]>
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 | 16 | |
7c673cae FG |
17 | /** |
18 | * Defines virtual interface to be implemented by key value store | |
19 | * | |
20 | * Kyoto Cabinet or LevelDB should implement this | |
21 | */ | |
11fdf7f2 | 22 | class KeyValueDB { |
7c673cae FG |
23 | public: |
24 | class TransactionImpl { | |
25 | public: | |
26 | /// Set Keys | |
27 | void set( | |
11fdf7f2 | 28 | const std::string &prefix, ///< [in] Prefix for keys, or CF name |
9f95a23c | 29 | const std::map<std::string, ceph::buffer::list> &to_set ///< [in] keys/values to set |
7c673cae | 30 | ) { |
9f95a23c | 31 | for (auto it = to_set.cbegin(); it != to_set.cend(); ++it) |
7c673cae FG |
32 | set(prefix, it->first, it->second); |
33 | } | |
34 | ||
9f95a23c | 35 | /// Set Keys (via encoded ceph::buffer::list) |
7c673cae | 36 | void set( |
11fdf7f2 | 37 | const std::string &prefix, ///< [in] prefix, or CF name |
9f95a23c | 38 | ceph::buffer::list& to_set_bl ///< [in] encoded key/values to set |
7c673cae | 39 | ) { |
9f95a23c | 40 | using ceph::decode; |
11fdf7f2 | 41 | auto p = std::cbegin(to_set_bl); |
7c673cae | 42 | uint32_t num; |
11fdf7f2 | 43 | decode(num, p); |
7c673cae | 44 | while (num--) { |
f67539c2 | 45 | std::string key; |
9f95a23c | 46 | ceph::buffer::list value; |
11fdf7f2 TL |
47 | decode(key, p); |
48 | decode(value, p); | |
7c673cae FG |
49 | set(prefix, key, value); |
50 | } | |
51 | } | |
52 | ||
53 | /// Set Key | |
54 | virtual void set( | |
11fdf7f2 | 55 | const std::string &prefix, ///< [in] Prefix or CF for the key |
7c673cae | 56 | const std::string &k, ///< [in] Key to set |
9f95a23c | 57 | const ceph::buffer::list &bl ///< [in] Value to set |
7c673cae FG |
58 | ) = 0; |
59 | virtual void set( | |
60 | const std::string &prefix, | |
61 | const char *k, | |
62 | size_t keylen, | |
9f95a23c | 63 | const ceph::buffer::list& bl) { |
f67539c2 | 64 | set(prefix, std::string(k, keylen), bl); |
7c673cae FG |
65 | } |
66 | ||
9f95a23c | 67 | /// Removes Keys (via encoded ceph::buffer::list) |
7c673cae | 68 | void rmkeys( |
11fdf7f2 | 69 | const std::string &prefix, ///< [in] Prefix or CF to search for |
9f95a23c | 70 | ceph::buffer::list &keys_bl ///< [in] Keys to remove |
7c673cae | 71 | ) { |
9f95a23c | 72 | using ceph::decode; |
11fdf7f2 | 73 | auto p = std::cbegin(keys_bl); |
7c673cae | 74 | uint32_t num; |
11fdf7f2 | 75 | decode(num, p); |
7c673cae | 76 | while (num--) { |
f67539c2 | 77 | std::string key; |
11fdf7f2 | 78 | decode(key, p); |
7c673cae FG |
79 | rmkey(prefix, key); |
80 | } | |
81 | } | |
82 | ||
83 | /// Removes Keys | |
84 | void rmkeys( | |
11fdf7f2 | 85 | const std::string &prefix, ///< [in] Prefix/CF to search for |
7c673cae FG |
86 | const std::set<std::string> &keys ///< [in] Keys to remove |
87 | ) { | |
9f95a23c | 88 | for (auto it = keys.cbegin(); it != keys.cend(); ++it) |
7c673cae FG |
89 | rmkey(prefix, *it); |
90 | } | |
91 | ||
92 | /// Remove Key | |
93 | virtual void rmkey( | |
11fdf7f2 TL |
94 | const std::string &prefix, ///< [in] Prefix/CF to search for |
95 | const std::string &k ///< [in] Key to remove | |
7c673cae FG |
96 | ) = 0; |
97 | virtual void rmkey( | |
98 | const std::string &prefix, ///< [in] Prefix to search for | |
99 | const char *k, ///< [in] Key to remove | |
100 | size_t keylen | |
101 | ) { | |
f67539c2 | 102 | rmkey(prefix, std::string(k, keylen)); |
7c673cae FG |
103 | } |
104 | ||
105 | /// Remove Single Key which exists and was not overwritten. | |
106 | /// This API is only related to performance optimization, and should only be | |
107 | /// re-implemented by log-insert-merge tree based keyvalue stores(such as RocksDB). | |
108 | /// If a key is overwritten (by calling set multiple times), then the result | |
109 | /// of calling rm_single_key on this key is undefined. | |
110 | virtual void rm_single_key( | |
11fdf7f2 | 111 | const std::string &prefix, ///< [in] Prefix/CF to search for |
7c673cae FG |
112 | const std::string &k ///< [in] Key to remove |
113 | ) { return rmkey(prefix, k);} | |
114 | ||
115 | /// Removes keys beginning with prefix | |
116 | virtual void rmkeys_by_prefix( | |
11fdf7f2 | 117 | const std::string &prefix ///< [in] Prefix/CF by which to remove keys |
7c673cae FG |
118 | ) = 0; |
119 | ||
120 | virtual void rm_range_keys( | |
f67539c2 TL |
121 | const std::string &prefix, ///< [in] Prefix by which to remove keys |
122 | const std::string &start, ///< [in] The start bound of remove keys | |
123 | const std::string &end ///< [in] The start bound of remove keys | |
7c673cae FG |
124 | ) = 0; |
125 | ||
126 | /// Merge value into key | |
127 | virtual void merge( | |
11fdf7f2 | 128 | const std::string &prefix, ///< [in] Prefix/CF ==> MUST match some established merge operator |
7c673cae | 129 | const std::string &key, ///< [in] Key to be merged |
9f95a23c | 130 | const ceph::buffer::list &value ///< [in] value to be merged into key |
11fdf7f2 | 131 | ) { ceph_abort_msg("Not implemented"); } |
7c673cae FG |
132 | |
133 | virtual ~TransactionImpl() {} | |
134 | }; | |
11fdf7f2 | 135 | typedef std::shared_ptr< TransactionImpl > Transaction; |
7c673cae FG |
136 | |
137 | /// create a new instance | |
138 | static KeyValueDB *create(CephContext *cct, const std::string& type, | |
139 | const std::string& dir, | |
9f95a23c | 140 | std::map<std::string,std::string> options = {}, |
7c673cae FG |
141 | void *p = NULL); |
142 | ||
143 | /// test whether we can successfully initialize; may have side effects (e.g., create) | |
144 | static int test_init(const std::string& type, const std::string& dir); | |
f67539c2 TL |
145 | virtual int init(std::string option_str="") = 0; |
146 | virtual int open(std::ostream &out, const std::string& cfs="") = 0; | |
9f95a23c | 147 | // std::vector cfs contains column families to be created when db is created. |
f67539c2 | 148 | virtual int create_and_open(std::ostream &out, const std::string& cfs="") = 0; |
11fdf7f2 | 149 | |
f67539c2 | 150 | virtual int open_read_only(std::ostream &out, const std::string& cfs="") { |
11fdf7f2 TL |
151 | return -ENOTSUP; |
152 | } | |
153 | ||
7c673cae FG |
154 | virtual void close() { } |
155 | ||
11fdf7f2 TL |
156 | /// Try to repair K/V database. leveldb and rocksdb require that database must be not opened. |
157 | virtual int repair(std::ostream &out) { return 0; } | |
158 | ||
7c673cae FG |
159 | virtual Transaction get_transaction() = 0; |
160 | virtual int submit_transaction(Transaction) = 0; | |
161 | virtual int submit_transaction_sync(Transaction t) { | |
162 | return submit_transaction(t); | |
163 | } | |
164 | ||
165 | /// Retrieve Keys | |
166 | virtual int get( | |
11fdf7f2 TL |
167 | const std::string &prefix, ///< [in] Prefix/CF for key |
168 | const std::set<std::string> &key, ///< [in] Key to retrieve | |
9f95a23c | 169 | std::map<std::string, ceph::buffer::list> *out ///< [out] Key value retrieved |
7c673cae | 170 | ) = 0; |
11fdf7f2 | 171 | virtual int get(const std::string &prefix, ///< [in] prefix or CF name |
7c673cae | 172 | const std::string &key, ///< [in] key |
9f95a23c | 173 | ceph::buffer::list *value) { ///< [out] value |
7c673cae FG |
174 | std::set<std::string> ks; |
175 | ks.insert(key); | |
9f95a23c | 176 | std::map<std::string,ceph::buffer::list> om; |
7c673cae FG |
177 | int r = get(prefix, ks, &om); |
178 | if (om.find(key) != om.end()) { | |
11fdf7f2 | 179 | *value = std::move(om[key]); |
7c673cae | 180 | } else { |
9f95a23c | 181 | *value = ceph::buffer::list(); |
7c673cae FG |
182 | r = -ENOENT; |
183 | } | |
184 | return r; | |
185 | } | |
f67539c2 | 186 | virtual int get(const std::string &prefix, |
7c673cae | 187 | const char *key, size_t keylen, |
9f95a23c | 188 | ceph::buffer::list *value) { |
f67539c2 | 189 | return get(prefix, std::string(key, keylen), value); |
7c673cae FG |
190 | } |
191 | ||
11fdf7f2 TL |
192 | // This superclass is used both by kv iterators *and* by the ObjectMap |
193 | // omap iterator. The class hierarchies are unfortunately tied together | |
194 | // by the legacy DBOjectMap implementation :(. | |
195 | class SimplestIteratorImpl { | |
7c673cae FG |
196 | public: |
197 | virtual int seek_to_first() = 0; | |
198 | virtual int upper_bound(const std::string &after) = 0; | |
199 | virtual int lower_bound(const std::string &to) = 0; | |
200 | virtual bool valid() = 0; | |
11fdf7f2 | 201 | virtual int next() = 0; |
7c673cae | 202 | virtual std::string key() = 0; |
9f95a23c TL |
203 | virtual std::string tail_key() { |
204 | return ""; | |
205 | } | |
206 | virtual ceph::buffer::list value() = 0; | |
7c673cae | 207 | virtual int status() = 0; |
11fdf7f2 TL |
208 | virtual ~SimplestIteratorImpl() {} |
209 | }; | |
210 | ||
211 | class IteratorImpl : public SimplestIteratorImpl { | |
212 | public: | |
213 | virtual ~IteratorImpl() {} | |
214 | virtual int seek_to_last() = 0; | |
215 | virtual int prev() = 0; | |
216 | virtual std::pair<std::string, std::string> raw_key() = 0; | |
9f95a23c TL |
217 | virtual ceph::buffer::ptr value_as_ptr() { |
218 | ceph::buffer::list bl = value(); | |
11fdf7f2 TL |
219 | if (bl.length() == 1) { |
220 | return *bl.buffers().begin(); | |
221 | } else if (bl.length() == 0) { | |
9f95a23c | 222 | return ceph::buffer::ptr(); |
11fdf7f2 TL |
223 | } else { |
224 | ceph_abort(); | |
225 | } | |
226 | } | |
7c673cae | 227 | }; |
11fdf7f2 | 228 | typedef std::shared_ptr< IteratorImpl > Iterator; |
7c673cae | 229 | |
11fdf7f2 | 230 | // This is the low-level iterator implemented by the underlying KV store. |
7c673cae FG |
231 | class WholeSpaceIteratorImpl { |
232 | public: | |
233 | virtual int seek_to_first() = 0; | |
234 | virtual int seek_to_first(const std::string &prefix) = 0; | |
235 | virtual int seek_to_last() = 0; | |
236 | virtual int seek_to_last(const std::string &prefix) = 0; | |
237 | virtual int upper_bound(const std::string &prefix, const std::string &after) = 0; | |
238 | virtual int lower_bound(const std::string &prefix, const std::string &to) = 0; | |
239 | virtual bool valid() = 0; | |
240 | virtual int next() = 0; | |
241 | virtual int prev() = 0; | |
242 | virtual std::string key() = 0; | |
243 | virtual std::pair<std::string,std::string> raw_key() = 0; | |
244 | virtual bool raw_key_is_prefixed(const std::string &prefix) = 0; | |
9f95a23c TL |
245 | virtual ceph::buffer::list value() = 0; |
246 | virtual ceph::buffer::ptr value_as_ptr() { | |
247 | ceph::buffer::list bl = value(); | |
7c673cae FG |
248 | if (bl.length()) { |
249 | return *bl.buffers().begin(); | |
250 | } else { | |
9f95a23c | 251 | return ceph::buffer::ptr(); |
7c673cae FG |
252 | } |
253 | } | |
254 | virtual int status() = 0; | |
255 | virtual size_t key_size() { | |
256 | return 0; | |
257 | } | |
258 | virtual size_t value_size() { | |
259 | return 0; | |
260 | } | |
261 | virtual ~WholeSpaceIteratorImpl() { } | |
262 | }; | |
11fdf7f2 | 263 | typedef std::shared_ptr< WholeSpaceIteratorImpl > WholeSpaceIterator; |
7c673cae | 264 | |
11fdf7f2 TL |
265 | private: |
266 | // This class filters a WholeSpaceIterator by a prefix. | |
267 | class PrefixIteratorImpl : public IteratorImpl { | |
7c673cae FG |
268 | const std::string prefix; |
269 | WholeSpaceIterator generic_iter; | |
270 | public: | |
11fdf7f2 | 271 | PrefixIteratorImpl(const std::string &prefix, WholeSpaceIterator iter) : |
7c673cae | 272 | prefix(prefix), generic_iter(iter) { } |
11fdf7f2 | 273 | ~PrefixIteratorImpl() override { } |
7c673cae FG |
274 | |
275 | int seek_to_first() override { | |
276 | return generic_iter->seek_to_first(prefix); | |
277 | } | |
11fdf7f2 | 278 | int seek_to_last() override { |
7c673cae FG |
279 | return generic_iter->seek_to_last(prefix); |
280 | } | |
281 | int upper_bound(const std::string &after) override { | |
282 | return generic_iter->upper_bound(prefix, after); | |
283 | } | |
284 | int lower_bound(const std::string &to) override { | |
285 | return generic_iter->lower_bound(prefix, to); | |
286 | } | |
287 | bool valid() override { | |
288 | if (!generic_iter->valid()) | |
289 | return false; | |
290 | return generic_iter->raw_key_is_prefixed(prefix); | |
291 | } | |
11fdf7f2 TL |
292 | int next() override { |
293 | return generic_iter->next(); | |
7c673cae | 294 | } |
11fdf7f2 TL |
295 | int prev() override { |
296 | return generic_iter->prev(); | |
7c673cae FG |
297 | } |
298 | std::string key() override { | |
299 | return generic_iter->key(); | |
300 | } | |
11fdf7f2 | 301 | std::pair<std::string, std::string> raw_key() override { |
7c673cae FG |
302 | return generic_iter->raw_key(); |
303 | } | |
9f95a23c | 304 | ceph::buffer::list value() override { |
7c673cae FG |
305 | return generic_iter->value(); |
306 | } | |
9f95a23c | 307 | ceph::buffer::ptr value_as_ptr() override { |
7c673cae FG |
308 | return generic_iter->value_as_ptr(); |
309 | } | |
310 | int status() override { | |
311 | return generic_iter->status(); | |
312 | } | |
313 | }; | |
11fdf7f2 | 314 | public: |
f67539c2 TL |
315 | typedef uint32_t IteratorOpts; |
316 | static const uint32_t ITERATOR_NOCACHE = 1; | |
317 | virtual WholeSpaceIterator get_wholespace_iterator(IteratorOpts opts = 0) = 0; | |
318 | virtual Iterator get_iterator(const std::string &prefix, IteratorOpts opts = 0) { | |
11fdf7f2 TL |
319 | return std::make_shared<PrefixIteratorImpl>( |
320 | prefix, | |
f67539c2 | 321 | get_wholespace_iterator(opts)); |
7c673cae FG |
322 | } |
323 | ||
324 | virtual uint64_t get_estimated_size(std::map<std::string,uint64_t> &extra) = 0; | |
325 | virtual int get_statfs(struct store_statfs_t *buf) { | |
326 | return -EOPNOTSUPP; | |
327 | } | |
328 | ||
31f18b77 FG |
329 | virtual int set_cache_size(uint64_t) { |
330 | return -EOPNOTSUPP; | |
331 | } | |
332 | ||
11fdf7f2 | 333 | virtual int set_cache_high_pri_pool_ratio(double ratio) { |
91327a77 AA |
334 | return -EOPNOTSUPP; |
335 | } | |
336 | ||
11fdf7f2 | 337 | virtual int64_t get_cache_usage() const { |
91327a77 AA |
338 | return -EOPNOTSUPP; |
339 | } | |
340 | ||
f67539c2 TL |
341 | virtual int64_t get_cache_usage(std::string prefix) const { |
342 | return -EOPNOTSUPP; | |
343 | } | |
344 | ||
11fdf7f2 TL |
345 | virtual std::shared_ptr<PriorityCache::PriCache> get_priority_cache() const { |
346 | return nullptr; | |
91327a77 AA |
347 | } |
348 | ||
f67539c2 TL |
349 | virtual std::shared_ptr<PriorityCache::PriCache> get_priority_cache(std::string prefix) const { |
350 | return nullptr; | |
351 | } | |
352 | ||
353 | ||
354 | ||
11fdf7f2 | 355 | virtual ~KeyValueDB() {} |
91327a77 | 356 | |
11fdf7f2 | 357 | /// estimate space utilization for a prefix (in bytes) |
f67539c2 TL |
358 | virtual int64_t estimate_prefix_size(const std::string& prefix, |
359 | const std::string& key_prefix) { | |
11fdf7f2 | 360 | return 0; |
91327a77 AA |
361 | } |
362 | ||
7c673cae FG |
363 | /// compact the underlying store |
364 | virtual void compact() {} | |
365 | ||
11fdf7f2 TL |
366 | /// compact the underlying store in async mode |
367 | virtual void compact_async() {} | |
368 | ||
7c673cae FG |
369 | /// compact db for all keys with a given prefix |
370 | virtual void compact_prefix(const std::string& prefix) {} | |
371 | /// compact db for all keys with a given prefix, async | |
372 | virtual void compact_prefix_async(const std::string& prefix) {} | |
373 | virtual void compact_range(const std::string& prefix, | |
374 | const std::string& start, const std::string& end) {} | |
375 | virtual void compact_range_async(const std::string& prefix, | |
376 | const std::string& start, const std::string& end) {} | |
377 | ||
378 | // See RocksDB merge operator definition, we support the basic | |
379 | // associative merge only right now. | |
380 | class MergeOperator { | |
381 | public: | |
382 | /// Merge into a key that doesn't exist | |
383 | virtual void merge_nonexistent( | |
384 | const char *rdata, size_t rlen, | |
385 | std::string *new_value) = 0; | |
386 | /// Merge into a key that does exist | |
387 | virtual void merge( | |
388 | const char *ldata, size_t llen, | |
389 | const char *rdata, size_t rlen, | |
390 | std::string *new_value) = 0; | |
391 | /// We use each operator name and each prefix to construct the overall RocksDB operator name for consistency check at open time. | |
91327a77 | 392 | virtual const char *name() const = 0; |
7c673cae FG |
393 | |
394 | virtual ~MergeOperator() {} | |
395 | }; | |
396 | ||
397 | /// Setup one or more operators, this needs to be done BEFORE the DB is opened. | |
398 | virtual int set_merge_operator(const std::string& prefix, | |
399 | std::shared_ptr<MergeOperator> mop) { | |
400 | return -EOPNOTSUPP; | |
401 | } | |
402 | ||
9f95a23c | 403 | virtual void get_statistics(ceph::Formatter *f) { |
7c673cae FG |
404 | return; |
405 | } | |
3efd9988 FG |
406 | |
407 | /** | |
408 | * Return your perf counters if you have any. Subclasses are not | |
409 | * required to implement this, and callers must respect a null return | |
410 | * value. | |
411 | */ | |
412 | virtual PerfCounters *get_perf_counters() { | |
413 | return nullptr; | |
414 | } | |
9f95a23c TL |
415 | |
416 | /** | |
417 | * Access implementation specific integral property corresponding | |
418 | * to passed property and prefic. | |
419 | * Return value is true if property is valid for prefix, populates out. | |
420 | */ | |
421 | virtual bool get_property( | |
422 | const std::string &property, | |
423 | uint64_t *out) { | |
424 | return false; | |
425 | } | |
7c673cae | 426 | protected: |
11fdf7f2 | 427 | /// List of matching prefixes/ColumnFamilies and merge operators |
7c673cae FG |
428 | std::vector<std::pair<std::string, |
429 | std::shared_ptr<MergeOperator> > > merge_ops; | |
430 | ||
7c673cae FG |
431 | }; |
432 | ||
433 | #endif |