pack-objects: implement bitmap writing
[git/git.git] / pack-bitmap-write.c
1 #include "cache.h"
2 #include "commit.h"
3 #include "tag.h"
4 #include "diff.h"
5 #include "revision.h"
6 #include "list-objects.h"
7 #include "progress.h"
8 #include "pack-revindex.h"
9 #include "pack.h"
10 #include "pack-bitmap.h"
11 #include "sha1-lookup.h"
12 #include "pack-objects.h"
13
14 struct bitmapped_commit {
15 struct commit *commit;
16 struct ewah_bitmap *bitmap;
17 struct ewah_bitmap *write_as;
18 int flags;
19 int xor_offset;
20 uint32_t commit_pos;
21 };
22
23 struct bitmap_writer {
24 struct ewah_bitmap *commits;
25 struct ewah_bitmap *trees;
26 struct ewah_bitmap *blobs;
27 struct ewah_bitmap *tags;
28
29 khash_sha1 *bitmaps;
30 khash_sha1 *reused;
31 struct packing_data *to_pack;
32
33 struct bitmapped_commit *selected;
34 unsigned int selected_nr, selected_alloc;
35
36 struct progress *progress;
37 int show_progress;
38 unsigned char pack_checksum[20];
39 };
40
41 static struct bitmap_writer writer;
42
43 void bitmap_writer_show_progress(int show)
44 {
45 writer.show_progress = show;
46 }
47
48 /**
49 * Build the initial type index for the packfile
50 */
51 void bitmap_writer_build_type_index(struct pack_idx_entry **index,
52 uint32_t index_nr)
53 {
54 uint32_t i;
55
56 writer.commits = ewah_new();
57 writer.trees = ewah_new();
58 writer.blobs = ewah_new();
59 writer.tags = ewah_new();
60
61 for (i = 0; i < index_nr; ++i) {
62 struct object_entry *entry = (struct object_entry *)index[i];
63 enum object_type real_type;
64
65 entry->in_pack_pos = i;
66
67 switch (entry->type) {
68 case OBJ_COMMIT:
69 case OBJ_TREE:
70 case OBJ_BLOB:
71 case OBJ_TAG:
72 real_type = entry->type;
73 break;
74
75 default:
76 real_type = sha1_object_info(entry->idx.sha1, NULL);
77 break;
78 }
79
80 switch (real_type) {
81 case OBJ_COMMIT:
82 ewah_set(writer.commits, i);
83 break;
84
85 case OBJ_TREE:
86 ewah_set(writer.trees, i);
87 break;
88
89 case OBJ_BLOB:
90 ewah_set(writer.blobs, i);
91 break;
92
93 case OBJ_TAG:
94 ewah_set(writer.tags, i);
95 break;
96
97 default:
98 die("Missing type information for %s (%d/%d)",
99 sha1_to_hex(entry->idx.sha1), real_type, entry->type);
100 }
101 }
102 }
103
104 /**
105 * Compute the actual bitmaps
106 */
107 static struct object **seen_objects;
108 static unsigned int seen_objects_nr, seen_objects_alloc;
109
110 static inline void push_bitmapped_commit(struct commit *commit, struct ewah_bitmap *reused)
111 {
112 if (writer.selected_nr >= writer.selected_alloc) {
113 writer.selected_alloc = (writer.selected_alloc + 32) * 2;
114 writer.selected = xrealloc(writer.selected,
115 writer.selected_alloc * sizeof(struct bitmapped_commit));
116 }
117
118 writer.selected[writer.selected_nr].commit = commit;
119 writer.selected[writer.selected_nr].bitmap = reused;
120 writer.selected[writer.selected_nr].flags = 0;
121
122 writer.selected_nr++;
123 }
124
125 static inline void mark_as_seen(struct object *object)
126 {
127 ALLOC_GROW(seen_objects, seen_objects_nr + 1, seen_objects_alloc);
128 seen_objects[seen_objects_nr++] = object;
129 }
130
131 static inline void reset_all_seen(void)
132 {
133 unsigned int i;
134 for (i = 0; i < seen_objects_nr; ++i) {
135 seen_objects[i]->flags &= ~(SEEN | ADDED | SHOWN);
136 }
137 seen_objects_nr = 0;
138 }
139
140 static uint32_t find_object_pos(const unsigned char *sha1)
141 {
142 struct object_entry *entry = packlist_find(writer.to_pack, sha1, NULL);
143
144 if (!entry) {
145 die("Failed to write bitmap index. Packfile doesn't have full closure "
146 "(object %s is missing)", sha1_to_hex(sha1));
147 }
148
149 return entry->in_pack_pos;
150 }
151
152 static void show_object(struct object *object, const struct name_path *path,
153 const char *last, void *data)
154 {
155 struct bitmap *base = data;
156 bitmap_set(base, find_object_pos(object->sha1));
157 mark_as_seen(object);
158 }
159
160 static void show_commit(struct commit *commit, void *data)
161 {
162 mark_as_seen((struct object *)commit);
163 }
164
165 static int
166 add_to_include_set(struct bitmap *base, struct commit *commit)
167 {
168 khiter_t hash_pos;
169 uint32_t bitmap_pos = find_object_pos(commit->object.sha1);
170
171 if (bitmap_get(base, bitmap_pos))
172 return 0;
173
174 hash_pos = kh_get_sha1(writer.bitmaps, commit->object.sha1);
175 if (hash_pos < kh_end(writer.bitmaps)) {
176 struct bitmapped_commit *bc = kh_value(writer.bitmaps, hash_pos);
177 bitmap_or_ewah(base, bc->bitmap);
178 return 0;
179 }
180
181 bitmap_set(base, bitmap_pos);
182 return 1;
183 }
184
185 static int
186 should_include(struct commit *commit, void *_data)
187 {
188 struct bitmap *base = _data;
189
190 if (!add_to_include_set(base, commit)) {
191 struct commit_list *parent = commit->parents;
192
193 mark_as_seen((struct object *)commit);
194
195 while (parent) {
196 parent->item->object.flags |= SEEN;
197 mark_as_seen((struct object *)parent->item);
198 parent = parent->next;
199 }
200
201 return 0;
202 }
203
204 return 1;
205 }
206
207 static void compute_xor_offsets(void)
208 {
209 static const int MAX_XOR_OFFSET_SEARCH = 10;
210
211 int i, next = 0;
212
213 while (next < writer.selected_nr) {
214 struct bitmapped_commit *stored = &writer.selected[next];
215
216 int best_offset = 0;
217 struct ewah_bitmap *best_bitmap = stored->bitmap;
218 struct ewah_bitmap *test_xor;
219
220 for (i = 1; i <= MAX_XOR_OFFSET_SEARCH; ++i) {
221 int curr = next - i;
222
223 if (curr < 0)
224 break;
225
226 test_xor = ewah_pool_new();
227 ewah_xor(writer.selected[curr].bitmap, stored->bitmap, test_xor);
228
229 if (test_xor->buffer_size < best_bitmap->buffer_size) {
230 if (best_bitmap != stored->bitmap)
231 ewah_pool_free(best_bitmap);
232
233 best_bitmap = test_xor;
234 best_offset = i;
235 } else {
236 ewah_pool_free(test_xor);
237 }
238 }
239
240 stored->xor_offset = best_offset;
241 stored->write_as = best_bitmap;
242
243 next++;
244 }
245 }
246
247 void bitmap_writer_build(struct packing_data *to_pack)
248 {
249 static const double REUSE_BITMAP_THRESHOLD = 0.2;
250
251 int i, reuse_after, need_reset;
252 struct bitmap *base = bitmap_new();
253 struct rev_info revs;
254
255 writer.bitmaps = kh_init_sha1();
256 writer.to_pack = to_pack;
257
258 if (writer.show_progress)
259 writer.progress = start_progress("Building bitmaps", writer.selected_nr);
260
261 init_revisions(&revs, NULL);
262 revs.tag_objects = 1;
263 revs.tree_objects = 1;
264 revs.blob_objects = 1;
265 revs.no_walk = 0;
266
267 revs.include_check = should_include;
268 reset_revision_walk();
269
270 reuse_after = writer.selected_nr * REUSE_BITMAP_THRESHOLD;
271 need_reset = 0;
272
273 for (i = writer.selected_nr - 1; i >= 0; --i) {
274 struct bitmapped_commit *stored;
275 struct object *object;
276
277 khiter_t hash_pos;
278 int hash_ret;
279
280 stored = &writer.selected[i];
281 object = (struct object *)stored->commit;
282
283 if (stored->bitmap == NULL) {
284 if (i < writer.selected_nr - 1 &&
285 (need_reset ||
286 !in_merge_bases(writer.selected[i + 1].commit,
287 stored->commit))) {
288 bitmap_reset(base);
289 reset_all_seen();
290 }
291
292 add_pending_object(&revs, object, "");
293 revs.include_check_data = base;
294
295 if (prepare_revision_walk(&revs))
296 die("revision walk setup failed");
297
298 traverse_commit_list(&revs, show_commit, show_object, base);
299
300 revs.pending.nr = 0;
301 revs.pending.alloc = 0;
302 revs.pending.objects = NULL;
303
304 stored->bitmap = bitmap_to_ewah(base);
305 need_reset = 0;
306 } else
307 need_reset = 1;
308
309 if (i >= reuse_after)
310 stored->flags |= BITMAP_FLAG_REUSE;
311
312 hash_pos = kh_put_sha1(writer.bitmaps, object->sha1, &hash_ret);
313 if (hash_ret == 0)
314 die("Duplicate entry when writing index: %s",
315 sha1_to_hex(object->sha1));
316
317 kh_value(writer.bitmaps, hash_pos) = stored;
318 display_progress(writer.progress, writer.selected_nr - i);
319 }
320
321 bitmap_free(base);
322 stop_progress(&writer.progress);
323
324 compute_xor_offsets();
325 }
326
327 /**
328 * Select the commits that will be bitmapped
329 */
330 static inline unsigned int next_commit_index(unsigned int idx)
331 {
332 static const unsigned int MIN_COMMITS = 100;
333 static const unsigned int MAX_COMMITS = 5000;
334
335 static const unsigned int MUST_REGION = 100;
336 static const unsigned int MIN_REGION = 20000;
337
338 unsigned int offset, next;
339
340 if (idx <= MUST_REGION)
341 return 0;
342
343 if (idx <= MIN_REGION) {
344 offset = idx - MUST_REGION;
345 return (offset < MIN_COMMITS) ? offset : MIN_COMMITS;
346 }
347
348 offset = idx - MIN_REGION;
349 next = (offset < MAX_COMMITS) ? offset : MAX_COMMITS;
350
351 return (next > MIN_COMMITS) ? next : MIN_COMMITS;
352 }
353
354 static int date_compare(const void *_a, const void *_b)
355 {
356 struct commit *a = *(struct commit **)_a;
357 struct commit *b = *(struct commit **)_b;
358 return (long)b->date - (long)a->date;
359 }
360
361 void bitmap_writer_reuse_bitmaps(struct packing_data *to_pack)
362 {
363 if (prepare_bitmap_git() < 0)
364 return;
365
366 writer.reused = kh_init_sha1();
367 rebuild_existing_bitmaps(to_pack, writer.reused, writer.show_progress);
368 }
369
370 static struct ewah_bitmap *find_reused_bitmap(const unsigned char *sha1)
371 {
372 khiter_t hash_pos;
373
374 if (!writer.reused)
375 return NULL;
376
377 hash_pos = kh_get_sha1(writer.reused, sha1);
378 if (hash_pos >= kh_end(writer.reused))
379 return NULL;
380
381 return kh_value(writer.reused, hash_pos);
382 }
383
384 void bitmap_writer_select_commits(struct commit **indexed_commits,
385 unsigned int indexed_commits_nr,
386 int max_bitmaps)
387 {
388 unsigned int i = 0, j, next;
389
390 qsort(indexed_commits, indexed_commits_nr, sizeof(indexed_commits[0]),
391 date_compare);
392
393 if (writer.show_progress)
394 writer.progress = start_progress("Selecting bitmap commits", 0);
395
396 if (indexed_commits_nr < 100) {
397 for (i = 0; i < indexed_commits_nr; ++i)
398 push_bitmapped_commit(indexed_commits[i], NULL);
399 return;
400 }
401
402 for (;;) {
403 struct ewah_bitmap *reused_bitmap = NULL;
404 struct commit *chosen = NULL;
405
406 next = next_commit_index(i);
407
408 if (i + next >= indexed_commits_nr)
409 break;
410
411 if (max_bitmaps > 0 && writer.selected_nr >= max_bitmaps) {
412 writer.selected_nr = max_bitmaps;
413 break;
414 }
415
416 if (next == 0) {
417 chosen = indexed_commits[i];
418 reused_bitmap = find_reused_bitmap(chosen->object.sha1);
419 } else {
420 chosen = indexed_commits[i + next];
421
422 for (j = 0; j <= next; ++j) {
423 struct commit *cm = indexed_commits[i + j];
424
425 reused_bitmap = find_reused_bitmap(cm->object.sha1);
426 if (reused_bitmap || (cm->object.flags & NEEDS_BITMAP) != 0) {
427 chosen = cm;
428 break;
429 }
430
431 if (cm->parents && cm->parents->next)
432 chosen = cm;
433 }
434 }
435
436 push_bitmapped_commit(chosen, reused_bitmap);
437
438 i += next + 1;
439 display_progress(writer.progress, i);
440 }
441
442 stop_progress(&writer.progress);
443 }
444
445
446 static int sha1write_ewah_helper(void *f, const void *buf, size_t len)
447 {
448 /* sha1write will die on error */
449 sha1write(f, buf, len);
450 return len;
451 }
452
453 /**
454 * Write the bitmap index to disk
455 */
456 static inline void dump_bitmap(struct sha1file *f, struct ewah_bitmap *bitmap)
457 {
458 if (ewah_serialize_to(bitmap, sha1write_ewah_helper, f) < 0)
459 die("Failed to write bitmap index");
460 }
461
462 static const unsigned char *sha1_access(size_t pos, void *table)
463 {
464 struct pack_idx_entry **index = table;
465 return index[pos]->sha1;
466 }
467
468 static void write_selected_commits_v1(struct sha1file *f,
469 struct pack_idx_entry **index,
470 uint32_t index_nr)
471 {
472 int i;
473
474 for (i = 0; i < writer.selected_nr; ++i) {
475 struct bitmapped_commit *stored = &writer.selected[i];
476 struct bitmap_disk_entry on_disk;
477
478 int commit_pos =
479 sha1_pos(stored->commit->object.sha1, index, index_nr, sha1_access);
480
481 if (commit_pos < 0)
482 die("BUG: trying to write commit not in index");
483
484 on_disk.object_pos = htonl(commit_pos);
485 on_disk.xor_offset = stored->xor_offset;
486 on_disk.flags = stored->flags;
487
488 sha1write(f, &on_disk, sizeof(on_disk));
489 dump_bitmap(f, stored->write_as);
490 }
491 }
492
493 void bitmap_writer_set_checksum(unsigned char *sha1)
494 {
495 hashcpy(writer.pack_checksum, sha1);
496 }
497
498 void bitmap_writer_finish(struct pack_idx_entry **index,
499 uint32_t index_nr,
500 const char *filename)
501 {
502 static char tmp_file[PATH_MAX];
503 static uint16_t default_version = 1;
504 static uint16_t flags = BITMAP_OPT_FULL_DAG;
505 struct sha1file *f;
506
507 struct bitmap_disk_header header;
508
509 int fd = odb_mkstemp(tmp_file, sizeof(tmp_file), "pack/tmp_bitmap_XXXXXX");
510
511 if (fd < 0)
512 die_errno("unable to create '%s'", tmp_file);
513 f = sha1fd(fd, tmp_file);
514
515 memcpy(header.magic, BITMAP_IDX_SIGNATURE, sizeof(BITMAP_IDX_SIGNATURE));
516 header.version = htons(default_version);
517 header.options = htons(flags);
518 header.entry_count = htonl(writer.selected_nr);
519 memcpy(header.checksum, writer.pack_checksum, 20);
520
521 sha1write(f, &header, sizeof(header));
522 dump_bitmap(f, writer.commits);
523 dump_bitmap(f, writer.trees);
524 dump_bitmap(f, writer.blobs);
525 dump_bitmap(f, writer.tags);
526 write_selected_commits_v1(f, index, index_nr);
527
528 sha1close(f, NULL, CSUM_FSYNC);
529
530 if (adjust_shared_perm(tmp_file))
531 die_errno("unable to make temporary bitmap file readable");
532
533 if (rename(tmp_file, filename))
534 die_errno("unable to rename temporary bitmap file to '%s'", filename);
535 }