Merge branch 'jt/http-redact-cookies' into maint
[git.git] / name-hash.c
blob45c98db0a057e93b5f3f84eba2008da86130010d
1 /*
2 * name-hash.c
4 * Hashing names in the index state
6 * Copyright (C) 2008 Linus Torvalds
7 */
8 #define NO_THE_INDEX_COMPATIBILITY_MACROS
9 #include "cache.h"
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];
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)
24 const struct dir_entry *e1 = entry;
25 const struct dir_entry *e2 = entry_or_key;
26 const char *name = keydata;
28 return e1->namelen != e2->namelen || strncasecmp(e1->name,
29 name ? name : e2->name, e1->namelen);
32 static struct dir_entry *find_dir_entry__hash(struct index_state *istate,
33 const char *name, unsigned int namelen, unsigned int hash)
35 struct dir_entry key;
36 hashmap_entry_init(&key, hash);
37 key.namelen = namelen;
38 return hashmap_get(&istate->dir_hash, &key, name);
41 static struct dir_entry *find_dir_entry(struct index_state *istate,
42 const char *name, unsigned int namelen)
44 return find_dir_entry__hash(istate, name, namelen, memihash(name, namelen));
47 static struct dir_entry *hash_dir_entry(struct index_state *istate,
48 struct cache_entry *ce, int namelen)
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).
57 struct dir_entry *dir;
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--;
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);
75 /* recursively add missing parent directories */
76 dir->parent = hash_dir_entry(istate, ce, namelen);
78 return dir;
81 static void add_dir_entry(struct index_state *istate, struct cache_entry *ce)
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;
89 static void remove_dir_entry(struct index_state *istate, struct cache_entry *ce)
92 * Release reference to the directory entry. If 0, remove and continue
93 * with parent directory.
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;
104 static void hash_index_entry(struct index_state *istate, struct cache_entry *ce)
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);
112 if (ignore_case)
113 add_dir_entry(istate, ce);
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)
121 const struct cache_entry *ce1 = entry;
122 const struct cache_entry *ce2 = entry_or_key;
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.
128 return remove ? !(ce1 == ce2) : 0;
131 static int lazy_try_threaded = 1;
132 static int lazy_nr_dir_threads;
134 #ifdef NO_PTHREADS
136 static inline int lookup_lazy_params(struct index_state *istate)
138 return 0;
141 static inline void threaded_lazy_init_name_hash(
142 struct index_state *istate)
146 #else
148 #include "thread-utils.h"
151 * Set a minimum number of cache_entries that we will handle per
152 * thread and use that to decide how many threads to run (upto
153 * the number on the system).
155 * For guidance setting the lower per-thread bound, see:
156 * t/helper/test-lazy-init-name-hash --analyze
158 #define LAZY_THREAD_COST (2000)
161 * We use n mutexes to guard n partitions of the "istate->dir_hash"
162 * hashtable. Since "find" and "insert" operations will hash to a
163 * particular bucket and modify/search a single chain, we can say
164 * that "all chains mod n" are guarded by the same mutex -- rather
165 * than having a single mutex to guard the entire table. (This does
166 * require that we disable "rehashing" on the hashtable.)
168 * So, a larger value here decreases the probability of a collision
169 * and the time that each thread must wait for the mutex.
171 #define LAZY_MAX_MUTEX (32)
173 static pthread_mutex_t *lazy_dir_mutex_array;
176 * An array of lazy_entry items is used by the n threads in
177 * the directory parse (first) phase to (lock-free) store the
178 * intermediate results. These values are then referenced by
179 * the 2 threads in the second phase.
181 struct lazy_entry {
182 struct dir_entry *dir;
183 unsigned int hash_dir;
184 unsigned int hash_name;
188 * Decide if we want to use threads (if available) to load
189 * the hash tables. We set "lazy_nr_dir_threads" to zero when
190 * it is not worth it.
192 static int lookup_lazy_params(struct index_state *istate)
194 int nr_cpus;
196 lazy_nr_dir_threads = 0;
198 if (!lazy_try_threaded)
199 return 0;
202 * If we are respecting case, just use the original
203 * code to build the "istate->name_hash". We don't
204 * need the complexity here.
206 if (!ignore_case)
207 return 0;
209 nr_cpus = online_cpus();
210 if (nr_cpus < 2)
211 return 0;
213 if (istate->cache_nr < 2 * LAZY_THREAD_COST)
214 return 0;
216 if (istate->cache_nr < nr_cpus * LAZY_THREAD_COST)
217 nr_cpus = istate->cache_nr / LAZY_THREAD_COST;
218 lazy_nr_dir_threads = nr_cpus;
219 return lazy_nr_dir_threads;
223 * Initialize n mutexes for use when searching and inserting
224 * into "istate->dir_hash". All "dir" threads are trying
225 * to insert partial pathnames into the hash as they iterate
226 * over their portions of the index, so lock contention is
227 * high.
229 * However, the hashmap is going to put items into bucket
230 * chains based on their hash values. Use that to create n
231 * mutexes and lock on mutex[bucket(hash) % n]. This will
232 * decrease the collision rate by (hopefully) by a factor of n.
234 static void init_dir_mutex(void)
236 int j;
238 lazy_dir_mutex_array = xcalloc(LAZY_MAX_MUTEX, sizeof(pthread_mutex_t));
240 for (j = 0; j < LAZY_MAX_MUTEX; j++)
241 init_recursive_mutex(&lazy_dir_mutex_array[j]);
244 static void cleanup_dir_mutex(void)
246 int j;
248 for (j = 0; j < LAZY_MAX_MUTEX; j++)
249 pthread_mutex_destroy(&lazy_dir_mutex_array[j]);
251 free(lazy_dir_mutex_array);
254 static void lock_dir_mutex(int j)
256 pthread_mutex_lock(&lazy_dir_mutex_array[j]);
259 static void unlock_dir_mutex(int j)
261 pthread_mutex_unlock(&lazy_dir_mutex_array[j]);
264 static inline int compute_dir_lock_nr(
265 const struct hashmap *map,
266 unsigned int hash)
268 return hashmap_bucket(map, hash) % LAZY_MAX_MUTEX;
271 static struct dir_entry *hash_dir_entry_with_parent_and_prefix(
272 struct index_state *istate,
273 struct dir_entry *parent,
274 struct strbuf *prefix)
276 struct dir_entry *dir;
277 unsigned int hash;
278 int lock_nr;
281 * Either we have a parent directory and path with slash(es)
282 * or the directory is an immediate child of the root directory.
284 assert((parent != NULL) ^ (strchr(prefix->buf, '/') == NULL));
286 if (parent)
287 hash = memihash_cont(parent->ent.hash,
288 prefix->buf + parent->namelen,
289 prefix->len - parent->namelen);
290 else
291 hash = memihash(prefix->buf, prefix->len);
293 lock_nr = compute_dir_lock_nr(&istate->dir_hash, hash);
294 lock_dir_mutex(lock_nr);
296 dir = find_dir_entry__hash(istate, prefix->buf, prefix->len, hash);
297 if (!dir) {
298 FLEX_ALLOC_MEM(dir, name, prefix->buf, prefix->len);
299 hashmap_entry_init(dir, hash);
300 dir->namelen = prefix->len;
301 dir->parent = parent;
302 hashmap_add(&istate->dir_hash, dir);
304 if (parent) {
305 unlock_dir_mutex(lock_nr);
307 /* All I really need here is an InterlockedIncrement(&(parent->nr)) */
308 lock_nr = compute_dir_lock_nr(&istate->dir_hash, parent->ent.hash);
309 lock_dir_mutex(lock_nr);
310 parent->nr++;
314 unlock_dir_mutex(lock_nr);
316 return dir;
320 * handle_range_1() and handle_range_dir() are derived from
321 * clear_ce_flags_1() and clear_ce_flags_dir() in unpack-trees.c
322 * and handle the iteration over the entire array of index entries.
323 * They use recursion for adjacent entries in the same parent
324 * directory.
326 static int handle_range_1(
327 struct index_state *istate,
328 int k_start,
329 int k_end,
330 struct dir_entry *parent,
331 struct strbuf *prefix,
332 struct lazy_entry *lazy_entries);
334 static int handle_range_dir(
335 struct index_state *istate,
336 int k_start,
337 int k_end,
338 struct dir_entry *parent,
339 struct strbuf *prefix,
340 struct lazy_entry *lazy_entries,
341 struct dir_entry **dir_new_out)
343 int rc, k;
344 int input_prefix_len = prefix->len;
345 struct dir_entry *dir_new;
347 dir_new = hash_dir_entry_with_parent_and_prefix(istate, parent, prefix);
349 strbuf_addch(prefix, '/');
352 * Scan forward in the index array for index entries having the same
353 * path prefix (that are also in this directory).
355 if (k_start + 1 >= k_end)
356 k = k_end;
357 else if (strncmp(istate->cache[k_start + 1]->name, prefix->buf, prefix->len) > 0)
358 k = k_start + 1;
359 else if (strncmp(istate->cache[k_end - 1]->name, prefix->buf, prefix->len) == 0)
360 k = k_end;
361 else {
362 int begin = k_start;
363 int end = k_end;
364 while (begin < end) {
365 int mid = (begin + end) >> 1;
366 int cmp = strncmp(istate->cache[mid]->name, prefix->buf, prefix->len);
367 if (cmp == 0) /* mid has same prefix; look in second part */
368 begin = mid + 1;
369 else if (cmp > 0) /* mid is past group; look in first part */
370 end = mid;
371 else
372 die("cache entry out of order");
374 k = begin;
378 * Recurse and process what we can of this subset [k_start, k).
380 rc = handle_range_1(istate, k_start, k, dir_new, prefix, lazy_entries);
382 strbuf_setlen(prefix, input_prefix_len);
384 *dir_new_out = dir_new;
385 return rc;
388 static int handle_range_1(
389 struct index_state *istate,
390 int k_start,
391 int k_end,
392 struct dir_entry *parent,
393 struct strbuf *prefix,
394 struct lazy_entry *lazy_entries)
396 int input_prefix_len = prefix->len;
397 int k = k_start;
399 while (k < k_end) {
400 struct cache_entry *ce_k = istate->cache[k];
401 const char *name, *slash;
403 if (prefix->len && strncmp(ce_k->name, prefix->buf, prefix->len))
404 break;
406 name = ce_k->name + prefix->len;
407 slash = strchr(name, '/');
409 if (slash) {
410 int len = slash - name;
411 int processed;
412 struct dir_entry *dir_new;
414 strbuf_add(prefix, name, len);
415 processed = handle_range_dir(istate, k, k_end, parent, prefix, lazy_entries, &dir_new);
416 if (processed) {
417 k += processed;
418 strbuf_setlen(prefix, input_prefix_len);
419 continue;
422 strbuf_addch(prefix, '/');
423 processed = handle_range_1(istate, k, k_end, dir_new, prefix, lazy_entries);
424 k += processed;
425 strbuf_setlen(prefix, input_prefix_len);
426 continue;
430 * It is too expensive to take a lock to insert "ce_k"
431 * into "istate->name_hash" and increment the ref-count
432 * on the "parent" dir. So we defer actually updating
433 * permanent data structures until phase 2 (where we
434 * can change the locking requirements) and simply
435 * accumulate our current results into the lazy_entries
436 * data array).
438 * We do not need to lock the lazy_entries array because
439 * we have exclusive access to the cells in the range
440 * [k_start,k_end) that this thread was given.
442 lazy_entries[k].dir = parent;
443 if (parent) {
444 lazy_entries[k].hash_name = memihash_cont(
445 parent->ent.hash,
446 ce_k->name + parent->namelen,
447 ce_namelen(ce_k) - parent->namelen);
448 lazy_entries[k].hash_dir = parent->ent.hash;
449 } else {
450 lazy_entries[k].hash_name = memihash(ce_k->name, ce_namelen(ce_k));
453 k++;
456 return k - k_start;
459 struct lazy_dir_thread_data {
460 pthread_t pthread;
461 struct index_state *istate;
462 struct lazy_entry *lazy_entries;
463 int k_start;
464 int k_end;
467 static void *lazy_dir_thread_proc(void *_data)
469 struct lazy_dir_thread_data *d = _data;
470 struct strbuf prefix = STRBUF_INIT;
471 handle_range_1(d->istate, d->k_start, d->k_end, NULL, &prefix, d->lazy_entries);
472 strbuf_release(&prefix);
473 return NULL;
476 struct lazy_name_thread_data {
477 pthread_t pthread;
478 struct index_state *istate;
479 struct lazy_entry *lazy_entries;
482 static void *lazy_name_thread_proc(void *_data)
484 struct lazy_name_thread_data *d = _data;
485 int k;
487 for (k = 0; k < d->istate->cache_nr; k++) {
488 struct cache_entry *ce_k = d->istate->cache[k];
489 ce_k->ce_flags |= CE_HASHED;
490 hashmap_entry_init(ce_k, d->lazy_entries[k].hash_name);
491 hashmap_add(&d->istate->name_hash, ce_k);
494 return NULL;
497 static inline void lazy_update_dir_ref_counts(
498 struct index_state *istate,
499 struct lazy_entry *lazy_entries)
501 int k;
503 for (k = 0; k < istate->cache_nr; k++) {
504 if (lazy_entries[k].dir)
505 lazy_entries[k].dir->nr++;
509 static void threaded_lazy_init_name_hash(
510 struct index_state *istate)
512 int nr_each;
513 int k_start;
514 int t;
515 struct lazy_entry *lazy_entries;
516 struct lazy_dir_thread_data *td_dir;
517 struct lazy_name_thread_data *td_name;
519 k_start = 0;
520 nr_each = DIV_ROUND_UP(istate->cache_nr, lazy_nr_dir_threads);
522 lazy_entries = xcalloc(istate->cache_nr, sizeof(struct lazy_entry));
523 td_dir = xcalloc(lazy_nr_dir_threads, sizeof(struct lazy_dir_thread_data));
524 td_name = xcalloc(1, sizeof(struct lazy_name_thread_data));
526 init_dir_mutex();
529 * Phase 1:
530 * Build "istate->dir_hash" using n "dir" threads (and a read-only index).
532 for (t = 0; t < lazy_nr_dir_threads; t++) {
533 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
534 td_dir_t->istate = istate;
535 td_dir_t->lazy_entries = lazy_entries;
536 td_dir_t->k_start = k_start;
537 k_start += nr_each;
538 if (k_start > istate->cache_nr)
539 k_start = istate->cache_nr;
540 td_dir_t->k_end = k_start;
541 if (pthread_create(&td_dir_t->pthread, NULL, lazy_dir_thread_proc, td_dir_t))
542 die("unable to create lazy_dir_thread");
544 for (t = 0; t < lazy_nr_dir_threads; t++) {
545 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
546 if (pthread_join(td_dir_t->pthread, NULL))
547 die("unable to join lazy_dir_thread");
551 * Phase 2:
552 * Iterate over all index entries and add them to the "istate->name_hash"
553 * using a single "name" background thread.
554 * (Testing showed it wasn't worth running more than 1 thread for this.)
556 * Meanwhile, finish updating the parent directory ref-counts for each
557 * index entry using the current thread. (This step is very fast and
558 * doesn't need threading.)
560 td_name->istate = istate;
561 td_name->lazy_entries = lazy_entries;
562 if (pthread_create(&td_name->pthread, NULL, lazy_name_thread_proc, td_name))
563 die("unable to create lazy_name_thread");
565 lazy_update_dir_ref_counts(istate, lazy_entries);
567 if (pthread_join(td_name->pthread, NULL))
568 die("unable to join lazy_name_thread");
570 cleanup_dir_mutex();
572 free(td_name);
573 free(td_dir);
574 free(lazy_entries);
577 #endif
579 static void lazy_init_name_hash(struct index_state *istate)
581 if (istate->name_hash_initialized)
582 return;
583 hashmap_init(&istate->name_hash, cache_entry_cmp, NULL, istate->cache_nr);
584 hashmap_init(&istate->dir_hash, dir_entry_cmp, NULL, istate->cache_nr);
586 if (lookup_lazy_params(istate)) {
588 * Disable item counting and automatic rehashing because
589 * we do per-chain (mod n) locking rather than whole hashmap
590 * locking and we need to prevent the table-size from changing
591 * and bucket items from being redistributed.
593 hashmap_disable_item_counting(&istate->dir_hash);
594 threaded_lazy_init_name_hash(istate);
595 hashmap_enable_item_counting(&istate->dir_hash);
596 } else {
597 int nr;
598 for (nr = 0; nr < istate->cache_nr; nr++)
599 hash_index_entry(istate, istate->cache[nr]);
602 istate->name_hash_initialized = 1;
606 * A test routine for t/helper/ sources.
608 * Returns the number of threads used or 0 when
609 * the non-threaded code path was used.
611 * Requesting threading WILL NOT override guards
612 * in lookup_lazy_params().
614 int test_lazy_init_name_hash(struct index_state *istate, int try_threaded)
616 lazy_nr_dir_threads = 0;
617 lazy_try_threaded = try_threaded;
619 lazy_init_name_hash(istate);
621 return lazy_nr_dir_threads;
624 void add_name_hash(struct index_state *istate, struct cache_entry *ce)
626 if (istate->name_hash_initialized)
627 hash_index_entry(istate, ce);
630 void remove_name_hash(struct index_state *istate, struct cache_entry *ce)
632 if (!istate->name_hash_initialized || !(ce->ce_flags & CE_HASHED))
633 return;
634 ce->ce_flags &= ~CE_HASHED;
635 hashmap_remove(&istate->name_hash, ce, ce);
637 if (ignore_case)
638 remove_dir_entry(istate, ce);
641 static int slow_same_name(const char *name1, int len1, const char *name2, int len2)
643 if (len1 != len2)
644 return 0;
646 while (len1) {
647 unsigned char c1 = *name1++;
648 unsigned char c2 = *name2++;
649 len1--;
650 if (c1 != c2) {
651 c1 = toupper(c1);
652 c2 = toupper(c2);
653 if (c1 != c2)
654 return 0;
657 return 1;
660 static int same_name(const struct cache_entry *ce, const char *name, int namelen, int icase)
662 int len = ce_namelen(ce);
665 * Always do exact compare, even if we want a case-ignoring comparison;
666 * we do the quick exact one first, because it will be the common case.
668 if (len == namelen && !memcmp(name, ce->name, len))
669 return 1;
671 if (!icase)
672 return 0;
674 return slow_same_name(name, namelen, ce->name, len);
677 int index_dir_exists(struct index_state *istate, const char *name, int namelen)
679 struct dir_entry *dir;
681 lazy_init_name_hash(istate);
682 dir = find_dir_entry(istate, name, namelen);
683 return dir && dir->nr;
686 void adjust_dirname_case(struct index_state *istate, char *name)
688 const char *startPtr = name;
689 const char *ptr = startPtr;
691 lazy_init_name_hash(istate);
692 while (*ptr) {
693 while (*ptr && *ptr != '/')
694 ptr++;
696 if (*ptr == '/') {
697 struct dir_entry *dir;
699 ptr++;
700 dir = find_dir_entry(istate, name, ptr - name + 1);
701 if (dir) {
702 memcpy((void *)startPtr, dir->name + (startPtr - name), ptr - startPtr);
703 startPtr = ptr;
709 struct cache_entry *index_file_exists(struct index_state *istate, const char *name, int namelen, int icase)
711 struct cache_entry *ce;
713 lazy_init_name_hash(istate);
715 ce = hashmap_get_from_hash(&istate->name_hash,
716 memihash(name, namelen), NULL);
717 while (ce) {
718 if (same_name(ce, name, namelen, icase))
719 return ce;
720 ce = hashmap_get_next(&istate->name_hash, ce);
722 return NULL;
725 void free_name_hash(struct index_state *istate)
727 if (!istate->name_hash_initialized)
728 return;
729 istate->name_hash_initialized = 0;
731 hashmap_free(&istate->name_hash, 0);
732 hashmap_free(&istate->dir_hash, 1);