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