]>
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" | |
16 | #include "rgw_common.h" | |
17 | #include "cls/rgw/cls_rgw_types.h" | |
18 | #include "cls/version/cls_version_types.h" | |
19 | #include "cls/log/cls_log_types.h" | |
20 | #include "cls/statelog/cls_statelog_types.h" | |
21 | #include "cls/timeindex/cls_timeindex_types.h" | |
22 | #include "rgw_log.h" | |
23 | #include "rgw_metadata.h" | |
24 | #include "rgw_meta_sync_status.h" | |
25 | #include "rgw_period_puller.h" | |
26 | #include "rgw_sync_module.h" | |
b32b8144 | 27 | #include "rgw_sync_log_trim.h" |
7c673cae FG |
28 | |
29 | class RGWWatcher; | |
30 | class SafeTimer; | |
31 | class ACLOwner; | |
32 | class RGWGC; | |
33 | class RGWMetaNotifier; | |
34 | class RGWDataNotifier; | |
35 | class RGWLC; | |
36 | class RGWObjectExpirer; | |
37 | class RGWMetaSyncProcessorThread; | |
38 | class RGWDataSyncProcessorThread; | |
39 | class RGWSyncLogTrimThread; | |
40 | class RGWRESTConn; | |
41 | struct RGWZoneGroup; | |
42 | struct RGWZoneParams; | |
31f18b77 FG |
43 | class RGWReshard; |
44 | class RGWReshardWait; | |
7c673cae FG |
45 | |
46 | /* flags for put_obj_meta() */ | |
47 | #define PUT_OBJ_CREATE 0x01 | |
48 | #define PUT_OBJ_EXCL 0x02 | |
49 | #define PUT_OBJ_CREATE_EXCL (PUT_OBJ_CREATE | PUT_OBJ_EXCL) | |
50 | ||
51 | #define RGW_OBJ_NS_MULTIPART "multipart" | |
52 | #define RGW_OBJ_NS_SHADOW "shadow" | |
53 | ||
54 | #define RGW_BUCKET_INSTANCE_MD_PREFIX ".bucket.meta." | |
55 | ||
56 | #define RGW_NO_SHARD -1 | |
57 | ||
31f18b77 FG |
58 | #define RGW_SHARDS_PRIME_0 7877 |
59 | #define RGW_SHARDS_PRIME_1 65521 | |
60 | ||
1adf2230 | 61 | // only called by rgw_shard_id and rgw_bucket_shard_index |
31f18b77 FG |
62 | static inline int rgw_shards_mod(unsigned hval, int max_shards) |
63 | { | |
64 | if (max_shards <= RGW_SHARDS_PRIME_0) { | |
65 | return hval % RGW_SHARDS_PRIME_0 % max_shards; | |
66 | } | |
67 | return hval % RGW_SHARDS_PRIME_1 % max_shards; | |
68 | } | |
69 | ||
1adf2230 AA |
70 | // used for logging and tagging |
71 | static inline int rgw_shard_id(const string& key, int max_shards) | |
31f18b77 | 72 | { |
1adf2230 AA |
73 | return rgw_shards_mod(ceph_str_hash_linux(key.c_str(), key.size()), |
74 | max_shards); | |
75 | } | |
76 | ||
77 | // used for bucket indices | |
78 | static inline uint32_t rgw_bucket_shard_index(const std::string& key, | |
79 | int num_shards) { | |
80 | uint32_t sid = ceph_str_hash_linux(key.c_str(), key.size()); | |
81 | uint32_t sid2 = sid ^ ((sid & 0xFF) << 24); | |
82 | return rgw_shards_mod(sid2, num_shards); | |
31f18b77 FG |
83 | } |
84 | ||
85 | static inline int rgw_shards_max() | |
86 | { | |
87 | return RGW_SHARDS_PRIME_1; | |
88 | } | |
7c673cae FG |
89 | |
90 | static inline void prepend_bucket_marker(const rgw_bucket& bucket, const string& orig_oid, string& oid) | |
91 | { | |
92 | if (bucket.marker.empty() || orig_oid.empty()) { | |
93 | oid = orig_oid; | |
94 | } else { | |
95 | oid = bucket.marker; | |
96 | oid.append("_"); | |
97 | oid.append(orig_oid); | |
98 | } | |
99 | } | |
100 | ||
101 | static inline void get_obj_bucket_and_oid_loc(const rgw_obj& obj, string& oid, string& locator) | |
102 | { | |
103 | const rgw_bucket& bucket = obj.bucket; | |
104 | prepend_bucket_marker(bucket, obj.get_oid(), oid); | |
105 | const string& loc = obj.key.get_loc(); | |
106 | if (!loc.empty()) { | |
107 | prepend_bucket_marker(bucket, loc, locator); | |
108 | } else { | |
109 | locator.clear(); | |
110 | } | |
111 | } | |
112 | ||
113 | int rgw_init_ioctx(librados::Rados *rados, const rgw_pool& pool, librados::IoCtx& ioctx, bool create = false); | |
114 | ||
115 | int rgw_policy_from_attrset(CephContext *cct, map<string, bufferlist>& attrset, RGWAccessControlPolicy *policy); | |
116 | ||
117 | static inline bool rgw_raw_obj_to_obj(const rgw_bucket& bucket, const rgw_raw_obj& raw_obj, rgw_obj *obj) | |
118 | { | |
119 | ssize_t pos = raw_obj.oid.find('_'); | |
120 | if (pos < 0) { | |
121 | return false; | |
122 | } | |
123 | ||
124 | if (!rgw_obj_key::parse_raw_oid(raw_obj.oid.substr(pos + 1), &obj->key)) { | |
125 | return false; | |
126 | } | |
127 | obj->bucket = bucket; | |
128 | ||
129 | return true; | |
130 | } | |
131 | ||
132 | struct rgw_bucket_placement { | |
133 | string placement_rule; | |
134 | rgw_bucket bucket; | |
135 | ||
136 | void dump(Formatter *f) const; | |
137 | }; | |
138 | ||
139 | class rgw_obj_select { | |
140 | string placement_rule; | |
141 | rgw_obj obj; | |
142 | rgw_raw_obj raw_obj; | |
143 | bool is_raw; | |
144 | ||
145 | public: | |
146 | rgw_obj_select() : is_raw(false) {} | |
147 | rgw_obj_select(const rgw_obj& _obj) : obj(_obj), is_raw(false) {} | |
148 | rgw_obj_select(const rgw_raw_obj& _raw_obj) : raw_obj(_raw_obj), is_raw(true) {} | |
149 | rgw_obj_select(const rgw_obj_select& rhs) { | |
c07f9fc5 | 150 | placement_rule = rhs.placement_rule; |
7c673cae FG |
151 | is_raw = rhs.is_raw; |
152 | if (is_raw) { | |
153 | raw_obj = rhs.raw_obj; | |
154 | } else { | |
155 | obj = rhs.obj; | |
156 | } | |
157 | } | |
158 | ||
159 | rgw_raw_obj get_raw_obj(const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params) const; | |
160 | rgw_raw_obj get_raw_obj(RGWRados *store) const; | |
161 | ||
162 | rgw_obj_select& operator=(const rgw_obj& rhs) { | |
163 | obj = rhs; | |
164 | is_raw = false; | |
165 | return *this; | |
166 | } | |
167 | ||
168 | rgw_obj_select& operator=(const rgw_raw_obj& rhs) { | |
169 | raw_obj = rhs; | |
170 | is_raw = true; | |
171 | return *this; | |
172 | } | |
173 | ||
174 | void set_placement_rule(const string& rule) { | |
175 | placement_rule = rule; | |
176 | } | |
177 | }; | |
178 | ||
179 | struct compression_block { | |
180 | uint64_t old_ofs; | |
181 | uint64_t new_ofs; | |
182 | uint64_t len; | |
183 | ||
184 | void encode(bufferlist& bl) const { | |
185 | ENCODE_START(1, 1, bl); | |
186 | ::encode(old_ofs, bl); | |
187 | ::encode(new_ofs, bl); | |
188 | ::encode(len, bl); | |
189 | ENCODE_FINISH(bl); | |
190 | } | |
191 | ||
192 | void decode(bufferlist::iterator& bl) { | |
193 | DECODE_START(1, bl); | |
194 | ::decode(old_ofs, bl); | |
195 | ::decode(new_ofs, bl); | |
196 | ::decode(len, bl); | |
197 | DECODE_FINISH(bl); | |
198 | } | |
199 | }; | |
200 | WRITE_CLASS_ENCODER(compression_block) | |
201 | ||
202 | struct RGWCompressionInfo { | |
203 | string compression_type; | |
204 | uint64_t orig_size; | |
205 | vector<compression_block> blocks; | |
206 | ||
207 | RGWCompressionInfo() : compression_type("none"), orig_size(0) {} | |
208 | RGWCompressionInfo(const RGWCompressionInfo& cs_info) : compression_type(cs_info.compression_type), | |
209 | orig_size(cs_info.orig_size), | |
210 | blocks(cs_info.blocks) {} | |
211 | ||
212 | void encode(bufferlist& bl) const { | |
213 | ENCODE_START(1, 1, bl); | |
214 | ::encode(compression_type, bl); | |
215 | ::encode(orig_size, bl); | |
216 | ::encode(blocks, bl); | |
217 | ENCODE_FINISH(bl); | |
218 | } | |
219 | ||
220 | void decode(bufferlist::iterator& bl) { | |
221 | DECODE_START(1, bl); | |
222 | ::decode(compression_type, bl); | |
223 | ::decode(orig_size, bl); | |
224 | ::decode(blocks, bl); | |
225 | DECODE_FINISH(bl); | |
226 | } | |
227 | }; | |
228 | WRITE_CLASS_ENCODER(RGWCompressionInfo) | |
229 | ||
230 | int rgw_compression_info_from_attrset(map<string, bufferlist>& attrs, bool& need_decompress, RGWCompressionInfo& cs_info); | |
231 | ||
232 | struct RGWOLHInfo { | |
233 | rgw_obj target; | |
234 | bool removed; | |
235 | ||
236 | RGWOLHInfo() : removed(false) {} | |
237 | ||
238 | void encode(bufferlist& bl) const { | |
239 | ENCODE_START(1, 1, bl); | |
240 | ::encode(target, bl); | |
241 | ::encode(removed, bl); | |
242 | ENCODE_FINISH(bl); | |
243 | } | |
244 | ||
245 | void decode(bufferlist::iterator& bl) { | |
246 | DECODE_START(1, bl); | |
247 | ::decode(target, bl); | |
248 | ::decode(removed, bl); | |
249 | DECODE_FINISH(bl); | |
250 | } | |
251 | static void generate_test_instances(list<RGWOLHInfo*>& o); | |
252 | void dump(Formatter *f) const; | |
253 | }; | |
254 | WRITE_CLASS_ENCODER(RGWOLHInfo) | |
255 | ||
256 | struct RGWOLHPendingInfo { | |
257 | ceph::real_time time; | |
258 | ||
259 | RGWOLHPendingInfo() {} | |
260 | ||
261 | void encode(bufferlist& bl) const { | |
262 | ENCODE_START(1, 1, bl); | |
263 | ::encode(time, bl); | |
264 | ENCODE_FINISH(bl); | |
265 | } | |
266 | ||
267 | void decode(bufferlist::iterator& bl) { | |
268 | DECODE_START(1, bl); | |
269 | ::decode(time, bl); | |
270 | DECODE_FINISH(bl); | |
271 | } | |
272 | ||
273 | void dump(Formatter *f) const; | |
274 | }; | |
275 | WRITE_CLASS_ENCODER(RGWOLHPendingInfo) | |
276 | ||
277 | struct RGWUsageBatch { | |
278 | map<ceph::real_time, rgw_usage_log_entry> m; | |
279 | ||
280 | void insert(ceph::real_time& t, rgw_usage_log_entry& entry, bool *account) { | |
281 | bool exists = m.find(t) != m.end(); | |
282 | *account = !exists; | |
283 | m[t].aggregate(entry); | |
284 | } | |
285 | }; | |
286 | ||
287 | struct RGWUsageIter { | |
288 | string read_iter; | |
289 | uint32_t index; | |
290 | ||
291 | RGWUsageIter() : index(0) {} | |
292 | }; | |
293 | ||
294 | class RGWGetDataCB { | |
295 | protected: | |
296 | uint64_t extra_data_len; | |
297 | public: | |
298 | virtual int handle_data(bufferlist& bl, off_t bl_ofs, off_t bl_len) = 0; | |
299 | RGWGetDataCB() : extra_data_len(0) {} | |
300 | virtual ~RGWGetDataCB() {} | |
301 | virtual void set_extra_data_len(uint64_t len) { | |
302 | extra_data_len = len; | |
303 | } | |
304 | /** | |
305 | * Flushes any cached data. Used by RGWGetObjFilter. | |
306 | * Return logic same as handle_data. | |
307 | */ | |
308 | virtual int flush() { | |
309 | return 0; | |
310 | } | |
311 | /** | |
312 | * Allows to extend fetch range of RGW object. Used by RGWGetObjFilter. | |
313 | */ | |
314 | virtual int fixup_range(off_t& bl_ofs, off_t& bl_end) { | |
315 | return 0; | |
316 | } | |
317 | }; | |
318 | ||
319 | class RGWAccessListFilter { | |
320 | public: | |
321 | virtual ~RGWAccessListFilter() {} | |
322 | virtual bool filter(string& name, string& key) = 0; | |
323 | }; | |
324 | ||
325 | struct RGWCloneRangeInfo { | |
326 | rgw_obj src; | |
327 | off_t src_ofs; | |
328 | off_t dst_ofs; | |
329 | uint64_t len; | |
330 | }; | |
331 | ||
332 | struct RGWObjManifestPart { | |
333 | rgw_obj loc; /* the object where the data is located */ | |
334 | uint64_t loc_ofs; /* the offset at that object where the data is located */ | |
335 | uint64_t size; /* the part size */ | |
336 | ||
337 | RGWObjManifestPart() : loc_ofs(0), size(0) {} | |
338 | ||
339 | void encode(bufferlist& bl) const { | |
340 | ENCODE_START(2, 2, bl); | |
341 | ::encode(loc, bl); | |
342 | ::encode(loc_ofs, bl); | |
343 | ::encode(size, bl); | |
344 | ENCODE_FINISH(bl); | |
345 | } | |
346 | ||
347 | void decode(bufferlist::iterator& bl) { | |
348 | DECODE_START_LEGACY_COMPAT_LEN_32(2, 2, 2, bl); | |
349 | ::decode(loc, bl); | |
350 | ::decode(loc_ofs, bl); | |
351 | ::decode(size, bl); | |
352 | DECODE_FINISH(bl); | |
353 | } | |
354 | ||
355 | void dump(Formatter *f) const; | |
356 | static void generate_test_instances(list<RGWObjManifestPart*>& o); | |
357 | }; | |
358 | WRITE_CLASS_ENCODER(RGWObjManifestPart) | |
359 | ||
360 | /* | |
361 | The manifest defines a set of rules for structuring the object parts. | |
362 | There are a few terms to note: | |
363 | - head: the head part of the object, which is the part that contains | |
364 | the first chunk of data. An object might not have a head (as in the | |
365 | case of multipart-part objects). | |
366 | - stripe: data portion of a single rgw object that resides on a single | |
367 | rados object. | |
368 | - part: a collection of stripes that make a contiguous part of an | |
369 | object. A regular object will only have one part (although might have | |
370 | many stripes), a multipart object might have many parts. Each part | |
371 | has a fixed stripe size, although the last stripe of a part might | |
372 | be smaller than that. Consecutive parts may be merged if their stripe | |
373 | value is the same. | |
374 | */ | |
375 | ||
376 | struct RGWObjManifestRule { | |
377 | uint32_t start_part_num; | |
378 | uint64_t start_ofs; | |
379 | uint64_t part_size; /* each part size, 0 if there's no part size, meaning it's unlimited */ | |
380 | uint64_t stripe_max_size; /* underlying obj max size */ | |
381 | string override_prefix; | |
382 | ||
383 | RGWObjManifestRule() : start_part_num(0), start_ofs(0), part_size(0), stripe_max_size(0) {} | |
384 | RGWObjManifestRule(uint32_t _start_part_num, uint64_t _start_ofs, uint64_t _part_size, uint64_t _stripe_max_size) : | |
385 | start_part_num(_start_part_num), start_ofs(_start_ofs), part_size(_part_size), stripe_max_size(_stripe_max_size) {} | |
386 | ||
387 | void encode(bufferlist& bl) const { | |
388 | ENCODE_START(2, 1, bl); | |
389 | ::encode(start_part_num, bl); | |
390 | ::encode(start_ofs, bl); | |
391 | ::encode(part_size, bl); | |
392 | ::encode(stripe_max_size, bl); | |
393 | ::encode(override_prefix, bl); | |
394 | ENCODE_FINISH(bl); | |
395 | } | |
396 | ||
397 | void decode(bufferlist::iterator& bl) { | |
398 | DECODE_START(2, bl); | |
399 | ::decode(start_part_num, bl); | |
400 | ::decode(start_ofs, bl); | |
401 | ::decode(part_size, bl); | |
402 | ::decode(stripe_max_size, bl); | |
403 | if (struct_v >= 2) | |
404 | ::decode(override_prefix, bl); | |
405 | DECODE_FINISH(bl); | |
406 | } | |
407 | void dump(Formatter *f) const; | |
408 | }; | |
409 | WRITE_CLASS_ENCODER(RGWObjManifestRule) | |
410 | ||
411 | class RGWObjManifest { | |
412 | protected: | |
413 | bool explicit_objs; /* old manifest? */ | |
414 | map<uint64_t, RGWObjManifestPart> objs; | |
415 | ||
416 | uint64_t obj_size; | |
417 | ||
418 | rgw_obj obj; | |
419 | uint64_t head_size; | |
420 | string head_placement_rule; | |
421 | ||
422 | uint64_t max_head_size; | |
423 | string prefix; | |
424 | rgw_bucket_placement tail_placement; /* might be different than the original bucket, | |
425 | as object might have been copied across pools */ | |
426 | map<uint64_t, RGWObjManifestRule> rules; | |
427 | ||
428 | string tail_instance; /* tail object's instance */ | |
429 | ||
430 | void convert_to_explicit(const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params); | |
431 | int append_explicit(RGWObjManifest& m, const RGWZoneGroup& zonegroup, const RGWZoneParams& zone_params); | |
432 | void append_rules(RGWObjManifest& m, map<uint64_t, RGWObjManifestRule>::iterator& iter, string *override_prefix); | |
433 | ||
434 | void update_iterators() { | |
435 | begin_iter.seek(0); | |
436 | end_iter.seek(obj_size); | |
437 | } | |
438 | public: | |
439 | ||
440 | RGWObjManifest() : explicit_objs(false), obj_size(0), head_size(0), max_head_size(0), | |
441 | begin_iter(this), end_iter(this) {} | |
442 | RGWObjManifest(const RGWObjManifest& rhs) { | |
443 | *this = rhs; | |
444 | } | |
445 | RGWObjManifest& operator=(const RGWObjManifest& rhs) { | |
446 | explicit_objs = rhs.explicit_objs; | |
447 | objs = rhs.objs; | |
448 | obj_size = rhs.obj_size; | |
449 | obj = rhs.obj; | |
450 | head_size = rhs.head_size; | |
451 | max_head_size = rhs.max_head_size; | |
452 | prefix = rhs.prefix; | |
453 | tail_placement = rhs.tail_placement; | |
454 | rules = rhs.rules; | |
455 | tail_instance = rhs.tail_instance; | |
456 | ||
457 | begin_iter.set_manifest(this); | |
458 | end_iter.set_manifest(this); | |
459 | ||
460 | begin_iter.seek(rhs.begin_iter.get_ofs()); | |
461 | end_iter.seek(rhs.end_iter.get_ofs()); | |
462 | ||
463 | return *this; | |
464 | } | |
465 | ||
466 | map<uint64_t, RGWObjManifestPart>& get_explicit_objs() { | |
467 | return objs; | |
468 | } | |
469 | ||
470 | ||
471 | void set_explicit(uint64_t _size, map<uint64_t, RGWObjManifestPart>& _objs) { | |
472 | explicit_objs = true; | |
473 | obj_size = _size; | |
474 | objs.swap(_objs); | |
475 | } | |
476 | ||
477 | void get_implicit_location(uint64_t cur_part_id, uint64_t cur_stripe, uint64_t ofs, string *override_prefix, rgw_obj_select *location); | |
478 | ||
479 | void set_trivial_rule(uint64_t tail_ofs, uint64_t stripe_max_size) { | |
480 | RGWObjManifestRule rule(0, tail_ofs, 0, stripe_max_size); | |
481 | rules[0] = rule; | |
482 | max_head_size = tail_ofs; | |
483 | } | |
484 | ||
485 | void set_multipart_part_rule(uint64_t stripe_max_size, uint64_t part_num) { | |
486 | RGWObjManifestRule rule(0, 0, 0, stripe_max_size); | |
487 | rule.start_part_num = part_num; | |
488 | rules[0] = rule; | |
489 | max_head_size = 0; | |
490 | } | |
491 | ||
492 | void encode(bufferlist& bl) const { | |
493 | ENCODE_START(7, 6, bl); | |
494 | ::encode(obj_size, bl); | |
495 | ::encode(objs, bl); | |
496 | ::encode(explicit_objs, bl); | |
497 | ::encode(obj, bl); | |
498 | ::encode(head_size, bl); | |
499 | ::encode(max_head_size, bl); | |
500 | ::encode(prefix, bl); | |
501 | ::encode(rules, bl); | |
502 | bool encode_tail_bucket = !(tail_placement.bucket == obj.bucket); | |
503 | ::encode(encode_tail_bucket, bl); | |
504 | if (encode_tail_bucket) { | |
505 | ::encode(tail_placement.bucket, bl); | |
506 | } | |
507 | bool encode_tail_instance = (tail_instance != obj.key.instance); | |
508 | ::encode(encode_tail_instance, bl); | |
509 | if (encode_tail_instance) { | |
510 | ::encode(tail_instance, bl); | |
511 | } | |
512 | ::encode(head_placement_rule, bl); | |
513 | ::encode(tail_placement.placement_rule, bl); | |
514 | ENCODE_FINISH(bl); | |
515 | } | |
516 | ||
517 | void decode(bufferlist::iterator& bl) { | |
518 | DECODE_START_LEGACY_COMPAT_LEN_32(7, 2, 2, bl); | |
519 | ::decode(obj_size, bl); | |
520 | ::decode(objs, bl); | |
521 | if (struct_v >= 3) { | |
522 | ::decode(explicit_objs, bl); | |
523 | ::decode(obj, bl); | |
524 | ::decode(head_size, bl); | |
525 | ::decode(max_head_size, bl); | |
526 | ::decode(prefix, bl); | |
527 | ::decode(rules, bl); | |
528 | } else { | |
529 | explicit_objs = true; | |
530 | if (!objs.empty()) { | |
531 | map<uint64_t, RGWObjManifestPart>::iterator iter = objs.begin(); | |
532 | obj = iter->second.loc; | |
533 | head_size = iter->second.size; | |
534 | max_head_size = head_size; | |
535 | } | |
536 | } | |
537 | ||
538 | if (explicit_objs && head_size > 0 && !objs.empty()) { | |
539 | /* patch up manifest due to issue 16435: | |
540 | * the first object in the explicit objs list might not be the one we need to access, use the | |
541 | * head object instead if set. This would happen if we had an old object that was created | |
542 | * when the explicit objs manifest was around, and it got copied. | |
543 | */ | |
544 | rgw_obj& obj_0 = objs[0].loc; | |
545 | if (!obj_0.get_oid().empty() && obj_0.key.ns.empty()) { | |
546 | objs[0].loc = obj; | |
547 | objs[0].size = head_size; | |
548 | } | |
549 | } | |
550 | ||
551 | if (struct_v >= 4) { | |
552 | if (struct_v < 6) { | |
553 | ::decode(tail_placement.bucket, bl); | |
554 | } else { | |
555 | bool need_to_decode; | |
556 | ::decode(need_to_decode, bl); | |
557 | if (need_to_decode) { | |
558 | ::decode(tail_placement.bucket, bl); | |
559 | } else { | |
560 | tail_placement.bucket = obj.bucket; | |
561 | } | |
562 | } | |
563 | } | |
564 | ||
565 | if (struct_v >= 5) { | |
566 | if (struct_v < 6) { | |
567 | ::decode(tail_instance, bl); | |
568 | } else { | |
569 | bool need_to_decode; | |
570 | ::decode(need_to_decode, bl); | |
571 | if (need_to_decode) { | |
572 | ::decode(tail_instance, bl); | |
573 | } else { | |
574 | tail_instance = obj.key.instance; | |
575 | } | |
576 | } | |
577 | } else { // old object created before 'tail_instance' field added to manifest | |
578 | tail_instance = obj.key.instance; | |
579 | } | |
580 | ||
581 | if (struct_v >= 7) { | |
582 | ::decode(head_placement_rule, bl); | |
583 | ::decode(tail_placement.placement_rule, bl); | |
584 | } | |
585 | ||
586 | update_iterators(); | |
587 | DECODE_FINISH(bl); | |
588 | } | |
589 | ||
590 | void dump(Formatter *f) const; | |
591 | static void generate_test_instances(list<RGWObjManifest*>& o); | |
592 | ||
593 | int append(RGWObjManifest& m, RGWZoneGroup& zonegroup, RGWZoneParams& zone_params); | |
594 | int append(RGWObjManifest& m, RGWRados *store); | |
595 | ||
596 | bool get_rule(uint64_t ofs, RGWObjManifestRule *rule); | |
597 | ||
598 | bool empty() { | |
599 | if (explicit_objs) | |
600 | return objs.empty(); | |
601 | return rules.empty(); | |
602 | } | |
603 | ||
604 | bool has_explicit_objs() { | |
605 | return explicit_objs; | |
606 | } | |
607 | ||
608 | bool has_tail() { | |
609 | if (explicit_objs) { | |
610 | if (objs.size() == 1) { | |
611 | map<uint64_t, RGWObjManifestPart>::iterator iter = objs.begin(); | |
612 | rgw_obj& o = iter->second.loc; | |
613 | return !(obj == o); | |
614 | } | |
615 | return (objs.size() >= 2); | |
616 | } | |
617 | return (obj_size > head_size); | |
618 | } | |
619 | ||
620 | void set_head(const string& placement_rule, const rgw_obj& _o, uint64_t _s) { | |
621 | head_placement_rule = placement_rule; | |
622 | obj = _o; | |
623 | head_size = _s; | |
624 | ||
625 | if (explicit_objs && head_size > 0) { | |
626 | objs[0].loc = obj; | |
627 | objs[0].size = head_size; | |
628 | } | |
629 | } | |
630 | ||
631 | const rgw_obj& get_obj() { | |
632 | return obj; | |
633 | } | |
634 | ||
635 | void set_tail_placement(const string& placement_rule, const rgw_bucket& _b) { | |
636 | tail_placement.placement_rule = placement_rule; | |
637 | tail_placement.bucket = _b; | |
638 | } | |
639 | ||
640 | const rgw_bucket_placement& get_tail_placement() { | |
641 | return tail_placement; | |
642 | } | |
643 | ||
644 | const string& get_head_placement_rule() { | |
645 | return head_placement_rule; | |
646 | } | |
647 | ||
648 | void set_prefix(const string& _p) { | |
649 | prefix = _p; | |
650 | } | |
651 | ||
652 | const string& get_prefix() { | |
653 | return prefix; | |
654 | } | |
655 | ||
656 | void set_tail_instance(const string& _ti) { | |
657 | tail_instance = _ti; | |
658 | } | |
659 | ||
660 | const string& get_tail_instance() { | |
661 | return tail_instance; | |
662 | } | |
663 | ||
664 | void set_head_size(uint64_t _s) { | |
665 | head_size = _s; | |
666 | } | |
667 | ||
668 | void set_obj_size(uint64_t s) { | |
669 | obj_size = s; | |
670 | ||
671 | update_iterators(); | |
672 | } | |
673 | ||
674 | uint64_t get_obj_size() { | |
675 | return obj_size; | |
676 | } | |
677 | ||
678 | uint64_t get_head_size() { | |
679 | return head_size; | |
680 | } | |
681 | ||
682 | void set_max_head_size(uint64_t s) { | |
683 | max_head_size = s; | |
684 | } | |
685 | ||
686 | uint64_t get_max_head_size() { | |
687 | return max_head_size; | |
688 | } | |
689 | ||
690 | class obj_iterator { | |
691 | RGWObjManifest *manifest; | |
692 | uint64_t part_ofs; /* where current part starts */ | |
693 | uint64_t stripe_ofs; /* where current stripe starts */ | |
694 | uint64_t ofs; /* current position within the object */ | |
695 | uint64_t stripe_size; /* current part size */ | |
696 | ||
697 | int cur_part_id; | |
698 | int cur_stripe; | |
699 | string cur_override_prefix; | |
700 | ||
701 | rgw_obj_select location; | |
702 | ||
703 | map<uint64_t, RGWObjManifestRule>::iterator rule_iter; | |
704 | map<uint64_t, RGWObjManifestRule>::iterator next_rule_iter; | |
705 | ||
706 | map<uint64_t, RGWObjManifestPart>::iterator explicit_iter; | |
707 | ||
708 | void init() { | |
709 | part_ofs = 0; | |
710 | stripe_ofs = 0; | |
711 | ofs = 0; | |
712 | stripe_size = 0; | |
713 | cur_part_id = 0; | |
714 | cur_stripe = 0; | |
715 | } | |
716 | ||
717 | void update_explicit_pos(); | |
718 | ||
719 | ||
720 | protected: | |
721 | ||
722 | void set_manifest(RGWObjManifest *m) { | |
723 | manifest = m; | |
724 | } | |
725 | ||
726 | public: | |
727 | obj_iterator() : manifest(NULL) { | |
728 | init(); | |
729 | } | |
730 | explicit obj_iterator(RGWObjManifest *_m) : manifest(_m) { | |
731 | init(); | |
732 | if (!manifest->empty()) { | |
733 | seek(0); | |
734 | } | |
735 | } | |
736 | obj_iterator(RGWObjManifest *_m, uint64_t _ofs) : manifest(_m) { | |
737 | init(); | |
738 | if (!manifest->empty()) { | |
739 | seek(_ofs); | |
740 | } | |
741 | } | |
742 | void seek(uint64_t ofs); | |
743 | ||
744 | void operator++(); | |
745 | bool operator==(const obj_iterator& rhs) { | |
746 | return (ofs == rhs.ofs); | |
747 | } | |
748 | bool operator!=(const obj_iterator& rhs) { | |
749 | return (ofs != rhs.ofs); | |
750 | } | |
751 | const rgw_obj_select& get_location() { | |
752 | return location; | |
753 | } | |
754 | ||
755 | /* start of current stripe */ | |
756 | uint64_t get_stripe_ofs() { | |
757 | if (manifest->explicit_objs) { | |
758 | return explicit_iter->first; | |
759 | } | |
760 | return stripe_ofs; | |
761 | } | |
762 | ||
763 | /* current ofs relative to start of rgw object */ | |
764 | uint64_t get_ofs() const { | |
765 | return ofs; | |
766 | } | |
767 | ||
768 | /* stripe number */ | |
769 | int get_cur_stripe() const { | |
770 | return cur_stripe; | |
771 | } | |
772 | ||
773 | /* current stripe size */ | |
774 | uint64_t get_stripe_size() { | |
775 | if (manifest->explicit_objs) { | |
776 | return explicit_iter->second.size; | |
777 | } | |
778 | return stripe_size; | |
779 | } | |
780 | ||
781 | /* offset where data starts within current stripe */ | |
782 | uint64_t location_ofs() { | |
783 | if (manifest->explicit_objs) { | |
784 | return explicit_iter->second.loc_ofs; | |
785 | } | |
786 | return 0; /* all stripes start at zero offset */ | |
787 | } | |
788 | ||
789 | void update_location(); | |
790 | ||
791 | friend class RGWObjManifest; | |
792 | }; | |
793 | ||
794 | const obj_iterator& obj_begin(); | |
795 | const obj_iterator& obj_end(); | |
796 | obj_iterator obj_find(uint64_t ofs); | |
797 | ||
798 | obj_iterator begin_iter; | |
799 | obj_iterator end_iter; | |
800 | ||
801 | /* | |
802 | * simple object generator. Using a simple single rule manifest. | |
803 | */ | |
804 | class generator { | |
805 | RGWObjManifest *manifest; | |
806 | uint64_t last_ofs; | |
807 | uint64_t cur_part_ofs; | |
808 | int cur_part_id; | |
809 | int cur_stripe; | |
810 | uint64_t cur_stripe_size; | |
811 | string cur_oid; | |
812 | ||
813 | string oid_prefix; | |
814 | ||
815 | rgw_obj_select cur_obj; | |
7c673cae FG |
816 | |
817 | RGWObjManifestRule rule; | |
818 | ||
819 | public: | |
820 | generator() : manifest(NULL), last_ofs(0), cur_part_ofs(0), cur_part_id(0), | |
821 | cur_stripe(0), cur_stripe_size(0) {} | |
822 | int create_begin(CephContext *cct, RGWObjManifest *manifest, const string& placement_rule, rgw_bucket& bucket, rgw_obj& obj); | |
823 | ||
824 | int create_next(uint64_t ofs); | |
825 | ||
826 | rgw_raw_obj get_cur_obj(RGWZoneGroup& zonegroup, RGWZoneParams& zone_params) { return cur_obj.get_raw_obj(zonegroup, zone_params); } | |
827 | rgw_raw_obj get_cur_obj(RGWRados *store) { return cur_obj.get_raw_obj(store); } | |
828 | ||
829 | /* total max size of current stripe (including head obj) */ | |
830 | uint64_t cur_stripe_max_size() { | |
831 | return cur_stripe_size; | |
832 | } | |
833 | }; | |
834 | }; | |
835 | WRITE_CLASS_ENCODER(RGWObjManifest) | |
836 | ||
837 | struct RGWUploadPartInfo { | |
838 | uint32_t num; | |
839 | uint64_t size; | |
840 | uint64_t accounted_size{0}; | |
841 | string etag; | |
842 | ceph::real_time modified; | |
843 | RGWObjManifest manifest; | |
844 | RGWCompressionInfo cs_info; | |
845 | ||
846 | RGWUploadPartInfo() : num(0), size(0) {} | |
847 | ||
848 | void encode(bufferlist& bl) const { | |
849 | ENCODE_START(4, 2, bl); | |
850 | ::encode(num, bl); | |
851 | ::encode(size, bl); | |
852 | ::encode(etag, bl); | |
853 | ::encode(modified, bl); | |
854 | ::encode(manifest, bl); | |
855 | ::encode(cs_info, bl); | |
856 | ::encode(accounted_size, bl); | |
857 | ENCODE_FINISH(bl); | |
858 | } | |
859 | void decode(bufferlist::iterator& bl) { | |
860 | DECODE_START_LEGACY_COMPAT_LEN(4, 2, 2, bl); | |
861 | ::decode(num, bl); | |
862 | ::decode(size, bl); | |
863 | ::decode(etag, bl); | |
864 | ::decode(modified, bl); | |
865 | if (struct_v >= 3) | |
866 | ::decode(manifest, bl); | |
867 | if (struct_v >= 4) { | |
868 | ::decode(cs_info, bl); | |
869 | ::decode(accounted_size, bl); | |
870 | } else { | |
871 | accounted_size = size; | |
872 | } | |
873 | DECODE_FINISH(bl); | |
874 | } | |
875 | void dump(Formatter *f) const; | |
876 | static void generate_test_instances(list<RGWUploadPartInfo*>& o); | |
877 | }; | |
878 | WRITE_CLASS_ENCODER(RGWUploadPartInfo) | |
879 | ||
880 | struct RGWObjState { | |
881 | rgw_obj obj; | |
882 | bool is_atomic; | |
883 | bool has_attrs; | |
884 | bool exists; | |
885 | uint64_t size; //< size of raw object | |
886 | uint64_t accounted_size{0}; //< size before compression, encryption | |
887 | ceph::real_time mtime; | |
888 | uint64_t epoch; | |
889 | bufferlist obj_tag; | |
181888fb | 890 | bufferlist tail_tag; |
7c673cae FG |
891 | string write_tag; |
892 | bool fake_tag; | |
893 | RGWObjManifest manifest; | |
894 | bool has_manifest; | |
895 | string shadow_obj; | |
896 | bool has_data; | |
897 | bufferlist data; | |
898 | bool prefetch_data; | |
899 | bool keep_tail; | |
900 | bool is_olh; | |
901 | bufferlist olh_tag; | |
902 | uint64_t pg_ver; | |
903 | uint32_t zone_short_id; | |
904 | ||
905 | /* important! don't forget to update copy constructor */ | |
906 | ||
907 | RGWObjVersionTracker objv_tracker; | |
908 | ||
909 | map<string, bufferlist> attrset; | |
910 | RGWObjState() : is_atomic(false), has_attrs(0), exists(false), | |
911 | size(0), epoch(0), fake_tag(false), has_manifest(false), | |
912 | has_data(false), prefetch_data(false), keep_tail(false), is_olh(false), | |
913 | pg_ver(0), zone_short_id(0) {} | |
914 | RGWObjState(const RGWObjState& rhs) : obj (rhs.obj) { | |
915 | is_atomic = rhs.is_atomic; | |
916 | has_attrs = rhs.has_attrs; | |
917 | exists = rhs.exists; | |
918 | size = rhs.size; | |
919 | accounted_size = rhs.accounted_size; | |
920 | mtime = rhs.mtime; | |
921 | epoch = rhs.epoch; | |
922 | if (rhs.obj_tag.length()) { | |
923 | obj_tag = rhs.obj_tag; | |
924 | } | |
181888fb FG |
925 | if (rhs.tail_tag.length()) { |
926 | tail_tag = rhs.tail_tag; | |
927 | } | |
7c673cae FG |
928 | write_tag = rhs.write_tag; |
929 | fake_tag = rhs.fake_tag; | |
930 | if (rhs.has_manifest) { | |
931 | manifest = rhs.manifest; | |
932 | } | |
933 | has_manifest = rhs.has_manifest; | |
934 | shadow_obj = rhs.shadow_obj; | |
935 | has_data = rhs.has_data; | |
936 | if (rhs.data.length()) { | |
937 | data = rhs.data; | |
938 | } | |
939 | prefetch_data = rhs.prefetch_data; | |
940 | keep_tail = rhs.keep_tail; | |
941 | is_olh = rhs.is_olh; | |
942 | objv_tracker = rhs.objv_tracker; | |
943 | pg_ver = rhs.pg_ver; | |
944 | } | |
945 | ||
946 | bool get_attr(string name, bufferlist& dest) { | |
947 | map<string, bufferlist>::iterator iter = attrset.find(name); | |
948 | if (iter != attrset.end()) { | |
949 | dest = iter->second; | |
950 | return true; | |
951 | } | |
952 | return false; | |
953 | } | |
954 | }; | |
955 | ||
956 | struct RGWRawObjState { | |
957 | rgw_raw_obj obj; | |
958 | bool has_attrs{false}; | |
959 | bool exists{false}; | |
960 | uint64_t size{0}; | |
961 | ceph::real_time mtime; | |
962 | uint64_t epoch; | |
963 | bufferlist obj_tag; | |
964 | bool has_data{false}; | |
965 | bufferlist data; | |
966 | bool prefetch_data{false}; | |
967 | uint64_t pg_ver{0}; | |
968 | ||
969 | /* important! don't forget to update copy constructor */ | |
970 | ||
971 | RGWObjVersionTracker objv_tracker; | |
972 | ||
973 | map<string, bufferlist> attrset; | |
974 | RGWRawObjState() {} | |
975 | RGWRawObjState(const RGWRawObjState& rhs) : obj (rhs.obj) { | |
976 | has_attrs = rhs.has_attrs; | |
977 | exists = rhs.exists; | |
978 | size = rhs.size; | |
979 | mtime = rhs.mtime; | |
980 | epoch = rhs.epoch; | |
981 | if (rhs.obj_tag.length()) { | |
982 | obj_tag = rhs.obj_tag; | |
983 | } | |
984 | has_data = rhs.has_data; | |
985 | if (rhs.data.length()) { | |
986 | data = rhs.data; | |
987 | } | |
988 | prefetch_data = rhs.prefetch_data; | |
989 | pg_ver = rhs.pg_ver; | |
990 | objv_tracker = rhs.objv_tracker; | |
991 | } | |
992 | }; | |
993 | ||
994 | struct RGWPoolIterCtx { | |
995 | librados::IoCtx io_ctx; | |
996 | librados::NObjectIterator iter; | |
997 | }; | |
998 | ||
999 | struct RGWListRawObjsCtx { | |
1000 | bool initialized; | |
1001 | RGWPoolIterCtx iter_ctx; | |
1002 | ||
1003 | RGWListRawObjsCtx() : initialized(false) {} | |
1004 | }; | |
1005 | ||
1006 | struct RGWDefaultSystemMetaObjInfo { | |
1007 | string default_id; | |
1008 | ||
1009 | void encode(bufferlist& bl) const { | |
1010 | ENCODE_START(1, 1, bl); | |
1011 | ::encode(default_id, bl); | |
1012 | ENCODE_FINISH(bl); | |
1013 | } | |
1014 | ||
1015 | void decode(bufferlist::iterator& bl) { | |
1016 | DECODE_START(1, bl); | |
1017 | ::decode(default_id, bl); | |
1018 | DECODE_FINISH(bl); | |
1019 | } | |
1020 | ||
1021 | void dump(Formatter *f) const; | |
1022 | void decode_json(JSONObj *obj); | |
1023 | }; | |
1024 | WRITE_CLASS_ENCODER(RGWDefaultSystemMetaObjInfo) | |
1025 | ||
1026 | struct RGWNameToId { | |
1027 | string obj_id; | |
1028 | ||
1029 | void encode(bufferlist& bl) const { | |
1030 | ENCODE_START(1, 1, bl); | |
1031 | ::encode(obj_id, bl); | |
1032 | ENCODE_FINISH(bl); | |
1033 | } | |
1034 | ||
1035 | void decode(bufferlist::iterator& bl) { | |
1036 | DECODE_START(1, bl); | |
1037 | ::decode(obj_id, bl); | |
1038 | DECODE_FINISH(bl); | |
1039 | } | |
1040 | ||
1041 | void dump(Formatter *f) const; | |
1042 | void decode_json(JSONObj *obj); | |
1043 | }; | |
1044 | WRITE_CLASS_ENCODER(RGWNameToId) | |
1045 | ||
1046 | class RGWSystemMetaObj { | |
1047 | protected: | |
1048 | string id; | |
1049 | string name; | |
1050 | ||
1051 | CephContext *cct; | |
1052 | RGWRados *store; | |
1053 | ||
1054 | int store_name(bool exclusive); | |
1055 | int store_info(bool exclusive); | |
1056 | int read_info(const string& obj_id, bool old_format = false); | |
1057 | int read_id(const string& obj_name, string& obj_id); | |
1058 | int read_default(RGWDefaultSystemMetaObjInfo& default_info, | |
1059 | const string& oid); | |
1060 | /* read and use default id */ | |
1061 | int use_default(bool old_format = false); | |
1062 | ||
1063 | public: | |
1064 | RGWSystemMetaObj() : cct(NULL), store(NULL) {} | |
1065 | RGWSystemMetaObj(const string& _name): name(_name), cct(NULL), store(NULL) {} | |
1066 | RGWSystemMetaObj(const string& _id, const string& _name) : id(_id), name(_name), cct(NULL), store(NULL) {} | |
1067 | RGWSystemMetaObj(CephContext *_cct, RGWRados *_store): cct(_cct), store(_store){} | |
1068 | RGWSystemMetaObj(const string& _name, CephContext *_cct, RGWRados *_store): name(_name), cct(_cct), store(_store){} | |
1069 | const string& get_name() const { return name; } | |
1070 | const string& get_id() const { return id; } | |
1071 | ||
1072 | void set_name(const string& _name) { name = _name;} | |
1073 | void set_id(const string& _id) { id = _id;} | |
1074 | void clear_id() { id.clear(); } | |
1075 | ||
1076 | virtual ~RGWSystemMetaObj() {} | |
1077 | ||
1078 | virtual void encode(bufferlist& bl) const { | |
1079 | ENCODE_START(1, 1, bl); | |
1080 | ::encode(id, bl); | |
1081 | ::encode(name, bl); | |
1082 | ENCODE_FINISH(bl); | |
1083 | } | |
1084 | ||
1085 | virtual void decode(bufferlist::iterator& bl) { | |
1086 | DECODE_START(1, bl); | |
1087 | ::decode(id, bl); | |
1088 | ::decode(name, bl); | |
1089 | DECODE_FINISH(bl); | |
1090 | } | |
1091 | ||
1092 | void reinit_instance(CephContext *_cct, RGWRados *_store) { | |
1093 | cct = _cct; | |
1094 | store = _store; | |
1095 | } | |
1096 | int init(CephContext *_cct, RGWRados *_store, bool setup_obj = true, bool old_format = false); | |
1097 | virtual int read_default_id(string& default_id, bool old_format = false); | |
1098 | virtual int set_as_default(bool exclusive = false); | |
1099 | int delete_default(); | |
1100 | virtual int create(bool exclusive = true); | |
1101 | int delete_obj(bool old_format = false); | |
1102 | int rename(const string& new_name); | |
1103 | int update() { return store_info(false);} | |
1104 | int update_name() { return store_name(false);} | |
1105 | int read(); | |
1106 | int write(bool exclusive); | |
1107 | ||
1108 | virtual rgw_pool get_pool(CephContext *cct) = 0; | |
1109 | virtual const string get_default_oid(bool old_format = false) = 0; | |
1110 | virtual const string& get_names_oid_prefix() = 0; | |
1111 | virtual const string& get_info_oid_prefix(bool old_format = false) = 0; | |
1112 | virtual const string& get_predefined_name(CephContext *cct) = 0; | |
1113 | ||
1114 | void dump(Formatter *f) const; | |
1115 | void decode_json(JSONObj *obj); | |
1116 | }; | |
1117 | WRITE_CLASS_ENCODER(RGWSystemMetaObj) | |
1118 | ||
1119 | struct RGWZonePlacementInfo { | |
1120 | rgw_pool index_pool; | |
1121 | rgw_pool data_pool; | |
1122 | rgw_pool data_extra_pool; /* if not set we should use data_pool */ | |
1123 | RGWBucketIndexType index_type; | |
1124 | std::string compression_type; | |
1125 | ||
1126 | RGWZonePlacementInfo() : index_type(RGWBIType_Normal) {} | |
1127 | ||
1128 | void encode(bufferlist& bl) const { | |
1129 | ENCODE_START(6, 1, bl); | |
1130 | ::encode(index_pool.to_str(), bl); | |
1131 | ::encode(data_pool.to_str(), bl); | |
1132 | ::encode(data_extra_pool.to_str(), bl); | |
1133 | ::encode((uint32_t)index_type, bl); | |
1134 | ::encode(compression_type, bl); | |
1135 | ENCODE_FINISH(bl); | |
1136 | } | |
1137 | ||
1138 | void decode(bufferlist::iterator& bl) { | |
1139 | DECODE_START(6, bl); | |
1140 | string index_pool_str; | |
1141 | string data_pool_str; | |
1142 | ::decode(index_pool_str, bl); | |
1143 | index_pool = rgw_pool(index_pool_str); | |
1144 | ::decode(data_pool_str, bl); | |
1145 | data_pool = rgw_pool(data_pool_str); | |
1146 | if (struct_v >= 4) { | |
1147 | string data_extra_pool_str; | |
1148 | ::decode(data_extra_pool_str, bl); | |
1149 | data_extra_pool = rgw_pool(data_extra_pool_str); | |
1150 | } | |
1151 | if (struct_v >= 5) { | |
1152 | uint32_t it; | |
1153 | ::decode(it, bl); | |
1154 | index_type = (RGWBucketIndexType)it; | |
1155 | } | |
1156 | if (struct_v >= 6) { | |
1157 | ::decode(compression_type, bl); | |
1158 | } | |
1159 | DECODE_FINISH(bl); | |
1160 | } | |
31f18b77 | 1161 | const rgw_pool& get_data_extra_pool() const { |
7c673cae FG |
1162 | if (data_extra_pool.empty()) { |
1163 | return data_pool; | |
1164 | } | |
1165 | return data_extra_pool; | |
1166 | } | |
1167 | void dump(Formatter *f) const; | |
1168 | void decode_json(JSONObj *obj); | |
1169 | }; | |
1170 | WRITE_CLASS_ENCODER(RGWZonePlacementInfo) | |
1171 | ||
1172 | struct RGWZoneParams : RGWSystemMetaObj { | |
1173 | rgw_pool domain_root; | |
1174 | rgw_pool metadata_heap; | |
1175 | rgw_pool control_pool; | |
1176 | rgw_pool gc_pool; | |
1177 | rgw_pool lc_pool; | |
1178 | rgw_pool log_pool; | |
1179 | rgw_pool intent_log_pool; | |
1180 | rgw_pool usage_log_pool; | |
1181 | ||
1182 | rgw_pool user_keys_pool; | |
1183 | rgw_pool user_email_pool; | |
1184 | rgw_pool user_swift_pool; | |
1185 | rgw_pool user_uid_pool; | |
1186 | rgw_pool roles_pool; | |
31f18b77 | 1187 | rgw_pool reshard_pool; |
7c673cae FG |
1188 | |
1189 | RGWAccessKey system_key; | |
1190 | ||
1191 | map<string, RGWZonePlacementInfo> placement_pools; | |
1192 | ||
1193 | string realm_id; | |
1194 | ||
31f18b77 | 1195 | map<string, string, ltstr_nocase> tier_config; |
7c673cae FG |
1196 | |
1197 | RGWZoneParams() : RGWSystemMetaObj() {} | |
1198 | RGWZoneParams(const string& name) : RGWSystemMetaObj(name){} | |
1199 | RGWZoneParams(const string& id, const string& name) : RGWSystemMetaObj(id, name) {} | |
1200 | RGWZoneParams(const string& id, const string& name, const string& _realm_id) | |
1201 | : RGWSystemMetaObj(id, name), realm_id(_realm_id) {} | |
1202 | ||
1203 | rgw_pool get_pool(CephContext *cct); | |
1204 | const string get_default_oid(bool old_format = false) override; | |
1205 | const string& get_names_oid_prefix() override; | |
1206 | const string& get_info_oid_prefix(bool old_format = false) override; | |
1207 | const string& get_predefined_name(CephContext *cct) override; | |
1208 | ||
1209 | int init(CephContext *_cct, RGWRados *_store, bool setup_obj = true, | |
1210 | bool old_format = false); | |
1211 | using RGWSystemMetaObj::init; | |
1212 | int read_default_id(string& default_id, bool old_format = false) override; | |
1213 | int set_as_default(bool exclusive = false) override; | |
1214 | int create_default(bool old_format = false); | |
1215 | int create(bool exclusive = true) override; | |
1216 | int fix_pool_names(); | |
1217 | ||
1218 | const string& get_compression_type(const string& placement_rule) const; | |
1219 | ||
1220 | void encode(bufferlist& bl) const override { | |
31f18b77 | 1221 | ENCODE_START(10, 1, bl); |
7c673cae FG |
1222 | ::encode(domain_root, bl); |
1223 | ::encode(control_pool, bl); | |
1224 | ::encode(gc_pool, bl); | |
1225 | ::encode(log_pool, bl); | |
1226 | ::encode(intent_log_pool, bl); | |
1227 | ::encode(usage_log_pool, bl); | |
1228 | ::encode(user_keys_pool, bl); | |
1229 | ::encode(user_email_pool, bl); | |
1230 | ::encode(user_swift_pool, bl); | |
1231 | ::encode(user_uid_pool, bl); | |
1232 | RGWSystemMetaObj::encode(bl); | |
1233 | ::encode(system_key, bl); | |
1234 | ::encode(placement_pools, bl); | |
1235 | ::encode(metadata_heap, bl); | |
1236 | ::encode(realm_id, bl); | |
1237 | ::encode(lc_pool, bl); | |
1238 | ::encode(tier_config, bl); | |
1239 | ::encode(roles_pool, bl); | |
31f18b77 | 1240 | ::encode(reshard_pool, bl); |
7c673cae FG |
1241 | ENCODE_FINISH(bl); |
1242 | } | |
1243 | ||
1244 | void decode(bufferlist::iterator& bl) override { | |
31f18b77 | 1245 | DECODE_START(10, bl); |
7c673cae FG |
1246 | ::decode(domain_root, bl); |
1247 | ::decode(control_pool, bl); | |
1248 | ::decode(gc_pool, bl); | |
1249 | ::decode(log_pool, bl); | |
1250 | ::decode(intent_log_pool, bl); | |
1251 | ::decode(usage_log_pool, bl); | |
1252 | ::decode(user_keys_pool, bl); | |
1253 | ::decode(user_email_pool, bl); | |
1254 | ::decode(user_swift_pool, bl); | |
1255 | ::decode(user_uid_pool, bl); | |
1256 | if (struct_v >= 6) { | |
1257 | RGWSystemMetaObj::decode(bl); | |
1258 | } else if (struct_v >= 2) { | |
1259 | ::decode(name, bl); | |
1260 | id = name; | |
1261 | } | |
1262 | if (struct_v >= 3) | |
1263 | ::decode(system_key, bl); | |
1264 | if (struct_v >= 4) | |
1265 | ::decode(placement_pools, bl); | |
1266 | if (struct_v >= 5) | |
1267 | ::decode(metadata_heap, bl); | |
1268 | if (struct_v >= 6) { | |
1269 | ::decode(realm_id, bl); | |
1270 | } | |
1271 | if (struct_v >= 7) { | |
1272 | ::decode(lc_pool, bl); | |
1273 | } else { | |
d2e6a577 | 1274 | lc_pool = log_pool.name + ":lc"; |
7c673cae FG |
1275 | } |
1276 | if (struct_v >= 8) { | |
1277 | ::decode(tier_config, bl); | |
1278 | } | |
1279 | if (struct_v >= 9) { | |
1280 | ::decode(roles_pool, bl); | |
1281 | } else { | |
d2e6a577 | 1282 | roles_pool = name + ".rgw.meta:roles"; |
7c673cae | 1283 | } |
31f18b77 FG |
1284 | if (struct_v >= 10) { |
1285 | ::decode(reshard_pool, bl); | |
1286 | } else { | |
c07f9fc5 | 1287 | reshard_pool = log_pool.name + ":reshard"; |
31f18b77 | 1288 | } |
7c673cae FG |
1289 | DECODE_FINISH(bl); |
1290 | } | |
1291 | void dump(Formatter *f) const; | |
1292 | void decode_json(JSONObj *obj); | |
1293 | static void generate_test_instances(list<RGWZoneParams*>& o); | |
1294 | ||
7c673cae FG |
1295 | bool get_placement(const string& placement_id, RGWZonePlacementInfo *placement) const { |
1296 | auto iter = placement_pools.find(placement_id); | |
1297 | if (iter == placement_pools.end()) { | |
1298 | return false; | |
1299 | } | |
1300 | *placement = iter->second; | |
1301 | return true; | |
1302 | } | |
1303 | ||
1304 | /* | |
1305 | * return data pool of the head object | |
1306 | */ | |
1307 | bool get_head_data_pool(const string& placement_id, const rgw_obj& obj, rgw_pool *pool) const { | |
1308 | const rgw_data_placement_target& explicit_placement = obj.bucket.explicit_placement; | |
1309 | if (!explicit_placement.data_pool.empty()) { | |
1310 | if (!obj.in_extra_data) { | |
1311 | *pool = explicit_placement.data_pool; | |
1312 | } else { | |
1313 | *pool = explicit_placement.get_data_extra_pool(); | |
1314 | } | |
1315 | return true; | |
1316 | } | |
1317 | if (placement_id.empty()) { | |
1318 | return false; | |
1319 | } | |
1320 | auto iter = placement_pools.find(placement_id); | |
1321 | if (iter == placement_pools.end()) { | |
1322 | return false; | |
1323 | } | |
1324 | if (!obj.in_extra_data) { | |
1325 | *pool = iter->second.data_pool; | |
1326 | } else { | |
31f18b77 | 1327 | *pool = iter->second.get_data_extra_pool(); |
7c673cae FG |
1328 | } |
1329 | return true; | |
1330 | } | |
1331 | }; | |
1332 | WRITE_CLASS_ENCODER(RGWZoneParams) | |
1333 | ||
1334 | struct RGWZone { | |
1335 | string id; | |
1336 | string name; | |
1337 | list<string> endpoints; | |
1338 | bool log_meta; | |
1339 | bool log_data; | |
1340 | bool read_only; | |
1341 | string tier_type; | |
1342 | ||
1343 | /** | |
1344 | * Represents the number of shards for the bucket index object, a value of zero | |
1345 | * indicates there is no sharding. By default (no sharding, the name of the object | |
1346 | * is '.dir.{marker}', with sharding, the name is '.dir.{marker}.{sharding_id}', | |
1347 | * sharding_id is zero-based value. It is not recommended to set a too large value | |
1348 | * (e.g. thousand) as it increases the cost for bucket listing. | |
1349 | */ | |
1350 | uint32_t bucket_index_max_shards; | |
1351 | ||
1352 | bool sync_from_all; | |
1353 | set<string> sync_from; /* list of zones to sync from */ | |
1354 | ||
1355 | RGWZone() : log_meta(false), log_data(false), read_only(false), bucket_index_max_shards(0), | |
1356 | sync_from_all(true) {} | |
1357 | ||
1358 | void encode(bufferlist& bl) const { | |
1359 | ENCODE_START(6, 1, bl); | |
1360 | ::encode(name, bl); | |
1361 | ::encode(endpoints, bl); | |
1362 | ::encode(log_meta, bl); | |
1363 | ::encode(log_data, bl); | |
1364 | ::encode(bucket_index_max_shards, bl); | |
1365 | ::encode(id, bl); | |
1366 | ::encode(read_only, bl); | |
1367 | ::encode(tier_type, bl); | |
1368 | ::encode(sync_from_all, bl); | |
1369 | ::encode(sync_from, bl); | |
1370 | ENCODE_FINISH(bl); | |
1371 | } | |
1372 | ||
1373 | void decode(bufferlist::iterator& bl) { | |
1374 | DECODE_START(6, bl); | |
1375 | ::decode(name, bl); | |
1376 | if (struct_v < 4) { | |
1377 | id = name; | |
1378 | } | |
1379 | ::decode(endpoints, bl); | |
1380 | if (struct_v >= 2) { | |
1381 | ::decode(log_meta, bl); | |
1382 | ::decode(log_data, bl); | |
1383 | } | |
1384 | if (struct_v >= 3) { | |
1385 | ::decode(bucket_index_max_shards, bl); | |
1386 | } | |
1387 | if (struct_v >= 4) { | |
1388 | ::decode(id, bl); | |
1389 | ::decode(read_only, bl); | |
1390 | } | |
1391 | if (struct_v >= 5) { | |
1392 | ::decode(tier_type, bl); | |
1393 | } | |
1394 | if (struct_v >= 6) { | |
1395 | ::decode(sync_from_all, bl); | |
1396 | ::decode(sync_from, bl); | |
1397 | } | |
1398 | DECODE_FINISH(bl); | |
1399 | } | |
1400 | void dump(Formatter *f) const; | |
1401 | void decode_json(JSONObj *obj); | |
1402 | static void generate_test_instances(list<RGWZone*>& o); | |
1403 | ||
1404 | bool is_read_only() { return read_only; } | |
1405 | ||
28e407b8 | 1406 | bool syncs_from(const string& zone_id) const { |
7c673cae FG |
1407 | return (sync_from_all || sync_from.find(zone_id) != sync_from.end()); |
1408 | } | |
1409 | }; | |
1410 | WRITE_CLASS_ENCODER(RGWZone) | |
1411 | ||
1412 | struct RGWDefaultZoneGroupInfo { | |
1413 | string default_zonegroup; | |
1414 | ||
1415 | void encode(bufferlist& bl) const { | |
1416 | ENCODE_START(1, 1, bl); | |
1417 | ::encode(default_zonegroup, bl); | |
1418 | ENCODE_FINISH(bl); | |
1419 | } | |
1420 | ||
1421 | void decode(bufferlist::iterator& bl) { | |
1422 | DECODE_START(1, bl); | |
1423 | ::decode(default_zonegroup, bl); | |
1424 | DECODE_FINISH(bl); | |
1425 | } | |
1426 | void dump(Formatter *f) const; | |
1427 | void decode_json(JSONObj *obj); | |
1428 | //todo: implement ceph-dencoder | |
1429 | }; | |
1430 | WRITE_CLASS_ENCODER(RGWDefaultZoneGroupInfo) | |
1431 | ||
1432 | struct RGWZoneGroupPlacementTarget { | |
1433 | string name; | |
1434 | set<string> tags; | |
1435 | ||
c07f9fc5 | 1436 | bool user_permitted(list<string>& user_tags) const { |
7c673cae FG |
1437 | if (tags.empty()) { |
1438 | return true; | |
1439 | } | |
1440 | for (auto& rule : user_tags) { | |
1441 | if (tags.find(rule) != tags.end()) { | |
1442 | return true; | |
1443 | } | |
1444 | } | |
1445 | return false; | |
1446 | } | |
1447 | ||
1448 | void encode(bufferlist& bl) const { | |
1449 | ENCODE_START(1, 1, bl); | |
1450 | ::encode(name, bl); | |
1451 | ::encode(tags, bl); | |
1452 | ENCODE_FINISH(bl); | |
1453 | } | |
1454 | ||
1455 | void decode(bufferlist::iterator& bl) { | |
1456 | DECODE_START(1, bl); | |
1457 | ::decode(name, bl); | |
1458 | ::decode(tags, bl); | |
1459 | DECODE_FINISH(bl); | |
1460 | } | |
1461 | void dump(Formatter *f) const; | |
1462 | void decode_json(JSONObj *obj); | |
1463 | }; | |
1464 | WRITE_CLASS_ENCODER(RGWZoneGroupPlacementTarget) | |
1465 | ||
1466 | ||
1467 | struct RGWZoneGroup : public RGWSystemMetaObj { | |
1468 | string api_name; | |
1469 | list<string> endpoints; | |
1470 | bool is_master; | |
1471 | ||
1472 | string master_zone; | |
1473 | map<string, RGWZone> zones; | |
1474 | ||
1475 | map<string, RGWZoneGroupPlacementTarget> placement_targets; | |
1476 | string default_placement; | |
1477 | ||
1478 | list<string> hostnames; | |
1479 | list<string> hostnames_s3website; | |
1480 | // TODO: Maybe convert hostnames to a map<string,list<string>> for | |
1481 | // endpoint_type->hostnames | |
1482 | /* | |
1483 | 20:05 < _robbat21irssi> maybe I do someting like: if (hostname_map.empty()) { populate all map keys from hostnames; }; | |
1484 | 20:05 < _robbat21irssi> but that's a later compatability migration planning bit | |
1485 | 20:06 < yehudasa> more like if (!hostnames.empty()) { | |
1486 | 20:06 < yehudasa> for (list<string>::iterator iter = hostnames.begin(); iter != hostnames.end(); ++iter) { | |
1487 | 20:06 < yehudasa> hostname_map["s3"].append(iter->second); | |
1488 | 20:07 < yehudasa> hostname_map["s3website"].append(iter->second); | |
1489 | 20:07 < yehudasa> s/append/push_back/g | |
1490 | 20:08 < _robbat21irssi> inner loop over APIs | |
1491 | 20:08 < yehudasa> yeah, probably | |
1492 | 20:08 < _robbat21irssi> s3, s3website, swift, swith_auth, swift_website | |
1493 | */ | |
1494 | map<string, list<string> > api_hostname_map; | |
1495 | map<string, list<string> > api_endpoints_map; | |
1496 | ||
1497 | string realm_id; | |
1498 | ||
1499 | RGWZoneGroup(): is_master(false){} | |
1500 | RGWZoneGroup(const std::string &id, const std::string &name):RGWSystemMetaObj(id, name) {} | |
1501 | RGWZoneGroup(const std::string &_name):RGWSystemMetaObj(_name) {} | |
1502 | RGWZoneGroup(const std::string &_name, bool _is_master, CephContext *cct, RGWRados* store, | |
1503 | const string& _realm_id, const list<string>& _endpoints) | |
1504 | : RGWSystemMetaObj(_name, cct , store), endpoints(_endpoints), is_master(_is_master), | |
1505 | realm_id(_realm_id) {} | |
1506 | ||
1507 | bool is_master_zonegroup() const { return is_master;} | |
1508 | void update_master(bool _is_master) { | |
1509 | is_master = _is_master; | |
1510 | post_process_params(); | |
1511 | } | |
1512 | void post_process_params(); | |
1513 | ||
1514 | void encode(bufferlist& bl) const override { | |
1515 | ENCODE_START(4, 1, bl); | |
1516 | ::encode(name, bl); | |
1517 | ::encode(api_name, bl); | |
1518 | ::encode(is_master, bl); | |
1519 | ::encode(endpoints, bl); | |
1520 | ::encode(master_zone, bl); | |
1521 | ::encode(zones, bl); | |
1522 | ::encode(placement_targets, bl); | |
1523 | ::encode(default_placement, bl); | |
1524 | ::encode(hostnames, bl); | |
1525 | ::encode(hostnames_s3website, bl); | |
1526 | RGWSystemMetaObj::encode(bl); | |
1527 | ::encode(realm_id, bl); | |
1528 | ENCODE_FINISH(bl); | |
1529 | } | |
1530 | ||
1531 | void decode(bufferlist::iterator& bl) override { | |
1532 | DECODE_START(4, bl); | |
1533 | ::decode(name, bl); | |
1534 | ::decode(api_name, bl); | |
1535 | ::decode(is_master, bl); | |
1536 | ::decode(endpoints, bl); | |
1537 | ::decode(master_zone, bl); | |
1538 | ::decode(zones, bl); | |
1539 | ::decode(placement_targets, bl); | |
1540 | ::decode(default_placement, bl); | |
1541 | if (struct_v >= 2) { | |
1542 | ::decode(hostnames, bl); | |
1543 | } | |
1544 | if (struct_v >= 3) { | |
1545 | ::decode(hostnames_s3website, bl); | |
1546 | } | |
1547 | if (struct_v >= 4) { | |
1548 | RGWSystemMetaObj::decode(bl); | |
1549 | ::decode(realm_id, bl); | |
1550 | } else { | |
1551 | id = name; | |
1552 | } | |
1553 | DECODE_FINISH(bl); | |
1554 | } | |
1555 | ||
1556 | int read_default_id(string& default_id, bool old_format = false) override; | |
1557 | int set_as_default(bool exclusive = false) override; | |
1558 | int create_default(bool old_format = false); | |
1559 | int equals(const string& other_zonegroup) const; | |
1560 | int add_zone(const RGWZoneParams& zone_params, bool *is_master, bool *read_only, | |
1561 | const list<string>& endpoints, const string *ptier_type, | |
1562 | bool *psync_from_all, list<string>& sync_from, list<string>& sync_from_rm); | |
1563 | int remove_zone(const std::string& zone_id); | |
1564 | int rename_zone(const RGWZoneParams& zone_params); | |
1565 | rgw_pool get_pool(CephContext *cct); | |
1566 | const string get_default_oid(bool old_region_format = false) override; | |
1567 | const string& get_info_oid_prefix(bool old_region_format = false) override; | |
1568 | const string& get_names_oid_prefix() override; | |
1569 | const string& get_predefined_name(CephContext *cct) override; | |
1570 | ||
1571 | void dump(Formatter *f) const; | |
1572 | void decode_json(JSONObj *obj); | |
1573 | static void generate_test_instances(list<RGWZoneGroup*>& o); | |
1574 | }; | |
1575 | WRITE_CLASS_ENCODER(RGWZoneGroup) | |
1576 | ||
1577 | struct RGWPeriodMap | |
1578 | { | |
1579 | string id; | |
1580 | map<string, RGWZoneGroup> zonegroups; | |
1581 | map<string, RGWZoneGroup> zonegroups_by_api; | |
1582 | map<string, uint32_t> short_zone_ids; | |
1583 | ||
1584 | string master_zonegroup; | |
1585 | ||
1586 | void encode(bufferlist& bl) const; | |
1587 | void decode(bufferlist::iterator& bl); | |
1588 | ||
1589 | int update(const RGWZoneGroup& zonegroup, CephContext *cct); | |
1590 | ||
1591 | void dump(Formatter *f) const; | |
1592 | void decode_json(JSONObj *obj); | |
1593 | ||
1594 | void reset() { | |
1595 | zonegroups.clear(); | |
1596 | zonegroups_by_api.clear(); | |
1597 | master_zonegroup.clear(); | |
1598 | } | |
1599 | ||
1600 | uint32_t get_zone_short_id(const string& zone_id) const; | |
1601 | }; | |
1602 | WRITE_CLASS_ENCODER(RGWPeriodMap) | |
1603 | ||
1604 | struct RGWPeriodConfig | |
1605 | { | |
1606 | RGWQuotaInfo bucket_quota; | |
1607 | RGWQuotaInfo user_quota; | |
1608 | ||
1609 | void encode(bufferlist& bl) const { | |
1610 | ENCODE_START(1, 1, bl); | |
1611 | ::encode(bucket_quota, bl); | |
1612 | ::encode(user_quota, bl); | |
1613 | ENCODE_FINISH(bl); | |
1614 | } | |
1615 | ||
1616 | void decode(bufferlist::iterator& bl) { | |
1617 | DECODE_START(1, bl); | |
1618 | ::decode(bucket_quota, bl); | |
1619 | ::decode(user_quota, bl); | |
1620 | DECODE_FINISH(bl); | |
1621 | } | |
1622 | ||
1623 | void dump(Formatter *f) const; | |
1624 | void decode_json(JSONObj *obj); | |
1625 | ||
1626 | // the period config must be stored in a local object outside of the period, | |
1627 | // so that it can be used in a default configuration where no realm/period | |
1628 | // exists | |
1629 | int read(RGWRados *store, const std::string& realm_id); | |
1630 | int write(RGWRados *store, const std::string& realm_id); | |
1631 | ||
1632 | static std::string get_oid(const std::string& realm_id); | |
1633 | static rgw_pool get_pool(CephContext *cct); | |
1634 | }; | |
1635 | WRITE_CLASS_ENCODER(RGWPeriodConfig) | |
1636 | ||
1637 | /* for backward comaptability */ | |
1638 | struct RGWRegionMap { | |
1639 | ||
1640 | map<string, RGWZoneGroup> regions; | |
1641 | ||
1642 | string master_region; | |
1643 | ||
1644 | RGWQuotaInfo bucket_quota; | |
1645 | RGWQuotaInfo user_quota; | |
1646 | ||
1647 | void encode(bufferlist& bl) const; | |
1648 | void decode(bufferlist::iterator& bl); | |
1649 | ||
1650 | void dump(Formatter *f) const; | |
1651 | void decode_json(JSONObj *obj); | |
1652 | }; | |
1653 | WRITE_CLASS_ENCODER(RGWRegionMap) | |
1654 | ||
1655 | struct RGWZoneGroupMap { | |
1656 | ||
1657 | map<string, RGWZoneGroup> zonegroups; | |
1658 | map<string, RGWZoneGroup> zonegroups_by_api; | |
1659 | ||
1660 | string master_zonegroup; | |
1661 | ||
1662 | RGWQuotaInfo bucket_quota; | |
1663 | RGWQuotaInfo user_quota; | |
1664 | ||
1665 | /* constract the map */ | |
1666 | int read(CephContext *cct, RGWRados *store); | |
1667 | ||
1668 | void encode(bufferlist& bl) const; | |
1669 | void decode(bufferlist::iterator& bl); | |
1670 | ||
1671 | void dump(Formatter *f) const; | |
1672 | void decode_json(JSONObj *obj); | |
1673 | }; | |
1674 | WRITE_CLASS_ENCODER(RGWZoneGroupMap) | |
1675 | ||
1676 | class RGWRealm; | |
1677 | ||
1678 | struct objexp_hint_entry { | |
1679 | string tenant; | |
1680 | string bucket_name; | |
1681 | string bucket_id; | |
1682 | rgw_obj_key obj_key; | |
1683 | ceph::real_time exp_time; | |
1684 | ||
1685 | void encode(bufferlist& bl) const { | |
1686 | ENCODE_START(2, 1, bl); | |
1687 | ::encode(bucket_name, bl); | |
1688 | ::encode(bucket_id, bl); | |
1689 | ::encode(obj_key, bl); | |
1690 | ::encode(exp_time, bl); | |
1691 | ::encode(tenant, bl); | |
1692 | ENCODE_FINISH(bl); | |
1693 | } | |
1694 | ||
1695 | void decode(bufferlist::iterator& bl) { | |
1696 | // XXX Do we want DECODE_START_LEGACY_COMPAT_LEN(2, 1, 1, bl); ? | |
1697 | DECODE_START(2, bl); | |
1698 | ::decode(bucket_name, bl); | |
1699 | ::decode(bucket_id, bl); | |
1700 | ::decode(obj_key, bl); | |
1701 | ::decode(exp_time, bl); | |
1702 | if (struct_v >= 2) { | |
1703 | ::decode(tenant, bl); | |
1704 | } else { | |
1705 | tenant.clear(); | |
1706 | } | |
1707 | DECODE_FINISH(bl); | |
1708 | } | |
1709 | }; | |
1710 | WRITE_CLASS_ENCODER(objexp_hint_entry) | |
1711 | ||
1712 | class RGWPeriod; | |
1713 | ||
1714 | class RGWRealm : public RGWSystemMetaObj | |
1715 | { | |
1716 | string current_period; | |
1717 | epoch_t epoch{0}; //< realm epoch, incremented for each new period | |
1718 | ||
1719 | int create_control(bool exclusive); | |
1720 | int delete_control(); | |
1721 | public: | |
1722 | RGWRealm() {} | |
1723 | RGWRealm(const string& _id, const string& _name = "") : RGWSystemMetaObj(_id, _name) {} | |
1724 | RGWRealm(CephContext *_cct, RGWRados *_store): RGWSystemMetaObj(_cct, _store) {} | |
1725 | RGWRealm(const string& _name, CephContext *_cct, RGWRados *_store): RGWSystemMetaObj(_name, _cct, _store){} | |
1726 | ||
1727 | void encode(bufferlist& bl) const override { | |
1728 | ENCODE_START(1, 1, bl); | |
1729 | RGWSystemMetaObj::encode(bl); | |
1730 | ::encode(current_period, bl); | |
1731 | ::encode(epoch, bl); | |
1732 | ENCODE_FINISH(bl); | |
1733 | } | |
1734 | ||
1735 | void decode(bufferlist::iterator& bl) override { | |
1736 | DECODE_START(1, bl); | |
1737 | RGWSystemMetaObj::decode(bl); | |
1738 | ::decode(current_period, bl); | |
1739 | ::decode(epoch, bl); | |
1740 | DECODE_FINISH(bl); | |
1741 | } | |
1742 | ||
1743 | int create(bool exclusive = true) override; | |
1744 | int delete_obj(); | |
1745 | rgw_pool get_pool(CephContext *cct); | |
1746 | const string get_default_oid(bool old_format = false) override; | |
1747 | const string& get_names_oid_prefix() override; | |
1748 | const string& get_info_oid_prefix(bool old_format = false) override; | |
1749 | const string& get_predefined_name(CephContext *cct) override; | |
1750 | ||
1751 | using RGWSystemMetaObj::read_id; // expose as public for radosgw-admin | |
1752 | ||
1753 | void dump(Formatter *f) const; | |
1754 | void decode_json(JSONObj *obj); | |
1755 | ||
1756 | const string& get_current_period() const { | |
1757 | return current_period; | |
1758 | } | |
1759 | int set_current_period(RGWPeriod& period); | |
1760 | void clear_current_period_and_epoch() { | |
1761 | current_period.clear(); | |
1762 | epoch = 0; | |
1763 | } | |
1764 | epoch_t get_epoch() const { return epoch; } | |
1765 | ||
1766 | string get_control_oid(); | |
1767 | /// send a notify on the realm control object | |
1768 | int notify_zone(bufferlist& bl); | |
1769 | /// notify the zone of a new period | |
1770 | int notify_new_period(const RGWPeriod& period); | |
1771 | }; | |
1772 | WRITE_CLASS_ENCODER(RGWRealm) | |
1773 | ||
1774 | struct RGWPeriodLatestEpochInfo { | |
1775 | epoch_t epoch; | |
1776 | ||
1777 | void encode(bufferlist& bl) const { | |
1778 | ENCODE_START(1, 1, bl); | |
1779 | ::encode(epoch, bl); | |
1780 | ENCODE_FINISH(bl); | |
1781 | } | |
1782 | ||
1783 | void decode(bufferlist::iterator& bl) { | |
1784 | DECODE_START(1, bl); | |
1785 | ::decode(epoch, bl); | |
1786 | DECODE_FINISH(bl); | |
1787 | } | |
1788 | ||
1789 | void dump(Formatter *f) const; | |
1790 | void decode_json(JSONObj *obj); | |
1791 | }; | |
1792 | WRITE_CLASS_ENCODER(RGWPeriodLatestEpochInfo) | |
1793 | ||
1794 | class RGWPeriod | |
1795 | { | |
1796 | string id; | |
1797 | epoch_t epoch; | |
1798 | string predecessor_uuid; | |
1799 | std::vector<std::string> sync_status; | |
1800 | RGWPeriodMap period_map; | |
1801 | RGWPeriodConfig period_config; | |
1802 | string master_zonegroup; | |
1803 | string master_zone; | |
1804 | ||
1805 | string realm_id; | |
1806 | string realm_name; | |
1807 | epoch_t realm_epoch{1}; //< realm epoch when period was made current | |
1808 | ||
1809 | CephContext *cct; | |
1810 | RGWRados *store; | |
1811 | ||
1812 | int read_info(); | |
224ce89b WB |
1813 | int read_latest_epoch(RGWPeriodLatestEpochInfo& epoch_info, |
1814 | RGWObjVersionTracker *objv = nullptr); | |
7c673cae FG |
1815 | int use_latest_epoch(); |
1816 | int use_current_period(); | |
1817 | ||
1818 | const string get_period_oid(); | |
1819 | const string get_period_oid_prefix(); | |
1820 | ||
1821 | // gather the metadata sync status for each shard; only for use on master zone | |
1822 | int update_sync_status(const RGWPeriod ¤t_period, | |
1823 | std::ostream& error_stream, bool force_if_stale); | |
1824 | ||
1825 | public: | |
1826 | RGWPeriod() : epoch(0), cct(NULL), store(NULL) {} | |
1827 | ||
1828 | RGWPeriod(const string& period_id, epoch_t _epoch = 0) | |
1829 | : id(period_id), epoch(_epoch), | |
1830 | cct(NULL), store(NULL) {} | |
1831 | ||
1832 | const string& get_id() const { return id; } | |
1833 | epoch_t get_epoch() const { return epoch; } | |
1834 | epoch_t get_realm_epoch() const { return realm_epoch; } | |
1835 | const string& get_predecessor() const { return predecessor_uuid; } | |
1836 | const string& get_master_zone() const { return master_zone; } | |
1837 | const string& get_master_zonegroup() const { return master_zonegroup; } | |
1838 | const string& get_realm() const { return realm_id; } | |
1839 | const RGWPeriodMap& get_map() const { return period_map; } | |
1840 | RGWPeriodConfig& get_config() { return period_config; } | |
1841 | const RGWPeriodConfig& get_config() const { return period_config; } | |
1842 | const std::vector<std::string>& get_sync_status() const { return sync_status; } | |
1843 | rgw_pool get_pool(CephContext *cct); | |
1844 | const string& get_latest_epoch_oid(); | |
1845 | const string& get_info_oid_prefix(); | |
1846 | ||
1847 | void set_user_quota(RGWQuotaInfo& user_quota) { | |
1848 | period_config.user_quota = user_quota; | |
1849 | } | |
1850 | ||
1851 | void set_bucket_quota(RGWQuotaInfo& bucket_quota) { | |
1852 | period_config.bucket_quota = bucket_quota; | |
1853 | } | |
1854 | ||
1855 | void set_id(const string& id) { | |
1856 | this->id = id; | |
1857 | period_map.id = id; | |
1858 | } | |
1859 | void set_epoch(epoch_t epoch) { this->epoch = epoch; } | |
1860 | void set_realm_epoch(epoch_t epoch) { realm_epoch = epoch; } | |
1861 | ||
1862 | void set_predecessor(const string& predecessor) | |
1863 | { | |
1864 | predecessor_uuid = predecessor; | |
1865 | } | |
1866 | ||
1867 | void set_realm_id(const string& _realm_id) { | |
1868 | realm_id = _realm_id; | |
1869 | } | |
1870 | ||
1871 | int reflect(); | |
1872 | ||
1873 | int get_zonegroup(RGWZoneGroup& zonegroup, | |
1874 | const string& zonegroup_id); | |
1875 | ||
3efd9988 | 1876 | bool is_single_zonegroup() const |
224ce89b WB |
1877 | { |
1878 | return (period_map.zonegroups.size() == 1); | |
1879 | } | |
1880 | ||
1881 | /* | |
1882 | returns true if there are several zone groups with a least one zone | |
1883 | */ | |
1884 | bool is_multi_zonegroups_with_zones() | |
1885 | { | |
1886 | int count = 0; | |
1887 | for (const auto& zg: period_map.zonegroups) { | |
1888 | if (zg.second.zones.size() > 0) { | |
1889 | if (count++ > 0) { | |
1890 | return true; | |
1891 | } | |
1892 | } | |
1893 | } | |
1894 | return false; | |
1895 | } | |
7c673cae FG |
1896 | |
1897 | int get_latest_epoch(epoch_t& epoch); | |
224ce89b WB |
1898 | int set_latest_epoch(epoch_t epoch, bool exclusive = false, |
1899 | RGWObjVersionTracker *objv = nullptr); | |
1900 | // update latest_epoch if the given epoch is higher, else return -EEXIST | |
1901 | int update_latest_epoch(epoch_t epoch); | |
7c673cae FG |
1902 | |
1903 | int init(CephContext *_cct, RGWRados *_store, const string &period_realm_id, const string &period_realm_name = "", | |
1904 | bool setup_obj = true); | |
1905 | int init(CephContext *_cct, RGWRados *_store, bool setup_obj = true); | |
7c673cae FG |
1906 | |
1907 | int create(bool exclusive = true); | |
1908 | int delete_obj(); | |
1909 | int store_info(bool exclusive); | |
1910 | int add_zonegroup(const RGWZoneGroup& zonegroup); | |
1911 | ||
1912 | void fork(); | |
1913 | int update(); | |
1914 | ||
1915 | // commit a staging period; only for use on master zone | |
1916 | int commit(RGWRealm& realm, const RGWPeriod ¤t_period, | |
1917 | std::ostream& error_stream, bool force_if_stale = false); | |
1918 | ||
1919 | void encode(bufferlist& bl) const { | |
1920 | ENCODE_START(1, 1, bl); | |
1921 | ::encode(id, bl); | |
1922 | ::encode(epoch, bl); | |
1923 | ::encode(realm_epoch, bl); | |
1924 | ::encode(predecessor_uuid, bl); | |
1925 | ::encode(sync_status, bl); | |
1926 | ::encode(period_map, bl); | |
1927 | ::encode(master_zone, bl); | |
1928 | ::encode(master_zonegroup, bl); | |
1929 | ::encode(period_config, bl); | |
1930 | ::encode(realm_id, bl); | |
1931 | ::encode(realm_name, bl); | |
1932 | ENCODE_FINISH(bl); | |
1933 | } | |
1934 | ||
1935 | void decode(bufferlist::iterator& bl) { | |
1936 | DECODE_START(1, bl); | |
1937 | ::decode(id, bl); | |
1938 | ::decode(epoch, bl); | |
1939 | ::decode(realm_epoch, bl); | |
1940 | ::decode(predecessor_uuid, bl); | |
1941 | ::decode(sync_status, bl); | |
1942 | ::decode(period_map, bl); | |
1943 | ::decode(master_zone, bl); | |
1944 | ::decode(master_zonegroup, bl); | |
1945 | ::decode(period_config, bl); | |
1946 | ::decode(realm_id, bl); | |
1947 | ::decode(realm_name, bl); | |
1948 | DECODE_FINISH(bl); | |
1949 | } | |
1950 | void dump(Formatter *f) const; | |
1951 | void decode_json(JSONObj *obj); | |
1952 | ||
1953 | static string get_staging_id(const string& realm_id) { | |
1954 | return realm_id + ":staging"; | |
1955 | } | |
1956 | }; | |
1957 | WRITE_CLASS_ENCODER(RGWPeriod) | |
1958 | ||
1959 | class RGWDataChangesLog; | |
1960 | class RGWMetaSyncStatusManager; | |
1961 | class RGWDataSyncStatusManager; | |
1962 | class RGWReplicaLogger; | |
1963 | class RGWCoroutinesManagerRegistry; | |
1964 | ||
1965 | class RGWStateLog { | |
1966 | RGWRados *store; | |
1967 | int num_shards; | |
1968 | string module_name; | |
1969 | ||
1970 | void oid_str(int shard, string& oid); | |
1971 | int get_shard_num(const string& object); | |
1972 | string get_oid(const string& object); | |
1973 | int open_ioctx(librados::IoCtx& ioctx); | |
1974 | ||
1975 | struct list_state { | |
1976 | int cur_shard; | |
1977 | int max_shard; | |
1978 | string marker; | |
1979 | string client_id; | |
1980 | string op_id; | |
1981 | string object; | |
1982 | ||
1983 | list_state() : cur_shard(0), max_shard(0) {} | |
1984 | }; | |
1985 | ||
1986 | protected: | |
1987 | virtual bool dump_entry_internal(const cls_statelog_entry& entry, Formatter *f) { | |
1988 | return false; | |
1989 | } | |
1990 | ||
1991 | public: | |
1992 | RGWStateLog(RGWRados *_store, int _num_shards, const string& _module_name) : | |
1993 | store(_store), num_shards(_num_shards), module_name(_module_name) {} | |
1994 | virtual ~RGWStateLog() {} | |
1995 | ||
1996 | int store_entry(const string& client_id, const string& op_id, const string& object, | |
1997 | uint32_t state, bufferlist *bl, uint32_t *check_state); | |
1998 | ||
1999 | int remove_entry(const string& client_id, const string& op_id, const string& object); | |
2000 | ||
2001 | void init_list_entries(const string& client_id, const string& op_id, const string& object, | |
2002 | void **handle); | |
2003 | ||
2004 | int list_entries(void *handle, int max_entries, list<cls_statelog_entry>& entries, bool *done); | |
2005 | ||
2006 | void finish_list_entries(void *handle); | |
2007 | ||
2008 | virtual void dump_entry(const cls_statelog_entry& entry, Formatter *f); | |
2009 | }; | |
2010 | ||
2011 | /* | |
2012 | * state transitions: | |
2013 | * | |
2014 | * unknown -> in-progress -> complete | |
2015 | * -> error | |
2016 | * | |
2017 | * user can try setting the 'abort' state, and it can only succeed if state is | |
2018 | * in-progress. | |
2019 | * | |
2020 | * state renewal cannot switch state (stays in the same state) | |
2021 | * | |
2022 | * rgw can switch from in-progress to complete | |
2023 | * rgw can switch from in-progress to error | |
2024 | * | |
2025 | * rgw can switch from abort to cancelled | |
2026 | * | |
2027 | */ | |
2028 | ||
2029 | class RGWOpState : public RGWStateLog { | |
2030 | protected: | |
2031 | bool dump_entry_internal(const cls_statelog_entry& entry, Formatter *f) override; | |
2032 | public: | |
2033 | ||
2034 | enum OpState { | |
2035 | OPSTATE_UNKNOWN = 0, | |
2036 | OPSTATE_IN_PROGRESS = 1, | |
2037 | OPSTATE_COMPLETE = 2, | |
2038 | OPSTATE_ERROR = 3, | |
2039 | OPSTATE_ABORT = 4, | |
2040 | OPSTATE_CANCELLED = 5, | |
2041 | }; | |
2042 | ||
2043 | explicit RGWOpState(RGWRados *_store); | |
2044 | ||
2045 | int state_from_str(const string& s, OpState *state); | |
2046 | int set_state(const string& client_id, const string& op_id, const string& object, OpState state); | |
2047 | int renew_state(const string& client_id, const string& op_id, const string& object, OpState state); | |
2048 | }; | |
2049 | ||
2050 | class RGWOpStateSingleOp | |
2051 | { | |
2052 | RGWOpState os; | |
2053 | string client_id; | |
2054 | string op_id; | |
2055 | string object; | |
2056 | ||
2057 | CephContext *cct; | |
2058 | ||
2059 | RGWOpState::OpState cur_state; | |
2060 | ceph::real_time last_update; | |
2061 | ||
2062 | public: | |
2063 | RGWOpStateSingleOp(RGWRados *store, const string& cid, const string& oid, const string& obj); | |
2064 | ||
2065 | int set_state(RGWOpState::OpState state); | |
2066 | int renew_state(); | |
2067 | }; | |
2068 | ||
2069 | class RGWGetBucketStats_CB : public RefCountedObject { | |
2070 | protected: | |
2071 | rgw_bucket bucket; | |
2072 | map<RGWObjCategory, RGWStorageStats> *stats; | |
2073 | public: | |
224ce89b | 2074 | explicit RGWGetBucketStats_CB(const rgw_bucket& _bucket) : bucket(_bucket), stats(NULL) {} |
7c673cae FG |
2075 | ~RGWGetBucketStats_CB() override {} |
2076 | virtual void handle_response(int r) = 0; | |
2077 | virtual void set_response(map<RGWObjCategory, RGWStorageStats> *_stats) { | |
2078 | stats = _stats; | |
2079 | } | |
2080 | }; | |
2081 | ||
2082 | class RGWGetUserStats_CB : public RefCountedObject { | |
2083 | protected: | |
2084 | rgw_user user; | |
2085 | RGWStorageStats stats; | |
2086 | public: | |
2087 | explicit RGWGetUserStats_CB(const rgw_user& _user) : user(_user) {} | |
2088 | ~RGWGetUserStats_CB() override {} | |
2089 | virtual void handle_response(int r) = 0; | |
2090 | virtual void set_response(RGWStorageStats& _stats) { | |
2091 | stats = _stats; | |
2092 | } | |
2093 | }; | |
2094 | ||
2095 | class RGWGetDirHeader_CB; | |
2096 | class RGWGetUserHeader_CB; | |
2097 | ||
2098 | struct rgw_rados_ref { | |
2099 | rgw_pool pool; | |
2100 | string oid; | |
2101 | string key; | |
2102 | librados::IoCtx ioctx; | |
2103 | }; | |
2104 | ||
2105 | class RGWChainedCache { | |
2106 | public: | |
2107 | virtual ~RGWChainedCache() {} | |
2108 | virtual void chain_cb(const string& key, void *data) = 0; | |
2109 | virtual void invalidate(const string& key) = 0; | |
2110 | virtual void invalidate_all() = 0; | |
2111 | ||
2112 | struct Entry { | |
2113 | RGWChainedCache *cache; | |
2114 | const string& key; | |
2115 | void *data; | |
2116 | ||
2117 | Entry(RGWChainedCache *_c, const string& _k, void *_d) : cache(_c), key(_k), data(_d) {} | |
2118 | }; | |
2119 | }; | |
2120 | ||
2121 | template <class T, class S> | |
2122 | class RGWObjectCtxImpl { | |
2123 | RGWRados *store; | |
2124 | std::map<T, S> objs_state; | |
2125 | RWLock lock; | |
2126 | ||
2127 | public: | |
2128 | RGWObjectCtxImpl(RGWRados *_store) : store(_store), lock("RGWObjectCtxImpl") {} | |
2129 | ||
2130 | S *get_state(const T& obj) { | |
2131 | S *result; | |
2132 | typename std::map<T, S>::iterator iter; | |
2133 | lock.get_read(); | |
2134 | assert (!obj.empty()); | |
2135 | iter = objs_state.find(obj); | |
2136 | if (iter != objs_state.end()) { | |
2137 | result = &iter->second; | |
2138 | lock.unlock(); | |
2139 | } else { | |
2140 | lock.unlock(); | |
2141 | lock.get_write(); | |
2142 | result = &objs_state[obj]; | |
2143 | lock.unlock(); | |
2144 | } | |
2145 | return result; | |
2146 | } | |
2147 | ||
2148 | void set_atomic(T& obj) { | |
2149 | RWLock::WLocker wl(lock); | |
2150 | assert (!obj.empty()); | |
2151 | objs_state[obj].is_atomic = true; | |
2152 | } | |
2153 | void set_prefetch_data(T& obj) { | |
2154 | RWLock::WLocker wl(lock); | |
2155 | assert (!obj.empty()); | |
2156 | objs_state[obj].prefetch_data = true; | |
2157 | } | |
2158 | void invalidate(T& obj) { | |
2159 | RWLock::WLocker wl(lock); | |
2160 | auto iter = objs_state.find(obj); | |
2161 | if (iter == objs_state.end()) { | |
2162 | return; | |
2163 | } | |
2164 | bool is_atomic = iter->second.is_atomic; | |
2165 | bool prefetch_data = iter->second.prefetch_data; | |
2166 | ||
2167 | objs_state.erase(iter); | |
2168 | ||
2169 | if (is_atomic || prefetch_data) { | |
2170 | auto& s = objs_state[obj]; | |
2171 | s.is_atomic = is_atomic; | |
2172 | s.prefetch_data = prefetch_data; | |
2173 | } | |
2174 | } | |
2175 | }; | |
2176 | ||
2177 | template<> | |
2178 | void RGWObjectCtxImpl<rgw_obj, RGWObjState>::invalidate(rgw_obj& obj); | |
2179 | ||
2180 | template<> | |
2181 | void RGWObjectCtxImpl<rgw_raw_obj, RGWRawObjState>::invalidate(rgw_raw_obj& obj); | |
2182 | ||
2183 | struct RGWObjectCtx { | |
2184 | RGWRados *store; | |
2185 | void *user_ctx; | |
2186 | ||
2187 | RGWObjectCtxImpl<rgw_obj, RGWObjState> obj; | |
2188 | RGWObjectCtxImpl<rgw_raw_obj, RGWRawObjState> raw; | |
2189 | ||
2190 | explicit RGWObjectCtx(RGWRados *_store) : store(_store), user_ctx(NULL), obj(store), raw(store) { } | |
2191 | RGWObjectCtx(RGWRados *_store, void *_user_ctx) : store(_store), user_ctx(_user_ctx), obj(store), raw(store) { } | |
2192 | }; | |
2193 | ||
2194 | class Finisher; | |
2195 | class RGWAsyncRadosProcessor; | |
2196 | ||
2197 | template <class T> | |
2198 | class RGWChainedCacheImpl; | |
2199 | ||
2200 | struct bucket_info_entry { | |
2201 | RGWBucketInfo info; | |
2202 | real_time mtime; | |
2203 | map<string, bufferlist> attrs; | |
2204 | }; | |
2205 | ||
2206 | struct tombstone_entry { | |
2207 | ceph::real_time mtime; | |
2208 | uint32_t zone_short_id; | |
2209 | uint64_t pg_ver; | |
2210 | ||
2211 | tombstone_entry() = default; | |
2212 | tombstone_entry(const RGWObjState& state) | |
2213 | : mtime(state.mtime), zone_short_id(state.zone_short_id), | |
2214 | pg_ver(state.pg_ver) {} | |
2215 | }; | |
2216 | ||
31f18b77 FG |
2217 | class RGWIndexCompletionManager; |
2218 | ||
3a9019d9 | 2219 | class RGWRados : public AdminSocketHook |
7c673cae FG |
2220 | { |
2221 | friend class RGWGC; | |
2222 | friend class RGWMetaNotifier; | |
2223 | friend class RGWDataNotifier; | |
2224 | friend class RGWLC; | |
2225 | friend class RGWObjectExpirer; | |
2226 | friend class RGWMetaSyncProcessorThread; | |
2227 | friend class RGWDataSyncProcessorThread; | |
2228 | friend class RGWStateLog; | |
2229 | friend class RGWReplicaLogger; | |
31f18b77 FG |
2230 | friend class RGWReshard; |
2231 | friend class RGWBucketReshard; | |
f64942e4 | 2232 | friend class RGWBucketReshardLock; |
31f18b77 | 2233 | friend class BucketIndexLockGuard; |
d2e6a577 | 2234 | friend class RGWCompleteMultipart; |
7c673cae | 2235 | |
3a9019d9 FG |
2236 | static const char* admin_commands[4][3]; |
2237 | ||
7c673cae FG |
2238 | /** Open the pool used as root for this gateway */ |
2239 | int open_root_pool_ctx(); | |
2240 | int open_gc_pool_ctx(); | |
2241 | int open_lc_pool_ctx(); | |
2242 | int open_objexp_pool_ctx(); | |
31f18b77 | 2243 | int open_reshard_pool_ctx(); |
7c673cae FG |
2244 | |
2245 | int open_pool_ctx(const rgw_pool& pool, librados::IoCtx& io_ctx); | |
2246 | int open_bucket_index_ctx(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx); | |
2247 | int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, string& bucket_oid); | |
2248 | int open_bucket_index_base(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
2249 | string& bucket_oid_base); | |
2250 | int open_bucket_index_shard(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
2251 | const string& obj_key, string *bucket_obj, int *shard_id); | |
2252 | int open_bucket_index_shard(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
2253 | int shard_id, string *bucket_obj); | |
2254 | int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
2255 | map<int, string>& bucket_objs, int shard_id = -1, map<int, string> *bucket_instance_ids = NULL); | |
2256 | template<typename T> | |
2257 | int open_bucket_index(const RGWBucketInfo& bucket_info, librados::IoCtx& index_ctx, | |
2258 | map<int, string>& oids, map<int, T>& bucket_objs, | |
2259 | int shard_id = -1, map<int, string> *bucket_instance_ids = NULL); | |
2260 | void build_bucket_index_marker(const string& shard_id_str, const string& shard_marker, | |
2261 | string *marker); | |
2262 | ||
2263 | void get_bucket_instance_ids(const RGWBucketInfo& bucket_info, int shard_id, map<int, string> *result); | |
2264 | ||
2265 | std::atomic<int64_t> max_req_id = { 0 }; | |
2266 | Mutex lock; | |
2267 | Mutex watchers_lock; | |
2268 | SafeTimer *timer; | |
2269 | ||
2270 | RGWGC *gc; | |
2271 | RGWLC *lc; | |
2272 | RGWObjectExpirer *obj_expirer; | |
2273 | bool use_gc_thread; | |
2274 | bool use_lc_thread; | |
2275 | bool quota_threads; | |
2276 | bool run_sync_thread; | |
31f18b77 | 2277 | bool run_reshard_thread; |
7c673cae FG |
2278 | |
2279 | RGWAsyncRadosProcessor* async_rados; | |
2280 | ||
2281 | RGWMetaNotifier *meta_notifier; | |
2282 | RGWDataNotifier *data_notifier; | |
2283 | RGWMetaSyncProcessorThread *meta_sync_processor_thread; | |
2284 | map<string, RGWDataSyncProcessorThread *> data_sync_processor_threads; | |
2285 | ||
b32b8144 | 2286 | boost::optional<rgw::BucketTrimManager> bucket_trim; |
7c673cae FG |
2287 | RGWSyncLogTrimThread *sync_log_trimmer{nullptr}; |
2288 | ||
2289 | Mutex meta_sync_thread_lock; | |
2290 | Mutex data_sync_thread_lock; | |
2291 | ||
2292 | int num_watchers; | |
2293 | RGWWatcher **watchers; | |
2294 | std::set<int> watchers_set; | |
2295 | librados::IoCtx root_pool_ctx; // .rgw | |
2296 | librados::IoCtx control_pool_ctx; // .rgw.control | |
2297 | bool watch_initialized; | |
2298 | ||
2299 | friend class RGWWatcher; | |
2300 | ||
2301 | Mutex bucket_id_lock; | |
2302 | ||
2303 | // This field represents the number of bucket index object shards | |
2304 | uint32_t bucket_index_max_shards; | |
2305 | ||
2306 | int get_obj_head_ioctx(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::IoCtx *ioctx); | |
2307 | int get_obj_head_ref(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_rados_ref *ref); | |
224ce89b | 2308 | int get_system_obj_ref(const rgw_raw_obj& obj, rgw_rados_ref *ref); |
7c673cae FG |
2309 | uint64_t max_bucket_id; |
2310 | ||
2311 | int get_olh_target_state(RGWObjectCtx& rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
2312 | RGWObjState *olh_state, RGWObjState **target_state); | |
2313 | int get_system_obj_state_impl(RGWObjectCtx *rctx, rgw_raw_obj& obj, RGWRawObjState **state, RGWObjVersionTracker *objv_tracker); | |
2314 | int get_obj_state_impl(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state, | |
2315 | bool follow_olh, bool assume_noent = false); | |
2316 | int append_atomic_test(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
2317 | librados::ObjectOperation& op, RGWObjState **state); | |
2318 | ||
2319 | int update_placement_map(); | |
2320 | int store_bucket_info(RGWBucketInfo& info, map<string, bufferlist> *pattrs, RGWObjVersionTracker *objv_tracker, bool exclusive); | |
2321 | ||
2322 | void remove_rgw_head_obj(librados::ObjectWriteOperation& op); | |
2323 | void cls_obj_check_prefix_exist(librados::ObjectOperation& op, const string& prefix, bool fail_if_exist); | |
2324 | void cls_obj_check_mtime(librados::ObjectOperation& op, const real_time& mtime, bool high_precision_time, RGWCheckMTimeType type); | |
2325 | protected: | |
2326 | CephContext *cct; | |
2327 | ||
2328 | std::vector<librados::Rados> rados; | |
2329 | uint32_t next_rados_handle; | |
2330 | RWLock handle_lock; | |
2331 | std::map<pthread_t, int> rados_map; | |
2332 | ||
2333 | using RGWChainedCacheImpl_bucket_info_entry = RGWChainedCacheImpl<bucket_info_entry>; | |
2334 | RGWChainedCacheImpl_bucket_info_entry *binfo_cache; | |
2335 | ||
2336 | using tombstone_cache_t = lru_map<rgw_obj, tombstone_entry>; | |
2337 | tombstone_cache_t *obj_tombstone_cache; | |
2338 | ||
2339 | librados::IoCtx gc_pool_ctx; // .rgw.gc | |
2340 | librados::IoCtx lc_pool_ctx; // .rgw.lc | |
2341 | librados::IoCtx objexp_pool_ctx; | |
31f18b77 | 2342 | librados::IoCtx reshard_pool_ctx; |
7c673cae FG |
2343 | |
2344 | bool pools_initialized; | |
2345 | ||
7c673cae FG |
2346 | string trans_id_suffix; |
2347 | ||
2348 | RGWQuotaHandler *quota_handler; | |
2349 | ||
2350 | Finisher *finisher; | |
31f18b77 | 2351 | |
7c673cae FG |
2352 | RGWCoroutinesManagerRegistry *cr_registry; |
2353 | ||
2354 | RGWSyncModulesManager *sync_modules_manager{nullptr}; | |
2355 | RGWSyncModuleInstanceRef sync_module; | |
2356 | bool writeable_zone{false}; | |
2357 | ||
2358 | RGWZoneGroup zonegroup; | |
2359 | RGWZone zone_public_config; /* external zone params, e.g., entrypoints, log flags, etc. */ | |
2360 | RGWZoneParams zone_params; /* internal zone params, e.g., rados pools */ | |
2361 | uint32_t zone_short_id; | |
2362 | ||
2363 | RGWPeriod current_period; | |
31f18b77 FG |
2364 | |
2365 | RGWIndexCompletionManager *index_completion_manager{nullptr}; | |
7c673cae FG |
2366 | public: |
2367 | RGWRados() : lock("rados_timer_lock"), watchers_lock("watchers_lock"), timer(NULL), | |
2368 | gc(NULL), lc(NULL), obj_expirer(NULL), use_gc_thread(false), use_lc_thread(false), quota_threads(false), | |
31f18b77 | 2369 | run_sync_thread(false), run_reshard_thread(false), async_rados(nullptr), meta_notifier(NULL), |
7c673cae FG |
2370 | data_notifier(NULL), meta_sync_processor_thread(NULL), |
2371 | meta_sync_thread_lock("meta_sync_thread_lock"), data_sync_thread_lock("data_sync_thread_lock"), | |
2372 | num_watchers(0), watchers(NULL), | |
2373 | watch_initialized(false), | |
2374 | bucket_id_lock("rados_bucket_id"), | |
2375 | bucket_index_max_shards(0), | |
2376 | max_bucket_id(0), cct(NULL), | |
2377 | next_rados_handle(0), | |
2378 | handle_lock("rados_handle_lock"), | |
2379 | binfo_cache(NULL), obj_tombstone_cache(nullptr), | |
2380 | pools_initialized(false), | |
2381 | quota_handler(NULL), | |
2382 | finisher(NULL), | |
2383 | cr_registry(NULL), | |
2384 | zone_short_id(0), | |
2385 | rest_master_conn(NULL), | |
31f18b77 | 2386 | meta_mgr(NULL), data_log(NULL), reshard(NULL) {} |
7c673cae FG |
2387 | |
2388 | uint64_t get_new_req_id() { | |
2389 | return ++max_req_id; | |
2390 | } | |
2391 | ||
2392 | librados::IoCtx* get_lc_pool_ctx() { | |
2393 | return &lc_pool_ctx; | |
2394 | } | |
2395 | void set_context(CephContext *_cct) { | |
2396 | cct = _cct; | |
2397 | } | |
2398 | ||
2399 | /** | |
2400 | * AmazonS3 errors contain a HostId string, but is an opaque base64 blob; we | |
2401 | * try to be more transparent. This has a wrapper so we can update it when zonegroup/zone are changed. | |
2402 | */ | |
2403 | void init_host_id() { | |
2404 | /* uint64_t needs 16, two '-' separators and a trailing null */ | |
2405 | const string& zone_name = get_zone().name; | |
2406 | const string& zonegroup_name = zonegroup.get_name(); | |
2407 | char charbuf[16 + zone_name.size() + zonegroup_name.size() + 2 + 1]; | |
2408 | snprintf(charbuf, sizeof(charbuf), "%llx-%s-%s", (unsigned long long)instance_id(), zone_name.c_str(), zonegroup_name.c_str()); | |
2409 | string s(charbuf); | |
2410 | host_id = s; | |
2411 | } | |
2412 | ||
2413 | string host_id; | |
2414 | ||
2415 | RGWRealm realm; | |
2416 | ||
2417 | RGWRESTConn *rest_master_conn; | |
2418 | map<string, RGWRESTConn *> zone_conn_map; | |
2419 | map<string, RGWRESTConn *> zone_data_sync_from_map; | |
2420 | map<string, RGWRESTConn *> zone_data_notify_to_map; | |
2421 | map<string, RGWRESTConn *> zonegroup_conn_map; | |
2422 | ||
2423 | map<string, string> zone_id_by_name; | |
2424 | map<string, RGWZone> zone_by_id; | |
2425 | ||
2426 | RGWRESTConn *get_zone_conn_by_id(const string& id) { | |
2427 | auto citer = zone_conn_map.find(id); | |
2428 | if (citer == zone_conn_map.end()) { | |
2429 | return NULL; | |
2430 | } | |
2431 | ||
2432 | return citer->second; | |
2433 | } | |
2434 | ||
2435 | RGWRESTConn *get_zone_conn_by_name(const string& name) { | |
2436 | auto i = zone_id_by_name.find(name); | |
2437 | if (i == zone_id_by_name.end()) { | |
2438 | return NULL; | |
2439 | } | |
2440 | ||
2441 | return get_zone_conn_by_id(i->second); | |
2442 | } | |
2443 | ||
2444 | bool find_zone_id_by_name(const string& name, string *id) { | |
2445 | auto i = zone_id_by_name.find(name); | |
2446 | if (i == zone_id_by_name.end()) { | |
2447 | return false; | |
2448 | } | |
2449 | *id = i->second; | |
2450 | return true; | |
2451 | } | |
2452 | ||
2453 | int get_zonegroup(const string& id, RGWZoneGroup& zonegroup) { | |
2454 | int ret = 0; | |
2455 | if (id == get_zonegroup().get_id()) { | |
2456 | zonegroup = get_zonegroup(); | |
2457 | } else if (!current_period.get_id().empty()) { | |
2458 | ret = current_period.get_zonegroup(zonegroup, id); | |
2459 | } | |
2460 | return ret; | |
2461 | } | |
2462 | ||
2463 | RGWRealm& get_realm() { | |
2464 | return realm; | |
2465 | } | |
2466 | ||
2467 | RGWZoneParams& get_zone_params() { return zone_params; } | |
2468 | RGWZoneGroup& get_zonegroup() { | |
2469 | return zonegroup; | |
2470 | } | |
2471 | RGWZone& get_zone() { | |
2472 | return zone_public_config; | |
2473 | } | |
2474 | ||
2475 | bool zone_is_writeable() { | |
2476 | return writeable_zone && !get_zone().is_read_only(); | |
2477 | } | |
2478 | ||
2479 | uint32_t get_zone_short_id() const { | |
2480 | return zone_short_id; | |
2481 | } | |
2482 | ||
2483 | bool zone_syncs_from(RGWZone& target_zone, RGWZone& source_zone); | |
2484 | ||
2485 | const RGWQuotaInfo& get_bucket_quota() { | |
2486 | return current_period.get_config().bucket_quota; | |
2487 | } | |
2488 | ||
2489 | const RGWQuotaInfo& get_user_quota() { | |
2490 | return current_period.get_config().user_quota; | |
2491 | } | |
2492 | ||
2493 | const string& get_current_period_id() { | |
2494 | return current_period.get_id(); | |
2495 | } | |
31f18b77 FG |
2496 | |
2497 | bool has_zonegroup_api(const std::string& api) const { | |
2498 | if (!current_period.get_id().empty()) { | |
2499 | const auto& zonegroups_by_api = current_period.get_map().zonegroups_by_api; | |
2500 | if (zonegroups_by_api.find(api) != zonegroups_by_api.end()) | |
2501 | return true; | |
2502 | } | |
2503 | return false; | |
2504 | } | |
2505 | ||
7c673cae FG |
2506 | // pulls missing periods for period_history |
2507 | std::unique_ptr<RGWPeriodPuller> period_puller; | |
2508 | // maintains a connected history of periods | |
2509 | std::unique_ptr<RGWPeriodHistory> period_history; | |
2510 | ||
2511 | RGWAsyncRadosProcessor* get_async_rados() const { return async_rados; }; | |
2512 | ||
2513 | RGWMetadataManager *meta_mgr; | |
2514 | ||
2515 | RGWDataChangesLog *data_log; | |
2516 | ||
31f18b77 FG |
2517 | RGWReshard *reshard; |
2518 | std::shared_ptr<RGWReshardWait> reshard_wait; | |
2519 | ||
7c673cae FG |
2520 | virtual ~RGWRados() = default; |
2521 | ||
2522 | tombstone_cache_t *get_tombstone_cache() { | |
2523 | return obj_tombstone_cache; | |
2524 | } | |
2525 | ||
2526 | RGWSyncModulesManager *get_sync_modules_manager() { | |
2527 | return sync_modules_manager; | |
2528 | } | |
2529 | const RGWSyncModuleInstanceRef& get_sync_module() { | |
2530 | return sync_module; | |
2531 | } | |
2532 | ||
2533 | int get_required_alignment(const rgw_pool& pool, uint64_t *alignment); | |
2534 | int get_max_chunk_size(const rgw_pool& pool, uint64_t *max_chunk_size); | |
2535 | int get_max_chunk_size(const string& placement_rule, const rgw_obj& obj, uint64_t *max_chunk_size); | |
2536 | ||
2537 | uint32_t get_max_bucket_shards() { | |
31f18b77 | 2538 | return rgw_shards_max(); |
7c673cae FG |
2539 | } |
2540 | ||
181888fb | 2541 | |
224ce89b | 2542 | int get_raw_obj_ref(const rgw_raw_obj& obj, rgw_rados_ref *ref); |
7c673cae | 2543 | |
181888fb FG |
2544 | int list_raw_objects_init(const rgw_pool& pool, const string& marker, RGWListRawObjsCtx *ctx); |
2545 | int list_raw_objects_next(const string& prefix_filter, int max, | |
2546 | RGWListRawObjsCtx& ctx, list<string>& oids, | |
2547 | bool *is_truncated); | |
7c673cae FG |
2548 | int list_raw_objects(const rgw_pool& pool, const string& prefix_filter, int max, |
2549 | RGWListRawObjsCtx& ctx, list<string>& oids, | |
2550 | bool *is_truncated); | |
181888fb | 2551 | string list_raw_objs_get_cursor(RGWListRawObjsCtx& ctx); |
7c673cae FG |
2552 | |
2553 | int list_raw_prefixed_objs(const rgw_pool& pool, const string& prefix, list<string>& result); | |
2554 | int list_zonegroups(list<string>& zonegroups); | |
2555 | int list_regions(list<string>& regions); | |
2556 | int list_zones(list<string>& zones); | |
2557 | int list_realms(list<string>& realms); | |
2558 | int list_periods(list<string>& periods); | |
2559 | int list_periods(const string& current_period, list<string>& periods); | |
2560 | void tick(); | |
2561 | ||
2562 | CephContext *ctx() { return cct; } | |
2563 | /** do all necessary setup of the storage device */ | |
31f18b77 | 2564 | int initialize(CephContext *_cct, bool _use_gc_thread, bool _use_lc_thread, bool _quota_threads, bool _run_sync_thread, bool _run_reshard_thread) { |
7c673cae FG |
2565 | set_context(_cct); |
2566 | use_gc_thread = _use_gc_thread; | |
2567 | use_lc_thread = _use_lc_thread; | |
2568 | quota_threads = _quota_threads; | |
2569 | run_sync_thread = _run_sync_thread; | |
31f18b77 | 2570 | run_reshard_thread = _run_reshard_thread; |
7c673cae FG |
2571 | return initialize(); |
2572 | } | |
2573 | /** Initialize the RADOS instance and prepare to do other ops */ | |
2574 | virtual int init_rados(); | |
2575 | int init_zg_from_period(bool *initialized); | |
2576 | int init_zg_from_local(bool *creating_defaults); | |
2577 | int init_complete(); | |
2578 | int replace_region_with_zonegroup(); | |
2579 | int convert_regionmap(); | |
2580 | int initialize(); | |
2581 | void finalize(); | |
2582 | ||
224ce89b WB |
2583 | int register_to_service_map(const string& daemon_type, const map<string, string>& meta); |
2584 | ||
7c673cae FG |
2585 | void schedule_context(Context *c); |
2586 | ||
2587 | /** set up a bucket listing. handle is filled in. */ | |
2588 | int list_buckets_init(RGWAccessHandle *handle); | |
2589 | /** | |
2590 | * get the next bucket in the listing. obj is filled in, | |
2591 | * handle is updated. | |
2592 | */ | |
2593 | int list_buckets_next(rgw_bucket_dir_entry& obj, RGWAccessHandle *handle); | |
2594 | ||
2595 | /// list logs | |
2596 | int log_list_init(const string& prefix, RGWAccessHandle *handle); | |
2597 | int log_list_next(RGWAccessHandle handle, string *name); | |
2598 | ||
2599 | /// remove log | |
2600 | int log_remove(const string& name); | |
2601 | ||
2602 | /// show log | |
2603 | int log_show_init(const string& name, RGWAccessHandle *handle); | |
2604 | int log_show_next(RGWAccessHandle handle, rgw_log_entry *entry); | |
2605 | ||
2606 | // log bandwidth info | |
2607 | int log_usage(map<rgw_user_bucket, RGWUsageBatch>& usage_info); | |
2608 | int read_usage(const rgw_user& user, uint64_t start_epoch, uint64_t end_epoch, uint32_t max_entries, | |
2609 | bool *is_truncated, RGWUsageIter& read_iter, map<rgw_user_bucket, rgw_usage_log_entry>& usage); | |
2610 | int trim_usage(rgw_user& user, uint64_t start_epoch, uint64_t end_epoch); | |
2611 | ||
2612 | int create_pool(const rgw_pool& pool); | |
2613 | ||
7c673cae | 2614 | int init_bucket_index(RGWBucketInfo& bucket_info, int num_shards); |
f64942e4 | 2615 | int clean_bucket_index(RGWBucketInfo& bucket_info, int num_shards); |
7c673cae FG |
2616 | int select_bucket_placement(RGWUserInfo& user_info, const string& zonegroup_id, const string& rule, |
2617 | string *pselected_rule_name, RGWZonePlacementInfo *rule_info); | |
2618 | int select_legacy_bucket_placement(RGWZonePlacementInfo *rule_info); | |
2619 | int select_new_bucket_location(RGWUserInfo& user_info, const string& zonegroup_id, const string& rule, | |
2620 | string *pselected_rule_name, RGWZonePlacementInfo *rule_info); | |
2621 | int select_bucket_location_by_rule(const string& location_rule, RGWZonePlacementInfo *rule_info); | |
2622 | void create_bucket_id(string *bucket_id); | |
2623 | ||
2624 | bool get_obj_data_pool(const string& placement_rule, const rgw_obj& obj, rgw_pool *pool); | |
2625 | bool obj_to_raw(const string& placement_rule, const rgw_obj& obj, rgw_raw_obj *raw_obj); | |
2626 | ||
2627 | int create_bucket(RGWUserInfo& owner, rgw_bucket& bucket, | |
2628 | const string& zonegroup_id, | |
2629 | const string& placement_rule, | |
2630 | const string& swift_ver_location, | |
2631 | const RGWQuotaInfo * pquota_info, | |
2632 | map<std::string,bufferlist>& attrs, | |
2633 | RGWBucketInfo& bucket_info, | |
2634 | obj_version *pobjv, | |
2635 | obj_version *pep_objv, | |
2636 | ceph::real_time creation_time, | |
2637 | rgw_bucket *master_bucket, | |
2638 | uint32_t *master_num_shards, | |
2639 | bool exclusive = true); | |
2640 | int add_bucket_placement(const rgw_pool& new_pool); | |
2641 | int remove_bucket_placement(const rgw_pool& new_pool); | |
2642 | int list_placement_set(set<rgw_pool>& names); | |
2643 | int create_pools(vector<rgw_pool>& pools, vector<int>& retcodes); | |
2644 | ||
2645 | RGWCoroutinesManagerRegistry *get_cr_registry() { return cr_registry; } | |
2646 | ||
2647 | class SystemObject { | |
2648 | RGWRados *store; | |
2649 | RGWObjectCtx& ctx; | |
2650 | rgw_raw_obj obj; | |
2651 | ||
2652 | RGWObjState *state; | |
2653 | ||
2654 | protected: | |
2655 | int get_state(RGWRawObjState **pstate, RGWObjVersionTracker *objv_tracker); | |
2656 | ||
2657 | public: | |
2658 | SystemObject(RGWRados *_store, RGWObjectCtx& _ctx, rgw_raw_obj& _obj) : store(_store), ctx(_ctx), obj(_obj), state(NULL) {} | |
2659 | ||
2660 | void invalidate_state(); | |
2661 | ||
2662 | RGWRados *get_store() { return store; } | |
2663 | rgw_raw_obj& get_obj() { return obj; } | |
2664 | RGWObjectCtx& get_ctx() { return ctx; } | |
2665 | ||
2666 | struct Read { | |
2667 | RGWRados::SystemObject *source; | |
2668 | ||
2669 | struct GetObjState { | |
2670 | rgw_rados_ref ref; | |
2671 | bool has_ref{false}; | |
2672 | uint64_t last_ver{0}; | |
2673 | ||
2674 | GetObjState() {} | |
2675 | ||
2676 | int get_ref(RGWRados *store, rgw_raw_obj& obj, rgw_rados_ref **pref); | |
2677 | } state; | |
2678 | ||
2679 | struct StatParams { | |
2680 | ceph::real_time *lastmod; | |
2681 | uint64_t *obj_size; | |
2682 | map<string, bufferlist> *attrs; | |
7c673cae | 2683 | |
31f18b77 | 2684 | StatParams() : lastmod(NULL), obj_size(NULL), attrs(NULL) {} |
7c673cae FG |
2685 | } stat_params; |
2686 | ||
2687 | struct ReadParams { | |
224ce89b | 2688 | rgw_cache_entry_info *cache_info{nullptr}; |
7c673cae FG |
2689 | map<string, bufferlist> *attrs; |
2690 | ||
2691 | ReadParams() : attrs(NULL) {} | |
2692 | } read_params; | |
2693 | ||
2694 | explicit Read(RGWRados::SystemObject *_source) : source(_source) {} | |
2695 | ||
2696 | int stat(RGWObjVersionTracker *objv_tracker); | |
b32b8144 FG |
2697 | int read(int64_t ofs, int64_t end, bufferlist& bl, RGWObjVersionTracker *objv_tracker, |
2698 | boost::optional<obj_version> refresh_version = boost::none); | |
7c673cae FG |
2699 | int get_attr(const char *name, bufferlist& dest); |
2700 | }; | |
2701 | }; | |
2702 | ||
2703 | struct BucketShard { | |
2704 | RGWRados *store; | |
2705 | rgw_bucket bucket; | |
2706 | int shard_id; | |
2707 | librados::IoCtx index_ctx; | |
2708 | string bucket_obj; | |
2709 | ||
2710 | explicit BucketShard(RGWRados *_store) : store(_store), shard_id(-1) {} | |
f64942e4 AA |
2711 | int init(const rgw_bucket& _bucket, const rgw_obj& obj, RGWBucketInfo* out); |
2712 | int init(const rgw_bucket& _bucket, int sid, RGWBucketInfo* out); | |
a8e16298 | 2713 | int init(const RGWBucketInfo& bucket_info, const rgw_obj& obj); |
b32b8144 | 2714 | int init(const RGWBucketInfo& bucket_info, int sid); |
7c673cae FG |
2715 | }; |
2716 | ||
2717 | class Object { | |
2718 | RGWRados *store; | |
2719 | RGWBucketInfo bucket_info; | |
2720 | RGWObjectCtx& ctx; | |
2721 | rgw_obj obj; | |
2722 | ||
2723 | BucketShard bs; | |
2724 | ||
2725 | RGWObjState *state; | |
2726 | ||
2727 | bool versioning_disabled; | |
2728 | ||
2729 | bool bs_initialized; | |
2730 | ||
2731 | protected: | |
2732 | int get_state(RGWObjState **pstate, bool follow_olh, bool assume_noent = false); | |
2733 | void invalidate_state(); | |
2734 | ||
2735 | int prepare_atomic_modification(librados::ObjectWriteOperation& op, bool reset_obj, const string *ptag, | |
181888fb | 2736 | const char *ifmatch, const char *ifnomatch, bool removal_op, bool modify_tail); |
7c673cae FG |
2737 | int complete_atomic_modification(); |
2738 | ||
2739 | public: | |
2740 | Object(RGWRados *_store, const RGWBucketInfo& _bucket_info, RGWObjectCtx& _ctx, const rgw_obj& _obj) : store(_store), bucket_info(_bucket_info), | |
2741 | ctx(_ctx), obj(_obj), bs(store), | |
2742 | state(NULL), versioning_disabled(false), | |
2743 | bs_initialized(false) {} | |
2744 | ||
2745 | RGWRados *get_store() { return store; } | |
2746 | rgw_obj& get_obj() { return obj; } | |
2747 | RGWObjectCtx& get_ctx() { return ctx; } | |
2748 | RGWBucketInfo& get_bucket_info() { return bucket_info; } | |
2749 | int get_manifest(RGWObjManifest **pmanifest); | |
2750 | ||
2751 | int get_bucket_shard(BucketShard **pbs) { | |
2752 | if (!bs_initialized) { | |
f64942e4 AA |
2753 | int r = |
2754 | bs.init(bucket_info.bucket, obj, nullptr /* no RGWBucketInfo */); | |
7c673cae FG |
2755 | if (r < 0) { |
2756 | return r; | |
2757 | } | |
2758 | bs_initialized = true; | |
2759 | } | |
2760 | *pbs = &bs; | |
2761 | return 0; | |
2762 | } | |
2763 | ||
2764 | void set_versioning_disabled(bool status) { | |
2765 | versioning_disabled = status; | |
2766 | } | |
2767 | ||
2768 | bool versioning_enabled() { | |
2769 | return (!versioning_disabled && bucket_info.versioning_enabled()); | |
2770 | } | |
2771 | ||
2772 | struct Read { | |
2773 | RGWRados::Object *source; | |
2774 | ||
2775 | struct GetObjState { | |
2776 | librados::IoCtx io_ctx; | |
2777 | rgw_obj obj; | |
2778 | rgw_raw_obj head_obj; | |
2779 | } state; | |
2780 | ||
2781 | struct ConditionParams { | |
2782 | const ceph::real_time *mod_ptr; | |
2783 | const ceph::real_time *unmod_ptr; | |
2784 | bool high_precision_time; | |
2785 | uint32_t mod_zone_id; | |
2786 | uint64_t mod_pg_ver; | |
2787 | const char *if_match; | |
2788 | const char *if_nomatch; | |
2789 | ||
2790 | ConditionParams() : | |
2791 | mod_ptr(NULL), unmod_ptr(NULL), high_precision_time(false), mod_zone_id(0), mod_pg_ver(0), | |
2792 | if_match(NULL), if_nomatch(NULL) {} | |
2793 | } conds; | |
2794 | ||
2795 | struct Params { | |
2796 | ceph::real_time *lastmod; | |
2797 | uint64_t *obj_size; | |
2798 | map<string, bufferlist> *attrs; | |
7c673cae | 2799 | |
31f18b77 | 2800 | Params() : lastmod(NULL), obj_size(NULL), attrs(NULL) {} |
7c673cae FG |
2801 | } params; |
2802 | ||
2803 | explicit Read(RGWRados::Object *_source) : source(_source) {} | |
2804 | ||
2805 | int prepare(); | |
2806 | static int range_to_ofs(uint64_t obj_size, int64_t &ofs, int64_t &end); | |
2807 | int read(int64_t ofs, int64_t end, bufferlist& bl); | |
2808 | int iterate(int64_t ofs, int64_t end, RGWGetDataCB *cb); | |
2809 | int get_attr(const char *name, bufferlist& dest); | |
2810 | }; | |
2811 | ||
2812 | struct Write { | |
2813 | RGWRados::Object *target; | |
2814 | ||
2815 | struct MetaParams { | |
2816 | ceph::real_time *mtime; | |
2817 | map<std::string, bufferlist>* rmattrs; | |
2818 | const bufferlist *data; | |
2819 | RGWObjManifest *manifest; | |
2820 | const string *ptag; | |
2821 | list<rgw_obj_index_key> *remove_objs; | |
2822 | ceph::real_time set_mtime; | |
2823 | rgw_user owner; | |
2824 | RGWObjCategory category; | |
2825 | int flags; | |
2826 | const char *if_match; | |
2827 | const char *if_nomatch; | |
91327a77 | 2828 | boost::optional<uint64_t> olh_epoch; |
7c673cae FG |
2829 | ceph::real_time delete_at; |
2830 | bool canceled; | |
2831 | const string *user_data; | |
31f18b77 | 2832 | rgw_zone_set *zones_trace; |
181888fb | 2833 | bool modify_tail; |
3efd9988 | 2834 | bool completeMultipart; |
7c673cae FG |
2835 | |
2836 | MetaParams() : mtime(NULL), rmattrs(NULL), data(NULL), manifest(NULL), ptag(NULL), | |
2837 | remove_objs(NULL), category(RGW_OBJ_CATEGORY_MAIN), flags(0), | |
91327a77 | 2838 | if_match(NULL), if_nomatch(NULL), canceled(false), user_data(nullptr), zones_trace(nullptr), |
3efd9988 | 2839 | modify_tail(false), completeMultipart(false) {} |
7c673cae FG |
2840 | } meta; |
2841 | ||
2842 | explicit Write(RGWRados::Object *_target) : target(_target) {} | |
2843 | ||
2844 | int _do_write_meta(uint64_t size, uint64_t accounted_size, | |
2845 | map<std::string, bufferlist>& attrs, | |
181888fb | 2846 | bool modify_tail, bool assume_noent, |
7c673cae FG |
2847 | void *index_op); |
2848 | int write_meta(uint64_t size, uint64_t accounted_size, | |
2849 | map<std::string, bufferlist>& attrs); | |
2850 | int write_data(const char *data, uint64_t ofs, uint64_t len, bool exclusive); | |
2851 | }; | |
2852 | ||
2853 | struct Delete { | |
2854 | RGWRados::Object *target; | |
2855 | ||
2856 | struct DeleteParams { | |
2857 | rgw_user bucket_owner; | |
2858 | int versioning_status; | |
2859 | ACLOwner obj_owner; /* needed for creation of deletion marker */ | |
2860 | uint64_t olh_epoch; | |
2861 | string marker_version_id; | |
2862 | uint32_t bilog_flags; | |
2863 | list<rgw_obj_index_key> *remove_objs; | |
2864 | ceph::real_time expiration_time; | |
2865 | ceph::real_time unmod_since; | |
2866 | ceph::real_time mtime; /* for setting delete marker mtime */ | |
2867 | bool high_precision_time; | |
31f18b77 | 2868 | rgw_zone_set *zones_trace; |
7c673cae | 2869 | |
31f18b77 | 2870 | DeleteParams() : versioning_status(0), olh_epoch(0), bilog_flags(0), remove_objs(NULL), high_precision_time(false), zones_trace(nullptr) {} |
7c673cae FG |
2871 | } params; |
2872 | ||
2873 | struct DeleteResult { | |
2874 | bool delete_marker; | |
2875 | string version_id; | |
2876 | ||
2877 | DeleteResult() : delete_marker(false) {} | |
2878 | } result; | |
2879 | ||
2880 | explicit Delete(RGWRados::Object *_target) : target(_target) {} | |
2881 | ||
2882 | int delete_obj(); | |
2883 | }; | |
2884 | ||
2885 | struct Stat { | |
2886 | RGWRados::Object *source; | |
2887 | ||
2888 | struct Result { | |
2889 | rgw_obj obj; | |
2890 | RGWObjManifest manifest; | |
2891 | bool has_manifest; | |
2892 | uint64_t size; | |
2893 | struct timespec mtime; | |
2894 | map<string, bufferlist> attrs; | |
2895 | ||
2896 | Result() : has_manifest(false), size(0) {} | |
2897 | } result; | |
2898 | ||
2899 | struct State { | |
2900 | librados::IoCtx io_ctx; | |
2901 | librados::AioCompletion *completion; | |
2902 | int ret; | |
2903 | ||
2904 | State() : completion(NULL), ret(0) {} | |
2905 | } state; | |
2906 | ||
2907 | ||
2908 | explicit Stat(RGWRados::Object *_source) : source(_source) {} | |
2909 | ||
2910 | int stat_async(); | |
2911 | int wait(); | |
2912 | int stat(); | |
2913 | private: | |
2914 | int finish(); | |
2915 | }; | |
2916 | }; | |
2917 | ||
2918 | class Bucket { | |
2919 | RGWRados *store; | |
2920 | RGWBucketInfo bucket_info; | |
2921 | rgw_bucket& bucket; | |
2922 | int shard_id; | |
2923 | ||
2924 | public: | |
2925 | Bucket(RGWRados *_store, const RGWBucketInfo& _bucket_info) : store(_store), bucket_info(_bucket_info), bucket(bucket_info.bucket), | |
2926 | shard_id(RGW_NO_SHARD) {} | |
2927 | RGWRados *get_store() { return store; } | |
2928 | rgw_bucket& get_bucket() { return bucket; } | |
2929 | RGWBucketInfo& get_bucket_info() { return bucket_info; } | |
2930 | ||
31f18b77 FG |
2931 | int update_bucket_id(const string& new_bucket_id); |
2932 | ||
7c673cae FG |
2933 | int get_shard_id() { return shard_id; } |
2934 | void set_shard_id(int id) { | |
2935 | shard_id = id; | |
2936 | } | |
2937 | ||
2938 | class UpdateIndex { | |
2939 | RGWRados::Bucket *target; | |
2940 | string optag; | |
2941 | rgw_obj obj; | |
2942 | uint16_t bilog_flags{0}; | |
2943 | BucketShard bs; | |
2944 | bool bs_initialized{false}; | |
2945 | bool blind; | |
2946 | bool prepared{false}; | |
31f18b77 FG |
2947 | rgw_zone_set *zones_trace{nullptr}; |
2948 | ||
2949 | int init_bs() { | |
f64942e4 AA |
2950 | int r = |
2951 | bs.init(target->get_bucket(), obj, nullptr /* no RGWBucketInfo */); | |
31f18b77 FG |
2952 | if (r < 0) { |
2953 | return r; | |
2954 | } | |
2955 | bs_initialized = true; | |
2956 | return 0; | |
2957 | } | |
2958 | ||
2959 | void invalidate_bs() { | |
2960 | bs_initialized = false; | |
2961 | } | |
2962 | ||
2963 | int guard_reshard(BucketShard **pbs, std::function<int(BucketShard *)> call); | |
7c673cae FG |
2964 | public: |
2965 | ||
2966 | UpdateIndex(RGWRados::Bucket *_target, const rgw_obj& _obj) : target(_target), obj(_obj), | |
2967 | bs(target->get_store()) { | |
2968 | blind = (target->get_bucket_info().index_type == RGWBIType_Indexless); | |
2969 | } | |
2970 | ||
2971 | int get_bucket_shard(BucketShard **pbs) { | |
2972 | if (!bs_initialized) { | |
31f18b77 | 2973 | int r = init_bs(); |
7c673cae FG |
2974 | if (r < 0) { |
2975 | return r; | |
2976 | } | |
7c673cae FG |
2977 | } |
2978 | *pbs = &bs; | |
2979 | return 0; | |
2980 | } | |
2981 | ||
2982 | void set_bilog_flags(uint16_t flags) { | |
2983 | bilog_flags = flags; | |
2984 | } | |
31f18b77 FG |
2985 | |
2986 | void set_zones_trace(rgw_zone_set *_zones_trace) { | |
2987 | zones_trace = _zones_trace; | |
2988 | } | |
7c673cae FG |
2989 | |
2990 | int prepare(RGWModifyOp, const string *write_tag); | |
2991 | int complete(int64_t poolid, uint64_t epoch, uint64_t size, | |
2992 | uint64_t accounted_size, ceph::real_time& ut, | |
2993 | const string& etag, const string& content_type, | |
2994 | bufferlist *acl_bl, RGWObjCategory category, | |
2995 | list<rgw_obj_index_key> *remove_objs, const string *user_data = nullptr); | |
2996 | int complete_del(int64_t poolid, uint64_t epoch, | |
2997 | ceph::real_time& removed_mtime, /* mtime of removed object */ | |
2998 | list<rgw_obj_index_key> *remove_objs); | |
2999 | int cancel(); | |
3000 | ||
3001 | const string *get_optag() { return &optag; } | |
3002 | ||
3003 | bool is_prepared() { return prepared; } | |
1adf2230 AA |
3004 | }; // class UpdateIndex |
3005 | ||
3006 | class List { | |
3007 | protected: | |
7c673cae | 3008 | |
7c673cae FG |
3009 | RGWRados::Bucket *target; |
3010 | rgw_obj_key next_marker; | |
3011 | ||
1adf2230 AA |
3012 | int list_objects_ordered(int64_t max, |
3013 | vector<rgw_bucket_dir_entry> *result, | |
3014 | map<string, bool> *common_prefixes, | |
3015 | bool *is_truncated); | |
3016 | int list_objects_unordered(int64_t max, | |
3017 | vector<rgw_bucket_dir_entry> *result, | |
3018 | map<string, bool> *common_prefixes, | |
3019 | bool *is_truncated); | |
3020 | ||
3021 | public: | |
3022 | ||
7c673cae FG |
3023 | struct Params { |
3024 | string prefix; | |
3025 | string delim; | |
3026 | rgw_obj_key marker; | |
3027 | rgw_obj_key end_marker; | |
3028 | string ns; | |
3029 | bool enforce_ns; | |
3030 | RGWAccessListFilter *filter; | |
3031 | bool list_versions; | |
1adf2230 AA |
3032 | bool allow_unordered; |
3033 | ||
3034 | Params() : | |
3035 | enforce_ns(true), | |
3036 | filter(NULL), | |
3037 | list_versions(false), | |
3038 | allow_unordered(false) | |
3039 | {} | |
7c673cae FG |
3040 | } params; |
3041 | ||
7c673cae FG |
3042 | explicit List(RGWRados::Bucket *_target) : target(_target) {} |
3043 | ||
1adf2230 AA |
3044 | int list_objects(int64_t max, |
3045 | vector<rgw_bucket_dir_entry> *result, | |
3046 | map<string, bool> *common_prefixes, | |
3047 | bool *is_truncated) { | |
3048 | if (params.allow_unordered) { | |
3049 | return list_objects_unordered(max, result, common_prefixes, | |
3050 | is_truncated); | |
3051 | } else { | |
3052 | return list_objects_ordered(max, result, common_prefixes, | |
3053 | is_truncated); | |
3054 | } | |
3055 | } | |
7c673cae FG |
3056 | rgw_obj_key& get_next_marker() { |
3057 | return next_marker; | |
3058 | } | |
1adf2230 AA |
3059 | }; // class List |
3060 | }; // class Bucket | |
7c673cae FG |
3061 | |
3062 | /** Write/overwrite an object to the bucket storage. */ | |
3063 | virtual int put_system_obj_impl(rgw_raw_obj& obj, uint64_t size, ceph::real_time *mtime, | |
3064 | map<std::string, bufferlist>& attrs, int flags, | |
3065 | bufferlist& data, | |
3066 | RGWObjVersionTracker *objv_tracker, | |
3067 | ceph::real_time set_mtime /* 0 for don't set */); | |
3068 | ||
3069 | virtual int put_system_obj_data(void *ctx, rgw_raw_obj& obj, bufferlist& bl, | |
3070 | off_t ofs, bool exclusive, | |
3071 | RGWObjVersionTracker *objv_tracker = nullptr); | |
3072 | int aio_put_obj_data(void *ctx, rgw_raw_obj& obj, bufferlist& bl, | |
3073 | off_t ofs, bool exclusive, void **handle); | |
3074 | ||
3075 | int put_system_obj(void *ctx, rgw_raw_obj& obj, const char *data, size_t len, bool exclusive, | |
3076 | ceph::real_time *mtime, map<std::string, bufferlist>& attrs, RGWObjVersionTracker *objv_tracker, | |
3077 | ceph::real_time set_mtime) { | |
3078 | bufferlist bl; | |
3079 | bl.append(data, len); | |
3080 | int flags = PUT_OBJ_CREATE; | |
3081 | if (exclusive) | |
3082 | flags |= PUT_OBJ_EXCL; | |
3083 | ||
3084 | return put_system_obj_impl(obj, len, mtime, attrs, flags, bl, objv_tracker, set_mtime); | |
3085 | } | |
3086 | int aio_wait(void *handle); | |
3087 | bool aio_completed(void *handle); | |
3088 | ||
3089 | int on_last_entry_in_listing(RGWBucketInfo& bucket_info, | |
3090 | const std::string& obj_prefix, | |
3091 | const std::string& obj_delim, | |
3092 | std::function<int(const rgw_bucket_dir_entry&)> handler); | |
3093 | ||
3094 | bool swift_versioning_enabled(const RGWBucketInfo& bucket_info) const { | |
3095 | return bucket_info.has_swift_versioning() && | |
3096 | bucket_info.swift_ver_location.size(); | |
3097 | } | |
3098 | ||
3099 | int swift_versioning_copy(RGWObjectCtx& obj_ctx, /* in/out */ | |
3100 | const rgw_user& user, /* in */ | |
3101 | RGWBucketInfo& bucket_info, /* in */ | |
3102 | rgw_obj& obj); /* in */ | |
3103 | int swift_versioning_restore(RGWObjectCtx& obj_ctx, /* in/out */ | |
3104 | const rgw_user& user, /* in */ | |
3105 | RGWBucketInfo& bucket_info, /* in */ | |
3106 | rgw_obj& obj, /* in */ | |
3107 | bool& restored); /* out */ | |
3108 | int copy_obj_to_remote_dest(RGWObjState *astate, | |
3109 | map<string, bufferlist>& src_attrs, | |
3110 | RGWRados::Object::Read& read_op, | |
3111 | const rgw_user& user_id, | |
3112 | rgw_obj& dest_obj, | |
3113 | ceph::real_time *mtime); | |
3114 | ||
3115 | enum AttrsMod { | |
3116 | ATTRSMOD_NONE = 0, | |
3117 | ATTRSMOD_REPLACE = 1, | |
3118 | ATTRSMOD_MERGE = 2 | |
3119 | }; | |
3120 | ||
3121 | int rewrite_obj(RGWBucketInfo& dest_bucket_info, rgw_obj& obj); | |
3122 | ||
3123 | int stat_remote_obj(RGWObjectCtx& obj_ctx, | |
3124 | const rgw_user& user_id, | |
3125 | const string& client_id, | |
3126 | req_info *info, | |
3127 | const string& source_zone, | |
3128 | rgw_obj& src_obj, | |
3129 | RGWBucketInfo& src_bucket_info, | |
3130 | real_time *src_mtime, | |
3131 | uint64_t *psize, | |
3132 | const real_time *mod_ptr, | |
3133 | const real_time *unmod_ptr, | |
3134 | bool high_precision_time, | |
3135 | const char *if_match, | |
3136 | const char *if_nomatch, | |
3137 | map<string, bufferlist> *pattrs, | |
3138 | string *version_id, | |
3139 | string *ptag, | |
3140 | string *petag); | |
3141 | ||
3142 | int fetch_remote_obj(RGWObjectCtx& obj_ctx, | |
3143 | const rgw_user& user_id, | |
3144 | const string& client_id, | |
3145 | const string& op_id, | |
3146 | bool record_op_state, | |
3147 | req_info *info, | |
3148 | const string& source_zone, | |
3149 | rgw_obj& dest_obj, | |
3150 | rgw_obj& src_obj, | |
3151 | RGWBucketInfo& dest_bucket_info, | |
3152 | RGWBucketInfo& src_bucket_info, | |
3153 | ceph::real_time *src_mtime, | |
3154 | ceph::real_time *mtime, | |
3155 | const ceph::real_time *mod_ptr, | |
3156 | const ceph::real_time *unmod_ptr, | |
3157 | bool high_precision_time, | |
3158 | const char *if_match, | |
3159 | const char *if_nomatch, | |
3160 | AttrsMod attrs_mod, | |
3161 | bool copy_if_newer, | |
3162 | map<string, bufferlist>& attrs, | |
3163 | RGWObjCategory category, | |
91327a77 | 3164 | boost::optional<uint64_t> olh_epoch, |
7c673cae FG |
3165 | ceph::real_time delete_at, |
3166 | string *version_id, | |
3167 | string *ptag, | |
3168 | ceph::buffer::list *petag, | |
7c673cae | 3169 | void (*progress_cb)(off_t, void *), |
31f18b77 FG |
3170 | void *progress_data, |
3171 | rgw_zone_set *zones_trace= nullptr); | |
7c673cae FG |
3172 | /** |
3173 | * Copy an object. | |
3174 | * dest_obj: the object to copy into | |
3175 | * src_obj: the object to copy from | |
3176 | * attrs: usage depends on attrs_mod parameter | |
3177 | * attrs_mod: the modification mode of the attrs, may have the following values: | |
3178 | * ATTRSMOD_NONE - the attributes of the source object will be | |
3179 | * copied without modifications, attrs parameter is ignored; | |
3180 | * ATTRSMOD_REPLACE - new object will have the attributes provided by attrs | |
3181 | * parameter, source object attributes are not copied; | |
3182 | * ATTRSMOD_MERGE - any conflicting meta keys on the source object's attributes | |
3183 | * are overwritten by values contained in attrs parameter. | |
7c673cae FG |
3184 | * Returns: 0 on success, -ERR# otherwise. |
3185 | */ | |
3186 | int copy_obj(RGWObjectCtx& obj_ctx, | |
3187 | const rgw_user& user_id, | |
3188 | const string& client_id, | |
3189 | const string& op_id, | |
3190 | req_info *info, | |
3191 | const string& source_zone, | |
3192 | rgw_obj& dest_obj, | |
3193 | rgw_obj& src_obj, | |
3194 | RGWBucketInfo& dest_bucket_info, | |
3195 | RGWBucketInfo& src_bucket_info, | |
3196 | ceph::real_time *src_mtime, | |
3197 | ceph::real_time *mtime, | |
3198 | const ceph::real_time *mod_ptr, | |
3199 | const ceph::real_time *unmod_ptr, | |
3200 | bool high_precision_time, | |
3201 | const char *if_match, | |
3202 | const char *if_nomatch, | |
3203 | AttrsMod attrs_mod, | |
3204 | bool copy_if_newer, | |
3205 | map<std::string, bufferlist>& attrs, | |
3206 | RGWObjCategory category, | |
3207 | uint64_t olh_epoch, | |
3208 | ceph::real_time delete_at, | |
3209 | string *version_id, | |
3210 | string *ptag, | |
3211 | ceph::buffer::list *petag, | |
7c673cae FG |
3212 | void (*progress_cb)(off_t, void *), |
3213 | void *progress_data); | |
3214 | ||
3215 | int copy_obj_data(RGWObjectCtx& obj_ctx, | |
3216 | RGWBucketInfo& dest_bucket_info, | |
3217 | RGWRados::Object::Read& read_op, off_t end, | |
3218 | rgw_obj& dest_obj, | |
3219 | rgw_obj& src_obj, | |
3220 | uint64_t max_chunk_size, | |
3221 | ceph::real_time *mtime, | |
3222 | ceph::real_time set_mtime, | |
3223 | map<string, bufferlist>& attrs, | |
3224 | RGWObjCategory category, | |
3225 | uint64_t olh_epoch, | |
3226 | ceph::real_time delete_at, | |
3227 | string *version_id, | |
3228 | string *ptag, | |
31f18b77 | 3229 | ceph::buffer::list *petag); |
7c673cae FG |
3230 | |
3231 | int check_bucket_empty(RGWBucketInfo& bucket_info); | |
3232 | ||
3233 | /** | |
3234 | * Delete a bucket. | |
3235 | * bucket: the name of the bucket to delete | |
3236 | * Returns 0 on success, -ERR# otherwise. | |
3237 | */ | |
3238 | int delete_bucket(RGWBucketInfo& bucket_info, RGWObjVersionTracker& objv_tracker, bool check_empty = true); | |
3239 | ||
3240 | bool is_meta_master(); | |
3241 | ||
3242 | /** | |
3243 | * Check to see if the bucket metadata is synced | |
3244 | */ | |
3245 | bool is_syncing_bucket_meta(const rgw_bucket& bucket); | |
3246 | void wakeup_meta_sync_shards(set<int>& shard_ids); | |
3247 | void wakeup_data_sync_shards(const string& source_zone, map<int, set<string> >& shard_ids); | |
3248 | ||
3249 | RGWMetaSyncStatusManager* get_meta_sync_manager(); | |
3250 | RGWDataSyncStatusManager* get_data_sync_manager(const std::string& source_zone); | |
3251 | ||
3252 | int set_bucket_owner(rgw_bucket& bucket, ACLOwner& owner); | |
3253 | int set_buckets_enabled(std::vector<rgw_bucket>& buckets, bool enabled); | |
3254 | int bucket_suspended(rgw_bucket& bucket, bool *suspended); | |
3255 | ||
3256 | /** Delete an object.*/ | |
3257 | int delete_obj(RGWObjectCtx& obj_ctx, | |
3258 | const RGWBucketInfo& bucket_owner, | |
3259 | const rgw_obj& src_obj, | |
3260 | int versioning_status, | |
3261 | uint16_t bilog_flags = 0, | |
31f18b77 FG |
3262 | const ceph::real_time& expiration_time = ceph::real_time(), |
3263 | rgw_zone_set *zones_trace = nullptr); | |
7c673cae FG |
3264 | |
3265 | /** Delete a raw object.*/ | |
3266 | int delete_raw_obj(const rgw_raw_obj& obj); | |
3267 | ||
3268 | /* Delete a system object */ | |
3269 | virtual int delete_system_obj(rgw_raw_obj& src_obj, RGWObjVersionTracker *objv_tracker = NULL); | |
3270 | ||
3271 | /** Remove an object from the bucket index */ | |
3272 | int delete_obj_index(const rgw_obj& obj); | |
3273 | ||
3274 | /** | |
31f18b77 FG |
3275 | * Get an attribute for a system object. |
3276 | * obj: the object to get attr | |
7c673cae FG |
3277 | * name: name of the attr to retrieve |
3278 | * dest: bufferlist to store the result in | |
3279 | * Returns: 0 on success, -ERR# otherwise. | |
3280 | */ | |
3281 | virtual int system_obj_get_attr(rgw_raw_obj& obj, const char *name, bufferlist& dest); | |
3282 | ||
3283 | int system_obj_set_attr(void *ctx, rgw_raw_obj& obj, const char *name, bufferlist& bl, | |
3284 | RGWObjVersionTracker *objv_tracker); | |
3285 | virtual int system_obj_set_attrs(void *ctx, rgw_raw_obj& obj, | |
3286 | map<string, bufferlist>& attrs, | |
3287 | map<string, bufferlist>* rmattrs, | |
3288 | RGWObjVersionTracker *objv_tracker); | |
3289 | ||
3290 | /** | |
3291 | * Set an attr on an object. | |
3292 | * bucket: name of the bucket holding the object | |
3293 | * obj: name of the object to set the attr on | |
3294 | * name: the attr to set | |
3295 | * bl: the contents of the attr | |
3296 | * Returns: 0 on success, -ERR# otherwise. | |
3297 | */ | |
3298 | int set_attr(void *ctx, const RGWBucketInfo& bucket_info, rgw_obj& obj, const char *name, bufferlist& bl); | |
3299 | ||
3300 | int set_attrs(void *ctx, const RGWBucketInfo& bucket_info, rgw_obj& obj, | |
3301 | map<string, bufferlist>& attrs, | |
3302 | map<string, bufferlist>* rmattrs); | |
3303 | ||
3304 | int get_system_obj_state(RGWObjectCtx *rctx, rgw_raw_obj& obj, RGWRawObjState **state, RGWObjVersionTracker *objv_tracker); | |
3305 | int get_obj_state(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state, | |
3306 | bool follow_olh, bool assume_noent = false); | |
3307 | int get_obj_state(RGWObjectCtx *rctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWObjState **state) { | |
3308 | return get_obj_state(rctx, bucket_info, obj, state, true); | |
3309 | } | |
3310 | ||
3311 | virtual int stat_system_obj(RGWObjectCtx& obj_ctx, | |
3312 | RGWRados::SystemObject::Read::GetObjState& state, | |
3313 | rgw_raw_obj& obj, | |
3314 | map<string, bufferlist> *attrs, | |
3315 | ceph::real_time *lastmod, | |
3316 | uint64_t *obj_size, | |
3317 | RGWObjVersionTracker *objv_tracker); | |
3318 | ||
3319 | virtual int get_system_obj(RGWObjectCtx& obj_ctx, RGWRados::SystemObject::Read::GetObjState& read_state, | |
3320 | RGWObjVersionTracker *objv_tracker, rgw_raw_obj& obj, | |
3321 | bufferlist& bl, off_t ofs, off_t end, | |
3322 | map<string, bufferlist> *attrs, | |
b32b8144 FG |
3323 | rgw_cache_entry_info *cache_info, |
3324 | boost::optional<obj_version> refresh_version = | |
3325 | boost::none); | |
7c673cae FG |
3326 | |
3327 | virtual void register_chained_cache(RGWChainedCache *cache) {} | |
3328 | virtual bool chain_cache_entry(list<rgw_cache_entry_info *>& cache_info_entries, RGWChainedCache::Entry *chained_entry) { return false; } | |
3329 | ||
3330 | int iterate_obj(RGWObjectCtx& ctx, | |
3331 | const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
3332 | off_t ofs, off_t end, | |
3333 | uint64_t max_chunk_size, | |
3334 | int (*iterate_obj_cb)(const RGWBucketInfo& bucket_info, const rgw_obj& obj, const rgw_raw_obj&, off_t, off_t, off_t, bool, RGWObjState *, void *), | |
3335 | void *arg); | |
3336 | ||
3337 | int flush_read_list(struct get_obj_data *d); | |
3338 | ||
3339 | int get_obj_iterate_cb(RGWObjectCtx *ctx, RGWObjState *astate, | |
3340 | const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
3341 | const rgw_raw_obj& read_obj, | |
3342 | off_t obj_ofs, off_t read_ofs, off_t len, | |
3343 | bool is_head_obj, void *arg); | |
3344 | ||
3345 | void get_obj_aio_completion_cb(librados::completion_t cb, void *arg); | |
3346 | ||
3347 | /** | |
3348 | * a simple object read without keeping state | |
3349 | */ | |
3350 | ||
3351 | virtual int raw_obj_stat(rgw_raw_obj& obj, uint64_t *psize, ceph::real_time *pmtime, uint64_t *epoch, | |
3352 | map<string, bufferlist> *attrs, bufferlist *first_chunk, | |
3353 | RGWObjVersionTracker *objv_tracker); | |
3354 | ||
3355 | int obj_operate(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::ObjectWriteOperation *op); | |
3356 | int obj_operate(const RGWBucketInfo& bucket_info, const rgw_obj& obj, librados::ObjectReadOperation *op); | |
3357 | ||
f64942e4 AA |
3358 | int guard_reshard(BucketShard *bs, |
3359 | const rgw_obj& obj_instance, | |
3360 | const RGWBucketInfo& bucket_info, | |
3361 | std::function<int(BucketShard *)> call); | |
3362 | int block_while_resharding(RGWRados::BucketShard *bs, | |
3363 | string *new_bucket_id, | |
3364 | const RGWBucketInfo& bucket_info); | |
31f18b77 | 3365 | |
7c673cae FG |
3366 | void bucket_index_guard_olh_op(RGWObjState& olh_state, librados::ObjectOperation& op); |
3367 | int olh_init_modification(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, string *op_tag); | |
3368 | int olh_init_modification_impl(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, string *op_tag); | |
3369 | int bucket_index_link_olh(const RGWBucketInfo& bucket_info, RGWObjState& olh_state, | |
3370 | const rgw_obj& obj_instance, bool delete_marker, | |
3371 | const string& op_tag, struct rgw_bucket_dir_entry_meta *meta, | |
3372 | uint64_t olh_epoch, | |
91327a77 AA |
3373 | ceph::real_time unmod_since, bool high_precision_time, |
3374 | rgw_zone_set *zones_trace = nullptr, | |
3375 | bool log_data_change = false); | |
31f18b77 | 3376 | 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 |
3377 | int bucket_index_read_olh_log(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& obj_instance, uint64_t ver_marker, |
3378 | map<uint64_t, vector<rgw_bucket_olh_log_entry> > *log, bool *is_truncated); | |
3379 | int bucket_index_trim_olh_log(const RGWBucketInfo& bucket_info, RGWObjState& obj_state, const rgw_obj& obj_instance, uint64_t ver); | |
3380 | int bucket_index_clear_olh(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& obj_instance); | |
3381 | int apply_olh_log(RGWObjectCtx& ctx, RGWObjState& obj_state, const RGWBucketInfo& bucket_info, const rgw_obj& obj, | |
3382 | bufferlist& obj_tag, map<uint64_t, vector<rgw_bucket_olh_log_entry> >& log, | |
31f18b77 FG |
3383 | uint64_t *plast_ver, rgw_zone_set *zones_trace = nullptr); |
3384 | int update_olh(RGWObjectCtx& obj_ctx, RGWObjState *state, const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 3385 | 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 |
3386 | uint64_t olh_epoch, ceph::real_time unmod_since, bool high_precision_time, |
3387 | rgw_zone_set *zones_trace = nullptr, bool log_data_change = false); | |
a8e16298 TL |
3388 | int repair_olh(RGWObjState* state, const RGWBucketInfo& bucket_info, |
3389 | const rgw_obj& obj); | |
7c673cae | 3390 | int unlink_obj_instance(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, const rgw_obj& target_obj, |
31f18b77 | 3391 | uint64_t olh_epoch, rgw_zone_set *zones_trace = nullptr); |
7c673cae FG |
3392 | |
3393 | void check_pending_olh_entries(map<string, bufferlist>& pending_entries, map<string, bufferlist> *rm_pending_entries); | |
3394 | int remove_olh_pending_entries(const RGWBucketInfo& bucket_info, RGWObjState& state, const rgw_obj& olh_obj, map<string, bufferlist>& pending_attrs); | |
3395 | int follow_olh(const RGWBucketInfo& bucket_info, RGWObjectCtx& ctx, RGWObjState *state, const rgw_obj& olh_obj, rgw_obj *target); | |
3396 | int get_olh(const RGWBucketInfo& bucket_info, const rgw_obj& obj, RGWOLHInfo *olh); | |
3397 | ||
3398 | void gen_rand_obj_instance_name(rgw_obj *target); | |
3399 | ||
3400 | int omap_get_vals(rgw_raw_obj& obj, bufferlist& header, const std::string& marker, uint64_t count, std::map<string, bufferlist>& m); | |
3401 | int omap_get_all(rgw_raw_obj& obj, bufferlist& header, std::map<string, bufferlist>& m); | |
3402 | int omap_set(rgw_raw_obj& obj, const std::string& key, bufferlist& bl); | |
3403 | int omap_set(rgw_raw_obj& obj, map<std::string, bufferlist>& m); | |
3404 | int omap_del(rgw_raw_obj& obj, const std::string& key); | |
3405 | int update_containers_stats(map<string, RGWBucketEnt>& m); | |
3406 | int append_async(rgw_raw_obj& obj, size_t size, bufferlist& bl); | |
3407 | ||
3408 | int watch(const string& oid, uint64_t *watch_handle, librados::WatchCtx2 *ctx); | |
3409 | int unwatch(uint64_t watch_handle); | |
3410 | void add_watcher(int i); | |
3411 | void remove_watcher(int i); | |
3412 | virtual bool need_watch_notify() { return false; } | |
3413 | int init_watch(); | |
3414 | void finalize_watch(); | |
3415 | int distribute(const string& key, bufferlist& bl); | |
3416 | virtual int watch_cb(uint64_t notify_id, | |
3417 | uint64_t cookie, | |
3418 | uint64_t notifier_id, | |
3419 | bufferlist& bl) { return 0; } | |
3420 | void pick_control_oid(const string& key, string& notify_oid); | |
3421 | ||
3422 | virtual void set_cache_enabled(bool state) {} | |
3423 | ||
3424 | void set_atomic(void *ctx, rgw_obj& obj) { | |
3425 | RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx); | |
3426 | rctx->obj.set_atomic(obj); | |
3427 | } | |
3428 | void set_prefetch_data(void *ctx, rgw_obj& obj) { | |
3429 | RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx); | |
3430 | rctx->obj.set_prefetch_data(obj); | |
3431 | } | |
3432 | void set_prefetch_data(void *ctx, rgw_raw_obj& obj) { | |
3433 | RGWObjectCtx *rctx = static_cast<RGWObjectCtx *>(ctx); | |
3434 | rctx->raw.set_prefetch_data(obj); | |
3435 | } | |
3436 | ||
3437 | int decode_policy(bufferlist& bl, ACLOwner *owner); | |
3438 | int get_bucket_stats(RGWBucketInfo& bucket_info, int shard_id, string *bucket_ver, string *master_ver, | |
c07f9fc5 | 3439 | map<RGWObjCategory, RGWStorageStats>& stats, string *max_marker, bool* syncstopped = NULL); |
7c673cae FG |
3440 | int get_bucket_stats_async(RGWBucketInfo& bucket_info, int shard_id, RGWGetBucketStats_CB *cb); |
3441 | int get_user_stats(const rgw_user& user, RGWStorageStats& stats); | |
3442 | int get_user_stats_async(const rgw_user& user, RGWGetUserStats_CB *cb); | |
3443 | void get_bucket_instance_obj(const rgw_bucket& bucket, rgw_raw_obj& obj); | |
3444 | void get_bucket_meta_oid(const rgw_bucket& bucket, string& oid); | |
3445 | ||
3446 | int put_bucket_entrypoint_info(const string& tenant_name, const string& bucket_name, RGWBucketEntryPoint& entry_point, | |
3447 | bool exclusive, RGWObjVersionTracker& objv_tracker, ceph::real_time mtime, | |
3448 | map<string, bufferlist> *pattrs); | |
3449 | int put_bucket_instance_info(RGWBucketInfo& info, bool exclusive, ceph::real_time mtime, map<string, bufferlist> *pattrs); | |
3450 | int get_bucket_entrypoint_info(RGWObjectCtx& obj_ctx, const string& tenant_name, const string& bucket_name, | |
3451 | RGWBucketEntryPoint& entry_point, RGWObjVersionTracker *objv_tracker, | |
b32b8144 FG |
3452 | ceph::real_time *pmtime, map<string, bufferlist> *pattrs, rgw_cache_entry_info *cache_info = NULL, |
3453 | boost::optional<obj_version> refresh_version = boost::none); | |
7c673cae FG |
3454 | int get_bucket_instance_info(RGWObjectCtx& obj_ctx, const string& meta_key, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs); |
3455 | int get_bucket_instance_info(RGWObjectCtx& obj_ctx, const rgw_bucket& bucket, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs); | |
31f18b77 | 3456 | int get_bucket_instance_from_oid(RGWObjectCtx& obj_ctx, const string& oid, RGWBucketInfo& info, ceph::real_time *pmtime, map<string, bufferlist> *pattrs, |
b32b8144 FG |
3457 | rgw_cache_entry_info *cache_info = NULL, |
3458 | boost::optional<obj_version> refresh_version = boost::none); | |
7c673cae FG |
3459 | |
3460 | int convert_old_bucket_info(RGWObjectCtx& obj_ctx, const string& tenant_name, const string& bucket_name); | |
3461 | static void make_bucket_entry_name(const string& tenant_name, const string& bucket_name, string& bucket_entry); | |
b32b8144 FG |
3462 | |
3463 | ||
3464 | private: | |
3465 | int _get_bucket_info(RGWObjectCtx& obj_ctx, const string& tenant, | |
3466 | const string& bucket_name, RGWBucketInfo& info, | |
3467 | real_time *pmtime, | |
3468 | map<string, bufferlist> *pattrs, | |
3469 | boost::optional<obj_version> refresh_version); | |
3470 | public: | |
3471 | ||
3a9019d9 FG |
3472 | bool call(std::string command, cmdmap_t& cmdmap, std::string format, |
3473 | bufferlist& out) override final; | |
3474 | ||
3475 | // Should really be protected, but some older GCCs don't handle | |
3476 | // access control properly with lambdas defined in member functions | |
3477 | // of child classes. | |
3478 | ||
3479 | void cache_list_dump_helper(Formatter* f, | |
3480 | const std::string& name, | |
3481 | const ceph::real_time mtime, | |
3482 | const std::uint64_t size) { | |
3483 | f->dump_string("name", name); | |
3484 | f->dump_string("mtime", ceph::to_iso_8601(mtime)); | |
3485 | f->dump_unsigned("size", size); | |
3486 | } | |
3487 | ||
3488 | protected: | |
3489 | ||
3490 | // `call_list` must iterate over all cache entries and call | |
3491 | // `cache_list_dump_helper` with the supplied Formatter on any that | |
3492 | // include `filter` as a substring. | |
3493 | // | |
3494 | virtual void call_list(const boost::optional<std::string>& filter, | |
3495 | Formatter* format); | |
3496 | // `call_inspect` must look up the requested target and, if found, | |
3497 | // dump it to the supplied Formatter and return true. If not found, | |
3498 | // it must return false. | |
3499 | // | |
3500 | virtual bool call_inspect(const std::string& target, Formatter* format); | |
3501 | ||
3502 | // `call_erase` must erase the requested target and return true. If | |
3503 | // the requested target does not exist, it should return false. | |
3504 | virtual bool call_erase(const std::string& target); | |
3505 | ||
3506 | // `call_zap` must erase the cache. | |
3507 | virtual void call_zap(); | |
3508 | public: | |
b32b8144 | 3509 | |
7c673cae | 3510 | int get_bucket_info(RGWObjectCtx& obj_ctx, |
b32b8144 FG |
3511 | const string& tenant_name, const string& bucket_name, |
3512 | RGWBucketInfo& info, | |
3513 | ceph::real_time *pmtime, map<string, bufferlist> *pattrs = NULL); | |
3514 | ||
3515 | // Returns true on successful refresh. Returns false if there was an | |
3516 | // error or the version stored on the OSD is the same as that | |
3517 | // presented in the BucketInfo structure. | |
3518 | // | |
3519 | int try_refresh_bucket_info(RGWBucketInfo& info, | |
3520 | ceph::real_time *pmtime, | |
3521 | map<string, bufferlist> *pattrs = nullptr); | |
3522 | ||
7c673cae | 3523 | int put_linked_bucket_info(RGWBucketInfo& info, bool exclusive, ceph::real_time mtime, obj_version *pep_objv, |
b32b8144 | 3524 | map<string, bufferlist> *pattrs, bool create_entry_point); |
7c673cae | 3525 | |
31f18b77 FG |
3526 | int cls_obj_prepare_op(BucketShard& bs, RGWModifyOp op, string& tag, rgw_obj& obj, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); |
3527 | int cls_obj_complete_op(BucketShard& bs, const rgw_obj& obj, RGWModifyOp op, string& tag, int64_t pool, uint64_t epoch, | |
3528 | rgw_bucket_dir_entry& ent, RGWObjCategory category, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); | |
3529 | int cls_obj_complete_add(BucketShard& bs, const rgw_obj& obj, string& tag, int64_t pool, uint64_t epoch, rgw_bucket_dir_entry& ent, | |
3530 | RGWObjCategory category, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 3531 | int cls_obj_complete_del(BucketShard& bs, string& tag, int64_t pool, uint64_t epoch, rgw_obj& obj, |
31f18b77 FG |
3532 | ceph::real_time& removed_mtime, list<rgw_obj_index_key> *remove_objs, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); |
3533 | int cls_obj_complete_cancel(BucketShard& bs, string& tag, rgw_obj& obj, uint16_t bilog_flags, rgw_zone_set *zones_trace = nullptr); | |
7c673cae | 3534 | int cls_obj_set_bucket_tag_timeout(RGWBucketInfo& bucket_info, uint64_t timeout); |
1adf2230 AA |
3535 | int cls_bucket_list_ordered(RGWBucketInfo& bucket_info, int shard_id, |
3536 | rgw_obj_index_key& start, const string& prefix, | |
3537 | uint32_t num_entries, bool list_versions, | |
3538 | map<string, rgw_bucket_dir_entry>& m, | |
3539 | bool *is_truncated, | |
3540 | rgw_obj_index_key *last_entry, | |
3541 | bool (*force_check_filter)(const string& name) = nullptr); | |
3542 | int cls_bucket_list_unordered(RGWBucketInfo& bucket_info, int shard_id, | |
3543 | rgw_obj_index_key& start, const string& prefix, | |
3544 | uint32_t num_entries, bool list_versions, | |
3545 | vector<rgw_bucket_dir_entry>& ent_list, | |
3546 | bool *is_truncated, rgw_obj_index_key *last_entry, | |
3547 | bool (*force_check_filter)(const string& name) = nullptr); | |
a8e16298 | 3548 | 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 |
3549 | int cls_bucket_head_async(const RGWBucketInfo& bucket_info, int shard_id, RGWGetDirHeader_CB *ctx, int *num_aio); |
3550 | 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); | |
3551 | int trim_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id, string& marker, string& end_marker); | |
c07f9fc5 FG |
3552 | int resync_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id); |
3553 | int stop_bi_log_entries(RGWBucketInfo& bucket_info, int shard_id); | |
7c673cae FG |
3554 | int get_bi_log_status(RGWBucketInfo& bucket_info, int shard_id, map<int, string>& max_marker); |
3555 | ||
a8e16298 TL |
3556 | int bi_get_instance(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_bucket_dir_entry *dirent); |
3557 | int bi_get_olh(const RGWBucketInfo& bucket_info, const rgw_obj& obj, rgw_bucket_olh_entry *olh); | |
3558 | int bi_get(const RGWBucketInfo& bucket_info, const rgw_obj& obj, BIIndexType index_type, rgw_cls_bi_entry *entry); | |
7c673cae FG |
3559 | void bi_put(librados::ObjectWriteOperation& op, BucketShard& bs, rgw_cls_bi_entry& entry); |
3560 | int bi_put(BucketShard& bs, rgw_cls_bi_entry& entry); | |
3561 | int bi_put(rgw_bucket& bucket, rgw_obj& obj, rgw_cls_bi_entry& entry); | |
3562 | 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); | |
3563 | int bi_list(BucketShard& bs, const string& filter_obj, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated); | |
3564 | int bi_list(rgw_bucket& bucket, const string& obj_name, const string& marker, uint32_t max, | |
3565 | list<rgw_cls_bi_entry> *entries, bool *is_truncated); | |
3566 | int bi_remove(BucketShard& bs); | |
3567 | ||
3568 | int cls_obj_usage_log_add(const string& oid, rgw_usage_log_info& info); | |
3569 | int cls_obj_usage_log_read(string& oid, string& user, uint64_t start_epoch, uint64_t end_epoch, uint32_t max_entries, | |
3570 | string& read_iter, map<rgw_user_bucket, rgw_usage_log_entry>& usage, bool *is_truncated); | |
3571 | int cls_obj_usage_log_trim(string& oid, string& user, uint64_t start_epoch, uint64_t end_epoch); | |
3572 | ||
3573 | int key_to_shard_id(const string& key, int max_shards); | |
3574 | void shard_name(const string& prefix, unsigned max_shards, const string& key, string& name, int *shard_id); | |
3575 | void shard_name(const string& prefix, unsigned max_shards, const string& section, const string& key, string& name); | |
3576 | void shard_name(const string& prefix, unsigned shard_id, string& name); | |
3577 | int get_target_shard_id(const RGWBucketInfo& bucket_info, const string& obj_key, int *shard_id); | |
3578 | void time_log_prepare_entry(cls_log_entry& entry, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl); | |
3579 | int time_log_add_init(librados::IoCtx& io_ctx); | |
3580 | int time_log_add(const string& oid, list<cls_log_entry>& entries, | |
3581 | librados::AioCompletion *completion, bool monotonic_inc = true); | |
3582 | int time_log_add(const string& oid, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl); | |
3583 | int time_log_list(const string& oid, const ceph::real_time& start_time, const ceph::real_time& end_time, | |
3584 | int max_entries, list<cls_log_entry>& entries, | |
3585 | const string& marker, string *out_marker, bool *truncated); | |
3586 | int time_log_info(const string& oid, cls_log_header *header); | |
3587 | int time_log_info_async(librados::IoCtx& io_ctx, const string& oid, cls_log_header *header, librados::AioCompletion *completion); | |
3588 | int time_log_trim(const string& oid, const ceph::real_time& start_time, const ceph::real_time& end_time, | |
3589 | const string& from_marker, const string& to_marker, | |
3590 | librados::AioCompletion *completion = nullptr); | |
3591 | ||
3592 | string objexp_hint_get_shardname(int shard_num); | |
3593 | int objexp_key_shard(const rgw_obj_index_key& key); | |
3594 | void objexp_get_shard(int shard_num, | |
3595 | string& shard); /* out */ | |
3596 | int objexp_hint_add(const ceph::real_time& delete_at, | |
3597 | const string& tenant_name, | |
3598 | const string& bucket_name, | |
3599 | const string& bucket_id, | |
3600 | const rgw_obj_index_key& obj_key); | |
3601 | int objexp_hint_list(const string& oid, | |
3602 | const ceph::real_time& start_time, | |
3603 | const ceph::real_time& end_time, | |
3604 | const int max_entries, | |
3605 | const string& marker, | |
3606 | list<cls_timeindex_entry>& entries, /* out */ | |
3607 | string *out_marker, /* out */ | |
3608 | bool *truncated); /* out */ | |
3609 | int objexp_hint_parse(cls_timeindex_entry &ti_entry, | |
3610 | objexp_hint_entry& hint_entry); /* out */ | |
3611 | int objexp_hint_trim(const string& oid, | |
3612 | const ceph::real_time& start_time, | |
3613 | const ceph::real_time& end_time, | |
3614 | const string& from_marker = std::string(), | |
3615 | const string& to_marker = std::string()); | |
3616 | ||
3617 | int lock_exclusive(rgw_pool& pool, const string& oid, ceph::timespan& duration, string& zone_id, string& owner_id); | |
3618 | int unlock(rgw_pool& pool, const string& oid, string& zone_id, string& owner_id); | |
3619 | ||
3620 | void update_gc_chain(rgw_obj& head_obj, RGWObjManifest& manifest, cls_rgw_obj_chain *chain); | |
3621 | int send_chain_to_gc(cls_rgw_obj_chain& chain, const string& tag, bool sync); | |
3622 | int gc_operate(string& oid, librados::ObjectWriteOperation *op); | |
3623 | int gc_aio_operate(string& oid, librados::ObjectWriteOperation *op); | |
3624 | int gc_operate(string& oid, librados::ObjectReadOperation *op, bufferlist *pbl); | |
3625 | ||
3626 | int list_gc_objs(int *index, string& marker, uint32_t max, bool expired_only, std::list<cls_rgw_gc_obj_info>& result, bool *truncated); | |
3627 | int process_gc(); | |
1adf2230 | 3628 | bool process_expire_objects(); |
7c673cae FG |
3629 | int defer_gc(void *ctx, const RGWBucketInfo& bucket_info, const rgw_obj& obj); |
3630 | ||
3631 | int process_lc(); | |
3632 | int list_lc_progress(const string& marker, uint32_t max_entries, map<string, int> *progress_map); | |
3633 | ||
3634 | int bucket_check_index(RGWBucketInfo& bucket_info, | |
3635 | map<RGWObjCategory, RGWStorageStats> *existing_stats, | |
3636 | map<RGWObjCategory, RGWStorageStats> *calculated_stats); | |
3637 | int bucket_rebuild_index(RGWBucketInfo& bucket_info); | |
f64942e4 | 3638 | int bucket_set_reshard(const RGWBucketInfo& bucket_info, const cls_rgw_bucket_instance_entry& entry); |
7c673cae FG |
3639 | int remove_objs_from_index(RGWBucketInfo& bucket_info, list<rgw_obj_index_key>& oid_list); |
3640 | int move_rados_obj(librados::IoCtx& src_ioctx, | |
3641 | const string& src_oid, const string& src_locator, | |
3642 | librados::IoCtx& dst_ioctx, | |
3643 | const string& dst_oid, const string& dst_locator); | |
3644 | int fix_head_obj_locator(const RGWBucketInfo& bucket_info, bool copy_obj, bool remove_bad, rgw_obj_key& key); | |
3645 | int fix_tail_obj_locator(const RGWBucketInfo& bucket_info, rgw_obj_key& key, bool fix, bool *need_fix); | |
3646 | ||
3647 | int cls_user_get_header(const string& user_id, cls_user_header *header); | |
94b18763 | 3648 | int cls_user_reset_stats(const string& user_id); |
7c673cae FG |
3649 | int cls_user_get_header_async(const string& user_id, RGWGetUserHeader_CB *ctx); |
3650 | int cls_user_sync_bucket_stats(rgw_raw_obj& user_obj, const RGWBucketInfo& bucket_info); | |
3651 | int cls_user_list_buckets(rgw_raw_obj& obj, | |
3652 | const string& in_marker, | |
3653 | const string& end_marker, | |
3654 | int max_entries, | |
3655 | list<cls_user_bucket_entry>& entries, | |
3656 | string *out_marker, | |
3657 | bool *truncated); | |
3658 | int cls_user_add_bucket(rgw_raw_obj& obj, const cls_user_bucket_entry& entry); | |
3659 | int cls_user_update_buckets(rgw_raw_obj& obj, list<cls_user_bucket_entry>& entries, bool add); | |
3660 | int cls_user_complete_stats_sync(rgw_raw_obj& obj); | |
3661 | int complete_sync_user_stats(const rgw_user& user_id); | |
3662 | int cls_user_add_bucket(rgw_raw_obj& obj, list<cls_user_bucket_entry>& entries); | |
3663 | int cls_user_remove_bucket(rgw_raw_obj& obj, const cls_user_bucket& bucket); | |
c07f9fc5 | 3664 | int cls_user_get_bucket_stats(const rgw_bucket& bucket, cls_user_bucket_entry& entry); |
7c673cae FG |
3665 | |
3666 | int check_quota(const rgw_user& bucket_owner, rgw_bucket& bucket, | |
3667 | RGWQuotaInfo& user_quota, RGWQuotaInfo& bucket_quota, uint64_t obj_size); | |
3668 | ||
224ce89b | 3669 | int check_bucket_shards(const RGWBucketInfo& bucket_info, const rgw_bucket& bucket, |
31f18b77 FG |
3670 | RGWQuotaInfo& bucket_quota); |
3671 | ||
3672 | int add_bucket_to_reshard(const RGWBucketInfo& bucket_info, uint32_t new_num_shards); | |
3673 | ||
7c673cae | 3674 | uint64_t instance_id(); |
224ce89b WB |
3675 | const string& zone_name() { |
3676 | return get_zone_params().get_name(); | |
3677 | } | |
7c673cae FG |
3678 | const string& zone_id() { |
3679 | return get_zone_params().get_id(); | |
3680 | } | |
3681 | string unique_id(uint64_t unique_num) { | |
3682 | char buf[32]; | |
3683 | snprintf(buf, sizeof(buf), ".%llu.%llu", (unsigned long long)instance_id(), (unsigned long long)unique_num); | |
3684 | string s = get_zone_params().get_id() + buf; | |
3685 | return s; | |
3686 | } | |
3687 | ||
3688 | void init_unique_trans_id_deps() { | |
3689 | char buf[16 + 2 + 1]; /* uint64_t needs 16, 2 hyphens add further 2 */ | |
3690 | ||
3691 | snprintf(buf, sizeof(buf), "-%llx-", (unsigned long long)instance_id()); | |
3692 | url_encode(string(buf) + get_zone_params().get_name(), trans_id_suffix); | |
3693 | } | |
3694 | ||
3695 | /* In order to preserve compability with Swift API, transaction ID | |
3696 | * should contain at least 32 characters satisfying following spec: | |
3697 | * - first 21 chars must be in range [0-9a-f]. Swift uses this | |
3698 | * space for storing fragment of UUID obtained through a call to | |
3699 | * uuid4() function of Python's uuid module; | |
3700 | * - char no. 22 must be a hyphen; | |
3701 | * - at least 10 next characters constitute hex-formatted timestamp | |
3702 | * padded with zeroes if necessary. All bytes must be in [0-9a-f] | |
3703 | * range; | |
3704 | * - last, optional part of transaction ID is any url-encoded string | |
3705 | * without restriction on length. */ | |
3706 | string unique_trans_id(const uint64_t unique_num) { | |
3707 | char buf[41]; /* 2 + 21 + 1 + 16 (timestamp can consume up to 16) + 1 */ | |
3708 | time_t timestamp = time(NULL); | |
3709 | ||
3710 | snprintf(buf, sizeof(buf), "tx%021llx-%010llx", | |
3711 | (unsigned long long)unique_num, | |
3712 | (unsigned long long)timestamp); | |
3713 | ||
3714 | return string(buf) + trans_id_suffix; | |
3715 | } | |
3716 | ||
3717 | void get_log_pool(rgw_pool& pool) { | |
3718 | pool = get_zone_params().log_pool; | |
3719 | } | |
3720 | ||
3721 | bool need_to_log_data() { | |
3722 | return get_zone().log_data; | |
3723 | } | |
3724 | ||
3725 | bool need_to_log_metadata() { | |
224ce89b WB |
3726 | return is_meta_master() && |
3727 | (get_zonegroup().zones.size() > 1 || current_period.is_multi_zonegroups_with_zones()); | |
7c673cae FG |
3728 | } |
3729 | ||
3efd9988 FG |
3730 | bool can_reshard() const { |
3731 | return current_period.get_id().empty() || | |
3732 | (zonegroup.zones.size() == 1 && current_period.is_single_zonegroup()); | |
3733 | } | |
3734 | ||
7c673cae FG |
3735 | librados::Rados* get_rados_handle(); |
3736 | ||
3737 | int delete_raw_obj_aio(const rgw_raw_obj& obj, list<librados::AioCompletion *>& handles); | |
3738 | int delete_obj_aio(const rgw_obj& obj, RGWBucketInfo& info, RGWObjState *astate, | |
3739 | list<librados::AioCompletion *>& handles, bool keep_index_consistent); | |
3740 | private: | |
3741 | /** | |
3742 | * This is a helper method, it generates a list of bucket index objects with the given | |
3743 | * bucket base oid and number of shards. | |
3744 | * | |
3745 | * bucket_oid_base [in] - base name of the bucket index object; | |
3746 | * num_shards [in] - number of bucket index object shards. | |
3747 | * bucket_objs [out] - filled by this method, a list of bucket index objects. | |
3748 | */ | |
3749 | void get_bucket_index_objects(const string& bucket_oid_base, uint32_t num_shards, | |
3750 | map<int, string>& bucket_objs, int shard_id = -1); | |
3751 | ||
3752 | /** | |
3753 | * Get the bucket index object with the given base bucket index object and object key, | |
3754 | * and the number of bucket index shards. | |
3755 | * | |
3756 | * bucket_oid_base [in] - bucket object base name. | |
3757 | * obj_key [in] - object key. | |
3758 | * num_shards [in] - number of bucket index shards. | |
3759 | * hash_type [in] - type of hash to find the shard ID. | |
3760 | * bucket_obj [out] - the bucket index object for the given object. | |
3761 | * | |
3762 | * Return 0 on success, a failure code otherwise. | |
3763 | */ | |
3764 | int get_bucket_index_object(const string& bucket_oid_base, const string& obj_key, | |
3765 | uint32_t num_shards, RGWBucketInfo::BIShardsHashType hash_type, string *bucket_obj, int *shard); | |
3766 | ||
3767 | void get_bucket_index_object(const string& bucket_oid_base, uint32_t num_shards, | |
3768 | int shard_id, string *bucket_obj); | |
3769 | ||
3770 | /** | |
3771 | * Check the actual on-disk state of the object specified | |
3772 | * by list_state, and fill in the time and size of object. | |
3773 | * Then append any changes to suggested_updates for | |
3774 | * the rgw class' dir_suggest_changes function. | |
3775 | * | |
3776 | * Note that this can maul list_state; don't use it afterwards. Also | |
3777 | * it expects object to already be filled in from list_state; it only | |
3778 | * sets the size and mtime. | |
3779 | * | |
3780 | * Returns 0 on success, -ENOENT if the object doesn't exist on disk, | |
3781 | * and -errno on other failures. (-ENOENT is not a failure, and it | |
3782 | * will encode that info as a suggested update.) | |
3783 | */ | |
3784 | int check_disk_state(librados::IoCtx io_ctx, | |
3785 | const RGWBucketInfo& bucket_info, | |
3786 | rgw_bucket_dir_entry& list_state, | |
3787 | rgw_bucket_dir_entry& object, | |
3788 | bufferlist& suggested_updates); | |
3789 | ||
3790 | /** | |
3791 | * Init pool iteration | |
31f18b77 | 3792 | * pool: pool to use for the ctx initialization |
7c673cae FG |
3793 | * ctx: context object to use for the iteration |
3794 | * Returns: 0 on success, -ERR# otherwise. | |
3795 | */ | |
3796 | int pool_iterate_begin(const rgw_pool& pool, RGWPoolIterCtx& ctx); | |
31f18b77 | 3797 | |
181888fb FG |
3798 | /** |
3799 | * Init pool iteration | |
3800 | * pool: pool to use | |
3801 | * cursor: position to start iteration | |
3802 | * ctx: context object to use for the iteration | |
3803 | * Returns: 0 on success, -ERR# otherwise. | |
3804 | */ | |
3805 | int pool_iterate_begin(const rgw_pool& pool, const string& cursor, RGWPoolIterCtx& ctx); | |
3806 | ||
3807 | /** | |
3808 | * Get pool iteration position | |
3809 | * ctx: context object to use for the iteration | |
3810 | * Returns: string representation of position | |
3811 | */ | |
3812 | string pool_iterate_get_cursor(RGWPoolIterCtx& ctx); | |
3813 | ||
7c673cae FG |
3814 | /** |
3815 | * Iterate over pool return object names, use optional filter | |
3816 | * ctx: iteration context, initialized with pool_iterate_begin() | |
3817 | * num: max number of objects to return | |
3818 | * objs: a vector that the results will append into | |
3819 | * is_truncated: if not NULL, will hold true iff iteration is complete | |
3820 | * filter: if not NULL, will be used to filter returned objects | |
3821 | * Returns: 0 on success, -ERR# otherwise. | |
3822 | */ | |
3823 | int pool_iterate(RGWPoolIterCtx& ctx, uint32_t num, vector<rgw_bucket_dir_entry>& objs, | |
3824 | bool *is_truncated, RGWAccessListFilter *filter); | |
3825 | ||
3826 | uint64_t next_bucket_id(); | |
3827 | }; | |
3828 | ||
3829 | class RGWStoreManager { | |
3830 | public: | |
3831 | RGWStoreManager() {} | |
28e407b8 AA |
3832 | static RGWRados *get_storage(CephContext *cct, bool use_gc_thread, bool use_lc_thread, bool quota_threads, |
3833 | bool run_sync_thread, bool run_reshard_thread, bool use_cache = true) { | |
31f18b77 | 3834 | RGWRados *store = init_storage_provider(cct, use_gc_thread, use_lc_thread, quota_threads, run_sync_thread, |
28e407b8 | 3835 | run_reshard_thread, use_cache); |
7c673cae FG |
3836 | return store; |
3837 | } | |
3838 | static RGWRados *get_raw_storage(CephContext *cct) { | |
3839 | RGWRados *store = init_raw_storage_provider(cct); | |
3840 | return store; | |
3841 | } | |
28e407b8 | 3842 | 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 |
3843 | static RGWRados *init_raw_storage_provider(CephContext *cct); |
3844 | static void close_storage(RGWRados *store); | |
3845 | ||
3846 | }; | |
3847 | ||
3848 | template <class T> | |
3849 | class RGWChainedCacheImpl : public RGWChainedCache { | |
b32b8144 | 3850 | ceph::timespan expiry; |
7c673cae FG |
3851 | RWLock lock; |
3852 | ||
b32b8144 | 3853 | map<string, std::pair<T, ceph::coarse_mono_time>> entries; |
7c673cae FG |
3854 | |
3855 | public: | |
3856 | RGWChainedCacheImpl() : lock("RGWChainedCacheImpl::lock") {} | |
3857 | ||
3858 | void init(RGWRados *store) { | |
3859 | store->register_chained_cache(this); | |
b32b8144 FG |
3860 | expiry = std::chrono::seconds(store->ctx()->_conf->get_val<uint64_t>( |
3861 | "rgw_cache_expiry_interval")); | |
7c673cae FG |
3862 | } |
3863 | ||
3864 | bool find(const string& key, T *entry) { | |
3865 | RWLock::RLocker rl(lock); | |
b32b8144 | 3866 | auto iter = entries.find(key); |
7c673cae FG |
3867 | if (iter == entries.end()) { |
3868 | return false; | |
3869 | } | |
b32b8144 FG |
3870 | if (expiry.count() && |
3871 | (ceph::coarse_mono_clock::now() - iter->second.second) > expiry) { | |
3872 | return false; | |
3873 | } | |
7c673cae | 3874 | |
b32b8144 | 3875 | *entry = iter->second.first; |
7c673cae FG |
3876 | return true; |
3877 | } | |
3878 | ||
3879 | bool put(RGWRados *store, const string& key, T *entry, list<rgw_cache_entry_info *>& cache_info_entries) { | |
3880 | Entry chain_entry(this, key, entry); | |
3881 | ||
3882 | /* we need the store cache to call us under its lock to maintain lock ordering */ | |
3883 | return store->chain_cache_entry(cache_info_entries, &chain_entry); | |
3884 | } | |
3885 | ||
3886 | void chain_cb(const string& key, void *data) override { | |
3887 | T *entry = static_cast<T *>(data); | |
3888 | RWLock::WLocker wl(lock); | |
b32b8144 FG |
3889 | entries[key].first = *entry; |
3890 | if (expiry.count() > 0) { | |
3891 | entries[key].second = ceph::coarse_mono_clock::now(); | |
3892 | } | |
7c673cae FG |
3893 | } |
3894 | ||
3895 | void invalidate(const string& key) override { | |
3896 | RWLock::WLocker wl(lock); | |
3897 | entries.erase(key); | |
3898 | } | |
3899 | ||
3900 | void invalidate_all() override { | |
3901 | RWLock::WLocker wl(lock); | |
3902 | entries.clear(); | |
3903 | } | |
3904 | }; /* RGWChainedCacheImpl */ | |
3905 | ||
3906 | /** | |
3907 | * Base of PUT operation. | |
3908 | * Allow to create chained data transformers like compresors and encryptors. | |
3909 | */ | |
3910 | class RGWPutObjDataProcessor | |
3911 | { | |
3912 | public: | |
3913 | RGWPutObjDataProcessor(){} | |
3914 | virtual ~RGWPutObjDataProcessor(){} | |
3915 | virtual int handle_data(bufferlist& bl, off_t ofs, void **phandle, rgw_raw_obj *pobj, bool *again) = 0; | |
3916 | virtual int throttle_data(void *handle, const rgw_raw_obj& obj, uint64_t size, bool need_to_wait) = 0; | |
3917 | }; /* RGWPutObjDataProcessor */ | |
3918 | ||
3919 | ||
3920 | class RGWPutObjProcessor : public RGWPutObjDataProcessor | |
3921 | { | |
3922 | protected: | |
3923 | RGWRados *store; | |
3924 | RGWObjectCtx& obj_ctx; | |
3925 | bool is_complete; | |
3926 | RGWBucketInfo bucket_info; | |
3927 | bool canceled; | |
3928 | ||
3929 | virtual int do_complete(size_t accounted_size, const string& etag, | |
3930 | ceph::real_time *mtime, ceph::real_time set_mtime, | |
3931 | map<string, bufferlist>& attrs, ceph::real_time delete_at, | |
31f18b77 FG |
3932 | const char *if_match, const char *if_nomatch, const string *user_data, |
3933 | rgw_zone_set* zones_trace = nullptr) = 0; | |
7c673cae FG |
3934 | |
3935 | public: | |
3936 | RGWPutObjProcessor(RGWObjectCtx& _obj_ctx, RGWBucketInfo& _bi) : store(NULL), | |
3937 | obj_ctx(_obj_ctx), | |
3938 | is_complete(false), | |
3939 | bucket_info(_bi), | |
3940 | canceled(false) {} | |
3941 | ~RGWPutObjProcessor() override {} | |
3942 | virtual int prepare(RGWRados *_store, string *oid_rand) { | |
3943 | store = _store; | |
3944 | return 0; | |
3945 | } | |
3946 | ||
3947 | int complete(size_t accounted_size, const string& etag, | |
3948 | ceph::real_time *mtime, ceph::real_time set_mtime, | |
3949 | map<string, bufferlist>& attrs, ceph::real_time delete_at, | |
31f18b77 FG |
3950 | const char *if_match = NULL, const char *if_nomatch = NULL, const string *user_data = nullptr, |
3951 | rgw_zone_set *zones_trace = nullptr); | |
7c673cae FG |
3952 | |
3953 | CephContext *ctx(); | |
3954 | ||
3955 | bool is_canceled() { return canceled; } | |
3956 | }; /* RGWPutObjProcessor */ | |
3957 | ||
3958 | struct put_obj_aio_info { | |
3959 | void *handle; | |
3960 | rgw_raw_obj obj; | |
3961 | uint64_t size; | |
3962 | }; | |
3963 | ||
3964 | #define RGW_PUT_OBJ_MIN_WINDOW_SIZE_DEFAULT (16 * 1024 * 1024) | |
3965 | ||
3966 | class RGWPutObjProcessor_Aio : public RGWPutObjProcessor | |
3967 | { | |
3968 | list<struct put_obj_aio_info> pending; | |
3969 | uint64_t window_size{RGW_PUT_OBJ_MIN_WINDOW_SIZE_DEFAULT}; | |
3970 | uint64_t pending_size{0}; | |
3971 | ||
3972 | struct put_obj_aio_info pop_pending(); | |
3973 | int wait_pending_front(); | |
3974 | bool pending_has_completed(); | |
3975 | ||
3976 | rgw_raw_obj last_written_obj; | |
3977 | ||
3978 | protected: | |
3979 | uint64_t obj_len{0}; | |
3980 | ||
3981 | set<rgw_raw_obj> written_objs; | |
3982 | rgw_obj head_obj; | |
3983 | ||
3984 | void add_written_obj(const rgw_raw_obj& obj) { | |
3985 | written_objs.insert(obj); | |
3986 | } | |
3987 | ||
3988 | int drain_pending(); | |
3989 | int handle_obj_data(rgw_raw_obj& obj, bufferlist& bl, off_t ofs, off_t abs_ofs, void **phandle, bool exclusive); | |
3990 | ||
3991 | public: | |
3992 | int prepare(RGWRados *store, string *oid_rand) override; | |
3993 | int throttle_data(void *handle, const rgw_raw_obj& obj, uint64_t size, bool need_to_wait) override; | |
3994 | ||
3995 | RGWPutObjProcessor_Aio(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info) : RGWPutObjProcessor(obj_ctx, bucket_info) {} | |
3996 | ~RGWPutObjProcessor_Aio() override; | |
3997 | }; /* RGWPutObjProcessor_Aio */ | |
3998 | ||
3999 | class RGWPutObjProcessor_Atomic : public RGWPutObjProcessor_Aio | |
4000 | { | |
4001 | bufferlist first_chunk; | |
4002 | uint64_t part_size; | |
4003 | off_t cur_part_ofs; | |
4004 | off_t next_part_ofs; | |
4005 | int cur_part_id; | |
4006 | off_t data_ofs; | |
4007 | ||
4008 | bufferlist pending_data_bl; | |
4009 | uint64_t max_chunk_size; | |
4010 | ||
4011 | bool versioned_object; | |
91327a77 | 4012 | boost::optional<uint64_t> olh_epoch; |
7c673cae FG |
4013 | string version_id; |
4014 | ||
4015 | protected: | |
4016 | rgw_bucket bucket; | |
4017 | string obj_str; | |
4018 | ||
4019 | string unique_tag; | |
4020 | ||
4021 | rgw_raw_obj cur_obj; | |
4022 | RGWObjManifest manifest; | |
4023 | RGWObjManifest::generator manifest_gen; | |
4024 | ||
4025 | int write_data(bufferlist& bl, off_t ofs, void **phandle, rgw_raw_obj *pobj, bool exclusive); | |
4026 | int do_complete(size_t accounted_size, const string& etag, | |
4027 | ceph::real_time *mtime, ceph::real_time set_mtime, | |
4028 | map<string, bufferlist>& attrs, ceph::real_time delete_at, | |
31f18b77 | 4029 | const char *if_match, const char *if_nomatch, const string *user_data, rgw_zone_set *zones_trace) override; |
7c673cae FG |
4030 | |
4031 | int prepare_next_part(off_t ofs); | |
4032 | int complete_parts(); | |
4033 | int complete_writing_data(); | |
4034 | ||
4035 | int prepare_init(RGWRados *store, string *oid_rand); | |
4036 | ||
4037 | public: | |
4038 | ~RGWPutObjProcessor_Atomic() override {} | |
4039 | RGWPutObjProcessor_Atomic(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, | |
4040 | rgw_bucket& _b, const string& _o, uint64_t _p, const string& _t, bool versioned) : | |
4041 | RGWPutObjProcessor_Aio(obj_ctx, bucket_info), | |
4042 | part_size(_p), | |
4043 | cur_part_ofs(0), | |
4044 | next_part_ofs(_p), | |
4045 | cur_part_id(0), | |
4046 | data_ofs(0), | |
4047 | max_chunk_size(0), | |
4048 | versioned_object(versioned), | |
7c673cae FG |
4049 | bucket(_b), |
4050 | obj_str(_o), | |
4051 | unique_tag(_t) {} | |
4052 | int prepare(RGWRados *store, string *oid_rand) override; | |
4053 | virtual bool immutable_head() { return false; } | |
4054 | int handle_data(bufferlist& bl, off_t ofs, void **phandle, rgw_raw_obj *pobj, bool *again) override; | |
4055 | ||
4056 | void set_olh_epoch(uint64_t epoch) { | |
4057 | olh_epoch = epoch; | |
4058 | } | |
4059 | ||
4060 | void set_version_id(const string& vid) { | |
4061 | version_id = vid; | |
4062 | } | |
4063 | }; /* RGWPutObjProcessor_Atomic */ | |
4064 | ||
4065 | #define MP_META_SUFFIX ".meta" | |
4066 | ||
4067 | class RGWMPObj { | |
4068 | string oid; | |
4069 | string prefix; | |
4070 | string meta; | |
4071 | string upload_id; | |
4072 | public: | |
4073 | RGWMPObj() {} | |
4074 | RGWMPObj(const string& _oid, const string& _upload_id) { | |
4075 | init(_oid, _upload_id, _upload_id); | |
4076 | } | |
4077 | void init(const string& _oid, const string& _upload_id) { | |
4078 | init(_oid, _upload_id, _upload_id); | |
4079 | } | |
4080 | void init(const string& _oid, const string& _upload_id, const string& part_unique_str) { | |
4081 | if (_oid.empty()) { | |
4082 | clear(); | |
4083 | return; | |
4084 | } | |
4085 | oid = _oid; | |
4086 | upload_id = _upload_id; | |
4087 | prefix = oid + "."; | |
4088 | meta = prefix + upload_id + MP_META_SUFFIX; | |
4089 | prefix.append(part_unique_str); | |
4090 | } | |
4091 | string& get_meta() { return meta; } | |
4092 | string get_part(int num) { | |
4093 | char buf[16]; | |
4094 | snprintf(buf, 16, ".%d", num); | |
4095 | string s = prefix; | |
4096 | s.append(buf); | |
4097 | return s; | |
4098 | } | |
4099 | string get_part(string& part) { | |
4100 | string s = prefix; | |
4101 | s.append("."); | |
4102 | s.append(part); | |
4103 | return s; | |
4104 | } | |
4105 | string& get_upload_id() { | |
4106 | return upload_id; | |
4107 | } | |
4108 | string& get_key() { | |
4109 | return oid; | |
4110 | } | |
4111 | bool from_meta(string& meta) { | |
4112 | int end_pos = meta.rfind('.'); // search for ".meta" | |
4113 | if (end_pos < 0) | |
4114 | return false; | |
4115 | int mid_pos = meta.rfind('.', end_pos - 1); // <key>.<upload_id> | |
4116 | if (mid_pos < 0) | |
4117 | return false; | |
4118 | oid = meta.substr(0, mid_pos); | |
4119 | upload_id = meta.substr(mid_pos + 1, end_pos - mid_pos - 1); | |
4120 | init(oid, upload_id, upload_id); | |
4121 | return true; | |
4122 | } | |
4123 | void clear() { | |
4124 | oid = ""; | |
4125 | prefix = ""; | |
4126 | meta = ""; | |
4127 | upload_id = ""; | |
4128 | } | |
4129 | }; | |
4130 | ||
4131 | class RGWPutObjProcessor_Multipart : public RGWPutObjProcessor_Atomic | |
4132 | { | |
4133 | string part_num; | |
4134 | RGWMPObj mp; | |
4135 | req_state *s; | |
4136 | string upload_id; | |
4137 | ||
4138 | protected: | |
4139 | int prepare(RGWRados *store, string *oid_rand); | |
4140 | int do_complete(size_t accounted_size, const string& etag, | |
4141 | ceph::real_time *mtime, ceph::real_time set_mtime, | |
4142 | map<string, bufferlist>& attrs, ceph::real_time delete_at, | |
31f18b77 FG |
4143 | const char *if_match, const char *if_nomatch, const string *user_data, |
4144 | rgw_zone_set *zones_trace) override; | |
7c673cae FG |
4145 | public: |
4146 | bool immutable_head() { return true; } | |
4147 | RGWPutObjProcessor_Multipart(RGWObjectCtx& obj_ctx, RGWBucketInfo& bucket_info, uint64_t _p, req_state *_s) : | |
4148 | RGWPutObjProcessor_Atomic(obj_ctx, bucket_info, _s->bucket, _s->object.name, _p, _s->req_id, false), s(_s) {} | |
4149 | void get_mp(RGWMPObj** _mp); | |
4150 | }; /* RGWPutObjProcessor_Multipart */ | |
4151 | #endif |