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