commit-graph: lazy-load trees for commits
[git.git] / commit-graph.c
blob70fa1b25fd3d07a935cc0e7ba4cf70bc24ca3d21
1 #include "cache.h"
2 #include "config.h"
3 #include "git-compat-util.h"
4 #include "lockfile.h"
5 #include "pack.h"
6 #include "packfile.h"
7 #include "commit.h"
8 #include "object.h"
9 #include "revision.h"
10 #include "sha1-lookup.h"
11 #include "commit-graph.h"
13 #define GRAPH_SIGNATURE 0x43475048 /* "CGPH" */
14 #define GRAPH_CHUNKID_OIDFANOUT 0x4f494446 /* "OIDF" */
15 #define GRAPH_CHUNKID_OIDLOOKUP 0x4f49444c /* "OIDL" */
16 #define GRAPH_CHUNKID_DATA 0x43444154 /* "CDAT" */
17 #define GRAPH_CHUNKID_LARGEEDGES 0x45444745 /* "EDGE" */
19 #define GRAPH_DATA_WIDTH 36
21 #define GRAPH_VERSION_1 0x1
22 #define GRAPH_VERSION GRAPH_VERSION_1
24 #define GRAPH_OID_VERSION_SHA1 1
25 #define GRAPH_OID_LEN_SHA1 GIT_SHA1_RAWSZ
26 #define GRAPH_OID_VERSION GRAPH_OID_VERSION_SHA1
27 #define GRAPH_OID_LEN GRAPH_OID_LEN_SHA1
29 #define GRAPH_OCTOPUS_EDGES_NEEDED 0x80000000
30 #define GRAPH_PARENT_MISSING 0x7fffffff
31 #define GRAPH_EDGE_LAST_MASK 0x7fffffff
32 #define GRAPH_PARENT_NONE 0x70000000
34 #define GRAPH_LAST_EDGE 0x80000000
36 #define GRAPH_FANOUT_SIZE (4 * 256)
37 #define GRAPH_CHUNKLOOKUP_WIDTH 12
38 #define GRAPH_MIN_SIZE (5 * GRAPH_CHUNKLOOKUP_WIDTH + GRAPH_FANOUT_SIZE + \
39 GRAPH_OID_LEN + 8)
41 char *get_commit_graph_filename(const char *obj_dir)
43 return xstrfmt("%s/info/commit-graph", obj_dir);
46 static struct commit_graph *alloc_commit_graph(void)
48 struct commit_graph *g = xcalloc(1, sizeof(*g));
49 g->graph_fd = -1;
51 return g;
54 struct commit_graph *load_commit_graph_one(const char *graph_file)
56 void *graph_map;
57 const unsigned char *data, *chunk_lookup;
58 size_t graph_size;
59 struct stat st;
60 uint32_t i;
61 struct commit_graph *graph;
62 int fd = git_open(graph_file);
63 uint64_t last_chunk_offset;
64 uint32_t last_chunk_id;
65 uint32_t graph_signature;
66 unsigned char graph_version, hash_version;
68 if (fd < 0)
69 return NULL;
70 if (fstat(fd, &st)) {
71 close(fd);
72 return NULL;
74 graph_size = xsize_t(st.st_size);
76 if (graph_size < GRAPH_MIN_SIZE) {
77 close(fd);
78 die("graph file %s is too small", graph_file);
80 graph_map = xmmap(NULL, graph_size, PROT_READ, MAP_PRIVATE, fd, 0);
81 data = (const unsigned char *)graph_map;
83 graph_signature = get_be32(data);
84 if (graph_signature != GRAPH_SIGNATURE) {
85 error("graph signature %X does not match signature %X",
86 graph_signature, GRAPH_SIGNATURE);
87 goto cleanup_fail;
90 graph_version = *(unsigned char*)(data + 4);
91 if (graph_version != GRAPH_VERSION) {
92 error("graph version %X does not match version %X",
93 graph_version, GRAPH_VERSION);
94 goto cleanup_fail;
97 hash_version = *(unsigned char*)(data + 5);
98 if (hash_version != GRAPH_OID_VERSION) {
99 error("hash version %X does not match version %X",
100 hash_version, GRAPH_OID_VERSION);
101 goto cleanup_fail;
104 graph = alloc_commit_graph();
106 graph->hash_len = GRAPH_OID_LEN;
107 graph->num_chunks = *(unsigned char*)(data + 6);
108 graph->graph_fd = fd;
109 graph->data = graph_map;
110 graph->data_len = graph_size;
112 last_chunk_id = 0;
113 last_chunk_offset = 8;
114 chunk_lookup = data + 8;
115 for (i = 0; i < graph->num_chunks; i++) {
116 uint32_t chunk_id = get_be32(chunk_lookup + 0);
117 uint64_t chunk_offset = get_be64(chunk_lookup + 4);
118 int chunk_repeated = 0;
120 chunk_lookup += GRAPH_CHUNKLOOKUP_WIDTH;
122 if (chunk_offset > graph_size - GIT_MAX_RAWSZ) {
123 error("improper chunk offset %08x%08x", (uint32_t)(chunk_offset >> 32),
124 (uint32_t)chunk_offset);
125 goto cleanup_fail;
128 switch (chunk_id) {
129 case GRAPH_CHUNKID_OIDFANOUT:
130 if (graph->chunk_oid_fanout)
131 chunk_repeated = 1;
132 else
133 graph->chunk_oid_fanout = (uint32_t*)(data + chunk_offset);
134 break;
136 case GRAPH_CHUNKID_OIDLOOKUP:
137 if (graph->chunk_oid_lookup)
138 chunk_repeated = 1;
139 else
140 graph->chunk_oid_lookup = data + chunk_offset;
141 break;
143 case GRAPH_CHUNKID_DATA:
144 if (graph->chunk_commit_data)
145 chunk_repeated = 1;
146 else
147 graph->chunk_commit_data = data + chunk_offset;
148 break;
150 case GRAPH_CHUNKID_LARGEEDGES:
151 if (graph->chunk_large_edges)
152 chunk_repeated = 1;
153 else
154 graph->chunk_large_edges = data + chunk_offset;
155 break;
158 if (chunk_repeated) {
159 error("chunk id %08x appears multiple times", chunk_id);
160 goto cleanup_fail;
163 if (last_chunk_id == GRAPH_CHUNKID_OIDLOOKUP)
165 graph->num_commits = (chunk_offset - last_chunk_offset)
166 / graph->hash_len;
169 last_chunk_id = chunk_id;
170 last_chunk_offset = chunk_offset;
173 return graph;
175 cleanup_fail:
176 munmap(graph_map, graph_size);
177 close(fd);
178 exit(1);
181 /* global storage */
182 static struct commit_graph *commit_graph = NULL;
184 static void prepare_commit_graph_one(const char *obj_dir)
186 char *graph_name;
188 if (commit_graph)
189 return;
191 graph_name = get_commit_graph_filename(obj_dir);
192 commit_graph = load_commit_graph_one(graph_name);
194 FREE_AND_NULL(graph_name);
197 static int prepare_commit_graph_run_once = 0;
198 static void prepare_commit_graph(void)
200 struct alternate_object_database *alt;
201 char *obj_dir;
203 if (prepare_commit_graph_run_once)
204 return;
205 prepare_commit_graph_run_once = 1;
207 obj_dir = get_object_directory();
208 prepare_commit_graph_one(obj_dir);
209 prepare_alt_odb();
210 for (alt = alt_odb_list; !commit_graph && alt; alt = alt->next)
211 prepare_commit_graph_one(alt->path);
214 static void close_commit_graph(void)
216 if (!commit_graph)
217 return;
219 if (commit_graph->graph_fd >= 0) {
220 munmap((void *)commit_graph->data, commit_graph->data_len);
221 commit_graph->data = NULL;
222 close(commit_graph->graph_fd);
225 FREE_AND_NULL(commit_graph);
228 static int bsearch_graph(struct commit_graph *g, struct object_id *oid, uint32_t *pos)
230 return bsearch_hash(oid->hash, g->chunk_oid_fanout,
231 g->chunk_oid_lookup, g->hash_len, pos);
234 static struct commit_list **insert_parent_or_die(struct commit_graph *g,
235 uint64_t pos,
236 struct commit_list **pptr)
238 struct commit *c;
239 struct object_id oid;
240 hashcpy(oid.hash, g->chunk_oid_lookup + g->hash_len * pos);
241 c = lookup_commit(&oid);
242 if (!c)
243 die("could not find commit %s", oid_to_hex(&oid));
244 c->graph_pos = pos;
245 return &commit_list_insert(c, pptr)->next;
248 static int fill_commit_in_graph(struct commit *item, struct commit_graph *g, uint32_t pos)
250 uint32_t edge_value;
251 uint32_t *parent_data_ptr;
252 uint64_t date_low, date_high;
253 struct commit_list **pptr;
254 const unsigned char *commit_data = g->chunk_commit_data + (g->hash_len + 16) * pos;
256 item->object.parsed = 1;
257 item->graph_pos = pos;
259 item->maybe_tree = NULL;
261 date_high = get_be32(commit_data + g->hash_len + 8) & 0x3;
262 date_low = get_be32(commit_data + g->hash_len + 12);
263 item->date = (timestamp_t)((date_high << 32) | date_low);
265 pptr = &item->parents;
267 edge_value = get_be32(commit_data + g->hash_len);
268 if (edge_value == GRAPH_PARENT_NONE)
269 return 1;
270 pptr = insert_parent_or_die(g, edge_value, pptr);
272 edge_value = get_be32(commit_data + g->hash_len + 4);
273 if (edge_value == GRAPH_PARENT_NONE)
274 return 1;
275 if (!(edge_value & GRAPH_OCTOPUS_EDGES_NEEDED)) {
276 pptr = insert_parent_or_die(g, edge_value, pptr);
277 return 1;
280 parent_data_ptr = (uint32_t*)(g->chunk_large_edges +
281 4 * (uint64_t)(edge_value & GRAPH_EDGE_LAST_MASK));
282 do {
283 edge_value = get_be32(parent_data_ptr);
284 pptr = insert_parent_or_die(g,
285 edge_value & GRAPH_EDGE_LAST_MASK,
286 pptr);
287 parent_data_ptr++;
288 } while (!(edge_value & GRAPH_LAST_EDGE));
290 return 1;
293 int parse_commit_in_graph(struct commit *item)
295 if (!core_commit_graph)
296 return 0;
297 if (item->object.parsed)
298 return 1;
300 prepare_commit_graph();
301 if (commit_graph) {
302 uint32_t pos;
303 int found;
304 if (item->graph_pos != COMMIT_NOT_FROM_GRAPH) {
305 pos = item->graph_pos;
306 found = 1;
307 } else {
308 found = bsearch_graph(commit_graph, &(item->object.oid), &pos);
311 if (found)
312 return fill_commit_in_graph(item, commit_graph, pos);
315 return 0;
318 static struct tree *load_tree_for_commit(struct commit_graph *g, struct commit *c)
320 struct object_id oid;
321 const unsigned char *commit_data = g->chunk_commit_data +
322 GRAPH_DATA_WIDTH * (c->graph_pos);
324 hashcpy(oid.hash, commit_data);
325 c->maybe_tree = lookup_tree(&oid);
327 return c->maybe_tree;
330 struct tree *get_commit_tree_in_graph(const struct commit *c)
332 if (c->maybe_tree)
333 return c->maybe_tree;
334 if (c->graph_pos == COMMIT_NOT_FROM_GRAPH)
335 BUG("get_commit_tree_in_graph called from non-commit-graph commit");
337 return load_tree_for_commit(commit_graph, (struct commit *)c);
340 static void write_graph_chunk_fanout(struct hashfile *f,
341 struct commit **commits,
342 int nr_commits)
344 int i, count = 0;
345 struct commit **list = commits;
348 * Write the first-level table (the list is sorted,
349 * but we use a 256-entry lookup to be able to avoid
350 * having to do eight extra binary search iterations).
352 for (i = 0; i < 256; i++) {
353 while (count < nr_commits) {
354 if ((*list)->object.oid.hash[0] != i)
355 break;
356 count++;
357 list++;
360 hashwrite_be32(f, count);
364 static void write_graph_chunk_oids(struct hashfile *f, int hash_len,
365 struct commit **commits, int nr_commits)
367 struct commit **list = commits;
368 int count;
369 for (count = 0; count < nr_commits; count++, list++)
370 hashwrite(f, (*list)->object.oid.hash, (int)hash_len);
373 static const unsigned char *commit_to_sha1(size_t index, void *table)
375 struct commit **commits = table;
376 return commits[index]->object.oid.hash;
379 static void write_graph_chunk_data(struct hashfile *f, int hash_len,
380 struct commit **commits, int nr_commits)
382 struct commit **list = commits;
383 struct commit **last = commits + nr_commits;
384 uint32_t num_extra_edges = 0;
386 while (list < last) {
387 struct commit_list *parent;
388 int edge_value;
389 uint32_t packedDate[2];
391 parse_commit(*list);
392 hashwrite(f, get_commit_tree_oid(*list)->hash, hash_len);
394 parent = (*list)->parents;
396 if (!parent)
397 edge_value = GRAPH_PARENT_NONE;
398 else {
399 edge_value = sha1_pos(parent->item->object.oid.hash,
400 commits,
401 nr_commits,
402 commit_to_sha1);
404 if (edge_value < 0)
405 edge_value = GRAPH_PARENT_MISSING;
408 hashwrite_be32(f, edge_value);
410 if (parent)
411 parent = parent->next;
413 if (!parent)
414 edge_value = GRAPH_PARENT_NONE;
415 else if (parent->next)
416 edge_value = GRAPH_OCTOPUS_EDGES_NEEDED | num_extra_edges;
417 else {
418 edge_value = sha1_pos(parent->item->object.oid.hash,
419 commits,
420 nr_commits,
421 commit_to_sha1);
422 if (edge_value < 0)
423 edge_value = GRAPH_PARENT_MISSING;
426 hashwrite_be32(f, edge_value);
428 if (edge_value & GRAPH_OCTOPUS_EDGES_NEEDED) {
429 do {
430 num_extra_edges++;
431 parent = parent->next;
432 } while (parent);
435 if (sizeof((*list)->date) > 4)
436 packedDate[0] = htonl(((*list)->date >> 32) & 0x3);
437 else
438 packedDate[0] = 0;
440 packedDate[1] = htonl((*list)->date);
441 hashwrite(f, packedDate, 8);
443 list++;
447 static void write_graph_chunk_large_edges(struct hashfile *f,
448 struct commit **commits,
449 int nr_commits)
451 struct commit **list = commits;
452 struct commit **last = commits + nr_commits;
453 struct commit_list *parent;
455 while (list < last) {
456 int num_parents = 0;
457 for (parent = (*list)->parents; num_parents < 3 && parent;
458 parent = parent->next)
459 num_parents++;
461 if (num_parents <= 2) {
462 list++;
463 continue;
466 /* Since num_parents > 2, this initializer is safe. */
467 for (parent = (*list)->parents->next; parent; parent = parent->next) {
468 int edge_value = sha1_pos(parent->item->object.oid.hash,
469 commits,
470 nr_commits,
471 commit_to_sha1);
473 if (edge_value < 0)
474 edge_value = GRAPH_PARENT_MISSING;
475 else if (!parent->next)
476 edge_value |= GRAPH_LAST_EDGE;
478 hashwrite_be32(f, edge_value);
481 list++;
485 static int commit_compare(const void *_a, const void *_b)
487 const struct object_id *a = (const struct object_id *)_a;
488 const struct object_id *b = (const struct object_id *)_b;
489 return oidcmp(a, b);
492 struct packed_commit_list {
493 struct commit **list;
494 int nr;
495 int alloc;
498 struct packed_oid_list {
499 struct object_id *list;
500 int nr;
501 int alloc;
504 static int add_packed_commits(const struct object_id *oid,
505 struct packed_git *pack,
506 uint32_t pos,
507 void *data)
509 struct packed_oid_list *list = (struct packed_oid_list*)data;
510 enum object_type type;
511 off_t offset = nth_packed_object_offset(pack, pos);
512 struct object_info oi = OBJECT_INFO_INIT;
514 oi.typep = &type;
515 if (packed_object_info(pack, offset, &oi) < 0)
516 die("unable to get type of object %s", oid_to_hex(oid));
518 if (type != OBJ_COMMIT)
519 return 0;
521 ALLOC_GROW(list->list, list->nr + 1, list->alloc);
522 oidcpy(&(list->list[list->nr]), oid);
523 list->nr++;
525 return 0;
528 static void add_missing_parents(struct packed_oid_list *oids, struct commit *commit)
530 struct commit_list *parent;
531 for (parent = commit->parents; parent; parent = parent->next) {
532 if (!(parent->item->object.flags & UNINTERESTING)) {
533 ALLOC_GROW(oids->list, oids->nr + 1, oids->alloc);
534 oidcpy(&oids->list[oids->nr], &(parent->item->object.oid));
535 oids->nr++;
536 parent->item->object.flags |= UNINTERESTING;
541 static void close_reachable(struct packed_oid_list *oids)
543 int i;
544 struct commit *commit;
546 for (i = 0; i < oids->nr; i++) {
547 commit = lookup_commit(&oids->list[i]);
548 if (commit)
549 commit->object.flags |= UNINTERESTING;
553 * As this loop runs, oids->nr may grow, but not more
554 * than the number of missing commits in the reachable
555 * closure.
557 for (i = 0; i < oids->nr; i++) {
558 commit = lookup_commit(&oids->list[i]);
560 if (commit && !parse_commit(commit))
561 add_missing_parents(oids, commit);
564 for (i = 0; i < oids->nr; i++) {
565 commit = lookup_commit(&oids->list[i]);
567 if (commit)
568 commit->object.flags &= ~UNINTERESTING;
572 void write_commit_graph(const char *obj_dir,
573 const char **pack_indexes,
574 int nr_packs,
575 const char **commit_hex,
576 int nr_commits,
577 int append)
579 struct packed_oid_list oids;
580 struct packed_commit_list commits;
581 struct hashfile *f;
582 uint32_t i, count_distinct = 0;
583 char *graph_name;
584 int fd;
585 struct lock_file lk = LOCK_INIT;
586 uint32_t chunk_ids[5];
587 uint64_t chunk_offsets[5];
588 int num_chunks;
589 int num_extra_edges;
590 struct commit_list *parent;
592 oids.nr = 0;
593 oids.alloc = approximate_object_count() / 4;
595 if (append) {
596 prepare_commit_graph_one(obj_dir);
597 if (commit_graph)
598 oids.alloc += commit_graph->num_commits;
601 if (oids.alloc < 1024)
602 oids.alloc = 1024;
603 ALLOC_ARRAY(oids.list, oids.alloc);
605 if (append && commit_graph) {
606 for (i = 0; i < commit_graph->num_commits; i++) {
607 const unsigned char *hash = commit_graph->chunk_oid_lookup +
608 commit_graph->hash_len * i;
609 hashcpy(oids.list[oids.nr++].hash, hash);
613 if (pack_indexes) {
614 struct strbuf packname = STRBUF_INIT;
615 int dirlen;
616 strbuf_addf(&packname, "%s/pack/", obj_dir);
617 dirlen = packname.len;
618 for (i = 0; i < nr_packs; i++) {
619 struct packed_git *p;
620 strbuf_setlen(&packname, dirlen);
621 strbuf_addstr(&packname, pack_indexes[i]);
622 p = add_packed_git(packname.buf, packname.len, 1);
623 if (!p)
624 die("error adding pack %s", packname.buf);
625 if (open_pack_index(p))
626 die("error opening index for %s", packname.buf);
627 for_each_object_in_pack(p, add_packed_commits, &oids);
628 close_pack(p);
630 strbuf_release(&packname);
633 if (commit_hex) {
634 for (i = 0; i < nr_commits; i++) {
635 const char *end;
636 struct object_id oid;
637 struct commit *result;
639 if (commit_hex[i] && parse_oid_hex(commit_hex[i], &oid, &end))
640 continue;
642 result = lookup_commit_reference_gently(&oid, 1);
644 if (result) {
645 ALLOC_GROW(oids.list, oids.nr + 1, oids.alloc);
646 oidcpy(&oids.list[oids.nr], &(result->object.oid));
647 oids.nr++;
652 if (!pack_indexes && !commit_hex)
653 for_each_packed_object(add_packed_commits, &oids, 0);
655 close_reachable(&oids);
657 QSORT(oids.list, oids.nr, commit_compare);
659 count_distinct = 1;
660 for (i = 1; i < oids.nr; i++) {
661 if (oidcmp(&oids.list[i-1], &oids.list[i]))
662 count_distinct++;
665 if (count_distinct >= GRAPH_PARENT_MISSING)
666 die(_("the commit graph format cannot write %d commits"), count_distinct);
668 commits.nr = 0;
669 commits.alloc = count_distinct;
670 ALLOC_ARRAY(commits.list, commits.alloc);
672 num_extra_edges = 0;
673 for (i = 0; i < oids.nr; i++) {
674 int num_parents = 0;
675 if (i > 0 && !oidcmp(&oids.list[i-1], &oids.list[i]))
676 continue;
678 commits.list[commits.nr] = lookup_commit(&oids.list[i]);
679 parse_commit(commits.list[commits.nr]);
681 for (parent = commits.list[commits.nr]->parents;
682 parent; parent = parent->next)
683 num_parents++;
685 if (num_parents > 2)
686 num_extra_edges += num_parents - 1;
688 commits.nr++;
690 num_chunks = num_extra_edges ? 4 : 3;
692 if (commits.nr >= GRAPH_PARENT_MISSING)
693 die(_("too many commits to write graph"));
695 graph_name = get_commit_graph_filename(obj_dir);
696 fd = hold_lock_file_for_update(&lk, graph_name, 0);
698 if (fd < 0) {
699 struct strbuf folder = STRBUF_INIT;
700 strbuf_addstr(&folder, graph_name);
701 strbuf_setlen(&folder, strrchr(folder.buf, '/') - folder.buf);
703 if (mkdir(folder.buf, 0777) < 0)
704 die_errno(_("cannot mkdir %s"), folder.buf);
705 strbuf_release(&folder);
707 fd = hold_lock_file_for_update(&lk, graph_name, LOCK_DIE_ON_ERROR);
709 if (fd < 0)
710 die_errno("unable to create '%s'", graph_name);
713 f = hashfd(lk.tempfile->fd, lk.tempfile->filename.buf);
715 hashwrite_be32(f, GRAPH_SIGNATURE);
717 hashwrite_u8(f, GRAPH_VERSION);
718 hashwrite_u8(f, GRAPH_OID_VERSION);
719 hashwrite_u8(f, num_chunks);
720 hashwrite_u8(f, 0); /* unused padding byte */
722 chunk_ids[0] = GRAPH_CHUNKID_OIDFANOUT;
723 chunk_ids[1] = GRAPH_CHUNKID_OIDLOOKUP;
724 chunk_ids[2] = GRAPH_CHUNKID_DATA;
725 if (num_extra_edges)
726 chunk_ids[3] = GRAPH_CHUNKID_LARGEEDGES;
727 else
728 chunk_ids[3] = 0;
729 chunk_ids[4] = 0;
731 chunk_offsets[0] = 8 + (num_chunks + 1) * GRAPH_CHUNKLOOKUP_WIDTH;
732 chunk_offsets[1] = chunk_offsets[0] + GRAPH_FANOUT_SIZE;
733 chunk_offsets[2] = chunk_offsets[1] + GRAPH_OID_LEN * commits.nr;
734 chunk_offsets[3] = chunk_offsets[2] + (GRAPH_OID_LEN + 16) * commits.nr;
735 chunk_offsets[4] = chunk_offsets[3] + 4 * num_extra_edges;
737 for (i = 0; i <= num_chunks; i++) {
738 uint32_t chunk_write[3];
740 chunk_write[0] = htonl(chunk_ids[i]);
741 chunk_write[1] = htonl(chunk_offsets[i] >> 32);
742 chunk_write[2] = htonl(chunk_offsets[i] & 0xffffffff);
743 hashwrite(f, chunk_write, 12);
746 write_graph_chunk_fanout(f, commits.list, commits.nr);
747 write_graph_chunk_oids(f, GRAPH_OID_LEN, commits.list, commits.nr);
748 write_graph_chunk_data(f, GRAPH_OID_LEN, commits.list, commits.nr);
749 write_graph_chunk_large_edges(f, commits.list, commits.nr);
751 close_commit_graph();
752 finalize_hashfile(f, NULL, CSUM_HASH_IN_STREAM | CSUM_FSYNC);
753 commit_lock_file(&lk);
755 free(oids.list);
756 oids.alloc = 0;
757 oids.nr = 0;