]> git.ipfire.org Git - thirdparty/git.git/blob - name-hash.c
hashmap: introduce hashmap_free_entries
[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 struct hashmap_entry *eptr,
21 const struct hashmap_entry *entry_or_key,
22 const void *keydata)
23 {
24 const struct dir_entry *e1, *e2;
25 const char *name = keydata;
26
27 e1 = container_of(eptr, const struct dir_entry, ent);
28 e2 = container_of(entry_or_key, const struct dir_entry, ent);
29
30 return e1->namelen != e2->namelen || strncasecmp(e1->name,
31 name ? name : e2->name, e1->namelen);
32 }
33
34 static struct dir_entry *find_dir_entry__hash(struct index_state *istate,
35 const char *name, unsigned int namelen, unsigned int hash)
36 {
37 struct dir_entry key;
38 hashmap_entry_init(&key.ent, hash);
39 key.namelen = namelen;
40 return hashmap_get_entry(&istate->dir_hash, &key, name,
41 struct dir_entry, ent);
42 }
43
44 static struct dir_entry *find_dir_entry(struct index_state *istate,
45 const char *name, unsigned int namelen)
46 {
47 return find_dir_entry__hash(istate, name, namelen, memihash(name, namelen));
48 }
49
50 static struct dir_entry *hash_dir_entry(struct index_state *istate,
51 struct cache_entry *ce, int namelen)
52 {
53 /*
54 * Throw each directory component in the hash for quick lookup
55 * during a git status. Directory components are stored without their
56 * closing slash. Despite submodules being a directory, they never
57 * reach this point, because they are stored
58 * in index_state.name_hash (as ordinary cache_entries).
59 */
60 struct dir_entry *dir;
61
62 /* get length of parent directory */
63 while (namelen > 0 && !is_dir_sep(ce->name[namelen - 1]))
64 namelen--;
65 if (namelen <= 0)
66 return NULL;
67 namelen--;
68
69 /* lookup existing entry for that directory */
70 dir = find_dir_entry(istate, ce->name, namelen);
71 if (!dir) {
72 /* not found, create it and add to hash table */
73 FLEX_ALLOC_MEM(dir, name, ce->name, namelen);
74 hashmap_entry_init(&dir->ent, memihash(ce->name, namelen));
75 dir->namelen = namelen;
76 hashmap_add(&istate->dir_hash, &dir->ent);
77
78 /* recursively add missing parent directories */
79 dir->parent = hash_dir_entry(istate, ce, namelen);
80 }
81 return dir;
82 }
83
84 static void add_dir_entry(struct index_state *istate, struct cache_entry *ce)
85 {
86 /* Add reference to the directory entry (and parents if 0). */
87 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
88 while (dir && !(dir->nr++))
89 dir = dir->parent;
90 }
91
92 static void remove_dir_entry(struct index_state *istate, struct cache_entry *ce)
93 {
94 /*
95 * Release reference to the directory entry. If 0, remove and continue
96 * with parent directory.
97 */
98 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
99 while (dir && !(--dir->nr)) {
100 struct dir_entry *parent = dir->parent;
101 hashmap_remove(&istate->dir_hash, &dir->ent, NULL);
102 free(dir);
103 dir = parent;
104 }
105 }
106
107 static void hash_index_entry(struct index_state *istate, struct cache_entry *ce)
108 {
109 if (ce->ce_flags & CE_HASHED)
110 return;
111 ce->ce_flags |= CE_HASHED;
112 hashmap_entry_init(&ce->ent, memihash(ce->name, ce_namelen(ce)));
113 hashmap_add(&istate->name_hash, &ce->ent);
114
115 if (ignore_case)
116 add_dir_entry(istate, ce);
117 }
118
119 static int cache_entry_cmp(const void *unused_cmp_data,
120 const struct hashmap_entry *eptr,
121 const struct hashmap_entry *entry_or_key,
122 const void *remove)
123 {
124 const struct cache_entry *ce1, *ce2;
125
126 ce1 = container_of(eptr, const struct cache_entry, ent);
127 ce2 = container_of(entry_or_key, const struct cache_entry, ent);
128
129 /*
130 * For remove_name_hash, find the exact entry (pointer equality); for
131 * index_file_exists, find all entries with matching hash code and
132 * decide whether the entry matches in same_name.
133 */
134 return remove ? !(ce1 == ce2) : 0;
135 }
136
137 static int lazy_try_threaded = 1;
138 static int lazy_nr_dir_threads;
139
140 /*
141 * Set a minimum number of cache_entries that we will handle per
142 * thread and use that to decide how many threads to run (upto
143 * the number on the system).
144 *
145 * For guidance setting the lower per-thread bound, see:
146 * t/helper/test-lazy-init-name-hash --analyze
147 */
148 #define LAZY_THREAD_COST (2000)
149
150 /*
151 * We use n mutexes to guard n partitions of the "istate->dir_hash"
152 * hashtable. Since "find" and "insert" operations will hash to a
153 * particular bucket and modify/search a single chain, we can say
154 * that "all chains mod n" are guarded by the same mutex -- rather
155 * than having a single mutex to guard the entire table. (This does
156 * require that we disable "rehashing" on the hashtable.)
157 *
158 * So, a larger value here decreases the probability of a collision
159 * and the time that each thread must wait for the mutex.
160 */
161 #define LAZY_MAX_MUTEX (32)
162
163 static pthread_mutex_t *lazy_dir_mutex_array;
164
165 /*
166 * An array of lazy_entry items is used by the n threads in
167 * the directory parse (first) phase to (lock-free) store the
168 * intermediate results. These values are then referenced by
169 * the 2 threads in the second phase.
170 */
171 struct lazy_entry {
172 struct dir_entry *dir;
173 unsigned int hash_dir;
174 unsigned int hash_name;
175 };
176
177 /*
178 * Decide if we want to use threads (if available) to load
179 * the hash tables. We set "lazy_nr_dir_threads" to zero when
180 * it is not worth it.
181 */
182 static int lookup_lazy_params(struct index_state *istate)
183 {
184 int nr_cpus;
185
186 lazy_nr_dir_threads = 0;
187
188 if (!lazy_try_threaded)
189 return 0;
190
191 /*
192 * If we are respecting case, just use the original
193 * code to build the "istate->name_hash". We don't
194 * need the complexity here.
195 */
196 if (!ignore_case)
197 return 0;
198
199 nr_cpus = online_cpus();
200 if (nr_cpus < 2)
201 return 0;
202
203 if (istate->cache_nr < 2 * LAZY_THREAD_COST)
204 return 0;
205
206 if (istate->cache_nr < nr_cpus * LAZY_THREAD_COST)
207 nr_cpus = istate->cache_nr / LAZY_THREAD_COST;
208 lazy_nr_dir_threads = nr_cpus;
209 return lazy_nr_dir_threads;
210 }
211
212 /*
213 * Initialize n mutexes for use when searching and inserting
214 * into "istate->dir_hash". All "dir" threads are trying
215 * to insert partial pathnames into the hash as they iterate
216 * over their portions of the index, so lock contention is
217 * high.
218 *
219 * However, the hashmap is going to put items into bucket
220 * chains based on their hash values. Use that to create n
221 * mutexes and lock on mutex[bucket(hash) % n]. This will
222 * decrease the collision rate by (hopefully) by a factor of n.
223 */
224 static void init_dir_mutex(void)
225 {
226 int j;
227
228 lazy_dir_mutex_array = xcalloc(LAZY_MAX_MUTEX, sizeof(pthread_mutex_t));
229
230 for (j = 0; j < LAZY_MAX_MUTEX; j++)
231 init_recursive_mutex(&lazy_dir_mutex_array[j]);
232 }
233
234 static void cleanup_dir_mutex(void)
235 {
236 int j;
237
238 for (j = 0; j < LAZY_MAX_MUTEX; j++)
239 pthread_mutex_destroy(&lazy_dir_mutex_array[j]);
240
241 free(lazy_dir_mutex_array);
242 }
243
244 static void lock_dir_mutex(int j)
245 {
246 pthread_mutex_lock(&lazy_dir_mutex_array[j]);
247 }
248
249 static void unlock_dir_mutex(int j)
250 {
251 pthread_mutex_unlock(&lazy_dir_mutex_array[j]);
252 }
253
254 static inline int compute_dir_lock_nr(
255 const struct hashmap *map,
256 unsigned int hash)
257 {
258 return hashmap_bucket(map, hash) % LAZY_MAX_MUTEX;
259 }
260
261 static struct dir_entry *hash_dir_entry_with_parent_and_prefix(
262 struct index_state *istate,
263 struct dir_entry *parent,
264 struct strbuf *prefix)
265 {
266 struct dir_entry *dir;
267 unsigned int hash;
268 int lock_nr;
269
270 /*
271 * Either we have a parent directory and path with slash(es)
272 * or the directory is an immediate child of the root directory.
273 */
274 assert((parent != NULL) ^ (strchr(prefix->buf, '/') == NULL));
275
276 if (parent)
277 hash = memihash_cont(parent->ent.hash,
278 prefix->buf + parent->namelen,
279 prefix->len - parent->namelen);
280 else
281 hash = memihash(prefix->buf, prefix->len);
282
283 lock_nr = compute_dir_lock_nr(&istate->dir_hash, hash);
284 lock_dir_mutex(lock_nr);
285
286 dir = find_dir_entry__hash(istate, prefix->buf, prefix->len, hash);
287 if (!dir) {
288 FLEX_ALLOC_MEM(dir, name, prefix->buf, prefix->len);
289 hashmap_entry_init(&dir->ent, hash);
290 dir->namelen = prefix->len;
291 dir->parent = parent;
292 hashmap_add(&istate->dir_hash, &dir->ent);
293
294 if (parent) {
295 unlock_dir_mutex(lock_nr);
296
297 /* All I really need here is an InterlockedIncrement(&(parent->nr)) */
298 lock_nr = compute_dir_lock_nr(&istate->dir_hash, parent->ent.hash);
299 lock_dir_mutex(lock_nr);
300 parent->nr++;
301 }
302 }
303
304 unlock_dir_mutex(lock_nr);
305
306 return dir;
307 }
308
309 /*
310 * handle_range_1() and handle_range_dir() are derived from
311 * clear_ce_flags_1() and clear_ce_flags_dir() in unpack-trees.c
312 * and handle the iteration over the entire array of index entries.
313 * They use recursion for adjacent entries in the same parent
314 * directory.
315 */
316 static int handle_range_1(
317 struct index_state *istate,
318 int k_start,
319 int k_end,
320 struct dir_entry *parent,
321 struct strbuf *prefix,
322 struct lazy_entry *lazy_entries);
323
324 static int handle_range_dir(
325 struct index_state *istate,
326 int k_start,
327 int k_end,
328 struct dir_entry *parent,
329 struct strbuf *prefix,
330 struct lazy_entry *lazy_entries,
331 struct dir_entry **dir_new_out)
332 {
333 int rc, k;
334 int input_prefix_len = prefix->len;
335 struct dir_entry *dir_new;
336
337 dir_new = hash_dir_entry_with_parent_and_prefix(istate, parent, prefix);
338
339 strbuf_addch(prefix, '/');
340
341 /*
342 * Scan forward in the index array for index entries having the same
343 * path prefix (that are also in this directory).
344 */
345 if (k_start + 1 >= k_end)
346 k = k_end;
347 else if (strncmp(istate->cache[k_start + 1]->name, prefix->buf, prefix->len) > 0)
348 k = k_start + 1;
349 else if (strncmp(istate->cache[k_end - 1]->name, prefix->buf, prefix->len) == 0)
350 k = k_end;
351 else {
352 int begin = k_start;
353 int end = k_end;
354 assert(begin >= 0);
355 while (begin < end) {
356 int mid = begin + ((end - begin) >> 1);
357 int cmp = strncmp(istate->cache[mid]->name, prefix->buf, prefix->len);
358 if (cmp == 0) /* mid has same prefix; look in second part */
359 begin = mid + 1;
360 else if (cmp > 0) /* mid is past group; look in first part */
361 end = mid;
362 else
363 die("cache entry out of order");
364 }
365 k = begin;
366 }
367
368 /*
369 * Recurse and process what we can of this subset [k_start, k).
370 */
371 rc = handle_range_1(istate, k_start, k, dir_new, prefix, lazy_entries);
372
373 strbuf_setlen(prefix, input_prefix_len);
374
375 *dir_new_out = dir_new;
376 return rc;
377 }
378
379 static int handle_range_1(
380 struct index_state *istate,
381 int k_start,
382 int k_end,
383 struct dir_entry *parent,
384 struct strbuf *prefix,
385 struct lazy_entry *lazy_entries)
386 {
387 int input_prefix_len = prefix->len;
388 int k = k_start;
389
390 while (k < k_end) {
391 struct cache_entry *ce_k = istate->cache[k];
392 const char *name, *slash;
393
394 if (prefix->len && strncmp(ce_k->name, prefix->buf, prefix->len))
395 break;
396
397 name = ce_k->name + prefix->len;
398 slash = strchr(name, '/');
399
400 if (slash) {
401 int len = slash - name;
402 int processed;
403 struct dir_entry *dir_new;
404
405 strbuf_add(prefix, name, len);
406 processed = handle_range_dir(istate, k, k_end, parent, prefix, lazy_entries, &dir_new);
407 if (processed) {
408 k += processed;
409 strbuf_setlen(prefix, input_prefix_len);
410 continue;
411 }
412
413 strbuf_addch(prefix, '/');
414 processed = handle_range_1(istate, k, k_end, dir_new, prefix, lazy_entries);
415 k += processed;
416 strbuf_setlen(prefix, input_prefix_len);
417 continue;
418 }
419
420 /*
421 * It is too expensive to take a lock to insert "ce_k"
422 * into "istate->name_hash" and increment the ref-count
423 * on the "parent" dir. So we defer actually updating
424 * permanent data structures until phase 2 (where we
425 * can change the locking requirements) and simply
426 * accumulate our current results into the lazy_entries
427 * data array).
428 *
429 * We do not need to lock the lazy_entries array because
430 * we have exclusive access to the cells in the range
431 * [k_start,k_end) that this thread was given.
432 */
433 lazy_entries[k].dir = parent;
434 if (parent) {
435 lazy_entries[k].hash_name = memihash_cont(
436 parent->ent.hash,
437 ce_k->name + parent->namelen,
438 ce_namelen(ce_k) - parent->namelen);
439 lazy_entries[k].hash_dir = parent->ent.hash;
440 } else {
441 lazy_entries[k].hash_name = memihash(ce_k->name, ce_namelen(ce_k));
442 }
443
444 k++;
445 }
446
447 return k - k_start;
448 }
449
450 struct lazy_dir_thread_data {
451 pthread_t pthread;
452 struct index_state *istate;
453 struct lazy_entry *lazy_entries;
454 int k_start;
455 int k_end;
456 };
457
458 static void *lazy_dir_thread_proc(void *_data)
459 {
460 struct lazy_dir_thread_data *d = _data;
461 struct strbuf prefix = STRBUF_INIT;
462 handle_range_1(d->istate, d->k_start, d->k_end, NULL, &prefix, d->lazy_entries);
463 strbuf_release(&prefix);
464 return NULL;
465 }
466
467 struct lazy_name_thread_data {
468 pthread_t pthread;
469 struct index_state *istate;
470 struct lazy_entry *lazy_entries;
471 };
472
473 static void *lazy_name_thread_proc(void *_data)
474 {
475 struct lazy_name_thread_data *d = _data;
476 int k;
477
478 for (k = 0; k < d->istate->cache_nr; k++) {
479 struct cache_entry *ce_k = d->istate->cache[k];
480 ce_k->ce_flags |= CE_HASHED;
481 hashmap_entry_init(&ce_k->ent, d->lazy_entries[k].hash_name);
482 hashmap_add(&d->istate->name_hash, &ce_k->ent);
483 }
484
485 return NULL;
486 }
487
488 static inline void lazy_update_dir_ref_counts(
489 struct index_state *istate,
490 struct lazy_entry *lazy_entries)
491 {
492 int k;
493
494 for (k = 0; k < istate->cache_nr; k++) {
495 if (lazy_entries[k].dir)
496 lazy_entries[k].dir->nr++;
497 }
498 }
499
500 static void threaded_lazy_init_name_hash(
501 struct index_state *istate)
502 {
503 int err;
504 int nr_each;
505 int k_start;
506 int t;
507 struct lazy_entry *lazy_entries;
508 struct lazy_dir_thread_data *td_dir;
509 struct lazy_name_thread_data *td_name;
510
511 if (!HAVE_THREADS)
512 return;
513
514 k_start = 0;
515 nr_each = DIV_ROUND_UP(istate->cache_nr, lazy_nr_dir_threads);
516
517 lazy_entries = xcalloc(istate->cache_nr, sizeof(struct lazy_entry));
518 td_dir = xcalloc(lazy_nr_dir_threads, sizeof(struct lazy_dir_thread_data));
519 td_name = xcalloc(1, sizeof(struct lazy_name_thread_data));
520
521 init_dir_mutex();
522
523 /*
524 * Phase 1:
525 * Build "istate->dir_hash" using n "dir" threads (and a read-only index).
526 */
527 for (t = 0; t < lazy_nr_dir_threads; t++) {
528 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
529 td_dir_t->istate = istate;
530 td_dir_t->lazy_entries = lazy_entries;
531 td_dir_t->k_start = k_start;
532 k_start += nr_each;
533 if (k_start > istate->cache_nr)
534 k_start = istate->cache_nr;
535 td_dir_t->k_end = k_start;
536 err = pthread_create(&td_dir_t->pthread, NULL, lazy_dir_thread_proc, td_dir_t);
537 if (err)
538 die(_("unable to create lazy_dir thread: %s"), strerror(err));
539 }
540 for (t = 0; t < lazy_nr_dir_threads; t++) {
541 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
542 if (pthread_join(td_dir_t->pthread, NULL))
543 die("unable to join lazy_dir_thread");
544 }
545
546 /*
547 * Phase 2:
548 * Iterate over all index entries and add them to the "istate->name_hash"
549 * using a single "name" background thread.
550 * (Testing showed it wasn't worth running more than 1 thread for this.)
551 *
552 * Meanwhile, finish updating the parent directory ref-counts for each
553 * index entry using the current thread. (This step is very fast and
554 * doesn't need threading.)
555 */
556 td_name->istate = istate;
557 td_name->lazy_entries = lazy_entries;
558 err = pthread_create(&td_name->pthread, NULL, lazy_name_thread_proc, td_name);
559 if (err)
560 die(_("unable to create lazy_name thread: %s"), strerror(err));
561
562 lazy_update_dir_ref_counts(istate, lazy_entries);
563
564 err = pthread_join(td_name->pthread, NULL);
565 if (err)
566 die(_("unable to join lazy_name thread: %s"), strerror(err));
567
568 cleanup_dir_mutex();
569
570 free(td_name);
571 free(td_dir);
572 free(lazy_entries);
573 }
574
575 static void lazy_init_name_hash(struct index_state *istate)
576 {
577
578 if (istate->name_hash_initialized)
579 return;
580 trace_performance_enter();
581 hashmap_init(&istate->name_hash, cache_entry_cmp, NULL, istate->cache_nr);
582 hashmap_init(&istate->dir_hash, dir_entry_cmp, NULL, istate->cache_nr);
583
584 if (lookup_lazy_params(istate)) {
585 /*
586 * Disable item counting and automatic rehashing because
587 * we do per-chain (mod n) locking rather than whole hashmap
588 * locking and we need to prevent the table-size from changing
589 * and bucket items from being redistributed.
590 */
591 hashmap_disable_item_counting(&istate->dir_hash);
592 threaded_lazy_init_name_hash(istate);
593 hashmap_enable_item_counting(&istate->dir_hash);
594 } else {
595 int nr;
596 for (nr = 0; nr < istate->cache_nr; nr++)
597 hash_index_entry(istate, istate->cache[nr]);
598 }
599
600 istate->name_hash_initialized = 1;
601 trace_performance_leave("initialize name hash");
602 }
603
604 /*
605 * A test routine for t/helper/ sources.
606 *
607 * Returns the number of threads used or 0 when
608 * the non-threaded code path was used.
609 *
610 * Requesting threading WILL NOT override guards
611 * in lookup_lazy_params().
612 */
613 int test_lazy_init_name_hash(struct index_state *istate, int try_threaded)
614 {
615 lazy_nr_dir_threads = 0;
616 lazy_try_threaded = try_threaded;
617
618 lazy_init_name_hash(istate);
619
620 return lazy_nr_dir_threads;
621 }
622
623 void add_name_hash(struct index_state *istate, struct cache_entry *ce)
624 {
625 if (istate->name_hash_initialized)
626 hash_index_entry(istate, ce);
627 }
628
629 void remove_name_hash(struct index_state *istate, struct cache_entry *ce)
630 {
631 if (!istate->name_hash_initialized || !(ce->ce_flags & CE_HASHED))
632 return;
633 ce->ce_flags &= ~CE_HASHED;
634 hashmap_remove(&istate->name_hash, &ce->ent, ce);
635
636 if (ignore_case)
637 remove_dir_entry(istate, ce);
638 }
639
640 static int slow_same_name(const char *name1, int len1, const char *name2, int len2)
641 {
642 if (len1 != len2)
643 return 0;
644
645 while (len1) {
646 unsigned char c1 = *name1++;
647 unsigned char c2 = *name2++;
648 len1--;
649 if (c1 != c2) {
650 c1 = toupper(c1);
651 c2 = toupper(c2);
652 if (c1 != c2)
653 return 0;
654 }
655 }
656 return 1;
657 }
658
659 static int same_name(const struct cache_entry *ce, const char *name, int namelen, int icase)
660 {
661 int len = ce_namelen(ce);
662
663 /*
664 * Always do exact compare, even if we want a case-ignoring comparison;
665 * we do the quick exact one first, because it will be the common case.
666 */
667 if (len == namelen && !memcmp(name, ce->name, len))
668 return 1;
669
670 if (!icase)
671 return 0;
672
673 return slow_same_name(name, namelen, ce->name, len);
674 }
675
676 int index_dir_exists(struct index_state *istate, const char *name, int namelen)
677 {
678 struct dir_entry *dir;
679
680 lazy_init_name_hash(istate);
681 dir = find_dir_entry(istate, name, namelen);
682 return dir && dir->nr;
683 }
684
685 void adjust_dirname_case(struct index_state *istate, char *name)
686 {
687 const char *startPtr = name;
688 const char *ptr = startPtr;
689
690 lazy_init_name_hash(istate);
691 while (*ptr) {
692 while (*ptr && *ptr != '/')
693 ptr++;
694
695 if (*ptr == '/') {
696 struct dir_entry *dir;
697
698 dir = find_dir_entry(istate, name, ptr - name);
699 if (dir) {
700 memcpy((void *)startPtr, dir->name + (startPtr - name), ptr - startPtr);
701 startPtr = ptr + 1;
702 }
703 ptr++;
704 }
705 }
706 }
707
708 struct cache_entry *index_file_exists(struct index_state *istate, const char *name, int namelen, int icase)
709 {
710 struct cache_entry *ce;
711 unsigned int hash = memihash(name, namelen);
712
713 lazy_init_name_hash(istate);
714
715 ce = hashmap_get_entry_from_hash(&istate->name_hash, hash, NULL,
716 struct cache_entry, ent);
717 hashmap_for_each_entry_from(&istate->name_hash, ce,
718 struct cache_entry, ent) {
719 if (same_name(ce, name, namelen, icase))
720 return ce;
721 }
722 return NULL;
723 }
724
725 void free_name_hash(struct index_state *istate)
726 {
727 if (!istate->name_hash_initialized)
728 return;
729 istate->name_hash_initialized = 0;
730
731 hashmap_free(&istate->name_hash);
732 hashmap_free_entries(&istate->dir_hash, struct dir_entry, ent);
733 }