Line data Source code
1 : // Copyright 2012 The LevelDB-Go and Pebble Authors. All rights reserved. Use
2 : // of this source code is governed by a BSD-style license that can be found in
3 : // the LICENSE file.
4 :
5 : package pebble
6 :
7 : import (
8 : "context"
9 : "io"
10 : "math"
11 : "sync"
12 : "time"
13 :
14 : "github.com/cockroachdb/pebble/internal/base"
15 : "github.com/cockroachdb/pebble/internal/invariants"
16 : "github.com/cockroachdb/pebble/rangekey"
17 : "github.com/cockroachdb/pebble/sstable/block"
18 : )
19 :
20 : // Snapshot provides a read-only point-in-time view of the DB state.
21 : type Snapshot struct {
22 : // The db the snapshot was created from.
23 : db *DB
24 : seqNum base.SeqNum
25 :
26 : // Set if part of an EventuallyFileOnlySnapshot.
27 : efos *EventuallyFileOnlySnapshot
28 :
29 : // The list the snapshot is linked into.
30 : list *snapshotList
31 :
32 : // The next/prev link for the snapshotList doubly-linked list of snapshots.
33 : prev, next *Snapshot
34 : }
35 :
36 : var _ Reader = (*Snapshot)(nil)
37 :
38 : // Get gets the value for the given key. It returns ErrNotFound if the Snapshot
39 : // does not contain the key.
40 : //
41 : // The caller should not modify the contents of the returned slice, but it is
42 : // safe to modify the contents of the argument after Get returns. The returned
43 : // slice will remain valid until the returned Closer is closed. On success, the
44 : // caller MUST call closer.Close() or a memory leak will occur.
45 1 : func (s *Snapshot) Get(key []byte) ([]byte, io.Closer, error) {
46 1 : if s.db == nil {
47 0 : panic(ErrClosed)
48 : }
49 1 : return s.db.getInternal(key, nil /* batch */, s)
50 : }
51 :
52 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
53 : // return false). The iterator can be positioned via a call to SeekGE,
54 : // SeekLT, First or Last.
55 1 : func (s *Snapshot) NewIter(o *IterOptions) (*Iterator, error) {
56 1 : return s.NewIterWithContext(context.Background(), o)
57 1 : }
58 :
59 : // NewIterWithContext is like NewIter, and additionally accepts a context for
60 : // tracing.
61 1 : func (s *Snapshot) NewIterWithContext(ctx context.Context, o *IterOptions) (*Iterator, error) {
62 1 : if s.db == nil {
63 0 : panic(ErrClosed)
64 : }
65 1 : return s.db.newIter(ctx, nil /* batch */, newIterOpts{
66 1 : snapshot: snapshotIterOpts{seqNum: s.seqNum},
67 1 : }, o), nil
68 : }
69 :
70 : // ScanInternal scans all internal keys within the specified bounds, truncating
71 : // any rangedels and rangekeys to those bounds. For use when an external user
72 : // needs to be aware of all internal keys that make up a key range.
73 : //
74 : // See comment on db.ScanInternal for the behaviour that can be expected of
75 : // point keys deleted by range dels and keys masked by range keys.
76 : func (s *Snapshot) ScanInternal(
77 : ctx context.Context,
78 : category block.Category,
79 : lower, upper []byte,
80 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error,
81 : visitRangeDel func(start, end []byte, seqNum base.SeqNum) error,
82 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error,
83 : visitSharedFile func(sst *SharedSSTMeta) error,
84 : visitExternalFile func(sst *ExternalFile) error,
85 0 : ) error {
86 0 : if s.db == nil {
87 0 : panic(ErrClosed)
88 : }
89 0 : scanInternalOpts := &scanInternalOptions{
90 0 : category: category,
91 0 : visitPointKey: visitPointKey,
92 0 : visitRangeDel: visitRangeDel,
93 0 : visitRangeKey: visitRangeKey,
94 0 : visitSharedFile: visitSharedFile,
95 0 : visitExternalFile: visitExternalFile,
96 0 : IterOptions: IterOptions{
97 0 : KeyTypes: IterKeyTypePointsAndRanges,
98 0 : LowerBound: lower,
99 0 : UpperBound: upper,
100 0 : },
101 0 : }
102 0 :
103 0 : iter, err := s.db.newInternalIter(ctx, snapshotIterOpts{seqNum: s.seqNum}, scanInternalOpts)
104 0 : if err != nil {
105 0 : return err
106 0 : }
107 0 : defer iter.close()
108 0 :
109 0 : return scanInternalImpl(ctx, lower, upper, iter, scanInternalOpts)
110 : }
111 :
112 : // closeLocked is similar to Close(), except it requires that db.mu be held
113 : // by the caller.
114 1 : func (s *Snapshot) closeLocked() error {
115 1 : s.db.mu.snapshots.remove(s)
116 1 :
117 1 : // If s was the previous earliest snapshot, we might be able to reclaim
118 1 : // disk space by dropping obsolete records that were pinned by s.
119 1 : if e := s.db.mu.snapshots.earliest(); e > s.seqNum {
120 1 : s.db.maybeScheduleCompactionPicker(pickElisionOnly)
121 1 : }
122 1 : s.db = nil
123 1 : return nil
124 : }
125 :
126 : // Close closes the snapshot, releasing its resources. Close must be called.
127 : // Failure to do so will result in a tiny memory leak and a large leak of
128 : // resources on disk due to the entries the snapshot is preventing from being
129 : // deleted.
130 : //
131 : // d.mu must NOT be held by the caller.
132 1 : func (s *Snapshot) Close() error {
133 1 : db := s.db
134 1 : if db == nil {
135 0 : panic(ErrClosed)
136 : }
137 1 : db.mu.Lock()
138 1 : defer db.mu.Unlock()
139 1 : return s.closeLocked()
140 : }
141 :
142 : type snapshotList struct {
143 : root Snapshot
144 : }
145 :
146 1 : func (l *snapshotList) init() {
147 1 : l.root.next = &l.root
148 1 : l.root.prev = &l.root
149 1 : }
150 :
151 1 : func (l *snapshotList) empty() bool {
152 1 : return l.root.next == &l.root
153 1 : }
154 :
155 1 : func (l *snapshotList) count() int {
156 1 : if l.empty() {
157 1 : return 0
158 1 : }
159 0 : var count int
160 0 : for i := l.root.next; i != &l.root; i = i.next {
161 0 : count++
162 0 : }
163 0 : return count
164 : }
165 :
166 1 : func (l *snapshotList) earliest() base.SeqNum {
167 1 : v := base.SeqNum(math.MaxUint64)
168 1 : if !l.empty() {
169 1 : v = l.root.next.seqNum
170 1 : }
171 1 : return v
172 : }
173 :
174 1 : func (l *snapshotList) toSlice() []base.SeqNum {
175 1 : if l.empty() {
176 1 : return nil
177 1 : }
178 1 : var results []base.SeqNum
179 1 : for i := l.root.next; i != &l.root; i = i.next {
180 1 : results = append(results, i.seqNum)
181 1 : }
182 1 : return results
183 : }
184 :
185 1 : func (l *snapshotList) pushBack(s *Snapshot) {
186 1 : if s.list != nil || s.prev != nil || s.next != nil {
187 0 : panic("pebble: snapshot list is inconsistent")
188 : }
189 1 : s.prev = l.root.prev
190 1 : s.prev.next = s
191 1 : s.next = &l.root
192 1 : s.next.prev = s
193 1 : s.list = l
194 : }
195 :
196 1 : func (l *snapshotList) remove(s *Snapshot) {
197 1 : if s == &l.root {
198 0 : panic("pebble: cannot remove snapshot list root node")
199 : }
200 1 : if s.list != l {
201 0 : panic("pebble: snapshot list is inconsistent")
202 : }
203 1 : s.prev.next = s.next
204 1 : s.next.prev = s.prev
205 1 : s.next = nil // avoid memory leaks
206 1 : s.prev = nil // avoid memory leaks
207 1 : s.list = nil // avoid memory leaks
208 : }
209 :
210 : // EventuallyFileOnlySnapshot (aka EFOS) provides a read-only point-in-time view
211 : // of the database state, similar to Snapshot. An EventuallyFileOnlySnapshot
212 : // induces less write amplification than Snapshot, at the cost of increased space
213 : // amplification. While a Snapshot may increase write amplification across all
214 : // flushes and compactions for the duration of its lifetime, an
215 : // EventuallyFileOnlySnapshot only incurs that cost for flushes/compactions if
216 : // memtables at the time of EFOS instantiation contained keys that the EFOS is
217 : // interested in (i.e. its protectedRanges). In that case, the EFOS prevents
218 : // elision of keys visible to it, similar to a Snapshot, until those memtables
219 : // are flushed, and once that happens, the "EventuallyFileOnlySnapshot"
220 : // transitions to a file-only snapshot state in which it pins zombies sstables
221 : // like an open Iterator would, without pinning any memtables. Callers that can
222 : // tolerate the increased space amplification of pinning zombie sstables until
223 : // the snapshot is closed may prefer EventuallyFileOnlySnapshots for their
224 : // reduced write amplification. Callers that desire the benefits of the file-only
225 : // state that requires no pinning of memtables should call
226 : // `WaitForFileOnlySnapshot()` before relying on the EFOS to keep producing iterators
227 : // with zero write-amp and zero pinning of memtables in memory.
228 : //
229 : // EventuallyFileOnlySnapshots interact with the IngestAndExcise operation in
230 : // subtle ways. The IngestAndExcise can force the transition of an EFOS to a
231 : // file-only snapshot if an excise overlaps with the EFOS bounds.
232 : type EventuallyFileOnlySnapshot struct {
233 : mu struct {
234 : // NB: If both this mutex and db.mu are being grabbed, db.mu should be
235 : // grabbed _before_ grabbing this one.
236 : sync.Mutex
237 :
238 : // Either the snap field is set below, or the version is set at any given
239 : // point of time. If a snapshot is referenced, this is not a file-only
240 : // snapshot yet, and if a version is set (and ref'd) this is a file-only
241 : // snapshot.
242 :
243 : // The wrapped regular snapshot, if not a file-only snapshot yet.
244 : snap *Snapshot
245 : // The wrapped version reference, if a file-only snapshot.
246 : vers *version
247 : }
248 :
249 : // Key ranges to watch for an excise on.
250 : protectedRanges []KeyRange
251 :
252 : // The db the snapshot was created from.
253 : db *DB
254 : seqNum base.SeqNum
255 : closed chan struct{}
256 : }
257 :
258 1 : func (d *DB) makeEventuallyFileOnlySnapshot(keyRanges []KeyRange) *EventuallyFileOnlySnapshot {
259 1 : isFileOnly := true
260 1 :
261 1 : d.mu.Lock()
262 1 : defer d.mu.Unlock()
263 1 : seqNum := d.mu.versions.visibleSeqNum.Load()
264 1 : // Check if any of the keyRanges overlap with a memtable.
265 1 : for i := range d.mu.mem.queue {
266 1 : d.mu.mem.queue[i].computePossibleOverlaps(func(bounded) shouldContinue {
267 1 : isFileOnly = false
268 1 : return stopIteration
269 1 : }, sliceAsBounded(keyRanges)...)
270 : }
271 1 : es := &EventuallyFileOnlySnapshot{
272 1 : db: d,
273 1 : seqNum: seqNum,
274 1 : protectedRanges: keyRanges,
275 1 : closed: make(chan struct{}),
276 1 : }
277 1 : if isFileOnly {
278 1 : es.mu.vers = d.mu.versions.currentVersion()
279 1 : es.mu.vers.Ref()
280 1 : } else {
281 1 : s := &Snapshot{
282 1 : db: d,
283 1 : seqNum: seqNum,
284 1 : }
285 1 : s.efos = es
286 1 : es.mu.snap = s
287 1 : d.mu.snapshots.pushBack(s)
288 1 : }
289 1 : return es
290 : }
291 :
292 : // Transitions this EventuallyFileOnlySnapshot to a file-only snapshot. Requires
293 : // earliestUnflushedSeqNum and vers to correspond to the same Version from the
294 : // current or a past acquisition of db.mu. vers must have been Ref()'d before
295 : // that mutex was released, if it was released.
296 : //
297 : // NB: The caller is expected to check for es.excised before making this
298 : // call.
299 : //
300 : // d.mu must be held when calling this method.
301 1 : func (es *EventuallyFileOnlySnapshot) transitionToFileOnlySnapshot(vers *version) error {
302 1 : es.mu.Lock()
303 1 : select {
304 1 : case <-es.closed:
305 1 : vers.UnrefLocked()
306 1 : es.mu.Unlock()
307 1 : return ErrClosed
308 1 : default:
309 : }
310 1 : if es.mu.snap == nil {
311 0 : es.mu.Unlock()
312 0 : panic("pebble: tried to transition an eventually-file-only-snapshot twice")
313 : }
314 : // The caller has already called Ref() on vers.
315 1 : es.mu.vers = vers
316 1 : // NB: The callers should have already done a check of es.excised.
317 1 : oldSnap := es.mu.snap
318 1 : es.mu.snap = nil
319 1 : es.mu.Unlock()
320 1 : return oldSnap.closeLocked()
321 : }
322 :
323 : // hasTransitioned returns true if this EFOS has transitioned to a file-only
324 : // snapshot.
325 1 : func (es *EventuallyFileOnlySnapshot) hasTransitioned() bool {
326 1 : es.mu.Lock()
327 1 : defer es.mu.Unlock()
328 1 : return es.mu.vers != nil
329 1 : }
330 :
331 : // waitForFlush waits for a flush on any memtables that need to be flushed
332 : // before this EFOS can transition to a file-only snapshot. If this EFOS is
333 : // waiting on a flush of the mutable memtable, it forces a rotation within
334 : // `dur` duration. For immutable memtables, it schedules a flush and waits for
335 : // it to finish.
336 0 : func (es *EventuallyFileOnlySnapshot) waitForFlush(ctx context.Context, dur time.Duration) error {
337 0 : es.db.mu.Lock()
338 0 : defer es.db.mu.Unlock()
339 0 :
340 0 : earliestUnflushedSeqNum := es.db.getEarliestUnflushedSeqNumLocked()
341 0 : for earliestUnflushedSeqNum < es.seqNum {
342 0 : select {
343 0 : case <-es.closed:
344 0 : return ErrClosed
345 0 : case <-ctx.Done():
346 0 : return ctx.Err()
347 0 : default:
348 : }
349 : // Check if the current mutable memtable contains keys less than seqNum.
350 : // If so, rotate it.
351 0 : if es.db.mu.mem.mutable.logSeqNum < es.seqNum && dur.Nanoseconds() > 0 {
352 0 : es.db.maybeScheduleDelayedFlush(es.db.mu.mem.mutable, dur)
353 0 : } else {
354 0 : // Find the last memtable that contains seqNums less than es.seqNum,
355 0 : // and force a flush on it.
356 0 : var mem *flushableEntry
357 0 : for i := range es.db.mu.mem.queue {
358 0 : if es.db.mu.mem.queue[i].logSeqNum < es.seqNum {
359 0 : mem = es.db.mu.mem.queue[i]
360 0 : }
361 : }
362 0 : mem.flushForced = true
363 0 : es.db.maybeScheduleFlush()
364 : }
365 0 : es.db.mu.compact.cond.Wait()
366 0 :
367 0 : earliestUnflushedSeqNum = es.db.getEarliestUnflushedSeqNumLocked()
368 : }
369 0 : return nil
370 : }
371 :
372 : // WaitForFileOnlySnapshot blocks the calling goroutine until this snapshot
373 : // has been converted into a file-only snapshot (i.e. all memtables containing
374 : // keys < seqNum are flushed). A duration can be passed in, and if nonzero,
375 : // a delayed flush will be scheduled at that duration if necessary.
376 : //
377 : // Idempotent; can be called multiple times with no side effects.
378 : func (es *EventuallyFileOnlySnapshot) WaitForFileOnlySnapshot(
379 : ctx context.Context, dur time.Duration,
380 0 : ) error {
381 0 : if es.hasTransitioned() {
382 0 : return nil
383 0 : }
384 :
385 0 : if err := es.waitForFlush(ctx, dur); err != nil {
386 0 : return err
387 0 : }
388 :
389 0 : if invariants.Enabled {
390 0 : // Since we aren't returning an error, we _must_ have transitioned to a
391 0 : // file-only snapshot by now.
392 0 : if !es.hasTransitioned() {
393 0 : panic("expected EFOS to have transitioned to file-only snapshot after flush")
394 : }
395 : }
396 0 : return nil
397 : }
398 :
399 : // Close closes the file-only snapshot and releases all referenced resources.
400 : // Not idempotent.
401 1 : func (es *EventuallyFileOnlySnapshot) Close() error {
402 1 : close(es.closed)
403 1 : es.db.mu.Lock()
404 1 : defer es.db.mu.Unlock()
405 1 : es.mu.Lock()
406 1 : defer es.mu.Unlock()
407 1 :
408 1 : if es.mu.snap != nil {
409 1 : if err := es.mu.snap.closeLocked(); err != nil {
410 0 : return err
411 0 : }
412 : }
413 1 : if es.mu.vers != nil {
414 1 : es.mu.vers.UnrefLocked()
415 1 : }
416 1 : return nil
417 : }
418 :
419 : // Get implements the Reader interface.
420 1 : func (es *EventuallyFileOnlySnapshot) Get(key []byte) (value []byte, closer io.Closer, err error) {
421 1 : // TODO(jackson): Use getInternal.
422 1 : iter, err := es.NewIter(nil)
423 1 : if err != nil {
424 0 : return nil, nil, err
425 0 : }
426 1 : if !iter.SeekPrefixGE(key) {
427 1 : if err = firstError(iter.Error(), iter.Close()); err != nil {
428 0 : return nil, nil, err
429 0 : }
430 1 : return nil, nil, ErrNotFound
431 : }
432 1 : if !es.db.equal(iter.Key(), key) {
433 1 : return nil, nil, firstError(iter.Close(), ErrNotFound)
434 1 : }
435 1 : return iter.Value(), iter, nil
436 : }
437 :
438 : // NewIter returns an iterator that is unpositioned (Iterator.Valid() will
439 : // return false). The iterator can be positioned via a call to SeekGE,
440 : // SeekLT, First or Last.
441 1 : func (es *EventuallyFileOnlySnapshot) NewIter(o *IterOptions) (*Iterator, error) {
442 1 : return es.NewIterWithContext(context.Background(), o)
443 1 : }
444 :
445 : // NewIterWithContext is like NewIter, and additionally accepts a context for
446 : // tracing.
447 : func (es *EventuallyFileOnlySnapshot) NewIterWithContext(
448 : ctx context.Context, o *IterOptions,
449 1 : ) (*Iterator, error) {
450 1 : select {
451 0 : case <-es.closed:
452 0 : panic(ErrClosed)
453 1 : default:
454 : }
455 :
456 1 : es.mu.Lock()
457 1 : defer es.mu.Unlock()
458 1 : if es.mu.vers != nil {
459 1 : sOpts := snapshotIterOpts{seqNum: es.seqNum, vers: es.mu.vers}
460 1 : return es.db.newIter(ctx, nil /* batch */, newIterOpts{snapshot: sOpts}, o), nil
461 1 : }
462 :
463 1 : sOpts := snapshotIterOpts{seqNum: es.seqNum}
464 1 : iter := es.db.newIter(ctx, nil /* batch */, newIterOpts{snapshot: sOpts}, o)
465 1 : return iter, nil
466 : }
467 :
468 : // ScanInternal scans all internal keys within the specified bounds, truncating
469 : // any rangedels and rangekeys to those bounds. For use when an external user
470 : // needs to be aware of all internal keys that make up a key range.
471 : //
472 : // See comment on db.ScanInternal for the behaviour that can be expected of
473 : // point keys deleted by range dels and keys masked by range keys.
474 : func (es *EventuallyFileOnlySnapshot) ScanInternal(
475 : ctx context.Context,
476 : category block.Category,
477 : lower, upper []byte,
478 : visitPointKey func(key *InternalKey, value LazyValue, iterInfo IteratorLevel) error,
479 : visitRangeDel func(start, end []byte, seqNum base.SeqNum) error,
480 : visitRangeKey func(start, end []byte, keys []rangekey.Key) error,
481 : visitSharedFile func(sst *SharedSSTMeta) error,
482 : visitExternalFile func(sst *ExternalFile) error,
483 0 : ) error {
484 0 : if es.db == nil {
485 0 : panic(ErrClosed)
486 : }
487 0 : var sOpts snapshotIterOpts
488 0 : opts := &scanInternalOptions{
489 0 : category: category,
490 0 : IterOptions: IterOptions{
491 0 : KeyTypes: IterKeyTypePointsAndRanges,
492 0 : LowerBound: lower,
493 0 : UpperBound: upper,
494 0 : },
495 0 : visitPointKey: visitPointKey,
496 0 : visitRangeDel: visitRangeDel,
497 0 : visitRangeKey: visitRangeKey,
498 0 : visitSharedFile: visitSharedFile,
499 0 : visitExternalFile: visitExternalFile,
500 0 : }
501 0 : es.mu.Lock()
502 0 : if es.mu.vers != nil {
503 0 : sOpts = snapshotIterOpts{
504 0 : seqNum: es.seqNum,
505 0 : vers: es.mu.vers,
506 0 : }
507 0 : } else {
508 0 : sOpts = snapshotIterOpts{
509 0 : seqNum: es.seqNum,
510 0 : }
511 0 : }
512 0 : es.mu.Unlock()
513 0 : iter, err := es.db.newInternalIter(ctx, sOpts, opts)
514 0 : if err != nil {
515 0 : return err
516 0 : }
517 0 : defer iter.close()
518 0 :
519 0 : return scanInternalImpl(ctx, lower, upper, iter, opts)
520 : }
|
