537 | return false; | |
2538 | oid = meta.substr(0, mid_pos); | |
2539 | upload_id = meta.substr(mid_pos + 1, end_pos - mid_pos - 1); | |
2540 | init(oid, upload_id, upload_id); | |
2541 | return true; | |
2542 | } | |
2543 | void clear() { | |
2544 | oid = ""; | |
2545 | prefix = ""; | |
2546 | meta = ""; | |
2547 | upload_id = ""; | |
2548 | } | |
11fdf7f2 | 2549 | }; // class RGWMPObj |
7c673cae | 2550 | |
11fdf7f2 TL |
2551 | |
2552 | class RGWRadosThread { | |
2553 | class Worker : public Thread { | |
2554 | CephContext *cct; | |
2555 | RGWRadosThread *processor; | |
2556 | Mutex lock; | |
2557 | Cond cond; | |
2558 | ||
2559 | void wait() { | |
2560 | Mutex::Locker l(lock); | |
2561 | cond.Wait(lock); | |
2562 | }; | |
2563 | ||
2564 | void wait_interval(const utime_t& wait_time) { | |
2565 | Mutex::Locker l(lock); | |
2566 | cond.WaitInterval(lock, wait_time); | |
2567 | } | |
2568 | ||
2569 | public: | |
2570 | Worker(CephContext *_cct, RGWRadosThread *_p) : cct(_cct), processor(_p), lock("RGWRadosThread::Worker") {} | |
2571 | void *entry() override; | |
2572 | void signal() { | |
2573 | Mutex::Locker l(lock); | |
2574 | cond.Signal(); | |
2575 | } | |
2576 | }; | |
2577 | ||
2578 | Worker *worker; | |
7c673cae FG |
2579 | |
2580 | protected: | |
11fdf7f2 TL |
2581 | CephContext *cct; |
2582 | RGWRados *store; | |
2583 | ||
2584 | std::atomic<bool> down_flag = { false }; | |
2585 | ||
2586 | string thread_name; | |
2587 | ||
2588 | virtual uint64_t interval_msec() = 0; | |
2589 | virtual void stop_process() {} | |
7c673cae | 2590 | public: |
11fdf7f2 TL |
2591 | RGWRadosThread(RGWRados *_store, const string& thread_name = "radosgw") |
2592 | : worker(NULL), cct(_store->ctx()), store(_store), thread_name(thread_name) {} | |
2593 | virtual ~RGWRadosThread() { | |
2594 | stop(); | |
2595 | } | |
2596 | ||
2597 | virtual int init() { return 0; } | |
2598 | virtual int process() = 0; | |
2599 | ||
2600 | bool going_down() { return down_flag; } | |
2601 | ||
2602 | void start(); | |
2603 | void stop(); | |
2604 | ||
2605 | void signal() { | |
2606 | if (worker) { | |
2607 | worker->signal(); | |
2608 | } | |
2609 | } | |
2610 | }; | |
2611 | ||
7c673cae | 2612 | #endif |