]>
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) 2013- Sage Weil <sage@inktank.com> | |
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 | ||
16 | #ifndef CEPH_MEMSTORE_H | |
17 | #define CEPH_MEMSTORE_H | |
18 | ||
19 | #include <mutex> | |
20 | #include <boost/intrusive_ptr.hpp> | |
21 | ||
22 | #include "include/unordered_map.h" | |
7c673cae FG |
23 | #include "common/Finisher.h" |
24 | #include "common/RefCountedObj.h" | |
25 | #include "common/RWLock.h" | |
26 | #include "os/ObjectStore.h" | |
27 | #include "PageSet.h" | |
11fdf7f2 | 28 | #include "include/ceph_assert.h" |
7c673cae FG |
29 | |
30 | class MemStore : public ObjectStore { | |
31 | public: | |
32 | struct Object : public RefCountedObject { | |
11fdf7f2 TL |
33 | ceph::mutex xattr_mutex{ceph::make_mutex("MemStore::Object::xattr_mutex")}; |
34 | ceph::mutex omap_mutex{ceph::make_mutex("MemStore::Object::omap_mutex")}; | |
7c673cae FG |
35 | map<string,bufferptr> xattr; |
36 | bufferlist omap_header; | |
37 | map<string,bufferlist> omap; | |
38 | ||
39 | typedef boost::intrusive_ptr<Object> Ref; | |
40 | friend void intrusive_ptr_add_ref(Object *o) { o->get(); } | |
41 | friend void intrusive_ptr_release(Object *o) { o->put(); } | |
42 | ||
43 | Object() : RefCountedObject(nullptr, 0) {} | |
44 | // interface for object data | |
45 | virtual size_t get_size() const = 0; | |
46 | virtual int read(uint64_t offset, uint64_t len, bufferlist &bl) = 0; | |
47 | virtual int write(uint64_t offset, const bufferlist &bl) = 0; | |
48 | virtual int clone(Object *src, uint64_t srcoff, uint64_t len, | |
49 | uint64_t dstoff) = 0; | |
50 | virtual int truncate(uint64_t offset) = 0; | |
51 | virtual void encode(bufferlist& bl) const = 0; | |
11fdf7f2 | 52 | virtual void decode(bufferlist::const_iterator& p) = 0; |
7c673cae FG |
53 | |
54 | void encode_base(bufferlist& bl) const { | |
11fdf7f2 TL |
55 | using ceph::encode; |
56 | encode(xattr, bl); | |
57 | encode(omap_header, bl); | |
58 | encode(omap, bl); | |
7c673cae | 59 | } |
11fdf7f2 TL |
60 | void decode_base(bufferlist::const_iterator& p) { |
61 | using ceph::decode; | |
62 | decode(xattr, p); | |
63 | decode(omap_header, p); | |
64 | decode(omap, p); | |
7c673cae FG |
65 | } |
66 | ||
67 | void dump(Formatter *f) const { | |
68 | f->dump_int("data_len", get_size()); | |
69 | f->dump_int("omap_header_len", omap_header.length()); | |
70 | ||
71 | f->open_array_section("xattrs"); | |
72 | for (map<string,bufferptr>::const_iterator p = xattr.begin(); | |
73 | p != xattr.end(); | |
74 | ++p) { | |
75 | f->open_object_section("xattr"); | |
76 | f->dump_string("name", p->first); | |
77 | f->dump_int("length", p->second.length()); | |
78 | f->close_section(); | |
79 | } | |
80 | f->close_section(); | |
81 | ||
82 | f->open_array_section("omap"); | |
83 | for (map<string,bufferlist>::const_iterator p = omap.begin(); | |
84 | p != omap.end(); | |
85 | ++p) { | |
86 | f->open_object_section("pair"); | |
87 | f->dump_string("key", p->first); | |
88 | f->dump_int("length", p->second.length()); | |
89 | f->close_section(); | |
90 | } | |
91 | f->close_section(); | |
92 | } | |
93 | }; | |
94 | typedef Object::Ref ObjectRef; | |
95 | ||
96 | struct PageSetObject; | |
97 | struct Collection : public CollectionImpl { | |
11fdf7f2 | 98 | int bits = 0; |
7c673cae FG |
99 | CephContext *cct; |
100 | bool use_page_set; | |
101 | ceph::unordered_map<ghobject_t, ObjectRef> object_hash; ///< for lookup | |
102 | map<ghobject_t, ObjectRef> object_map; ///< for iteration | |
103 | map<string,bufferptr> xattr; | |
11fdf7f2 TL |
104 | /// for object_{map,hash} |
105 | ceph::shared_mutex lock{ | |
106 | ceph::make_shared_mutex("MemStore::Collection::lock", true, false)}; | |
107 | ||
108 | bool exists = true; | |
109 | ceph::mutex sequencer_mutex{ | |
110 | ceph::make_mutex("MemStore::Collection::sequencer_mutex")}; | |
7c673cae FG |
111 | |
112 | typedef boost::intrusive_ptr<Collection> Ref; | |
113 | friend void intrusive_ptr_add_ref(Collection *c) { c->get(); } | |
114 | friend void intrusive_ptr_release(Collection *c) { c->put(); } | |
115 | ||
7c673cae FG |
116 | ObjectRef create_object() const; |
117 | ||
118 | // NOTE: The lock only needs to protect the object_map/hash, not the | |
119 | // contents of individual objects. The osd is already sequencing | |
120 | // reads and writes, so we will never see them concurrently at this | |
121 | // level. | |
122 | ||
123 | ObjectRef get_object(ghobject_t oid) { | |
11fdf7f2 | 124 | std::shared_lock l{lock}; |
7c673cae FG |
125 | auto o = object_hash.find(oid); |
126 | if (o == object_hash.end()) | |
127 | return ObjectRef(); | |
128 | return o->second; | |
129 | } | |
130 | ||
131 | ObjectRef get_or_create_object(ghobject_t oid) { | |
11fdf7f2 | 132 | std::lock_guard l{lock}; |
7c673cae FG |
133 | auto result = object_hash.emplace(oid, ObjectRef()); |
134 | if (result.second) | |
135 | object_map[oid] = result.first->second = create_object(); | |
136 | return result.first->second; | |
137 | } | |
138 | ||
139 | void encode(bufferlist& bl) const { | |
140 | ENCODE_START(1, 1, bl); | |
11fdf7f2 TL |
141 | encode(xattr, bl); |
142 | encode(use_page_set, bl); | |
7c673cae | 143 | uint32_t s = object_map.size(); |
11fdf7f2 | 144 | encode(s, bl); |
7c673cae FG |
145 | for (map<ghobject_t, ObjectRef>::const_iterator p = object_map.begin(); |
146 | p != object_map.end(); | |
147 | ++p) { | |
11fdf7f2 | 148 | encode(p->first, bl); |
7c673cae FG |
149 | p->second->encode(bl); |
150 | } | |
151 | ENCODE_FINISH(bl); | |
152 | } | |
11fdf7f2 | 153 | void decode(bufferlist::const_iterator& p) { |
7c673cae | 154 | DECODE_START(1, p); |
11fdf7f2 TL |
155 | decode(xattr, p); |
156 | decode(use_page_set, p); | |
7c673cae | 157 | uint32_t s; |
11fdf7f2 | 158 | decode(s, p); |
7c673cae FG |
159 | while (s--) { |
160 | ghobject_t k; | |
11fdf7f2 | 161 | decode(k, p); |
7c673cae FG |
162 | auto o = create_object(); |
163 | o->decode(p); | |
164 | object_map.insert(make_pair(k, o)); | |
165 | object_hash.insert(make_pair(k, o)); | |
166 | } | |
167 | DECODE_FINISH(p); | |
168 | } | |
169 | ||
170 | uint64_t used_bytes() const { | |
171 | uint64_t result = 0; | |
172 | for (map<ghobject_t, ObjectRef>::const_iterator p = object_map.begin(); | |
173 | p != object_map.end(); | |
174 | ++p) { | |
175 | result += p->second->get_size(); | |
176 | } | |
177 | ||
178 | return result; | |
179 | } | |
180 | ||
11fdf7f2 TL |
181 | void flush() override { |
182 | } | |
183 | bool flush_commit(Context *c) override { | |
184 | return true; | |
185 | } | |
186 | ||
7c673cae | 187 | explicit Collection(CephContext *cct, coll_t c) |
11fdf7f2 | 188 | : CollectionImpl(c), |
7c673cae | 189 | cct(cct), |
11fdf7f2 | 190 | use_page_set(cct->_conf->memstore_page_set) {} |
7c673cae FG |
191 | }; |
192 | typedef Collection::Ref CollectionRef; | |
193 | ||
194 | private: | |
195 | class OmapIteratorImpl; | |
196 | ||
197 | ||
198 | ceph::unordered_map<coll_t, CollectionRef> coll_map; | |
11fdf7f2 TL |
199 | /// rwlock to protect coll_map |
200 | ceph::shared_mutex coll_lock{ | |
201 | ceph::make_shared_mutex("MemStore::coll_lock")}; | |
202 | map<coll_t,CollectionRef> new_coll_map; | |
7c673cae FG |
203 | |
204 | CollectionRef get_collection(const coll_t& cid); | |
205 | ||
206 | Finisher finisher; | |
207 | ||
208 | uint64_t used_bytes; | |
209 | ||
210 | void _do_transaction(Transaction& t); | |
211 | ||
212 | int _touch(const coll_t& cid, const ghobject_t& oid); | |
213 | int _write(const coll_t& cid, const ghobject_t& oid, uint64_t offset, size_t len, | |
214 | const bufferlist& bl, uint32_t fadvise_flags = 0); | |
215 | int _zero(const coll_t& cid, const ghobject_t& oid, uint64_t offset, size_t len); | |
216 | int _truncate(const coll_t& cid, const ghobject_t& oid, uint64_t size); | |
217 | int _remove(const coll_t& cid, const ghobject_t& oid); | |
218 | int _setattrs(const coll_t& cid, const ghobject_t& oid, map<string,bufferptr>& aset); | |
219 | int _rmattr(const coll_t& cid, const ghobject_t& oid, const char *name); | |
220 | int _rmattrs(const coll_t& cid, const ghobject_t& oid); | |
221 | int _clone(const coll_t& cid, const ghobject_t& oldoid, const ghobject_t& newoid); | |
222 | int _clone_range(const coll_t& cid, const ghobject_t& oldoid, | |
223 | const ghobject_t& newoid, | |
224 | uint64_t srcoff, uint64_t len, uint64_t dstoff); | |
225 | int _omap_clear(const coll_t& cid, const ghobject_t &oid); | |
226 | int _omap_setkeys(const coll_t& cid, const ghobject_t &oid, bufferlist& aset_bl); | |
227 | int _omap_rmkeys(const coll_t& cid, const ghobject_t &oid, bufferlist& keys_bl); | |
228 | int _omap_rmkeyrange(const coll_t& cid, const ghobject_t &oid, | |
229 | const string& first, const string& last); | |
230 | int _omap_setheader(const coll_t& cid, const ghobject_t &oid, const bufferlist &bl); | |
231 | ||
232 | int _collection_hint_expected_num_objs(const coll_t& cid, uint32_t pg_num, | |
233 | uint64_t num_objs) const { return 0; } | |
234 | int _create_collection(const coll_t& c, int bits); | |
235 | int _destroy_collection(const coll_t& c); | |
236 | int _collection_add(const coll_t& cid, const coll_t& ocid, const ghobject_t& oid); | |
237 | int _collection_move_rename(const coll_t& oldcid, const ghobject_t& oldoid, | |
238 | coll_t cid, const ghobject_t& o); | |
239 | int _split_collection(const coll_t& cid, uint32_t bits, uint32_t rem, coll_t dest); | |
11fdf7f2 | 240 | int _merge_collection(const coll_t& cid, uint32_t bits, coll_t dest); |
7c673cae FG |
241 | |
242 | int _save(); | |
243 | int _load(); | |
244 | ||
245 | void dump(Formatter *f); | |
246 | void dump_all(); | |
247 | ||
248 | public: | |
249 | MemStore(CephContext *cct, const string& path) | |
250 | : ObjectStore(cct, path), | |
7c673cae FG |
251 | finisher(cct), |
252 | used_bytes(0) {} | |
253 | ~MemStore() override { } | |
254 | ||
255 | string get_type() override { | |
256 | return "memstore"; | |
257 | } | |
258 | ||
259 | bool test_mount_in_use() override { | |
260 | return false; | |
261 | } | |
262 | ||
263 | int mount() override; | |
264 | int umount() override; | |
265 | ||
266 | int fsck(bool deep) override { | |
267 | return 0; | |
268 | } | |
269 | ||
270 | int validate_hobject_key(const hobject_t &obj) const override { | |
271 | return 0; | |
272 | } | |
273 | unsigned get_max_attr_name_length() override { | |
274 | return 256; // arbitrary; there is no real limit internally | |
275 | } | |
276 | ||
277 | int mkfs() override; | |
278 | int mkjournal() override { | |
279 | return 0; | |
280 | } | |
281 | bool wants_journal() override { | |
282 | return false; | |
283 | } | |
284 | bool allows_journal() override { | |
285 | return false; | |
286 | } | |
287 | bool needs_journal() override { | |
288 | return false; | |
289 | } | |
290 | ||
11fdf7f2 TL |
291 | int get_devices(set<string> *ls) override { |
292 | // no devices for us! | |
293 | return 0; | |
294 | } | |
295 | ||
296 | int statfs(struct store_statfs_t *buf, | |
297 | osd_alert_list_t* alerts = nullptr) override; | |
298 | int pool_statfs(uint64_t pool_id, struct store_statfs_t *buf) override; | |
7c673cae | 299 | |
7c673cae | 300 | bool exists(CollectionHandle &c, const ghobject_t& oid) override; |
7c673cae FG |
301 | int stat(CollectionHandle &c, const ghobject_t& oid, |
302 | struct stat *st, bool allow_eio = false) override; | |
303 | int set_collection_opts( | |
11fdf7f2 | 304 | CollectionHandle& c, |
7c673cae | 305 | const pool_opts_t& opts) override; |
7c673cae FG |
306 | int read( |
307 | CollectionHandle &c, | |
308 | const ghobject_t& oid, | |
309 | uint64_t offset, | |
310 | size_t len, | |
311 | bufferlist& bl, | |
224ce89b | 312 | uint32_t op_flags = 0) override; |
7c673cae | 313 | using ObjectStore::fiemap; |
11fdf7f2 TL |
314 | int fiemap(CollectionHandle& c, const ghobject_t& oid, |
315 | uint64_t offset, size_t len, bufferlist& bl) override; | |
316 | int fiemap(CollectionHandle& c, const ghobject_t& oid, uint64_t offset, | |
317 | size_t len, map<uint64_t, uint64_t>& destmap) override; | |
7c673cae FG |
318 | int getattr(CollectionHandle &c, const ghobject_t& oid, const char *name, |
319 | bufferptr& value) override; | |
7c673cae FG |
320 | int getattrs(CollectionHandle &c, const ghobject_t& oid, |
321 | map<string,bufferptr>& aset) override; | |
322 | ||
323 | int list_collections(vector<coll_t>& ls) override; | |
324 | ||
325 | CollectionHandle open_collection(const coll_t& c) override { | |
326 | return get_collection(c); | |
327 | } | |
11fdf7f2 TL |
328 | CollectionHandle create_new_collection(const coll_t& c) override; |
329 | ||
330 | void set_collection_commit_queue(const coll_t& cid, | |
331 | ContextQueue *commit_queue) override { | |
332 | } | |
333 | ||
7c673cae | 334 | bool collection_exists(const coll_t& c) override; |
11fdf7f2 TL |
335 | int collection_empty(CollectionHandle& c, bool *empty) override; |
336 | int collection_bits(CollectionHandle& c) override; | |
337 | int collection_list(CollectionHandle& cid, | |
7c673cae FG |
338 | const ghobject_t& start, const ghobject_t& end, int max, |
339 | vector<ghobject_t> *ls, ghobject_t *next) override; | |
340 | ||
341 | using ObjectStore::omap_get; | |
342 | int omap_get( | |
11fdf7f2 | 343 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae FG |
344 | const ghobject_t &oid, ///< [in] Object containing omap |
345 | bufferlist *header, ///< [out] omap header | |
346 | map<string, bufferlist> *out /// < [out] Key to value map | |
347 | ) override; | |
348 | ||
349 | using ObjectStore::omap_get_header; | |
350 | /// Get omap header | |
351 | int omap_get_header( | |
11fdf7f2 | 352 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae FG |
353 | const ghobject_t &oid, ///< [in] Object containing omap |
354 | bufferlist *header, ///< [out] omap header | |
355 | bool allow_eio = false ///< [in] don't assert on eio | |
356 | ) override; | |
357 | ||
358 | using ObjectStore::omap_get_keys; | |
359 | /// Get keys defined on oid | |
360 | int omap_get_keys( | |
11fdf7f2 | 361 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae FG |
362 | const ghobject_t &oid, ///< [in] Object containing omap |
363 | set<string> *keys ///< [out] Keys defined on oid | |
364 | ) override; | |
365 | ||
366 | using ObjectStore::omap_get_values; | |
367 | /// Get key values | |
368 | int omap_get_values( | |
11fdf7f2 | 369 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae FG |
370 | const ghobject_t &oid, ///< [in] Object containing omap |
371 | const set<string> &keys, ///< [in] Keys to get | |
372 | map<string, bufferlist> *out ///< [out] Returned keys and values | |
373 | ) override; | |
374 | ||
375 | using ObjectStore::omap_check_keys; | |
376 | /// Filters keys into out which are defined on oid | |
377 | int omap_check_keys( | |
11fdf7f2 | 378 | CollectionHandle& c, ///< [in] Collection containing oid |
7c673cae FG |
379 | const ghobject_t &oid, ///< [in] Object containing omap |
380 | const set<string> &keys, ///< [in] Keys to check | |
381 | set<string> *out ///< [out] Subset of keys defined on oid | |
382 | ) override; | |
383 | ||
384 | using ObjectStore::get_omap_iterator; | |
385 | ObjectMap::ObjectMapIterator get_omap_iterator( | |
11fdf7f2 | 386 | CollectionHandle& c, ///< [in] collection |
7c673cae FG |
387 | const ghobject_t &oid ///< [in] object |
388 | ) override; | |
389 | ||
390 | void set_fsid(uuid_d u) override; | |
391 | uuid_d get_fsid() override; | |
392 | ||
393 | uint64_t estimate_objects_overhead(uint64_t num_objects) override { | |
394 | return 0; //do not care | |
395 | } | |
396 | ||
397 | objectstore_perf_stat_t get_cur_stats() override; | |
398 | ||
399 | const PerfCounters* get_perf_counters() const override { | |
400 | return nullptr; | |
401 | } | |
402 | ||
403 | ||
404 | int queue_transactions( | |
11fdf7f2 TL |
405 | CollectionHandle& ch, |
406 | vector<Transaction>& tls, | |
7c673cae FG |
407 | TrackedOpRef op = TrackedOpRef(), |
408 | ThreadPool::TPHandle *handle = NULL) override; | |
409 | }; | |
410 | ||
411 | ||
412 | ||
413 | ||
414 | #endif |