Merge branch 'bw/clone-recursive-quiet' into maint
[git.git] / name-hash.c
blob0e10f3eab804cdc08d3a2f97021e8b18edcb552a
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 struct dir_entry *e1,
21 const struct dir_entry *e2,
22 const char *name)
24 return e1->namelen != e2->namelen || strncasecmp(e1->name,
25 name ? name : e2->name, e1->namelen);
28 static struct dir_entry *find_dir_entry__hash(struct index_state *istate,
29 const char *name, unsigned int namelen, unsigned int hash)
31 struct dir_entry key;
32 hashmap_entry_init(&key, hash);
33 key.namelen = namelen;
34 return hashmap_get(&istate->dir_hash, &key, name);
37 static struct dir_entry *find_dir_entry(struct index_state *istate,
38 const char *name, unsigned int namelen)
40 return find_dir_entry__hash(istate, name, namelen, memihash(name, namelen));
43 static struct dir_entry *hash_dir_entry(struct index_state *istate,
44 struct cache_entry *ce, int namelen)
47 * Throw each directory component in the hash for quick lookup
48 * during a git status. Directory components are stored without their
49 * closing slash. Despite submodules being a directory, they never
50 * reach this point, because they are stored
51 * in index_state.name_hash (as ordinary cache_entries).
53 struct dir_entry *dir;
55 /* get length of parent directory */
56 while (namelen > 0 && !is_dir_sep(ce->name[namelen - 1]))
57 namelen--;
58 if (namelen <= 0)
59 return NULL;
60 namelen--;
62 /* lookup existing entry for that directory */
63 dir = find_dir_entry(istate, ce->name, namelen);
64 if (!dir) {
65 /* not found, create it and add to hash table */
66 FLEX_ALLOC_MEM(dir, name, ce->name, namelen);
67 hashmap_entry_init(dir, memihash(ce->name, namelen));
68 dir->namelen = namelen;
69 hashmap_add(&istate->dir_hash, dir);
71 /* recursively add missing parent directories */
72 dir->parent = hash_dir_entry(istate, ce, namelen);
74 return dir;
77 static void add_dir_entry(struct index_state *istate, struct cache_entry *ce)
79 /* Add reference to the directory entry (and parents if 0). */
80 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
81 while (dir && !(dir->nr++))
82 dir = dir->parent;
85 static void remove_dir_entry(struct index_state *istate, struct cache_entry *ce)
88 * Release reference to the directory entry. If 0, remove and continue
89 * with parent directory.
91 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
92 while (dir && !(--dir->nr)) {
93 struct dir_entry *parent = dir->parent;
94 hashmap_remove(&istate->dir_hash, dir, NULL);
95 free(dir);
96 dir = parent;
100 static void hash_index_entry(struct index_state *istate, struct cache_entry *ce)
102 if (ce->ce_flags & CE_HASHED)
103 return;
104 ce->ce_flags |= CE_HASHED;
105 hashmap_entry_init(ce, memihash(ce->name, ce_namelen(ce)));
106 hashmap_add(&istate->name_hash, ce);
108 if (ignore_case)
109 add_dir_entry(istate, ce);
112 static int cache_entry_cmp(const void *unused_cmp_data,
113 const struct cache_entry *ce1,
114 const struct cache_entry *ce2,
115 const void *remove)
118 * For remove_name_hash, find the exact entry (pointer equality); for
119 * index_file_exists, find all entries with matching hash code and
120 * decide whether the entry matches in same_name.
122 return remove ? !(ce1 == ce2) : 0;
125 static int lazy_try_threaded = 1;
126 static int lazy_nr_dir_threads;
128 #ifdef NO_PTHREADS
130 static inline int lookup_lazy_params(struct index_state *istate)
132 return 0;
135 static inline void threaded_lazy_init_name_hash(
136 struct index_state *istate)
140 #else
142 #include "thread-utils.h"
145 * Set a minimum number of cache_entries that we will handle per
146 * thread and use that to decide how many threads to run (upto
147 * the number on the system).
149 * For guidance setting the lower per-thread bound, see:
150 * t/helper/test-lazy-init-name-hash --analyze
152 #define LAZY_THREAD_COST (2000)
155 * We use n mutexes to guard n partitions of the "istate->dir_hash"
156 * hashtable. Since "find" and "insert" operations will hash to a
157 * particular bucket and modify/search a single chain, we can say
158 * that "all chains mod n" are guarded by the same mutex -- rather
159 * than having a single mutex to guard the entire table. (This does
160 * require that we disable "rehashing" on the hashtable.)
162 * So, a larger value here decreases the probability of a collision
163 * and the time that each thread must wait for the mutex.
165 #define LAZY_MAX_MUTEX (32)
167 static pthread_mutex_t *lazy_dir_mutex_array;
170 * An array of lazy_entry items is used by the n threads in
171 * the directory parse (first) phase to (lock-free) store the
172 * intermediate results. These values are then referenced by
173 * the 2 threads in the second phase.
175 struct lazy_entry {
176 struct dir_entry *dir;
177 unsigned int hash_dir;
178 unsigned int hash_name;
182 * Decide if we want to use threads (if available) to load
183 * the hash tables. We set "lazy_nr_dir_threads" to zero when
184 * it is not worth it.
186 static int lookup_lazy_params(struct index_state *istate)
188 int nr_cpus;
190 lazy_nr_dir_threads = 0;
192 if (!lazy_try_threaded)
193 return 0;
196 * If we are respecting case, just use the original
197 * code to build the "istate->name_hash". We don't
198 * need the complexity here.
200 if (!ignore_case)
201 return 0;
203 nr_cpus = online_cpus();
204 if (nr_cpus < 2)
205 return 0;
207 if (istate->cache_nr < 2 * LAZY_THREAD_COST)
208 return 0;
210 if (istate->cache_nr < nr_cpus * LAZY_THREAD_COST)
211 nr_cpus = istate->cache_nr / LAZY_THREAD_COST;
212 lazy_nr_dir_threads = nr_cpus;
213 return lazy_nr_dir_threads;
217 * Initialize n mutexes for use when searching and inserting
218 * into "istate->dir_hash". All "dir" threads are trying
219 * to insert partial pathnames into the hash as they iterate
220 * over their portions of the index, so lock contention is
221 * high.
223 * However, the hashmap is going to put items into bucket
224 * chains based on their hash values. Use that to create n
225 * mutexes and lock on mutex[bucket(hash) % n]. This will
226 * decrease the collision rate by (hopefully) by a factor of n.
228 static void init_dir_mutex(void)
230 int j;
232 lazy_dir_mutex_array = xcalloc(LAZY_MAX_MUTEX, sizeof(pthread_mutex_t));
234 for (j = 0; j < LAZY_MAX_MUTEX; j++)
235 init_recursive_mutex(&lazy_dir_mutex_array[j]);
238 static void cleanup_dir_mutex(void)
240 int j;
242 for (j = 0; j < LAZY_MAX_MUTEX; j++)
243 pthread_mutex_destroy(&lazy_dir_mutex_array[j]);
245 free(lazy_dir_mutex_array);
248 static void lock_dir_mutex(int j)
250 pthread_mutex_lock(&lazy_dir_mutex_array[j]);
253 static void unlock_dir_mutex(int j)
255 pthread_mutex_unlock(&lazy_dir_mutex_array[j]);
258 static inline int compute_dir_lock_nr(
259 const struct hashmap *map,
260 unsigned int hash)
262 return hashmap_bucket(map, hash) % LAZY_MAX_MUTEX;
265 static struct dir_entry *hash_dir_entry_with_parent_and_prefix(
266 struct index_state *istate,
267 struct dir_entry *parent,
268 struct strbuf *prefix)
270 struct dir_entry *dir;
271 unsigned int hash;
272 int lock_nr;
275 * Either we have a parent directory and path with slash(es)
276 * or the directory is an immediate child of the root directory.
278 assert((parent != NULL) ^ (strchr(prefix->buf, '/') == NULL));
280 if (parent)
281 hash = memihash_cont(parent->ent.hash,
282 prefix->buf + parent->namelen,
283 prefix->len - parent->namelen);
284 else
285 hash = memihash(prefix->buf, prefix->len);
287 lock_nr = compute_dir_lock_nr(&istate->dir_hash, hash);
288 lock_dir_mutex(lock_nr);
290 dir = find_dir_entry__hash(istate, prefix->buf, prefix->len, hash);
291 if (!dir) {
292 FLEX_ALLOC_MEM(dir, name, prefix->buf, prefix->len);
293 hashmap_entry_init(dir, hash);
294 dir->namelen = prefix->len;
295 dir->parent = parent;
296 hashmap_add(&istate->dir_hash, dir);
298 if (parent) {
299 unlock_dir_mutex(lock_nr);
301 /* All I really need here is an InterlockedIncrement(&(parent->nr)) */
302 lock_nr = compute_dir_lock_nr(&istate->dir_hash, parent->ent.hash);
303 lock_dir_mutex(lock_nr);
304 parent->nr++;
308 unlock_dir_mutex(lock_nr);
310 return dir;
314 * handle_range_1() and handle_range_dir() are derived from
315 * clear_ce_flags_1() and clear_ce_flags_dir() in unpack-trees.c
316 * and handle the iteration over the entire array of index entries.
317 * They use recursion for adjacent entries in the same parent
318 * directory.
320 static int handle_range_1(
321 struct index_state *istate,
322 int k_start,
323 int k_end,
324 struct dir_entry *parent,
325 struct strbuf *prefix,
326 struct lazy_entry *lazy_entries);
328 static int handle_range_dir(
329 struct index_state *istate,
330 int k_start,
331 int k_end,
332 struct dir_entry *parent,
333 struct strbuf *prefix,
334 struct lazy_entry *lazy_entries,
335 struct dir_entry **dir_new_out)
337 int rc, k;
338 int input_prefix_len = prefix->len;
339 struct dir_entry *dir_new;
341 dir_new = hash_dir_entry_with_parent_and_prefix(istate, parent, prefix);
343 strbuf_addch(prefix, '/');
346 * Scan forward in the index array for index entries having the same
347 * path prefix (that are also in this directory).
349 if (k_start + 1 >= k_end)
350 k = k_end;
351 else if (strncmp(istate->cache[k_start + 1]->name, prefix->buf, prefix->len) > 0)
352 k = k_start + 1;
353 else if (strncmp(istate->cache[k_end - 1]->name, prefix->buf, prefix->len) == 0)
354 k = k_end;
355 else {
356 int begin = k_start;
357 int end = k_end;
358 while (begin < end) {
359 int mid = (begin + end) >> 1;
360 int cmp = strncmp(istate->cache[mid]->name, prefix->buf, prefix->len);
361 if (cmp == 0) /* mid has same prefix; look in second part */
362 begin = mid + 1;
363 else if (cmp > 0) /* mid is past group; look in first part */
364 end = mid;
365 else
366 die("cache entry out of order");
368 k = begin;
372 * Recurse and process what we can of this subset [k_start, k).
374 rc = handle_range_1(istate, k_start, k, dir_new, prefix, lazy_entries);
376 strbuf_setlen(prefix, input_prefix_len);
378 *dir_new_out = dir_new;
379 return rc;
382 static int handle_range_1(
383 struct index_state *istate,
384 int k_start,
385 int k_end,
386 struct dir_entry *parent,
387 struct strbuf *prefix,
388 struct lazy_entry *lazy_entries)
390 int input_prefix_len = prefix->len;
391 int k = k_start;
393 while (k < k_end) {
394 struct cache_entry *ce_k = istate->cache[k];
395 const char *name, *slash;
397 if (prefix->len && strncmp(ce_k->name, prefix->buf, prefix->len))
398 break;
400 name = ce_k->name + prefix->len;
401 slash = strchr(name, '/');
403 if (slash) {
404 int len = slash - name;
405 int processed;
406 struct dir_entry *dir_new;
408 strbuf_add(prefix, name, len);
409 processed = handle_range_dir(istate, k, k_end, parent, prefix, lazy_entries, &dir_new);
410 if (processed) {
411 k += processed;
412 strbuf_setlen(prefix, input_prefix_len);
413 continue;
416 strbuf_addch(prefix, '/');
417 processed = handle_range_1(istate, k, k_end, dir_new, prefix, lazy_entries);
418 k += processed;
419 strbuf_setlen(prefix, input_prefix_len);
420 continue;
424 * It is too expensive to take a lock to insert "ce_k"
425 * into "istate->name_hash" and increment the ref-count
426 * on the "parent" dir. So we defer actually updating
427 * permanent data structures until phase 2 (where we
428 * can change the locking requirements) and simply
429 * accumulate our current results into the lazy_entries
430 * data array).
432 * We do not need to lock the lazy_entries array because
433 * we have exclusive access to the cells in the range
434 * [k_start,k_end) that this thread was given.
436 lazy_entries[k].dir = parent;
437 if (parent) {
438 lazy_entries[k].hash_name = memihash_cont(
439 parent->ent.hash,
440 ce_k->name + parent->namelen,
441 ce_namelen(ce_k) - parent->namelen);
442 lazy_entries[k].hash_dir = parent->ent.hash;
443 } else {
444 lazy_entries[k].hash_name = memihash(ce_k->name, ce_namelen(ce_k));
447 k++;
450 return k - k_start;
453 struct lazy_dir_thread_data {
454 pthread_t pthread;
455 struct index_state *istate;
456 struct lazy_entry *lazy_entries;
457 int k_start;
458 int k_end;
461 static void *lazy_dir_thread_proc(void *_data)
463 struct lazy_dir_thread_data *d = _data;
464 struct strbuf prefix = STRBUF_INIT;
465 handle_range_1(d->istate, d->k_start, d->k_end, NULL, &prefix, d->lazy_entries);
466 strbuf_release(&prefix);
467 return NULL;
470 struct lazy_name_thread_data {
471 pthread_t pthread;
472 struct index_state *istate;
473 struct lazy_entry *lazy_entries;
476 static void *lazy_name_thread_proc(void *_data)
478 struct lazy_name_thread_data *d = _data;
479 int k;
481 for (k = 0; k < d->istate->cache_nr; k++) {
482 struct cache_entry *ce_k = d->istate->cache[k];
483 ce_k->ce_flags |= CE_HASHED;
484 hashmap_entry_init(ce_k, d->lazy_entries[k].hash_name);
485 hashmap_add(&d->istate->name_hash, ce_k);
488 return NULL;
491 static inline void lazy_update_dir_ref_counts(
492 struct index_state *istate,
493 struct lazy_entry *lazy_entries)
495 int k;
497 for (k = 0; k < istate->cache_nr; k++) {
498 if (lazy_entries[k].dir)
499 lazy_entries[k].dir->nr++;
503 static void threaded_lazy_init_name_hash(
504 struct index_state *istate)
506 int nr_each;
507 int k_start;
508 int t;
509 struct lazy_entry *lazy_entries;
510 struct lazy_dir_thread_data *td_dir;
511 struct lazy_name_thread_data *td_name;
513 k_start = 0;
514 nr_each = DIV_ROUND_UP(istate->cache_nr, lazy_nr_dir_threads);
516 lazy_entries = xcalloc(istate->cache_nr, sizeof(struct lazy_entry));
517 td_dir = xcalloc(lazy_nr_dir_threads, sizeof(struct lazy_dir_thread_data));
518 td_name = xcalloc(1, sizeof(struct lazy_name_thread_data));
520 init_dir_mutex();
523 * Phase 1:
524 * Build "istate->dir_hash" using n "dir" threads (and a read-only index).
526 for (t = 0; t < lazy_nr_dir_threads; t++) {
527 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
528 td_dir_t->istate = istate;
529 td_dir_t->lazy_entries = lazy_entries;
530 td_dir_t->k_start = k_start;
531 k_start += nr_each;
532 if (k_start > istate->cache_nr)
533 k_start = istate->cache_nr;
534 td_dir_t->k_end = k_start;
535 if (pthread_create(&td_dir_t->pthread, NULL, lazy_dir_thread_proc, td_dir_t))
536 die("unable to create lazy_dir_thread");
538 for (t = 0; t < lazy_nr_dir_threads; t++) {
539 struct lazy_dir_thread_data *td_dir_t = td_dir + t;
540 if (pthread_join(td_dir_t->pthread, NULL))
541 die("unable to join lazy_dir_thread");
545 * Phase 2:
546 * Iterate over all index entries and add them to the "istate->name_hash"
547 * using a single "name" background thread.
548 * (Testing showed it wasn't worth running more than 1 thread for this.)
550 * Meanwhile, finish updating the parent directory ref-counts for each
551 * index entry using the current thread. (This step is very fast and
552 * doesn't need threading.)
554 td_name->istate = istate;
555 td_name->lazy_entries = lazy_entries;
556 if (pthread_create(&td_name->pthread, NULL, lazy_name_thread_proc, td_name))
557 die("unable to create lazy_name_thread");
559 lazy_update_dir_ref_counts(istate, lazy_entries);
561 if (pthread_join(td_name->pthread, NULL))
562 die("unable to join lazy_name_thread");
564 cleanup_dir_mutex();
566 free(td_name);
567 free(td_dir);
568 free(lazy_entries);
571 #endif
573 static void lazy_init_name_hash(struct index_state *istate)
575 if (istate->name_hash_initialized)
576 return;
577 hashmap_init(&istate->name_hash, (hashmap_cmp_fn) cache_entry_cmp,
578 NULL, istate->cache_nr);
579 hashmap_init(&istate->dir_hash, (hashmap_cmp_fn) dir_entry_cmp,
580 NULL, istate->cache_nr);
582 if (lookup_lazy_params(istate)) {
583 hashmap_disallow_rehash(&istate->dir_hash, 1);
584 threaded_lazy_init_name_hash(istate);
585 hashmap_disallow_rehash(&istate->dir_hash, 0);
586 } else {
587 int nr;
588 for (nr = 0; nr < istate->cache_nr; nr++)
589 hash_index_entry(istate, istate->cache[nr]);
592 istate->name_hash_initialized = 1;
596 * A test routine for t/helper/ sources.
598 * Returns the number of threads used or 0 when
599 * the non-threaded code path was used.
601 * Requesting threading WILL NOT override guards
602 * in lookup_lazy_params().
604 int test_lazy_init_name_hash(struct index_state *istate, int try_threaded)
606 lazy_nr_dir_threads = 0;
607 lazy_try_threaded = try_threaded;
609 lazy_init_name_hash(istate);
611 return lazy_nr_dir_threads;
614 void add_name_hash(struct index_state *istate, struct cache_entry *ce)
616 if (istate->name_hash_initialized)
617 hash_index_entry(istate, ce);
620 void remove_name_hash(struct index_state *istate, struct cache_entry *ce)
622 if (!istate->name_hash_initialized || !(ce->ce_flags & CE_HASHED))
623 return;
624 ce->ce_flags &= ~CE_HASHED;
625 hashmap_remove(&istate->name_hash, ce, ce);
627 if (ignore_case)
628 remove_dir_entry(istate, ce);
631 static int slow_same_name(const char *name1, int len1, const char *name2, int len2)
633 if (len1 != len2)
634 return 0;
636 while (len1) {
637 unsigned char c1 = *name1++;
638 unsigned char c2 = *name2++;
639 len1--;
640 if (c1 != c2) {
641 c1 = toupper(c1);
642 c2 = toupper(c2);
643 if (c1 != c2)
644 return 0;
647 return 1;
650 static int same_name(const struct cache_entry *ce, const char *name, int namelen, int icase)
652 int len = ce_namelen(ce);
655 * Always do exact compare, even if we want a case-ignoring comparison;
656 * we do the quick exact one first, because it will be the common case.
658 if (len == namelen && !memcmp(name, ce->name, len))
659 return 1;
661 if (!icase)
662 return 0;
664 return slow_same_name(name, namelen, ce->name, len);
667 int index_dir_exists(struct index_state *istate, const char *name, int namelen)
669 struct dir_entry *dir;
671 lazy_init_name_hash(istate);
672 dir = find_dir_entry(istate, name, namelen);
673 return dir && dir->nr;
676 void adjust_dirname_case(struct index_state *istate, char *name)
678 const char *startPtr = name;
679 const char *ptr = startPtr;
681 lazy_init_name_hash(istate);
682 while (*ptr) {
683 while (*ptr && *ptr != '/')
684 ptr++;
686 if (*ptr == '/') {
687 struct dir_entry *dir;
689 ptr++;
690 dir = find_dir_entry(istate, name, ptr - name + 1);
691 if (dir) {
692 memcpy((void *)startPtr, dir->name + (startPtr - name), ptr - startPtr);
693 startPtr = ptr;
699 struct cache_entry *index_file_exists(struct index_state *istate, const char *name, int namelen, int icase)
701 struct cache_entry *ce;
703 lazy_init_name_hash(istate);
705 ce = hashmap_get_from_hash(&istate->name_hash,
706 memihash(name, namelen), NULL);
707 while (ce) {
708 if (same_name(ce, name, namelen, icase))
709 return ce;
710 ce = hashmap_get_next(&istate->name_hash, ce);
712 return NULL;
715 void free_name_hash(struct index_state *istate)
717 if (!istate->name_hash_initialized)
718 return;
719 istate->name_hash_initialized = 0;
721 hashmap_free(&istate->name_hash, 0);
722 hashmap_free(&istate->dir_hash, 1);