]>
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 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2014 Red Hat | |
7 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #ifndef CEPH_OSD_KSTORE_H | |
16 | #define CEPH_OSD_KSTORE_H | |
17 | ||
18 | #include "acconfig.h" | |
19 | ||
20 | #include <unistd.h> | |
21 | ||
22 | #include <atomic> | |
23 | #include <mutex> | |
24 | #include <condition_variable> | |
25 | ||
11fdf7f2 | 26 | #include "include/ceph_assert.h" |
7c673cae | 27 | #include "include/unordered_map.h" |
7c673cae FG |
28 | #include "common/Finisher.h" |
29 | #include "common/RWLock.h" | |
11fdf7f2 | 30 | #include "common/Throttle.h" |
7c673cae FG |
31 | #include "common/WorkQueue.h" |
32 | #include "os/ObjectStore.h" | |
33 | #include "common/perf_counters.h" | |
34 | #include "os/fs/FS.h" | |
35 | #include "kv/KeyValueDB.h" | |
36 | ||
37 | #include "kstore_types.h" | |
38 | ||
39 | #include "boost/intrusive/list.hpp" | |
40 | ||
41 | enum { | |
42 | l_kstore_first = 832430, | |
43 | l_kstore_state_prepare_lat, | |
44 | l_kstore_state_kv_queued_lat, | |
45 | l_kstore_state_kv_done_lat, | |
46 | l_kstore_state_finishing_lat, | |
47 | l_kstore_state_done_lat, | |
48 | l_kstore_last | |
49 | }; | |
50 | ||
51 | class KStore : public ObjectStore { | |
52 | // ----------------------------------------------------- | |
53 | // types | |
54 | public: | |
55 | ||
f67539c2 | 56 | struct TransContext; |
7c673cae FG |
57 | |
58 | /// an in-memory object | |
59 | struct Onode { | |
60 | CephContext* cct; | |
61 | std::atomic_int nref; ///< reference count | |
62 | ||
63 | ghobject_t oid; | |
f67539c2 | 64 | std::string key; ///< key under PREFIX_OBJ where we are stored |
7c673cae FG |
65 | boost::intrusive::list_member_hook<> lru_item; |
66 | ||
67 | kstore_onode_t onode; ///< metadata stored as value in kv store | |
68 | bool dirty; // ??? | |
69 | bool exists; | |
70 | ||
71 | std::mutex flush_lock; ///< protect flush_txns | |
72 | std::condition_variable flush_cond; ///< wait here for unapplied txns | |
f67539c2 | 73 | std::set<TransContext*> flush_txns; ///< committing txns |
7c673cae FG |
74 | |
75 | uint64_t tail_offset; | |
f67539c2 | 76 | ceph::buffer::list tail_bl; |
7c673cae | 77 | |
f67539c2 | 78 | std::map<uint64_t,ceph::buffer::list> pending_stripes; ///< unwritten stripes |
7c673cae | 79 | |
f67539c2 | 80 | Onode(CephContext* cct, const ghobject_t& o, const std::string& k) |
7c673cae FG |
81 | : cct(cct), |
82 | nref(0), | |
83 | oid(o), | |
84 | key(k), | |
85 | dirty(false), | |
86 | exists(false), | |
87 | tail_offset(0) { | |
88 | } | |
89 | ||
90 | void flush(); | |
91 | void get() { | |
92 | ++nref; | |
93 | } | |
94 | void put() { | |
95 | if (--nref == 0) | |
96 | delete this; | |
97 | } | |
98 | ||
99 | void clear_tail() { | |
100 | tail_offset = 0; | |
101 | tail_bl.clear(); | |
102 | } | |
103 | void clear_pending_stripes() { | |
104 | pending_stripes.clear(); | |
105 | } | |
106 | }; | |
107 | typedef boost::intrusive_ptr<Onode> OnodeRef; | |
108 | ||
109 | struct OnodeHashLRU { | |
110 | CephContext* cct; | |
111 | typedef boost::intrusive::list< | |
112 | Onode, | |
113 | boost::intrusive::member_hook< | |
114 | Onode, | |
115 | boost::intrusive::list_member_hook<>, | |
116 | &Onode::lru_item> > lru_list_t; | |
117 | ||
118 | std::mutex lock; | |
119 | ceph::unordered_map<ghobject_t,OnodeRef> onode_map; ///< forward lookups | |
120 | lru_list_t lru; ///< lru | |
121 | ||
122 | OnodeHashLRU(CephContext* cct) : cct(cct) {} | |
123 | ||
124 | void add(const ghobject_t& oid, OnodeRef o); | |
125 | void _touch(OnodeRef o); | |
126 | OnodeRef lookup(const ghobject_t& o); | |
127 | void rename(const ghobject_t& old_oid, const ghobject_t& new_oid); | |
128 | void clear(); | |
f67539c2 | 129 | bool get_next(const ghobject_t& after, std::pair<ghobject_t,OnodeRef> *next); |
7c673cae FG |
130 | int trim(int max=-1); |
131 | }; | |
132 | ||
11fdf7f2 TL |
133 | class OpSequencer; |
134 | typedef boost::intrusive_ptr<OpSequencer> OpSequencerRef; | |
135 | ||
7c673cae FG |
136 | struct Collection : public CollectionImpl { |
137 | KStore *store; | |
7c673cae | 138 | kstore_cnode_t cnode; |
9f95a23c TL |
139 | ceph::shared_mutex lock = |
140 | ceph::make_shared_mutex("KStore::Collection::lock", true, false); | |
7c673cae | 141 | |
11fdf7f2 TL |
142 | OpSequencerRef osr; |
143 | ||
7c673cae FG |
144 | // cache onodes on a per-collection basis to avoid lock |
145 | // contention. | |
146 | OnodeHashLRU onode_map; | |
147 | ||
148 | OnodeRef get_onode(const ghobject_t& oid, bool create); | |
149 | ||
7c673cae FG |
150 | bool contains(const ghobject_t& oid) { |
151 | if (cid.is_meta()) | |
152 | return oid.hobj.pool == -1; | |
153 | spg_t spgid; | |
154 | if (cid.is_pg(&spgid)) | |
155 | return | |
156 | spgid.pgid.contains(cnode.bits, oid) && | |
157 | oid.shard_id == spgid.shard; | |
158 | return false; | |
159 | } | |
160 | ||
11fdf7f2 TL |
161 | void flush() override; |
162 | bool flush_commit(Context *c) override; | |
163 | ||
9f95a23c TL |
164 | private: |
165 | FRIEND_MAKE_REF(Collection); | |
7c673cae FG |
166 | Collection(KStore *ns, coll_t c); |
167 | }; | |
9f95a23c | 168 | using CollectionRef = ceph::ref_t<Collection>; |
7c673cae FG |
169 | |
170 | class OmapIteratorImpl : public ObjectMap::ObjectMapIteratorImpl { | |
171 | CollectionRef c; | |
172 | OnodeRef o; | |
173 | KeyValueDB::Iterator it; | |
f67539c2 | 174 | std::string head, tail; |
7c673cae FG |
175 | public: |
176 | OmapIteratorImpl(CollectionRef c, OnodeRef o, KeyValueDB::Iterator it); | |
177 | int seek_to_first() override; | |
f67539c2 TL |
178 | int upper_bound(const std::string &after) override; |
179 | int lower_bound(const std::string &to) override; | |
7c673cae | 180 | bool valid() override; |
11fdf7f2 | 181 | int next() override; |
f67539c2 TL |
182 | std::string key() override; |
183 | ceph::buffer::list value() override; | |
7c673cae FG |
184 | int status() override { |
185 | return 0; | |
186 | } | |
187 | }; | |
188 | ||
7c673cae FG |
189 | struct TransContext { |
190 | typedef enum { | |
191 | STATE_PREPARE, | |
192 | STATE_AIO_WAIT, | |
193 | STATE_IO_DONE, | |
194 | STATE_KV_QUEUED, | |
195 | STATE_KV_COMMITTING, | |
196 | STATE_KV_DONE, | |
197 | STATE_FINISHING, | |
198 | STATE_DONE, | |
199 | } state_t; | |
200 | ||
201 | state_t state; | |
202 | ||
203 | const char *get_state_name() { | |
204 | switch (state) { | |
205 | case STATE_PREPARE: return "prepare"; | |
206 | case STATE_AIO_WAIT: return "aio_wait"; | |
207 | case STATE_IO_DONE: return "io_done"; | |
208 | case STATE_KV_QUEUED: return "kv_queued"; | |
209 | case STATE_KV_COMMITTING: return "kv_committing"; | |
210 | case STATE_KV_DONE: return "kv_done"; | |
211 | case STATE_FINISHING: return "finishing"; | |
212 | case STATE_DONE: return "done"; | |
213 | } | |
214 | return "???"; | |
215 | } | |
216 | ||
217 | void log_state_latency(PerfCounters *logger, int state) { | |
218 | utime_t lat, now = ceph_clock_now(); | |
219 | lat = now - start; | |
220 | logger->tinc(state, lat); | |
221 | start = now; | |
222 | } | |
223 | ||
11fdf7f2 | 224 | CollectionRef ch; |
7c673cae FG |
225 | OpSequencerRef osr; |
226 | boost::intrusive::list_member_hook<> sequencer_item; | |
227 | ||
228 | uint64_t ops, bytes; | |
229 | ||
f67539c2 | 230 | std::set<OnodeRef> onodes; ///< these onodes need to be updated/written |
7c673cae FG |
231 | KeyValueDB::Transaction t; ///< then we will commit this |
232 | Context *oncommit; ///< signal on commit | |
233 | Context *onreadable; ///< signal on readable | |
234 | Context *onreadable_sync; ///< signal on readable | |
f67539c2 TL |
235 | std::list<Context*> oncommits; ///< more commit completions |
236 | std::list<CollectionRef> removed_collections; ///< colls we removed | |
7c673cae FG |
237 | |
238 | CollectionRef first_collection; ///< first referenced collection | |
239 | utime_t start; | |
240 | explicit TransContext(OpSequencer *o) | |
241 | : state(STATE_PREPARE), | |
242 | osr(o), | |
243 | ops(0), | |
244 | bytes(0), | |
245 | oncommit(NULL), | |
246 | onreadable(NULL), | |
247 | onreadable_sync(NULL), | |
248 | start(ceph_clock_now()){ | |
249 | //cout << "txc new " << this << std::endl; | |
250 | } | |
251 | ~TransContext() { | |
252 | //cout << "txc del " << this << std::endl; | |
253 | } | |
254 | ||
255 | void write_onode(OnodeRef &o) { | |
256 | onodes.insert(o); | |
257 | } | |
258 | }; | |
259 | ||
11fdf7f2 | 260 | class OpSequencer : public RefCountedObject { |
7c673cae FG |
261 | public: |
262 | std::mutex qlock; | |
263 | std::condition_variable qcond; | |
264 | typedef boost::intrusive::list< | |
265 | TransContext, | |
266 | boost::intrusive::member_hook< | |
267 | TransContext, | |
268 | boost::intrusive::list_member_hook<>, | |
269 | &TransContext::sequencer_item> > q_list_t; | |
270 | q_list_t q; ///< transactions | |
271 | ||
11fdf7f2 TL |
272 | ~OpSequencer() { |
273 | ceph_assert(q.empty()); | |
7c673cae FG |
274 | } |
275 | ||
276 | void queue_new(TransContext *txc) { | |
277 | std::lock_guard<std::mutex> l(qlock); | |
278 | q.push_back(*txc); | |
279 | } | |
280 | ||
11fdf7f2 | 281 | void flush() { |
7c673cae FG |
282 | std::unique_lock<std::mutex> l(qlock); |
283 | while (!q.empty()) | |
284 | qcond.wait(l); | |
285 | } | |
286 | ||
11fdf7f2 | 287 | bool flush_commit(Context *c) { |
7c673cae FG |
288 | std::lock_guard<std::mutex> l(qlock); |
289 | if (q.empty()) { | |
290 | return true; | |
291 | } | |
292 | TransContext *txc = &q.back(); | |
293 | if (txc->state >= TransContext::STATE_KV_DONE) { | |
294 | return true; | |
295 | } | |
11fdf7f2 | 296 | ceph_assert(txc->state < TransContext::STATE_KV_DONE); |
7c673cae FG |
297 | txc->oncommits.push_back(c); |
298 | return false; | |
299 | } | |
300 | }; | |
301 | ||
302 | struct KVSyncThread : public Thread { | |
303 | KStore *store; | |
304 | explicit KVSyncThread(KStore *s) : store(s) {} | |
305 | void *entry() override { | |
306 | store->_kv_sync_thread(); | |
307 | return NULL; | |
308 | } | |
309 | }; | |
310 | ||
311 | // -------------------------------------------------------- | |
312 | // members | |
313 | private: | |
314 | KeyValueDB *db; | |
315 | uuid_d fsid; | |
f67539c2 | 316 | std::string basedir; |
7c673cae FG |
317 | int path_fd; ///< open handle to $path |
318 | int fsid_fd; ///< open handle (locked) to $path/fsid | |
319 | bool mounted; | |
320 | ||
9f95a23c TL |
321 | /// rwlock to protect coll_map |
322 | ceph::shared_mutex coll_lock = ceph::make_shared_mutex("KStore::coll_lock"); | |
7c673cae | 323 | ceph::unordered_map<coll_t, CollectionRef> coll_map; |
f67539c2 | 324 | std::map<coll_t,CollectionRef> new_coll_map; |
7c673cae FG |
325 | |
326 | std::mutex nid_lock; | |
327 | uint64_t nid_last; | |
328 | uint64_t nid_max; | |
329 | ||
330 | Throttle throttle_ops, throttle_bytes; ///< submit to commit | |
331 | ||
332 | Finisher finisher; | |
333 | ||
334 | KVSyncThread kv_sync_thread; | |
335 | std::mutex kv_lock; | |
336 | std::condition_variable kv_cond, kv_sync_cond; | |
337 | bool kv_stop; | |
f67539c2 | 338 | std::deque<TransContext*> kv_queue, kv_committing; |
7c673cae FG |
339 | |
340 | //Logger *logger; | |
341 | PerfCounters *logger; | |
342 | std::mutex reap_lock; | |
f67539c2 | 343 | std::list<CollectionRef> removed_collections; |
7c673cae FG |
344 | |
345 | ||
346 | // -------------------------------------------------------- | |
347 | // private methods | |
348 | ||
349 | void _init_logger(); | |
350 | void _shutdown_logger(); | |
351 | ||
352 | int _open_path(); | |
353 | void _close_path(); | |
354 | int _open_fsid(bool create); | |
355 | int _lock_fsid(); | |
356 | int _read_fsid(uuid_d *f); | |
357 | int _write_fsid(); | |
358 | void _close_fsid(); | |
359 | int _open_db(bool create); | |
360 | void _close_db(); | |
361 | int _open_collections(int *errors=0); | |
362 | void _close_collections(); | |
363 | ||
364 | int _open_super_meta(); | |
365 | ||
366 | CollectionRef _get_collection(coll_t cid); | |
367 | void _queue_reap_collection(CollectionRef& c); | |
368 | void _reap_collections(); | |
369 | ||
370 | void _assign_nid(TransContext *txc, OnodeRef o); | |
371 | ||
372 | void _dump_onode(OnodeRef o); | |
373 | ||
374 | TransContext *_txc_create(OpSequencer *osr); | |
375 | void _txc_release(TransContext *txc, uint64_t offset, uint64_t length); | |
376 | void _txc_add_transaction(TransContext *txc, Transaction *t); | |
377 | void _txc_finalize(OpSequencer *osr, TransContext *txc); | |
378 | void _txc_state_proc(TransContext *txc); | |
379 | void _txc_finish_kv(TransContext *txc); | |
380 | void _txc_finish(TransContext *txc); | |
381 | ||
382 | void _osr_reap_done(OpSequencer *osr); | |
383 | ||
384 | void _kv_sync_thread(); | |
385 | void _kv_stop() { | |
386 | { | |
387 | std::lock_guard<std::mutex> l(kv_lock); | |
388 | kv_stop = true; | |
389 | kv_cond.notify_all(); | |
390 | } | |
391 | kv_sync_thread.join(); | |
392 | kv_stop = false; | |
393 | } | |
394 | ||
f67539c2 | 395 | void _do_read_stripe(OnodeRef o, uint64_t offset, ceph::buffer::list *pbl, bool do_cache); |
7c673cae | 396 | void _do_write_stripe(TransContext *txc, OnodeRef o, |
f67539c2 | 397 | uint64_t offset, ceph::buffer::list& bl); |
7c673cae FG |
398 | void _do_remove_stripe(TransContext *txc, OnodeRef o, uint64_t offset); |
399 | ||
400 | int _collection_list( | |
401 | Collection *c, const ghobject_t& start, const ghobject_t& end, | |
f67539c2 | 402 | int max, std::vector<ghobject_t> *ls, ghobject_t *next); |
7c673cae FG |
403 | |
404 | public: | |
f67539c2 | 405 | KStore(CephContext *cct, const std::string& path); |
7c673cae FG |
406 | ~KStore() override; |
407 | ||
f67539c2 | 408 | std::string get_type() override { |
7c673cae FG |
409 | return "kstore"; |
410 | } | |
411 | ||
412 | bool needs_journal() override { return false; }; | |
413 | bool wants_journal() override { return false; }; | |
414 | bool allows_journal() override { return false; }; | |
415 | ||
f67539c2 | 416 | static int get_block_device_fsid(const std::string& path, uuid_d *fsid); |
7c673cae FG |
417 | |
418 | bool test_mount_in_use() override; | |
419 | ||
420 | int mount() override; | |
421 | int umount() override; | |
422 | void _sync(); | |
423 | ||
424 | int fsck(bool deep) override; | |
425 | ||
426 | ||
427 | int validate_hobject_key(const hobject_t &obj) const override { | |
428 | return 0; | |
429 | } | |
430 | unsigned get_max_attr_name_length() override { | |
431 | return 256; // arbitrary; there is no real limit internally | |
432 | } | |
433 | ||
434 | int mkfs() override; | |
435 | int mkjournal() override { | |
436 | return 0; | |
437 | } | |
f67539c2 | 438 | void dump_perf_counters(ceph::Formatter *f) override { |
7c673cae FG |
439 | f->open_object_section("perf_counters"); |
440 | logger->dump_formatted(f, false); | |
441 | f->close_section(); | |
442 | } | |
f67539c2 | 443 | void get_db_statistics(ceph::Formatter *f) override { |
11fdf7f2 TL |
444 | db->get_statistics(f); |
445 | } | |
446 | int statfs(struct store_statfs_t *buf, | |
447 | osd_alert_list_t* alerts = nullptr) override; | |
9f95a23c TL |
448 | int pool_statfs(uint64_t pool_id, struct store_statfs_t *buf, |
449 | bool *per_pool_omap) override; | |
11fdf7f2 TL |
450 | |
451 | CollectionHandle open_collection(const coll_t& c) override; | |
452 | CollectionHandle create_new_collection(const coll_t& c) override; | |
453 | void set_collection_commit_queue(const coll_t& cid, | |
454 | ContextQueue *commit_queue) override { | |
455 | } | |
7c673cae FG |
456 | |
457 | using ObjectStore::exists; | |
11fdf7f2 | 458 | bool exists(CollectionHandle& c, const ghobject_t& oid) override; |
7c673cae FG |
459 | using ObjectStore::stat; |
460 | int stat( | |
11fdf7f2 | 461 | CollectionHandle& c, |
7c673cae FG |
462 | const ghobject_t& oid, |
463 | struct stat *st, | |
464 | bool allow_eio = false) override; // struct stat? | |
465 | int set_collection_opts( | |
11fdf7f2 | 466 | CollectionHandle& c, |
7c673cae FG |
467 | const pool_opts_t& opts) override; |
468 | using ObjectStore::read; | |
469 | int read( | |
11fdf7f2 | 470 | CollectionHandle& c, |
7c673cae FG |
471 | const ghobject_t& oid, |
472 | uint64_t offset, | |
473 | size_t len, | |
f67539c2 | 474 | ceph::buffer::list& bl, |
224ce89b | 475 | uint32_t op_flags = 0) override; |
7c673cae FG |
476 | int _do_read( |
477 | OnodeRef o, | |
478 | uint64_t offset, | |
479 | size_t len, | |
f67539c2 | 480 | ceph::buffer::list& bl, |
9f95a23c | 481 | bool do_cache, |
7c673cae FG |
482 | uint32_t op_flags = 0); |
483 | ||
484 | using ObjectStore::fiemap; | |
f67539c2 TL |
485 | int fiemap(CollectionHandle& c, const ghobject_t& oid, uint64_t offset, size_t len, std::map<uint64_t, uint64_t>& destmap) override; |
486 | int fiemap(CollectionHandle& c, const ghobject_t& oid, uint64_t offset, size_t len, ceph::buffer::list& outbl) override; | |
7c673cae | 487 | using ObjectStore::getattr; |
f67539c2 | 488 | int getattr(CollectionHandle& c, const ghobject_t& oid, const char *name, ceph::buffer::ptr& value) override; |
7c673cae | 489 | using ObjectStore::getattrs; |
f67539c2 | 490 | int getattrs(CollectionHandle& c, const ghobject_t& oid, std::map<std::string,ceph::buffer::ptr>& aset) override; |
7c673cae | 491 | |
f67539c2 | 492 | int list_collections(std::vector<coll_t>& ls) override; |
7c673cae | 493 | bool collection_exists(const coll_t& c) override; |
11fdf7f2 TL |
494 | int collection_empty(CollectionHandle& c, bool *empty) override; |
495 | int collection_bits(CollectionHandle& c) override; | |
7c673cae FG |
496 | int collection_list( |
497 | CollectionHandle &c, const ghobject_t& start, const ghobject_t& end, | |
498 | int max, | |
f67539c2 | 499 | std::vector<ghobject_t> *ls, ghobject_t *next) override; |
7c673cae FG |
500 | |
501 | using ObjectStore::omap_get; | |
502 | int omap_get( | |
11fdf7f2 | 503 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae | 504 | const ghobject_t &oid, ///< [in] Object containing omap |
f67539c2 TL |
505 | ceph::buffer::list *header, ///< [out] omap header |
506 | std::map<std::string, ceph::buffer::list> *out /// < [out] Key to value std::map | |
7c673cae FG |
507 | ) override; |
508 | ||
509 | using ObjectStore::omap_get_header; | |
510 | /// Get omap header | |
511 | int omap_get_header( | |
11fdf7f2 | 512 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae | 513 | const ghobject_t &oid, ///< [in] Object containing omap |
f67539c2 | 514 | ceph::buffer::list *header, ///< [out] omap header |
7c673cae FG |
515 | bool allow_eio = false ///< [in] don't assert on eio |
516 | ) override; | |
517 | ||
518 | using ObjectStore::omap_get_keys; | |
519 | /// Get keys defined on oid | |
520 | int omap_get_keys( | |
11fdf7f2 | 521 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae | 522 | const ghobject_t &oid, ///< [in] Object containing omap |
f67539c2 | 523 | std::set<std::string> *keys ///< [out] Keys defined on oid |
7c673cae FG |
524 | ) override; |
525 | ||
526 | using ObjectStore::omap_get_values; | |
527 | /// Get key values | |
528 | int omap_get_values( | |
11fdf7f2 | 529 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae | 530 | const ghobject_t &oid, ///< [in] Object containing omap |
f67539c2 TL |
531 | const std::set<std::string> &keys, ///< [in] Keys to get |
532 | std::map<std::string, ceph::buffer::list> *out ///< [out] Returned keys and values | |
7c673cae FG |
533 | ) override; |
534 | ||
535 | using ObjectStore::omap_check_keys; | |
536 | /// Filters keys into out which are defined on oid | |
537 | int omap_check_keys( | |
11fdf7f2 | 538 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae | 539 | const ghobject_t &oid, ///< [in] Object containing omap |
f67539c2 TL |
540 | const std::set<std::string> &keys, ///< [in] Keys to check |
541 | std::set<std::string> *out ///< [out] Subset of keys defined on oid | |
7c673cae FG |
542 | ) override; |
543 | ||
544 | using ObjectStore::get_omap_iterator; | |
545 | ObjectMap::ObjectMapIterator get_omap_iterator( | |
11fdf7f2 | 546 | CollectionHandle& c, ///< [in] collection |
7c673cae FG |
547 | const ghobject_t &oid ///< [in] object |
548 | ) override; | |
549 | ||
550 | void set_fsid(uuid_d u) override { | |
551 | fsid = u; | |
552 | } | |
553 | uuid_d get_fsid() override { | |
554 | return fsid; | |
555 | } | |
556 | ||
557 | uint64_t estimate_objects_overhead(uint64_t num_objects) override { | |
558 | return num_objects * 300; //assuming per-object overhead is 300 bytes | |
559 | } | |
560 | ||
561 | objectstore_perf_stat_t get_cur_stats() override { | |
562 | return objectstore_perf_stat_t(); | |
563 | } | |
564 | const PerfCounters* get_perf_counters() const override { | |
565 | return logger; | |
566 | } | |
567 | ||
568 | ||
569 | int queue_transactions( | |
11fdf7f2 | 570 | CollectionHandle& ch, |
f67539c2 | 571 | std::vector<Transaction>& tls, |
7c673cae FG |
572 | TrackedOpRef op = TrackedOpRef(), |
573 | ThreadPool::TPHandle *handle = NULL) override; | |
574 | ||
224ce89b | 575 | void compact () override { |
11fdf7f2 | 576 | ceph_assert(db); |
224ce89b WB |
577 | db->compact(); |
578 | } | |
579 | ||
7c673cae FG |
580 | private: |
581 | // -------------------------------------------------------- | |
582 | // write ops | |
583 | ||
7c673cae FG |
584 | int _write(TransContext *txc, |
585 | CollectionRef& c, | |
586 | OnodeRef& o, | |
587 | uint64_t offset, size_t len, | |
f67539c2 | 588 | ceph::buffer::list& bl, |
7c673cae FG |
589 | uint32_t fadvise_flags); |
590 | int _do_write(TransContext *txc, | |
591 | OnodeRef o, | |
592 | uint64_t offset, uint64_t length, | |
f67539c2 | 593 | ceph::buffer::list& bl, |
7c673cae FG |
594 | uint32_t fadvise_flags); |
595 | int _touch(TransContext *txc, | |
596 | CollectionRef& c, | |
597 | OnodeRef& o); | |
598 | int _zero(TransContext *txc, | |
599 | CollectionRef& c, | |
600 | OnodeRef& o, | |
601 | uint64_t offset, size_t len); | |
602 | int _do_truncate(TransContext *txc, | |
603 | OnodeRef o, | |
604 | uint64_t offset); | |
605 | int _truncate(TransContext *txc, | |
606 | CollectionRef& c, | |
607 | OnodeRef& o, | |
608 | uint64_t offset); | |
609 | int _remove(TransContext *txc, | |
610 | CollectionRef& c, | |
611 | OnodeRef& o); | |
612 | int _do_remove(TransContext *txc, | |
613 | OnodeRef o); | |
614 | int _setattr(TransContext *txc, | |
615 | CollectionRef& c, | |
616 | OnodeRef& o, | |
f67539c2 TL |
617 | const std::string& name, |
618 | ceph::buffer::ptr& val); | |
7c673cae FG |
619 | int _setattrs(TransContext *txc, |
620 | CollectionRef& c, | |
621 | OnodeRef& o, | |
f67539c2 | 622 | const std::map<std::string,ceph::buffer::ptr>& aset); |
7c673cae FG |
623 | int _rmattr(TransContext *txc, |
624 | CollectionRef& c, | |
625 | OnodeRef& o, | |
f67539c2 | 626 | const std::string& name); |
7c673cae FG |
627 | int _rmattrs(TransContext *txc, |
628 | CollectionRef& c, | |
629 | OnodeRef& o); | |
630 | void _do_omap_clear(TransContext *txc, uint64_t id); | |
631 | int _omap_clear(TransContext *txc, | |
632 | CollectionRef& c, | |
633 | OnodeRef& o); | |
634 | int _omap_setkeys(TransContext *txc, | |
635 | CollectionRef& c, | |
636 | OnodeRef& o, | |
f67539c2 | 637 | ceph::buffer::list& bl); |
7c673cae FG |
638 | int _omap_setheader(TransContext *txc, |
639 | CollectionRef& c, | |
640 | OnodeRef& o, | |
f67539c2 | 641 | ceph::buffer::list& header); |
7c673cae FG |
642 | int _omap_rmkeys(TransContext *txc, |
643 | CollectionRef& c, | |
644 | OnodeRef& o, | |
f67539c2 | 645 | const ceph::buffer::list& bl); |
7c673cae FG |
646 | int _omap_rmkey_range(TransContext *txc, |
647 | CollectionRef& c, | |
648 | OnodeRef& o, | |
f67539c2 | 649 | const std::string& first, const std::string& last); |
7c673cae FG |
650 | int _setallochint(TransContext *txc, |
651 | CollectionRef& c, | |
652 | OnodeRef& o, | |
653 | uint64_t expected_object_size, | |
654 | uint64_t expected_write_size, | |
655 | uint32_t flags); | |
656 | int _clone(TransContext *txc, | |
657 | CollectionRef& c, | |
658 | OnodeRef& oldo, | |
659 | OnodeRef& newo); | |
660 | int _clone_range(TransContext *txc, | |
661 | CollectionRef& c, | |
662 | OnodeRef& oldo, | |
663 | OnodeRef& newo, | |
664 | uint64_t srcoff, uint64_t length, uint64_t dstoff); | |
665 | int _rename(TransContext *txc, | |
666 | CollectionRef& c, | |
667 | OnodeRef& oldo, | |
668 | OnodeRef& newo, | |
669 | const ghobject_t& new_oid); | |
670 | int _create_collection(TransContext *txc, coll_t cid, unsigned bits, | |
671 | CollectionRef *c); | |
672 | int _remove_collection(TransContext *txc, coll_t cid, CollectionRef *c); | |
673 | int _split_collection(TransContext *txc, | |
674 | CollectionRef& c, | |
675 | CollectionRef& d, | |
676 | unsigned bits, int rem); | |
11fdf7f2 TL |
677 | int _merge_collection(TransContext *txc, |
678 | CollectionRef *c, | |
679 | CollectionRef& d, | |
680 | unsigned bits); | |
7c673cae FG |
681 | |
682 | }; | |
683 | ||
7c673cae FG |
684 | static inline void intrusive_ptr_add_ref(KStore::Onode *o) { |
685 | o->get(); | |
686 | } | |
687 | static inline void intrusive_ptr_release(KStore::Onode *o) { | |
688 | o->put(); | |
689 | } | |
690 | ||
691 | static inline void intrusive_ptr_add_ref(KStore::OpSequencer *o) { | |
692 | o->get(); | |
693 | } | |
694 | static inline void intrusive_ptr_release(KStore::OpSequencer *o) { | |
695 | o->put(); | |
696 | } | |
697 | ||
698 | #endif |