1 // Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
2 // This source code is licensed under the BSD-style license found in the
3 // LICENSE file in the root directory of this source tree. An additional grant
4 // of patent rights can be found in the PATENTS file in the same directory.
6 // Copyright (c) 2011 The LevelDB Authors. All rights reserved.
7 // Use of this source code is governed by a BSD-style license that can be
8 // found in the LICENSE file. See the AUTHORS file for names of contributors.
12 #include <unordered_map>
17 #include "db/memtable_list.h"
18 #include "db/table_cache.h"
19 #include "db/table_properties_collector.h"
20 #include "db/write_batch_internal.h"
21 #include "db/write_controller.h"
22 #include "options/cf_options.h"
23 #include "rocksdb/compaction_job_stats.h"
24 #include "rocksdb/db.h"
25 #include "rocksdb/env.h"
26 #include "rocksdb/options.h"
27 #include "util/thread_local.h"
34 class MemTableListVersion
;
35 class CompactionPicker
;
39 class ColumnFamilyData
;
42 class InstrumentedMutex
;
43 class InstrumentedMutexLock
;
45 extern const double kIncSlowdownRatio
;
47 // ColumnFamilyHandleImpl is the class that clients use to access different
48 // column families. It has non-trivial destructor, which gets called when client
49 // is done using the column family
50 class ColumnFamilyHandleImpl
: public ColumnFamilyHandle
{
52 // create while holding the mutex
53 ColumnFamilyHandleImpl(
54 ColumnFamilyData
* cfd
, DBImpl
* db
, InstrumentedMutex
* mutex
);
55 // destroy without mutex
56 virtual ~ColumnFamilyHandleImpl();
57 virtual ColumnFamilyData
* cfd() const { return cfd_
; }
59 virtual uint32_t GetID() const override
;
60 virtual const std::string
& GetName() const override
;
61 virtual Status
GetDescriptor(ColumnFamilyDescriptor
* desc
) override
;
62 virtual const Comparator
* GetComparator() const override
;
65 ColumnFamilyData
* cfd_
;
67 InstrumentedMutex
* mutex_
;
70 // Does not ref-count ColumnFamilyData
71 // We use this dummy ColumnFamilyHandleImpl because sometimes MemTableInserter
72 // calls DBImpl methods. When this happens, MemTableInserter need access to
73 // ColumnFamilyHandle (same as the client would need). In that case, we feed
74 // MemTableInserter dummy ColumnFamilyHandle and enable it to call DBImpl
76 class ColumnFamilyHandleInternal
: public ColumnFamilyHandleImpl
{
78 ColumnFamilyHandleInternal()
79 : ColumnFamilyHandleImpl(nullptr, nullptr, nullptr) {}
81 void SetCFD(ColumnFamilyData
* _cfd
) { internal_cfd_
= _cfd
; }
82 virtual ColumnFamilyData
* cfd() const override
{ return internal_cfd_
; }
85 ColumnFamilyData
* internal_cfd_
;
88 // holds references to memtable, all immutable memtables and version
90 // Accessing members of this class is not thread-safe and requires external
91 // synchronization (ie db mutex held or on write thread).
93 MemTableListVersion
* imm
;
95 MutableCFOptions mutable_cf_options
;
96 // Version number of the current SuperVersion
97 uint64_t version_number
;
99 InstrumentedMutex
* db_mutex
;
101 // should be called outside the mutex
102 SuperVersion() = default;
105 // If Unref() returns true, Cleanup() should be called with mutex held
106 // before deleting this SuperVersion.
109 // call these two methods with db mutex held
110 // Cleanup unrefs mem, imm and current. Also, it stores all memtables
111 // that needs to be deleted in to_delete vector. Unrefing those
112 // objects needs to be done in the mutex
114 void Init(MemTable
* new_mem
, MemTableListVersion
* new_imm
,
115 Version
* new_current
);
117 // The value of dummy is not actually used. kSVInUse takes its address as a
118 // mark in the thread local storage to indicate the SuperVersion is in use
119 // by thread. This way, the value of kSVInUse is guaranteed to have no
120 // conflict with SuperVersion object address and portable on different
123 static void* const kSVInUse
;
124 static void* const kSVObsolete
;
127 std::atomic
<uint32_t> refs
;
128 // We need to_delete because during Cleanup(), imm->Unref() returns
129 // all memtables that we need to free through this vector. We then
130 // delete all those memtables outside of mutex, during destruction
131 autovector
<MemTable
*> to_delete
;
134 extern Status
CheckCompressionSupported(const ColumnFamilyOptions
& cf_options
);
136 extern Status
CheckConcurrentWritesSupported(
137 const ColumnFamilyOptions
& cf_options
);
139 extern ColumnFamilyOptions
SanitizeOptions(const ImmutableDBOptions
& db_options
,
140 const ColumnFamilyOptions
& src
);
141 // Wrap user defined table proproties collector factories `from cf_options`
142 // into internal ones in int_tbl_prop_collector_factories. Add a system internal
144 extern void GetIntTblPropCollectorFactory(
145 const ImmutableCFOptions
& ioptions
,
146 std::vector
<std::unique_ptr
<IntTblPropCollectorFactory
>>*
147 int_tbl_prop_collector_factories
);
149 class ColumnFamilySet
;
151 // This class keeps all the data that a column family needs.
152 // Most methods require DB mutex held, unless otherwise noted
153 class ColumnFamilyData
{
158 uint32_t GetID() const { return id_
; }
160 const std::string
& GetName() const { return name_
; }
162 // Ref() can only be called from a context where the caller can guarantee
163 // that ColumnFamilyData is alive (while holding a non-zero ref already,
164 // holding a DB mutex, or as the leader in a write batch group).
165 void Ref() { refs_
.fetch_add(1, std::memory_order_relaxed
); }
167 // Unref decreases the reference count, but does not handle deletion
168 // when the count goes to 0. If this method returns true then the
169 // caller should delete the instance immediately, or later, by calling
170 // FreeDeadColumnFamilies(). Unref() can only be called while holding
171 // a DB mutex, or during single-threaded recovery.
173 int old_refs
= refs_
.fetch_sub(1, std::memory_order_relaxed
);
174 assert(old_refs
> 0);
175 return old_refs
== 1;
178 // SetDropped() can only be called under following conditions:
179 // 1) Holding a DB mutex,
180 // 2) from single-threaded write thread, AND
181 // 3) from single-threaded VersionSet::LogAndApply()
182 // After dropping column family no other operation on that column family
183 // will be executed. All the files and memory will be, however, kept around
184 // until client drops the column family handle. That way, client can still
185 // access data from dropped column family.
186 // Column family can be dropped and still alive. In that state:
187 // *) Compaction and flush is not executed on the dropped column family.
188 // *) Client can continue reading from column family. Writes will fail unless
189 // WriteOptions::ignore_missing_column_families is true
190 // When the dropped column family is unreferenced, then we:
191 // *) Remove column family from the linked list maintained by ColumnFamilySet
192 // *) delete all memory associated with that column family
193 // *) delete all the files associated with that column family
195 bool IsDropped() const { return dropped_
; }
198 int NumberLevels() const { return ioptions_
.num_levels
; }
200 void SetLogNumber(uint64_t log_number
) { log_number_
= log_number
; }
201 uint64_t GetLogNumber() const { return log_number_
; }
204 const EnvOptions
* soptions() const;
205 const ImmutableCFOptions
* ioptions() const { return &ioptions_
; }
206 // REQUIRES: DB mutex held
207 // This returns the MutableCFOptions used by current SuperVersion
208 // You should use this API to reference MutableCFOptions most of the time.
209 const MutableCFOptions
* GetCurrentMutableCFOptions() const {
210 return &(super_version_
->mutable_cf_options
);
212 // REQUIRES: DB mutex held
213 // This returns the latest MutableCFOptions, which may be not in effect yet.
214 const MutableCFOptions
* GetLatestMutableCFOptions() const {
215 return &mutable_cf_options_
;
218 // REQUIRES: DB mutex held
219 // Build ColumnFamiliesOptions with immutable options and latest mutable
221 ColumnFamilyOptions
GetLatestCFOptions() const;
223 bool is_delete_range_supported() { return is_delete_range_supported_
; }
226 // REQUIRES: DB mutex held
228 const std::unordered_map
<std::string
, std::string
>& options_map
);
229 #endif // ROCKSDB_LITE
231 InternalStats
* internal_stats() { return internal_stats_
.get(); }
233 MemTableList
* imm() { return &imm_
; }
234 MemTable
* mem() { return mem_
; }
235 Version
* current() { return current_
; }
236 Version
* dummy_versions() { return dummy_versions_
; }
237 void SetCurrent(Version
* _current
);
238 uint64_t GetNumLiveVersions() const; // REQUIRE: DB mutex held
239 uint64_t GetTotalSstFilesSize() const; // REQUIRE: DB mutex held
240 void SetMemtable(MemTable
* new_mem
) { mem_
= new_mem
; }
242 // calculate the oldest log needed for the durability of this column family
243 uint64_t OldestLogToKeep();
245 // See Memtable constructor for explanation of earliest_seq param.
246 MemTable
* ConstructNewMemtable(const MutableCFOptions
& mutable_cf_options
,
247 SequenceNumber earliest_seq
);
248 void CreateNewMemtable(const MutableCFOptions
& mutable_cf_options
,
249 SequenceNumber earliest_seq
);
251 TableCache
* table_cache() const { return table_cache_
.get(); }
253 // See documentation in compaction_picker.h
254 // REQUIRES: DB mutex held
255 bool NeedsCompaction() const;
256 // REQUIRES: DB mutex held
257 Compaction
* PickCompaction(const MutableCFOptions
& mutable_options
,
258 LogBuffer
* log_buffer
);
260 // Check if the passed range overlap with any running compactions.
261 // REQUIRES: DB mutex held
262 bool RangeOverlapWithCompaction(const Slice
& smallest_user_key
,
263 const Slice
& largest_user_key
,
266 // A flag to tell a manual compaction is to compact all levels together
267 // instad of for specific level.
268 static const int kCompactAllLevels
;
269 // A flag to tell a manual compaction's output is base level.
270 static const int kCompactToBaseLevel
;
271 // REQUIRES: DB mutex held
272 Compaction
* CompactRange(const MutableCFOptions
& mutable_cf_options
,
273 int input_level
, int output_level
,
274 uint32_t output_path_id
, const InternalKey
* begin
,
275 const InternalKey
* end
, InternalKey
** compaction_end
,
276 bool* manual_conflict
);
278 CompactionPicker
* compaction_picker() { return compaction_picker_
.get(); }
280 const Comparator
* user_comparator() const {
281 return internal_comparator_
.user_comparator();
284 const InternalKeyComparator
& internal_comparator() const {
285 return internal_comparator_
;
288 const std::vector
<std::unique_ptr
<IntTblPropCollectorFactory
>>*
289 int_tbl_prop_collector_factories() const {
290 return &int_tbl_prop_collector_factories_
;
293 SuperVersion
* GetSuperVersion() { return super_version_
; }
295 // Return a already referenced SuperVersion to be used safely.
296 SuperVersion
* GetReferencedSuperVersion(InstrumentedMutex
* db_mutex
);
298 // Get SuperVersion stored in thread local storage. If it does not exist,
299 // get a reference from a current SuperVersion.
300 SuperVersion
* GetThreadLocalSuperVersion(InstrumentedMutex
* db_mutex
);
301 // Try to return SuperVersion back to thread local storage. Retrun true on
302 // success and false on failure. It fails when the thread local storage
303 // contains anything other than SuperVersion::kSVInUse flag.
304 bool ReturnThreadLocalSuperVersion(SuperVersion
* sv
);
306 uint64_t GetSuperVersionNumber() const {
307 return super_version_number_
.load();
309 // will return a pointer to SuperVersion* if previous SuperVersion
310 // if its reference count is zero and needs deletion or nullptr if not
311 // As argument takes a pointer to allocated SuperVersion to enable
312 // the clients to allocate SuperVersion outside of mutex.
313 // IMPORTANT: Only call this from DBImpl::InstallSuperVersion()
314 SuperVersion
* InstallSuperVersion(SuperVersion
* new_superversion
,
315 InstrumentedMutex
* db_mutex
,
316 const MutableCFOptions
& mutable_cf_options
);
317 SuperVersion
* InstallSuperVersion(SuperVersion
* new_superversion
,
318 InstrumentedMutex
* db_mutex
);
320 void ResetThreadLocalSuperVersions();
322 // Protected by DB mutex
323 void set_pending_flush(bool value
) { pending_flush_
= value
; }
324 void set_pending_compaction(bool value
) { pending_compaction_
= value
; }
325 bool pending_flush() { return pending_flush_
; }
326 bool pending_compaction() { return pending_compaction_
; }
328 // Recalculate some small conditions, which are changed only during
329 // compaction, adding new memtable and/or
330 // recalculation of compaction score. These values are used in
331 // DBImpl::MakeRoomForWrite function to decide, if it need to make
333 void RecalculateWriteStallConditions(
334 const MutableCFOptions
& mutable_cf_options
);
337 friend class ColumnFamilySet
;
338 ColumnFamilyData(uint32_t id
, const std::string
& name
,
339 Version
* dummy_versions
, Cache
* table_cache
,
340 WriteBufferManager
* write_buffer_manager
,
341 const ColumnFamilyOptions
& options
,
342 const ImmutableDBOptions
& db_options
,
343 const EnvOptions
& env_options
,
344 ColumnFamilySet
* column_family_set
);
347 const std::string name_
;
348 Version
* dummy_versions_
; // Head of circular doubly-linked list of versions.
349 Version
* current_
; // == dummy_versions->prev_
351 std::atomic
<int> refs_
; // outstanding references to ColumnFamilyData
352 bool dropped_
; // true if client dropped it
354 const InternalKeyComparator internal_comparator_
;
355 std::vector
<std::unique_ptr
<IntTblPropCollectorFactory
>>
356 int_tbl_prop_collector_factories_
;
358 const ColumnFamilyOptions initial_cf_options_
;
359 const ImmutableCFOptions ioptions_
;
360 MutableCFOptions mutable_cf_options_
;
362 const bool is_delete_range_supported_
;
364 std::unique_ptr
<TableCache
> table_cache_
;
366 std::unique_ptr
<InternalStats
> internal_stats_
;
368 WriteBufferManager
* write_buffer_manager_
;
372 SuperVersion
* super_version_
;
374 // An ordinal representing the current SuperVersion. Updated by
375 // InstallSuperVersion(), i.e. incremented every time super_version_
377 std::atomic
<uint64_t> super_version_number_
;
379 // Thread's local copy of SuperVersion pointer
380 // This needs to be destructed before mutex_
381 std::unique_ptr
<ThreadLocalPtr
> local_sv_
;
383 // pointers for a circular linked list. we use it to support iterations over
384 // all column families that are alive (note: dropped column families can also
385 // be alive as long as client holds a reference)
386 ColumnFamilyData
* next_
;
387 ColumnFamilyData
* prev_
;
389 // This is the earliest log file number that contains data from this
390 // Column Family. All earlier log files must be ignored and not
392 uint64_t log_number_
;
394 // An object that keeps all the compaction stats
395 // and picks the next compaction
396 std::unique_ptr
<CompactionPicker
> compaction_picker_
;
398 ColumnFamilySet
* column_family_set_
;
400 std::unique_ptr
<WriteControllerToken
> write_controller_token_
;
402 // If true --> this ColumnFamily is currently present in DBImpl::flush_queue_
405 // If true --> this ColumnFamily is currently present in
406 // DBImpl::compaction_queue_
407 bool pending_compaction_
;
409 uint64_t prev_compaction_needed_bytes_
;
411 // if the database was opened with 2pc enabled
415 // ColumnFamilySet has interesting thread-safety requirements
416 // * CreateColumnFamily() or RemoveColumnFamily() -- need to be protected by DB
417 // mutex AND executed in the write thread.
418 // CreateColumnFamily() should ONLY be called from VersionSet::LogAndApply() AND
419 // single-threaded write thread. It is also called during Recovery and in
421 // RemoveColumnFamily() is only called from SetDropped(). DB mutex needs to be
422 // held and it needs to be executed from the write thread. SetDropped() also
423 // guarantees that it will be called only from single-threaded LogAndApply(),
424 // but this condition is not that important.
425 // * Iteration -- hold DB mutex, but you can release it in the body of
426 // iteration. If you release DB mutex in body, reference the column
427 // family before the mutex and unreference after you unlock, since the column
428 // family might get dropped when the DB mutex is released
429 // * GetDefault() -- thread safe
430 // * GetColumnFamily() -- either inside of DB mutex or from a write thread
431 // * GetNextColumnFamilyID(), GetMaxColumnFamily(), UpdateMaxColumnFamily(),
432 // NumberOfColumnFamilies -- inside of DB mutex
433 class ColumnFamilySet
{
435 // ColumnFamilySet supports iteration
438 explicit iterator(ColumnFamilyData
* cfd
)
440 iterator
& operator++() {
441 // dropped column families might still be included in this iteration
442 // (we're only removing them when client drops the last reference to the
444 // dummy is never dead, so this will never be infinite
446 current_
= current_
->next_
;
447 } while (current_
->refs_
.load(std::memory_order_relaxed
) == 0);
450 bool operator!=(const iterator
& other
) {
451 return this->current_
!= other
.current_
;
453 ColumnFamilyData
* operator*() { return current_
; }
456 ColumnFamilyData
* current_
;
459 ColumnFamilySet(const std::string
& dbname
,
460 const ImmutableDBOptions
* db_options
,
461 const EnvOptions
& env_options
, Cache
* table_cache
,
462 WriteBufferManager
* write_buffer_manager
,
463 WriteController
* write_controller
);
466 ColumnFamilyData
* GetDefault() const;
467 // GetColumnFamily() calls return nullptr if column family is not found
468 ColumnFamilyData
* GetColumnFamily(uint32_t id
) const;
469 ColumnFamilyData
* GetColumnFamily(const std::string
& name
) const;
470 // this call will return the next available column family ID. it guarantees
471 // that there is no column family with id greater than or equal to the
472 // returned value in the current running instance or anytime in RocksDB
474 uint32_t GetNextColumnFamilyID();
475 uint32_t GetMaxColumnFamily();
476 void UpdateMaxColumnFamily(uint32_t new_max_column_family
);
477 size_t NumberOfColumnFamilies() const;
479 ColumnFamilyData
* CreateColumnFamily(const std::string
& name
, uint32_t id
,
480 Version
* dummy_version
,
481 const ColumnFamilyOptions
& options
);
483 iterator
begin() { return iterator(dummy_cfd_
->next_
); }
484 iterator
end() { return iterator(dummy_cfd_
); }
486 // REQUIRES: DB mutex held
487 // Don't call while iterating over ColumnFamilySet
488 void FreeDeadColumnFamilies();
490 Cache
* get_table_cache() { return table_cache_
; }
493 friend class ColumnFamilyData
;
494 // helper function that gets called from cfd destructor
495 // REQUIRES: DB mutex held
496 void RemoveColumnFamily(ColumnFamilyData
* cfd
);
498 // column_families_ and column_family_data_ need to be protected:
499 // * when mutating both conditions have to be satisfied:
500 // 1. DB mutex locked
501 // 2. thread currently in single-threaded write thread
502 // * when reading, at least one condition needs to be satisfied:
503 // 1. DB mutex locked
504 // 2. accessed from a single-threaded write thread
505 std::unordered_map
<std::string
, uint32_t> column_families_
;
506 std::unordered_map
<uint32_t, ColumnFamilyData
*> column_family_data_
;
508 uint32_t max_column_family_
;
509 ColumnFamilyData
* dummy_cfd_
;
510 // We don't hold the refcount here, since default column family always exists
511 // We are also not responsible for cleaning up default_cfd_cache_. This is
512 // just a cache that makes common case (accessing default column family)
514 ColumnFamilyData
* default_cfd_cache_
;
516 const std::string db_name_
;
517 const ImmutableDBOptions
* const db_options_
;
518 const EnvOptions env_options_
;
520 WriteBufferManager
* write_buffer_manager_
;
521 WriteController
* write_controller_
;
524 // We use ColumnFamilyMemTablesImpl to provide WriteBatch a way to access
525 // memtables of different column families (specified by ID in the write batch)
526 class ColumnFamilyMemTablesImpl
: public ColumnFamilyMemTables
{
528 explicit ColumnFamilyMemTablesImpl(ColumnFamilySet
* column_family_set
)
529 : column_family_set_(column_family_set
), current_(nullptr) {}
531 // Constructs a ColumnFamilyMemTablesImpl equivalent to one constructed
532 // with the arguments used to construct *orig.
533 explicit ColumnFamilyMemTablesImpl(ColumnFamilyMemTablesImpl
* orig
)
534 : column_family_set_(orig
->column_family_set_
), current_(nullptr) {}
536 // sets current_ to ColumnFamilyData with column_family_id
537 // returns false if column family doesn't exist
538 // REQUIRES: use this function of DBImpl::column_family_memtables_ should be
539 // under a DB mutex OR from a write thread
540 bool Seek(uint32_t column_family_id
) override
;
542 // Returns log number of the selected column family
543 // REQUIRES: under a DB mutex OR from a write thread
544 uint64_t GetLogNumber() const override
;
546 // REQUIRES: Seek() called first
547 // REQUIRES: use this function of DBImpl::column_family_memtables_ should be
548 // under a DB mutex OR from a write thread
549 virtual MemTable
* GetMemTable() const override
;
551 // Returns column family handle for the selected column family
552 // REQUIRES: use this function of DBImpl::column_family_memtables_ should be
553 // under a DB mutex OR from a write thread
554 virtual ColumnFamilyHandle
* GetColumnFamilyHandle() override
;
556 // Cannot be called while another thread is calling Seek().
557 // REQUIRES: use this function of DBImpl::column_family_memtables_ should be
558 // under a DB mutex OR from a write thread
559 virtual ColumnFamilyData
* current() override
{ return current_
; }
562 ColumnFamilySet
* column_family_set_
;
563 ColumnFamilyData
* current_
;
564 ColumnFamilyHandleInternal handle_
;
567 extern uint32_t GetColumnFamilyID(ColumnFamilyHandle
* column_family
);
569 extern const Comparator
* GetColumnFamilyUserComparator(
570 ColumnFamilyHandle
* column_family
);
572 } // namespace rocksdb