]>
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_RGW_METADATA_H | |
5 | #define CEPH_RGW_METADATA_H | |
6 | ||
7 | #include <string> | |
8 | #include <boost/optional.hpp> | |
9 | ||
10 | #include "include/types.h" | |
11 | #include "rgw_common.h" | |
12 | #include "rgw_period_history.h" | |
13 | #include "cls/version/cls_version_types.h" | |
14 | #include "cls/log/cls_log_types.h" | |
15 | #include "common/RWLock.h" | |
16 | #include "common/RefCountedObj.h" | |
17 | #include "common/ceph_time.h" | |
18 | ||
19 | ||
20 | class RGWRados; | |
21 | class RGWCoroutine; | |
22 | class JSONObj; | |
23 | struct RGWObjVersionTracker; | |
24 | ||
25 | struct obj_version; | |
26 | ||
27 | ||
28 | enum RGWMDLogStatus { | |
29 | MDLOG_STATUS_UNKNOWN, | |
30 | MDLOG_STATUS_WRITE, | |
31 | MDLOG_STATUS_SETATTRS, | |
32 | MDLOG_STATUS_REMOVE, | |
33 | MDLOG_STATUS_COMPLETE, | |
34 | MDLOG_STATUS_ABORT, | |
35 | }; | |
36 | ||
37 | class RGWMetadataObject { | |
38 | protected: | |
39 | obj_version objv; | |
40 | ceph::real_time mtime; | |
41 | ||
42 | public: | |
43 | RGWMetadataObject() {} | |
44 | virtual ~RGWMetadataObject() {} | |
45 | obj_version& get_version(); | |
46 | real_time get_mtime() { return mtime; } | |
47 | ||
48 | virtual void dump(Formatter *f) const = 0; | |
49 | }; | |
50 | ||
51 | class RGWMetadataManager; | |
52 | ||
53 | class RGWMetadataHandler { | |
54 | friend class RGWMetadataManager; | |
55 | ||
56 | public: | |
57 | enum sync_type_t { | |
58 | APPLY_ALWAYS, | |
59 | APPLY_UPDATES, | |
60 | APPLY_NEWER | |
61 | }; | |
62 | static bool string_to_sync_type(const string& sync_string, | |
63 | sync_type_t& type) { | |
64 | if (sync_string.compare("update-by-version") == 0) | |
65 | type = APPLY_UPDATES; | |
66 | else if (sync_string.compare("update-by-timestamp") == 0) | |
67 | type = APPLY_NEWER; | |
68 | else if (sync_string.compare("always") == 0) | |
69 | type = APPLY_ALWAYS; | |
70 | else | |
71 | return false; | |
72 | return true; | |
73 | } | |
74 | virtual ~RGWMetadataHandler() {} | |
75 | virtual string get_type() = 0; | |
76 | ||
77 | virtual int get(RGWRados *store, string& entry, RGWMetadataObject **obj) = 0; | |
78 | virtual int put(RGWRados *store, string& entry, RGWObjVersionTracker& objv_tracker, | |
79 | real_time mtime, JSONObj *obj, sync_type_t type) = 0; | |
80 | virtual int remove(RGWRados *store, string& entry, RGWObjVersionTracker& objv_tracker) = 0; | |
81 | ||
82 | virtual int list_keys_init(RGWRados *store, void **phandle) = 0; | |
83 | virtual int list_keys_next(void *handle, int max, list<string>& keys, bool *truncated) = 0; | |
84 | virtual void list_keys_complete(void *handle) = 0; | |
85 | ||
86 | /* key to use for hashing entries for log shard placement */ | |
87 | virtual void get_hash_key(const string& section, const string& key, string& hash_key) { | |
88 | hash_key = section + ":" + key; | |
89 | } | |
90 | ||
91 | protected: | |
92 | virtual void get_pool_and_oid(RGWRados *store, const string& key, rgw_pool& pool, string& oid) = 0; | |
93 | /** | |
94 | * Compare an incoming versus on-disk tag/version+mtime combo against | |
95 | * the sync mode to see if the new one should replace the on-disk one. | |
96 | * | |
97 | * @return true if the update should proceed, false otherwise. | |
98 | */ | |
99 | bool check_versions(const obj_version& ondisk, const real_time& ondisk_time, | |
100 | const obj_version& incoming, const real_time& incoming_time, | |
101 | sync_type_t sync_mode) { | |
102 | switch (sync_mode) { | |
103 | case APPLY_UPDATES: | |
104 | if ((ondisk.tag != incoming.tag) || | |
105 | (ondisk.ver >= incoming.ver)) | |
106 | return false; | |
107 | break; | |
108 | case APPLY_NEWER: | |
109 | if (ondisk_time >= incoming_time) | |
110 | return false; | |
111 | break; | |
112 | case APPLY_ALWAYS: //deliberate fall-thru -- we always apply! | |
113 | default: break; | |
114 | } | |
115 | return true; | |
116 | } | |
117 | ||
118 | /* | |
119 | * The tenant_name is always returned on purpose. May be empty, of course. | |
120 | */ | |
121 | static void parse_bucket(const string& bucket, | |
122 | string *tenant_name, | |
123 | string *bucket_name, | |
124 | string *bucket_instance = nullptr /* optional */) | |
125 | { | |
126 | int pos = bucket.find('/'); | |
127 | if (pos >= 0) { | |
128 | *tenant_name = bucket.substr(0, pos); | |
129 | } else { | |
130 | tenant_name->clear(); | |
131 | } | |
132 | string bn = bucket.substr(pos + 1); | |
133 | pos = bn.find (':'); | |
134 | if (pos < 0) { | |
135 | *bucket_name = std::move(bn); | |
136 | return; | |
137 | } | |
138 | *bucket_name = bn.substr(0, pos); | |
139 | if (bucket_instance) { | |
140 | *bucket_instance = bn.substr(pos + 1); | |
141 | } | |
142 | } | |
143 | }; | |
144 | ||
145 | #define META_LOG_OBJ_PREFIX "meta.log." | |
146 | ||
147 | struct RGWMetadataLogInfo { | |
148 | string marker; | |
149 | real_time last_update; | |
150 | ||
151 | void dump(Formatter *f) const; | |
152 | void decode_json(JSONObj *obj); | |
153 | }; | |
154 | ||
155 | class RGWCompletionManager; | |
156 | ||
157 | class RGWMetadataLogInfoCompletion : public RefCountedObject { | |
158 | public: | |
159 | using info_callback_t = std::function<void(int, const cls_log_header&)>; | |
160 | private: | |
161 | cls_log_header header; | |
162 | librados::IoCtx io_ctx; | |
163 | librados::AioCompletion *completion; | |
164 | std::mutex mutex; //< protects callback between cancel/complete | |
165 | boost::optional<info_callback_t> callback; //< cleared on cancel | |
166 | public: | |
167 | RGWMetadataLogInfoCompletion(info_callback_t callback); | |
168 | ~RGWMetadataLogInfoCompletion() override; | |
169 | ||
170 | librados::IoCtx& get_io_ctx() { return io_ctx; } | |
171 | cls_log_header& get_header() { return header; } | |
172 | librados::AioCompletion* get_completion() { return completion; } | |
173 | ||
174 | void finish(librados::completion_t cb) { | |
175 | std::lock_guard<std::mutex> lock(mutex); | |
176 | if (callback) { | |
177 | (*callback)(completion->get_return_value(), header); | |
178 | } | |
179 | } | |
180 | void cancel() { | |
181 | std::lock_guard<std::mutex> lock(mutex); | |
182 | callback = boost::none; | |
183 | } | |
184 | }; | |
185 | ||
186 | class RGWMetadataLog { | |
187 | CephContext *cct; | |
188 | RGWRados *store; | |
189 | const string prefix; | |
190 | ||
191 | static std::string make_prefix(const std::string& period) { | |
192 | if (period.empty()) | |
193 | return META_LOG_OBJ_PREFIX; | |
194 | return META_LOG_OBJ_PREFIX + period + "."; | |
195 | } | |
196 | ||
197 | RWLock lock; | |
198 | set<int> modified_shards; | |
199 | ||
200 | void mark_modified(int shard_id); | |
201 | public: | |
202 | RGWMetadataLog(CephContext *_cct, RGWRados *_store, const std::string& period) | |
203 | : cct(_cct), store(_store), | |
204 | prefix(make_prefix(period)), | |
205 | lock("RGWMetaLog::lock") {} | |
206 | ||
207 | void get_shard_oid(int id, string& oid) const { | |
208 | char buf[16]; | |
209 | snprintf(buf, sizeof(buf), "%d", id); | |
210 | oid = prefix + buf; | |
211 | } | |
212 | ||
213 | int add_entry(RGWMetadataHandler *handler, const string& section, const string& key, bufferlist& bl); | |
214 | int store_entries_in_shard(list<cls_log_entry>& entries, int shard_id, librados::AioCompletion *completion); | |
215 | ||
216 | struct LogListCtx { | |
217 | int cur_shard; | |
218 | string marker; | |
219 | real_time from_time; | |
220 | real_time end_time; | |
221 | ||
222 | string cur_oid; | |
223 | ||
224 | bool done; | |
225 | ||
226 | LogListCtx() : cur_shard(0), done(false) {} | |
227 | }; | |
228 | ||
229 | void init_list_entries(int shard_id, const real_time& from_time, const real_time& end_time, string& marker, void **handle); | |
230 | void complete_list_entries(void *handle); | |
231 | int list_entries(void *handle, | |
232 | int max_entries, | |
233 | list<cls_log_entry>& entries, | |
234 | string *out_marker, | |
235 | bool *truncated); | |
236 | ||
237 | int trim(int shard_id, const real_time& from_time, const real_time& end_time, const string& start_marker, const string& end_marker); | |
238 | int get_info(int shard_id, RGWMetadataLogInfo *info); | |
239 | int get_info_async(int shard_id, RGWMetadataLogInfoCompletion *completion); | |
240 | int lock_exclusive(int shard_id, timespan duration, string&zone_id, string& owner_id); | |
241 | int unlock(int shard_id, string& zone_id, string& owner_id); | |
242 | ||
243 | int update_shards(list<int>& shards); | |
244 | ||
245 | void read_clear_modified(set<int> &modified); | |
246 | }; | |
247 | ||
248 | struct LogStatusDump { | |
249 | RGWMDLogStatus status; | |
250 | ||
251 | explicit LogStatusDump(RGWMDLogStatus _status) : status(_status) {} | |
252 | void dump(Formatter *f) const; | |
253 | }; | |
254 | ||
255 | struct RGWMetadataLogData { | |
256 | obj_version read_version; | |
257 | obj_version write_version; | |
258 | RGWMDLogStatus status; | |
259 | ||
260 | RGWMetadataLogData() : status(MDLOG_STATUS_UNKNOWN) {} | |
261 | ||
262 | void encode(bufferlist& bl) const; | |
263 | void decode(bufferlist::iterator& bl); | |
264 | void dump(Formatter *f) const; | |
265 | void decode_json(JSONObj *obj); | |
266 | }; | |
267 | WRITE_CLASS_ENCODER(RGWMetadataLogData) | |
268 | ||
269 | struct RGWMetadataLogHistory { | |
270 | epoch_t oldest_realm_epoch; | |
271 | std::string oldest_period_id; | |
272 | ||
273 | void encode(bufferlist& bl) const { | |
274 | ENCODE_START(1, 1, bl); | |
275 | ::encode(oldest_realm_epoch, bl); | |
276 | ::encode(oldest_period_id, bl); | |
277 | ENCODE_FINISH(bl); | |
278 | } | |
279 | void decode(bufferlist::iterator& p) { | |
280 | DECODE_START(1, p); | |
281 | ::decode(oldest_realm_epoch, p); | |
282 | ::decode(oldest_period_id, p); | |
283 | DECODE_FINISH(p); | |
284 | } | |
285 | ||
286 | static const std::string oid; | |
287 | }; | |
288 | WRITE_CLASS_ENCODER(RGWMetadataLogHistory) | |
289 | ||
290 | class RGWMetadataManager { | |
291 | map<string, RGWMetadataHandler *> handlers; | |
292 | CephContext *cct; | |
293 | RGWRados *store; | |
294 | ||
295 | // maintain a separate metadata log for each period | |
296 | std::map<std::string, RGWMetadataLog> md_logs; | |
297 | // use the current period's log for mutating operations | |
298 | RGWMetadataLog* current_log = nullptr; | |
299 | ||
300 | void parse_metadata_key(const string& metadata_key, string& type, string& entry); | |
301 | ||
302 | int find_handler(const string& metadata_key, RGWMetadataHandler **handler, string& entry); | |
303 | int pre_modify(RGWMetadataHandler *handler, string& section, const string& key, | |
304 | RGWMetadataLogData& log_data, RGWObjVersionTracker *objv_tracker, | |
305 | RGWMDLogStatus op_type); | |
306 | int post_modify(RGWMetadataHandler *handler, const string& section, const string& key, RGWMetadataLogData& log_data, | |
307 | RGWObjVersionTracker *objv_tracker, int ret); | |
308 | ||
309 | string heap_oid(RGWMetadataHandler *handler, const string& key, const obj_version& objv); | |
310 | int store_in_heap(RGWMetadataHandler *handler, const string& key, bufferlist& bl, | |
311 | RGWObjVersionTracker *objv_tracker, real_time mtime, | |
312 | map<string, bufferlist> *pattrs); | |
313 | int remove_from_heap(RGWMetadataHandler *handler, const string& key, RGWObjVersionTracker *objv_tracker); | |
314 | public: | |
315 | RGWMetadataManager(CephContext *_cct, RGWRados *_store); | |
316 | ~RGWMetadataManager(); | |
317 | ||
318 | int init(const std::string& current_period); | |
319 | ||
320 | /// initialize the oldest log period if it doesn't exist, and attach it to | |
321 | /// our current history | |
322 | RGWPeriodHistory::Cursor init_oldest_log_period(); | |
323 | ||
324 | /// read the oldest log period, and return a cursor to it in our existing | |
325 | /// period history | |
326 | RGWPeriodHistory::Cursor read_oldest_log_period() const; | |
327 | ||
328 | /// read the oldest log period asynchronously and write its result to the | |
329 | /// given cursor pointer | |
330 | RGWCoroutine* read_oldest_log_period_cr(RGWPeriodHistory::Cursor *period, | |
331 | RGWObjVersionTracker *objv) const; | |
332 | ||
333 | /// try to advance the oldest log period when the given period is trimmed, | |
334 | /// using a rados lock to provide atomicity | |
335 | RGWCoroutine* trim_log_period_cr(RGWPeriodHistory::Cursor period, | |
336 | RGWObjVersionTracker *objv) const; | |
337 | ||
338 | /// find or create the metadata log for the given period | |
339 | RGWMetadataLog* get_log(const std::string& period); | |
340 | ||
341 | int register_handler(RGWMetadataHandler *handler); | |
342 | ||
343 | RGWMetadataHandler *get_handler(const string& type); | |
344 | ||
345 | int put_entry(RGWMetadataHandler *handler, const string& key, bufferlist& bl, bool exclusive, | |
346 | RGWObjVersionTracker *objv_tracker, real_time mtime, map<string, bufferlist> *pattrs = NULL); | |
347 | int remove_entry(RGWMetadataHandler *handler, string& key, RGWObjVersionTracker *objv_tracker); | |
348 | int get(string& metadata_key, Formatter *f); | |
349 | int put(string& metadata_key, bufferlist& bl, | |
350 | RGWMetadataHandler::sync_type_t sync_mode, | |
351 | obj_version *existing_version = NULL); | |
352 | int remove(string& metadata_key); | |
353 | ||
354 | int list_keys_init(string& section, void **phandle); | |
355 | int list_keys_next(void *handle, int max, list<string>& keys, bool *truncated); | |
356 | void list_keys_complete(void *handle); | |
357 | ||
358 | void dump_log_entry(cls_log_entry& entry, Formatter *f); | |
359 | ||
360 | void get_sections(list<string>& sections); | |
361 | int lock_exclusive(string& metadata_key, timespan duration, string& owner_id); | |
362 | int unlock(string& metadata_key, string& owner_id); | |
363 | ||
364 | int get_log_shard_id(const string& section, const string& key, int *shard_id); | |
365 | }; | |
366 | ||
367 | #endif |