fork of https://github.com/sourcegraph/zoekt
1// Copyright 2016 Google Inc. All rights reserved.
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7// http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15package search
16
17import (
18 "context"
19 "fmt"
20 "log"
21 "math"
22 "os"
23 "runtime"
24 "runtime/debug"
25 "slices"
26 "sort"
27 "strconv"
28 "sync"
29 "time"
30
31 "github.com/prometheus/client_golang/prometheus"
32 "github.com/prometheus/client_golang/prometheus/promauto"
33 "go.uber.org/atomic"
34 "golang.org/x/sync/semaphore"
35
36 "github.com/sourcegraph/zoekt"
37 "github.com/sourcegraph/zoekt/index"
38 "github.com/sourcegraph/zoekt/internal/tenant/systemtenant"
39 "github.com/sourcegraph/zoekt/internal/trace"
40 "github.com/sourcegraph/zoekt/query"
41)
42
43var (
44 metricShardsLoaded = promauto.NewGauge(prometheus.GaugeOpts{
45 Name: "zoekt_shards_loaded",
46 Help: "The number of shards currently loaded",
47 })
48 metricShardsLoadedTotal = promauto.NewCounter(prometheus.CounterOpts{
49 Name: "zoekt_shards_loaded_total",
50 Help: "The total number of shards loaded",
51 })
52 metricShardsLoadFailedTotal = promauto.NewCounter(prometheus.CounterOpts{
53 Name: "zoekt_shards_load_failed_total",
54 Help: "The total number of shard loads that failed",
55 })
56
57 metricSearchRunning = promauto.NewGauge(prometheus.GaugeOpts{
58 Name: "zoekt_search_running",
59 Help: "The number of concurrent search requests running",
60 })
61 metricSearchShardRunning = promauto.NewGauge(prometheus.GaugeOpts{
62 Name: "zoekt_search_shard_running",
63 Help: "The number of concurrent search requests in a shard running",
64 })
65 metricSearchFailedTotal = promauto.NewCounter(prometheus.CounterOpts{
66 Name: "zoekt_search_failed_total",
67 Help: "The total number of search requests that failed",
68 })
69 metricSearchDuration = promauto.NewHistogram(prometheus.HistogramOpts{
70 Name: "zoekt_search_duration_seconds",
71 Help: "The duration a search request took in seconds",
72 Buckets: prometheus.DefBuckets, // DefBuckets good for service timings
73 })
74
75 // A Counter per Stat. Name should match field in zoekt.Stats.
76 metricSearchContentBytesLoadedTotal = promauto.NewCounter(prometheus.CounterOpts{
77 Name: "zoekt_search_content_loaded_bytes_total",
78 Help: "Total amount of I/O for reading contents",
79 })
80 metricSearchIndexBytesLoadedTotal = promauto.NewCounter(prometheus.CounterOpts{
81 Name: "zoekt_search_index_loaded_bytes_total",
82 Help: "Total amount of I/O for reading from index",
83 })
84 metricSearchCrashesTotal = promauto.NewCounter(prometheus.CounterOpts{
85 Name: "zoekt_search_crashes_total",
86 Help: "Total number of search shards that had a crash",
87 })
88 metricSearchFileCountTotal = promauto.NewCounter(prometheus.CounterOpts{
89 Name: "zoekt_search_file_count_total",
90 Help: "Total number of files containing a match",
91 })
92 metricSearchShardFilesConsideredTotal = promauto.NewCounter(prometheus.CounterOpts{
93 Name: "zoekt_search_shard_files_considered_total",
94 Help: "Total number of files in shards that we considered",
95 })
96 metricSearchFilesConsideredTotal = promauto.NewCounter(prometheus.CounterOpts{
97 Name: "zoekt_search_files_considered_total",
98 Help: "Total files that we evaluated. Equivalent to files for which all atom matches (including negations) evaluated to true",
99 })
100 metricSearchFilesLoadedTotal = promauto.NewCounter(prometheus.CounterOpts{
101 Name: "zoekt_search_files_loaded_total",
102 Help: "Total files for which we loaded file content to verify substring matches",
103 })
104 metricSearchFilesSkippedTotal = promauto.NewCounter(prometheus.CounterOpts{
105 Name: "zoekt_search_files_skipped_total",
106 Help: "Total candidate files whose contents weren't examined because we gathered enough matches",
107 })
108 metricSearchShardsSkippedTotal = promauto.NewCounter(prometheus.CounterOpts{
109 Name: "zoekt_search_shards_skipped_total",
110 Help: "Total shards that we did not process because a query was canceled",
111 })
112 metricSearchMatchCountTotal = promauto.NewCounter(prometheus.CounterOpts{
113 Name: "zoekt_search_match_count_total",
114 Help: "Total number of non-overlapping matches",
115 })
116 metricSearchNgramMatchesTotal = promauto.NewCounter(prometheus.CounterOpts{
117 Name: "zoekt_search_ngram_matches_total",
118 Help: "Total number of candidate matches as a result of searching ngrams",
119 })
120 metricSearchNgramLookupsTotal = promauto.NewCounter(prometheus.CounterOpts{
121 Name: "zoekt_search_ngram_lookups_total",
122 Help: "Total number of times we accessed an ngram in the index",
123 })
124 metricSearchRegexpsConsideredTotal = promauto.NewCounter(prometheus.CounterOpts{
125 Name: "zoekt_search_regexps_considered_total",
126 Help: "Total number of times regexp was called on files that we evaluated",
127 })
128
129 metricListRunning = promauto.NewGauge(prometheus.GaugeOpts{
130 Name: "zoekt_list_running",
131 Help: "The number of concurrent list requests running",
132 })
133 metricListShardRunning = promauto.NewGauge(prometheus.GaugeOpts{
134 Name: "zoekt_list_shard_running",
135 Help: "The number of concurrent list requests in a shard running",
136 })
137 metricShardsBatchReplaceDurationSeconds = promauto.NewHistogram(prometheus.HistogramOpts{
138 Name: "zoekt_shards_batch_replace_duration_seconds",
139 Help: "The time it takes to replace a batch of Searchers.",
140 Buckets: []float64{0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1, 2, 5, 10, 30},
141 })
142 metricListAllRepos = promauto.NewGauge(prometheus.GaugeOpts{
143 Name: "zoekt_list_all_stats_repos",
144 Help: "The last List(true) value for RepoStats.Repos. Repos is used for aggregrating the number of repositories.",
145 })
146 metricListAllShards = promauto.NewGauge(prometheus.GaugeOpts{
147 Name: "zoekt_list_all_stats_shards",
148 Help: "The last List(true) value for RepoStats.Shards. Shards is the total number of search shards.",
149 })
150 metricListAllDocuments = promauto.NewGauge(prometheus.GaugeOpts{
151 Name: "zoekt_list_all_stats_documents",
152 Help: "The last List(true) value for RepoStats.Documents. Documents holds the number of documents or files.",
153 })
154 metricListAllIndexBytes = promauto.NewGauge(prometheus.GaugeOpts{
155 Name: "zoekt_list_all_stats_index_bytes",
156 Help: "The last List(true) value for RepoStats.IndexBytes. IndexBytes is the amount of RAM used for index overhead.",
157 })
158 metricListAllContentBytes = promauto.NewGauge(prometheus.GaugeOpts{
159 Name: "zoekt_list_all_stats_content_bytes",
160 Help: "The last List(true) value for RepoStats.ContentBytes. ContentBytes is the amount of RAM used for raw content.",
161 })
162 metricListAllNewLinesCount = promauto.NewGauge(prometheus.GaugeOpts{
163 Name: "zoekt_list_all_stats_new_lines_count",
164 Help: "The last List(true) value for RepoStats.NewLinesCount.",
165 })
166 metricListAllDefaultBranchNewLinesCount = promauto.NewGauge(prometheus.GaugeOpts{
167 Name: "zoekt_list_all_stats_default_branch_new_lines_count",
168 Help: "The last List(true) value for RepoStats.DefaultBranchNewLinesCount.",
169 })
170 metricListAllOtherBranchesNewLinesCount = promauto.NewGauge(prometheus.GaugeOpts{
171 Name: "zoekt_list_all_stats_other_branches_new_lines_count",
172 Help: "The last List(true) value for RepoStats.OtherBranchesNewLinesCount.",
173 })
174)
175
176type rankedShard struct {
177 zoekt.Searcher
178
179 priority float64 // maximum priority across all repos in the shard
180
181 // We have out of band ranking on compound shards which can change even if
182 // the shard file does not. So we compute a rank in getShards. We store
183 // repos here to avoid the cost of List in the search request path.
184 //
185 // repos is nil only if that call failed.
186 repos []*zoekt.Repository
187}
188
189// loaded stores the state we compute when updating the state of shards from
190// disk.
191type loaded struct {
192 // shards is the currently loaded shards sorted by decreasing rank and
193 // should not be mutated.
194 shards []*rankedShard
195
196 // ready is true if sharded searcher has finished loading all initial
197 // shards on startup.
198 ready bool
199}
200
201type shardedSearcher struct {
202 // Limit the number of parallel queries. Since searching is
203 // CPU bound, we can't do better than #CPU queries in
204 // parallel. If we do so, we just create more memory
205 // pressure.
206 sched scheduler
207
208 mu sync.Mutex // protects writes to shards
209 shards map[string]*rankedShard
210
211 ready atomic.Bool
212 ranked atomic.Value
213}
214
215func newShardedSearcher(n int64) *shardedSearcher {
216 ss := &shardedSearcher{
217 shards: make(map[string]*rankedShard),
218 sched: newScheduler(n),
219 }
220 return ss
221}
222
223// NewDirectorySearcher returns a searcher instance that loads all
224// shards corresponding to a glob into memory.
225func NewDirectorySearcher(dir string) (zoekt.Streamer, error) {
226 return newDirectorySearcher(dir, true)
227}
228
229// NewDirectorySearcherFast is like NewDirectorySearcher, but does not block
230// on the initial loading of shards.
231//
232// This exists since in the case of zoekt-webserver we are happy with having
233// partial availability since that is better than no availability on large
234// instances.
235func NewDirectorySearcherFast(dir string) (zoekt.Streamer, error) {
236 return newDirectorySearcher(dir, false)
237}
238
239func newDirectorySearcher(dir string, waitUntilReady bool) (zoekt.Streamer, error) {
240 ss := newShardedSearcher(int64(runtime.GOMAXPROCS(0)))
241 tl := &loader{
242 ss: ss,
243 }
244 dw, err := newDirectoryWatcher(dir, tl)
245 if err != nil {
246 return nil, err
247 }
248
249 if waitUntilReady {
250 if err := dw.WaitUntilReady(); err != nil {
251 return nil, err
252 }
253 }
254
255 ds := &directorySearcher{
256 Streamer: ss,
257 directoryWatcher: dw,
258 }
259
260 return &typeRepoSearcher{Streamer: ds}, nil
261}
262
263type directorySearcher struct {
264 zoekt.Streamer
265
266 directoryWatcher *DirectoryWatcher
267}
268
269func (s *directorySearcher) Close() {
270 // We need to Stop directoryWatcher first since it calls load/unload on
271 // Searcher.
272 s.directoryWatcher.Stop()
273 s.Streamer.Close()
274}
275
276type loader struct {
277 ss *shardedSearcher
278}
279
280func (tl *loader) load(keys ...string) {
281 // This is called with all keys on startup, so once this function has
282 // finished running shardedSearcher will be ready.
283 defer tl.ss.markReady()
284
285 if len(keys) == 0 {
286 // If there's nothing to load, we exit early here, but we want to mark
287 // ourselves as ready.
288 return
289 }
290
291 var (
292 mu sync.Mutex // synchronizes writes to the shards map
293 wg sync.WaitGroup // used to wait for all shards to load
294 sem = semaphore.NewWeighted(int64(runtime.GOMAXPROCS(0)))
295 loadedShards = make(map[string]zoekt.Searcher)
296 )
297
298 publishLoaded := func() {
299 mu.Lock()
300 chunk := loadedShards
301 loadedShards = make(map[string]zoekt.Searcher)
302 mu.Unlock()
303 tl.ss.replace(chunk)
304 }
305
306 log.Printf("[INFO] loading %d shard(s): %s", len(keys), humanTruncateList(keys, 5))
307
308 lastProgress := time.Now()
309 for i, key := range keys {
310 // If taking a while to start-up occasionally give a progress message
311 if time.Since(lastProgress) > 5*time.Second {
312 log.Printf("[INFO] still need to load %d shards...", len(keys)-i)
313 lastProgress = time.Now()
314
315 publishLoaded()
316 }
317
318 _ = sem.Acquire(context.Background(), 1)
319 wg.Add(1)
320
321 go func(key string) {
322 defer sem.Release(1)
323 defer wg.Done()
324
325 shard, err := loadShard(key)
326 if err != nil {
327 metricShardsLoadFailedTotal.Inc()
328 log.Printf("[ERROR] reloading: %s, err %v ", key, err)
329 return
330 }
331 metricShardsLoadedTotal.Inc()
332
333 mu.Lock()
334 loadedShards[key] = shard
335 mu.Unlock()
336 }(key)
337 }
338
339 wg.Wait()
340
341 publishLoaded()
342}
343
344func (tl *loader) drop(keys ...string) {
345 shards := make(map[string]zoekt.Searcher, len(keys))
346 for _, key := range keys {
347 shards[key] = nil
348 }
349 tl.ss.replace(shards)
350}
351
352func (ss *shardedSearcher) String() string {
353 return "shardedSearcher"
354}
355
356// Close closes references to open files. It may be called only once.
357func (ss *shardedSearcher) Close() {
358 ss.mu.Lock()
359 shards := make(map[string]zoekt.Searcher, len(ss.shards))
360 for k := range ss.shards {
361 shards[k] = nil
362 }
363 ss.mu.Unlock()
364
365 ss.replace(shards)
366}
367
368func selectRepoSet(shards []*rankedShard, q query.Q) ([]*rankedShard, query.Q) {
369 and, ok := q.(*query.And)
370 if ok {
371 return doSelectRepoSet(shards, and)
372 }
373
374 // We have queries which look like (reposet ...) and we want to do the same
375 // optimizations. To simplify we just always wrap the query in And and then
376 // on the return value call Simplify to unwrap. In particular this is
377 // important for List calls.
378 and = &query.And{Children: []query.Q{q}}
379 shards, q = doSelectRepoSet(shards, and)
380 return shards, query.Simplify(q)
381}
382
383func doSelectRepoSet(shards []*rankedShard, and *query.And) ([]*rankedShard, query.Q) {
384 // (and (reposet ...) (q))
385 // (and true (q)) with a filtered shards
386 // (and false) // noop
387
388 // (and (repobranches ...) (q))
389 // (and (repobranches ...) (q))
390
391 // Note: we also support (and (repo ...) (q)) even though sourcegraph does
392 // not generate those sorts of queries. This is to support manual testing.
393
394 hasReposForPredicate := func(pred func(repo *zoekt.Repository) bool) func(repos []*zoekt.Repository) (any, all bool) {
395 return func(repos []*zoekt.Repository) (any, all bool) {
396 any = false
397 all = true
398 for _, repo := range repos {
399 b := pred(repo)
400 any = any || b
401 all = all && b
402 }
403 return any, all
404 }
405 }
406
407 for i, c := range and.Children {
408 var setSize int
409 var hasRepos func([]*zoekt.Repository) (bool, bool)
410 switch setQuery := c.(type) {
411 case *query.RepoSet:
412 setSize = len(setQuery.Set)
413 hasRepos = hasReposForPredicate(func(repo *zoekt.Repository) bool {
414 return setQuery.Set[repo.Name]
415 })
416 case *query.RepoIDs:
417 setSize = int(setQuery.Repos.GetCardinality())
418 hasRepos = hasReposForPredicate(func(repo *zoekt.Repository) bool {
419 return setQuery.Repos.Contains(repo.ID)
420 })
421 case *query.Repo:
422 setSize = 0
423 hasRepos = hasReposForPredicate(func(repo *zoekt.Repository) bool {
424 return setQuery.Regexp.MatchString(repo.Name)
425 })
426 case *query.BranchesRepos:
427 for _, br := range setQuery.List {
428 setSize += int(br.Repos.GetCardinality())
429 }
430
431 hasRepos = hasReposForPredicate(func(repo *zoekt.Repository) bool {
432 for _, br := range setQuery.List {
433 if br.Repos.Contains(repo.ID) {
434 return true
435 }
436 }
437 return false
438 })
439 default:
440 continue
441 }
442
443 // setSize may be larger than the number of shards we have. The size of
444 // filtered is bounded by min(len(set), len(shards))
445 if setSize > len(shards) {
446 setSize = len(shards)
447 }
448
449 filtered := make([]*rankedShard, 0, setSize)
450 filteredAll := true
451
452 for _, s := range shards {
453 if s.repos == nil {
454 // repos is nil if we failed to List the shard. This shouldn't
455 // happen, but if it does we don't know what is in it and must search
456 // it without simplifying the query.
457 filtered = append(filtered, s)
458 filteredAll = false
459 } else if any, all := hasRepos(s.repos); any {
460 filtered = append(filtered, s)
461 filteredAll = filteredAll && all
462 }
463 }
464
465 // We don't need to adjust the query since we are returning an empty set
466 // of shards to search.
467 if len(filtered) == 0 {
468 return filtered, and
469 }
470
471 // We can't simplify the query since we are searching shards which contain
472 // repos we aren't supposed to search.
473 if !filteredAll {
474 return filtered, and
475 }
476
477 // We don't want to mutate the original and, so we clone it before
478 // mutating it.
479 and = &query.And{Children: slices.Clone(and.Children)}
480
481 // This optimization allows us to avoid the work done by
482 // indexData.simplify for each shard.
483 //
484 // For example if our query is (and (reposet foo bar) (content baz))
485 // then at this point filtered is [foo bar] and q is the same. For each
486 // shard indexData.simplify will simplify to (and true (content baz)) ->
487 // (content baz). This work can be done now once, rather than per shard.
488 switch c := c.(type) {
489 case *query.RepoSet, *query.RepoIDs, *query.Repo:
490 and.Children[i] = &query.Const{Value: true}
491 return filtered, query.Simplify(and)
492
493 case *query.BranchesRepos:
494 // We can only replace if all the repos want the same branches. We
495 // simplify and just check that we are requesting 1 branch. The common
496 // case is just asking for HEAD, so this should be effective.
497 if len(c.List) != 1 {
498 return filtered, and
499 }
500
501 // Every repo wants the same branches, so we can replace RepoBranches
502 // with a list of branch queries.
503 and.Children[i] = &query.Branch{Pattern: c.List[0].Branch, Exact: true}
504 return filtered, query.Simplify(and)
505 }
506
507 // Stop after first RepoSet, otherwise we might append duplicate
508 // shards to `filtered`
509 return filtered, and
510 }
511
512 return shards, and
513}
514
515func (ss *shardedSearcher) Search(ctx context.Context, q query.Q, opts *zoekt.SearchOptions) (sr *zoekt.SearchResult, err error) {
516 tr, ctx := trace.New(ctx, "shardedSearcher.Search", "")
517 tr.LazyLog(q, true)
518 tr.LazyPrintf("opts: %+v", opts)
519 defer func() {
520 if sr != nil {
521 tr.LazyPrintf("num files: %d", len(sr.Files))
522 tr.LazyPrintf("stats: %+v", sr.Stats)
523 }
524 if err != nil {
525 tr.LazyPrintf("error: %v", err)
526 tr.SetError(err)
527 }
528 tr.Finish()
529 }()
530 ctx, cancel := context.WithCancel(ctx)
531 defer cancel()
532
533 collectSender := newCollectSender(opts)
534
535 start := time.Now()
536 proc, err := ss.sched.Acquire(ctx)
537 if err != nil {
538 return nil, err
539 }
540 defer proc.Release()
541 tr.LazyPrintf("acquired process")
542
543 wait := time.Since(start)
544 start = time.Now()
545
546 loaded := ss.getLoaded()
547 done, err := streamSearch(ctx, proc, q, opts, loaded.shards, collectSender)
548 defer done()
549 if err != nil {
550 return nil, err
551 }
552
553 aggregate, ok := collectSender.Done()
554 if !ok {
555 aggregate = &zoekt.SearchResult{
556 RepoURLs: map[string]string{},
557 LineFragments: map[string]string{},
558 }
559 }
560
561 copyFiles(aggregate)
562
563 if !loaded.ready {
564 // We may have missed results due to not being fully loaded.
565 aggregate.Stats.Crashes++
566 }
567
568 aggregate.Stats.Wait = wait
569 aggregate.Stats.Duration = time.Since(start)
570
571 return aggregate, nil
572}
573
574func (ss *shardedSearcher) StreamSearch(ctx context.Context, q query.Q, opts *zoekt.SearchOptions, sender zoekt.Sender) (err error) {
575 tr, ctx := trace.New(ctx, "shardedSearcher.StreamSearch", "")
576 defer func() {
577 if err != nil {
578 tr.LazyPrintf("error: %v", err)
579 tr.SetError(err)
580 }
581 tr.Finish()
582 }()
583
584 start := time.Now()
585 proc, err := ss.sched.Acquire(ctx)
586 if err != nil {
587 return err
588 }
589 defer proc.Release()
590 tr.LazyPrintf("acquired process")
591
592 loaded := ss.getLoaded()
593 shards := loaded.shards
594
595 maxPendingPriority := math.Inf(-1)
596 if len(shards) > 0 {
597 maxPendingPriority = shards[0].priority
598 }
599
600 stillLoadingCrashes := 0
601 if !loaded.ready {
602 // We may have missed results due to not being fully loaded.
603 stillLoadingCrashes++
604 }
605
606 sender.Send(&zoekt.SearchResult{
607 Stats: zoekt.Stats{
608 Crashes: stillLoadingCrashes,
609 Wait: time.Since(start),
610 },
611 Progress: zoekt.Progress{
612 MaxPendingPriority: maxPendingPriority,
613 },
614 })
615
616 // Matches flow from the shards up the stack in the following order:
617 //
618 // 1. Search shards
619 // 2. flushCollectSender (aggregate)
620 // 3. limitSender (limit)
621 // 4. copyFileSender (copy)
622 //
623 // For streaming, the wrapping has to happen in the inverted order.
624 sender = copyFileSender(sender)
625
626 if truncator, hasLimits := index.NewDisplayTruncator(opts); hasLimits {
627 var cancel context.CancelFunc
628 ctx, cancel = context.WithCancel(ctx)
629 defer cancel()
630 sender = limitSender(cancel, sender, truncator)
631 }
632
633 sender, flush := newFlushCollectSender(opts, sender)
634
635 done, err := streamSearch(ctx, proc, q, opts, shards, sender)
636
637 // Even though streaming is done, we may have results sitting in a buffer we
638 // need to flush. So we need to send those before calling done.
639 flush()
640 done()
641
642 return err
643}
644
645// streamSearch is an internal helper since both Search and StreamSearch are
646// largely similar.
647//
648// done must always be called, even if err is non-nil. The SearchResults sent
649// via sender contain references to the underlying mmap data that the garbage
650// collector can't see. Calling done informs the garbage collector it is free
651// to collect those shards. The caller must call copyFiles on any
652// SearchResults it returns/streams out before calling done.
653func streamSearch(ctx context.Context, proc *process, q query.Q, opts *zoekt.SearchOptions, shards []*rankedShard, sender zoekt.Sender) (done func(), err error) {
654 tr, ctx := trace.New(ctx, "shardedSearcher.streamSearch", "")
655 overallStart := time.Now()
656 metricSearchRunning.Inc()
657 defer func() {
658 metricSearchRunning.Dec()
659 metricSearchDuration.Observe(time.Since(overallStart).Seconds())
660 if err != nil {
661 metricSearchFailedTotal.Inc()
662
663 tr.LazyPrintf("error: %v", err)
664 tr.SetError(err)
665 }
666 tr.Finish()
667 }()
668
669 // Select the subset of shards that we will search over for the given query.
670 {
671 beforeLen := len(shards)
672 beforeQ := q
673 shards, q = selectRepoSet(shards, q)
674 tr.LazyPrintf("selectRepoSet shards=%d->%d q=%s->%s", beforeLen, len(shards), beforeQ, q)
675 }
676
677 if len(shards) == 0 {
678 return func() {}, nil
679 }
680
681 var cancel context.CancelFunc
682 if opts.MaxWallTime == 0 {
683 ctx, cancel = context.WithCancel(ctx)
684 } else {
685 ctx, cancel = context.WithTimeout(ctx, opts.MaxWallTime)
686 }
687
688 defer cancel()
689
690 // We set the number of workers to GOMAXPROCS, or the number of shards,
691 // whichever is smaller.
692 workers := min(runtime.GOMAXPROCS(0), len(shards))
693
694 type result struct {
695 priority float64
696 *zoekt.SearchResult
697 err error
698 }
699
700 var (
701 // buffered channels to continue searching when sending back results
702 // takes a while / blocks. The maximum pending result set is workers * 2.
703 results = make(chan *result, workers)
704 search = make(chan *rankedShard, workers)
705 wg sync.WaitGroup
706 )
707
708 // Start workers that receive shards from the search channel, search them,
709 // and send the results to the results channel. This process is repeated
710 // until the search channel is closed.
711 //
712 // Note: Making "search" a buffered channel has the effect of limiting the number of parallel shard searches.
713 // Since searching is mostly CPU bound, limiting parallel shard searches also reduces the peak working set.
714 wg.Add(workers)
715 for range workers {
716 go func() {
717 defer wg.Done()
718 for s := range search {
719 sr, err := searchOneShard(ctx, s, q, opts)
720 r := &result{priority: s.priority, SearchResult: sr, err: err}
721 results <- r
722 }
723 }()
724 }
725
726 go func() {
727 wg.Wait()
728 close(results)
729 }()
730
731 var (
732 pending = make(prioritySlice, 0, workers)
733 shard = 0
734 next = shards[shard]
735
736 // We need a separate nil-able reference to the same channel so we can close(search) for the worker
737 // go-routines to finish but also set work to nil in order for the select statement below to ignore
738 // that case when we want to stop a search. This is needed because sending on a closed channel panics.
739 work = search
740 )
741
742 stop := func() {
743 if work != nil {
744 close(search)
745 work = nil
746 next = nil
747 }
748 }
749
750 // tracked so we can stop when we hit TotalMaxMatchCount
751 var totalMatchCount int
752
753search:
754 for {
755 // At the top of each iteration, have the proc associated with this search yield its won "timeslice"
756 // to possibly allow other searches to make progress
757 _ = proc.Yield(ctx) // Note: we let searchOneShard handle context errors
758
759 select {
760 case work <- next: // is there a worker available to search the next shard?
761 pending.append(next.priority)
762
763 shard++
764 if shard == len(shards) {
765 stop()
766 } else {
767 next = shards[shard]
768 }
769 case r, ok := <-results: // is there a result to send back?
770 if !ok {
771 break search
772 }
773
774 // delete this result's priority from pending before computing the new max pending priority
775 pending.remove(r.priority)
776
777 if r.err != nil {
778 // Set final error and stop searching new shards, but consume any pending
779 // search results.
780 stop()
781 err = r.err
782 continue
783 }
784
785 // Update the match count statistics and stop searching new shards if we've
786 // reached the limit set in the options.
787 totalMatchCount += r.SearchResult.Stats.MatchCount
788 if opts.TotalMaxMatchCount > 0 && totalMatchCount > opts.TotalMaxMatchCount {
789 stop()
790 }
791
792 observeMetrics(r.SearchResult)
793
794 r.Priority = r.priority
795 r.MaxPendingPriority = pending.max()
796
797 sendByRepository(r.SearchResult, opts, sender) // send the result back to the client
798 }
799 }
800
801 return func() { runtime.KeepAlive(shards) }, err
802}
803
804// sendByRepository splits a zoekt.SearchResult by repository and calls
805// sender.Send for each batch. Ranking in Sourcegraph expects zoekt.SearchResult
806// to contain results with the same zoekt.SearchResult.priority only.
807//
808// We split by repository instead of by priority because it is easier to set
809// RepoURLs and LineFragments in zoekt.SearchResult.
810func sendByRepository(result *zoekt.SearchResult, opts *zoekt.SearchOptions, sender zoekt.Sender) {
811 if len(result.RepoURLs) <= 1 || len(result.Files) == 0 {
812 index.SortFiles(result.Files)
813 sender.Send(result)
814 return
815 }
816
817 send := func(repoName string, a, b int, stats zoekt.Stats) {
818 index.SortFiles(result.Files[a:b])
819
820 filteredRepoURLs := map[string]string{repoName: result.RepoURLs[repoName]}
821 filteredLineFragments := map[string]string{repoName: result.LineFragments[repoName]}
822
823 // Filter RepoURLs and LineFragments to only those of repoName and its
824 // subRepositories if there are subRepositories
825 for _, file := range result.Files[a:b] {
826 name := file.SubRepositoryName
827 if name == "" {
828 continue
829 }
830 _, repoSet := filteredRepoURLs[name]
831 url, ok := result.RepoURLs[name]
832 if !repoSet && ok {
833 filteredRepoURLs[name] = url
834 }
835 _, fragSet := filteredLineFragments[name]
836 frag, ok := result.LineFragments[name]
837 if !fragSet && ok {
838 filteredLineFragments[name] = frag
839 }
840 }
841
842 sender.Send(&zoekt.SearchResult{
843 Stats: stats,
844 Progress: zoekt.Progress{
845 Priority: result.Files[a].RepositoryPriority,
846 MaxPendingPriority: result.MaxPendingPriority,
847 },
848 Files: result.Files[a:b],
849 RepoURLs: filteredRepoURLs,
850 LineFragments: filteredLineFragments,
851 })
852 }
853
854 var startIndex, endIndex int
855 curRepoID := result.Files[0].RepositoryID
856 curRepoName := result.Files[0].Repository
857
858 fm := zoekt.FileMatch{}
859 for endIndex, fm = range result.Files {
860 if curRepoID != fm.RepositoryID {
861 // Stats must stay aggregate-able, hence we sent the aggregate stats with the
862 // last event.
863 send(curRepoName, startIndex, endIndex, zoekt.Stats{})
864
865 startIndex = endIndex
866 curRepoID = fm.RepositoryID
867 curRepoName = fm.Repository
868 }
869 }
870
871 send(curRepoName, startIndex, endIndex+1, result.Stats)
872}
873
874func observeMetrics(sr *zoekt.SearchResult) {
875 metricSearchContentBytesLoadedTotal.Add(float64(sr.Stats.ContentBytesLoaded))
876 metricSearchIndexBytesLoadedTotal.Add(float64(sr.Stats.IndexBytesLoaded))
877 metricSearchCrashesTotal.Add(float64(sr.Stats.Crashes))
878 metricSearchFileCountTotal.Add(float64(sr.Stats.FileCount))
879 metricSearchShardFilesConsideredTotal.Add(float64(sr.Stats.ShardFilesConsidered))
880 metricSearchFilesConsideredTotal.Add(float64(sr.Stats.FilesConsidered))
881 metricSearchFilesLoadedTotal.Add(float64(sr.Stats.FilesLoaded))
882 metricSearchFilesSkippedTotal.Add(float64(sr.Stats.FilesSkipped))
883 metricSearchShardsSkippedTotal.Add(float64(sr.Stats.ShardsSkipped))
884 metricSearchMatchCountTotal.Add(float64(sr.Stats.MatchCount))
885 metricSearchNgramMatchesTotal.Add(float64(sr.Stats.NgramMatches))
886 metricSearchNgramLookupsTotal.Add(float64(sr.Stats.NgramLookups))
887 metricSearchRegexpsConsideredTotal.Add(float64(sr.Stats.RegexpsConsidered))
888}
889
890func copySlice(src *[]byte) {
891 if *src == nil {
892 return
893 }
894 dst := make([]byte, len(*src))
895 copy(dst, *src)
896 *src = dst
897}
898
899func copyFiles(sr *zoekt.SearchResult) {
900 for i := range sr.Files {
901 copySlice(&sr.Files[i].Content)
902 copySlice(&sr.Files[i].Checksum)
903 for l := range sr.Files[i].LineMatches {
904 copySlice(&sr.Files[i].LineMatches[l].Line)
905 copySlice(&sr.Files[i].LineMatches[l].Before)
906 copySlice(&sr.Files[i].LineMatches[l].After)
907 }
908 for c := range sr.Files[i].ChunkMatches {
909 copySlice(&sr.Files[i].ChunkMatches[c].Content)
910 }
911 }
912}
913
914func searchOneShard(ctx context.Context, s zoekt.Searcher, q query.Q, opts *zoekt.SearchOptions) (sr *zoekt.SearchResult, err error) {
915 metricSearchShardRunning.Inc()
916 defer func() {
917 metricSearchShardRunning.Dec()
918 if e := recover(); e != nil {
919 log.Printf("[ERROR] crashed shard: %s: %#v, %s", s, e, debug.Stack())
920
921 if sr == nil {
922 sr = &zoekt.SearchResult{}
923 }
924 sr.Stats.Crashes = 1
925 }
926 }()
927
928 return s.Search(ctx, q, opts)
929}
930
931type shardListResult struct {
932 rl *zoekt.RepoList
933 err error
934}
935
936func listOneShard(ctx context.Context, s zoekt.Searcher, q query.Q, opts *zoekt.ListOptions, sink chan shardListResult) {
937 metricListShardRunning.Inc()
938 defer func() {
939 metricListShardRunning.Dec()
940 if r := recover(); r != nil {
941 log.Printf("[ERROR] crashed shard: %s: %s, %s", s.String(), r, debug.Stack())
942 sink <- shardListResult{
943 &zoekt.RepoList{Crashes: 1}, nil,
944 }
945 }
946 }()
947
948 ms, err := s.List(ctx, q, opts)
949 sink <- shardListResult{ms, err}
950}
951
952func (ss *shardedSearcher) List(ctx context.Context, q query.Q, opts *zoekt.ListOptions) (rl *zoekt.RepoList, err error) {
953 tr, ctx := trace.New(ctx, "shardedSearcher.List", "")
954 metricListRunning.Inc()
955 defer func() {
956 metricListRunning.Dec()
957 if rl != nil {
958 tr.LazyPrintf("repos.size=%d reposmap.size=%d crashes=%d stats=%+v", len(rl.Repos), len(rl.ReposMap), rl.Crashes, rl.Stats)
959 }
960 if err != nil {
961 tr.LazyPrintf("error: %v", err)
962 tr.SetError(err)
963 }
964 tr.Finish()
965 }()
966
967 q = query.Simplify(q)
968 isAll := false
969 if c, ok := q.(*query.Const); ok {
970 isAll = c.Value
971 }
972
973 proc, err := ss.sched.Acquire(ctx)
974 if err != nil {
975 return nil, err
976 }
977 defer proc.Release()
978 tr.LazyPrintf("acquired process")
979
980 loaded := ss.getLoaded()
981 shards := loaded.shards
982
983 // Setup what we return now, since we may short circuit if there are no
984 // shards to search.
985 stillLoadingCrashes := 0
986 if !loaded.ready {
987 // We may have missed results due to not being fully loaded.
988 stillLoadingCrashes++
989 }
990 agg := zoekt.RepoList{
991 Crashes: stillLoadingCrashes,
992 ReposMap: zoekt.ReposMap{},
993 Repos: []*zoekt.RepoListEntry{},
994 }
995
996 // PERF: Select the subset of shards that we will search over for the given
997 // query. A common List query only asks for a specific repo, so this is an
998 // important optimization.
999 {
1000 beforeLen := len(shards)
1001 beforeQ := q
1002 shards, q = selectRepoSet(shards, q)
1003 tr.LazyPrintf("selectRepoSet shards=%d->%d q=%s->%s", beforeLen, len(shards), beforeQ, q)
1004 }
1005
1006 if len(shards) == 0 {
1007 return &agg, nil
1008 }
1009
1010 shardCount := len(shards)
1011 all := make(chan shardListResult, shardCount)
1012 feeder := make(chan zoekt.Searcher, len(shards))
1013 for _, s := range shards {
1014 feeder <- s
1015 }
1016 close(feeder)
1017
1018 for range runtime.GOMAXPROCS(0) {
1019 go func() {
1020 for s := range feeder {
1021 listOneShard(ctx, s, q, opts, all)
1022 }
1023 }()
1024 }
1025
1026 uniq := map[string]*zoekt.RepoListEntry{}
1027
1028 for range shards {
1029 r := <-all
1030 if r.err != nil {
1031 return nil, r.err
1032 }
1033
1034 agg.Crashes += r.rl.Crashes
1035 agg.Stats.Add(&r.rl.Stats)
1036
1037 for _, r := range r.rl.Repos {
1038 prev, ok := uniq[r.Repository.Name]
1039 if !ok {
1040 cp := *r // We need to copy because we mutate r.Stats when merging duplicates
1041 uniq[r.Repository.Name] = &cp
1042 } else {
1043 prev.Stats.Add(&r.Stats)
1044 }
1045 }
1046
1047 for id, r := range r.rl.ReposMap {
1048 _, ok := agg.ReposMap[id]
1049 if !ok {
1050 agg.ReposMap[id] = r
1051 }
1052 }
1053 }
1054
1055 agg.Repos = make([]*zoekt.RepoListEntry, 0, len(uniq))
1056 for _, r := range uniq {
1057 agg.Repos = append(agg.Repos, r)
1058 }
1059
1060 // Only one of these fields is populated and in all cases the size of that
1061 // field is the number of Repos.
1062 //
1063 // Note: we don't just add individual Stats.Repos since a repository can
1064 // have multiple shards.
1065 agg.Stats.Repos = len(uniq) + len(agg.ReposMap)
1066
1067 if isAll && len(agg.Repos) > 0 {
1068 reportListAllMetrics(agg.Repos)
1069 }
1070
1071 return &agg, nil
1072}
1073
1074func reportListAllMetrics(repos []*zoekt.RepoListEntry) {
1075 var stats zoekt.RepoStats
1076 for _, r := range repos {
1077 stats.Add(&r.Stats)
1078 }
1079
1080 metricListAllRepos.Set(float64(stats.Repos))
1081 metricListAllIndexBytes.Set(float64(stats.IndexBytes))
1082 metricListAllContentBytes.Set(float64(stats.ContentBytes))
1083 metricListAllDocuments.Set(float64(stats.Documents))
1084 metricListAllShards.Set(float64(stats.Shards))
1085 metricListAllNewLinesCount.Set(float64(stats.NewLinesCount))
1086 metricListAllDefaultBranchNewLinesCount.Set(float64(stats.DefaultBranchNewLinesCount))
1087 metricListAllOtherBranchesNewLinesCount.Set(float64(stats.OtherBranchesNewLinesCount))
1088}
1089
1090// getLoaded returns the currently loaded shards. Shared so do not mutate.
1091func (s *shardedSearcher) getLoaded() loaded {
1092 // next commit will store the true value of this, for now we keep the
1093 // backwards compatible behaviour.
1094 ready := s.ready.Load()
1095 // ranked is loaded after ready to avoid a race were ready is true but
1096 // ranked is still not the final set of shards.
1097 ranked, _ := s.ranked.Load().([]*rankedShard)
1098 return loaded{
1099 shards: ranked,
1100 ready: ready,
1101 }
1102}
1103
1104func mkRankedShard(s zoekt.Searcher) *rankedShard {
1105 q := query.Const{Value: true}
1106 // We need to use WithUnsafeContext here, otherwise we cannot return a proper
1107 // rankedShard. On the user request path we use selectRepoSet which relies on
1108 // rankedShard.repos being set.
1109 result, err := s.List(systemtenant.WithUnsafeContext(context.Background()), &q, nil)
1110 if err != nil {
1111 log.Printf("[ERROR] mkRankedShard(%s): failed to cache repository list: %v", s, err)
1112 return &rankedShard{Searcher: s}
1113 }
1114
1115 var (
1116 maxPriority float64
1117 repos = make([]*zoekt.Repository, 0, len(result.Repos))
1118 )
1119 for i := range result.Repos {
1120 repo := &result.Repos[i].Repository
1121 repos = append(repos, repo)
1122 if repo.RawConfig != nil {
1123 priority, _ := strconv.ParseFloat(repo.RawConfig["priority"], 64)
1124 if priority > maxPriority {
1125 maxPriority = priority
1126 }
1127 }
1128 }
1129
1130 return &rankedShard{
1131 Searcher: s,
1132 repos: repos,
1133 priority: maxPriority,
1134 }
1135}
1136
1137// markReady should be called once all shards have been passed into replace on
1138// startup. Once s is marked as ready it stops reporting a Crash in the
1139// response Stats.
1140func (s *shardedSearcher) markReady() {
1141 s.ready.CompareAndSwap(false, true)
1142}
1143
1144func (s *shardedSearcher) replace(shards map[string]zoekt.Searcher) {
1145 if len(shards) == 0 {
1146 return
1147 }
1148
1149 defer func(began time.Time) {
1150 metricShardsBatchReplaceDurationSeconds.Observe(time.Since(began).Seconds())
1151 }(time.Now())
1152
1153 s.mu.Lock()
1154 defer s.mu.Unlock()
1155
1156 for key, shard := range shards {
1157 var r *rankedShard
1158 if shard != nil {
1159 r = mkRankedShard(shard)
1160 }
1161
1162 old := s.shards[key]
1163 if shard == nil {
1164 delete(s.shards, key)
1165 } else {
1166 s.shards[key] = r
1167 }
1168
1169 if old != nil && old.Searcher != nil {
1170 // _ ___ /^^\ /^\ /^^\_
1171 // _ _@)@) \ ,,/ '` ~ `'~~ ', `\.
1172 // _/o\_ _ _ _/~`.`...'~\ ./~~..,'`','',.,' ' ~:
1173 // / `,'.~,~.~ . , . , ~|, ,/ .,' , ,. .. ,,. `, ~\_
1174 // ( ' _' _ '_` _ ' . , `\_/ .' ..' ' ` ` `.. `, \_
1175 // ~V~ V~ V~ V~ ~\ ` ' . ' , ' .,.,''`.,.''`.,.``. ', \_
1176 // _/\ /\ /\ /\_/, . ' , `_/~\_ .' .,. ,, , _/~\_ `. `. '., \_
1177 // < ~ ~ '~`'~'`, ., . `_: ::: \_ ' `_/ ::: \_ `.,' . ', \_
1178 // \ ' `_ '`_ _ ',/ _::_::_ \ _ _/ _::_::_ \ `.,'.,`., \-,-,-,_,_,
1179 // `'~~ `'~~ `'~~ `'~~ \(_)(_)(_)/ `~~' \(_)(_)(_)/ ~'`\_.._,._,'_;_;_;_;_;
1180 //
1181 // We can't just call Close now, because there may be ongoing searches
1182 // which have old in the shards list. Previously we used an exclusive
1183 // lock to guarantee there were no concurrent searches. However, that
1184 // led to blocking on the read path.
1185 //
1186 // We could introduce granular locking per rankedShard to know when
1187 // there are no more references. However, this becomes tricky in
1188 // practice. Instead we rely on the garbage collector noticing old is no
1189 // longer used. We take care in our searchers to runtime.KeepAlive until
1190 // we have stopped referencing the underling mmap data.
1191 runtime.SetFinalizer(old, func(r *rankedShard) {
1192 r.Close()
1193 })
1194 }
1195 }
1196
1197 ranked := make([]*rankedShard, 0, len(s.shards))
1198 for _, r := range s.shards {
1199 ranked = append(ranked, r)
1200 }
1201
1202 sort.Slice(ranked, func(i, j int) bool {
1203 priorityDiff := ranked[i].priority - ranked[j].priority
1204 if priorityDiff != 0 {
1205 return priorityDiff > 0
1206 }
1207 if len(ranked[i].repos) == 0 || len(ranked[j].repos) == 0 {
1208 // Protect against empty names which can happen if we fail to List or
1209 // the shard is full of tombstones. Prefer the shard which has names.
1210 return len(ranked[i].repos) >= len(ranked[j].repos)
1211 }
1212 return ranked[i].repos[0].Name < ranked[j].repos[0].Name
1213 })
1214
1215 s.ranked.Store(ranked)
1216
1217 metricShardsLoaded.Set(float64(len(ranked)))
1218}
1219
1220func loadShard(fn string) (zoekt.Searcher, error) {
1221 f, err := os.Open(fn)
1222 if err != nil {
1223 return nil, err
1224 }
1225
1226 iFile, err := index.NewIndexFile(f)
1227 if err != nil {
1228 return nil, err
1229 }
1230 s, err := index.NewSearcher(iFile)
1231 if err != nil {
1232 iFile.Close()
1233 return nil, fmt.Errorf("NewSearcher(%s): %v", fn, err)
1234 }
1235
1236 return s, nil
1237}
1238
1239// prioritySlice is a trivial implementation of an array that provides three
1240// things: appending a value, removing a value, and getting the array's max.
1241// Operations take O(n) time, which is acceptable because N is restricted to
1242// GOMAXPROCS (i.e., number of cpu cores) by the shardedSearcher interface.
1243type prioritySlice []float64
1244
1245func (p *prioritySlice) append(pri float64) {
1246 *p = append(*p, pri)
1247}
1248
1249func (p *prioritySlice) remove(pri float64) {
1250 for i, opri := range *p {
1251 if opri == pri {
1252 if i != len(*p)-1 {
1253 // swap to make this element the tail
1254 (*p)[i] = (*p)[len(*p)-1]
1255 }
1256 // pop the end off
1257 *p = (*p)[:len(*p)-1]
1258 break
1259 }
1260 }
1261}
1262
1263func (p *prioritySlice) max() float64 {
1264 // remove() and max() could be combined, but this is easier to read and
1265 // the expected performance difference from the extra lock and loop is
1266 // almost certainly irrelevant.
1267 maxPri := math.Inf(-1)
1268 for _, pri := range *p {
1269 if pri > maxPri {
1270 maxPri = pri
1271 }
1272 }
1273 return maxPri
1274}