doc: describe git svn init --ignore-refs
[git/git-svn.git] / name-hash.c
blob39309efb7f9e21eb7a039b6b17cdb0d9f1859820
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 struct dir_entry *e1,
20 const struct dir_entry *e2, const char *name)
22 return e1->namelen != e2->namelen || strncasecmp(e1->name,
23 name ? name : e2->name, e1->namelen);
26 static struct dir_entry *find_dir_entry__hash(struct index_state *istate,
27 const char *name, unsigned int namelen, unsigned int hash)
29 struct dir_entry key;
30 hashmap_entry_init(&key, hash);
31 key.namelen = namelen;
32 return hashmap_get(&istate->dir_hash, &key, name);
35 static struct dir_entry *find_dir_entry(struct index_state *istate,
36 const char *name, unsigned int namelen)
38 return find_dir_entry__hash(istate, name, namelen, memihash(name, namelen));
41 static struct dir_entry *hash_dir_entry(struct index_state *istate,
42 struct cache_entry *ce, int namelen)
45 * Throw each directory component in the hash for quick lookup
46 * during a git status. Directory components are stored without their
47 * closing slash. Despite submodules being a directory, they never
48 * reach this point, because they are stored
49 * in index_state.name_hash (as ordinary cache_entries).
51 struct dir_entry *dir;
53 /* get length of parent directory */
54 while (namelen > 0 && !is_dir_sep(ce->name[namelen - 1]))
55 namelen--;
56 if (namelen <= 0)
57 return NULL;
58 namelen--;
60 /* lookup existing entry for that directory */
61 dir = find_dir_entry(istate, ce->name, namelen);
62 if (!dir) {
63 /* not found, create it and add to hash table */
64 FLEX_ALLOC_MEM(dir, name, ce->name, namelen);
65 hashmap_entry_init(dir, memihash(ce->name, namelen));
66 dir->namelen = namelen;
67 hashmap_add(&istate->dir_hash, dir);
69 /* recursively add missing parent directories */
70 dir->parent = hash_dir_entry(istate, ce, namelen);
72 return dir;
75 static void add_dir_entry(struct index_state *istate, struct cache_entry *ce)
77 /* Add reference to the directory entry (and parents if 0). */
78 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
79 while (dir && !(dir->nr++))
80 dir = dir->parent;
83 static void remove_dir_entry(struct index_state *istate, struct cache_entry *ce)
86 * Release reference to the directory entry. If 0, remove and continue
87 * with parent directory.
89 struct dir_entry *dir = hash_dir_entry(istate, ce, ce_namelen(ce));
90 while (dir && !(--dir->nr)) {
91 struct dir_entry *parent = dir->parent;
92 hashmap_remove(&istate->dir_hash, dir, NULL);
93 free(dir);
94 dir = parent;
98 static void hash_index_entry(struct index_state *istate, struct cache_entry *ce)
100 if (ce->ce_flags & CE_HASHED)
101 return;
102 ce->ce_flags |= CE_HASHED;
103 hashmap_entry_init(ce, memihash(ce->name, ce_namelen(ce)));
104 hashmap_add(&istate->name_hash, ce);
106 if (ignore_case)
107 add_dir_entry(istate, ce);
110 static int cache_entry_cmp(const struct cache_entry *ce1,
111 const struct cache_entry *ce2, const void *remove)
114 * For remove_name_hash, find the exact entry (pointer equality); for
115 * index_file_exists, find all entries with matching hash code and
116 * decide whether the entry matches in same_name.
118 return remove ? !(ce1 == ce2) : 0;
121 static int lazy_try_threaded = 1;
122 static int lazy_nr_dir_threads;
124 #ifdef NO_PTHREADS
126 static inline int lookup_lazy_params(struct index_state *istate)
128 return 0;
131 static inline void threaded_lazy_init_name_hash(
132 struct index_state *istate)
136 #else
138 #include "thread-utils.h"
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).
145 * For guidance setting the lower per-thread bound, see:
146 * t/helper/test-lazy-init-name-hash --analyze
148 #define LAZY_THREAD_COST (2000)
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.)
158 * So, a larger value here decreases the probability of a collision
159 * and the time that each thread must wait for the mutex.
161 #define LAZY_MAX_MUTEX (32)
163 static pthread_mutex_t *lazy_dir_mutex_array;
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.
171 struct lazy_entry {
172 struct dir_entry *dir;
173 unsigned int hash_dir;
174 unsigned int hash_name;
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.
182 static int lookup_lazy_params(struct index_state *istate)
184 int nr_cpus;
186 lazy_nr_dir_threads = 0;
188 if (!lazy_try_threaded)
189 return 0;
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.
196 if (!ignore_case)
197 return 0;
199 nr_cpus = online_cpus();
200 if (nr_cpus < 2)
201 return 0;
203 if (istate->cache_nr < 2 * LAZY_THREAD_COST)
204 return 0;
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;
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.
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.
224 static void init_dir_mutex(void)
226 int j;
228 lazy_dir_mutex_array = xcalloc(LAZY_MAX_MUTEX, sizeof(pthread_mutex_t));
230 for (j = 0; j < LAZY_MAX_MUTEX; j++)
231 init_recursive_mutex(&lazy_dir_mutex_array[j]);
234 static void cleanup_dir_mutex(void)
236 int j;
238 for (j = 0; j < LAZY_MAX_MUTEX; j++)
239 pthread_mutex_destroy(&lazy_dir_mutex_array[j]);
241 free(lazy_dir_mutex_array);
244 static void lock_dir_mutex(int j)
246 pthread_mutex_lock(&lazy_dir_mutex_array[j]);
249 static void unlock_dir_mutex(int j)
251 pthread_mutex_unlock(&lazy_dir_mutex_array[j]);
254 static inline int compute_dir_lock_nr(
255 const struct hashmap *map,
256 unsigned int hash)
258 return hashmap_bucket(map, hash) % LAZY_MAX_MUTEX;
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)
266 struct dir_entry *dir;
267 unsigned int hash;
268 int lock_nr;
271 * Either we have a parent directory and path with slash(es)
272 * or the directory is an immediate child of the root directory.
274 assert((parent != NULL) ^ (strchr(prefix->buf, '/') == NULL));
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);
283 lock_nr = compute_dir_lock_nr(&istate->dir_hash, hash);
284 lock_dir_mutex(lock_nr);
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, hash);
290 dir->namelen = prefix->len;
291 dir->parent = parent;
292 hashmap_add(&istate->dir_hash, dir);
294 if (parent) {
295 unlock_dir_mutex(lock_nr);
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++;
304 unlock_dir_mutex(lock_nr);
306 return dir;
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.
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);
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)
333 int rc, k;
334 int input_prefix_len = prefix->len;
335 struct dir_entry *dir_new;
337 dir_new = hash_dir_entry_with_parent_and_prefix(istate, parent, prefix);
339 strbuf_addch(prefix, '/');
342 * Scan forward in the index array for index entries having the same
343 * path prefix (that are also in this directory).
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 while (begin < end) {
355 int mid = (begin + end) >> 1;
356 int cmp = strncmp(istate->cache[mid]->name, prefix->buf, prefix->len);
357 if (cmp == 0) /* mid has same prefix; look in second part */
358 begin = mid + 1;
359 else if (cmp > 0) /* mid is past group; look in first part */
360 end = mid;
361 else
362 die("cache entry out of order");
364 k = begin;
368 * Recurse and process what we can of this subset [k_start, k).
370 rc = handle_range_1(istate, k_start, k, dir_new, prefix, lazy_entries);
372 strbuf_setlen(prefix, input_prefix_len);
374 *dir_new_out = dir_new;
375 return rc;
378 static int handle_range_1(
379 struct index_state *istate,
380 int k_start,
381 int k_end,
382 struct dir_entry *parent,
383 struct strbuf *prefix,
384 struct lazy_entry *lazy_entries)
386 int input_prefix_len = prefix->len;
387 int k = k_start;
389 while (k < k_end) {
390 struct cache_entry *ce_k = istate->cache[k];
391 const char *name, *slash;
393 if (prefix->len && strncmp(ce_k->name, prefix->buf, prefix->len))
394 break;
396 name = ce_k->name + prefix->len;
397 slash = strchr(name, '/');
399 if (slash) {
400 int len = slash - name;
401 int processed;
402 struct dir_entry *dir_new;
404 strbuf_add(prefix, name, len);
405 processed = handle_range_dir(istate, k, k_end, parent, prefix, lazy_entries, &dir_new);
406 if (processed) {
407 k += processed;
408 strbuf_setlen(prefix, input_prefix_len);
409 continue;
412 strbuf_addch(prefix, '/');
413 processed = handle_range_1(istate, k, k_end, dir_new, prefix, lazy_entries);
414 k += processed;
415 strbuf_setlen(prefix, input_prefix_len);
416 continue;
420 * It is too expensive to take a lock to insert "ce_k"
421 * into "istate->name_hash" and increment the ref-count
422 * on the "parent" dir. So we defer actually updating
423 * permanent data structures until phase 2 (where we
424 * can change the locking requirements) and simply
425 * accumulate our current results into the lazy_entries
426 * data array).
428 * We do not need to lock the lazy_entries array because
429 * we have exclusive access to the cells in the range
430 * [k_start,k_end) that this thread was given.
432 lazy_entries[k].dir = parent;
433 if (parent) {
434 lazy_entries[k].hash_name = memihash_cont(
435 parent->ent.hash,
436 ce_k->name + parent->namelen,
437 ce_namelen(ce_k) - parent->namelen);
438 lazy_entries[k].hash_dir = parent->ent.hash;
439 } else {
440 lazy_entries[k].hash_name = memihash(ce_k->name, ce_namelen(ce_k));
443 k++;
446 return k - k_start;
449 struct lazy_dir_thread_data {
450 pthread_t pthread;
451 struct index_state *istate;
452 struct lazy_entry *lazy_entries;
453 int k_start;
454 int k_end;
457 static void *lazy_dir_thread_proc(void *_data)
459 struct lazy_dir_thread_data *d = _data;
460 struct strbuf prefix = STRBUF_INIT;
461 handle_range_1(d->istate, d->k_start, d->k_end, NULL, &prefix, d->lazy_entries);
462 strbuf_release(&prefix);
463 return NULL;
466 struct lazy_name_thread_data {
467 pthread_t pthread;
468 struct index_state *istate;
469 struct lazy_entry *lazy_entries;
472 static void *lazy_name_thread_proc(void *_data)
474 struct lazy_name_thread_data *d = _data;
475 int k;
477 for (k = 0; k < d->istate->cache_nr; k++) {
478 struct cache_entry *ce_k = d->istate->cache[k];
479 ce_k->ce_flags |= CE_HASHED;
480 hashmap_entry_init(ce_k, d->lazy_entries[k].hash_name);
481 hashmap_add(&d->istate->name_hash, ce_k);
484 return NULL;
487 static inline void lazy_update_dir_ref_counts(
488 struct index_state *istate,
489 struct lazy_entry *lazy_entries)
491 int k;
493 for (k = 0; k < istate->cache_nr; k++) {
494 if (lazy_entries[k].dir)
495 lazy_entries[k].dir->nr++;
499 static void threaded_lazy_init_name_hash(
500 struct index_state *istate)
502 int nr_each;
503 int k_start;
504 int t;
505 struct lazy_entry *lazy_entries;
506 struct lazy_dir_thread_data *td_dir;
507 struct lazy_name_thread_data *td_name;
509 k_start = 0;
510 nr_each = DIV_ROUND_UP(istate->cache_nr, lazy_nr_dir_threads);
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));
516 init_dir_mutex();
519 * Phase 1:
520 * Build "istate->dir_hash" using n "dir" threads (and a read-only index).
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 if (pthread_create(&td_dir_t->pthread, NULL, lazy_dir_thread_proc, td_dir_t))
532 die("unable to create lazy_dir_thread");
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");
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.)
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.)
550 td_name->istate = istate;
551 td_name->lazy_entries = lazy_entries;
552 if (pthread_create(&td_name->pthread, NULL, lazy_name_thread_proc, td_name))
553 die("unable to create lazy_name_thread");
555 lazy_update_dir_ref_counts(istate, lazy_entries);
557 if (pthread_join(td_name->pthread, NULL))
558 die("unable to join lazy_name_thread");
560 cleanup_dir_mutex();
562 free(td_name);
563 free(td_dir);
564 free(lazy_entries);
567 #endif
569 static void lazy_init_name_hash(struct index_state *istate)
571 if (istate->name_hash_initialized)
572 return;
573 hashmap_init(&istate->name_hash, (hashmap_cmp_fn) cache_entry_cmp,
574 istate->cache_nr);
575 hashmap_init(&istate->dir_hash, (hashmap_cmp_fn) dir_entry_cmp,
576 istate->cache_nr);
578 if (lookup_lazy_params(istate)) {
579 hashmap_disallow_rehash(&istate->dir_hash, 1);
580 threaded_lazy_init_name_hash(istate);
581 hashmap_disallow_rehash(&istate->dir_hash, 0);
582 } else {
583 int nr;
584 for (nr = 0; nr < istate->cache_nr; nr++)
585 hash_index_entry(istate, istate->cache[nr]);
588 istate->name_hash_initialized = 1;
592 * A test routine for t/helper/ sources.
594 * Returns the number of threads used or 0 when
595 * the non-threaded code path was used.
597 * Requesting threading WILL NOT override guards
598 * in lookup_lazy_params().
600 int test_lazy_init_name_hash(struct index_state *istate, int try_threaded)
602 lazy_nr_dir_threads = 0;
603 lazy_try_threaded = try_threaded;
605 lazy_init_name_hash(istate);
607 return lazy_nr_dir_threads;
610 void add_name_hash(struct index_state *istate, struct cache_entry *ce)
612 if (istate->name_hash_initialized)
613 hash_index_entry(istate, ce);
616 void remove_name_hash(struct index_state *istate, struct cache_entry *ce)
618 if (!istate->name_hash_initialized || !(ce->ce_flags & CE_HASHED))
619 return;
620 ce->ce_flags &= ~CE_HASHED;
621 hashmap_remove(&istate->name_hash, ce, ce);
623 if (ignore_case)
624 remove_dir_entry(istate, ce);
627 static int slow_same_name(const char *name1, int len1, const char *name2, int len2)
629 if (len1 != len2)
630 return 0;
632 while (len1) {
633 unsigned char c1 = *name1++;
634 unsigned char c2 = *name2++;
635 len1--;
636 if (c1 != c2) {
637 c1 = toupper(c1);
638 c2 = toupper(c2);
639 if (c1 != c2)
640 return 0;
643 return 1;
646 static int same_name(const struct cache_entry *ce, const char *name, int namelen, int icase)
648 int len = ce_namelen(ce);
651 * Always do exact compare, even if we want a case-ignoring comparison;
652 * we do the quick exact one first, because it will be the common case.
654 if (len == namelen && !memcmp(name, ce->name, len))
655 return 1;
657 if (!icase)
658 return 0;
660 return slow_same_name(name, namelen, ce->name, len);
663 int index_dir_exists(struct index_state *istate, const char *name, int namelen)
665 struct dir_entry *dir;
667 lazy_init_name_hash(istate);
668 dir = find_dir_entry(istate, name, namelen);
669 return dir && dir->nr;
672 void adjust_dirname_case(struct index_state *istate, char *name)
674 const char *startPtr = name;
675 const char *ptr = startPtr;
677 lazy_init_name_hash(istate);
678 while (*ptr) {
679 while (*ptr && *ptr != '/')
680 ptr++;
682 if (*ptr == '/') {
683 struct dir_entry *dir;
685 ptr++;
686 dir = find_dir_entry(istate, name, ptr - name + 1);
687 if (dir) {
688 memcpy((void *)startPtr, dir->name + (startPtr - name), ptr - startPtr);
689 startPtr = ptr;
695 struct cache_entry *index_file_exists(struct index_state *istate, const char *name, int namelen, int icase)
697 struct cache_entry *ce;
699 lazy_init_name_hash(istate);
701 ce = hashmap_get_from_hash(&istate->name_hash,
702 memihash(name, namelen), NULL);
703 while (ce) {
704 if (same_name(ce, name, namelen, icase))
705 return ce;
706 ce = hashmap_get_next(&istate->name_hash, ce);
708 return NULL;
711 void free_name_hash(struct index_state *istate)
713 if (!istate->name_hash_initialized)
714 return;
715 istate->name_hash_initialized = 0;
717 hashmap_free(&istate->name_hash, 0);
718 hashmap_free(&istate->dir_hash, 1);