]>
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> | |
33c7a0ef | 10 | #include <optional> |
7c673cae | 11 | #include <string> |
7c673cae FG |
12 | #include <boost/scoped_ptr.hpp> |
13 | #include "include/encoding.h" | |
14 | #include "common/Formatter.h" | |
3efd9988 | 15 | #include "common/perf_counters.h" |
91327a77 | 16 | #include "common/PriorityCache.h" |
7c673cae | 17 | |
7c673cae FG |
18 | /** |
19 | * Defines virtual interface to be implemented by key value store | |
20 | * | |
20effc67 | 21 | * Kyoto Cabinet should implement this |
7c673cae | 22 | */ |
11fdf7f2 | 23 | class KeyValueDB { |
7c673cae FG |
24 | public: |
25 | class TransactionImpl { | |
26 | public: | |
27 | /// Set Keys | |
28 | void set( | |
11fdf7f2 | 29 | const std::string &prefix, ///< [in] Prefix for keys, or CF name |
9f95a23c | 30 | const std::map<std::string, ceph::buffer::list> &to_set ///< [in] keys/values to set |
7c673cae | 31 | ) { |
9f95a23c | 32 | for (auto it = to_set.cbegin(); it != to_set.cend(); ++it) |
7c673cae FG |
33 | set(prefix, it->first, it->second); |
34 | } | |
35 | ||
9f95a23c | 36 | /// Set Keys (via encoded ceph::buffer::list) |
7c673cae | 37 | void set( |
11fdf7f2 | 38 | const std::string &prefix, ///< [in] prefix, or CF name |
9f95a23c | 39 | ceph::buffer::list& to_set_bl ///< [in] encoded key/values to set |
7c673cae | 40 | ) { |
9f95a23c | 41 | using ceph::decode; |
11fdf7f2 | 42 | auto p = std::cbegin(to_set_bl); |
7c673cae | 43 | uint32_t num; |
11fdf7f2 | 44 | decode(num, p); |
7c673cae | 45 | while (num--) { |
f67539c2 | 46 | std::string key; |
9f95a23c | 47 | ceph::buffer::list value; |
11fdf7f2 TL |
48 | decode(key, p); |
49 | decode(value, p); | |
7c673cae FG |
50 | set(prefix, key, value); |
51 | } | |
52 | } | |
53 | ||
54 | /// Set Key | |
55 | virtual void set( | |
11fdf7f2 | 56 | const std::string &prefix, ///< [in] Prefix or CF for the key |
7c673cae | 57 | const std::string &k, ///< [in] Key to set |
9f95a23c | 58 | const ceph::buffer::list &bl ///< [in] Value to set |
7c673cae FG |
59 | ) = 0; |
60 | virtual void set( | |
61 | const std::string &prefix, | |
62 | const char *k, | |
63 | size_t keylen, | |
9f95a23c | 64 | const ceph::buffer::list& bl) { |
f67539c2 | 65 | set(prefix, std::string(k, keylen), bl); |
7c673cae FG |
66 | } |
67 | ||
9f95a23c | 68 | /// Removes Keys (via encoded ceph::buffer::list) |
7c673cae | 69 | void rmkeys( |
11fdf7f2 | 70 | const std::string &prefix, ///< [in] Prefix or CF to search for |
9f95a23c | 71 | ceph::buffer::list &keys_bl ///< [in] Keys to remove |
7c673cae | 72 | ) { |
9f95a23c | 73 | using ceph::decode; |
11fdf7f2 | 74 | auto p = std::cbegin(keys_bl); |
7c673cae | 75 | uint32_t num; |
11fdf7f2 | 76 | decode(num, p); |
7c673cae | 77 | while (num--) { |
f67539c2 | 78 | std::string key; |
11fdf7f2 | 79 | decode(key, p); |
7c673cae FG |
80 | rmkey(prefix, key); |
81 | } | |
82 | } | |
83 | ||
84 | /// Removes Keys | |
85 | void rmkeys( | |
11fdf7f2 | 86 | const std::string &prefix, ///< [in] Prefix/CF to search for |
7c673cae FG |
87 | const std::set<std::string> &keys ///< [in] Keys to remove |
88 | ) { | |
9f95a23c | 89 | for (auto it = keys.cbegin(); it != keys.cend(); ++it) |
7c673cae FG |
90 | rmkey(prefix, *it); |
91 | } | |
92 | ||
93 | /// Remove Key | |
94 | virtual void rmkey( | |
11fdf7f2 TL |
95 | const std::string &prefix, ///< [in] Prefix/CF to search for |
96 | const std::string &k ///< [in] Key to remove | |
7c673cae FG |
97 | ) = 0; |
98 | virtual void rmkey( | |
99 | const std::string &prefix, ///< [in] Prefix to search for | |
100 | const char *k, ///< [in] Key to remove | |
101 | size_t keylen | |
102 | ) { | |
f67539c2 | 103 | rmkey(prefix, std::string(k, keylen)); |
7c673cae FG |
104 | } |
105 | ||
106 | /// Remove Single Key which exists and was not overwritten. | |
107 | /// This API is only related to performance optimization, and should only be | |
108 | /// re-implemented by log-insert-merge tree based keyvalue stores(such as RocksDB). | |
109 | /// If a key is overwritten (by calling set multiple times), then the result | |
110 | /// of calling rm_single_key on this key is undefined. | |
111 | virtual void rm_single_key( | |
11fdf7f2 | 112 | const std::string &prefix, ///< [in] Prefix/CF to search for |
7c673cae FG |
113 | const std::string &k ///< [in] Key to remove |
114 | ) { return rmkey(prefix, k);} | |
115 | ||
116 | /// Removes keys beginning with prefix | |
117 | virtual void rmkeys_by_prefix( | |
11fdf7f2 | 118 | const std::string &prefix ///< [in] Prefix/CF by which to remove keys |
7c673cae FG |
119 | ) = 0; |
120 | ||
121 | virtual void rm_range_keys( | |
f67539c2 TL |
122 | const std::string &prefix, ///< [in] Prefix by which to remove keys |
123 | const std::string &start, ///< [in] The start bound of remove keys | |
124 | const std::string &end ///< [in] The start bound of remove keys | |
7c673cae FG |
125 | ) = 0; |
126 | ||
127 | /// Merge value into key | |
128 | virtual void merge( | |
11fdf7f2 | 129 | const std::string &prefix, ///< [in] Prefix/CF ==> MUST match some established merge operator |
7c673cae | 130 | const std::string &key, ///< [in] Key to be merged |
9f95a23c | 131 | const ceph::buffer::list &value ///< [in] value to be merged into key |
11fdf7f2 | 132 | ) { ceph_abort_msg("Not implemented"); } |
7c673cae FG |
133 | |
134 | virtual ~TransactionImpl() {} | |
135 | }; | |
11fdf7f2 | 136 | typedef std::shared_ptr< TransactionImpl > Transaction; |
7c673cae FG |
137 | |
138 | /// create a new instance | |
139 | static KeyValueDB *create(CephContext *cct, const std::string& type, | |
140 | const std::string& dir, | |
9f95a23c | 141 | std::map<std::string,std::string> options = {}, |
7c673cae FG |
142 | void *p = NULL); |
143 | ||
144 | /// test whether we can successfully initialize; may have side effects (e.g., create) | |
145 | static int test_init(const std::string& type, const std::string& dir); | |
f67539c2 TL |
146 | virtual int init(std::string option_str="") = 0; |
147 | virtual int open(std::ostream &out, const std::string& cfs="") = 0; | |
9f95a23c | 148 | // std::vector cfs contains column families to be created when db is created. |
f67539c2 | 149 | virtual int create_and_open(std::ostream &out, const std::string& cfs="") = 0; |
11fdf7f2 | 150 | |
f67539c2 | 151 | virtual int open_read_only(std::ostream &out, const std::string& cfs="") { |
11fdf7f2 TL |
152 | return -ENOTSUP; |
153 | } | |
154 | ||
7c673cae FG |
155 | virtual void close() { } |
156 | ||
20effc67 | 157 | /// Try to repair K/V database. rocksdb requires that database must be not opened. |
11fdf7f2 TL |
158 | virtual int repair(std::ostream &out) { return 0; } |
159 | ||
7c673cae FG |
160 | virtual Transaction get_transaction() = 0; |
161 | virtual int submit_transaction(Transaction) = 0; | |
162 | virtual int submit_transaction_sync(Transaction t) { | |
163 | return submit_transaction(t); | |
164 | } | |
165 | ||
166 | /// Retrieve Keys | |
167 | virtual int get( | |
11fdf7f2 TL |
168 | const std::string &prefix, ///< [in] Prefix/CF for key |
169 | const std::set<std::string> &key, ///< [in] Key to retrieve | |
9f95a23c | 170 | std::map<std::string, ceph::buffer::list> *out ///< [out] Key value retrieved |
7c673cae | 171 | ) = 0; |
11fdf7f2 | 172 | virtual int get(const std::string &prefix, ///< [in] prefix or CF name |
7c673cae | 173 | const std::string &key, ///< [in] key |
9f95a23c | 174 | ceph::buffer::list *value) { ///< [out] value |
7c673cae FG |
175 | std::set<std::string> ks; |
176 | ks.insert(key); | |
9f95a23c | 177 | std::map<std::string,ceph::buffer::list> om; |
7c673cae FG |
178 | int r = get(prefix, ks, &om); |
179 | if (om.find(key) != om.end()) { | |
11fdf7f2 | 180 | *value = std::move(om[key]); |
7c673cae | 181 | } else { |
9f95a23c | 182 | *value = ceph::buffer::list(); |
7c673cae FG |
183 | r = -ENOENT; |
184 | } | |
185 | return r; | |
186 | } | |
f67539c2 | 187 | virtual int get(const std::string &prefix, |
7c673cae | 188 | const char *key, size_t keylen, |
9f95a23c | 189 | ceph::buffer::list *value) { |
f67539c2 | 190 | return get(prefix, std::string(key, keylen), value); |
7c673cae FG |
191 | } |
192 | ||
11fdf7f2 TL |
193 | // This superclass is used both by kv iterators *and* by the ObjectMap |
194 | // omap iterator. The class hierarchies are unfortunately tied together | |
195 | // by the legacy DBOjectMap implementation :(. | |
196 | class SimplestIteratorImpl { | |
7c673cae FG |
197 | public: |
198 | virtual int seek_to_first() = 0; | |
199 | virtual int upper_bound(const std::string &after) = 0; | |
200 | virtual int lower_bound(const std::string &to) = 0; | |
201 | virtual bool valid() = 0; | |
11fdf7f2 | 202 | virtual int next() = 0; |
7c673cae | 203 | virtual std::string key() = 0; |
9f95a23c TL |
204 | virtual std::string tail_key() { |
205 | return ""; | |
206 | } | |
207 | virtual ceph::buffer::list value() = 0; | |
7c673cae | 208 | virtual int status() = 0; |
11fdf7f2 TL |
209 | virtual ~SimplestIteratorImpl() {} |
210 | }; | |
211 | ||
212 | class IteratorImpl : public SimplestIteratorImpl { | |
213 | public: | |
214 | virtual ~IteratorImpl() {} | |
215 | virtual int seek_to_last() = 0; | |
216 | virtual int prev() = 0; | |
217 | virtual std::pair<std::string, std::string> raw_key() = 0; | |
9f95a23c TL |
218 | virtual ceph::buffer::ptr value_as_ptr() { |
219 | ceph::buffer::list bl = value(); | |
11fdf7f2 TL |
220 | if (bl.length() == 1) { |
221 | return *bl.buffers().begin(); | |
222 | } else if (bl.length() == 0) { | |
9f95a23c | 223 | return ceph::buffer::ptr(); |
11fdf7f2 TL |
224 | } else { |
225 | ceph_abort(); | |
226 | } | |
227 | } | |
7c673cae | 228 | }; |
11fdf7f2 | 229 | typedef std::shared_ptr< IteratorImpl > Iterator; |
7c673cae | 230 | |
11fdf7f2 | 231 | // This is the low-level iterator implemented by the underlying KV store. |
7c673cae FG |
232 | class WholeSpaceIteratorImpl { |
233 | public: | |
234 | virtual int seek_to_first() = 0; | |
235 | virtual int seek_to_first(const std::string &prefix) = 0; | |
236 | virtual int seek_to_last() = 0; | |
237 | virtual int seek_to_last(const std::string &prefix) = 0; | |
238 | virtual int upper_bound(const std::string &prefix, const std::string &after) = 0; | |
239 | virtual int lower_bound(const std::string &prefix, const std::string &to) = 0; | |
240 | virtual bool valid() = 0; | |
241 | virtual int next() = 0; | |
242 | virtual int prev() = 0; | |
243 | virtual std::string key() = 0; | |
244 | virtual std::pair<std::string,std::string> raw_key() = 0; | |
245 | virtual bool raw_key_is_prefixed(const std::string &prefix) = 0; | |
9f95a23c TL |
246 | virtual ceph::buffer::list value() = 0; |
247 | virtual ceph::buffer::ptr value_as_ptr() { | |
248 | ceph::buffer::list bl = value(); | |
7c673cae FG |
249 | if (bl.length()) { |
250 | return *bl.buffers().begin(); | |
251 | } else { | |
9f95a23c | 252 | return ceph::buffer::ptr(); |
7c673cae FG |
253 | } |
254 | } | |
255 | virtual int status() = 0; | |
256 | virtual size_t key_size() { | |
257 | return 0; | |
258 | } | |
259 | virtual size_t value_size() { | |
260 | return 0; | |
261 | } | |
262 | virtual ~WholeSpaceIteratorImpl() { } | |
263 | }; | |
11fdf7f2 | 264 | typedef std::shared_ptr< WholeSpaceIteratorImpl > WholeSpaceIterator; |
7c673cae | 265 | |
11fdf7f2 TL |
266 | private: |
267 | // This class filters a WholeSpaceIterator by a prefix. | |
20effc67 TL |
268 | // Performs as a dummy wrapper over WholeSpaceIterator |
269 | // if prefix is empty | |
11fdf7f2 | 270 | class PrefixIteratorImpl : public IteratorImpl { |
7c673cae FG |
271 | const std::string prefix; |
272 | WholeSpaceIterator generic_iter; | |
273 | public: | |
11fdf7f2 | 274 | PrefixIteratorImpl(const std::string &prefix, WholeSpaceIterator iter) : |
7c673cae | 275 | prefix(prefix), generic_iter(iter) { } |
11fdf7f2 | 276 | ~PrefixIteratorImpl() override { } |
7c673cae FG |
277 | |
278 | int seek_to_first() override { | |
20effc67 TL |
279 | return prefix.empty() ? |
280 | generic_iter->seek_to_first() : | |
281 | generic_iter->seek_to_first(prefix); | |
7c673cae | 282 | } |
11fdf7f2 | 283 | int seek_to_last() override { |
20effc67 TL |
284 | return prefix.empty() ? |
285 | generic_iter->seek_to_last() : | |
286 | generic_iter->seek_to_last(prefix); | |
7c673cae FG |
287 | } |
288 | int upper_bound(const std::string &after) override { | |
289 | return generic_iter->upper_bound(prefix, after); | |
290 | } | |
291 | int lower_bound(const std::string &to) override { | |
292 | return generic_iter->lower_bound(prefix, to); | |
293 | } | |
294 | bool valid() override { | |
295 | if (!generic_iter->valid()) | |
296 | return false; | |
20effc67 TL |
297 | if (prefix.empty()) |
298 | return true; | |
299 | return prefix.empty() ? | |
300 | true : | |
301 | generic_iter->raw_key_is_prefixed(prefix); | |
7c673cae | 302 | } |
11fdf7f2 TL |
303 | int next() override { |
304 | return generic_iter->next(); | |
7c673cae | 305 | } |
11fdf7f2 TL |
306 | int prev() override { |
307 | return generic_iter->prev(); | |
7c673cae FG |
308 | } |
309 | std::string key() override { | |
310 | return generic_iter->key(); | |
311 | } | |
11fdf7f2 | 312 | std::pair<std::string, std::string> raw_key() override { |
7c673cae FG |
313 | return generic_iter->raw_key(); |
314 | } | |
9f95a23c | 315 | ceph::buffer::list value() override { |
7c673cae FG |
316 | return generic_iter->value(); |
317 | } | |
9f95a23c | 318 | ceph::buffer::ptr value_as_ptr() override { |
7c673cae FG |
319 | return generic_iter->value_as_ptr(); |
320 | } | |
321 | int status() override { | |
322 | return generic_iter->status(); | |
323 | } | |
324 | }; | |
1e59de90 TL |
325 | protected: |
326 | Iterator make_iterator(const std::string &prefix, WholeSpaceIterator w_iter) { | |
327 | return std::make_shared<PrefixIteratorImpl>( | |
328 | prefix, | |
329 | w_iter); | |
330 | } | |
11fdf7f2 | 331 | public: |
f67539c2 TL |
332 | typedef uint32_t IteratorOpts; |
333 | static const uint32_t ITERATOR_NOCACHE = 1; | |
33c7a0ef TL |
334 | |
335 | struct IteratorBounds { | |
336 | std::optional<std::string> lower_bound; | |
337 | std::optional<std::string> upper_bound; | |
338 | }; | |
339 | ||
f67539c2 | 340 | virtual WholeSpaceIterator get_wholespace_iterator(IteratorOpts opts = 0) = 0; |
33c7a0ef | 341 | virtual Iterator get_iterator(const std::string &prefix, IteratorOpts opts = 0, IteratorBounds bounds = IteratorBounds()) { |
1e59de90 | 342 | return make_iterator(prefix, |
f67539c2 | 343 | get_wholespace_iterator(opts)); |
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 | ||
f67539c2 TL |
363 | virtual int64_t get_cache_usage(std::string prefix) const { |
364 | return -EOPNOTSUPP; | |
365 | } | |
366 | ||
11fdf7f2 TL |
367 | virtual std::shared_ptr<PriorityCache::PriCache> get_priority_cache() const { |
368 | return nullptr; | |
91327a77 AA |
369 | } |
370 | ||
f67539c2 TL |
371 | virtual std::shared_ptr<PriorityCache::PriCache> get_priority_cache(std::string prefix) const { |
372 | return nullptr; | |
373 | } | |
374 | ||
375 | ||
376 | ||
11fdf7f2 | 377 | virtual ~KeyValueDB() {} |
91327a77 | 378 | |
11fdf7f2 | 379 | /// estimate space utilization for a prefix (in bytes) |
f67539c2 TL |
380 | virtual int64_t estimate_prefix_size(const std::string& prefix, |
381 | const std::string& key_prefix) { | |
11fdf7f2 | 382 | return 0; |
91327a77 AA |
383 | } |
384 | ||
7c673cae FG |
385 | /// compact the underlying store |
386 | virtual void compact() {} | |
387 | ||
11fdf7f2 TL |
388 | /// compact the underlying store in async mode |
389 | virtual void compact_async() {} | |
390 | ||
7c673cae FG |
391 | /// compact db for all keys with a given prefix |
392 | virtual void compact_prefix(const std::string& prefix) {} | |
393 | /// compact db for all keys with a given prefix, async | |
394 | virtual void compact_prefix_async(const std::string& prefix) {} | |
395 | virtual void compact_range(const std::string& prefix, | |
396 | const std::string& start, const std::string& end) {} | |
397 | virtual void compact_range_async(const std::string& prefix, | |
398 | const std::string& start, const std::string& end) {} | |
399 | ||
400 | // See RocksDB merge operator definition, we support the basic | |
401 | // associative merge only right now. | |
402 | class MergeOperator { | |
403 | public: | |
404 | /// Merge into a key that doesn't exist | |
405 | virtual void merge_nonexistent( | |
406 | const char *rdata, size_t rlen, | |
407 | std::string *new_value) = 0; | |
408 | /// Merge into a key that does exist | |
409 | virtual void merge( | |
410 | const char *ldata, size_t llen, | |
411 | const char *rdata, size_t rlen, | |
412 | std::string *new_value) = 0; | |
413 | /// We use each operator name and each prefix to construct the overall RocksDB operator name for consistency check at open time. | |
91327a77 | 414 | virtual const char *name() const = 0; |
7c673cae FG |
415 | |
416 | virtual ~MergeOperator() {} | |
417 | }; | |
418 | ||
419 | /// Setup one or more operators, this needs to be done BEFORE the DB is opened. | |
420 | virtual int set_merge_operator(const std::string& prefix, | |
421 | std::shared_ptr<MergeOperator> mop) { | |
422 | return -EOPNOTSUPP; | |
423 | } | |
424 | ||
9f95a23c | 425 | virtual void get_statistics(ceph::Formatter *f) { |
7c673cae FG |
426 | return; |
427 | } | |
3efd9988 FG |
428 | |
429 | /** | |
430 | * Return your perf counters if you have any. Subclasses are not | |
431 | * required to implement this, and callers must respect a null return | |
432 | * value. | |
433 | */ | |
434 | virtual PerfCounters *get_perf_counters() { | |
435 | return nullptr; | |
436 | } | |
9f95a23c TL |
437 | |
438 | /** | |
439 | * Access implementation specific integral property corresponding | |
440 | * to passed property and prefic. | |
441 | * Return value is true if property is valid for prefix, populates out. | |
442 | */ | |
443 | virtual bool get_property( | |
444 | const std::string &property, | |
445 | uint64_t *out) { | |
446 | return false; | |
447 | } | |
7c673cae | 448 | protected: |
11fdf7f2 | 449 | /// List of matching prefixes/ColumnFamilies and merge operators |
7c673cae FG |
450 | std::vector<std::pair<std::string, |
451 | std::shared_ptr<MergeOperator> > > merge_ops; | |
452 | ||
7c673cae FG |
453 | }; |
454 | ||
455 | #endif |