sparse-index: add guard to ensure full index
[git] / pack-bitmap-write.c
1 #include "cache.h"
2 #include "object-store.h"
3 #include "commit.h"
4 #include "tag.h"
5 #include "diff.h"
6 #include "revision.h"
7 #include "list-objects.h"
8 #include "progress.h"
9 #include "pack-revindex.h"
10 #include "pack.h"
11 #include "pack-bitmap.h"
12 #include "hash-lookup.h"
13 #include "pack-objects.h"
14 #include "commit-reach.h"
15 #include "prio-queue.h"
16
17 struct bitmapped_commit {
18         struct commit *commit;
19         struct ewah_bitmap *bitmap;
20         struct ewah_bitmap *write_as;
21         int flags;
22         int xor_offset;
23         uint32_t commit_pos;
24 };
25
26 struct bitmap_writer {
27         struct ewah_bitmap *commits;
28         struct ewah_bitmap *trees;
29         struct ewah_bitmap *blobs;
30         struct ewah_bitmap *tags;
31
32         kh_oid_map_t *bitmaps;
33         struct packing_data *to_pack;
34
35         struct bitmapped_commit *selected;
36         unsigned int selected_nr, selected_alloc;
37
38         struct progress *progress;
39         int show_progress;
40         unsigned char pack_checksum[GIT_MAX_RAWSZ];
41 };
42
43 static struct bitmap_writer writer;
44
45 void bitmap_writer_show_progress(int show)
46 {
47         writer.show_progress = show;
48 }
49
50 /**
51  * Build the initial type index for the packfile
52  */
53 void bitmap_writer_build_type_index(struct packing_data *to_pack,
54                                     struct pack_idx_entry **index,
55                                     uint32_t index_nr)
56 {
57         uint32_t i;
58
59         writer.commits = ewah_new();
60         writer.trees = ewah_new();
61         writer.blobs = ewah_new();
62         writer.tags = ewah_new();
63         ALLOC_ARRAY(to_pack->in_pack_pos, to_pack->nr_objects);
64
65         for (i = 0; i < index_nr; ++i) {
66                 struct object_entry *entry = (struct object_entry *)index[i];
67                 enum object_type real_type;
68
69                 oe_set_in_pack_pos(to_pack, entry, i);
70
71                 switch (oe_type(entry)) {
72                 case OBJ_COMMIT:
73                 case OBJ_TREE:
74                 case OBJ_BLOB:
75                 case OBJ_TAG:
76                         real_type = oe_type(entry);
77                         break;
78
79                 default:
80                         real_type = oid_object_info(to_pack->repo,
81                                                     &entry->idx.oid, NULL);
82                         break;
83                 }
84
85                 switch (real_type) {
86                 case OBJ_COMMIT:
87                         ewah_set(writer.commits, i);
88                         break;
89
90                 case OBJ_TREE:
91                         ewah_set(writer.trees, i);
92                         break;
93
94                 case OBJ_BLOB:
95                         ewah_set(writer.blobs, i);
96                         break;
97
98                 case OBJ_TAG:
99                         ewah_set(writer.tags, i);
100                         break;
101
102                 default:
103                         die("Missing type information for %s (%d/%d)",
104                             oid_to_hex(&entry->idx.oid), real_type,
105                             oe_type(entry));
106                 }
107         }
108 }
109
110 /**
111  * Compute the actual bitmaps
112  */
113
114 static inline void push_bitmapped_commit(struct commit *commit)
115 {
116         if (writer.selected_nr >= writer.selected_alloc) {
117                 writer.selected_alloc = (writer.selected_alloc + 32) * 2;
118                 REALLOC_ARRAY(writer.selected, writer.selected_alloc);
119         }
120
121         writer.selected[writer.selected_nr].commit = commit;
122         writer.selected[writer.selected_nr].bitmap = NULL;
123         writer.selected[writer.selected_nr].flags = 0;
124
125         writer.selected_nr++;
126 }
127
128 static uint32_t find_object_pos(const struct object_id *oid)
129 {
130         struct object_entry *entry = packlist_find(writer.to_pack, oid);
131
132         if (!entry) {
133                 die("Failed to write bitmap index. Packfile doesn't have full closure "
134                         "(object %s is missing)", oid_to_hex(oid));
135         }
136
137         return oe_in_pack_pos(writer.to_pack, entry);
138 }
139
140 static void compute_xor_offsets(void)
141 {
142         static const int MAX_XOR_OFFSET_SEARCH = 10;
143
144         int i, next = 0;
145
146         while (next < writer.selected_nr) {
147                 struct bitmapped_commit *stored = &writer.selected[next];
148
149                 int best_offset = 0;
150                 struct ewah_bitmap *best_bitmap = stored->bitmap;
151                 struct ewah_bitmap *test_xor;
152
153                 for (i = 1; i <= MAX_XOR_OFFSET_SEARCH; ++i) {
154                         int curr = next - i;
155
156                         if (curr < 0)
157                                 break;
158
159                         test_xor = ewah_pool_new();
160                         ewah_xor(writer.selected[curr].bitmap, stored->bitmap, test_xor);
161
162                         if (test_xor->buffer_size < best_bitmap->buffer_size) {
163                                 if (best_bitmap != stored->bitmap)
164                                         ewah_pool_free(best_bitmap);
165
166                                 best_bitmap = test_xor;
167                                 best_offset = i;
168                         } else {
169                                 ewah_pool_free(test_xor);
170                         }
171                 }
172
173                 stored->xor_offset = best_offset;
174                 stored->write_as = best_bitmap;
175
176                 next++;
177         }
178 }
179
180 struct bb_commit {
181         struct commit_list *reverse_edges;
182         struct bitmap *commit_mask;
183         struct bitmap *bitmap;
184         unsigned selected:1,
185                  maximal:1;
186         unsigned idx; /* within selected array */
187 };
188
189 define_commit_slab(bb_data, struct bb_commit);
190
191 struct bitmap_builder {
192         struct bb_data data;
193         struct commit **commits;
194         size_t commits_nr, commits_alloc;
195 };
196
197 static void bitmap_builder_init(struct bitmap_builder *bb,
198                                 struct bitmap_writer *writer,
199                                 struct bitmap_index *old_bitmap)
200 {
201         struct rev_info revs;
202         struct commit *commit;
203         struct commit_list *reusable = NULL;
204         struct commit_list *r;
205         unsigned int i, num_maximal = 0;
206
207         memset(bb, 0, sizeof(*bb));
208         init_bb_data(&bb->data);
209
210         reset_revision_walk();
211         repo_init_revisions(writer->to_pack->repo, &revs, NULL);
212         revs.topo_order = 1;
213         revs.first_parent_only = 1;
214
215         for (i = 0; i < writer->selected_nr; i++) {
216                 struct commit *c = writer->selected[i].commit;
217                 struct bb_commit *ent = bb_data_at(&bb->data, c);
218
219                 ent->selected = 1;
220                 ent->maximal = 1;
221                 ent->idx = i;
222
223                 ent->commit_mask = bitmap_new();
224                 bitmap_set(ent->commit_mask, i);
225
226                 add_pending_object(&revs, &c->object, "");
227         }
228
229         if (prepare_revision_walk(&revs))
230                 die("revision walk setup failed");
231
232         while ((commit = get_revision(&revs))) {
233                 struct commit_list *p = commit->parents;
234                 struct bb_commit *c_ent;
235
236                 parse_commit_or_die(commit);
237
238                 c_ent = bb_data_at(&bb->data, commit);
239
240                 /*
241                  * If there is no commit_mask, there is no reason to iterate
242                  * over this commit; it is not selected (if it were, it would
243                  * not have a blank commit mask) and all its children have
244                  * existing bitmaps (see the comment starting with "This commit
245                  * has an existing bitmap" below), so it does not contribute
246                  * anything to the final bitmap file or its descendants.
247                  */
248                 if (!c_ent->commit_mask)
249                         continue;
250
251                 if (old_bitmap && bitmap_for_commit(old_bitmap, commit)) {
252                         /*
253                          * This commit has an existing bitmap, so we can
254                          * get its bits immediately without an object
255                          * walk. That is, it is reusable as-is and there is no
256                          * need to continue walking beyond it.
257                          *
258                          * Mark it as such and add it to bb->commits separately
259                          * to avoid allocating a position in the commit mask.
260                          */
261                         commit_list_insert(commit, &reusable);
262                         goto next;
263                 }
264
265                 if (c_ent->maximal) {
266                         num_maximal++;
267                         ALLOC_GROW(bb->commits, bb->commits_nr + 1, bb->commits_alloc);
268                         bb->commits[bb->commits_nr++] = commit;
269                 }
270
271                 if (p) {
272                         struct bb_commit *p_ent = bb_data_at(&bb->data, p->item);
273                         int c_not_p, p_not_c;
274
275                         if (!p_ent->commit_mask) {
276                                 p_ent->commit_mask = bitmap_new();
277                                 c_not_p = 1;
278                                 p_not_c = 0;
279                         } else {
280                                 c_not_p = bitmap_is_subset(c_ent->commit_mask, p_ent->commit_mask);
281                                 p_not_c = bitmap_is_subset(p_ent->commit_mask, c_ent->commit_mask);
282                         }
283
284                         if (!c_not_p)
285                                 continue;
286
287                         bitmap_or(p_ent->commit_mask, c_ent->commit_mask);
288
289                         if (p_not_c)
290                                 p_ent->maximal = 1;
291                         else {
292                                 p_ent->maximal = 0;
293                                 free_commit_list(p_ent->reverse_edges);
294                                 p_ent->reverse_edges = NULL;
295                         }
296
297                         if (c_ent->maximal) {
298                                 commit_list_insert(commit, &p_ent->reverse_edges);
299                         } else {
300                                 struct commit_list *cc = c_ent->reverse_edges;
301
302                                 for (; cc; cc = cc->next) {
303                                         if (!commit_list_contains(cc->item, p_ent->reverse_edges))
304                                                 commit_list_insert(cc->item, &p_ent->reverse_edges);
305                                 }
306                         }
307                 }
308
309 next:
310                 bitmap_free(c_ent->commit_mask);
311                 c_ent->commit_mask = NULL;
312         }
313
314         for (r = reusable; r; r = r->next) {
315                 ALLOC_GROW(bb->commits, bb->commits_nr + 1, bb->commits_alloc);
316                 bb->commits[bb->commits_nr++] = r->item;
317         }
318
319         trace2_data_intmax("pack-bitmap-write", the_repository,
320                            "num_selected_commits", writer->selected_nr);
321         trace2_data_intmax("pack-bitmap-write", the_repository,
322                            "num_maximal_commits", num_maximal);
323
324         free_commit_list(reusable);
325 }
326
327 static void bitmap_builder_clear(struct bitmap_builder *bb)
328 {
329         clear_bb_data(&bb->data);
330         free(bb->commits);
331         bb->commits_nr = bb->commits_alloc = 0;
332 }
333
334 static void fill_bitmap_tree(struct bitmap *bitmap,
335                              struct tree *tree)
336 {
337         uint32_t pos;
338         struct tree_desc desc;
339         struct name_entry entry;
340
341         /*
342          * If our bit is already set, then there is nothing to do. Both this
343          * tree and all of its children will be set.
344          */
345         pos = find_object_pos(&tree->object.oid);
346         if (bitmap_get(bitmap, pos))
347                 return;
348         bitmap_set(bitmap, pos);
349
350         if (parse_tree(tree) < 0)
351                 die("unable to load tree object %s",
352                     oid_to_hex(&tree->object.oid));
353         init_tree_desc(&desc, tree->buffer, tree->size);
354
355         while (tree_entry(&desc, &entry)) {
356                 switch (object_type(entry.mode)) {
357                 case OBJ_TREE:
358                         fill_bitmap_tree(bitmap,
359                                          lookup_tree(the_repository, &entry.oid));
360                         break;
361                 case OBJ_BLOB:
362                         bitmap_set(bitmap, find_object_pos(&entry.oid));
363                         break;
364                 default:
365                         /* Gitlink, etc; not reachable */
366                         break;
367                 }
368         }
369
370         free_tree_buffer(tree);
371 }
372
373 static void fill_bitmap_commit(struct bb_commit *ent,
374                                struct commit *commit,
375                                struct prio_queue *queue,
376                                struct prio_queue *tree_queue,
377                                struct bitmap_index *old_bitmap,
378                                const uint32_t *mapping)
379 {
380         if (!ent->bitmap)
381                 ent->bitmap = bitmap_new();
382
383         prio_queue_put(queue, commit);
384
385         while (queue->nr) {
386                 struct commit_list *p;
387                 struct commit *c = prio_queue_get(queue);
388
389                 if (old_bitmap && mapping) {
390                         struct ewah_bitmap *old = bitmap_for_commit(old_bitmap, c);
391                         /*
392                          * If this commit has an old bitmap, then translate that
393                          * bitmap and add its bits to this one. No need to walk
394                          * parents or the tree for this commit.
395                          */
396                         if (old && !rebuild_bitmap(mapping, old, ent->bitmap))
397                                 continue;
398                 }
399
400                 /*
401                  * Mark ourselves and queue our tree. The commit
402                  * walk ensures we cover all parents.
403                  */
404                 bitmap_set(ent->bitmap, find_object_pos(&c->object.oid));
405                 prio_queue_put(tree_queue, get_commit_tree(c));
406
407                 for (p = c->parents; p; p = p->next) {
408                         int pos = find_object_pos(&p->item->object.oid);
409                         if (!bitmap_get(ent->bitmap, pos)) {
410                                 bitmap_set(ent->bitmap, pos);
411                                 prio_queue_put(queue, p->item);
412                         }
413                 }
414         }
415
416         while (tree_queue->nr)
417                 fill_bitmap_tree(ent->bitmap, prio_queue_get(tree_queue));
418 }
419
420 static void store_selected(struct bb_commit *ent, struct commit *commit)
421 {
422         struct bitmapped_commit *stored = &writer.selected[ent->idx];
423         khiter_t hash_pos;
424         int hash_ret;
425
426         stored->bitmap = bitmap_to_ewah(ent->bitmap);
427
428         hash_pos = kh_put_oid_map(writer.bitmaps, commit->object.oid, &hash_ret);
429         if (hash_ret == 0)
430                 die("Duplicate entry when writing index: %s",
431                     oid_to_hex(&commit->object.oid));
432         kh_value(writer.bitmaps, hash_pos) = stored;
433 }
434
435 void bitmap_writer_build(struct packing_data *to_pack)
436 {
437         struct bitmap_builder bb;
438         size_t i;
439         int nr_stored = 0; /* for progress */
440         struct prio_queue queue = { compare_commits_by_gen_then_commit_date };
441         struct prio_queue tree_queue = { NULL };
442         struct bitmap_index *old_bitmap;
443         uint32_t *mapping;
444
445         writer.bitmaps = kh_init_oid_map();
446         writer.to_pack = to_pack;
447
448         if (writer.show_progress)
449                 writer.progress = start_progress("Building bitmaps", writer.selected_nr);
450         trace2_region_enter("pack-bitmap-write", "building_bitmaps_total",
451                             the_repository);
452
453         old_bitmap = prepare_bitmap_git(to_pack->repo);
454         if (old_bitmap)
455                 mapping = create_bitmap_mapping(old_bitmap, to_pack);
456         else
457                 mapping = NULL;
458
459         bitmap_builder_init(&bb, &writer, old_bitmap);
460         for (i = bb.commits_nr; i > 0; i--) {
461                 struct commit *commit = bb.commits[i-1];
462                 struct bb_commit *ent = bb_data_at(&bb.data, commit);
463                 struct commit *child;
464                 int reused = 0;
465
466                 fill_bitmap_commit(ent, commit, &queue, &tree_queue,
467                                    old_bitmap, mapping);
468
469                 if (ent->selected) {
470                         store_selected(ent, commit);
471                         nr_stored++;
472                         display_progress(writer.progress, nr_stored);
473                 }
474
475                 while ((child = pop_commit(&ent->reverse_edges))) {
476                         struct bb_commit *child_ent =
477                                 bb_data_at(&bb.data, child);
478
479                         if (child_ent->bitmap)
480                                 bitmap_or(child_ent->bitmap, ent->bitmap);
481                         else if (reused)
482                                 child_ent->bitmap = bitmap_dup(ent->bitmap);
483                         else {
484                                 child_ent->bitmap = ent->bitmap;
485                                 reused = 1;
486                         }
487                 }
488                 if (!reused)
489                         bitmap_free(ent->bitmap);
490                 ent->bitmap = NULL;
491         }
492         clear_prio_queue(&queue);
493         clear_prio_queue(&tree_queue);
494         bitmap_builder_clear(&bb);
495         free(mapping);
496
497         trace2_region_leave("pack-bitmap-write", "building_bitmaps_total",
498                             the_repository);
499
500         stop_progress(&writer.progress);
501
502         compute_xor_offsets();
503 }
504
505 /**
506  * Select the commits that will be bitmapped
507  */
508 static inline unsigned int next_commit_index(unsigned int idx)
509 {
510         static const unsigned int MIN_COMMITS = 100;
511         static const unsigned int MAX_COMMITS = 5000;
512
513         static const unsigned int MUST_REGION = 100;
514         static const unsigned int MIN_REGION = 20000;
515
516         unsigned int offset, next;
517
518         if (idx <= MUST_REGION)
519                 return 0;
520
521         if (idx <= MIN_REGION) {
522                 offset = idx - MUST_REGION;
523                 return (offset < MIN_COMMITS) ? offset : MIN_COMMITS;
524         }
525
526         offset = idx - MIN_REGION;
527         next = (offset < MAX_COMMITS) ? offset : MAX_COMMITS;
528
529         return (next > MIN_COMMITS) ? next : MIN_COMMITS;
530 }
531
532 static int date_compare(const void *_a, const void *_b)
533 {
534         struct commit *a = *(struct commit **)_a;
535         struct commit *b = *(struct commit **)_b;
536         return (long)b->date - (long)a->date;
537 }
538
539 void bitmap_writer_select_commits(struct commit **indexed_commits,
540                                   unsigned int indexed_commits_nr,
541                                   int max_bitmaps)
542 {
543         unsigned int i = 0, j, next;
544
545         QSORT(indexed_commits, indexed_commits_nr, date_compare);
546
547         if (writer.show_progress)
548                 writer.progress = start_progress("Selecting bitmap commits", 0);
549
550         if (indexed_commits_nr < 100) {
551                 for (i = 0; i < indexed_commits_nr; ++i)
552                         push_bitmapped_commit(indexed_commits[i]);
553                 return;
554         }
555
556         for (;;) {
557                 struct commit *chosen = NULL;
558
559                 next = next_commit_index(i);
560
561                 if (i + next >= indexed_commits_nr)
562                         break;
563
564                 if (max_bitmaps > 0 && writer.selected_nr >= max_bitmaps) {
565                         writer.selected_nr = max_bitmaps;
566                         break;
567                 }
568
569                 if (next == 0) {
570                         chosen = indexed_commits[i];
571                 } else {
572                         chosen = indexed_commits[i + next];
573
574                         for (j = 0; j <= next; ++j) {
575                                 struct commit *cm = indexed_commits[i + j];
576
577                                 if ((cm->object.flags & NEEDS_BITMAP) != 0) {
578                                         chosen = cm;
579                                         break;
580                                 }
581
582                                 if (cm->parents && cm->parents->next)
583                                         chosen = cm;
584                         }
585                 }
586
587                 push_bitmapped_commit(chosen);
588
589                 i += next + 1;
590                 display_progress(writer.progress, i);
591         }
592
593         stop_progress(&writer.progress);
594 }
595
596
597 static int hashwrite_ewah_helper(void *f, const void *buf, size_t len)
598 {
599         /* hashwrite will die on error */
600         hashwrite(f, buf, len);
601         return len;
602 }
603
604 /**
605  * Write the bitmap index to disk
606  */
607 static inline void dump_bitmap(struct hashfile *f, struct ewah_bitmap *bitmap)
608 {
609         if (ewah_serialize_to(bitmap, hashwrite_ewah_helper, f) < 0)
610                 die("Failed to write bitmap index");
611 }
612
613 static const struct object_id *oid_access(size_t pos, const void *table)
614 {
615         const struct pack_idx_entry * const *index = table;
616         return &index[pos]->oid;
617 }
618
619 static void write_selected_commits_v1(struct hashfile *f,
620                                       struct pack_idx_entry **index,
621                                       uint32_t index_nr)
622 {
623         int i;
624
625         for (i = 0; i < writer.selected_nr; ++i) {
626                 struct bitmapped_commit *stored = &writer.selected[i];
627
628                 int commit_pos =
629                         oid_pos(&stored->commit->object.oid, index, index_nr, oid_access);
630
631                 if (commit_pos < 0)
632                         BUG("trying to write commit not in index");
633
634                 hashwrite_be32(f, commit_pos);
635                 hashwrite_u8(f, stored->xor_offset);
636                 hashwrite_u8(f, stored->flags);
637
638                 dump_bitmap(f, stored->write_as);
639         }
640 }
641
642 static void write_hash_cache(struct hashfile *f,
643                              struct pack_idx_entry **index,
644                              uint32_t index_nr)
645 {
646         uint32_t i;
647
648         for (i = 0; i < index_nr; ++i) {
649                 struct object_entry *entry = (struct object_entry *)index[i];
650                 hashwrite_be32(f, entry->hash);
651         }
652 }
653
654 void bitmap_writer_set_checksum(unsigned char *sha1)
655 {
656         hashcpy(writer.pack_checksum, sha1);
657 }
658
659 void bitmap_writer_finish(struct pack_idx_entry **index,
660                           uint32_t index_nr,
661                           const char *filename,
662                           uint16_t options)
663 {
664         static uint16_t default_version = 1;
665         static uint16_t flags = BITMAP_OPT_FULL_DAG;
666         struct strbuf tmp_file = STRBUF_INIT;
667         struct hashfile *f;
668
669         struct bitmap_disk_header header;
670
671         int fd = odb_mkstemp(&tmp_file, "pack/tmp_bitmap_XXXXXX");
672
673         f = hashfd(fd, tmp_file.buf);
674
675         memcpy(header.magic, BITMAP_IDX_SIGNATURE, sizeof(BITMAP_IDX_SIGNATURE));
676         header.version = htons(default_version);
677         header.options = htons(flags | options);
678         header.entry_count = htonl(writer.selected_nr);
679         hashcpy(header.checksum, writer.pack_checksum);
680
681         hashwrite(f, &header, sizeof(header) - GIT_MAX_RAWSZ + the_hash_algo->rawsz);
682         dump_bitmap(f, writer.commits);
683         dump_bitmap(f, writer.trees);
684         dump_bitmap(f, writer.blobs);
685         dump_bitmap(f, writer.tags);
686         write_selected_commits_v1(f, index, index_nr);
687
688         if (options & BITMAP_OPT_HASH_CACHE)
689                 write_hash_cache(f, index, index_nr);
690
691         finalize_hashfile(f, NULL, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
692
693         if (adjust_shared_perm(tmp_file.buf))
694                 die_errno("unable to make temporary bitmap file readable");
695
696         if (rename(tmp_file.buf, filename))
697                 die_errno("unable to rename temporary bitmap file to '%s'", filename);
698
699         strbuf_release(&tmp_file);
700 }