]> git.ipfire.org Git - thirdparty/git.git/blob - name-hash.c
path.c: don't call the match function without value in trie_find()
[thirdparty/git.git] / name-hash.c
1 /*
2 * name-hash.c
3 *
4 * Hashing names in the index state
5 *
6 * Copyright (C) 2008 Linus Torvalds
7 */
8 #include "cache.h"
9 #include "thread-utils.h"
10
11 struct dir_entry {
12 struct hashmap_entry ent;
13 struct dir_entry *parent;
14 int nr;
15 unsigned int namelen;
16 char name[FLEX_ARRAY];
17 };
18
19 static int dir_entry_cmp(const void *unused_cmp_data,
20 const void *entry,
21 const void *entry_or_key,
22 const void *keydata)
23 {
24 const struct dir_entry *e1 = entry;
25 const struct dir_entry *e2 = entry_or_key;
26 const char *name = keydata;
27
28 return e1->namelen != e2->namelen || strncasecmp(e1->name,
29 name ? name : e2->name, e1->namelen);
30 }
31
32 static struct dir_entry *find_dir_entry__hash(struct index_state *istate,
33 const char *name, unsigned int namelen, unsigned int hash)
34 {
35 struct dir_entry key;
36 hashmap_entry_init(&key, hash);
37 key.namelen = namelen;
38 return hashmap_get(&istate->dir_hash, &key, name);
39 }
40
41 static struct dir_entry *find_dir_entry(struct index_state *istate,
42 const char *name, unsigned int namelen)
43 {
44 return find_dir_entry__hash(istate, name, namelen, memihash(name, namelen));
45 }
46
47 static struct dir_entry *hash_dir_entry(struct index_state *istate,
48 struct cache_entry *ce, int namelen)
49 {
50 /*
51 * Throw each directory component in the hash for quick lookup
52 * during a git status. Directory components are stored without their
53 * closing slash. Despite submodules being a directory, they never
54 * reach this point, because they are stored
55 * in index_state.name_hash (as ordinary cache_entries).
56 */
57 struct dir_entry *dir;
58
59 /* get length of parent directory */
60 while (namelen > 0 && !is_dir_sep(ce->name[namelen - 1]))
61 namelen--;
62 if (namelen <= 0)
63 return NULL;
64 namelen--;
65
66 /* lookup existing entry for that directory */
67 dir = find_dir_entry(istate, ce->name, namelen);
68 if (!dir) {
69 /* not found, create it and add to hash table */
70 FLEX_ALLOC_MEM(dir, name, ce->name, namelen);
71 hashmap_entry_init(dir, memihash(ce->name, namelen));
72 dir->namelen = namelen;
73 hashmap_add(&istate->dir_hash, dir);
74
75 /* recursively add missing parent directories */
76 dir->parent = hash_dir_entry(istate, ce, namelen);
77 }
78 return dir;
79 }
80
81 static void add_dir_entry(struct index_state *istate, struct cache_entry *ce)
82 {
83 /* Add reference to the directory entry (and parents if 0). */
84 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
85 while (dir && !(dir->nr++))
86 dir = dir->parent;
87 }
88
89 static void remove_dir_entry(struct index_state *istate, struct cache_entry *ce)
90 {
91 /*
92 * Release reference to the directory entry. If 0, remove and continue
93 * with parent directory.
94 */
95 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
96 while (dir && !(--dir->nr)) {
97 struct dir_entry *parent = dir->parent;
98 hashmap_remove(&istate->dir_hash, dir, NULL);
99 free(dir);
100 dir = parent;
101 }
102 }
103
104 static void hash_index_entry(struct index_state *istate, struct cache_entry *ce)
105 {
106 if (ce->ce_flags & CE_HASHED)
107 return;
108 ce->ce_flags |= CE_HASHED;
109 hashmap_entry_init(ce, memihash(ce->name, ce_namelen(ce)));
110 hashmap_add(&istate->name_hash, ce);
111
112 if (ignore_case)
113 add_dir_entry(istate, ce);
114 }
115
116 static int cache_entry_cmp(const void *unused_cmp_data,
117 const void *entry,
118 const void *entry_or_key,
119 const void *remove)
120 {
121 const struct cache_entry *ce1 = entry;
122 const struct cache_entry *ce2 = entry_or_key;
123 /*
124 * For remove_name_hash, find the exact entry (pointer equality); for
125 * index_file_exists, find all entries with matching hash code and
126 * decide whether the entry matches in same_name.
127 */
128 return remove ? !(ce1 == ce2) : 0;
129 }
130
131 static int lazy_try_threaded = 1;
132 static int lazy_nr_dir_threads;
133
134 /*
135 * Set a minimum number of cache_entries that we will handle per
136 * thread and use that to decide how many threads to run (upto
137 * the number on the system).
138 *
139 * For guidance setting the lower per-thread bound, see:
140 * t/helper/test-lazy-init-name-hash --analyze
141 */
142 #define LAZY_THREAD_COST (2000)
143
144 /*
145 * We use n mutexes to guard n partitions of the "istate->dir_hash"
146 * hashtable. Since "find" and "insert" operations will hash to a
147 * particular bucket and modify/search a single chain, we can say
148 * that "all chains mod n" are guarded by the same mutex -- rather
149 * than having a single mutex to guard the entire table. (This does
150 * require that we disable "rehashing" on the hashtable.)
151 *
152 * So, a larger value here decreases the probability of a collision
153 * and the time that each thread must wait for the mutex.
154 */
155 #define LAZY_MAX_MUTEX (32)
156
157 static pthread_mutex_t *lazy_dir_mutex_array;
158
159 /*
160 * An array of lazy_entry items is used by the n threads in
161 * the directory parse (first) phase to (lock-free) store the
162 * intermediate results. These values are then referenced by
163 * the 2 threads in the second phase.
164 */
165 struct lazy_entry {
166 struct dir_entry *dir;
167 unsigned int hash_dir;
168 unsigned int hash_name;
169 };
170
171 /*
172 * Decide if we want to use threads (if available) to load
173 * the hash tables. We set "lazy_nr_dir_threads" to zero when
174 * it is not worth it.
175 */
176 static int lookup_lazy_params(struct index_state *istate)
177 {
178 int nr_cpus;
179
180 lazy_nr_dir_threads = 0;
181
182 if (!lazy_try_threaded)
183 return 0;
184
185 /*
186 * If we are respecting case, just use the original
187 * code to build the "istate->name_hash". We don't
188 * need the complexity here.
189 */
190 if (!ignore_case)
191 return 0;
192
193 nr_cpus = online_cpus();
194 if (nr_cpus < 2)
195 return 0;
196
197 if (istate->cache_nr < 2 * LAZY_THREAD_COST)
198 return 0;
199
200 if (istate->cache_nr < nr_cpus * LAZY_THREAD_COST)
201 nr_cpus = istate->cache_nr / LAZY_THREAD_COST;
202 lazy_nr_dir_threads = nr_cpus;
203 return lazy_nr_dir_threads;
204 }
205
206 /*
207 * Initialize n mutexes for use when searching and inserting
208 * into "istate->dir_hash". All "dir" threads are trying
209 * to insert partial pathnames into the hash as they iterate
210 * over their portions of the index, so lock contention is
211 * high.
212 *
213 * However, the hashmap is going to put items into bucket
214 * chains based on their hash values. Use that to create n
215 * mutexes and lock on mutex[bucket(hash) % n]. This will
216 * decrease the collision rate by (hopefully) by a factor of n.
217 */
218 static void init_dir_mutex(void)
219 {
220 int j;
221
222 lazy_dir_mutex_array = xcalloc(LAZY_MAX_MUTEX, sizeof(pthread_mutex_t));
223
224 for (j = 0; j < LAZY_MAX_MUTEX; j++)
225 init_recursive_mutex(&lazy_dir_mutex_array[j]);
226 }
227
228 static void cleanup_dir_mutex(void)
229 {
230 int j;
231
232 for (j = 0; j < LAZY_MAX_MUTEX; j++)
233 pthread_mutex_destroy(&lazy_dir_mutex_array[j]);
234
235 free(lazy_dir_mutex_array);
236 }
237
238 static void lock_dir_mutex(int j)
239 {
240 pthread_mutex_lock(&lazy_dir_mutex_array[j]);
241 }
242
243 static void unlock_dir_mutex(int j)
244 {
245 pthread_mutex_unlock(&lazy_dir_mutex_array[j]);
246 }
247
248 static inline int compute_dir_lock_nr(
249 const struct hashmap *map,
250 unsigned int hash)
251 {
252 return hashmap_bucket(map, hash) % LAZY_MAX_MUTEX;
253 }
254
255 static struct dir_entry *hash_dir_entry_with_parent_and_prefix(
256 struct index_state *istate,
257 struct dir_entry *parent,
258 struct strbuf *prefix)
259 {
260 struct dir_entry *dir;
261 unsigned int hash;
262 int lock_nr;
263
264 /*
265 * Either we have a parent directory and path with slash(es)
266 * or the directory is an immediate child of the root directory.
267 */
268 assert((parent != NULL) ^ (strchr(prefix->buf, '/') == NULL));
269
270 if (parent)
271 hash = memihash_cont(parent->ent.hash,
272 prefix->buf + parent->namelen,
273 prefix->len - parent->namelen);
274 else
275 hash = memihash(prefix->buf, prefix->len);
276
277 lock_nr = compute_dir_lock_nr(&istate->dir_hash, hash);
278 lock_dir_mutex(lock_nr);
279
280 dir = find_dir_entry__hash(istate, prefix->buf, prefix->len, hash);
281 if (!dir) {
282 FLEX_ALLOC_MEM(dir, name, prefix->buf, prefix->len);
283 hashmap_entry_init(dir, hash);
284 dir->namelen = prefix->len;
285 dir->parent = parent;
286 hashmap_add(&istate->dir_hash, dir);
287
288 if (parent) {
289 unlock_dir_mutex(lock_nr);
290
291 /* All I really need here is an InterlockedIncrement(&(parent->nr)) */
292 lock_nr = compute_dir_lock_nr(&istate->dir_hash, parent->ent.hash);
293 lock_dir_mutex(lock_nr);
294 parent->nr++;
295 }
296 }
297
298 unlock_dir_mutex(lock_nr);
299
300 return dir;
301 }
302
303 /*
304 * handle_range_1() and handle_range_dir() are derived from
305 * clear_ce_flags_1() and clear_ce_flags_dir() in unpack-trees.c
306 * and handle the iteration over the entire array of index entries.
307 * They use recursion for adjacent entries in the same parent
308 * directory.
309 */
310 static int handle_range_1(
311 struct index_state *istate,
312 int k_start,
313 int k_end,
314 struct dir_entry *parent,
315 struct strbuf *prefix,
316 struct lazy_entry *lazy_entries);
317
318 static int handle_range_dir(
319 struct index_state *istate,
320 int k_start,
321 int k_end,
322 struct dir_entry *parent,
323 struct strbuf *prefix,
324 struct lazy_entry *lazy_entries,
325 struct dir_entry **dir_new_out)
326 {
327 int rc, k;
328 int input_prefix_len = prefix->len;
329 struct dir_entry *dir_new;
330
331 dir_new = hash_dir_entry_with_parent_and_prefix(istate, parent, prefix);
332
333 strbuf_addch(prefix, '/');
334
335 /*
336 * Scan forward in the index array for index entries having the same
337 * path prefix (that are also in this directory).
338 */
339 if (k_start + 1 >= k_end)
340 k = k_end;
341 else if (strncmp(istate->cache[k_start + 1]->name, prefix->buf, prefix->len) > 0)
342 k = k_start + 1;
343 else if (strncmp(istate->cache[k_end - 1]->name, prefix->buf, prefix->len) == 0)
344 k = k_end;
345 else {
346 int begin = k_start;
347 int end = k_end;
348 assert(begin >= 0);
349 while (begin < end) {
350 int mid = begin + ((end - begin) >> 1);
351 int cmp = strncmp(istate->cache[mid]->name, prefix->buf, prefix->len);
352 if (cmp == 0) /* mid has same prefix; look in second part */
353 begin = mid + 1;
354 else if (cmp > 0) /* mid is past group; look in first part */
355 end = mid;
356 else
357 die("cache entry out of order");
358 }
359 k = begin;
360 }
361
362 /*
363 * Recurse and process what we can of this subset [k_start, k).
364 */
365 rc = handle_range_1(istate, k_start, k, dir_new, prefix, lazy_entries);
366
367 strbuf_setlen(prefix, input_prefix_len);
368
369 *dir_new_out = dir_new;
370 return rc;
371 }
372
373 static int handle_range_1(
374 struct index_state *istate,
375 int k_start,
376 int k_end,
377 struct dir_entry *parent,
378 struct strbuf *prefix,
379 struct lazy_entry *lazy_entries)
380 {
381 int input_prefix_len = prefix->len;
382 int k = k_start;
383
384 while (k < k_end) {
385 struct cache_entry *ce_k = istate->cache[k];
386 const char *name, *slash;
387
388 if (prefix->len && strncmp(ce_k->name, prefix->buf, prefix->len))
389 break;
390
391 name = ce_k->name + prefix->len;
392 slash = strchr(name, '/');
393
394 if (slash) {
395 int len = slash - name;
396 int processed;
397 struct dir_entry *dir_new;
398
399 strbuf_add(prefix, name, len);
400 processed = handle_range_dir(istate, k, k_end, parent, prefix, lazy_entries, &dir_new);
401 if (processed) {
402 k += processed;
403 strbuf_setlen(prefix, input_prefix_len);
404 continue;
405 }
406
407 strbuf_addch(prefix, '/');
408 processed = handle_range_1(istate, k, k_end, dir_new, prefix, lazy_entries);
409 k += processed;
410 strbuf_setlen(prefix, input_prefix_len);
411 continue;
412 }
413
414 /*
415 * It is too expensive to take a lock to insert "ce_k"
416 * into "istate->name_hash" and increment the ref-count
417 * on the "parent" dir. So we defer actually updating
418 * permanent data structures until phase 2 (where we
419 * can change the locking requirements) and simply
420 * accumulate our current results into the lazy_entries
421 * data array).
422 *
423 * We do not need to lock the lazy_entries array because
424 * we have exclusive access to the cells in the range
425 * [k_start,k_end) that this thread was given.
426 */
427 lazy_entries[k].dir = parent;
428 if (parent) {
429 lazy_entries[k].hash_name = memihash_cont(
430 parent->ent.hash,
431 ce_k->name + parent->namelen,
432 ce_namelen(ce_k) - parent->namelen);
433 lazy_entries[k].hash_dir = parent->ent.hash;
434 } else {
435 lazy_entries[k].hash_name = memihash(ce_k->name, ce_namelen(ce_k));
436 }
437
438 k++;
439 }
440
441 return k - k_start;
442 }
443
444 struct lazy_dir_thread_data {
445 pthread_t pthread;
446 struct index_state *istate;
447 struct lazy_entry *lazy_entries;
448 int k_start;
449 int k_end;
450 };
451
452 static void *lazy_dir_thread_proc(void *_data)
453 {
454 struct lazy_dir_thread_data *d = _data;
455 struct strbuf prefix = STRBUF_INIT;
456 handle_range_1(d->istate, d->k_start, d->k_end, NULL, &prefix, d->lazy_entries);
457 strbuf_release(&prefix);
458 return NULL;
459 }
460
461 struct lazy_name_thread_data {
462 pthread_t pthread;
463 struct index_state *istate;
464 struct lazy_entry *lazy_entries;
465 };
466
467 static void *lazy_name_thread_proc(void *_data)
468 {
469 struct lazy_name_thread_data *d = _data;
470 int k;
471
472 for (k = 0; k < d->istate->cache_nr; k++) {
473 struct cache_entry *ce_k = d->istate->cache[k];
474 ce_k->ce_flags |= CE_HASHED;
475 hashmap_entry_init(ce_k, d->lazy_entries[k].hash_name);
476 hashmap_add(&d->istate->name_hash, ce_k);
477 }
478
479 return NULL;
480 }
481
482 static inline void lazy_update_dir_ref_counts(
483 struct index_state *istate,
484 struct lazy_entry *lazy_entries)
485 {
486 int k;
487
488 for (k = 0; k < istate->cache_nr; k++) {
489 if (lazy_entries[k].dir)
490 lazy_entries[k].dir->nr++;
491 }
492 }
493
494 static void threaded_lazy_init_name_hash(
495 struct index_state *istate)
496 {
497 int err;
498 int nr_each;
499 int k_start;
500 int t;
501 struct lazy_entry *lazy_entries;
502 struct lazy_dir_thread_data *td_dir;
503 struct lazy_name_thread_data *td_name;
504
505 if (!HAVE_THREADS)
506 return;
507
508 k_start = 0;
509 nr_each = DIV_ROUND_UP(istate->cache_nr, lazy_nr_dir_threads);
510
511 lazy_entries = xcalloc(istate->cache_nr, sizeof(struct lazy_entry));
512 td_dir = xcalloc(lazy_nr_dir_threads, sizeof(struct lazy_dir_thread_data));
513 td_name = xcalloc(1, sizeof(struct lazy_name_thread_data));
514
515 init_dir_mutex();
516
517 /*
518 * Phase 1:
519 * Build "istate->dir_hash" using n "dir" threads (and a read-only index).
520 */
521 for (t = 0; t < lazy_nr_dir_threads; t++) {
522 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
523 td_dir_t->istate = istate;
524 td_dir_t->lazy_entries = lazy_entries;
525 td_dir_t->k_start = k_start;
526 k_start += nr_each;
527 if (k_start > istate->cache_nr)
528 k_start = istate->cache_nr;
529 td_dir_t->k_end = k_start;
530 err = pthread_create(&td_dir_t->pthread, NULL, lazy_dir_thread_proc, td_dir_t);
531 if (err)
532 die(_("unable to create lazy_dir thread: %s"), strerror(err));
533 }
534 for (t = 0; t < lazy_nr_dir_threads; t++) {
535 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
536 if (pthread_join(td_dir_t->pthread, NULL))
537 die("unable to join lazy_dir_thread");
538 }
539
540 /*
541 * Phase 2:
542 * Iterate over all index entries and add them to the "istate->name_hash"
543 * using a single "name" background thread.
544 * (Testing showed it wasn't worth running more than 1 thread for this.)
545 *
546 * Meanwhile, finish updating the parent directory ref-counts for each
547 * index entry using the current thread. (This step is very fast and
548 * doesn't need threading.)
549 */
550 td_name->istate = istate;
551 td_name->lazy_entries = lazy_entries;
552 err = pthread_create(&td_name->pthread, NULL, lazy_name_thread_proc, td_name);
553 if (err)
554 die(_("unable to create lazy_name thread: %s"), strerror(err));
555
556 lazy_update_dir_ref_counts(istate, lazy_entries);
557
558 err = pthread_join(td_name->pthread, NULL);
559 if (err)
560 die(_("unable to join lazy_name thread: %s"), strerror(err));
561
562 cleanup_dir_mutex();
563
564 free(td_name);
565 free(td_dir);
566 free(lazy_entries);
567 }
568
569 static void lazy_init_name_hash(struct index_state *istate)
570 {
571
572 if (istate->name_hash_initialized)
573 return;
574 trace_performance_enter();
575 hashmap_init(&istate->name_hash, cache_entry_cmp, NULL, istate->cache_nr);
576 hashmap_init(&istate->dir_hash, dir_entry_cmp, NULL, istate->cache_nr);
577
578 if (lookup_lazy_params(istate)) {
579 /*
580 * Disable item counting and automatic rehashing because
581 * we do per-chain (mod n) locking rather than whole hashmap
582 * locking and we need to prevent the table-size from changing
583 * and bucket items from being redistributed.
584 */
585 hashmap_disable_item_counting(&istate->dir_hash);
586 threaded_lazy_init_name_hash(istate);
587 hashmap_enable_item_counting(&istate->dir_hash);
588 } else {
589 int nr;
590 for (nr = 0; nr < istate->cache_nr; nr++)
591 hash_index_entry(istate, istate->cache[nr]);
592 }
593
594 istate->name_hash_initialized = 1;
595 trace_performance_leave("initialize name hash");
596 }
597
598 /*
599 * A test routine for t/helper/ sources.
600 *
601 * Returns the number of threads used or 0 when
602 * the non-threaded code path was used.
603 *
604 * Requesting threading WILL NOT override guards
605 * in lookup_lazy_params().
606 */
607 int test_lazy_init_name_hash(struct index_state *istate, int try_threaded)
608 {
609 lazy_nr_dir_threads = 0;
610 lazy_try_threaded = try_threaded;
611
612 lazy_init_name_hash(istate);
613
614 return lazy_nr_dir_threads;
615 }
616
617 void add_name_hash(struct index_state *istate, struct cache_entry *ce)
618 {
619 if (istate->name_hash_initialized)
620 hash_index_entry(istate, ce);
621 }
622
623 void remove_name_hash(struct index_state *istate, struct cache_entry *ce)
624 {
625 if (!istate->name_hash_initialized || !(ce->ce_flags & CE_HASHED))
626 return;
627 ce->ce_flags &= ~CE_HASHED;
628 hashmap_remove(&istate->name_hash, ce, ce);
629
630 if (ignore_case)
631 remove_dir_entry(istate, ce);
632 }
633
634 static int slow_same_name(const char *name1, int len1, const char *name2, int len2)
635 {
636 if (len1 != len2)
637 return 0;
638
639 while (len1) {
640 unsigned char c1 = *name1++;
641 unsigned char c2 = *name2++;
642 len1--;
643 if (c1 != c2) {
644 c1 = toupper(c1);
645 c2 = toupper(c2);
646 if (c1 != c2)
647 return 0;
648 }
649 }
650 return 1;
651 }
652
653 static int same_name(const struct cache_entry *ce, const char *name, int namelen, int icase)
654 {
655 int len = ce_namelen(ce);
656
657 /*
658 * Always do exact compare, even if we want a case-ignoring comparison;
659 * we do the quick exact one first, because it will be the common case.
660 */
661 if (len == namelen && !memcmp(name, ce->name, len))
662 return 1;
663
664 if (!icase)
665 return 0;
666
667 return slow_same_name(name, namelen, ce->name, len);
668 }
669
670 int index_dir_exists(struct index_state *istate, const char *name, int namelen)
671 {
672 struct dir_entry *dir;
673
674 lazy_init_name_hash(istate);
675 dir = find_dir_entry(istate, name, namelen);
676 return dir && dir->nr;
677 }
678
679 void adjust_dirname_case(struct index_state *istate, char *name)
680 {
681 const char *startPtr = name;
682 const char *ptr = startPtr;
683
684 lazy_init_name_hash(istate);
685 while (*ptr) {
686 while (*ptr && *ptr != '/')
687 ptr++;
688
689 if (*ptr == '/') {
690 struct dir_entry *dir;
691
692 dir = find_dir_entry(istate, name, ptr - name);
693 if (dir) {
694 memcpy((void *)startPtr, dir->name + (startPtr - name), ptr - startPtr);
695 startPtr = ptr + 1;
696 }
697 ptr++;
698 }
699 }
700 }
701
702 struct cache_entry *index_file_exists(struct index_state *istate, const char *name, int namelen, int icase)
703 {
704 struct cache_entry *ce;
705
706 lazy_init_name_hash(istate);
707
708 ce = hashmap_get_from_hash(&istate->name_hash,
709 memihash(name, namelen), NULL);
710 while (ce) {
711 if (same_name(ce, name, namelen, icase))
712 return ce;
713 ce = hashmap_get_next(&istate->name_hash, ce);
714 }
715 return NULL;
716 }
717
718 void free_name_hash(struct index_state *istate)
719 {
720 if (!istate->name_hash_initialized)
721 return;
722 istate->name_hash_initialized = 0;
723
724 hashmap_free(&istate->name_hash, 0);
725 hashmap_free(&istate->dir_hash, 1);
726 }