grep: skip UTF8 checks explicitly
[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_STR(m, object_dir, 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                         m->packs[i]->multi_pack_index = 0;
197         }
198         FREE_AND_NULL(m->packs);
199         FREE_AND_NULL(m->pack_names);
200 }
201
202 int prepare_midx_pack(struct repository *r, struct multi_pack_index *m, uint32_t pack_int_id)
203 {
204         struct strbuf pack_name = STRBUF_INIT;
205         struct packed_git *p;
206
207         if (pack_int_id >= m->num_packs)
208                 die(_("bad pack-int-id: %u (%u total packs)"),
209                     pack_int_id, m->num_packs);
210
211         if (m->packs[pack_int_id])
212                 return 0;
213
214         strbuf_addf(&pack_name, "%s/pack/%s", m->object_dir,
215                     m->pack_names[pack_int_id]);
216
217         p = add_packed_git(pack_name.buf, pack_name.len, m->local);
218         strbuf_release(&pack_name);
219
220         if (!p)
221                 return 1;
222
223         p->multi_pack_index = 1;
224         m->packs[pack_int_id] = p;
225         install_packed_git(r, p);
226         list_add_tail(&p->mru, &r->objects->packed_git_mru);
227
228         return 0;
229 }
230
231 int bsearch_midx(const struct object_id *oid, struct multi_pack_index *m, uint32_t *result)
232 {
233         return bsearch_hash(oid->hash, m->chunk_oid_fanout, m->chunk_oid_lookup,
234                             MIDX_HASH_LEN, result);
235 }
236
237 struct object_id *nth_midxed_object_oid(struct object_id *oid,
238                                         struct multi_pack_index *m,
239                                         uint32_t n)
240 {
241         if (n >= m->num_objects)
242                 return NULL;
243
244         hashcpy(oid->hash, m->chunk_oid_lookup + m->hash_len * n);
245         return oid;
246 }
247
248 static off_t nth_midxed_offset(struct multi_pack_index *m, uint32_t pos)
249 {
250         const unsigned char *offset_data;
251         uint32_t offset32;
252
253         offset_data = m->chunk_object_offsets + pos * MIDX_CHUNK_OFFSET_WIDTH;
254         offset32 = get_be32(offset_data + sizeof(uint32_t));
255
256         if (m->chunk_large_offsets && offset32 & MIDX_LARGE_OFFSET_NEEDED) {
257                 if (sizeof(off_t) < sizeof(uint64_t))
258                         die(_("multi-pack-index stores a 64-bit offset, but off_t is too small"));
259
260                 offset32 ^= MIDX_LARGE_OFFSET_NEEDED;
261                 return get_be64(m->chunk_large_offsets + sizeof(uint64_t) * offset32);
262         }
263
264         return offset32;
265 }
266
267 static uint32_t nth_midxed_pack_int_id(struct multi_pack_index *m, uint32_t pos)
268 {
269         return get_be32(m->chunk_object_offsets + pos * MIDX_CHUNK_OFFSET_WIDTH);
270 }
271
272 static int nth_midxed_pack_entry(struct repository *r,
273                                  struct multi_pack_index *m,
274                                  struct pack_entry *e,
275                                  uint32_t pos)
276 {
277         uint32_t pack_int_id;
278         struct packed_git *p;
279
280         if (pos >= m->num_objects)
281                 return 0;
282
283         pack_int_id = nth_midxed_pack_int_id(m, pos);
284
285         if (prepare_midx_pack(r, m, pack_int_id))
286                 die(_("error preparing packfile from multi-pack-index"));
287         p = m->packs[pack_int_id];
288
289         /*
290         * We are about to tell the caller where they can locate the
291         * requested object.  We better make sure the packfile is
292         * still here and can be accessed before supplying that
293         * answer, as it may have been deleted since the MIDX was
294         * loaded!
295         */
296         if (!is_pack_valid(p))
297                 return 0;
298
299         if (p->num_bad_objects) {
300                 uint32_t i;
301                 struct object_id oid;
302                 nth_midxed_object_oid(&oid, m, pos);
303                 for (i = 0; i < p->num_bad_objects; i++)
304                         if (hasheq(oid.hash,
305                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
306                                 return 0;
307         }
308
309         e->offset = nth_midxed_offset(m, pos);
310         e->p = p;
311
312         return 1;
313 }
314
315 int fill_midx_entry(struct repository * r,
316                     const struct object_id *oid,
317                     struct pack_entry *e,
318                     struct multi_pack_index *m)
319 {
320         uint32_t pos;
321
322         if (!bsearch_midx(oid, m, &pos))
323                 return 0;
324
325         return nth_midxed_pack_entry(r, m, e, pos);
326 }
327
328 /* Match "foo.idx" against either "foo.pack" _or_ "foo.idx". */
329 static int cmp_idx_or_pack_name(const char *idx_or_pack_name,
330                                 const char *idx_name)
331 {
332         /* Skip past any initial matching prefix. */
333         while (*idx_name && *idx_name == *idx_or_pack_name) {
334                 idx_name++;
335                 idx_or_pack_name++;
336         }
337
338         /*
339          * If we didn't match completely, we may have matched "pack-1234." and
340          * be left with "idx" and "pack" respectively, which is also OK. We do
341          * not have to check for "idx" and "idx", because that would have been
342          * a complete match (and in that case these strcmps will be false, but
343          * we'll correctly return 0 from the final strcmp() below.
344          *
345          * Technically this matches "fooidx" and "foopack", but we'd never have
346          * such names in the first place.
347          */
348         if (!strcmp(idx_name, "idx") && !strcmp(idx_or_pack_name, "pack"))
349                 return 0;
350
351         /*
352          * This not only checks for a complete match, but also orders based on
353          * the first non-identical character, which means our ordering will
354          * match a raw strcmp(). That makes it OK to use this to binary search
355          * a naively-sorted list.
356          */
357         return strcmp(idx_or_pack_name, idx_name);
358 }
359
360 int midx_contains_pack(struct multi_pack_index *m, const char *idx_or_pack_name)
361 {
362         uint32_t first = 0, last = m->num_packs;
363
364         while (first < last) {
365                 uint32_t mid = first + (last - first) / 2;
366                 const char *current;
367                 int cmp;
368
369                 current = m->pack_names[mid];
370                 cmp = cmp_idx_or_pack_name(idx_or_pack_name, current);
371                 if (!cmp)
372                         return 1;
373                 if (cmp > 0) {
374                         first = mid + 1;
375                         continue;
376                 }
377                 last = mid;
378         }
379
380         return 0;
381 }
382
383 int prepare_multi_pack_index_one(struct repository *r, const char *object_dir, int local)
384 {
385         struct multi_pack_index *m;
386         struct multi_pack_index *m_search;
387         int config_value;
388         static int env_value = -1;
389
390         if (env_value < 0)
391                 env_value = git_env_bool(GIT_TEST_MULTI_PACK_INDEX, 0);
392
393         if (!env_value &&
394             (repo_config_get_bool(r, "core.multipackindex", &config_value) ||
395             !config_value))
396                 return 0;
397
398         for (m_search = r->objects->multi_pack_index; m_search; m_search = m_search->next)
399                 if (!strcmp(object_dir, m_search->object_dir))
400                         return 1;
401
402         m = load_multi_pack_index(object_dir, local);
403
404         if (m) {
405                 m->next = r->objects->multi_pack_index;
406                 r->objects->multi_pack_index = m;
407                 return 1;
408         }
409
410         return 0;
411 }
412
413 static size_t write_midx_header(struct hashfile *f,
414                                 unsigned char num_chunks,
415                                 uint32_t num_packs)
416 {
417         unsigned char byte_values[4];
418
419         hashwrite_be32(f, MIDX_SIGNATURE);
420         byte_values[0] = MIDX_VERSION;
421         byte_values[1] = MIDX_HASH_VERSION;
422         byte_values[2] = num_chunks;
423         byte_values[3] = 0; /* unused */
424         hashwrite(f, byte_values, sizeof(byte_values));
425         hashwrite_be32(f, num_packs);
426
427         return MIDX_HEADER_SIZE;
428 }
429
430 struct pack_list {
431         struct packed_git **list;
432         char **names;
433         uint32_t nr;
434         uint32_t alloc_list;
435         uint32_t alloc_names;
436         size_t pack_name_concat_len;
437         struct multi_pack_index *m;
438 };
439
440 static void add_pack_to_midx(const char *full_path, size_t full_path_len,
441                              const char *file_name, void *data)
442 {
443         struct pack_list *packs = (struct pack_list *)data;
444
445         if (ends_with(file_name, ".idx")) {
446                 if (packs->m && midx_contains_pack(packs->m, file_name))
447                         return;
448
449                 ALLOC_GROW(packs->list, packs->nr + 1, packs->alloc_list);
450                 ALLOC_GROW(packs->names, packs->nr + 1, packs->alloc_names);
451
452                 packs->list[packs->nr] = add_packed_git(full_path,
453                                                         full_path_len,
454                                                         0);
455
456                 if (!packs->list[packs->nr]) {
457                         warning(_("failed to add packfile '%s'"),
458                                 full_path);
459                         return;
460                 }
461
462                 if (open_pack_index(packs->list[packs->nr])) {
463                         warning(_("failed to open pack-index '%s'"),
464                                 full_path);
465                         close_pack(packs->list[packs->nr]);
466                         FREE_AND_NULL(packs->list[packs->nr]);
467                         return;
468                 }
469
470                 packs->names[packs->nr] = xstrdup(file_name);
471                 packs->pack_name_concat_len += strlen(file_name) + 1;
472                 packs->nr++;
473         }
474 }
475
476 struct pack_pair {
477         uint32_t pack_int_id;
478         char *pack_name;
479 };
480
481 static int pack_pair_compare(const void *_a, const void *_b)
482 {
483         struct pack_pair *a = (struct pack_pair *)_a;
484         struct pack_pair *b = (struct pack_pair *)_b;
485         return strcmp(a->pack_name, b->pack_name);
486 }
487
488 static void sort_packs_by_name(char **pack_names, uint32_t nr_packs, uint32_t *perm)
489 {
490         uint32_t i;
491         struct pack_pair *pairs;
492
493         ALLOC_ARRAY(pairs, nr_packs);
494
495         for (i = 0; i < nr_packs; i++) {
496                 pairs[i].pack_int_id = i;
497                 pairs[i].pack_name = pack_names[i];
498         }
499
500         QSORT(pairs, nr_packs, pack_pair_compare);
501
502         for (i = 0; i < nr_packs; i++) {
503                 pack_names[i] = pairs[i].pack_name;
504                 perm[pairs[i].pack_int_id] = i;
505         }
506
507         free(pairs);
508 }
509
510 struct pack_midx_entry {
511         struct object_id oid;
512         uint32_t pack_int_id;
513         time_t pack_mtime;
514         uint64_t offset;
515 };
516
517 static int midx_oid_compare(const void *_a, const void *_b)
518 {
519         const struct pack_midx_entry *a = (const struct pack_midx_entry *)_a;
520         const struct pack_midx_entry *b = (const struct pack_midx_entry *)_b;
521         int cmp = oidcmp(&a->oid, &b->oid);
522
523         if (cmp)
524                 return cmp;
525
526         if (a->pack_mtime > b->pack_mtime)
527                 return -1;
528         else if (a->pack_mtime < b->pack_mtime)
529                 return 1;
530
531         return a->pack_int_id - b->pack_int_id;
532 }
533
534 static int nth_midxed_pack_midx_entry(struct multi_pack_index *m,
535                                       uint32_t *pack_perm,
536                                       struct pack_midx_entry *e,
537                                       uint32_t pos)
538 {
539         if (pos >= m->num_objects)
540                 return 1;
541
542         nth_midxed_object_oid(&e->oid, m, pos);
543         e->pack_int_id = pack_perm[nth_midxed_pack_int_id(m, pos)];
544         e->offset = nth_midxed_offset(m, pos);
545
546         /* consider objects in midx to be from "old" packs */
547         e->pack_mtime = 0;
548         return 0;
549 }
550
551 static void fill_pack_entry(uint32_t pack_int_id,
552                             struct packed_git *p,
553                             uint32_t cur_object,
554                             struct pack_midx_entry *entry)
555 {
556         if (!nth_packed_object_oid(&entry->oid, p, cur_object))
557                 die(_("failed to locate object %d in packfile"), cur_object);
558
559         entry->pack_int_id = pack_int_id;
560         entry->pack_mtime = p->mtime;
561
562         entry->offset = nth_packed_object_offset(p, cur_object);
563 }
564
565 /*
566  * It is possible to artificially get into a state where there are many
567  * duplicate copies of objects. That can create high memory pressure if
568  * we are to create a list of all objects before de-duplication. To reduce
569  * this memory pressure without a significant performance drop, automatically
570  * group objects by the first byte of their object id. Use the IDX fanout
571  * tables to group the data, copy to a local array, then sort.
572  *
573  * Copy only the de-duplicated entries (selected by most-recent modified time
574  * of a packfile containing the object).
575  */
576 static struct pack_midx_entry *get_sorted_entries(struct multi_pack_index *m,
577                                                   struct packed_git **p,
578                                                   uint32_t *perm,
579                                                   uint32_t nr_packs,
580                                                   uint32_t *nr_objects)
581 {
582         uint32_t cur_fanout, cur_pack, cur_object;
583         uint32_t alloc_fanout, alloc_objects, total_objects = 0;
584         struct pack_midx_entry *entries_by_fanout = NULL;
585         struct pack_midx_entry *deduplicated_entries = NULL;
586         uint32_t start_pack = m ? m->num_packs : 0;
587
588         for (cur_pack = start_pack; cur_pack < nr_packs; cur_pack++)
589                 total_objects += p[cur_pack]->num_objects;
590
591         /*
592          * As we de-duplicate by fanout value, we expect the fanout
593          * slices to be evenly distributed, with some noise. Hence,
594          * allocate slightly more than one 256th.
595          */
596         alloc_objects = alloc_fanout = total_objects > 3200 ? total_objects / 200 : 16;
597
598         ALLOC_ARRAY(entries_by_fanout, alloc_fanout);
599         ALLOC_ARRAY(deduplicated_entries, alloc_objects);
600         *nr_objects = 0;
601
602         for (cur_fanout = 0; cur_fanout < 256; cur_fanout++) {
603                 uint32_t nr_fanout = 0;
604
605                 if (m) {
606                         uint32_t start = 0, end;
607
608                         if (cur_fanout)
609                                 start = ntohl(m->chunk_oid_fanout[cur_fanout - 1]);
610                         end = ntohl(m->chunk_oid_fanout[cur_fanout]);
611
612                         for (cur_object = start; cur_object < end; cur_object++) {
613                                 ALLOC_GROW(entries_by_fanout, nr_fanout + 1, alloc_fanout);
614                                 nth_midxed_pack_midx_entry(m, perm,
615                                                            &entries_by_fanout[nr_fanout],
616                                                            cur_object);
617                                 nr_fanout++;
618                         }
619                 }
620
621                 for (cur_pack = start_pack; cur_pack < nr_packs; cur_pack++) {
622                         uint32_t start = 0, end;
623
624                         if (cur_fanout)
625                                 start = get_pack_fanout(p[cur_pack], cur_fanout - 1);
626                         end = get_pack_fanout(p[cur_pack], cur_fanout);
627
628                         for (cur_object = start; cur_object < end; cur_object++) {
629                                 ALLOC_GROW(entries_by_fanout, nr_fanout + 1, alloc_fanout);
630                                 fill_pack_entry(perm[cur_pack], p[cur_pack], cur_object, &entries_by_fanout[nr_fanout]);
631                                 nr_fanout++;
632                         }
633                 }
634
635                 QSORT(entries_by_fanout, nr_fanout, midx_oid_compare);
636
637                 /*
638                  * The batch is now sorted by OID and then mtime (descending).
639                  * Take only the first duplicate.
640                  */
641                 for (cur_object = 0; cur_object < nr_fanout; cur_object++) {
642                         if (cur_object && oideq(&entries_by_fanout[cur_object - 1].oid,
643                                                 &entries_by_fanout[cur_object].oid))
644                                 continue;
645
646                         ALLOC_GROW(deduplicated_entries, *nr_objects + 1, alloc_objects);
647                         memcpy(&deduplicated_entries[*nr_objects],
648                                &entries_by_fanout[cur_object],
649                                sizeof(struct pack_midx_entry));
650                         (*nr_objects)++;
651                 }
652         }
653
654         free(entries_by_fanout);
655         return deduplicated_entries;
656 }
657
658 static size_t write_midx_pack_names(struct hashfile *f,
659                                     char **pack_names,
660                                     uint32_t num_packs)
661 {
662         uint32_t i;
663         unsigned char padding[MIDX_CHUNK_ALIGNMENT];
664         size_t written = 0;
665
666         for (i = 0; i < num_packs; i++) {
667                 size_t writelen = strlen(pack_names[i]) + 1;
668
669                 if (i && strcmp(pack_names[i], pack_names[i - 1]) <= 0)
670                         BUG("incorrect pack-file order: %s before %s",
671                             pack_names[i - 1],
672                             pack_names[i]);
673
674                 hashwrite(f, pack_names[i], writelen);
675                 written += writelen;
676         }
677
678         /* add padding to be aligned */
679         i = MIDX_CHUNK_ALIGNMENT - (written % MIDX_CHUNK_ALIGNMENT);
680         if (i < MIDX_CHUNK_ALIGNMENT) {
681                 memset(padding, 0, sizeof(padding));
682                 hashwrite(f, padding, i);
683                 written += i;
684         }
685
686         return written;
687 }
688
689 static size_t write_midx_oid_fanout(struct hashfile *f,
690                                     struct pack_midx_entry *objects,
691                                     uint32_t nr_objects)
692 {
693         struct pack_midx_entry *list = objects;
694         struct pack_midx_entry *last = objects + nr_objects;
695         uint32_t count = 0;
696         uint32_t i;
697
698         /*
699         * Write the first-level table (the list is sorted,
700         * but we use a 256-entry lookup to be able to avoid
701         * having to do eight extra binary search iterations).
702         */
703         for (i = 0; i < 256; i++) {
704                 struct pack_midx_entry *next = list;
705
706                 while (next < last && next->oid.hash[0] == i) {
707                         count++;
708                         next++;
709                 }
710
711                 hashwrite_be32(f, count);
712                 list = next;
713         }
714
715         return MIDX_CHUNK_FANOUT_SIZE;
716 }
717
718 static size_t write_midx_oid_lookup(struct hashfile *f, unsigned char hash_len,
719                                     struct pack_midx_entry *objects,
720                                     uint32_t nr_objects)
721 {
722         struct pack_midx_entry *list = objects;
723         uint32_t i;
724         size_t written = 0;
725
726         for (i = 0; i < nr_objects; i++) {
727                 struct pack_midx_entry *obj = list++;
728
729                 if (i < nr_objects - 1) {
730                         struct pack_midx_entry *next = list;
731                         if (oidcmp(&obj->oid, &next->oid) >= 0)
732                                 BUG("OIDs not in order: %s >= %s",
733                                     oid_to_hex(&obj->oid),
734                                     oid_to_hex(&next->oid));
735                 }
736
737                 hashwrite(f, obj->oid.hash, (int)hash_len);
738                 written += hash_len;
739         }
740
741         return written;
742 }
743
744 static size_t write_midx_object_offsets(struct hashfile *f, int large_offset_needed,
745                                         struct pack_midx_entry *objects, uint32_t nr_objects)
746 {
747         struct pack_midx_entry *list = objects;
748         uint32_t i, nr_large_offset = 0;
749         size_t written = 0;
750
751         for (i = 0; i < nr_objects; i++) {
752                 struct pack_midx_entry *obj = list++;
753
754                 hashwrite_be32(f, obj->pack_int_id);
755
756                 if (large_offset_needed && obj->offset >> 31)
757                         hashwrite_be32(f, MIDX_LARGE_OFFSET_NEEDED | nr_large_offset++);
758                 else if (!large_offset_needed && obj->offset >> 32)
759                         BUG("object %s requires a large offset (%"PRIx64") but the MIDX is not writing large offsets!",
760                             oid_to_hex(&obj->oid),
761                             obj->offset);
762                 else
763                         hashwrite_be32(f, (uint32_t)obj->offset);
764
765                 written += MIDX_CHUNK_OFFSET_WIDTH;
766         }
767
768         return written;
769 }
770
771 static size_t write_midx_large_offsets(struct hashfile *f, uint32_t nr_large_offset,
772                                        struct pack_midx_entry *objects, uint32_t nr_objects)
773 {
774         struct pack_midx_entry *list = objects, *end = objects + nr_objects;
775         size_t written = 0;
776
777         while (nr_large_offset) {
778                 struct pack_midx_entry *obj;
779                 uint64_t offset;
780
781                 if (list >= end)
782                         BUG("too many large-offset objects");
783
784                 obj = list++;
785                 offset = obj->offset;
786
787                 if (!(offset >> 31))
788                         continue;
789
790                 hashwrite_be32(f, offset >> 32);
791                 hashwrite_be32(f, offset & 0xffffffffUL);
792                 written += 2 * sizeof(uint32_t);
793
794                 nr_large_offset--;
795         }
796
797         return written;
798 }
799
800 int write_midx_file(const char *object_dir)
801 {
802         unsigned char cur_chunk, num_chunks = 0;
803         char *midx_name;
804         uint32_t i;
805         struct hashfile *f = NULL;
806         struct lock_file lk;
807         struct pack_list packs;
808         uint32_t *pack_perm = NULL;
809         uint64_t written = 0;
810         uint32_t chunk_ids[MIDX_MAX_CHUNKS + 1];
811         uint64_t chunk_offsets[MIDX_MAX_CHUNKS + 1];
812         uint32_t nr_entries, num_large_offsets = 0;
813         struct pack_midx_entry *entries = NULL;
814         int large_offsets_needed = 0;
815
816         midx_name = get_midx_filename(object_dir);
817         if (safe_create_leading_directories(midx_name)) {
818                 UNLEAK(midx_name);
819                 die_errno(_("unable to create leading directories of %s"),
820                           midx_name);
821         }
822
823         packs.m = load_multi_pack_index(object_dir, 1);
824
825         packs.nr = 0;
826         packs.alloc_list = packs.m ? packs.m->num_packs : 16;
827         packs.alloc_names = packs.alloc_list;
828         packs.list = NULL;
829         packs.names = NULL;
830         packs.pack_name_concat_len = 0;
831         ALLOC_ARRAY(packs.list, packs.alloc_list);
832         ALLOC_ARRAY(packs.names, packs.alloc_names);
833
834         if (packs.m) {
835                 for (i = 0; i < packs.m->num_packs; i++) {
836                         ALLOC_GROW(packs.list, packs.nr + 1, packs.alloc_list);
837                         ALLOC_GROW(packs.names, packs.nr + 1, packs.alloc_names);
838
839                         packs.list[packs.nr] = NULL;
840                         packs.names[packs.nr] = xstrdup(packs.m->pack_names[i]);
841                         packs.pack_name_concat_len += strlen(packs.names[packs.nr]) + 1;
842                         packs.nr++;
843                 }
844         }
845
846         for_each_file_in_pack_dir(object_dir, add_pack_to_midx, &packs);
847
848         if (packs.m && packs.nr == packs.m->num_packs)
849                 goto cleanup;
850
851         if (packs.pack_name_concat_len % MIDX_CHUNK_ALIGNMENT)
852                 packs.pack_name_concat_len += MIDX_CHUNK_ALIGNMENT -
853                                               (packs.pack_name_concat_len % MIDX_CHUNK_ALIGNMENT);
854
855         ALLOC_ARRAY(pack_perm, packs.nr);
856         sort_packs_by_name(packs.names, packs.nr, pack_perm);
857
858         entries = get_sorted_entries(packs.m, packs.list, pack_perm, packs.nr, &nr_entries);
859
860         for (i = 0; i < nr_entries; i++) {
861                 if (entries[i].offset > 0x7fffffff)
862                         num_large_offsets++;
863                 if (entries[i].offset > 0xffffffff)
864                         large_offsets_needed = 1;
865         }
866
867         hold_lock_file_for_update(&lk, midx_name, LOCK_DIE_ON_ERROR);
868         f = hashfd(lk.tempfile->fd, lk.tempfile->filename.buf);
869         FREE_AND_NULL(midx_name);
870
871         if (packs.m)
872                 close_midx(packs.m);
873
874         cur_chunk = 0;
875         num_chunks = large_offsets_needed ? 5 : 4;
876
877         written = write_midx_header(f, num_chunks, packs.nr);
878
879         chunk_ids[cur_chunk] = MIDX_CHUNKID_PACKNAMES;
880         chunk_offsets[cur_chunk] = written + (num_chunks + 1) * MIDX_CHUNKLOOKUP_WIDTH;
881
882         cur_chunk++;
883         chunk_ids[cur_chunk] = MIDX_CHUNKID_OIDFANOUT;
884         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + packs.pack_name_concat_len;
885
886         cur_chunk++;
887         chunk_ids[cur_chunk] = MIDX_CHUNKID_OIDLOOKUP;
888         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + MIDX_CHUNK_FANOUT_SIZE;
889
890         cur_chunk++;
891         chunk_ids[cur_chunk] = MIDX_CHUNKID_OBJECTOFFSETS;
892         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + nr_entries * MIDX_HASH_LEN;
893
894         cur_chunk++;
895         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + nr_entries * MIDX_CHUNK_OFFSET_WIDTH;
896         if (large_offsets_needed) {
897                 chunk_ids[cur_chunk] = MIDX_CHUNKID_LARGEOFFSETS;
898
899                 cur_chunk++;
900                 chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] +
901                                            num_large_offsets * MIDX_CHUNK_LARGE_OFFSET_WIDTH;
902         }
903
904         chunk_ids[cur_chunk] = 0;
905
906         for (i = 0; i <= num_chunks; i++) {
907                 if (i && chunk_offsets[i] < chunk_offsets[i - 1])
908                         BUG("incorrect chunk offsets: %"PRIu64" before %"PRIu64,
909                             chunk_offsets[i - 1],
910                             chunk_offsets[i]);
911
912                 if (chunk_offsets[i] % MIDX_CHUNK_ALIGNMENT)
913                         BUG("chunk offset %"PRIu64" is not properly aligned",
914                             chunk_offsets[i]);
915
916                 hashwrite_be32(f, chunk_ids[i]);
917                 hashwrite_be32(f, chunk_offsets[i] >> 32);
918                 hashwrite_be32(f, chunk_offsets[i]);
919
920                 written += MIDX_CHUNKLOOKUP_WIDTH;
921         }
922
923         for (i = 0; i < num_chunks; i++) {
924                 if (written != chunk_offsets[i])
925                         BUG("incorrect chunk offset (%"PRIu64" != %"PRIu64") for chunk id %"PRIx32,
926                             chunk_offsets[i],
927                             written,
928                             chunk_ids[i]);
929
930                 switch (chunk_ids[i]) {
931                         case MIDX_CHUNKID_PACKNAMES:
932                                 written += write_midx_pack_names(f, packs.names, packs.nr);
933                                 break;
934
935                         case MIDX_CHUNKID_OIDFANOUT:
936                                 written += write_midx_oid_fanout(f, entries, nr_entries);
937                                 break;
938
939                         case MIDX_CHUNKID_OIDLOOKUP:
940                                 written += write_midx_oid_lookup(f, MIDX_HASH_LEN, entries, nr_entries);
941                                 break;
942
943                         case MIDX_CHUNKID_OBJECTOFFSETS:
944                                 written += write_midx_object_offsets(f, large_offsets_needed, entries, nr_entries);
945                                 break;
946
947                         case MIDX_CHUNKID_LARGEOFFSETS:
948                                 written += write_midx_large_offsets(f, num_large_offsets, entries, nr_entries);
949                                 break;
950
951                         default:
952                                 BUG("trying to write unknown chunk id %"PRIx32,
953                                     chunk_ids[i]);
954                 }
955         }
956
957         if (written != chunk_offsets[num_chunks])
958                 BUG("incorrect final offset %"PRIu64" != %"PRIu64,
959                     written,
960                     chunk_offsets[num_chunks]);
961
962         finalize_hashfile(f, NULL, CSUM_FSYNC | CSUM_HASH_IN_STREAM);
963         commit_lock_file(&lk);
964
965 cleanup:
966         for (i = 0; i < packs.nr; i++) {
967                 if (packs.list[i]) {
968                         close_pack(packs.list[i]);
969                         free(packs.list[i]);
970                 }
971                 free(packs.names[i]);
972         }
973
974         free(packs.list);
975         free(packs.names);
976         free(entries);
977         free(pack_perm);
978         free(midx_name);
979         return 0;
980 }
981
982 void clear_midx_file(struct repository *r)
983 {
984         char *midx = get_midx_filename(r->objects->odb->path);
985
986         if (r->objects && r->objects->multi_pack_index) {
987                 close_midx(r->objects->multi_pack_index);
988                 r->objects->multi_pack_index = NULL;
989         }
990
991         if (remove_path(midx)) {
992                 UNLEAK(midx);
993                 die(_("failed to clear multi-pack-index at %s"), midx);
994         }
995
996         free(midx);
997 }
998
999 static int verify_midx_error;
1000
1001 static void midx_report(const char *fmt, ...)
1002 {
1003         va_list ap;
1004         verify_midx_error = 1;
1005         va_start(ap, fmt);
1006         vfprintf(stderr, fmt, ap);
1007         fprintf(stderr, "\n");
1008         va_end(ap);
1009 }
1010
1011 struct pair_pos_vs_id
1012 {
1013         uint32_t pos;
1014         uint32_t pack_int_id;
1015 };
1016
1017 static int compare_pair_pos_vs_id(const void *_a, const void *_b)
1018 {
1019         struct pair_pos_vs_id *a = (struct pair_pos_vs_id *)_a;
1020         struct pair_pos_vs_id *b = (struct pair_pos_vs_id *)_b;
1021
1022         return b->pack_int_id - a->pack_int_id;
1023 }
1024
1025 /*
1026  * Limit calls to display_progress() for performance reasons.
1027  * The interval here was arbitrarily chosen.
1028  */
1029 #define SPARSE_PROGRESS_INTERVAL (1 << 12)
1030 #define midx_display_sparse_progress(progress, n) \
1031         do { \
1032                 uint64_t _n = (n); \
1033                 if ((_n & (SPARSE_PROGRESS_INTERVAL - 1)) == 0) \
1034                         display_progress(progress, _n); \
1035         } while (0)
1036
1037 int verify_midx_file(struct repository *r, const char *object_dir)
1038 {
1039         struct pair_pos_vs_id *pairs = NULL;
1040         uint32_t i;
1041         struct progress *progress;
1042         struct multi_pack_index *m = load_multi_pack_index(object_dir, 1);
1043         verify_midx_error = 0;
1044
1045         if (!m)
1046                 return 0;
1047
1048         progress = start_progress(_("Looking for referenced packfiles"),
1049                                   m->num_packs);
1050         for (i = 0; i < m->num_packs; i++) {
1051                 if (prepare_midx_pack(r, m, i))
1052                         midx_report("failed to load pack in position %d", i);
1053
1054                 display_progress(progress, i + 1);
1055         }
1056         stop_progress(&progress);
1057
1058         for (i = 0; i < 255; i++) {
1059                 uint32_t oid_fanout1 = ntohl(m->chunk_oid_fanout[i]);
1060                 uint32_t oid_fanout2 = ntohl(m->chunk_oid_fanout[i + 1]);
1061
1062                 if (oid_fanout1 > oid_fanout2)
1063                         midx_report(_("oid fanout out of order: fanout[%d] = %"PRIx32" > %"PRIx32" = fanout[%d]"),
1064                                     i, oid_fanout1, oid_fanout2, i + 1);
1065         }
1066
1067         progress = start_sparse_progress(_("Verifying OID order in MIDX"),
1068                                          m->num_objects - 1);
1069         for (i = 0; i < m->num_objects - 1; i++) {
1070                 struct object_id oid1, oid2;
1071
1072                 nth_midxed_object_oid(&oid1, m, i);
1073                 nth_midxed_object_oid(&oid2, m, i + 1);
1074
1075                 if (oidcmp(&oid1, &oid2) >= 0)
1076                         midx_report(_("oid lookup out of order: oid[%d] = %s >= %s = oid[%d]"),
1077                                     i, oid_to_hex(&oid1), oid_to_hex(&oid2), i + 1);
1078
1079                 midx_display_sparse_progress(progress, i + 1);
1080         }
1081         stop_progress(&progress);
1082
1083         /*
1084          * Create an array mapping each object to its packfile id.  Sort it
1085          * to group the objects by packfile.  Use this permutation to visit
1086          * each of the objects and only require 1 packfile to be open at a
1087          * time.
1088          */
1089         ALLOC_ARRAY(pairs, m->num_objects);
1090         for (i = 0; i < m->num_objects; i++) {
1091                 pairs[i].pos = i;
1092                 pairs[i].pack_int_id = nth_midxed_pack_int_id(m, i);
1093         }
1094
1095         progress = start_sparse_progress(_("Sorting objects by packfile"),
1096                                          m->num_objects);
1097         display_progress(progress, 0); /* TODO: Measure QSORT() progress */
1098         QSORT(pairs, m->num_objects, compare_pair_pos_vs_id);
1099         stop_progress(&progress);
1100
1101         progress = start_sparse_progress(_("Verifying object offsets"), m->num_objects);
1102         for (i = 0; i < m->num_objects; i++) {
1103                 struct object_id oid;
1104                 struct pack_entry e;
1105                 off_t m_offset, p_offset;
1106
1107                 if (i > 0 && pairs[i-1].pack_int_id != pairs[i].pack_int_id &&
1108                     m->packs[pairs[i-1].pack_int_id])
1109                 {
1110                         close_pack_fd(m->packs[pairs[i-1].pack_int_id]);
1111                         close_pack_index(m->packs[pairs[i-1].pack_int_id]);
1112                 }
1113
1114                 nth_midxed_object_oid(&oid, m, pairs[i].pos);
1115
1116                 if (!fill_midx_entry(r, &oid, &e, m)) {
1117                         midx_report(_("failed to load pack entry for oid[%d] = %s"),
1118                                     pairs[i].pos, oid_to_hex(&oid));
1119                         continue;
1120                 }
1121
1122                 if (open_pack_index(e.p)) {
1123                         midx_report(_("failed to load pack-index for packfile %s"),
1124                                     e.p->pack_name);
1125                         break;
1126                 }
1127
1128                 m_offset = e.offset;
1129                 p_offset = find_pack_entry_one(oid.hash, e.p);
1130
1131                 if (m_offset != p_offset)
1132                         midx_report(_("incorrect object offset for oid[%d] = %s: %"PRIx64" != %"PRIx64),
1133                                     pairs[i].pos, oid_to_hex(&oid), m_offset, p_offset);
1134
1135                 midx_display_sparse_progress(progress, i + 1);
1136         }
1137         stop_progress(&progress);
1138
1139         free(pairs);
1140
1141         return verify_midx_error;
1142 }