*.[ch]: manually align parameter lists
[git] / midx.c
1 #include "cache.h"
2 #include "config.h"
3 #include "csum-file.h"
4 #include "dir.h"
5 #include "lockfile.h"
6 #include "packfile.h"
7 #include "object-store.h"
8 #include "sha1-lookup.h"
9 #include "midx.h"
10 #include "progress.h"
11 #include "trace2.h"
12
13 #define MIDX_SIGNATURE 0x4d494458 /* "MIDX" */
14 #define MIDX_VERSION 1
15 #define MIDX_BYTE_FILE_VERSION 4
16 #define MIDX_BYTE_HASH_VERSION 5
17 #define MIDX_BYTE_NUM_CHUNKS 6
18 #define MIDX_BYTE_NUM_PACKS 8
19 #define MIDX_HASH_VERSION 1
20 #define MIDX_HEADER_SIZE 12
21 #define MIDX_HASH_LEN 20
22 #define MIDX_MIN_SIZE (MIDX_HEADER_SIZE + MIDX_HASH_LEN)
23
24 #define MIDX_MAX_CHUNKS 5
25 #define MIDX_CHUNK_ALIGNMENT 4
26 #define MIDX_CHUNKID_PACKNAMES 0x504e414d /* "PNAM" */
27 #define MIDX_CHUNKID_OIDFANOUT 0x4f494446 /* "OIDF" */
28 #define MIDX_CHUNKID_OIDLOOKUP 0x4f49444c /* "OIDL" */
29 #define MIDX_CHUNKID_OBJECTOFFSETS 0x4f4f4646 /* "OOFF" */
30 #define MIDX_CHUNKID_LARGEOFFSETS 0x4c4f4646 /* "LOFF" */
31 #define MIDX_CHUNKLOOKUP_WIDTH (sizeof(uint32_t) + sizeof(uint64_t))
32 #define MIDX_CHUNK_FANOUT_SIZE (sizeof(uint32_t) * 256)
33 #define MIDX_CHUNK_OFFSET_WIDTH (2 * sizeof(uint32_t))
34 #define MIDX_CHUNK_LARGE_OFFSET_WIDTH (sizeof(uint64_t))
35 #define MIDX_LARGE_OFFSET_NEEDED 0x80000000
36
37 static char *get_midx_filename(const char *object_dir)
38 {
39         return xstrfmt("%s/pack/multi-pack-index", object_dir);
40 }
41
42 struct multi_pack_index *load_multi_pack_index(const char *object_dir, int local)
43 {
44         struct multi_pack_index *m = NULL;
45         int fd;
46         struct stat st;
47         size_t midx_size;
48         void *midx_map = NULL;
49         uint32_t hash_version;
50         char *midx_name = get_midx_filename(object_dir);
51         uint32_t i;
52         const char *cur_pack_name;
53
54         fd = git_open(midx_name);
55
56         if (fd < 0)
57                 goto cleanup_fail;
58         if (fstat(fd, &st)) {
59                 error_errno(_("failed to read %s"), midx_name);
60                 goto cleanup_fail;
61         }
62
63         midx_size = xsize_t(st.st_size);
64
65         if (midx_size < MIDX_MIN_SIZE) {
66                 error(_("multi-pack-index file %s is too small"), midx_name);
67                 goto cleanup_fail;
68         }
69
70         FREE_AND_NULL(midx_name);
71
72         midx_map = xmmap(NULL, midx_size, PROT_READ, MAP_PRIVATE, fd, 0);
73
74         FLEX_ALLOC_MEM(m, object_dir, object_dir, strlen(object_dir));
75         m->fd = fd;
76         m->data = midx_map;
77         m->data_len = midx_size;
78         m->local = local;
79
80         m->signature = get_be32(m->data);
81         if (m->signature != MIDX_SIGNATURE)
82                 die(_("multi-pack-index signature 0x%08x does not match signature 0x%08x"),
83                       m->signature, MIDX_SIGNATURE);
84
85         m->version = m->data[MIDX_BYTE_FILE_VERSION];
86         if (m->version != MIDX_VERSION)
87                 die(_("multi-pack-index version %d not recognized"),
88                       m->version);
89
90         hash_version = m->data[MIDX_BYTE_HASH_VERSION];
91         if (hash_version != MIDX_HASH_VERSION)
92                 die(_("hash version %u does not match"), hash_version);
93         m->hash_len = MIDX_HASH_LEN;
94
95         m->num_chunks = m->data[MIDX_BYTE_NUM_CHUNKS];
96
97         m->num_packs = get_be32(m->data + MIDX_BYTE_NUM_PACKS);
98
99         for (i = 0; i < m->num_chunks; i++) {
100                 uint32_t chunk_id = get_be32(m->data + MIDX_HEADER_SIZE +
101                                              MIDX_CHUNKLOOKUP_WIDTH * i);
102                 uint64_t chunk_offset = get_be64(m->data + MIDX_HEADER_SIZE + 4 +
103                                                  MIDX_CHUNKLOOKUP_WIDTH * i);
104
105                 if (chunk_offset >= m->data_len)
106                         die(_("invalid chunk offset (too large)"));
107
108                 switch (chunk_id) {
109                         case MIDX_CHUNKID_PACKNAMES:
110                                 m->chunk_pack_names = m->data + chunk_offset;
111                                 break;
112
113                         case MIDX_CHUNKID_OIDFANOUT:
114                                 m->chunk_oid_fanout = (uint32_t *)(m->data + chunk_offset);
115                                 break;
116
117                         case MIDX_CHUNKID_OIDLOOKUP:
118                                 m->chunk_oid_lookup = m->data + chunk_offset;
119                                 break;
120
121                         case MIDX_CHUNKID_OBJECTOFFSETS:
122                                 m->chunk_object_offsets = m->data + chunk_offset;
123                                 break;
124
125                         case MIDX_CHUNKID_LARGEOFFSETS:
126                                 m->chunk_large_offsets = m->data + chunk_offset;
127                                 break;
128
129                         case 0:
130                                 die(_("terminating multi-pack-index chunk id appears earlier than expected"));
131                                 break;
132
133                         default:
134                                 /*
135                                  * Do nothing on unrecognized chunks, allowing future
136                                  * extensions to add optional chunks.
137                                  */
138                                 break;
139                 }
140         }
141
142         if (!m->chunk_pack_names)
143                 die(_("multi-pack-index missing required pack-name chunk"));
144         if (!m->chunk_oid_fanout)
145                 die(_("multi-pack-index missing required OID fanout chunk"));
146         if (!m->chunk_oid_lookup)
147                 die(_("multi-pack-index missing required OID lookup chunk"));
148         if (!m->chunk_object_offsets)
149                 die(_("multi-pack-index missing required object offsets chunk"));
150
151         m->num_objects = ntohl(m->chunk_oid_fanout[255]);
152
153         m->pack_names = xcalloc(m->num_packs, sizeof(*m->pack_names));
154         m->packs = xcalloc(m->num_packs, sizeof(*m->packs));
155
156         cur_pack_name = (const char *)m->chunk_pack_names;
157         for (i = 0; i < m->num_packs; i++) {
158                 m->pack_names[i] = cur_pack_name;
159
160                 cur_pack_name += strlen(cur_pack_name) + 1;
161
162                 if (i && strcmp(m->pack_names[i], m->pack_names[i - 1]) <= 0)
163                         die(_("multi-pack-index pack names out of order: '%s' before '%s'"),
164                               m->pack_names[i - 1],
165                               m->pack_names[i]);
166         }
167
168         trace2_data_intmax("midx", the_repository, "load/num_packs", m->num_packs);
169         trace2_data_intmax("midx", the_repository, "load/num_objects", m->num_objects);
170
171         return m;
172
173 cleanup_fail:
174         free(m);
175         free(midx_name);
176         if (midx_map)
177                 munmap(midx_map, midx_size);
178         if (0 <= fd)
179                 close(fd);
180         return NULL;
181 }
182
183 void close_midx(struct multi_pack_index *m)
184 {
185         uint32_t i;
186
187         if (!m)
188                 return;
189
190         munmap((unsigned char *)m->data, m->data_len);
191         close(m->fd);
192         m->fd = -1;
193
194         for (i = 0; i < m->num_packs; i++) {
195                 if (m->packs[i]) {
196                         close_pack(m->packs[i]);
197                         free(m->packs[i]);
198                 }
199         }
200         FREE_AND_NULL(m->packs);
201         FREE_AND_NULL(m->pack_names);
202 }
203
204 int prepare_midx_pack(struct multi_pack_index *m, uint32_t pack_int_id)
205 {
206         struct strbuf pack_name = STRBUF_INIT;
207
208         if (pack_int_id >= m->num_packs)
209                 die(_("bad pack-int-id: %u (%u total packs)"),
210                     pack_int_id, m->num_packs);
211
212         if (m->packs[pack_int_id])
213                 return 0;
214
215         strbuf_addf(&pack_name, "%s/pack/%s", m->object_dir,
216                     m->pack_names[pack_int_id]);
217
218         m->packs[pack_int_id] = add_packed_git(pack_name.buf, pack_name.len, m->local);
219         strbuf_release(&pack_name);
220         return !m->packs[pack_int_id];
221 }
222
223 int bsearch_midx(const struct object_id *oid, struct multi_pack_index *m, uint32_t *result)
224 {
225         return bsearch_hash(oid->hash, m->chunk_oid_fanout, m->chunk_oid_lookup,
226                             MIDX_HASH_LEN, result);
227 }
228
229 struct object_id *nth_midxed_object_oid(struct object_id *oid,
230                                         struct multi_pack_index *m,
231                                         uint32_t n)
232 {
233         if (n >= m->num_objects)
234                 return NULL;
235
236         hashcpy(oid->hash, m->chunk_oid_lookup + m->hash_len * n);
237         return oid;
238 }
239
240 static off_t nth_midxed_offset(struct multi_pack_index *m, uint32_t pos)
241 {
242         const unsigned char *offset_data;
243         uint32_t offset32;
244
245         offset_data = m->chunk_object_offsets + pos * MIDX_CHUNK_OFFSET_WIDTH;
246         offset32 = get_be32(offset_data + sizeof(uint32_t));
247
248         if (m->chunk_large_offsets && offset32 & MIDX_LARGE_OFFSET_NEEDED) {
249                 if (sizeof(off_t) < sizeof(uint64_t))
250                         die(_("multi-pack-index stores a 64-bit offset, but off_t is too small"));
251
252                 offset32 ^= MIDX_LARGE_OFFSET_NEEDED;
253                 return get_be64(m->chunk_large_offsets + sizeof(uint64_t) * offset32);
254         }
255
256         return offset32;
257 }
258
259 static uint32_t nth_midxed_pack_int_id(struct multi_pack_index *m, uint32_t pos)
260 {
261         return get_be32(m->chunk_object_offsets + pos * MIDX_CHUNK_OFFSET_WIDTH);
262 }
263
264 static int nth_midxed_pack_entry(struct multi_pack_index *m, struct pack_entry *e, uint32_t pos)
265 {
266         uint32_t pack_int_id;
267         struct packed_git *p;
268
269         if (pos >= m->num_objects)
270                 return 0;
271
272         pack_int_id = nth_midxed_pack_int_id(m, pos);
273
274         if (prepare_midx_pack(m, pack_int_id))
275                 die(_("error preparing packfile from multi-pack-index"));
276         p = m->packs[pack_int_id];
277
278         /*
279         * We are about to tell the caller where they can locate the
280         * requested object.  We better make sure the packfile is
281         * still here and can be accessed before supplying that
282         * answer, as it may have been deleted since the MIDX was
283         * loaded!
284         */
285         if (!is_pack_valid(p))
286                 return 0;
287
288         if (p->num_bad_objects) {
289                 uint32_t i;
290                 struct object_id oid;
291                 nth_midxed_object_oid(&oid, m, pos);
292                 for (i = 0; i < p->num_bad_objects; i++)
293                         if (hasheq(oid.hash,
294                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
295                                 return 0;
296         }
297
298         e->offset = nth_midxed_offset(m, pos);
299         e->p = p;
300
301         return 1;
302 }
303
304 int fill_midx_entry(const struct object_id *oid, struct pack_entry *e, struct multi_pack_index *m)
305 {
306         uint32_t pos;
307
308         if (!bsearch_midx(oid, m, &pos))
309                 return 0;
310
311         return nth_midxed_pack_entry(m, e, pos);
312 }
313
314 int midx_contains_pack(struct multi_pack_index *m, const char *idx_name)
315 {
316         uint32_t first = 0, last = m->num_packs;
317
318         while (first < last) {
319                 uint32_t mid = first + (last - first) / 2;
320                 const char *current;
321                 int cmp;
322
323                 current = m->pack_names[mid];
324                 cmp = strcmp(idx_name, current);
325                 if (!cmp)
326                         return 1;
327                 if (cmp > 0) {
328                         first = mid + 1;
329                         continue;
330                 }
331                 last = mid;
332         }
333
334         return 0;
335 }
336
337 int prepare_multi_pack_index_one(struct repository *r, const char *object_dir, int local)
338 {
339         struct multi_pack_index *m;
340         struct multi_pack_index *m_search;
341         int config_value;
342         static int env_value = -1;
343
344         if (env_value < 0)
345                 env_value = git_env_bool(GIT_TEST_MULTI_PACK_INDEX, 0);
346
347         if (!env_value &&
348             (repo_config_get_bool(r, "core.multipackindex", &config_value) ||
349             !config_value))
350                 return 0;
351
352         for (m_search = r->objects->multi_pack_index; m_search; m_search = m_search->next)
353                 if (!strcmp(object_dir, m_search->object_dir))
354                         return 1;
355
356         m = load_multi_pack_index(object_dir, local);
357
358         if (m) {
359                 m->next = r->objects->multi_pack_index;
360                 r->objects->multi_pack_index = m;
361                 return 1;
362         }
363
364         return 0;
365 }
366
367 static size_t write_midx_header(struct hashfile *f,
368                                 unsigned char num_chunks,
369                                 uint32_t num_packs)
370 {
371         unsigned char byte_values[4];
372
373         hashwrite_be32(f, MIDX_SIGNATURE);
374         byte_values[0] = MIDX_VERSION;
375         byte_values[1] = MIDX_HASH_VERSION;
376         byte_values[2] = num_chunks;
377         byte_values[3] = 0; /* unused */
378         hashwrite(f, byte_values, sizeof(byte_values));
379         hashwrite_be32(f, num_packs);
380
381         return MIDX_HEADER_SIZE;
382 }
383
384 struct pack_list {
385         struct packed_git **list;
386         char **names;
387         uint32_t nr;
388         uint32_t alloc_list;
389         uint32_t alloc_names;
390         size_t pack_name_concat_len;
391         struct multi_pack_index *m;
392 };
393
394 static void add_pack_to_midx(const char *full_path, size_t full_path_len,
395                              const char *file_name, void *data)
396 {
397         struct pack_list *packs = (struct pack_list *)data;
398
399         if (ends_with(file_name, ".idx")) {
400                 if (packs->m && midx_contains_pack(packs->m, file_name))
401                         return;
402
403                 ALLOC_GROW(packs->list, packs->nr + 1, packs->alloc_list);
404                 ALLOC_GROW(packs->names, packs->nr + 1, packs->alloc_names);
405
406                 packs->list[packs->nr] = add_packed_git(full_path,
407                                                         full_path_len,
408                                                         0);
409
410                 if (!packs->list[packs->nr]) {
411                         warning(_("failed to add packfile '%s'"),
412                                 full_path);
413                         return;
414                 }
415
416                 if (open_pack_index(packs->list[packs->nr])) {
417                         warning(_("failed to open pack-index '%s'"),
418                                 full_path);
419                         close_pack(packs->list[packs->nr]);
420                         FREE_AND_NULL(packs->list[packs->nr]);
421                         return;
422                 }
423
424                 packs->names[packs->nr] = xstrdup(file_name);
425                 packs->pack_name_concat_len += strlen(file_name) + 1;
426                 packs->nr++;
427         }
428 }
429
430 struct pack_pair {
431         uint32_t pack_int_id;
432         char *pack_name;
433 };
434
435 static int pack_pair_compare(const void *_a, const void *_b)
436 {
437         struct pack_pair *a = (struct pack_pair *)_a;
438         struct pack_pair *b = (struct pack_pair *)_b;
439         return strcmp(a->pack_name, b->pack_name);
440 }
441
442 static void sort_packs_by_name(char **pack_names, uint32_t nr_packs, uint32_t *perm)
443 {
444         uint32_t i;
445         struct pack_pair *pairs;
446
447         ALLOC_ARRAY(pairs, nr_packs);
448
449         for (i = 0; i < nr_packs; i++) {
450                 pairs[i].pack_int_id = i;
451                 pairs[i].pack_name = pack_names[i];
452         }
453
454         QSORT(pairs, nr_packs, pack_pair_compare);
455
456         for (i = 0; i < nr_packs; i++) {
457                 pack_names[i] = pairs[i].pack_name;
458                 perm[pairs[i].pack_int_id] = i;
459         }
460
461         free(pairs);
462 }
463
464 struct pack_midx_entry {
465         struct object_id oid;
466         uint32_t pack_int_id;
467         time_t pack_mtime;
468         uint64_t offset;
469 };
470
471 static int midx_oid_compare(const void *_a, const void *_b)
472 {
473         const struct pack_midx_entry *a = (const struct pack_midx_entry *)_a;
474         const struct pack_midx_entry *b = (const struct pack_midx_entry *)_b;
475         int cmp = oidcmp(&a->oid, &b->oid);
476
477         if (cmp)
478                 return cmp;
479
480         if (a->pack_mtime > b->pack_mtime)
481                 return -1;
482         else if (a->pack_mtime < b->pack_mtime)
483                 return 1;
484
485         return a->pack_int_id - b->pack_int_id;
486 }
487
488 static int nth_midxed_pack_midx_entry(struct multi_pack_index *m,
489                                       uint32_t *pack_perm,
490                                       struct pack_midx_entry *e,
491                                       uint32_t pos)
492 {
493         if (pos >= m->num_objects)
494                 return 1;
495
496         nth_midxed_object_oid(&e->oid, m, pos);
497         e->pack_int_id = pack_perm[nth_midxed_pack_int_id(m, pos)];
498         e->offset = nth_midxed_offset(m, pos);
499
500         /* consider objects in midx to be from "old" packs */
501         e->pack_mtime = 0;
502         return 0;
503 }
504
505 static void fill_pack_entry(uint32_t pack_int_id,
506                             struct packed_git *p,
507                             uint32_t cur_object,
508                             struct pack_midx_entry *entry)
509 {
510         if (!nth_packed_object_oid(&entry->oid, p, cur_object))
511                 die(_("failed to locate object %d in packfile"), cur_object);
512
513         entry->pack_int_id = pack_int_id;
514         entry->pack_mtime = p->mtime;
515
516         entry->offset = nth_packed_object_offset(p, cur_object);
517 }
518
519 /*
520  * It is possible to artificially get into a state where there are many
521  * duplicate copies of objects. That can create high memory pressure if
522  * we are to create a list of all objects before de-duplication. To reduce
523  * this memory pressure without a significant performance drop, automatically
524  * group objects by the first byte of their object id. Use the IDX fanout
525  * tables to group the data, copy to a local array, then sort.
526  *
527  * Copy only the de-duplicated entries (selected by most-recent modified time
528  * of a packfile containing the object).
529  */
530 static struct pack_midx_entry *get_sorted_entries(struct multi_pack_index *m,
531                                                   struct packed_git **p,
532                                                   uint32_t *perm,
533                                                   uint32_t nr_packs,
534                                                   uint32_t *nr_objects)
535 {
536         uint32_t cur_fanout, cur_pack, cur_object;
537         uint32_t alloc_fanout, alloc_objects, total_objects = 0;
538         struct pack_midx_entry *entries_by_fanout = NULL;
539         struct pack_midx_entry *deduplicated_entries = NULL;
540         uint32_t start_pack = m ? m->num_packs : 0;
541
542         for (cur_pack = start_pack; cur_pack < nr_packs; cur_pack++)
543                 total_objects += p[cur_pack]->num_objects;
544
545         /*
546          * As we de-duplicate by fanout value, we expect the fanout
547          * slices to be evenly distributed, with some noise. Hence,
548          * allocate slightly more than one 256th.
549          */
550         alloc_objects = alloc_fanout = total_objects > 3200 ? total_objects / 200 : 16;
551
552         ALLOC_ARRAY(entries_by_fanout, alloc_fanout);
553         ALLOC_ARRAY(deduplicated_entries, alloc_objects);
554         *nr_objects = 0;
555
556         for (cur_fanout = 0; cur_fanout < 256; cur_fanout++) {
557                 uint32_t nr_fanout = 0;
558
559                 if (m) {
560                         uint32_t start = 0, end;
561
562                         if (cur_fanout)
563                                 start = ntohl(m->chunk_oid_fanout[cur_fanout - 1]);
564                         end = ntohl(m->chunk_oid_fanout[cur_fanout]);
565
566                         for (cur_object = start; cur_object < end; cur_object++) {
567                                 ALLOC_GROW(entries_by_fanout, nr_fanout + 1, alloc_fanout);
568                                 nth_midxed_pack_midx_entry(m, perm,
569                                                            &entries_by_fanout[nr_fanout],
570                                                            cur_object);
571                                 nr_fanout++;
572                         }
573                 }
574
575                 for (cur_pack = start_pack; cur_pack < nr_packs; cur_pack++) {
576                         uint32_t start = 0, end;
577
578                         if (cur_fanout)
579                                 start = get_pack_fanout(p[cur_pack], cur_fanout - 1);
580                         end = get_pack_fanout(p[cur_pack], cur_fanout);
581
582                         for (cur_object = start; cur_object < end; cur_object++) {
583                                 ALLOC_GROW(entries_by_fanout, nr_fanout + 1, alloc_fanout);
584                                 fill_pack_entry(perm[cur_pack], p[cur_pack], cur_object, &entries_by_fanout[nr_fanout]);
585                                 nr_fanout++;
586                         }
587                 }
588
589                 QSORT(entries_by_fanout, nr_fanout, midx_oid_compare);
590
591                 /*
592                  * The batch is now sorted by OID and then mtime (descending).
593                  * Take only the first duplicate.
594                  */
595                 for (cur_object = 0; cur_object < nr_fanout; cur_object++) {
596                         if (cur_object && oideq(&entries_by_fanout[cur_object - 1].oid,
597                                                 &entries_by_fanout[cur_object].oid))
598                                 continue;
599
600                         ALLOC_GROW(deduplicated_entries, *nr_objects + 1, alloc_objects);
601                         memcpy(&deduplicated_entries[*nr_objects],
602                                &entries_by_fanout[cur_object],
603                                sizeof(struct pack_midx_entry));
604                         (*nr_objects)++;
605                 }
606         }
607
608         free(entries_by_fanout);
609         return deduplicated_entries;
610 }
611
612 static size_t write_midx_pack_names(struct hashfile *f,
613                                     char **pack_names,
614                                     uint32_t num_packs)
615 {
616         uint32_t i;
617         unsigned char padding[MIDX_CHUNK_ALIGNMENT];
618         size_t written = 0;
619
620         for (i = 0; i < num_packs; i++) {
621                 size_t writelen = strlen(pack_names[i]) + 1;
622
623                 if (i && strcmp(pack_names[i], pack_names[i - 1]) <= 0)
624                         BUG("incorrect pack-file order: %s before %s",
625                             pack_names[i - 1],
626                             pack_names[i]);
627
628                 hashwrite(f, pack_names[i], writelen);
629                 written += writelen;
630         }
631
632         /* add padding to be aligned */
633         i = MIDX_CHUNK_ALIGNMENT - (written % MIDX_CHUNK_ALIGNMENT);
634         if (i < MIDX_CHUNK_ALIGNMENT) {
635                 memset(padding, 0, sizeof(padding));
636                 hashwrite(f, padding, i);
637                 written += i;
638         }
639
640         return written;
641 }
642
643 static size_t write_midx_oid_fanout(struct hashfile *f,
644                                     struct pack_midx_entry *objects,
645                                     uint32_t nr_objects)
646 {
647         struct pack_midx_entry *list = objects;
648         struct pack_midx_entry *last = objects + nr_objects;
649         uint32_t count = 0;
650         uint32_t i;
651
652         /*
653         * Write the first-level table (the list is sorted,
654         * but we use a 256-entry lookup to be able to avoid
655         * having to do eight extra binary search iterations).
656         */
657         for (i = 0; i < 256; i++) {
658                 struct pack_midx_entry *next = list;
659
660                 while (next < last && next->oid.hash[0] == i) {
661                         count++;
662                         next++;
663                 }
664
665                 hashwrite_be32(f, count);
666                 list = next;
667         }
668
669         return MIDX_CHUNK_FANOUT_SIZE;
670 }
671
672 static size_t write_midx_oid_lookup(struct hashfile *f, unsigned char hash_len,
673                                     struct pack_midx_entry *objects,
674                                     uint32_t nr_objects)
675 {
676         struct pack_midx_entry *list = objects;
677         uint32_t i;
678         size_t written = 0;
679
680         for (i = 0; i < nr_objects; i++) {
681                 struct pack_midx_entry *obj = list++;
682
683                 if (i < nr_objects - 1) {
684                         struct pack_midx_entry *next = list;
685                         if (oidcmp(&obj->oid, &next->oid) >= 0)
686                                 BUG("OIDs not in order: %s >= %s",
687                                     oid_to_hex(&obj->oid),
688                                     oid_to_hex(&next->oid));
689                 }
690
691                 hashwrite(f, obj->oid.hash, (int)hash_len);
692                 written += hash_len;
693         }
694
695         return written;
696 }
697
698 static size_t write_midx_object_offsets(struct hashfile *f, int large_offset_needed,
699                                         struct pack_midx_entry *objects, uint32_t nr_objects)
700 {
701         struct pack_midx_entry *list = objects;
702         uint32_t i, nr_large_offset = 0;
703         size_t written = 0;
704
705         for (i = 0; i < nr_objects; i++) {
706                 struct pack_midx_entry *obj = list++;
707
708                 hashwrite_be32(f, obj->pack_int_id);
709
710                 if (large_offset_needed && obj->offset >> 31)
711                         hashwrite_be32(f, MIDX_LARGE_OFFSET_NEEDED | nr_large_offset++);
712                 else if (!large_offset_needed && obj->offset >> 32)
713                         BUG("object %s requires a large offset (%"PRIx64") but the MIDX is not writing large offsets!",
714                             oid_to_hex(&obj->oid),
715                             obj->offset);
716                 else
717                         hashwrite_be32(f, (uint32_t)obj->offset);
718
719                 written += MIDX_CHUNK_OFFSET_WIDTH;
720         }
721
722         return written;
723 }
724
725 static size_t write_midx_large_offsets(struct hashfile *f, uint32_t nr_large_offset,
726                                        struct pack_midx_entry *objects, uint32_t nr_objects)
727 {
728         struct pack_midx_entry *list = objects, *end = objects + nr_objects;
729         size_t written = 0;
730
731         while (nr_large_offset) {
732                 struct pack_midx_entry *obj;
733                 uint64_t offset;
734
735                 if (list >= end)
736                         BUG("too many large-offset objects");
737
738                 obj = list++;
739                 offset = obj->offset;
740
741                 if (!(offset >> 31))
742                         continue;
743
744                 hashwrite_be32(f, offset >> 32);
745                 hashwrite_be32(f, offset & 0xffffffffUL);
746                 written += 2 * sizeof(uint32_t);
747
748                 nr_large_offset--;
749         }
750
751         return written;
752 }
753
754 int write_midx_file(const char *object_dir)
755 {
756         unsigned char cur_chunk, num_chunks = 0;
757         char *midx_name;
758         uint32_t i;
759         struct hashfile *f = NULL;
760         struct lock_file lk;
761         struct pack_list packs;
762         uint32_t *pack_perm = NULL;
763         uint64_t written = 0;
764         uint32_t chunk_ids[MIDX_MAX_CHUNKS + 1];
765         uint64_t chunk_offsets[MIDX_MAX_CHUNKS + 1];
766         uint32_t nr_entries, num_large_offsets = 0;
767         struct pack_midx_entry *entries = NULL;
768         int large_offsets_needed = 0;
769
770         midx_name = get_midx_filename(object_dir);
771         if (safe_create_leading_directories(midx_name)) {
772                 UNLEAK(midx_name);
773                 die_errno(_("unable to create leading directories of %s"),
774                           midx_name);
775         }
776
777         packs.m = load_multi_pack_index(object_dir, 1);
778
779         packs.nr = 0;
780         packs.alloc_list = packs.m ? packs.m->num_packs : 16;
781         packs.alloc_names = packs.alloc_list;
782         packs.list = NULL;
783         packs.names = NULL;
784         packs.pack_name_concat_len = 0;
785         ALLOC_ARRAY(packs.list, packs.alloc_list);
786         ALLOC_ARRAY(packs.names, packs.alloc_names);
787
788         if (packs.m) {
789                 for (i = 0; i < packs.m->num_packs; i++) {
790                         ALLOC_GROW(packs.list, packs.nr + 1, packs.alloc_list);
791                         ALLOC_GROW(packs.names, packs.nr + 1, packs.alloc_names);
792
793                         packs.list[packs.nr] = NULL;
794                         packs.names[packs.nr] = xstrdup(packs.m->pack_names[i]);
795                         packs.pack_name_concat_len += strlen(packs.names[packs.nr]) + 1;
796                         packs.nr++;
797                 }
798         }
799
800         for_each_file_in_pack_dir(object_dir, add_pack_to_midx, &packs);
801
802         if (packs.m && packs.nr == packs.m->num_packs)
803                 goto cleanup;
804
805         if (packs.pack_name_concat_len % MIDX_CHUNK_ALIGNMENT)
806                 packs.pack_name_concat_len += MIDX_CHUNK_ALIGNMENT -
807                                               (packs.pack_name_concat_len % MIDX_CHUNK_ALIGNMENT);
808
809         ALLOC_ARRAY(pack_perm, packs.nr);
810         sort_packs_by_name(packs.names, packs.nr, pack_perm);
811
812         entries = get_sorted_entries(packs.m, packs.list, pack_perm, packs.nr, &nr_entries);
813
814         for (i = 0; i < nr_entries; i++) {
815                 if (entries[i].offset > 0x7fffffff)
816                         num_large_offsets++;
817                 if (entries[i].offset > 0xffffffff)
818                         large_offsets_needed = 1;
819         }
820
821         hold_lock_file_for_update(&lk, midx_name, LOCK_DIE_ON_ERROR);
822         f = hashfd(lk.tempfile->fd, lk.tempfile->filename.buf);
823         FREE_AND_NULL(midx_name);
824
825         if (packs.m)
826                 close_midx(packs.m);
827
828         cur_chunk = 0;
829         num_chunks = large_offsets_needed ? 5 : 4;
830
831         written = write_midx_header(f, num_chunks, packs.nr);
832
833         chunk_ids[cur_chunk] = MIDX_CHUNKID_PACKNAMES;
834         chunk_offsets[cur_chunk] = written + (num_chunks + 1) * MIDX_CHUNKLOOKUP_WIDTH;
835
836         cur_chunk++;
837         chunk_ids[cur_chunk] = MIDX_CHUNKID_OIDFANOUT;
838         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + packs.pack_name_concat_len;
839
840         cur_chunk++;
841         chunk_ids[cur_chunk] = MIDX_CHUNKID_OIDLOOKUP;
842         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + MIDX_CHUNK_FANOUT_SIZE;
843
844         cur_chunk++;
845         chunk_ids[cur_chunk] = MIDX_CHUNKID_OBJECTOFFSETS;
846         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + nr_entries * MIDX_HASH_LEN;
847
848         cur_chunk++;
849         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + nr_entries * MIDX_CHUNK_OFFSET_WIDTH;
850         if (large_offsets_needed) {
851                 chunk_ids[cur_chunk] = MIDX_CHUNKID_LARGEOFFSETS;
852
853                 cur_chunk++;
854                 chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] +
855                                            num_large_offsets * MIDX_CHUNK_LARGE_OFFSET_WIDTH;
856         }
857
858         chunk_ids[cur_chunk] = 0;
859
860         for (i = 0; i <= num_chunks; i++) {
861                 if (i && chunk_offsets[i] < chunk_offsets[i - 1])
862                         BUG("incorrect chunk offsets: %"PRIu64" before %"PRIu64,
863                             chunk_offsets[i - 1],
864                             chunk_offsets[i]);
865
866                 if (chunk_offsets[i] % MIDX_CHUNK_ALIGNMENT)
867                         BUG("chunk offset %"PRIu64" is not properly aligned",
868                             chunk_offsets[i]);
869
870                 hashwrite_be32(f, chunk_ids[i]);
871                 hashwrite_be32(f, chunk_offsets[i] >> 32);
872                 hashwrite_be32(f, chunk_offsets[i]);
873
874                 written += MIDX_CHUNKLOOKUP_WIDTH;
875         }
876
877         for (i = 0; i < num_chunks; i++) {
878                 if (written != chunk_offsets[i])
879                         BUG("incorrect chunk offset (%"PRIu64" != %"PRIu64") for chunk id %"PRIx32,
880                             chunk_offsets[i],
881                             written,
882                             chunk_ids[i]);
883
884                 switch (chunk_ids[i]) {
885                         case MIDX_CHUNKID_PACKNAMES:
886                                 written += write_midx_pack_names(f, packs.names, packs.nr);
887                                 break;
888
889                         case MIDX_CHUNKID_OIDFANOUT:
890                                 written += write_midx_oid_fanout(f, entries, nr_entries);
891                                 break;
892
893                         case MIDX_CHUNKID_OIDLOOKUP:
894                                 written += write_midx_oid_lookup(f, MIDX_HASH_LEN, entries, nr_entries);
895                                 break;
896
897                         case MIDX_CHUNKID_OBJECTOFFSETS:
898                                 written += write_midx_object_offsets(f, large_offsets_needed, entries, nr_entries);
899                                 break;
900
901                         case MIDX_CHUNKID_LARGEOFFSETS:
902                                 written += write_midx_large_offsets(f, num_large_offsets, entries, nr_entries);
903                                 break;
904
905                         default:
906                                 BUG("trying to write unknown chunk id %"PRIx32,
907                                     chunk_ids[i]);
908                 }
909         }
910
911         if (written != chunk_offsets[num_chunks])
912                 BUG("incorrect final offset %"PRIu64" != %"PRIu64,
913                     written,
914                     chunk_offsets[num_chunks]);
915
916         finalize_hashfile(f, NULL, CSUM_FSYNC | CSUM_HASH_IN_STREAM);
917         commit_lock_file(&lk);
918
919 cleanup:
920         for (i = 0; i < packs.nr; i++) {
921                 if (packs.list[i]) {
922                         close_pack(packs.list[i]);
923                         free(packs.list[i]);
924                 }
925                 free(packs.names[i]);
926         }
927
928         free(packs.list);
929         free(packs.names);
930         free(entries);
931         free(pack_perm);
932         free(midx_name);
933         return 0;
934 }
935
936 void clear_midx_file(struct repository *r)
937 {
938         char *midx = get_midx_filename(r->objects->odb->path);
939
940         if (r->objects && r->objects->multi_pack_index) {
941                 close_midx(r->objects->multi_pack_index);
942                 r->objects->multi_pack_index = NULL;
943         }
944
945         if (remove_path(midx)) {
946                 UNLEAK(midx);
947                 die(_("failed to clear multi-pack-index at %s"), midx);
948         }
949
950         free(midx);
951 }
952
953 static int verify_midx_error;
954
955 static void midx_report(const char *fmt, ...)
956 {
957         va_list ap;
958         verify_midx_error = 1;
959         va_start(ap, fmt);
960         vfprintf(stderr, fmt, ap);
961         fprintf(stderr, "\n");
962         va_end(ap);
963 }
964
965 struct pair_pos_vs_id
966 {
967         uint32_t pos;
968         uint32_t pack_int_id;
969 };
970
971 static int compare_pair_pos_vs_id(const void *_a, const void *_b)
972 {
973         struct pair_pos_vs_id *a = (struct pair_pos_vs_id *)_a;
974         struct pair_pos_vs_id *b = (struct pair_pos_vs_id *)_b;
975
976         return b->pack_int_id - a->pack_int_id;
977 }
978
979 /*
980  * Limit calls to display_progress() for performance reasons.
981  * The interval here was arbitrarily chosen.
982  */
983 #define SPARSE_PROGRESS_INTERVAL (1 << 12)
984 #define midx_display_sparse_progress(progress, n) \
985         do { \
986                 uint64_t _n = (n); \
987                 if ((_n & (SPARSE_PROGRESS_INTERVAL - 1)) == 0) \
988                         display_progress(progress, _n); \
989         } while (0)
990
991 int verify_midx_file(const char *object_dir)
992 {
993         struct pair_pos_vs_id *pairs = NULL;
994         uint32_t i;
995         struct progress *progress;
996         struct multi_pack_index *m = load_multi_pack_index(object_dir, 1);
997         verify_midx_error = 0;
998
999         if (!m)
1000                 return 0;
1001
1002         progress = start_progress(_("Looking for referenced packfiles"),
1003                                   m->num_packs);
1004         for (i = 0; i < m->num_packs; i++) {
1005                 if (prepare_midx_pack(m, i))
1006                         midx_report("failed to load pack in position %d", i);
1007
1008                 display_progress(progress, i + 1);
1009         }
1010         stop_progress(&progress);
1011
1012         for (i = 0; i < 255; i++) {
1013                 uint32_t oid_fanout1 = ntohl(m->chunk_oid_fanout[i]);
1014                 uint32_t oid_fanout2 = ntohl(m->chunk_oid_fanout[i + 1]);
1015
1016                 if (oid_fanout1 > oid_fanout2)
1017                         midx_report(_("oid fanout out of order: fanout[%d] = %"PRIx32" > %"PRIx32" = fanout[%d]"),
1018                                     i, oid_fanout1, oid_fanout2, i + 1);
1019         }
1020
1021         progress = start_sparse_progress(_("Verifying OID order in MIDX"),
1022                                          m->num_objects - 1);
1023         for (i = 0; i < m->num_objects - 1; i++) {
1024                 struct object_id oid1, oid2;
1025
1026                 nth_midxed_object_oid(&oid1, m, i);
1027                 nth_midxed_object_oid(&oid2, m, i + 1);
1028
1029                 if (oidcmp(&oid1, &oid2) >= 0)
1030                         midx_report(_("oid lookup out of order: oid[%d] = %s >= %s = oid[%d]"),
1031                                     i, oid_to_hex(&oid1), oid_to_hex(&oid2), i + 1);
1032
1033                 midx_display_sparse_progress(progress, i + 1);
1034         }
1035         stop_progress(&progress);
1036
1037         /*
1038          * Create an array mapping each object to its packfile id.  Sort it
1039          * to group the objects by packfile.  Use this permutation to visit
1040          * each of the objects and only require 1 packfile to be open at a
1041          * time.
1042          */
1043         ALLOC_ARRAY(pairs, m->num_objects);
1044         for (i = 0; i < m->num_objects; i++) {
1045                 pairs[i].pos = i;
1046                 pairs[i].pack_int_id = nth_midxed_pack_int_id(m, i);
1047         }
1048
1049         progress = start_sparse_progress(_("Sorting objects by packfile"),
1050                                          m->num_objects);
1051         display_progress(progress, 0); /* TODO: Measure QSORT() progress */
1052         QSORT(pairs, m->num_objects, compare_pair_pos_vs_id);
1053         stop_progress(&progress);
1054
1055         progress = start_sparse_progress(_("Verifying object offsets"), m->num_objects);
1056         for (i = 0; i < m->num_objects; i++) {
1057                 struct object_id oid;
1058                 struct pack_entry e;
1059                 off_t m_offset, p_offset;
1060
1061                 if (i > 0 && pairs[i-1].pack_int_id != pairs[i].pack_int_id &&
1062                     m->packs[pairs[i-1].pack_int_id])
1063                 {
1064                         close_pack_fd(m->packs[pairs[i-1].pack_int_id]);
1065                         close_pack_index(m->packs[pairs[i-1].pack_int_id]);
1066                 }
1067
1068                 nth_midxed_object_oid(&oid, m, pairs[i].pos);
1069
1070                 if (!fill_midx_entry(&oid, &e, m)) {
1071                         midx_report(_("failed to load pack entry for oid[%d] = %s"),
1072                                     pairs[i].pos, oid_to_hex(&oid));
1073                         continue;
1074                 }
1075
1076                 if (open_pack_index(e.p)) {
1077                         midx_report(_("failed to load pack-index for packfile %s"),
1078                                     e.p->pack_name);
1079                         break;
1080                 }
1081
1082                 m_offset = e.offset;
1083                 p_offset = find_pack_entry_one(oid.hash, e.p);
1084
1085                 if (m_offset != p_offset)
1086                         midx_report(_("incorrect object offset for oid[%d] = %s: %"PRIx64" != %"PRIx64),
1087                                     pairs[i].pos, oid_to_hex(&oid), m_offset, p_offset);
1088
1089                 midx_display_sparse_progress(progress, i + 1);
1090         }
1091         stop_progress(&progress);
1092
1093         free(pairs);
1094
1095         return verify_midx_error;
1096 }