1 #include "git-compat-util.h"
3 #include "environment.h"
6 #include "object-store.h"
11 #include "list-objects.h"
13 #include "pack-revindex.h"
15 #include "pack-bitmap.h"
16 #include "hash-lookup.h"
17 #include "pack-objects.h"
18 #include "commit-reach.h"
19 #include "prio-queue.h"
22 #include "tree-walk.h"
24 struct bitmapped_commit
{
25 struct commit
*commit
;
26 struct ewah_bitmap
*bitmap
;
27 struct ewah_bitmap
*write_as
;
33 struct bitmap_writer
{
34 struct ewah_bitmap
*commits
;
35 struct ewah_bitmap
*trees
;
36 struct ewah_bitmap
*blobs
;
37 struct ewah_bitmap
*tags
;
39 kh_oid_map_t
*bitmaps
;
40 struct packing_data
*to_pack
;
42 struct bitmapped_commit
*selected
;
43 unsigned int selected_nr
, selected_alloc
;
45 struct progress
*progress
;
47 unsigned char pack_checksum
[GIT_MAX_RAWSZ
];
50 static struct bitmap_writer writer
;
52 void bitmap_writer_show_progress(int show
)
54 writer
.show_progress
= show
;
58 * Build the initial type index for the packfile or multi-pack-index
60 void bitmap_writer_build_type_index(struct packing_data
*to_pack
,
61 struct pack_idx_entry
**index
,
66 writer
.commits
= ewah_new();
67 writer
.trees
= ewah_new();
68 writer
.blobs
= ewah_new();
69 writer
.tags
= ewah_new();
70 ALLOC_ARRAY(to_pack
->in_pack_pos
, to_pack
->nr_objects
);
72 for (i
= 0; i
< index_nr
; ++i
) {
73 struct object_entry
*entry
= (struct object_entry
*)index
[i
];
74 enum object_type real_type
;
76 oe_set_in_pack_pos(to_pack
, entry
, i
);
78 switch (oe_type(entry
)) {
83 real_type
= oe_type(entry
);
87 real_type
= oid_object_info(to_pack
->repo
,
88 &entry
->idx
.oid
, NULL
);
94 ewah_set(writer
.commits
, i
);
98 ewah_set(writer
.trees
, i
);
102 ewah_set(writer
.blobs
, i
);
106 ewah_set(writer
.tags
, i
);
110 die("Missing type information for %s (%d/%d)",
111 oid_to_hex(&entry
->idx
.oid
), real_type
,
118 * Compute the actual bitmaps
121 static inline void push_bitmapped_commit(struct commit
*commit
)
123 if (writer
.selected_nr
>= writer
.selected_alloc
) {
124 writer
.selected_alloc
= (writer
.selected_alloc
+ 32) * 2;
125 REALLOC_ARRAY(writer
.selected
, writer
.selected_alloc
);
128 writer
.selected
[writer
.selected_nr
].commit
= commit
;
129 writer
.selected
[writer
.selected_nr
].bitmap
= NULL
;
130 writer
.selected
[writer
.selected_nr
].flags
= 0;
132 writer
.selected_nr
++;
135 static uint32_t find_object_pos(const struct object_id
*oid
, int *found
)
137 struct object_entry
*entry
= packlist_find(writer
.to_pack
, oid
);
142 warning("Failed to write bitmap index. Packfile doesn't have full closure "
143 "(object %s is missing)", oid_to_hex(oid
));
149 return oe_in_pack_pos(writer
.to_pack
, entry
);
152 static void compute_xor_offsets(void)
154 static const int MAX_XOR_OFFSET_SEARCH
= 10;
158 while (next
< writer
.selected_nr
) {
159 struct bitmapped_commit
*stored
= &writer
.selected
[next
];
162 struct ewah_bitmap
*best_bitmap
= stored
->bitmap
;
163 struct ewah_bitmap
*test_xor
;
165 for (i
= 1; i
<= MAX_XOR_OFFSET_SEARCH
; ++i
) {
171 test_xor
= ewah_pool_new();
172 ewah_xor(writer
.selected
[curr
].bitmap
, stored
->bitmap
, test_xor
);
174 if (test_xor
->buffer_size
< best_bitmap
->buffer_size
) {
175 if (best_bitmap
!= stored
->bitmap
)
176 ewah_pool_free(best_bitmap
);
178 best_bitmap
= test_xor
;
181 ewah_pool_free(test_xor
);
185 stored
->xor_offset
= best_offset
;
186 stored
->write_as
= best_bitmap
;
193 struct commit_list
*reverse_edges
;
194 struct bitmap
*commit_mask
;
195 struct bitmap
*bitmap
;
198 unsigned idx
; /* within selected array */
201 define_commit_slab(bb_data
, struct bb_commit
);
203 struct bitmap_builder
{
205 struct commit
**commits
;
206 size_t commits_nr
, commits_alloc
;
209 static void bitmap_builder_init(struct bitmap_builder
*bb
,
210 struct bitmap_writer
*writer
,
211 struct bitmap_index
*old_bitmap
)
213 struct rev_info revs
;
214 struct commit
*commit
;
215 struct commit_list
*reusable
= NULL
;
216 struct commit_list
*r
;
217 unsigned int i
, num_maximal
= 0;
219 memset(bb
, 0, sizeof(*bb
));
220 init_bb_data(&bb
->data
);
222 reset_revision_walk();
223 repo_init_revisions(writer
->to_pack
->repo
, &revs
, NULL
);
225 revs
.first_parent_only
= 1;
227 for (i
= 0; i
< writer
->selected_nr
; i
++) {
228 struct commit
*c
= writer
->selected
[i
].commit
;
229 struct bb_commit
*ent
= bb_data_at(&bb
->data
, c
);
235 ent
->commit_mask
= bitmap_new();
236 bitmap_set(ent
->commit_mask
, i
);
238 add_pending_object(&revs
, &c
->object
, "");
241 if (prepare_revision_walk(&revs
))
242 die("revision walk setup failed");
244 while ((commit
= get_revision(&revs
))) {
245 struct commit_list
*p
= commit
->parents
;
246 struct bb_commit
*c_ent
;
248 parse_commit_or_die(commit
);
250 c_ent
= bb_data_at(&bb
->data
, commit
);
253 * If there is no commit_mask, there is no reason to iterate
254 * over this commit; it is not selected (if it were, it would
255 * not have a blank commit mask) and all its children have
256 * existing bitmaps (see the comment starting with "This commit
257 * has an existing bitmap" below), so it does not contribute
258 * anything to the final bitmap file or its descendants.
260 if (!c_ent
->commit_mask
)
263 if (old_bitmap
&& bitmap_for_commit(old_bitmap
, commit
)) {
265 * This commit has an existing bitmap, so we can
266 * get its bits immediately without an object
267 * walk. That is, it is reusable as-is and there is no
268 * need to continue walking beyond it.
270 * Mark it as such and add it to bb->commits separately
271 * to avoid allocating a position in the commit mask.
273 commit_list_insert(commit
, &reusable
);
277 if (c_ent
->maximal
) {
279 ALLOC_GROW(bb
->commits
, bb
->commits_nr
+ 1, bb
->commits_alloc
);
280 bb
->commits
[bb
->commits_nr
++] = commit
;
284 struct bb_commit
*p_ent
= bb_data_at(&bb
->data
, p
->item
);
285 int c_not_p
, p_not_c
;
287 if (!p_ent
->commit_mask
) {
288 p_ent
->commit_mask
= bitmap_new();
292 c_not_p
= bitmap_is_subset(c_ent
->commit_mask
, p_ent
->commit_mask
);
293 p_not_c
= bitmap_is_subset(p_ent
->commit_mask
, c_ent
->commit_mask
);
299 bitmap_or(p_ent
->commit_mask
, c_ent
->commit_mask
);
305 free_commit_list(p_ent
->reverse_edges
);
306 p_ent
->reverse_edges
= NULL
;
309 if (c_ent
->maximal
) {
310 commit_list_insert(commit
, &p_ent
->reverse_edges
);
312 struct commit_list
*cc
= c_ent
->reverse_edges
;
314 for (; cc
; cc
= cc
->next
) {
315 if (!commit_list_contains(cc
->item
, p_ent
->reverse_edges
))
316 commit_list_insert(cc
->item
, &p_ent
->reverse_edges
);
322 bitmap_free(c_ent
->commit_mask
);
323 c_ent
->commit_mask
= NULL
;
326 for (r
= reusable
; r
; r
= r
->next
) {
327 ALLOC_GROW(bb
->commits
, bb
->commits_nr
+ 1, bb
->commits_alloc
);
328 bb
->commits
[bb
->commits_nr
++] = r
->item
;
331 trace2_data_intmax("pack-bitmap-write", the_repository
,
332 "num_selected_commits", writer
->selected_nr
);
333 trace2_data_intmax("pack-bitmap-write", the_repository
,
334 "num_maximal_commits", num_maximal
);
336 release_revisions(&revs
);
337 free_commit_list(reusable
);
340 static void bitmap_builder_clear(struct bitmap_builder
*bb
)
342 clear_bb_data(&bb
->data
);
344 bb
->commits_nr
= bb
->commits_alloc
= 0;
347 static int fill_bitmap_tree(struct bitmap
*bitmap
,
352 struct tree_desc desc
;
353 struct name_entry entry
;
356 * If our bit is already set, then there is nothing to do. Both this
357 * tree and all of its children will be set.
359 pos
= find_object_pos(&tree
->object
.oid
, &found
);
362 if (bitmap_get(bitmap
, pos
))
364 bitmap_set(bitmap
, pos
);
366 if (parse_tree(tree
) < 0)
367 die("unable to load tree object %s",
368 oid_to_hex(&tree
->object
.oid
));
369 init_tree_desc(&desc
, tree
->buffer
, tree
->size
);
371 while (tree_entry(&desc
, &entry
)) {
372 switch (object_type(entry
.mode
)) {
374 if (fill_bitmap_tree(bitmap
,
375 lookup_tree(the_repository
, &entry
.oid
)) < 0)
379 pos
= find_object_pos(&entry
.oid
, &found
);
382 bitmap_set(bitmap
, pos
);
385 /* Gitlink, etc; not reachable */
390 free_tree_buffer(tree
);
394 static int reused_bitmaps_nr
;
396 static int fill_bitmap_commit(struct bb_commit
*ent
,
397 struct commit
*commit
,
398 struct prio_queue
*queue
,
399 struct prio_queue
*tree_queue
,
400 struct bitmap_index
*old_bitmap
,
401 const uint32_t *mapping
)
406 ent
->bitmap
= bitmap_new();
408 prio_queue_put(queue
, commit
);
411 struct commit_list
*p
;
412 struct commit
*c
= prio_queue_get(queue
);
414 if (old_bitmap
&& mapping
) {
415 struct ewah_bitmap
*old
= bitmap_for_commit(old_bitmap
, c
);
417 * If this commit has an old bitmap, then translate that
418 * bitmap and add its bits to this one. No need to walk
419 * parents or the tree for this commit.
421 if (old
&& !rebuild_bitmap(mapping
, old
, ent
->bitmap
)) {
428 * Mark ourselves and queue our tree. The commit
429 * walk ensures we cover all parents.
431 pos
= find_object_pos(&c
->object
.oid
, &found
);
434 bitmap_set(ent
->bitmap
, pos
);
435 prio_queue_put(tree_queue
,
436 repo_get_commit_tree(the_repository
, c
));
438 for (p
= c
->parents
; p
; p
= p
->next
) {
439 pos
= find_object_pos(&p
->item
->object
.oid
, &found
);
442 if (!bitmap_get(ent
->bitmap
, pos
)) {
443 bitmap_set(ent
->bitmap
, pos
);
444 prio_queue_put(queue
, p
->item
);
449 while (tree_queue
->nr
) {
450 if (fill_bitmap_tree(ent
->bitmap
,
451 prio_queue_get(tree_queue
)) < 0)
457 static void store_selected(struct bb_commit
*ent
, struct commit
*commit
)
459 struct bitmapped_commit
*stored
= &writer
.selected
[ent
->idx
];
463 stored
->bitmap
= bitmap_to_ewah(ent
->bitmap
);
465 hash_pos
= kh_put_oid_map(writer
.bitmaps
, commit
->object
.oid
, &hash_ret
);
467 die("Duplicate entry when writing index: %s",
468 oid_to_hex(&commit
->object
.oid
));
469 kh_value(writer
.bitmaps
, hash_pos
) = stored
;
472 int bitmap_writer_build(struct packing_data
*to_pack
)
474 struct bitmap_builder bb
;
476 int nr_stored
= 0; /* for progress */
477 struct prio_queue queue
= { compare_commits_by_gen_then_commit_date
};
478 struct prio_queue tree_queue
= { NULL
};
479 struct bitmap_index
*old_bitmap
;
481 int closed
= 1; /* until proven otherwise */
483 writer
.bitmaps
= kh_init_oid_map();
484 writer
.to_pack
= to_pack
;
486 if (writer
.show_progress
)
487 writer
.progress
= start_progress("Building bitmaps", writer
.selected_nr
);
488 trace2_region_enter("pack-bitmap-write", "building_bitmaps_total",
491 old_bitmap
= prepare_bitmap_git(to_pack
->repo
);
493 mapping
= create_bitmap_mapping(old_bitmap
, to_pack
);
497 bitmap_builder_init(&bb
, &writer
, old_bitmap
);
498 for (i
= bb
.commits_nr
; i
> 0; i
--) {
499 struct commit
*commit
= bb
.commits
[i
-1];
500 struct bb_commit
*ent
= bb_data_at(&bb
.data
, commit
);
501 struct commit
*child
;
504 if (fill_bitmap_commit(ent
, commit
, &queue
, &tree_queue
,
505 old_bitmap
, mapping
) < 0) {
511 store_selected(ent
, commit
);
513 display_progress(writer
.progress
, nr_stored
);
516 while ((child
= pop_commit(&ent
->reverse_edges
))) {
517 struct bb_commit
*child_ent
=
518 bb_data_at(&bb
.data
, child
);
520 if (child_ent
->bitmap
)
521 bitmap_or(child_ent
->bitmap
, ent
->bitmap
);
523 child_ent
->bitmap
= bitmap_dup(ent
->bitmap
);
525 child_ent
->bitmap
= ent
->bitmap
;
530 bitmap_free(ent
->bitmap
);
533 clear_prio_queue(&queue
);
534 clear_prio_queue(&tree_queue
);
535 bitmap_builder_clear(&bb
);
536 free_bitmap_index(old_bitmap
);
539 trace2_region_leave("pack-bitmap-write", "building_bitmaps_total",
541 trace2_data_intmax("pack-bitmap-write", the_repository
,
542 "building_bitmaps_reused", reused_bitmaps_nr
);
544 stop_progress(&writer
.progress
);
547 compute_xor_offsets();
548 return closed
? 0 : -1;
552 * Select the commits that will be bitmapped
554 static inline unsigned int next_commit_index(unsigned int idx
)
556 static const unsigned int MIN_COMMITS
= 100;
557 static const unsigned int MAX_COMMITS
= 5000;
559 static const unsigned int MUST_REGION
= 100;
560 static const unsigned int MIN_REGION
= 20000;
562 unsigned int offset
, next
;
564 if (idx
<= MUST_REGION
)
567 if (idx
<= MIN_REGION
) {
568 offset
= idx
- MUST_REGION
;
569 return (offset
< MIN_COMMITS
) ? offset
: MIN_COMMITS
;
572 offset
= idx
- MIN_REGION
;
573 next
= (offset
< MAX_COMMITS
) ? offset
: MAX_COMMITS
;
575 return (next
> MIN_COMMITS
) ? next
: MIN_COMMITS
;
578 static int date_compare(const void *_a
, const void *_b
)
580 struct commit
*a
= *(struct commit
**)_a
;
581 struct commit
*b
= *(struct commit
**)_b
;
582 return (long)b
->date
- (long)a
->date
;
585 void bitmap_writer_select_commits(struct commit
**indexed_commits
,
586 unsigned int indexed_commits_nr
,
589 unsigned int i
= 0, j
, next
;
591 QSORT(indexed_commits
, indexed_commits_nr
, date_compare
);
593 if (indexed_commits_nr
< 100) {
594 for (i
= 0; i
< indexed_commits_nr
; ++i
)
595 push_bitmapped_commit(indexed_commits
[i
]);
599 if (writer
.show_progress
)
600 writer
.progress
= start_progress("Selecting bitmap commits", 0);
603 struct commit
*chosen
= NULL
;
605 next
= next_commit_index(i
);
607 if (i
+ next
>= indexed_commits_nr
)
610 if (max_bitmaps
> 0 && writer
.selected_nr
>= max_bitmaps
) {
611 writer
.selected_nr
= max_bitmaps
;
616 chosen
= indexed_commits
[i
];
618 chosen
= indexed_commits
[i
+ next
];
620 for (j
= 0; j
<= next
; ++j
) {
621 struct commit
*cm
= indexed_commits
[i
+ j
];
623 if ((cm
->object
.flags
& NEEDS_BITMAP
) != 0) {
628 if (cm
->parents
&& cm
->parents
->next
)
633 push_bitmapped_commit(chosen
);
636 display_progress(writer
.progress
, i
);
639 stop_progress(&writer
.progress
);
643 static int hashwrite_ewah_helper(void *f
, const void *buf
, size_t len
)
645 /* hashwrite will die on error */
646 hashwrite(f
, buf
, len
);
651 * Write the bitmap index to disk
653 static inline void dump_bitmap(struct hashfile
*f
, struct ewah_bitmap
*bitmap
)
655 if (ewah_serialize_to(bitmap
, hashwrite_ewah_helper
, f
) < 0)
656 die("Failed to write bitmap index");
659 static const struct object_id
*oid_access(size_t pos
, const void *table
)
661 const struct pack_idx_entry
* const *index
= table
;
662 return &index
[pos
]->oid
;
665 static void write_selected_commits_v1(struct hashfile
*f
,
666 uint32_t *commit_positions
,
671 for (i
= 0; i
< writer
.selected_nr
; ++i
) {
672 struct bitmapped_commit
*stored
= &writer
.selected
[i
];
675 offsets
[i
] = hashfile_total(f
);
677 hashwrite_be32(f
, commit_positions
[i
]);
678 hashwrite_u8(f
, stored
->xor_offset
);
679 hashwrite_u8(f
, stored
->flags
);
681 dump_bitmap(f
, stored
->write_as
);
685 static int table_cmp(const void *_va
, const void *_vb
, void *_data
)
687 uint32_t *commit_positions
= _data
;
688 uint32_t a
= commit_positions
[*(uint32_t *)_va
];
689 uint32_t b
= commit_positions
[*(uint32_t *)_vb
];
699 static void write_lookup_table(struct hashfile
*f
,
700 uint32_t *commit_positions
,
704 uint32_t *table
, *table_inv
;
706 ALLOC_ARRAY(table
, writer
.selected_nr
);
707 ALLOC_ARRAY(table_inv
, writer
.selected_nr
);
709 for (i
= 0; i
< writer
.selected_nr
; i
++)
713 * At the end of this sort table[j] = i means that the i'th
714 * bitmap corresponds to j'th bitmapped commit (among the selected
715 * commits) in lex order of OIDs.
717 QSORT_S(table
, writer
.selected_nr
, table_cmp
, commit_positions
);
719 /* table_inv helps us discover that relationship (i'th bitmap
720 * to j'th commit by j = table_inv[i])
722 for (i
= 0; i
< writer
.selected_nr
; i
++)
723 table_inv
[table
[i
]] = i
;
725 trace2_region_enter("pack-bitmap-write", "writing_lookup_table", the_repository
);
726 for (i
= 0; i
< writer
.selected_nr
; i
++) {
727 struct bitmapped_commit
*selected
= &writer
.selected
[table
[i
]];
728 uint32_t xor_offset
= selected
->xor_offset
;
733 * xor_index stores the index (in the bitmap entries)
734 * of the corresponding xor bitmap. But we need to convert
735 * this index into lookup table's index. So, table_inv[xor_index]
736 * gives us the index position w.r.t. the lookup table.
738 * If "k = table[i] - xor_offset" then the xor base is the k'th
739 * bitmap. `table_inv[k]` gives us the position of that bitmap
740 * in the lookup table.
742 uint32_t xor_index
= table
[i
] - xor_offset
;
743 xor_row
= table_inv
[xor_index
];
745 xor_row
= 0xffffffff;
748 hashwrite_be32(f
, commit_positions
[table
[i
]]);
749 hashwrite_be64(f
, (uint64_t)offsets
[table
[i
]]);
750 hashwrite_be32(f
, xor_row
);
752 trace2_region_leave("pack-bitmap-write", "writing_lookup_table", the_repository
);
758 static void write_hash_cache(struct hashfile
*f
,
759 struct pack_idx_entry
**index
,
764 for (i
= 0; i
< index_nr
; ++i
) {
765 struct object_entry
*entry
= (struct object_entry
*)index
[i
];
766 hashwrite_be32(f
, entry
->hash
);
770 void bitmap_writer_set_checksum(const unsigned char *sha1
)
772 hashcpy(writer
.pack_checksum
, sha1
);
775 void bitmap_writer_finish(struct pack_idx_entry
**index
,
777 const char *filename
,
780 static uint16_t default_version
= 1;
781 static uint16_t flags
= BITMAP_OPT_FULL_DAG
;
782 struct strbuf tmp_file
= STRBUF_INIT
;
784 uint32_t *commit_positions
= NULL
;
785 off_t
*offsets
= NULL
;
788 struct bitmap_disk_header header
;
790 int fd
= odb_mkstemp(&tmp_file
, "pack/tmp_bitmap_XXXXXX");
792 f
= hashfd(fd
, tmp_file
.buf
);
794 memcpy(header
.magic
, BITMAP_IDX_SIGNATURE
, sizeof(BITMAP_IDX_SIGNATURE
));
795 header
.version
= htons(default_version
);
796 header
.options
= htons(flags
| options
);
797 header
.entry_count
= htonl(writer
.selected_nr
);
798 hashcpy(header
.checksum
, writer
.pack_checksum
);
800 hashwrite(f
, &header
, sizeof(header
) - GIT_MAX_RAWSZ
+ the_hash_algo
->rawsz
);
801 dump_bitmap(f
, writer
.commits
);
802 dump_bitmap(f
, writer
.trees
);
803 dump_bitmap(f
, writer
.blobs
);
804 dump_bitmap(f
, writer
.tags
);
806 if (options
& BITMAP_OPT_LOOKUP_TABLE
)
807 CALLOC_ARRAY(offsets
, index_nr
);
809 ALLOC_ARRAY(commit_positions
, writer
.selected_nr
);
811 for (i
= 0; i
< writer
.selected_nr
; i
++) {
812 struct bitmapped_commit
*stored
= &writer
.selected
[i
];
813 int commit_pos
= oid_pos(&stored
->commit
->object
.oid
, index
, index_nr
, oid_access
);
816 BUG(_("trying to write commit not in index"));
818 commit_positions
[i
] = commit_pos
;
821 write_selected_commits_v1(f
, commit_positions
, offsets
);
823 if (options
& BITMAP_OPT_LOOKUP_TABLE
)
824 write_lookup_table(f
, commit_positions
, offsets
);
826 if (options
& BITMAP_OPT_HASH_CACHE
)
827 write_hash_cache(f
, index
, index_nr
);
829 finalize_hashfile(f
, NULL
, FSYNC_COMPONENT_PACK_METADATA
,
830 CSUM_HASH_IN_STREAM
| CSUM_FSYNC
| CSUM_CLOSE
);
832 if (adjust_shared_perm(tmp_file
.buf
))
833 die_errno("unable to make temporary bitmap file readable");
835 if (rename(tmp_file
.buf
, filename
))
836 die_errno("unable to rename temporary bitmap file to '%s'", filename
);
838 strbuf_release(&tmp_file
);
839 free(commit_positions
);