pack-bitmap.c: support 'tree:0' filtering
[git] / pack-bitmap.c
1 #include "cache.h"
2 #include "commit.h"
3 #include "tag.h"
4 #include "diff.h"
5 #include "revision.h"
6 #include "progress.h"
7 #include "list-objects.h"
8 #include "pack.h"
9 #include "pack-bitmap.h"
10 #include "pack-revindex.h"
11 #include "pack-objects.h"
12 #include "packfile.h"
13 #include "repository.h"
14 #include "object-store.h"
15 #include "list-objects-filter-options.h"
16
17 /*
18  * An entry on the bitmap index, representing the bitmap for a given
19  * commit.
20  */
21 struct stored_bitmap {
22         struct object_id oid;
23         struct ewah_bitmap *root;
24         struct stored_bitmap *xor;
25         int flags;
26 };
27
28 /*
29  * The active bitmap index for a repository. By design, repositories only have
30  * a single bitmap index available (the index for the biggest packfile in
31  * the repository), since bitmap indexes need full closure.
32  *
33  * If there is more than one bitmap index available (e.g. because of alternates),
34  * the active bitmap index is the largest one.
35  */
36 struct bitmap_index {
37         /* Packfile to which this bitmap index belongs to */
38         struct packed_git *pack;
39
40         /*
41          * Mark the first `reuse_objects` in the packfile as reused:
42          * they will be sent as-is without using them for repacking
43          * calculations
44          */
45         uint32_t reuse_objects;
46
47         /* mmapped buffer of the whole bitmap index */
48         unsigned char *map;
49         size_t map_size; /* size of the mmaped buffer */
50         size_t map_pos; /* current position when loading the index */
51
52         /*
53          * Type indexes.
54          *
55          * Each bitmap marks which objects in the packfile  are of the given
56          * type. This provides type information when yielding the objects from
57          * the packfile during a walk, which allows for better delta bases.
58          */
59         struct ewah_bitmap *commits;
60         struct ewah_bitmap *trees;
61         struct ewah_bitmap *blobs;
62         struct ewah_bitmap *tags;
63
64         /* Map from object ID -> `stored_bitmap` for all the bitmapped commits */
65         kh_oid_map_t *bitmaps;
66
67         /* Number of bitmapped commits */
68         uint32_t entry_count;
69
70         /* If not NULL, this is a name-hash cache pointing into map. */
71         uint32_t *hashes;
72
73         /*
74          * Extended index.
75          *
76          * When trying to perform bitmap operations with objects that are not
77          * packed in `pack`, these objects are added to this "fake index" and
78          * are assumed to appear at the end of the packfile for all operations
79          */
80         struct eindex {
81                 struct object **objects;
82                 uint32_t *hashes;
83                 uint32_t count, alloc;
84                 kh_oid_pos_t *positions;
85         } ext_index;
86
87         /* Bitmap result of the last performed walk */
88         struct bitmap *result;
89
90         /* "have" bitmap from the last performed walk */
91         struct bitmap *haves;
92
93         /* Version of the bitmap index */
94         unsigned int version;
95 };
96
97 static struct ewah_bitmap *lookup_stored_bitmap(struct stored_bitmap *st)
98 {
99         struct ewah_bitmap *parent;
100         struct ewah_bitmap *composed;
101
102         if (st->xor == NULL)
103                 return st->root;
104
105         composed = ewah_pool_new();
106         parent = lookup_stored_bitmap(st->xor);
107         ewah_xor(st->root, parent, composed);
108
109         ewah_pool_free(st->root);
110         st->root = composed;
111         st->xor = NULL;
112
113         return composed;
114 }
115
116 /*
117  * Read a bitmap from the current read position on the mmaped
118  * index, and increase the read position accordingly
119  */
120 static struct ewah_bitmap *read_bitmap_1(struct bitmap_index *index)
121 {
122         struct ewah_bitmap *b = ewah_pool_new();
123
124         ssize_t bitmap_size = ewah_read_mmap(b,
125                 index->map + index->map_pos,
126                 index->map_size - index->map_pos);
127
128         if (bitmap_size < 0) {
129                 error("Failed to load bitmap index (corrupted?)");
130                 ewah_pool_free(b);
131                 return NULL;
132         }
133
134         index->map_pos += bitmap_size;
135         return b;
136 }
137
138 static int load_bitmap_header(struct bitmap_index *index)
139 {
140         struct bitmap_disk_header *header = (void *)index->map;
141
142         if (index->map_size < sizeof(*header) + the_hash_algo->rawsz)
143                 return error("Corrupted bitmap index (missing header data)");
144
145         if (memcmp(header->magic, BITMAP_IDX_SIGNATURE, sizeof(BITMAP_IDX_SIGNATURE)) != 0)
146                 return error("Corrupted bitmap index file (wrong header)");
147
148         index->version = ntohs(header->version);
149         if (index->version != 1)
150                 return error("Unsupported version for bitmap index file (%d)", index->version);
151
152         /* Parse known bitmap format options */
153         {
154                 uint32_t flags = ntohs(header->options);
155
156                 if ((flags & BITMAP_OPT_FULL_DAG) == 0)
157                         return error("Unsupported options for bitmap index file "
158                                 "(Git requires BITMAP_OPT_FULL_DAG)");
159
160                 if (flags & BITMAP_OPT_HASH_CACHE) {
161                         unsigned char *end = index->map + index->map_size - the_hash_algo->rawsz;
162                         index->hashes = ((uint32_t *)end) - index->pack->num_objects;
163                 }
164         }
165
166         index->entry_count = ntohl(header->entry_count);
167         index->map_pos += sizeof(*header) - GIT_MAX_RAWSZ + the_hash_algo->rawsz;
168         return 0;
169 }
170
171 static struct stored_bitmap *store_bitmap(struct bitmap_index *index,
172                                           struct ewah_bitmap *root,
173                                           const struct object_id *oid,
174                                           struct stored_bitmap *xor_with,
175                                           int flags)
176 {
177         struct stored_bitmap *stored;
178         khiter_t hash_pos;
179         int ret;
180
181         stored = xmalloc(sizeof(struct stored_bitmap));
182         stored->root = root;
183         stored->xor = xor_with;
184         stored->flags = flags;
185         oidcpy(&stored->oid, oid);
186
187         hash_pos = kh_put_oid_map(index->bitmaps, stored->oid, &ret);
188
189         /* a 0 return code means the insertion succeeded with no changes,
190          * because the SHA1 already existed on the map. this is bad, there
191          * shouldn't be duplicated commits in the index */
192         if (ret == 0) {
193                 error("Duplicate entry in bitmap index: %s", oid_to_hex(oid));
194                 return NULL;
195         }
196
197         kh_value(index->bitmaps, hash_pos) = stored;
198         return stored;
199 }
200
201 static inline uint32_t read_be32(const unsigned char *buffer, size_t *pos)
202 {
203         uint32_t result = get_be32(buffer + *pos);
204         (*pos) += sizeof(result);
205         return result;
206 }
207
208 static inline uint8_t read_u8(const unsigned char *buffer, size_t *pos)
209 {
210         return buffer[(*pos)++];
211 }
212
213 #define MAX_XOR_OFFSET 160
214
215 static int load_bitmap_entries_v1(struct bitmap_index *index)
216 {
217         uint32_t i;
218         struct stored_bitmap *recent_bitmaps[MAX_XOR_OFFSET] = { NULL };
219
220         for (i = 0; i < index->entry_count; ++i) {
221                 int xor_offset, flags;
222                 struct ewah_bitmap *bitmap = NULL;
223                 struct stored_bitmap *xor_bitmap = NULL;
224                 uint32_t commit_idx_pos;
225                 struct object_id oid;
226
227                 commit_idx_pos = read_be32(index->map, &index->map_pos);
228                 xor_offset = read_u8(index->map, &index->map_pos);
229                 flags = read_u8(index->map, &index->map_pos);
230
231                 nth_packed_object_id(&oid, index->pack, commit_idx_pos);
232
233                 bitmap = read_bitmap_1(index);
234                 if (!bitmap)
235                         return -1;
236
237                 if (xor_offset > MAX_XOR_OFFSET || xor_offset > i)
238                         return error("Corrupted bitmap pack index");
239
240                 if (xor_offset > 0) {
241                         xor_bitmap = recent_bitmaps[(i - xor_offset) % MAX_XOR_OFFSET];
242
243                         if (xor_bitmap == NULL)
244                                 return error("Invalid XOR offset in bitmap pack index");
245                 }
246
247                 recent_bitmaps[i % MAX_XOR_OFFSET] = store_bitmap(
248                         index, bitmap, &oid, xor_bitmap, flags);
249         }
250
251         return 0;
252 }
253
254 static char *pack_bitmap_filename(struct packed_git *p)
255 {
256         size_t len;
257
258         if (!strip_suffix(p->pack_name, ".pack", &len))
259                 BUG("pack_name does not end in .pack");
260         return xstrfmt("%.*s.bitmap", (int)len, p->pack_name);
261 }
262
263 static int open_pack_bitmap_1(struct bitmap_index *bitmap_git, struct packed_git *packfile)
264 {
265         int fd;
266         struct stat st;
267         char *idx_name;
268
269         if (open_pack_index(packfile))
270                 return -1;
271
272         idx_name = pack_bitmap_filename(packfile);
273         fd = git_open(idx_name);
274         free(idx_name);
275
276         if (fd < 0)
277                 return -1;
278
279         if (fstat(fd, &st)) {
280                 close(fd);
281                 return -1;
282         }
283
284         if (bitmap_git->pack) {
285                 warning("ignoring extra bitmap file: %s", packfile->pack_name);
286                 close(fd);
287                 return -1;
288         }
289
290         bitmap_git->pack = packfile;
291         bitmap_git->map_size = xsize_t(st.st_size);
292         bitmap_git->map = xmmap(NULL, bitmap_git->map_size, PROT_READ, MAP_PRIVATE, fd, 0);
293         bitmap_git->map_pos = 0;
294         close(fd);
295
296         if (load_bitmap_header(bitmap_git) < 0) {
297                 munmap(bitmap_git->map, bitmap_git->map_size);
298                 bitmap_git->map = NULL;
299                 bitmap_git->map_size = 0;
300                 return -1;
301         }
302
303         return 0;
304 }
305
306 static int load_pack_bitmap(struct bitmap_index *bitmap_git)
307 {
308         assert(bitmap_git->map);
309
310         bitmap_git->bitmaps = kh_init_oid_map();
311         bitmap_git->ext_index.positions = kh_init_oid_pos();
312         if (load_pack_revindex(bitmap_git->pack))
313                 goto failed;
314
315         if (!(bitmap_git->commits = read_bitmap_1(bitmap_git)) ||
316                 !(bitmap_git->trees = read_bitmap_1(bitmap_git)) ||
317                 !(bitmap_git->blobs = read_bitmap_1(bitmap_git)) ||
318                 !(bitmap_git->tags = read_bitmap_1(bitmap_git)))
319                 goto failed;
320
321         if (load_bitmap_entries_v1(bitmap_git) < 0)
322                 goto failed;
323
324         return 0;
325
326 failed:
327         munmap(bitmap_git->map, bitmap_git->map_size);
328         bitmap_git->map = NULL;
329         bitmap_git->map_size = 0;
330
331         kh_destroy_oid_map(bitmap_git->bitmaps);
332         bitmap_git->bitmaps = NULL;
333
334         kh_destroy_oid_pos(bitmap_git->ext_index.positions);
335         bitmap_git->ext_index.positions = NULL;
336
337         return -1;
338 }
339
340 static int open_pack_bitmap(struct repository *r,
341                             struct bitmap_index *bitmap_git)
342 {
343         struct packed_git *p;
344         int ret = -1;
345
346         assert(!bitmap_git->map);
347
348         for (p = get_all_packs(r); p; p = p->next) {
349                 if (open_pack_bitmap_1(bitmap_git, p) == 0)
350                         ret = 0;
351         }
352
353         return ret;
354 }
355
356 struct bitmap_index *prepare_bitmap_git(struct repository *r)
357 {
358         struct bitmap_index *bitmap_git = xcalloc(1, sizeof(*bitmap_git));
359
360         if (!open_pack_bitmap(r, bitmap_git) && !load_pack_bitmap(bitmap_git))
361                 return bitmap_git;
362
363         free_bitmap_index(bitmap_git);
364         return NULL;
365 }
366
367 struct include_data {
368         struct bitmap_index *bitmap_git;
369         struct bitmap *base;
370         struct bitmap *seen;
371 };
372
373 static inline int bitmap_position_extended(struct bitmap_index *bitmap_git,
374                                            const struct object_id *oid)
375 {
376         kh_oid_pos_t *positions = bitmap_git->ext_index.positions;
377         khiter_t pos = kh_get_oid_pos(positions, *oid);
378
379         if (pos < kh_end(positions)) {
380                 int bitmap_pos = kh_value(positions, pos);
381                 return bitmap_pos + bitmap_git->pack->num_objects;
382         }
383
384         return -1;
385 }
386
387 static inline int bitmap_position_packfile(struct bitmap_index *bitmap_git,
388                                            const struct object_id *oid)
389 {
390         off_t offset = find_pack_entry_one(oid->hash, bitmap_git->pack);
391         if (!offset)
392                 return -1;
393
394         return find_revindex_position(bitmap_git->pack, offset);
395 }
396
397 static int bitmap_position(struct bitmap_index *bitmap_git,
398                            const struct object_id *oid)
399 {
400         int pos = bitmap_position_packfile(bitmap_git, oid);
401         return (pos >= 0) ? pos : bitmap_position_extended(bitmap_git, oid);
402 }
403
404 static int ext_index_add_object(struct bitmap_index *bitmap_git,
405                                 struct object *object, const char *name)
406 {
407         struct eindex *eindex = &bitmap_git->ext_index;
408
409         khiter_t hash_pos;
410         int hash_ret;
411         int bitmap_pos;
412
413         hash_pos = kh_put_oid_pos(eindex->positions, object->oid, &hash_ret);
414         if (hash_ret > 0) {
415                 if (eindex->count >= eindex->alloc) {
416                         eindex->alloc = (eindex->alloc + 16) * 3 / 2;
417                         REALLOC_ARRAY(eindex->objects, eindex->alloc);
418                         REALLOC_ARRAY(eindex->hashes, eindex->alloc);
419                 }
420
421                 bitmap_pos = eindex->count;
422                 eindex->objects[eindex->count] = object;
423                 eindex->hashes[eindex->count] = pack_name_hash(name);
424                 kh_value(eindex->positions, hash_pos) = bitmap_pos;
425                 eindex->count++;
426         } else {
427                 bitmap_pos = kh_value(eindex->positions, hash_pos);
428         }
429
430         return bitmap_pos + bitmap_git->pack->num_objects;
431 }
432
433 struct bitmap_show_data {
434         struct bitmap_index *bitmap_git;
435         struct bitmap *base;
436 };
437
438 static void show_object(struct object *object, const char *name, void *data_)
439 {
440         struct bitmap_show_data *data = data_;
441         int bitmap_pos;
442
443         bitmap_pos = bitmap_position(data->bitmap_git, &object->oid);
444
445         if (bitmap_pos < 0)
446                 bitmap_pos = ext_index_add_object(data->bitmap_git, object,
447                                                   name);
448
449         bitmap_set(data->base, bitmap_pos);
450 }
451
452 static void show_commit(struct commit *commit, void *data)
453 {
454 }
455
456 static int add_to_include_set(struct bitmap_index *bitmap_git,
457                               struct include_data *data,
458                               const struct object_id *oid,
459                               int bitmap_pos)
460 {
461         khiter_t hash_pos;
462
463         if (data->seen && bitmap_get(data->seen, bitmap_pos))
464                 return 0;
465
466         if (bitmap_get(data->base, bitmap_pos))
467                 return 0;
468
469         hash_pos = kh_get_oid_map(bitmap_git->bitmaps, *oid);
470         if (hash_pos < kh_end(bitmap_git->bitmaps)) {
471                 struct stored_bitmap *st = kh_value(bitmap_git->bitmaps, hash_pos);
472                 bitmap_or_ewah(data->base, lookup_stored_bitmap(st));
473                 return 0;
474         }
475
476         bitmap_set(data->base, bitmap_pos);
477         return 1;
478 }
479
480 static int should_include(struct commit *commit, void *_data)
481 {
482         struct include_data *data = _data;
483         int bitmap_pos;
484
485         bitmap_pos = bitmap_position(data->bitmap_git, &commit->object.oid);
486         if (bitmap_pos < 0)
487                 bitmap_pos = ext_index_add_object(data->bitmap_git,
488                                                   (struct object *)commit,
489                                                   NULL);
490
491         if (!add_to_include_set(data->bitmap_git, data, &commit->object.oid,
492                                 bitmap_pos)) {
493                 struct commit_list *parent = commit->parents;
494
495                 while (parent) {
496                         parent->item->object.flags |= SEEN;
497                         parent = parent->next;
498                 }
499
500                 return 0;
501         }
502
503         return 1;
504 }
505
506 static struct bitmap *find_objects(struct bitmap_index *bitmap_git,
507                                    struct rev_info *revs,
508                                    struct object_list *roots,
509                                    struct bitmap *seen)
510 {
511         struct bitmap *base = NULL;
512         int needs_walk = 0;
513
514         struct object_list *not_mapped = NULL;
515
516         /*
517          * Go through all the roots for the walk. The ones that have bitmaps
518          * on the bitmap index will be `or`ed together to form an initial
519          * global reachability analysis.
520          *
521          * The ones without bitmaps in the index will be stored in the
522          * `not_mapped_list` for further processing.
523          */
524         while (roots) {
525                 struct object *object = roots->item;
526                 roots = roots->next;
527
528                 if (object->type == OBJ_COMMIT) {
529                         khiter_t pos = kh_get_oid_map(bitmap_git->bitmaps, object->oid);
530
531                         if (pos < kh_end(bitmap_git->bitmaps)) {
532                                 struct stored_bitmap *st = kh_value(bitmap_git->bitmaps, pos);
533                                 struct ewah_bitmap *or_with = lookup_stored_bitmap(st);
534
535                                 if (base == NULL)
536                                         base = ewah_to_bitmap(or_with);
537                                 else
538                                         bitmap_or_ewah(base, or_with);
539
540                                 object->flags |= SEEN;
541                                 continue;
542                         }
543                 }
544
545                 object_list_insert(object, &not_mapped);
546         }
547
548         /*
549          * Best case scenario: We found bitmaps for all the roots,
550          * so the resulting `or` bitmap has the full reachability analysis
551          */
552         if (not_mapped == NULL)
553                 return base;
554
555         roots = not_mapped;
556
557         /*
558          * Let's iterate through all the roots that don't have bitmaps to
559          * check if we can determine them to be reachable from the existing
560          * global bitmap.
561          *
562          * If we cannot find them in the existing global bitmap, we'll need
563          * to push them to an actual walk and run it until we can confirm
564          * they are reachable
565          */
566         while (roots) {
567                 struct object *object = roots->item;
568                 int pos;
569
570                 roots = roots->next;
571                 pos = bitmap_position(bitmap_git, &object->oid);
572
573                 if (pos < 0 || base == NULL || !bitmap_get(base, pos)) {
574                         object->flags &= ~UNINTERESTING;
575                         add_pending_object(revs, object, "");
576                         needs_walk = 1;
577                 } else {
578                         object->flags |= SEEN;
579                 }
580         }
581
582         if (needs_walk) {
583                 struct include_data incdata;
584                 struct bitmap_show_data show_data;
585
586                 if (base == NULL)
587                         base = bitmap_new();
588
589                 incdata.bitmap_git = bitmap_git;
590                 incdata.base = base;
591                 incdata.seen = seen;
592
593                 revs->include_check = should_include;
594                 revs->include_check_data = &incdata;
595
596                 if (prepare_revision_walk(revs))
597                         die("revision walk setup failed");
598
599                 show_data.bitmap_git = bitmap_git;
600                 show_data.base = base;
601
602                 traverse_commit_list(revs, show_commit, show_object,
603                                      &show_data);
604         }
605
606         return base;
607 }
608
609 static void show_extended_objects(struct bitmap_index *bitmap_git,
610                                   struct rev_info *revs,
611                                   show_reachable_fn show_reach)
612 {
613         struct bitmap *objects = bitmap_git->result;
614         struct eindex *eindex = &bitmap_git->ext_index;
615         uint32_t i;
616
617         for (i = 0; i < eindex->count; ++i) {
618                 struct object *obj;
619
620                 if (!bitmap_get(objects, bitmap_git->pack->num_objects + i))
621                         continue;
622
623                 obj = eindex->objects[i];
624                 if ((obj->type == OBJ_BLOB && !revs->blob_objects) ||
625                     (obj->type == OBJ_TREE && !revs->tree_objects) ||
626                     (obj->type == OBJ_TAG && !revs->tag_objects))
627                         continue;
628
629                 show_reach(&obj->oid, obj->type, 0, eindex->hashes[i], NULL, 0);
630         }
631 }
632
633 static void init_type_iterator(struct ewah_iterator *it,
634                                struct bitmap_index *bitmap_git,
635                                enum object_type type)
636 {
637         switch (type) {
638         case OBJ_COMMIT:
639                 ewah_iterator_init(it, bitmap_git->commits);
640                 break;
641
642         case OBJ_TREE:
643                 ewah_iterator_init(it, bitmap_git->trees);
644                 break;
645
646         case OBJ_BLOB:
647                 ewah_iterator_init(it, bitmap_git->blobs);
648                 break;
649
650         case OBJ_TAG:
651                 ewah_iterator_init(it, bitmap_git->tags);
652                 break;
653
654         default:
655                 BUG("object type %d not stored by bitmap type index", type);
656                 break;
657         }
658 }
659
660 static void show_objects_for_type(
661         struct bitmap_index *bitmap_git,
662         enum object_type object_type,
663         show_reachable_fn show_reach)
664 {
665         size_t i = 0;
666         uint32_t offset;
667
668         struct ewah_iterator it;
669         eword_t filter;
670
671         struct bitmap *objects = bitmap_git->result;
672
673         init_type_iterator(&it, bitmap_git, object_type);
674
675         for (i = 0; i < objects->word_alloc &&
676                         ewah_iterator_next(&filter, &it); i++) {
677                 eword_t word = objects->words[i] & filter;
678                 size_t pos = (i * BITS_IN_EWORD);
679
680                 if (!word)
681                         continue;
682
683                 for (offset = 0; offset < BITS_IN_EWORD; ++offset) {
684                         struct object_id oid;
685                         struct revindex_entry *entry;
686                         uint32_t hash = 0;
687
688                         if ((word >> offset) == 0)
689                                 break;
690
691                         offset += ewah_bit_ctz64(word >> offset);
692
693                         entry = &bitmap_git->pack->revindex[pos + offset];
694                         nth_packed_object_id(&oid, bitmap_git->pack, entry->nr);
695
696                         if (bitmap_git->hashes)
697                                 hash = get_be32(bitmap_git->hashes + entry->nr);
698
699                         show_reach(&oid, object_type, 0, hash, bitmap_git->pack, entry->offset);
700                 }
701         }
702 }
703
704 static int in_bitmapped_pack(struct bitmap_index *bitmap_git,
705                              struct object_list *roots)
706 {
707         while (roots) {
708                 struct object *object = roots->item;
709                 roots = roots->next;
710
711                 if (find_pack_entry_one(object->oid.hash, bitmap_git->pack) > 0)
712                         return 1;
713         }
714
715         return 0;
716 }
717
718 static struct bitmap *find_tip_objects(struct bitmap_index *bitmap_git,
719                                        struct object_list *tip_objects,
720                                        enum object_type type)
721 {
722         struct bitmap *result = bitmap_new();
723         struct object_list *p;
724
725         for (p = tip_objects; p; p = p->next) {
726                 int pos;
727
728                 if (p->item->type != type)
729                         continue;
730
731                 pos = bitmap_position(bitmap_git, &p->item->oid);
732                 if (pos < 0)
733                         continue;
734
735                 bitmap_set(result, pos);
736         }
737
738         return result;
739 }
740
741 static void filter_bitmap_exclude_type(struct bitmap_index *bitmap_git,
742                                        struct object_list *tip_objects,
743                                        struct bitmap *to_filter,
744                                        enum object_type type)
745 {
746         struct eindex *eindex = &bitmap_git->ext_index;
747         struct bitmap *tips;
748         struct ewah_iterator it;
749         eword_t mask;
750         uint32_t i;
751
752         if (type != OBJ_BLOB && type != OBJ_TREE)
753                 BUG("filter_bitmap_exclude_type: unsupported type '%d'", type);
754
755         /*
756          * The non-bitmap version of this filter never removes
757          * objects which the other side specifically asked for,
758          * so we must match that behavior.
759          */
760         tips = find_tip_objects(bitmap_git, tip_objects, type);
761
762         /*
763          * We can use the blob type-bitmap to work in whole words
764          * for the objects that are actually in the bitmapped packfile.
765          */
766         for (i = 0, init_type_iterator(&it, bitmap_git, type);
767              i < to_filter->word_alloc && ewah_iterator_next(&mask, &it);
768              i++) {
769                 if (i < tips->word_alloc)
770                         mask &= ~tips->words[i];
771                 to_filter->words[i] &= ~mask;
772         }
773
774         /*
775          * Clear any blobs that weren't in the packfile (and so would not have
776          * been caught by the loop above. We'll have to check them
777          * individually.
778          */
779         for (i = 0; i < eindex->count; i++) {
780                 uint32_t pos = i + bitmap_git->pack->num_objects;
781                 if (eindex->objects[i]->type == type &&
782                     bitmap_get(to_filter, pos) &&
783                     !bitmap_get(tips, pos))
784                         bitmap_unset(to_filter, pos);
785         }
786
787         bitmap_free(tips);
788 }
789
790 static void filter_bitmap_blob_none(struct bitmap_index *bitmap_git,
791                                     struct object_list *tip_objects,
792                                     struct bitmap *to_filter)
793 {
794         filter_bitmap_exclude_type(bitmap_git, tip_objects, to_filter,
795                                    OBJ_BLOB);
796 }
797
798 static unsigned long get_size_by_pos(struct bitmap_index *bitmap_git,
799                                      uint32_t pos)
800 {
801         struct packed_git *pack = bitmap_git->pack;
802         unsigned long size;
803         struct object_info oi = OBJECT_INFO_INIT;
804
805         oi.sizep = &size;
806
807         if (pos < pack->num_objects) {
808                 struct revindex_entry *entry = &pack->revindex[pos];
809                 if (packed_object_info(the_repository, pack,
810                                        entry->offset, &oi) < 0) {
811                         struct object_id oid;
812                         nth_packed_object_id(&oid, pack, entry->nr);
813                         die(_("unable to get size of %s"), oid_to_hex(&oid));
814                 }
815         } else {
816                 struct eindex *eindex = &bitmap_git->ext_index;
817                 struct object *obj = eindex->objects[pos - pack->num_objects];
818                 if (oid_object_info_extended(the_repository, &obj->oid, &oi, 0) < 0)
819                         die(_("unable to get size of %s"), oid_to_hex(&obj->oid));
820         }
821
822         return size;
823 }
824
825 static void filter_bitmap_blob_limit(struct bitmap_index *bitmap_git,
826                                      struct object_list *tip_objects,
827                                      struct bitmap *to_filter,
828                                      unsigned long limit)
829 {
830         struct eindex *eindex = &bitmap_git->ext_index;
831         struct bitmap *tips;
832         struct ewah_iterator it;
833         eword_t mask;
834         uint32_t i;
835
836         tips = find_tip_objects(bitmap_git, tip_objects, OBJ_BLOB);
837
838         for (i = 0, init_type_iterator(&it, bitmap_git, OBJ_BLOB);
839              i < to_filter->word_alloc && ewah_iterator_next(&mask, &it);
840              i++) {
841                 eword_t word = to_filter->words[i] & mask;
842                 unsigned offset;
843
844                 for (offset = 0; offset < BITS_IN_EWORD; offset++) {
845                         uint32_t pos;
846
847                         if ((word >> offset) == 0)
848                                 break;
849                         offset += ewah_bit_ctz64(word >> offset);
850                         pos = i * BITS_IN_EWORD + offset;
851
852                         if (!bitmap_get(tips, pos) &&
853                             get_size_by_pos(bitmap_git, pos) >= limit)
854                                 bitmap_unset(to_filter, pos);
855                 }
856         }
857
858         for (i = 0; i < eindex->count; i++) {
859                 uint32_t pos = i + bitmap_git->pack->num_objects;
860                 if (eindex->objects[i]->type == OBJ_BLOB &&
861                     bitmap_get(to_filter, pos) &&
862                     !bitmap_get(tips, pos) &&
863                     get_size_by_pos(bitmap_git, pos) >= limit)
864                         bitmap_unset(to_filter, pos);
865         }
866
867         bitmap_free(tips);
868 }
869
870 static void filter_bitmap_tree_depth(struct bitmap_index *bitmap_git,
871                                      struct object_list *tip_objects,
872                                      struct bitmap *to_filter,
873                                      unsigned long limit)
874 {
875         if (limit)
876                 BUG("filter_bitmap_tree_depth given non-zero limit");
877
878         filter_bitmap_exclude_type(bitmap_git, tip_objects, to_filter,
879                                    OBJ_TREE);
880         filter_bitmap_exclude_type(bitmap_git, tip_objects, to_filter,
881                                    OBJ_BLOB);
882 }
883
884 static int filter_bitmap(struct bitmap_index *bitmap_git,
885                          struct object_list *tip_objects,
886                          struct bitmap *to_filter,
887                          struct list_objects_filter_options *filter)
888 {
889         if (!filter || filter->choice == LOFC_DISABLED)
890                 return 0;
891
892         if (filter->choice == LOFC_BLOB_NONE) {
893                 if (bitmap_git)
894                         filter_bitmap_blob_none(bitmap_git, tip_objects,
895                                                 to_filter);
896                 return 0;
897         }
898
899         if (filter->choice == LOFC_BLOB_LIMIT) {
900                 if (bitmap_git)
901                         filter_bitmap_blob_limit(bitmap_git, tip_objects,
902                                                  to_filter,
903                                                  filter->blob_limit_value);
904                 return 0;
905         }
906
907         if (filter->choice == LOFC_TREE_DEPTH &&
908             filter->tree_exclude_depth == 0) {
909                 if (bitmap_git)
910                         filter_bitmap_tree_depth(bitmap_git, tip_objects,
911                                                  to_filter,
912                                                  filter->tree_exclude_depth);
913                 return 0;
914         }
915
916         /* filter choice not handled */
917         return -1;
918 }
919
920 static int can_filter_bitmap(struct list_objects_filter_options *filter)
921 {
922         return !filter_bitmap(NULL, NULL, NULL, filter);
923 }
924
925 struct bitmap_index *prepare_bitmap_walk(struct rev_info *revs,
926                                          struct list_objects_filter_options *filter)
927 {
928         unsigned int i;
929
930         struct object_list *wants = NULL;
931         struct object_list *haves = NULL;
932
933         struct bitmap *wants_bitmap = NULL;
934         struct bitmap *haves_bitmap = NULL;
935
936         struct bitmap_index *bitmap_git;
937
938         /*
939          * We can't do pathspec limiting with bitmaps, because we don't know
940          * which commits are associated with which object changes (let alone
941          * even which objects are associated with which paths).
942          */
943         if (revs->prune)
944                 return NULL;
945
946         if (!can_filter_bitmap(filter))
947                 return NULL;
948
949         /* try to open a bitmapped pack, but don't parse it yet
950          * because we may not need to use it */
951         bitmap_git = xcalloc(1, sizeof(*bitmap_git));
952         if (open_pack_bitmap(revs->repo, bitmap_git) < 0)
953                 goto cleanup;
954
955         for (i = 0; i < revs->pending.nr; ++i) {
956                 struct object *object = revs->pending.objects[i].item;
957
958                 if (object->type == OBJ_NONE)
959                         parse_object_or_die(&object->oid, NULL);
960
961                 while (object->type == OBJ_TAG) {
962                         struct tag *tag = (struct tag *) object;
963
964                         if (object->flags & UNINTERESTING)
965                                 object_list_insert(object, &haves);
966                         else
967                                 object_list_insert(object, &wants);
968
969                         object = parse_object_or_die(get_tagged_oid(tag), NULL);
970                 }
971
972                 if (object->flags & UNINTERESTING)
973                         object_list_insert(object, &haves);
974                 else
975                         object_list_insert(object, &wants);
976         }
977
978         /*
979          * if we have a HAVES list, but none of those haves is contained
980          * in the packfile that has a bitmap, we don't have anything to
981          * optimize here
982          */
983         if (haves && !in_bitmapped_pack(bitmap_git, haves))
984                 goto cleanup;
985
986         /* if we don't want anything, we're done here */
987         if (!wants)
988                 goto cleanup;
989
990         /*
991          * now we're going to use bitmaps, so load the actual bitmap entries
992          * from disk. this is the point of no return; after this the rev_list
993          * becomes invalidated and we must perform the revwalk through bitmaps
994          */
995         if (load_pack_bitmap(bitmap_git) < 0)
996                 goto cleanup;
997
998         object_array_clear(&revs->pending);
999
1000         if (haves) {
1001                 revs->ignore_missing_links = 1;
1002                 haves_bitmap = find_objects(bitmap_git, revs, haves, NULL);
1003                 reset_revision_walk();
1004                 revs->ignore_missing_links = 0;
1005
1006                 if (haves_bitmap == NULL)
1007                         BUG("failed to perform bitmap walk");
1008         }
1009
1010         wants_bitmap = find_objects(bitmap_git, revs, wants, haves_bitmap);
1011
1012         if (!wants_bitmap)
1013                 BUG("failed to perform bitmap walk");
1014
1015         if (haves_bitmap)
1016                 bitmap_and_not(wants_bitmap, haves_bitmap);
1017
1018         filter_bitmap(bitmap_git, wants, wants_bitmap, filter);
1019
1020         bitmap_git->result = wants_bitmap;
1021         bitmap_git->haves = haves_bitmap;
1022
1023         object_list_free(&wants);
1024         object_list_free(&haves);
1025
1026         return bitmap_git;
1027
1028 cleanup:
1029         free_bitmap_index(bitmap_git);
1030         object_list_free(&wants);
1031         object_list_free(&haves);
1032         return NULL;
1033 }
1034
1035 static void try_partial_reuse(struct bitmap_index *bitmap_git,
1036                               size_t pos,
1037                               struct bitmap *reuse,
1038                               struct pack_window **w_curs)
1039 {
1040         struct revindex_entry *revidx;
1041         off_t offset;
1042         enum object_type type;
1043         unsigned long size;
1044
1045         if (pos >= bitmap_git->pack->num_objects)
1046                 return; /* not actually in the pack */
1047
1048         revidx = &bitmap_git->pack->revindex[pos];
1049         offset = revidx->offset;
1050         type = unpack_object_header(bitmap_git->pack, w_curs, &offset, &size);
1051         if (type < 0)
1052                 return; /* broken packfile, punt */
1053
1054         if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA) {
1055                 off_t base_offset;
1056                 int base_pos;
1057
1058                 /*
1059                  * Find the position of the base object so we can look it up
1060                  * in our bitmaps. If we can't come up with an offset, or if
1061                  * that offset is not in the revidx, the pack is corrupt.
1062                  * There's nothing we can do, so just punt on this object,
1063                  * and the normal slow path will complain about it in
1064                  * more detail.
1065                  */
1066                 base_offset = get_delta_base(bitmap_git->pack, w_curs,
1067                                              &offset, type, revidx->offset);
1068                 if (!base_offset)
1069                         return;
1070                 base_pos = find_revindex_position(bitmap_git->pack, base_offset);
1071                 if (base_pos < 0)
1072                         return;
1073
1074                 /*
1075                  * We assume delta dependencies always point backwards. This
1076                  * lets us do a single pass, and is basically always true
1077                  * due to the way OFS_DELTAs work. You would not typically
1078                  * find REF_DELTA in a bitmapped pack, since we only bitmap
1079                  * packs we write fresh, and OFS_DELTA is the default). But
1080                  * let's double check to make sure the pack wasn't written with
1081                  * odd parameters.
1082                  */
1083                 if (base_pos >= pos)
1084                         return;
1085
1086                 /*
1087                  * And finally, if we're not sending the base as part of our
1088                  * reuse chunk, then don't send this object either. The base
1089                  * would come after us, along with other objects not
1090                  * necessarily in the pack, which means we'd need to convert
1091                  * to REF_DELTA on the fly. Better to just let the normal
1092                  * object_entry code path handle it.
1093                  */
1094                 if (!bitmap_get(reuse, base_pos))
1095                         return;
1096         }
1097
1098         /*
1099          * If we got here, then the object is OK to reuse. Mark it.
1100          */
1101         bitmap_set(reuse, pos);
1102 }
1103
1104 int reuse_partial_packfile_from_bitmap(struct bitmap_index *bitmap_git,
1105                                        struct packed_git **packfile_out,
1106                                        uint32_t *entries,
1107                                        struct bitmap **reuse_out)
1108 {
1109         struct bitmap *result = bitmap_git->result;
1110         struct bitmap *reuse;
1111         struct pack_window *w_curs = NULL;
1112         size_t i = 0;
1113         uint32_t offset;
1114
1115         assert(result);
1116
1117         while (i < result->word_alloc && result->words[i] == (eword_t)~0)
1118                 i++;
1119
1120         /* Don't mark objects not in the packfile */
1121         if (i > bitmap_git->pack->num_objects / BITS_IN_EWORD)
1122                 i = bitmap_git->pack->num_objects / BITS_IN_EWORD;
1123
1124         reuse = bitmap_word_alloc(i);
1125         memset(reuse->words, 0xFF, i * sizeof(eword_t));
1126
1127         for (; i < result->word_alloc; ++i) {
1128                 eword_t word = result->words[i];
1129                 size_t pos = (i * BITS_IN_EWORD);
1130
1131                 for (offset = 0; offset < BITS_IN_EWORD; ++offset) {
1132                         if ((word >> offset) == 0)
1133                                 break;
1134
1135                         offset += ewah_bit_ctz64(word >> offset);
1136                         try_partial_reuse(bitmap_git, pos + offset, reuse, &w_curs);
1137                 }
1138         }
1139
1140         unuse_pack(&w_curs);
1141
1142         *entries = bitmap_popcount(reuse);
1143         if (!*entries) {
1144                 bitmap_free(reuse);
1145                 return -1;
1146         }
1147
1148         /*
1149          * Drop any reused objects from the result, since they will not
1150          * need to be handled separately.
1151          */
1152         bitmap_and_not(result, reuse);
1153         *packfile_out = bitmap_git->pack;
1154         *reuse_out = reuse;
1155         return 0;
1156 }
1157
1158 int bitmap_walk_contains(struct bitmap_index *bitmap_git,
1159                          struct bitmap *bitmap, const struct object_id *oid)
1160 {
1161         int idx;
1162
1163         if (!bitmap)
1164                 return 0;
1165
1166         idx = bitmap_position(bitmap_git, oid);
1167         return idx >= 0 && bitmap_get(bitmap, idx);
1168 }
1169
1170 void traverse_bitmap_commit_list(struct bitmap_index *bitmap_git,
1171                                  struct rev_info *revs,
1172                                  show_reachable_fn show_reachable)
1173 {
1174         assert(bitmap_git->result);
1175
1176         show_objects_for_type(bitmap_git, OBJ_COMMIT, show_reachable);
1177         if (revs->tree_objects)
1178                 show_objects_for_type(bitmap_git, OBJ_TREE, show_reachable);
1179         if (revs->blob_objects)
1180                 show_objects_for_type(bitmap_git, OBJ_BLOB, show_reachable);
1181         if (revs->tag_objects)
1182                 show_objects_for_type(bitmap_git, OBJ_TAG, show_reachable);
1183
1184         show_extended_objects(bitmap_git, revs, show_reachable);
1185 }
1186
1187 static uint32_t count_object_type(struct bitmap_index *bitmap_git,
1188                                   enum object_type type)
1189 {
1190         struct bitmap *objects = bitmap_git->result;
1191         struct eindex *eindex = &bitmap_git->ext_index;
1192
1193         uint32_t i = 0, count = 0;
1194         struct ewah_iterator it;
1195         eword_t filter;
1196
1197         init_type_iterator(&it, bitmap_git, type);
1198
1199         while (i < objects->word_alloc && ewah_iterator_next(&filter, &it)) {
1200                 eword_t word = objects->words[i++] & filter;
1201                 count += ewah_bit_popcount64(word);
1202         }
1203
1204         for (i = 0; i < eindex->count; ++i) {
1205                 if (eindex->objects[i]->type == type &&
1206                         bitmap_get(objects, bitmap_git->pack->num_objects + i))
1207                         count++;
1208         }
1209
1210         return count;
1211 }
1212
1213 void count_bitmap_commit_list(struct bitmap_index *bitmap_git,
1214                               uint32_t *commits, uint32_t *trees,
1215                               uint32_t *blobs, uint32_t *tags)
1216 {
1217         assert(bitmap_git->result);
1218
1219         if (commits)
1220                 *commits = count_object_type(bitmap_git, OBJ_COMMIT);
1221
1222         if (trees)
1223                 *trees = count_object_type(bitmap_git, OBJ_TREE);
1224
1225         if (blobs)
1226                 *blobs = count_object_type(bitmap_git, OBJ_BLOB);
1227
1228         if (tags)
1229                 *tags = count_object_type(bitmap_git, OBJ_TAG);
1230 }
1231
1232 struct bitmap_test_data {
1233         struct bitmap_index *bitmap_git;
1234         struct bitmap *base;
1235         struct progress *prg;
1236         size_t seen;
1237 };
1238
1239 static void test_show_object(struct object *object, const char *name,
1240                              void *data)
1241 {
1242         struct bitmap_test_data *tdata = data;
1243         int bitmap_pos;
1244
1245         bitmap_pos = bitmap_position(tdata->bitmap_git, &object->oid);
1246         if (bitmap_pos < 0)
1247                 die("Object not in bitmap: %s\n", oid_to_hex(&object->oid));
1248
1249         bitmap_set(tdata->base, bitmap_pos);
1250         display_progress(tdata->prg, ++tdata->seen);
1251 }
1252
1253 static void test_show_commit(struct commit *commit, void *data)
1254 {
1255         struct bitmap_test_data *tdata = data;
1256         int bitmap_pos;
1257
1258         bitmap_pos = bitmap_position(tdata->bitmap_git,
1259                                      &commit->object.oid);
1260         if (bitmap_pos < 0)
1261                 die("Object not in bitmap: %s\n", oid_to_hex(&commit->object.oid));
1262
1263         bitmap_set(tdata->base, bitmap_pos);
1264         display_progress(tdata->prg, ++tdata->seen);
1265 }
1266
1267 void test_bitmap_walk(struct rev_info *revs)
1268 {
1269         struct object *root;
1270         struct bitmap *result = NULL;
1271         khiter_t pos;
1272         size_t result_popcnt;
1273         struct bitmap_test_data tdata;
1274         struct bitmap_index *bitmap_git;
1275
1276         if (!(bitmap_git = prepare_bitmap_git(revs->repo)))
1277                 die("failed to load bitmap indexes");
1278
1279         if (revs->pending.nr != 1)
1280                 die("you must specify exactly one commit to test");
1281
1282         fprintf(stderr, "Bitmap v%d test (%d entries loaded)\n",
1283                 bitmap_git->version, bitmap_git->entry_count);
1284
1285         root = revs->pending.objects[0].item;
1286         pos = kh_get_oid_map(bitmap_git->bitmaps, root->oid);
1287
1288         if (pos < kh_end(bitmap_git->bitmaps)) {
1289                 struct stored_bitmap *st = kh_value(bitmap_git->bitmaps, pos);
1290                 struct ewah_bitmap *bm = lookup_stored_bitmap(st);
1291
1292                 fprintf(stderr, "Found bitmap for %s. %d bits / %08x checksum\n",
1293                         oid_to_hex(&root->oid), (int)bm->bit_size, ewah_checksum(bm));
1294
1295                 result = ewah_to_bitmap(bm);
1296         }
1297
1298         if (result == NULL)
1299                 die("Commit %s doesn't have an indexed bitmap", oid_to_hex(&root->oid));
1300
1301         revs->tag_objects = 1;
1302         revs->tree_objects = 1;
1303         revs->blob_objects = 1;
1304
1305         result_popcnt = bitmap_popcount(result);
1306
1307         if (prepare_revision_walk(revs))
1308                 die("revision walk setup failed");
1309
1310         tdata.bitmap_git = bitmap_git;
1311         tdata.base = bitmap_new();
1312         tdata.prg = start_progress("Verifying bitmap entries", result_popcnt);
1313         tdata.seen = 0;
1314
1315         traverse_commit_list(revs, &test_show_commit, &test_show_object, &tdata);
1316
1317         stop_progress(&tdata.prg);
1318
1319         if (bitmap_equals(result, tdata.base))
1320                 fprintf(stderr, "OK!\n");
1321         else
1322                 fprintf(stderr, "Mismatch!\n");
1323
1324         free_bitmap_index(bitmap_git);
1325 }
1326
1327 static int rebuild_bitmap(uint32_t *reposition,
1328                           struct ewah_bitmap *source,
1329                           struct bitmap *dest)
1330 {
1331         uint32_t pos = 0;
1332         struct ewah_iterator it;
1333         eword_t word;
1334
1335         ewah_iterator_init(&it, source);
1336
1337         while (ewah_iterator_next(&word, &it)) {
1338                 uint32_t offset, bit_pos;
1339
1340                 for (offset = 0; offset < BITS_IN_EWORD; ++offset) {
1341                         if ((word >> offset) == 0)
1342                                 break;
1343
1344                         offset += ewah_bit_ctz64(word >> offset);
1345
1346                         bit_pos = reposition[pos + offset];
1347                         if (bit_pos > 0)
1348                                 bitmap_set(dest, bit_pos - 1);
1349                         else /* can't reuse, we don't have the object */
1350                                 return -1;
1351                 }
1352
1353                 pos += BITS_IN_EWORD;
1354         }
1355         return 0;
1356 }
1357
1358 int rebuild_existing_bitmaps(struct bitmap_index *bitmap_git,
1359                              struct packing_data *mapping,
1360                              kh_oid_map_t *reused_bitmaps,
1361                              int show_progress)
1362 {
1363         uint32_t i, num_objects;
1364         uint32_t *reposition;
1365         struct bitmap *rebuild;
1366         struct stored_bitmap *stored;
1367         struct progress *progress = NULL;
1368
1369         khiter_t hash_pos;
1370         int hash_ret;
1371
1372         num_objects = bitmap_git->pack->num_objects;
1373         reposition = xcalloc(num_objects, sizeof(uint32_t));
1374
1375         for (i = 0; i < num_objects; ++i) {
1376                 struct object_id oid;
1377                 struct revindex_entry *entry;
1378                 struct object_entry *oe;
1379
1380                 entry = &bitmap_git->pack->revindex[i];
1381                 nth_packed_object_id(&oid, bitmap_git->pack, entry->nr);
1382                 oe = packlist_find(mapping, &oid);
1383
1384                 if (oe)
1385                         reposition[i] = oe_in_pack_pos(mapping, oe) + 1;
1386         }
1387
1388         rebuild = bitmap_new();
1389         i = 0;
1390
1391         if (show_progress)
1392                 progress = start_progress("Reusing bitmaps", 0);
1393
1394         kh_foreach_value(bitmap_git->bitmaps, stored, {
1395                 if (stored->flags & BITMAP_FLAG_REUSE) {
1396                         if (!rebuild_bitmap(reposition,
1397                                             lookup_stored_bitmap(stored),
1398                                             rebuild)) {
1399                                 hash_pos = kh_put_oid_map(reused_bitmaps,
1400                                                           stored->oid,
1401                                                           &hash_ret);
1402                                 kh_value(reused_bitmaps, hash_pos) =
1403                                         bitmap_to_ewah(rebuild);
1404                         }
1405                         bitmap_reset(rebuild);
1406                         display_progress(progress, ++i);
1407                 }
1408         });
1409
1410         stop_progress(&progress);
1411
1412         free(reposition);
1413         bitmap_free(rebuild);
1414         return 0;
1415 }
1416
1417 void free_bitmap_index(struct bitmap_index *b)
1418 {
1419         if (!b)
1420                 return;
1421
1422         if (b->map)
1423                 munmap(b->map, b->map_size);
1424         ewah_pool_free(b->commits);
1425         ewah_pool_free(b->trees);
1426         ewah_pool_free(b->blobs);
1427         ewah_pool_free(b->tags);
1428         kh_destroy_oid_map(b->bitmaps);
1429         free(b->ext_index.objects);
1430         free(b->ext_index.hashes);
1431         bitmap_free(b->result);
1432         bitmap_free(b->haves);
1433         free(b);
1434 }
1435
1436 int bitmap_has_oid_in_uninteresting(struct bitmap_index *bitmap_git,
1437                                     const struct object_id *oid)
1438 {
1439         return bitmap_git &&
1440                 bitmap_walk_contains(bitmap_git, bitmap_git->haves, oid);
1441 }