]>
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 | #ifndef KEY_VALUE_DB_H | |
4 | #define KEY_VALUE_DB_H | |
5 | ||
6 | #include "include/buffer.h" | |
7 | #include <ostream> | |
8 | #include <set> | |
9 | #include <map> | |
10 | #include <string> | |
7c673cae FG |
11 | #include <boost/scoped_ptr.hpp> |
12 | #include "include/encoding.h" | |
13 | #include "common/Formatter.h" | |
3efd9988 | 14 | #include "common/perf_counters.h" |
91327a77 | 15 | #include "common/PriorityCache.h" |
7c673cae FG |
16 | |
17 | using std::string; | |
11fdf7f2 | 18 | using std::vector; |
7c673cae FG |
19 | /** |
20 | * Defines virtual interface to be implemented by key value store | |
21 | * | |
22 | * Kyoto Cabinet or LevelDB should implement this | |
23 | */ | |
11fdf7f2 | 24 | class KeyValueDB { |
7c673cae | 25 | public: |
11fdf7f2 TL |
26 | /* |
27 | * See RocksDB's definition of a column family(CF) and how to use it. | |
28 | * The interfaces of KeyValueDB is extended, when a column family is created. | |
29 | * Prefix will be the name of column family to use. | |
30 | */ | |
31 | struct ColumnFamily { | |
32 | string name; //< name of this individual column family | |
33 | string option; //< configure option string for this CF | |
34 | ColumnFamily(const string &name, const string &option) | |
35 | : name(name), option(option) {} | |
36 | }; | |
37 | ||
7c673cae FG |
38 | class TransactionImpl { |
39 | public: | |
40 | /// Set Keys | |
41 | void set( | |
11fdf7f2 | 42 | const std::string &prefix, ///< [in] Prefix for keys, or CF name |
9f95a23c | 43 | const std::map<std::string, ceph::buffer::list> &to_set ///< [in] keys/values to set |
7c673cae | 44 | ) { |
9f95a23c | 45 | for (auto it = to_set.cbegin(); it != to_set.cend(); ++it) |
7c673cae FG |
46 | set(prefix, it->first, it->second); |
47 | } | |
48 | ||
9f95a23c | 49 | /// Set Keys (via encoded ceph::buffer::list) |
7c673cae | 50 | void set( |
11fdf7f2 | 51 | const std::string &prefix, ///< [in] prefix, or CF name |
9f95a23c | 52 | ceph::buffer::list& to_set_bl ///< [in] encoded key/values to set |
7c673cae | 53 | ) { |
9f95a23c | 54 | using ceph::decode; |
11fdf7f2 | 55 | auto p = std::cbegin(to_set_bl); |
7c673cae | 56 | uint32_t num; |
11fdf7f2 | 57 | decode(num, p); |
7c673cae FG |
58 | while (num--) { |
59 | string key; | |
9f95a23c | 60 | ceph::buffer::list value; |
11fdf7f2 TL |
61 | decode(key, p); |
62 | decode(value, p); | |
7c673cae FG |
63 | set(prefix, key, value); |
64 | } | |
65 | } | |
66 | ||
67 | /// Set Key | |
68 | virtual void set( | |
11fdf7f2 | 69 | const std::string &prefix, ///< [in] Prefix or CF for the key |
7c673cae | 70 | const std::string &k, ///< [in] Key to set |
9f95a23c | 71 | const ceph::buffer::list &bl ///< [in] Value to set |
7c673cae FG |
72 | ) = 0; |
73 | virtual void set( | |
74 | const std::string &prefix, | |
75 | const char *k, | |
76 | size_t keylen, | |
9f95a23c | 77 | const ceph::buffer::list& bl) { |
7c673cae FG |
78 | set(prefix, string(k, keylen), bl); |
79 | } | |
80 | ||
9f95a23c | 81 | /// Removes Keys (via encoded ceph::buffer::list) |
7c673cae | 82 | void rmkeys( |
11fdf7f2 | 83 | const std::string &prefix, ///< [in] Prefix or CF to search for |
9f95a23c | 84 | ceph::buffer::list &keys_bl ///< [in] Keys to remove |
7c673cae | 85 | ) { |
9f95a23c | 86 | using ceph::decode; |
11fdf7f2 | 87 | auto p = std::cbegin(keys_bl); |
7c673cae | 88 | uint32_t num; |
11fdf7f2 | 89 | decode(num, p); |
7c673cae FG |
90 | while (num--) { |
91 | string key; | |
11fdf7f2 | 92 | decode(key, p); |
7c673cae FG |
93 | rmkey(prefix, key); |
94 | } | |
95 | } | |
96 | ||
97 | /// Removes Keys | |
98 | void rmkeys( | |
11fdf7f2 | 99 | const std::string &prefix, ///< [in] Prefix/CF to search for |
7c673cae FG |
100 | const std::set<std::string> &keys ///< [in] Keys to remove |
101 | ) { | |
9f95a23c | 102 | for (auto it = keys.cbegin(); it != keys.cend(); ++it) |
7c673cae FG |
103 | rmkey(prefix, *it); |
104 | } | |
105 | ||
106 | /// Remove Key | |
107 | virtual void rmkey( | |
11fdf7f2 TL |
108 | const std::string &prefix, ///< [in] Prefix/CF to search for |
109 | const std::string &k ///< [in] Key to remove | |
7c673cae FG |
110 | ) = 0; |
111 | virtual void rmkey( | |
112 | const std::string &prefix, ///< [in] Prefix to search for | |
113 | const char *k, ///< [in] Key to remove | |
114 | size_t keylen | |
115 | ) { | |
116 | rmkey(prefix, string(k, keylen)); | |
117 | } | |
118 | ||
119 | /// Remove Single Key which exists and was not overwritten. | |
120 | /// This API is only related to performance optimization, and should only be | |
121 | /// re-implemented by log-insert-merge tree based keyvalue stores(such as RocksDB). | |
122 | /// If a key is overwritten (by calling set multiple times), then the result | |
123 | /// of calling rm_single_key on this key is undefined. | |
124 | virtual void rm_single_key( | |
11fdf7f2 | 125 | const std::string &prefix, ///< [in] Prefix/CF to search for |
7c673cae FG |
126 | const std::string &k ///< [in] Key to remove |
127 | ) { return rmkey(prefix, k);} | |
128 | ||
129 | /// Removes keys beginning with prefix | |
130 | virtual void rmkeys_by_prefix( | |
11fdf7f2 | 131 | const std::string &prefix ///< [in] Prefix/CF by which to remove keys |
7c673cae FG |
132 | ) = 0; |
133 | ||
134 | virtual void rm_range_keys( | |
135 | const string &prefix, ///< [in] Prefix by which to remove keys | |
136 | const string &start, ///< [in] The start bound of remove keys | |
137 | const string &end ///< [in] The start bound of remove keys | |
138 | ) = 0; | |
139 | ||
140 | /// Merge value into key | |
141 | virtual void merge( | |
11fdf7f2 | 142 | const std::string &prefix, ///< [in] Prefix/CF ==> MUST match some established merge operator |
7c673cae | 143 | const std::string &key, ///< [in] Key to be merged |
9f95a23c | 144 | const ceph::buffer::list &value ///< [in] value to be merged into key |
11fdf7f2 | 145 | ) { ceph_abort_msg("Not implemented"); } |
7c673cae FG |
146 | |
147 | virtual ~TransactionImpl() {} | |
148 | }; | |
11fdf7f2 | 149 | typedef std::shared_ptr< TransactionImpl > Transaction; |
7c673cae FG |
150 | |
151 | /// create a new instance | |
152 | static KeyValueDB *create(CephContext *cct, const std::string& type, | |
153 | const std::string& dir, | |
9f95a23c | 154 | std::map<std::string,std::string> options = {}, |
7c673cae FG |
155 | void *p = NULL); |
156 | ||
157 | /// test whether we can successfully initialize; may have side effects (e.g., create) | |
158 | static int test_init(const std::string& type, const std::string& dir); | |
159 | virtual int init(string option_str="") = 0; | |
9f95a23c TL |
160 | virtual int open(std::ostream &out, const std::vector<ColumnFamily>& cfs = {}) = 0; |
161 | // std::vector cfs contains column families to be created when db is created. | |
11fdf7f2 | 162 | virtual int create_and_open(std::ostream &out, |
9f95a23c | 163 | const std::vector<ColumnFamily>& cfs = {}) = 0; |
11fdf7f2 | 164 | |
9f95a23c | 165 | virtual int open_read_only(std::ostream &out, const std::vector<ColumnFamily>& cfs = {}) { |
11fdf7f2 TL |
166 | return -ENOTSUP; |
167 | } | |
168 | ||
7c673cae FG |
169 | virtual void close() { } |
170 | ||
11fdf7f2 TL |
171 | /// Try to repair K/V database. leveldb and rocksdb require that database must be not opened. |
172 | virtual int repair(std::ostream &out) { return 0; } | |
173 | ||
7c673cae FG |
174 | virtual Transaction get_transaction() = 0; |
175 | virtual int submit_transaction(Transaction) = 0; | |
176 | virtual int submit_transaction_sync(Transaction t) { | |
177 | return submit_transaction(t); | |
178 | } | |
179 | ||
180 | /// Retrieve Keys | |
181 | virtual int get( | |
11fdf7f2 TL |
182 | const std::string &prefix, ///< [in] Prefix/CF for key |
183 | const std::set<std::string> &key, ///< [in] Key to retrieve | |
9f95a23c | 184 | std::map<std::string, ceph::buffer::list> *out ///< [out] Key value retrieved |
7c673cae | 185 | ) = 0; |
11fdf7f2 | 186 | virtual int get(const std::string &prefix, ///< [in] prefix or CF name |
7c673cae | 187 | const std::string &key, ///< [in] key |
9f95a23c | 188 | ceph::buffer::list *value) { ///< [out] value |
7c673cae FG |
189 | std::set<std::string> ks; |
190 | ks.insert(key); | |
9f95a23c | 191 | std::map<std::string,ceph::buffer::list> om; |
7c673cae FG |
192 | int r = get(prefix, ks, &om); |
193 | if (om.find(key) != om.end()) { | |
11fdf7f2 | 194 | *value = std::move(om[key]); |
7c673cae | 195 | } else { |
9f95a23c | 196 | *value = ceph::buffer::list(); |
7c673cae FG |
197 | r = -ENOENT; |
198 | } | |
199 | return r; | |
200 | } | |
201 | virtual int get(const string &prefix, | |
202 | const char *key, size_t keylen, | |
9f95a23c | 203 | ceph::buffer::list *value) { |
7c673cae FG |
204 | return get(prefix, string(key, keylen), value); |
205 | } | |
206 | ||
11fdf7f2 TL |
207 | // This superclass is used both by kv iterators *and* by the ObjectMap |
208 | // omap iterator. The class hierarchies are unfortunately tied together | |
209 | // by the legacy DBOjectMap implementation :(. | |
210 | class SimplestIteratorImpl { | |
7c673cae FG |
211 | public: |
212 | virtual int seek_to_first() = 0; | |
213 | virtual int upper_bound(const std::string &after) = 0; | |
214 | virtual int lower_bound(const std::string &to) = 0; | |
215 | virtual bool valid() = 0; | |
11fdf7f2 | 216 | virtual int next() = 0; |
7c673cae | 217 | virtual std::string key() = 0; |
9f95a23c TL |
218 | virtual std::string tail_key() { |
219 | return ""; | |
220 | } | |
221 | virtual ceph::buffer::list value() = 0; | |
7c673cae | 222 | virtual int status() = 0; |
11fdf7f2 TL |
223 | virtual ~SimplestIteratorImpl() {} |
224 | }; | |
225 | ||
226 | class IteratorImpl : public SimplestIteratorImpl { | |
227 | public: | |
228 | virtual ~IteratorImpl() {} | |
229 | virtual int seek_to_last() = 0; | |
230 | virtual int prev() = 0; | |
231 | virtual std::pair<std::string, std::string> raw_key() = 0; | |
9f95a23c TL |
232 | virtual ceph::buffer::ptr value_as_ptr() { |
233 | ceph::buffer::list bl = value(); | |
11fdf7f2 TL |
234 | if (bl.length() == 1) { |
235 | return *bl.buffers().begin(); | |
236 | } else if (bl.length() == 0) { | |
9f95a23c | 237 | return ceph::buffer::ptr(); |
11fdf7f2 TL |
238 | } else { |
239 | ceph_abort(); | |
240 | } | |
241 | } | |
7c673cae | 242 | }; |
11fdf7f2 | 243 | typedef std::shared_ptr< IteratorImpl > Iterator; |
7c673cae | 244 | |
11fdf7f2 | 245 | // This is the low-level iterator implemented by the underlying KV store. |
7c673cae FG |
246 | class WholeSpaceIteratorImpl { |
247 | public: | |
248 | virtual int seek_to_first() = 0; | |
249 | virtual int seek_to_first(const std::string &prefix) = 0; | |
250 | virtual int seek_to_last() = 0; | |
251 | virtual int seek_to_last(const std::string &prefix) = 0; | |
252 | virtual int upper_bound(const std::string &prefix, const std::string &after) = 0; | |
253 | virtual int lower_bound(const std::string &prefix, const std::string &to) = 0; | |
254 | virtual bool valid() = 0; | |
255 | virtual int next() = 0; | |
256 | virtual int prev() = 0; | |
257 | virtual std::string key() = 0; | |
258 | virtual std::pair<std::string,std::string> raw_key() = 0; | |
259 | virtual bool raw_key_is_prefixed(const std::string &prefix) = 0; | |
9f95a23c TL |
260 | virtual ceph::buffer::list value() = 0; |
261 | virtual ceph::buffer::ptr value_as_ptr() { | |
262 | ceph::buffer::list bl = value(); | |
7c673cae FG |
263 | if (bl.length()) { |
264 | return *bl.buffers().begin(); | |
265 | } else { | |
9f95a23c | 266 | return ceph::buffer::ptr(); |
7c673cae FG |
267 | } |
268 | } | |
269 | virtual int status() = 0; | |
270 | virtual size_t key_size() { | |
271 | return 0; | |
272 | } | |
273 | virtual size_t value_size() { | |
274 | return 0; | |
275 | } | |
276 | virtual ~WholeSpaceIteratorImpl() { } | |
277 | }; | |
11fdf7f2 | 278 | typedef std::shared_ptr< WholeSpaceIteratorImpl > WholeSpaceIterator; |
7c673cae | 279 | |
11fdf7f2 TL |
280 | private: |
281 | // This class filters a WholeSpaceIterator by a prefix. | |
282 | class PrefixIteratorImpl : public IteratorImpl { | |
7c673cae FG |
283 | const std::string prefix; |
284 | WholeSpaceIterator generic_iter; | |
285 | public: | |
11fdf7f2 | 286 | PrefixIteratorImpl(const std::string &prefix, WholeSpaceIterator iter) : |
7c673cae | 287 | prefix(prefix), generic_iter(iter) { } |
11fdf7f2 | 288 | ~PrefixIteratorImpl() override { } |
7c673cae FG |
289 | |
290 | int seek_to_first() override { | |
291 | return generic_iter->seek_to_first(prefix); | |
292 | } | |
11fdf7f2 | 293 | int seek_to_last() override { |
7c673cae FG |
294 | return generic_iter->seek_to_last(prefix); |
295 | } | |
296 | int upper_bound(const std::string &after) override { | |
297 | return generic_iter->upper_bound(prefix, after); | |
298 | } | |
299 | int lower_bound(const std::string &to) override { | |
300 | return generic_iter->lower_bound(prefix, to); | |
301 | } | |
302 | bool valid() override { | |
303 | if (!generic_iter->valid()) | |
304 | return false; | |
305 | return generic_iter->raw_key_is_prefixed(prefix); | |
306 | } | |
11fdf7f2 TL |
307 | int next() override { |
308 | return generic_iter->next(); | |
7c673cae | 309 | } |
11fdf7f2 TL |
310 | int prev() override { |
311 | return generic_iter->prev(); | |
7c673cae FG |
312 | } |
313 | std::string key() override { | |
314 | return generic_iter->key(); | |
315 | } | |
11fdf7f2 | 316 | std::pair<std::string, std::string> raw_key() override { |
7c673cae FG |
317 | return generic_iter->raw_key(); |
318 | } | |
9f95a23c | 319 | ceph::buffer::list value() override { |
7c673cae FG |
320 | return generic_iter->value(); |
321 | } | |
9f95a23c | 322 | ceph::buffer::ptr value_as_ptr() override { |
7c673cae FG |
323 | return generic_iter->value_as_ptr(); |
324 | } | |
325 | int status() override { | |
326 | return generic_iter->status(); | |
327 | } | |
328 | }; | |
11fdf7f2 | 329 | public: |
7c673cae | 330 | |
11fdf7f2 TL |
331 | virtual WholeSpaceIterator get_wholespace_iterator() = 0; |
332 | virtual Iterator get_iterator(const std::string &prefix) { | |
333 | return std::make_shared<PrefixIteratorImpl>( | |
334 | prefix, | |
335 | get_wholespace_iterator()); | |
336 | } | |
7c673cae | 337 | |
11fdf7f2 TL |
338 | void add_column_family(const std::string& cf_name, void *handle) { |
339 | cf_handles.insert(std::make_pair(cf_name, handle)); | |
7c673cae FG |
340 | } |
341 | ||
11fdf7f2 TL |
342 | bool is_column_family(const std::string& prefix) { |
343 | return cf_handles.count(prefix); | |
7c673cae FG |
344 | } |
345 | ||
346 | virtual uint64_t get_estimated_size(std::map<std::string,uint64_t> &extra) = 0; | |
347 | virtual int get_statfs(struct store_statfs_t *buf) { | |
348 | return -EOPNOTSUPP; | |
349 | } | |
350 | ||
31f18b77 FG |
351 | virtual int set_cache_size(uint64_t) { |
352 | return -EOPNOTSUPP; | |
353 | } | |
354 | ||
11fdf7f2 | 355 | virtual int set_cache_high_pri_pool_ratio(double ratio) { |
91327a77 AA |
356 | return -EOPNOTSUPP; |
357 | } | |
358 | ||
11fdf7f2 | 359 | virtual int64_t get_cache_usage() const { |
91327a77 AA |
360 | return -EOPNOTSUPP; |
361 | } | |
362 | ||
11fdf7f2 TL |
363 | virtual std::shared_ptr<PriorityCache::PriCache> get_priority_cache() const { |
364 | return nullptr; | |
91327a77 AA |
365 | } |
366 | ||
11fdf7f2 | 367 | virtual ~KeyValueDB() {} |
91327a77 | 368 | |
11fdf7f2 | 369 | /// estimate space utilization for a prefix (in bytes) |
9f95a23c TL |
370 | virtual int64_t estimate_prefix_size(const string& prefix, |
371 | const string& key_prefix) { | |
11fdf7f2 | 372 | return 0; |
91327a77 AA |
373 | } |
374 | ||
7c673cae FG |
375 | /// compact the underlying store |
376 | virtual void compact() {} | |
377 | ||
11fdf7f2 TL |
378 | /// compact the underlying store in async mode |
379 | virtual void compact_async() {} | |
380 | ||
7c673cae FG |
381 | /// compact db for all keys with a given prefix |
382 | virtual void compact_prefix(const std::string& prefix) {} | |
383 | /// compact db for all keys with a given prefix, async | |
384 | virtual void compact_prefix_async(const std::string& prefix) {} | |
385 | virtual void compact_range(const std::string& prefix, | |
386 | const std::string& start, const std::string& end) {} | |
387 | virtual void compact_range_async(const std::string& prefix, | |
388 | const std::string& start, const std::string& end) {} | |
389 | ||
390 | // See RocksDB merge operator definition, we support the basic | |
391 | // associative merge only right now. | |
392 | class MergeOperator { | |
393 | public: | |
394 | /// Merge into a key that doesn't exist | |
395 | virtual void merge_nonexistent( | |
396 | const char *rdata, size_t rlen, | |
397 | std::string *new_value) = 0; | |
398 | /// Merge into a key that does exist | |
399 | virtual void merge( | |
400 | const char *ldata, size_t llen, | |
401 | const char *rdata, size_t rlen, | |
402 | std::string *new_value) = 0; | |
403 | /// We use each operator name and each prefix to construct the overall RocksDB operator name for consistency check at open time. | |
91327a77 | 404 | virtual const char *name() const = 0; |
7c673cae FG |
405 | |
406 | virtual ~MergeOperator() {} | |
407 | }; | |
408 | ||
409 | /// Setup one or more operators, this needs to be done BEFORE the DB is opened. | |
410 | virtual int set_merge_operator(const std::string& prefix, | |
411 | std::shared_ptr<MergeOperator> mop) { | |
412 | return -EOPNOTSUPP; | |
413 | } | |
414 | ||
9f95a23c | 415 | virtual void get_statistics(ceph::Formatter *f) { |
7c673cae FG |
416 | return; |
417 | } | |
3efd9988 FG |
418 | |
419 | /** | |
420 | * Return your perf counters if you have any. Subclasses are not | |
421 | * required to implement this, and callers must respect a null return | |
422 | * value. | |
423 | */ | |
424 | virtual PerfCounters *get_perf_counters() { | |
425 | return nullptr; | |
426 | } | |
9f95a23c TL |
427 | |
428 | /** | |
429 | * Access implementation specific integral property corresponding | |
430 | * to passed property and prefic. | |
431 | * Return value is true if property is valid for prefix, populates out. | |
432 | */ | |
433 | virtual bool get_property( | |
434 | const std::string &property, | |
435 | uint64_t *out) { | |
436 | return false; | |
437 | } | |
7c673cae | 438 | protected: |
11fdf7f2 | 439 | /// List of matching prefixes/ColumnFamilies and merge operators |
7c673cae FG |
440 | std::vector<std::pair<std::string, |
441 | std::shared_ptr<MergeOperator> > > merge_ops; | |
442 | ||
11fdf7f2 TL |
443 | /// column families in use, name->handle |
444 | std::unordered_map<std::string, void *> cf_handles; | |
7c673cae FG |
445 | }; |
446 | ||
447 | #endif |