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