format-patch: tie file-opening logic to output_directory
[git] / builtin / index-pack.c
1 #include "builtin.h"
2 #include "config.h"
3 #include "delta.h"
4 #include "pack.h"
5 #include "csum-file.h"
6 #include "blob.h"
7 #include "commit.h"
8 #include "tag.h"
9 #include "tree.h"
10 #include "progress.h"
11 #include "fsck.h"
12 #include "exec-cmd.h"
13 #include "streaming.h"
14 #include "thread-utils.h"
15 #include "packfile.h"
16 #include "object-store.h"
17 #include "promisor-remote.h"
18
19 static const char index_pack_usage[] =
20 "git index-pack [-v] [-o <index-file>] [--keep | --keep=<msg>] [--verify] [--strict] (<pack-file> | --stdin [--fix-thin] [<pack-file>])";
21
22 struct object_entry {
23         struct pack_idx_entry idx;
24         unsigned long size;
25         unsigned char hdr_size;
26         signed char type;
27         signed char real_type;
28 };
29
30 struct object_stat {
31         unsigned delta_depth;
32         int base_object_no;
33 };
34
35 struct base_data {
36         /* Initialized by make_base(). */
37         struct base_data *base;
38         struct object_entry *obj;
39         int ref_first, ref_last;
40         int ofs_first, ofs_last;
41         /*
42          * Threads should increment retain_data if they are about to call
43          * patch_delta() using this struct's data as a base, and decrement this
44          * when they are done. While retain_data is nonzero, this struct's data
45          * will not be freed even if the delta base cache limit is exceeded.
46          */
47         int retain_data;
48         /*
49          * The number of direct children that have not been fully processed
50          * (entered work_head, entered done_head, left done_head). When this
51          * number reaches zero, this struct base_data can be freed.
52          */
53         int children_remaining;
54
55         /* Not initialized by make_base(). */
56         struct list_head list;
57         void *data;
58         unsigned long size;
59 };
60
61 /*
62  * Stack of struct base_data that have unprocessed children.
63  * threaded_second_pass() uses this as a source of work (the other being the
64  * objects array).
65  *
66  * Guarded by work_mutex.
67  */
68 static LIST_HEAD(work_head);
69
70 /*
71  * Stack of struct base_data that have children, all of whom have been
72  * processed or are being processed, and at least one child is being processed.
73  * These struct base_data must be kept around until the last child is
74  * processed.
75  *
76  * Guarded by work_mutex.
77  */
78 static LIST_HEAD(done_head);
79
80 /*
81  * All threads share one delta base cache.
82  *
83  * base_cache_used is guarded by work_mutex, and base_cache_limit is read-only
84  * in a thread.
85  */
86 static size_t base_cache_used;
87 static size_t base_cache_limit;
88
89 struct thread_local {
90         pthread_t thread;
91         int pack_fd;
92 };
93
94 /* Remember to update object flag allocation in object.h */
95 #define FLAG_LINK (1u<<20)
96 #define FLAG_CHECKED (1u<<21)
97
98 struct ofs_delta_entry {
99         off_t offset;
100         int obj_no;
101 };
102
103 struct ref_delta_entry {
104         struct object_id oid;
105         int obj_no;
106 };
107
108 static struct object_entry *objects;
109 static struct object_stat *obj_stat;
110 static struct ofs_delta_entry *ofs_deltas;
111 static struct ref_delta_entry *ref_deltas;
112 static struct thread_local nothread_data;
113 static int nr_objects;
114 static int nr_ofs_deltas;
115 static int nr_ref_deltas;
116 static int ref_deltas_alloc;
117 static int nr_resolved_deltas;
118 static int nr_threads;
119
120 static int from_stdin;
121 static int strict;
122 static int do_fsck_object;
123 static struct fsck_options fsck_options = FSCK_OPTIONS_STRICT;
124 static int verbose;
125 static int show_resolving_progress;
126 static int show_stat;
127 static int check_self_contained_and_connected;
128
129 static struct progress *progress;
130
131 /* We always read in 4kB chunks. */
132 static unsigned char input_buffer[4096];
133 static unsigned int input_offset, input_len;
134 static off_t consumed_bytes;
135 static off_t max_input_size;
136 static unsigned deepest_delta;
137 static git_hash_ctx input_ctx;
138 static uint32_t input_crc32;
139 static int input_fd, output_fd;
140 static const char *curr_pack;
141
142 static struct thread_local *thread_data;
143 static int nr_dispatched;
144 static int threads_active;
145
146 static pthread_mutex_t read_mutex;
147 #define read_lock()             lock_mutex(&read_mutex)
148 #define read_unlock()           unlock_mutex(&read_mutex)
149
150 static pthread_mutex_t counter_mutex;
151 #define counter_lock()          lock_mutex(&counter_mutex)
152 #define counter_unlock()        unlock_mutex(&counter_mutex)
153
154 static pthread_mutex_t work_mutex;
155 #define work_lock()             lock_mutex(&work_mutex)
156 #define work_unlock()           unlock_mutex(&work_mutex)
157
158 static pthread_mutex_t deepest_delta_mutex;
159 #define deepest_delta_lock()    lock_mutex(&deepest_delta_mutex)
160 #define deepest_delta_unlock()  unlock_mutex(&deepest_delta_mutex)
161
162 static pthread_key_t key;
163
164 static inline void lock_mutex(pthread_mutex_t *mutex)
165 {
166         if (threads_active)
167                 pthread_mutex_lock(mutex);
168 }
169
170 static inline void unlock_mutex(pthread_mutex_t *mutex)
171 {
172         if (threads_active)
173                 pthread_mutex_unlock(mutex);
174 }
175
176 /*
177  * Mutex and conditional variable can't be statically-initialized on Windows.
178  */
179 static void init_thread(void)
180 {
181         int i;
182         init_recursive_mutex(&read_mutex);
183         pthread_mutex_init(&counter_mutex, NULL);
184         pthread_mutex_init(&work_mutex, NULL);
185         if (show_stat)
186                 pthread_mutex_init(&deepest_delta_mutex, NULL);
187         pthread_key_create(&key, NULL);
188         thread_data = xcalloc(nr_threads, sizeof(*thread_data));
189         for (i = 0; i < nr_threads; i++) {
190                 thread_data[i].pack_fd = open(curr_pack, O_RDONLY);
191                 if (thread_data[i].pack_fd == -1)
192                         die_errno(_("unable to open %s"), curr_pack);
193         }
194
195         threads_active = 1;
196 }
197
198 static void cleanup_thread(void)
199 {
200         int i;
201         if (!threads_active)
202                 return;
203         threads_active = 0;
204         pthread_mutex_destroy(&read_mutex);
205         pthread_mutex_destroy(&counter_mutex);
206         pthread_mutex_destroy(&work_mutex);
207         if (show_stat)
208                 pthread_mutex_destroy(&deepest_delta_mutex);
209         for (i = 0; i < nr_threads; i++)
210                 close(thread_data[i].pack_fd);
211         pthread_key_delete(key);
212         free(thread_data);
213 }
214
215 static int mark_link(struct object *obj, int type, void *data, struct fsck_options *options)
216 {
217         if (!obj)
218                 return -1;
219
220         if (type != OBJ_ANY && obj->type != type)
221                 die(_("object type mismatch at %s"), oid_to_hex(&obj->oid));
222
223         obj->flags |= FLAG_LINK;
224         return 0;
225 }
226
227 /* The content of each linked object must have been checked
228    or it must be already present in the object database */
229 static unsigned check_object(struct object *obj)
230 {
231         if (!obj)
232                 return 0;
233
234         if (!(obj->flags & FLAG_LINK))
235                 return 0;
236
237         if (!(obj->flags & FLAG_CHECKED)) {
238                 unsigned long size;
239                 int type = oid_object_info(the_repository, &obj->oid, &size);
240                 if (type <= 0)
241                         die(_("did not receive expected object %s"),
242                               oid_to_hex(&obj->oid));
243                 if (type != obj->type)
244                         die(_("object %s: expected type %s, found %s"),
245                             oid_to_hex(&obj->oid),
246                             type_name(obj->type), type_name(type));
247                 obj->flags |= FLAG_CHECKED;
248                 return 1;
249         }
250
251         return 0;
252 }
253
254 static unsigned check_objects(void)
255 {
256         unsigned i, max, foreign_nr = 0;
257
258         max = get_max_object_index();
259
260         if (verbose)
261                 progress = start_delayed_progress(_("Checking objects"), max);
262
263         for (i = 0; i < max; i++) {
264                 foreign_nr += check_object(get_indexed_object(i));
265                 display_progress(progress, i + 1);
266         }
267
268         stop_progress(&progress);
269         return foreign_nr;
270 }
271
272
273 /* Discard current buffer used content. */
274 static void flush(void)
275 {
276         if (input_offset) {
277                 if (output_fd >= 0)
278                         write_or_die(output_fd, input_buffer, input_offset);
279                 the_hash_algo->update_fn(&input_ctx, input_buffer, input_offset);
280                 memmove(input_buffer, input_buffer + input_offset, input_len);
281                 input_offset = 0;
282         }
283 }
284
285 /*
286  * Make sure at least "min" bytes are available in the buffer, and
287  * return the pointer to the buffer.
288  */
289 static void *fill(int min)
290 {
291         if (min <= input_len)
292                 return input_buffer + input_offset;
293         if (min > sizeof(input_buffer))
294                 die(Q_("cannot fill %d byte",
295                        "cannot fill %d bytes",
296                        min),
297                     min);
298         flush();
299         do {
300                 ssize_t ret = xread(input_fd, input_buffer + input_len,
301                                 sizeof(input_buffer) - input_len);
302                 if (ret <= 0) {
303                         if (!ret)
304                                 die(_("early EOF"));
305                         die_errno(_("read error on input"));
306                 }
307                 input_len += ret;
308                 if (from_stdin)
309                         display_throughput(progress, consumed_bytes + input_len);
310         } while (input_len < min);
311         return input_buffer;
312 }
313
314 static void use(int bytes)
315 {
316         if (bytes > input_len)
317                 die(_("used more bytes than were available"));
318         input_crc32 = crc32(input_crc32, input_buffer + input_offset, bytes);
319         input_len -= bytes;
320         input_offset += bytes;
321
322         /* make sure off_t is sufficiently large not to wrap */
323         if (signed_add_overflows(consumed_bytes, bytes))
324                 die(_("pack too large for current definition of off_t"));
325         consumed_bytes += bytes;
326         if (max_input_size && consumed_bytes > max_input_size)
327                 die(_("pack exceeds maximum allowed size"));
328 }
329
330 static const char *open_pack_file(const char *pack_name)
331 {
332         if (from_stdin) {
333                 input_fd = 0;
334                 if (!pack_name) {
335                         struct strbuf tmp_file = STRBUF_INIT;
336                         output_fd = odb_mkstemp(&tmp_file,
337                                                 "pack/tmp_pack_XXXXXX");
338                         pack_name = strbuf_detach(&tmp_file, NULL);
339                 } else {
340                         output_fd = open(pack_name, O_CREAT|O_EXCL|O_RDWR, 0600);
341                         if (output_fd < 0)
342                                 die_errno(_("unable to create '%s'"), pack_name);
343                 }
344                 nothread_data.pack_fd = output_fd;
345         } else {
346                 input_fd = open(pack_name, O_RDONLY);
347                 if (input_fd < 0)
348                         die_errno(_("cannot open packfile '%s'"), pack_name);
349                 output_fd = -1;
350                 nothread_data.pack_fd = input_fd;
351         }
352         the_hash_algo->init_fn(&input_ctx);
353         return pack_name;
354 }
355
356 static void parse_pack_header(void)
357 {
358         struct pack_header *hdr = fill(sizeof(struct pack_header));
359
360         /* Header consistency check */
361         if (hdr->hdr_signature != htonl(PACK_SIGNATURE))
362                 die(_("pack signature mismatch"));
363         if (!pack_version_ok(hdr->hdr_version))
364                 die(_("pack version %"PRIu32" unsupported"),
365                         ntohl(hdr->hdr_version));
366
367         nr_objects = ntohl(hdr->hdr_entries);
368         use(sizeof(struct pack_header));
369 }
370
371 static NORETURN void bad_object(off_t offset, const char *format,
372                        ...) __attribute__((format (printf, 2, 3)));
373
374 static NORETURN void bad_object(off_t offset, const char *format, ...)
375 {
376         va_list params;
377         char buf[1024];
378
379         va_start(params, format);
380         vsnprintf(buf, sizeof(buf), format, params);
381         va_end(params);
382         die(_("pack has bad object at offset %"PRIuMAX": %s"),
383             (uintmax_t)offset, buf);
384 }
385
386 static inline struct thread_local *get_thread_data(void)
387 {
388         if (HAVE_THREADS) {
389                 if (threads_active)
390                         return pthread_getspecific(key);
391                 assert(!threads_active &&
392                        "This should only be reached when all threads are gone");
393         }
394         return &nothread_data;
395 }
396
397 static void set_thread_data(struct thread_local *data)
398 {
399         if (threads_active)
400                 pthread_setspecific(key, data);
401 }
402
403 static void free_base_data(struct base_data *c)
404 {
405         if (c->data) {
406                 FREE_AND_NULL(c->data);
407                 base_cache_used -= c->size;
408         }
409 }
410
411 static void prune_base_data(struct base_data *retain)
412 {
413         struct list_head *pos;
414
415         if (base_cache_used <= base_cache_limit)
416                 return;
417
418         list_for_each_prev(pos, &done_head) {
419                 struct base_data *b = list_entry(pos, struct base_data, list);
420                 if (b->retain_data || b == retain)
421                         continue;
422                 if (b->data) {
423                         free_base_data(b);
424                         if (base_cache_used <= base_cache_limit)
425                                 return;
426                 }
427         }
428
429         list_for_each_prev(pos, &work_head) {
430                 struct base_data *b = list_entry(pos, struct base_data, list);
431                 if (b->retain_data || b == retain)
432                         continue;
433                 if (b->data) {
434                         free_base_data(b);
435                         if (base_cache_used <= base_cache_limit)
436                                 return;
437                 }
438         }
439 }
440
441 static int is_delta_type(enum object_type type)
442 {
443         return (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA);
444 }
445
446 static void *unpack_entry_data(off_t offset, unsigned long size,
447                                enum object_type type, struct object_id *oid)
448 {
449         static char fixed_buf[8192];
450         int status;
451         git_zstream stream;
452         void *buf;
453         git_hash_ctx c;
454         char hdr[32];
455         int hdrlen;
456
457         if (!is_delta_type(type)) {
458                 hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %"PRIuMAX,
459                                    type_name(type),(uintmax_t)size) + 1;
460                 the_hash_algo->init_fn(&c);
461                 the_hash_algo->update_fn(&c, hdr, hdrlen);
462         } else
463                 oid = NULL;
464         if (type == OBJ_BLOB && size > big_file_threshold)
465                 buf = fixed_buf;
466         else
467                 buf = xmallocz(size);
468
469         memset(&stream, 0, sizeof(stream));
470         git_inflate_init(&stream);
471         stream.next_out = buf;
472         stream.avail_out = buf == fixed_buf ? sizeof(fixed_buf) : size;
473
474         do {
475                 unsigned char *last_out = stream.next_out;
476                 stream.next_in = fill(1);
477                 stream.avail_in = input_len;
478                 status = git_inflate(&stream, 0);
479                 use(input_len - stream.avail_in);
480                 if (oid)
481                         the_hash_algo->update_fn(&c, last_out, stream.next_out - last_out);
482                 if (buf == fixed_buf) {
483                         stream.next_out = buf;
484                         stream.avail_out = sizeof(fixed_buf);
485                 }
486         } while (status == Z_OK);
487         if (stream.total_out != size || status != Z_STREAM_END)
488                 bad_object(offset, _("inflate returned %d"), status);
489         git_inflate_end(&stream);
490         if (oid)
491                 the_hash_algo->final_fn(oid->hash, &c);
492         return buf == fixed_buf ? NULL : buf;
493 }
494
495 static void *unpack_raw_entry(struct object_entry *obj,
496                               off_t *ofs_offset,
497                               struct object_id *ref_oid,
498                               struct object_id *oid)
499 {
500         unsigned char *p;
501         unsigned long size, c;
502         off_t base_offset;
503         unsigned shift;
504         void *data;
505
506         obj->idx.offset = consumed_bytes;
507         input_crc32 = crc32(0, NULL, 0);
508
509         p = fill(1);
510         c = *p;
511         use(1);
512         obj->type = (c >> 4) & 7;
513         size = (c & 15);
514         shift = 4;
515         while (c & 0x80) {
516                 p = fill(1);
517                 c = *p;
518                 use(1);
519                 size += (c & 0x7f) << shift;
520                 shift += 7;
521         }
522         obj->size = size;
523
524         switch (obj->type) {
525         case OBJ_REF_DELTA:
526                 hashcpy(ref_oid->hash, fill(the_hash_algo->rawsz));
527                 use(the_hash_algo->rawsz);
528                 break;
529         case OBJ_OFS_DELTA:
530                 p = fill(1);
531                 c = *p;
532                 use(1);
533                 base_offset = c & 127;
534                 while (c & 128) {
535                         base_offset += 1;
536                         if (!base_offset || MSB(base_offset, 7))
537                                 bad_object(obj->idx.offset, _("offset value overflow for delta base object"));
538                         p = fill(1);
539                         c = *p;
540                         use(1);
541                         base_offset = (base_offset << 7) + (c & 127);
542                 }
543                 *ofs_offset = obj->idx.offset - base_offset;
544                 if (*ofs_offset <= 0 || *ofs_offset >= obj->idx.offset)
545                         bad_object(obj->idx.offset, _("delta base offset is out of bound"));
546                 break;
547         case OBJ_COMMIT:
548         case OBJ_TREE:
549         case OBJ_BLOB:
550         case OBJ_TAG:
551                 break;
552         default:
553                 bad_object(obj->idx.offset, _("unknown object type %d"), obj->type);
554         }
555         obj->hdr_size = consumed_bytes - obj->idx.offset;
556
557         data = unpack_entry_data(obj->idx.offset, obj->size, obj->type, oid);
558         obj->idx.crc32 = input_crc32;
559         return data;
560 }
561
562 static void *unpack_data(struct object_entry *obj,
563                          int (*consume)(const unsigned char *, unsigned long, void *),
564                          void *cb_data)
565 {
566         off_t from = obj[0].idx.offset + obj[0].hdr_size;
567         off_t len = obj[1].idx.offset - from;
568         unsigned char *data, *inbuf;
569         git_zstream stream;
570         int status;
571
572         data = xmallocz(consume ? 64*1024 : obj->size);
573         inbuf = xmalloc((len < 64*1024) ? (int)len : 64*1024);
574
575         memset(&stream, 0, sizeof(stream));
576         git_inflate_init(&stream);
577         stream.next_out = data;
578         stream.avail_out = consume ? 64*1024 : obj->size;
579
580         do {
581                 ssize_t n = (len < 64*1024) ? (ssize_t)len : 64*1024;
582                 n = xpread(get_thread_data()->pack_fd, inbuf, n, from);
583                 if (n < 0)
584                         die_errno(_("cannot pread pack file"));
585                 if (!n)
586                         die(Q_("premature end of pack file, %"PRIuMAX" byte missing",
587                                "premature end of pack file, %"PRIuMAX" bytes missing",
588                                (unsigned int)len),
589                             (uintmax_t)len);
590                 from += n;
591                 len -= n;
592                 stream.next_in = inbuf;
593                 stream.avail_in = n;
594                 if (!consume)
595                         status = git_inflate(&stream, 0);
596                 else {
597                         do {
598                                 status = git_inflate(&stream, 0);
599                                 if (consume(data, stream.next_out - data, cb_data)) {
600                                         free(inbuf);
601                                         free(data);
602                                         return NULL;
603                                 }
604                                 stream.next_out = data;
605                                 stream.avail_out = 64*1024;
606                         } while (status == Z_OK && stream.avail_in);
607                 }
608         } while (len && status == Z_OK && !stream.avail_in);
609
610         /* This has been inflated OK when first encountered, so... */
611         if (status != Z_STREAM_END || stream.total_out != obj->size)
612                 die(_("serious inflate inconsistency"));
613
614         git_inflate_end(&stream);
615         free(inbuf);
616         if (consume) {
617                 FREE_AND_NULL(data);
618         }
619         return data;
620 }
621
622 static void *get_data_from_pack(struct object_entry *obj)
623 {
624         return unpack_data(obj, NULL, NULL);
625 }
626
627 static int compare_ofs_delta_bases(off_t offset1, off_t offset2,
628                                    enum object_type type1,
629                                    enum object_type type2)
630 {
631         int cmp = type1 - type2;
632         if (cmp)
633                 return cmp;
634         return offset1 < offset2 ? -1 :
635                offset1 > offset2 ?  1 :
636                0;
637 }
638
639 static int find_ofs_delta(const off_t offset)
640 {
641         int first = 0, last = nr_ofs_deltas;
642
643         while (first < last) {
644                 int next = first + (last - first) / 2;
645                 struct ofs_delta_entry *delta = &ofs_deltas[next];
646                 int cmp;
647
648                 cmp = compare_ofs_delta_bases(offset, delta->offset,
649                                               OBJ_OFS_DELTA,
650                                               objects[delta->obj_no].type);
651                 if (!cmp)
652                         return next;
653                 if (cmp < 0) {
654                         last = next;
655                         continue;
656                 }
657                 first = next+1;
658         }
659         return -first-1;
660 }
661
662 static void find_ofs_delta_children(off_t offset,
663                                     int *first_index, int *last_index)
664 {
665         int first = find_ofs_delta(offset);
666         int last = first;
667         int end = nr_ofs_deltas - 1;
668
669         if (first < 0) {
670                 *first_index = 0;
671                 *last_index = -1;
672                 return;
673         }
674         while (first > 0 && ofs_deltas[first - 1].offset == offset)
675                 --first;
676         while (last < end && ofs_deltas[last + 1].offset == offset)
677                 ++last;
678         *first_index = first;
679         *last_index = last;
680 }
681
682 static int compare_ref_delta_bases(const struct object_id *oid1,
683                                    const struct object_id *oid2,
684                                    enum object_type type1,
685                                    enum object_type type2)
686 {
687         int cmp = type1 - type2;
688         if (cmp)
689                 return cmp;
690         return oidcmp(oid1, oid2);
691 }
692
693 static int find_ref_delta(const struct object_id *oid)
694 {
695         int first = 0, last = nr_ref_deltas;
696
697         while (first < last) {
698                 int next = first + (last - first) / 2;
699                 struct ref_delta_entry *delta = &ref_deltas[next];
700                 int cmp;
701
702                 cmp = compare_ref_delta_bases(oid, &delta->oid,
703                                               OBJ_REF_DELTA,
704                                               objects[delta->obj_no].type);
705                 if (!cmp)
706                         return next;
707                 if (cmp < 0) {
708                         last = next;
709                         continue;
710                 }
711                 first = next+1;
712         }
713         return -first-1;
714 }
715
716 static void find_ref_delta_children(const struct object_id *oid,
717                                     int *first_index, int *last_index)
718 {
719         int first = find_ref_delta(oid);
720         int last = first;
721         int end = nr_ref_deltas - 1;
722
723         if (first < 0) {
724                 *first_index = 0;
725                 *last_index = -1;
726                 return;
727         }
728         while (first > 0 && oideq(&ref_deltas[first - 1].oid, oid))
729                 --first;
730         while (last < end && oideq(&ref_deltas[last + 1].oid, oid))
731                 ++last;
732         *first_index = first;
733         *last_index = last;
734 }
735
736 struct compare_data {
737         struct object_entry *entry;
738         struct git_istream *st;
739         unsigned char *buf;
740         unsigned long buf_size;
741 };
742
743 static int compare_objects(const unsigned char *buf, unsigned long size,
744                            void *cb_data)
745 {
746         struct compare_data *data = cb_data;
747
748         if (data->buf_size < size) {
749                 free(data->buf);
750                 data->buf = xmalloc(size);
751                 data->buf_size = size;
752         }
753
754         while (size) {
755                 ssize_t len = read_istream(data->st, data->buf, size);
756                 if (len == 0)
757                         die(_("SHA1 COLLISION FOUND WITH %s !"),
758                             oid_to_hex(&data->entry->idx.oid));
759                 if (len < 0)
760                         die(_("unable to read %s"),
761                             oid_to_hex(&data->entry->idx.oid));
762                 if (memcmp(buf, data->buf, len))
763                         die(_("SHA1 COLLISION FOUND WITH %s !"),
764                             oid_to_hex(&data->entry->idx.oid));
765                 size -= len;
766                 buf += len;
767         }
768         return 0;
769 }
770
771 static int check_collison(struct object_entry *entry)
772 {
773         struct compare_data data;
774         enum object_type type;
775         unsigned long size;
776
777         if (entry->size <= big_file_threshold || entry->type != OBJ_BLOB)
778                 return -1;
779
780         memset(&data, 0, sizeof(data));
781         data.entry = entry;
782         data.st = open_istream(the_repository, &entry->idx.oid, &type, &size,
783                                NULL);
784         if (!data.st)
785                 return -1;
786         if (size != entry->size || type != entry->type)
787                 die(_("SHA1 COLLISION FOUND WITH %s !"),
788                     oid_to_hex(&entry->idx.oid));
789         unpack_data(entry, compare_objects, &data);
790         close_istream(data.st);
791         free(data.buf);
792         return 0;
793 }
794
795 static void sha1_object(const void *data, struct object_entry *obj_entry,
796                         unsigned long size, enum object_type type,
797                         const struct object_id *oid)
798 {
799         void *new_data = NULL;
800         int collision_test_needed = 0;
801
802         assert(data || obj_entry);
803
804         if (startup_info->have_repository) {
805                 read_lock();
806                 collision_test_needed =
807                         has_object_file_with_flags(oid, OBJECT_INFO_QUICK);
808                 read_unlock();
809         }
810
811         if (collision_test_needed && !data) {
812                 read_lock();
813                 if (!check_collison(obj_entry))
814                         collision_test_needed = 0;
815                 read_unlock();
816         }
817         if (collision_test_needed) {
818                 void *has_data;
819                 enum object_type has_type;
820                 unsigned long has_size;
821                 read_lock();
822                 has_type = oid_object_info(the_repository, oid, &has_size);
823                 if (has_type < 0)
824                         die(_("cannot read existing object info %s"), oid_to_hex(oid));
825                 if (has_type != type || has_size != size)
826                         die(_("SHA1 COLLISION FOUND WITH %s !"), oid_to_hex(oid));
827                 has_data = read_object_file(oid, &has_type, &has_size);
828                 read_unlock();
829                 if (!data)
830                         data = new_data = get_data_from_pack(obj_entry);
831                 if (!has_data)
832                         die(_("cannot read existing object %s"), oid_to_hex(oid));
833                 if (size != has_size || type != has_type ||
834                     memcmp(data, has_data, size) != 0)
835                         die(_("SHA1 COLLISION FOUND WITH %s !"), oid_to_hex(oid));
836                 free(has_data);
837         }
838
839         if (strict || do_fsck_object) {
840                 read_lock();
841                 if (type == OBJ_BLOB) {
842                         struct blob *blob = lookup_blob(the_repository, oid);
843                         if (blob)
844                                 blob->object.flags |= FLAG_CHECKED;
845                         else
846                                 die(_("invalid blob object %s"), oid_to_hex(oid));
847                         if (do_fsck_object &&
848                             fsck_object(&blob->object, (void *)data, size, &fsck_options))
849                                 die(_("fsck error in packed object"));
850                 } else {
851                         struct object *obj;
852                         int eaten;
853                         void *buf = (void *) data;
854
855                         assert(data && "data can only be NULL for large _blobs_");
856
857                         /*
858                          * we do not need to free the memory here, as the
859                          * buf is deleted by the caller.
860                          */
861                         obj = parse_object_buffer(the_repository, oid, type,
862                                                   size, buf,
863                                                   &eaten);
864                         if (!obj)
865                                 die(_("invalid %s"), type_name(type));
866                         if (do_fsck_object &&
867                             fsck_object(obj, buf, size, &fsck_options))
868                                 die(_("fsck error in packed object"));
869                         if (strict && fsck_walk(obj, NULL, &fsck_options))
870                                 die(_("Not all child objects of %s are reachable"), oid_to_hex(&obj->oid));
871
872                         if (obj->type == OBJ_TREE) {
873                                 struct tree *item = (struct tree *) obj;
874                                 item->buffer = NULL;
875                                 obj->parsed = 0;
876                         }
877                         if (obj->type == OBJ_COMMIT) {
878                                 struct commit *commit = (struct commit *) obj;
879                                 if (detach_commit_buffer(commit, NULL) != data)
880                                         BUG("parse_object_buffer transmogrified our buffer");
881                         }
882                         obj->flags |= FLAG_CHECKED;
883                 }
884                 read_unlock();
885         }
886
887         free(new_data);
888 }
889
890 /*
891  * Ensure that this node has been reconstructed and return its contents.
892  *
893  * In the typical and best case, this node would already be reconstructed
894  * (through the invocation to resolve_delta() in threaded_second_pass()) and it
895  * would not be pruned. However, if pruning of this node was necessary due to
896  * reaching delta_base_cache_limit, this function will find the closest
897  * ancestor with reconstructed data that has not been pruned (or if there is
898  * none, the ultimate base object), and reconstruct each node in the delta
899  * chain in order to generate the reconstructed data for this node.
900  */
901 static void *get_base_data(struct base_data *c)
902 {
903         if (!c->data) {
904                 struct object_entry *obj = c->obj;
905                 struct base_data **delta = NULL;
906                 int delta_nr = 0, delta_alloc = 0;
907
908                 while (is_delta_type(c->obj->type) && !c->data) {
909                         ALLOC_GROW(delta, delta_nr + 1, delta_alloc);
910                         delta[delta_nr++] = c;
911                         c = c->base;
912                 }
913                 if (!delta_nr) {
914                         c->data = get_data_from_pack(obj);
915                         c->size = obj->size;
916                         base_cache_used += c->size;
917                         prune_base_data(c);
918                 }
919                 for (; delta_nr > 0; delta_nr--) {
920                         void *base, *raw;
921                         c = delta[delta_nr - 1];
922                         obj = c->obj;
923                         base = get_base_data(c->base);
924                         raw = get_data_from_pack(obj);
925                         c->data = patch_delta(
926                                 base, c->base->size,
927                                 raw, obj->size,
928                                 &c->size);
929                         free(raw);
930                         if (!c->data)
931                                 bad_object(obj->idx.offset, _("failed to apply delta"));
932                         base_cache_used += c->size;
933                         prune_base_data(c);
934                 }
935                 free(delta);
936         }
937         return c->data;
938 }
939
940 static struct base_data *make_base(struct object_entry *obj,
941                                    struct base_data *parent)
942 {
943         struct base_data *base = xcalloc(1, sizeof(struct base_data));
944         base->base = parent;
945         base->obj = obj;
946         find_ref_delta_children(&obj->idx.oid,
947                                 &base->ref_first, &base->ref_last);
948         find_ofs_delta_children(obj->idx.offset,
949                                 &base->ofs_first, &base->ofs_last);
950         base->children_remaining = base->ref_last - base->ref_first +
951                 base->ofs_last - base->ofs_first + 2;
952         return base;
953 }
954
955 static struct base_data *resolve_delta(struct object_entry *delta_obj,
956                                        struct base_data *base)
957 {
958         void *delta_data, *result_data;
959         struct base_data *result;
960         unsigned long result_size;
961
962         if (show_stat) {
963                 int i = delta_obj - objects;
964                 int j = base->obj - objects;
965                 obj_stat[i].delta_depth = obj_stat[j].delta_depth + 1;
966                 deepest_delta_lock();
967                 if (deepest_delta < obj_stat[i].delta_depth)
968                         deepest_delta = obj_stat[i].delta_depth;
969                 deepest_delta_unlock();
970                 obj_stat[i].base_object_no = j;
971         }
972         delta_data = get_data_from_pack(delta_obj);
973         assert(base->data);
974         result_data = patch_delta(base->data, base->size,
975                                   delta_data, delta_obj->size, &result_size);
976         free(delta_data);
977         if (!result_data)
978                 bad_object(delta_obj->idx.offset, _("failed to apply delta"));
979         hash_object_file(the_hash_algo, result_data, result_size,
980                          type_name(delta_obj->real_type), &delta_obj->idx.oid);
981         sha1_object(result_data, NULL, result_size, delta_obj->real_type,
982                     &delta_obj->idx.oid);
983
984         result = make_base(delta_obj, base);
985         result->data = result_data;
986         result->size = result_size;
987
988         counter_lock();
989         nr_resolved_deltas++;
990         counter_unlock();
991
992         return result;
993 }
994
995 static int compare_ofs_delta_entry(const void *a, const void *b)
996 {
997         const struct ofs_delta_entry *delta_a = a;
998         const struct ofs_delta_entry *delta_b = b;
999
1000         return delta_a->offset < delta_b->offset ? -1 :
1001                delta_a->offset > delta_b->offset ?  1 :
1002                0;
1003 }
1004
1005 static int compare_ref_delta_entry(const void *a, const void *b)
1006 {
1007         const struct ref_delta_entry *delta_a = a;
1008         const struct ref_delta_entry *delta_b = b;
1009
1010         return oidcmp(&delta_a->oid, &delta_b->oid);
1011 }
1012
1013 static void *threaded_second_pass(void *data)
1014 {
1015         if (data)
1016                 set_thread_data(data);
1017         for (;;) {
1018                 struct base_data *parent = NULL;
1019                 struct object_entry *child_obj;
1020                 struct base_data *child;
1021
1022                 counter_lock();
1023                 display_progress(progress, nr_resolved_deltas);
1024                 counter_unlock();
1025
1026                 work_lock();
1027                 if (list_empty(&work_head)) {
1028                         /*
1029                          * Take an object from the object array.
1030                          */
1031                         while (nr_dispatched < nr_objects &&
1032                                is_delta_type(objects[nr_dispatched].type))
1033                                 nr_dispatched++;
1034                         if (nr_dispatched >= nr_objects) {
1035                                 work_unlock();
1036                                 break;
1037                         }
1038                         child_obj = &objects[nr_dispatched++];
1039                 } else {
1040                         /*
1041                          * Peek at the top of the stack, and take a child from
1042                          * it.
1043                          */
1044                         parent = list_first_entry(&work_head, struct base_data,
1045                                                   list);
1046
1047                         if (parent->ref_first <= parent->ref_last) {
1048                                 int offset = ref_deltas[parent->ref_first++].obj_no;
1049                                 child_obj = objects + offset;
1050                                 if (child_obj->real_type != OBJ_REF_DELTA)
1051                                         die("REF_DELTA at offset %"PRIuMAX" already resolved (duplicate base %s?)",
1052                                             (uintmax_t) child_obj->idx.offset,
1053                                             oid_to_hex(&parent->obj->idx.oid));
1054                                 child_obj->real_type = parent->obj->real_type;
1055                         } else {
1056                                 child_obj = objects +
1057                                         ofs_deltas[parent->ofs_first++].obj_no;
1058                                 assert(child_obj->real_type == OBJ_OFS_DELTA);
1059                                 child_obj->real_type = parent->obj->real_type;
1060                         }
1061
1062                         if (parent->ref_first > parent->ref_last &&
1063                             parent->ofs_first > parent->ofs_last) {
1064                                 /*
1065                                  * This parent has run out of children, so move
1066                                  * it to done_head.
1067                                  */
1068                                 list_del(&parent->list);
1069                                 list_add(&parent->list, &done_head);
1070                         }
1071
1072                         /*
1073                          * Ensure that the parent has data, since we will need
1074                          * it later.
1075                          *
1076                          * NEEDSWORK: If parent data needs to be reloaded, this
1077                          * prolongs the time that the current thread spends in
1078                          * the mutex. A mitigating factor is that parent data
1079                          * needs to be reloaded only if the delta base cache
1080                          * limit is exceeded, so in the typical case, this does
1081                          * not happen.
1082                          */
1083                         get_base_data(parent);
1084                         parent->retain_data++;
1085                 }
1086                 work_unlock();
1087
1088                 if (parent) {
1089                         child = resolve_delta(child_obj, parent);
1090                         if (!child->children_remaining)
1091                                 FREE_AND_NULL(child->data);
1092                 } else {
1093                         child = make_base(child_obj, NULL);
1094                         if (child->children_remaining) {
1095                                 /*
1096                                  * Since this child has its own delta children,
1097                                  * we will need this data in the future.
1098                                  * Inflate now so that future iterations will
1099                                  * have access to this object's data while
1100                                  * outside the work mutex.
1101                                  */
1102                                 child->data = get_data_from_pack(child_obj);
1103                                 child->size = child_obj->size;
1104                         }
1105                 }
1106
1107                 work_lock();
1108                 if (parent)
1109                         parent->retain_data--;
1110                 if (child->data) {
1111                         /*
1112                          * This child has its own children, so add it to
1113                          * work_head.
1114                          */
1115                         list_add(&child->list, &work_head);
1116                         base_cache_used += child->size;
1117                         prune_base_data(NULL);
1118                 } else {
1119                         /*
1120                          * This child does not have its own children. It may be
1121                          * the last descendant of its ancestors; free those
1122                          * that we can.
1123                          */
1124                         struct base_data *p = parent;
1125
1126                         while (p) {
1127                                 struct base_data *next_p;
1128
1129                                 p->children_remaining--;
1130                                 if (p->children_remaining)
1131                                         break;
1132
1133                                 next_p = p->base;
1134                                 free_base_data(p);
1135                                 list_del(&p->list);
1136                                 free(p);
1137
1138                                 p = next_p;
1139                         }
1140                 }
1141                 work_unlock();
1142         }
1143         return NULL;
1144 }
1145
1146 /*
1147  * First pass:
1148  * - find locations of all objects;
1149  * - calculate SHA1 of all non-delta objects;
1150  * - remember base (SHA1 or offset) for all deltas.
1151  */
1152 static void parse_pack_objects(unsigned char *hash)
1153 {
1154         int i, nr_delays = 0;
1155         struct ofs_delta_entry *ofs_delta = ofs_deltas;
1156         struct object_id ref_delta_oid;
1157         struct stat st;
1158
1159         if (verbose)
1160                 progress = start_progress(
1161                                 from_stdin ? _("Receiving objects") : _("Indexing objects"),
1162                                 nr_objects);
1163         for (i = 0; i < nr_objects; i++) {
1164                 struct object_entry *obj = &objects[i];
1165                 void *data = unpack_raw_entry(obj, &ofs_delta->offset,
1166                                               &ref_delta_oid,
1167                                               &obj->idx.oid);
1168                 obj->real_type = obj->type;
1169                 if (obj->type == OBJ_OFS_DELTA) {
1170                         nr_ofs_deltas++;
1171                         ofs_delta->obj_no = i;
1172                         ofs_delta++;
1173                 } else if (obj->type == OBJ_REF_DELTA) {
1174                         ALLOC_GROW(ref_deltas, nr_ref_deltas + 1, ref_deltas_alloc);
1175                         oidcpy(&ref_deltas[nr_ref_deltas].oid, &ref_delta_oid);
1176                         ref_deltas[nr_ref_deltas].obj_no = i;
1177                         nr_ref_deltas++;
1178                 } else if (!data) {
1179                         /* large blobs, check later */
1180                         obj->real_type = OBJ_BAD;
1181                         nr_delays++;
1182                 } else
1183                         sha1_object(data, NULL, obj->size, obj->type,
1184                                     &obj->idx.oid);
1185                 free(data);
1186                 display_progress(progress, i+1);
1187         }
1188         objects[i].idx.offset = consumed_bytes;
1189         stop_progress(&progress);
1190
1191         /* Check pack integrity */
1192         flush();
1193         the_hash_algo->final_fn(hash, &input_ctx);
1194         if (!hasheq(fill(the_hash_algo->rawsz), hash))
1195                 die(_("pack is corrupted (SHA1 mismatch)"));
1196         use(the_hash_algo->rawsz);
1197
1198         /* If input_fd is a file, we should have reached its end now. */
1199         if (fstat(input_fd, &st))
1200                 die_errno(_("cannot fstat packfile"));
1201         if (S_ISREG(st.st_mode) &&
1202                         lseek(input_fd, 0, SEEK_CUR) - input_len != st.st_size)
1203                 die(_("pack has junk at the end"));
1204
1205         for (i = 0; i < nr_objects; i++) {
1206                 struct object_entry *obj = &objects[i];
1207                 if (obj->real_type != OBJ_BAD)
1208                         continue;
1209                 obj->real_type = obj->type;
1210                 sha1_object(NULL, obj, obj->size, obj->type,
1211                             &obj->idx.oid);
1212                 nr_delays--;
1213         }
1214         if (nr_delays)
1215                 die(_("confusion beyond insanity in parse_pack_objects()"));
1216 }
1217
1218 /*
1219  * Second pass:
1220  * - for all non-delta objects, look if it is used as a base for
1221  *   deltas;
1222  * - if used as a base, uncompress the object and apply all deltas,
1223  *   recursively checking if the resulting object is used as a base
1224  *   for some more deltas.
1225  */
1226 static void resolve_deltas(void)
1227 {
1228         int i;
1229
1230         if (!nr_ofs_deltas && !nr_ref_deltas)
1231                 return;
1232
1233         /* Sort deltas by base SHA1/offset for fast searching */
1234         QSORT(ofs_deltas, nr_ofs_deltas, compare_ofs_delta_entry);
1235         QSORT(ref_deltas, nr_ref_deltas, compare_ref_delta_entry);
1236
1237         if (verbose || show_resolving_progress)
1238                 progress = start_progress(_("Resolving deltas"),
1239                                           nr_ref_deltas + nr_ofs_deltas);
1240
1241         nr_dispatched = 0;
1242         base_cache_limit = delta_base_cache_limit * nr_threads;
1243         if (nr_threads > 1 || getenv("GIT_FORCE_THREADS")) {
1244                 init_thread();
1245                 for (i = 0; i < nr_threads; i++) {
1246                         int ret = pthread_create(&thread_data[i].thread, NULL,
1247                                                  threaded_second_pass, thread_data + i);
1248                         if (ret)
1249                                 die(_("unable to create thread: %s"),
1250                                     strerror(ret));
1251                 }
1252                 for (i = 0; i < nr_threads; i++)
1253                         pthread_join(thread_data[i].thread, NULL);
1254                 cleanup_thread();
1255                 return;
1256         }
1257         threaded_second_pass(&nothread_data);
1258 }
1259
1260 /*
1261  * Third pass:
1262  * - append objects to convert thin pack to full pack if required
1263  * - write the final pack hash
1264  */
1265 static void fix_unresolved_deltas(struct hashfile *f);
1266 static void conclude_pack(int fix_thin_pack, const char *curr_pack, unsigned char *pack_hash)
1267 {
1268         if (nr_ref_deltas + nr_ofs_deltas == nr_resolved_deltas) {
1269                 stop_progress(&progress);
1270                 /* Flush remaining pack final hash. */
1271                 flush();
1272                 return;
1273         }
1274
1275         if (fix_thin_pack) {
1276                 struct hashfile *f;
1277                 unsigned char read_hash[GIT_MAX_RAWSZ], tail_hash[GIT_MAX_RAWSZ];
1278                 struct strbuf msg = STRBUF_INIT;
1279                 int nr_unresolved = nr_ofs_deltas + nr_ref_deltas - nr_resolved_deltas;
1280                 int nr_objects_initial = nr_objects;
1281                 if (nr_unresolved <= 0)
1282                         die(_("confusion beyond insanity"));
1283                 REALLOC_ARRAY(objects, nr_objects + nr_unresolved + 1);
1284                 memset(objects + nr_objects + 1, 0,
1285                        nr_unresolved * sizeof(*objects));
1286                 f = hashfd(output_fd, curr_pack);
1287                 fix_unresolved_deltas(f);
1288                 strbuf_addf(&msg, Q_("completed with %d local object",
1289                                      "completed with %d local objects",
1290                                      nr_objects - nr_objects_initial),
1291                             nr_objects - nr_objects_initial);
1292                 stop_progress_msg(&progress, msg.buf);
1293                 strbuf_release(&msg);
1294                 finalize_hashfile(f, tail_hash, 0);
1295                 hashcpy(read_hash, pack_hash);
1296                 fixup_pack_header_footer(output_fd, pack_hash,
1297                                          curr_pack, nr_objects,
1298                                          read_hash, consumed_bytes-the_hash_algo->rawsz);
1299                 if (!hasheq(read_hash, tail_hash))
1300                         die(_("Unexpected tail checksum for %s "
1301                               "(disk corruption?)"), curr_pack);
1302         }
1303         if (nr_ofs_deltas + nr_ref_deltas != nr_resolved_deltas)
1304                 die(Q_("pack has %d unresolved delta",
1305                        "pack has %d unresolved deltas",
1306                        nr_ofs_deltas + nr_ref_deltas - nr_resolved_deltas),
1307                     nr_ofs_deltas + nr_ref_deltas - nr_resolved_deltas);
1308 }
1309
1310 static int write_compressed(struct hashfile *f, void *in, unsigned int size)
1311 {
1312         git_zstream stream;
1313         int status;
1314         unsigned char outbuf[4096];
1315
1316         git_deflate_init(&stream, zlib_compression_level);
1317         stream.next_in = in;
1318         stream.avail_in = size;
1319
1320         do {
1321                 stream.next_out = outbuf;
1322                 stream.avail_out = sizeof(outbuf);
1323                 status = git_deflate(&stream, Z_FINISH);
1324                 hashwrite(f, outbuf, sizeof(outbuf) - stream.avail_out);
1325         } while (status == Z_OK);
1326
1327         if (status != Z_STREAM_END)
1328                 die(_("unable to deflate appended object (%d)"), status);
1329         size = stream.total_out;
1330         git_deflate_end(&stream);
1331         return size;
1332 }
1333
1334 static struct object_entry *append_obj_to_pack(struct hashfile *f,
1335                                const unsigned char *sha1, void *buf,
1336                                unsigned long size, enum object_type type)
1337 {
1338         struct object_entry *obj = &objects[nr_objects++];
1339         unsigned char header[10];
1340         unsigned long s = size;
1341         int n = 0;
1342         unsigned char c = (type << 4) | (s & 15);
1343         s >>= 4;
1344         while (s) {
1345                 header[n++] = c | 0x80;
1346                 c = s & 0x7f;
1347                 s >>= 7;
1348         }
1349         header[n++] = c;
1350         crc32_begin(f);
1351         hashwrite(f, header, n);
1352         obj[0].size = size;
1353         obj[0].hdr_size = n;
1354         obj[0].type = type;
1355         obj[0].real_type = type;
1356         obj[1].idx.offset = obj[0].idx.offset + n;
1357         obj[1].idx.offset += write_compressed(f, buf, size);
1358         obj[0].idx.crc32 = crc32_end(f);
1359         hashflush(f);
1360         hashcpy(obj->idx.oid.hash, sha1);
1361         return obj;
1362 }
1363
1364 static int delta_pos_compare(const void *_a, const void *_b)
1365 {
1366         struct ref_delta_entry *a = *(struct ref_delta_entry **)_a;
1367         struct ref_delta_entry *b = *(struct ref_delta_entry **)_b;
1368         return a->obj_no - b->obj_no;
1369 }
1370
1371 static void fix_unresolved_deltas(struct hashfile *f)
1372 {
1373         struct ref_delta_entry **sorted_by_pos;
1374         int i;
1375
1376         /*
1377          * Since many unresolved deltas may well be themselves base objects
1378          * for more unresolved deltas, we really want to include the
1379          * smallest number of base objects that would cover as much delta
1380          * as possible by picking the
1381          * trunc deltas first, allowing for other deltas to resolve without
1382          * additional base objects.  Since most base objects are to be found
1383          * before deltas depending on them, a good heuristic is to start
1384          * resolving deltas in the same order as their position in the pack.
1385          */
1386         ALLOC_ARRAY(sorted_by_pos, nr_ref_deltas);
1387         for (i = 0; i < nr_ref_deltas; i++)
1388                 sorted_by_pos[i] = &ref_deltas[i];
1389         QSORT(sorted_by_pos, nr_ref_deltas, delta_pos_compare);
1390
1391         if (has_promisor_remote()) {
1392                 /*
1393                  * Prefetch the delta bases.
1394                  */
1395                 struct oid_array to_fetch = OID_ARRAY_INIT;
1396                 for (i = 0; i < nr_ref_deltas; i++) {
1397                         struct ref_delta_entry *d = sorted_by_pos[i];
1398                         if (!oid_object_info_extended(the_repository, &d->oid,
1399                                                       NULL,
1400                                                       OBJECT_INFO_FOR_PREFETCH))
1401                                 continue;
1402                         oid_array_append(&to_fetch, &d->oid);
1403                 }
1404                 promisor_remote_get_direct(the_repository,
1405                                            to_fetch.oid, to_fetch.nr);
1406                 oid_array_clear(&to_fetch);
1407         }
1408
1409         for (i = 0; i < nr_ref_deltas; i++) {
1410                 struct ref_delta_entry *d = sorted_by_pos[i];
1411                 enum object_type type;
1412                 void *data;
1413                 unsigned long size;
1414
1415                 if (objects[d->obj_no].real_type != OBJ_REF_DELTA)
1416                         continue;
1417                 data = read_object_file(&d->oid, &type, &size);
1418                 if (!data)
1419                         continue;
1420
1421                 if (check_object_signature(the_repository, &d->oid,
1422                                            data, size,
1423                                            type_name(type)))
1424                         die(_("local object %s is corrupt"), oid_to_hex(&d->oid));
1425
1426                 /*
1427                  * Add this as an object to the objects array and call
1428                  * threaded_second_pass() (which will pick up the added
1429                  * object).
1430                  */
1431                 append_obj_to_pack(f, d->oid.hash, data, size, type);
1432                 threaded_second_pass(NULL);
1433
1434                 display_progress(progress, nr_resolved_deltas);
1435         }
1436         free(sorted_by_pos);
1437 }
1438
1439 static const char *derive_filename(const char *pack_name, const char *suffix,
1440                                    struct strbuf *buf)
1441 {
1442         size_t len;
1443         if (!strip_suffix(pack_name, ".pack", &len))
1444                 die(_("packfile name '%s' does not end with '.pack'"),
1445                     pack_name);
1446         strbuf_add(buf, pack_name, len);
1447         strbuf_addch(buf, '.');
1448         strbuf_addstr(buf, suffix);
1449         return buf->buf;
1450 }
1451
1452 static void write_special_file(const char *suffix, const char *msg,
1453                                const char *pack_name, const unsigned char *hash,
1454                                const char **report)
1455 {
1456         struct strbuf name_buf = STRBUF_INIT;
1457         const char *filename;
1458         int fd;
1459         int msg_len = strlen(msg);
1460
1461         if (pack_name)
1462                 filename = derive_filename(pack_name, suffix, &name_buf);
1463         else
1464                 filename = odb_pack_name(&name_buf, hash, suffix);
1465
1466         fd = odb_pack_keep(filename);
1467         if (fd < 0) {
1468                 if (errno != EEXIST)
1469                         die_errno(_("cannot write %s file '%s'"),
1470                                   suffix, filename);
1471         } else {
1472                 if (msg_len > 0) {
1473                         write_or_die(fd, msg, msg_len);
1474                         write_or_die(fd, "\n", 1);
1475                 }
1476                 if (close(fd) != 0)
1477                         die_errno(_("cannot close written %s file '%s'"),
1478                                   suffix, filename);
1479                 if (report)
1480                         *report = suffix;
1481         }
1482         strbuf_release(&name_buf);
1483 }
1484
1485 static void final(const char *final_pack_name, const char *curr_pack_name,
1486                   const char *final_index_name, const char *curr_index_name,
1487                   const char *keep_msg, const char *promisor_msg,
1488                   unsigned char *hash)
1489 {
1490         const char *report = "pack";
1491         struct strbuf pack_name = STRBUF_INIT;
1492         struct strbuf index_name = STRBUF_INIT;
1493         int err;
1494
1495         if (!from_stdin) {
1496                 close(input_fd);
1497         } else {
1498                 fsync_or_die(output_fd, curr_pack_name);
1499                 err = close(output_fd);
1500                 if (err)
1501                         die_errno(_("error while closing pack file"));
1502         }
1503
1504         if (keep_msg)
1505                 write_special_file("keep", keep_msg, final_pack_name, hash,
1506                                    &report);
1507         if (promisor_msg)
1508                 write_special_file("promisor", promisor_msg, final_pack_name,
1509                                    hash, NULL);
1510
1511         if (final_pack_name != curr_pack_name) {
1512                 if (!final_pack_name)
1513                         final_pack_name = odb_pack_name(&pack_name, hash, "pack");
1514                 if (finalize_object_file(curr_pack_name, final_pack_name))
1515                         die(_("cannot store pack file"));
1516         } else if (from_stdin)
1517                 chmod(final_pack_name, 0444);
1518
1519         if (final_index_name != curr_index_name) {
1520                 if (!final_index_name)
1521                         final_index_name = odb_pack_name(&index_name, hash, "idx");
1522                 if (finalize_object_file(curr_index_name, final_index_name))
1523                         die(_("cannot store index file"));
1524         } else
1525                 chmod(final_index_name, 0444);
1526
1527         if (do_fsck_object) {
1528                 struct packed_git *p;
1529                 p = add_packed_git(final_index_name, strlen(final_index_name), 0);
1530                 if (p)
1531                         install_packed_git(the_repository, p);
1532         }
1533
1534         if (!from_stdin) {
1535                 printf("%s\n", hash_to_hex(hash));
1536         } else {
1537                 struct strbuf buf = STRBUF_INIT;
1538
1539                 strbuf_addf(&buf, "%s\t%s\n", report, hash_to_hex(hash));
1540                 write_or_die(1, buf.buf, buf.len);
1541                 strbuf_release(&buf);
1542
1543                 /*
1544                  * Let's just mimic git-unpack-objects here and write
1545                  * the last part of the input buffer to stdout.
1546                  */
1547                 while (input_len) {
1548                         err = xwrite(1, input_buffer + input_offset, input_len);
1549                         if (err <= 0)
1550                                 break;
1551                         input_len -= err;
1552                         input_offset += err;
1553                 }
1554         }
1555
1556         strbuf_release(&index_name);
1557         strbuf_release(&pack_name);
1558 }
1559
1560 static int git_index_pack_config(const char *k, const char *v, void *cb)
1561 {
1562         struct pack_idx_option *opts = cb;
1563
1564         if (!strcmp(k, "pack.indexversion")) {
1565                 opts->version = git_config_int(k, v);
1566                 if (opts->version > 2)
1567                         die(_("bad pack.indexversion=%"PRIu32), opts->version);
1568                 return 0;
1569         }
1570         if (!strcmp(k, "pack.threads")) {
1571                 nr_threads = git_config_int(k, v);
1572                 if (nr_threads < 0)
1573                         die(_("invalid number of threads specified (%d)"),
1574                             nr_threads);
1575                 if (!HAVE_THREADS && nr_threads != 1) {
1576                         warning(_("no threads support, ignoring %s"), k);
1577                         nr_threads = 1;
1578                 }
1579                 return 0;
1580         }
1581         return git_default_config(k, v, cb);
1582 }
1583
1584 static int cmp_uint32(const void *a_, const void *b_)
1585 {
1586         uint32_t a = *((uint32_t *)a_);
1587         uint32_t b = *((uint32_t *)b_);
1588
1589         return (a < b) ? -1 : (a != b);
1590 }
1591
1592 static void read_v2_anomalous_offsets(struct packed_git *p,
1593                                       struct pack_idx_option *opts)
1594 {
1595         const uint32_t *idx1, *idx2;
1596         uint32_t i;
1597
1598         /* The address of the 4-byte offset table */
1599         idx1 = (((const uint32_t *)((const uint8_t *)p->index_data + p->crc_offset))
1600                 + p->num_objects /* CRC32 table */
1601                 );
1602
1603         /* The address of the 8-byte offset table */
1604         idx2 = idx1 + p->num_objects;
1605
1606         for (i = 0; i < p->num_objects; i++) {
1607                 uint32_t off = ntohl(idx1[i]);
1608                 if (!(off & 0x80000000))
1609                         continue;
1610                 off = off & 0x7fffffff;
1611                 check_pack_index_ptr(p, &idx2[off * 2]);
1612                 if (idx2[off * 2])
1613                         continue;
1614                 /*
1615                  * The real offset is ntohl(idx2[off * 2]) in high 4
1616                  * octets, and ntohl(idx2[off * 2 + 1]) in low 4
1617                  * octets.  But idx2[off * 2] is Zero!!!
1618                  */
1619                 ALLOC_GROW(opts->anomaly, opts->anomaly_nr + 1, opts->anomaly_alloc);
1620                 opts->anomaly[opts->anomaly_nr++] = ntohl(idx2[off * 2 + 1]);
1621         }
1622
1623         QSORT(opts->anomaly, opts->anomaly_nr, cmp_uint32);
1624 }
1625
1626 static void read_idx_option(struct pack_idx_option *opts, const char *pack_name)
1627 {
1628         struct packed_git *p = add_packed_git(pack_name, strlen(pack_name), 1);
1629
1630         if (!p)
1631                 die(_("Cannot open existing pack file '%s'"), pack_name);
1632         if (open_pack_index(p))
1633                 die(_("Cannot open existing pack idx file for '%s'"), pack_name);
1634
1635         /* Read the attributes from the existing idx file */
1636         opts->version = p->index_version;
1637
1638         if (opts->version == 2)
1639                 read_v2_anomalous_offsets(p, opts);
1640
1641         /*
1642          * Get rid of the idx file as we do not need it anymore.
1643          * NEEDSWORK: extract this bit from free_pack_by_name() in
1644          * sha1-file.c, perhaps?  It shouldn't matter very much as we
1645          * know we haven't installed this pack (hence we never have
1646          * read anything from it).
1647          */
1648         close_pack_index(p);
1649         free(p);
1650 }
1651
1652 static void show_pack_info(int stat_only)
1653 {
1654         int i, baseobjects = nr_objects - nr_ref_deltas - nr_ofs_deltas;
1655         unsigned long *chain_histogram = NULL;
1656
1657         if (deepest_delta)
1658                 chain_histogram = xcalloc(deepest_delta, sizeof(unsigned long));
1659
1660         for (i = 0; i < nr_objects; i++) {
1661                 struct object_entry *obj = &objects[i];
1662
1663                 if (is_delta_type(obj->type))
1664                         chain_histogram[obj_stat[i].delta_depth - 1]++;
1665                 if (stat_only)
1666                         continue;
1667                 printf("%s %-6s %"PRIuMAX" %"PRIuMAX" %"PRIuMAX,
1668                        oid_to_hex(&obj->idx.oid),
1669                        type_name(obj->real_type), (uintmax_t)obj->size,
1670                        (uintmax_t)(obj[1].idx.offset - obj->idx.offset),
1671                        (uintmax_t)obj->idx.offset);
1672                 if (is_delta_type(obj->type)) {
1673                         struct object_entry *bobj = &objects[obj_stat[i].base_object_no];
1674                         printf(" %u %s", obj_stat[i].delta_depth,
1675                                oid_to_hex(&bobj->idx.oid));
1676                 }
1677                 putchar('\n');
1678         }
1679
1680         if (baseobjects)
1681                 printf_ln(Q_("non delta: %d object",
1682                              "non delta: %d objects",
1683                              baseobjects),
1684                           baseobjects);
1685         for (i = 0; i < deepest_delta; i++) {
1686                 if (!chain_histogram[i])
1687                         continue;
1688                 printf_ln(Q_("chain length = %d: %lu object",
1689                              "chain length = %d: %lu objects",
1690                              chain_histogram[i]),
1691                           i + 1,
1692                           chain_histogram[i]);
1693         }
1694 }
1695
1696 int cmd_index_pack(int argc, const char **argv, const char *prefix)
1697 {
1698         int i, fix_thin_pack = 0, verify = 0, stat_only = 0;
1699         const char *curr_index;
1700         const char *index_name = NULL, *pack_name = NULL;
1701         const char *keep_msg = NULL;
1702         const char *promisor_msg = NULL;
1703         struct strbuf index_name_buf = STRBUF_INIT;
1704         struct pack_idx_entry **idx_objects;
1705         struct pack_idx_option opts;
1706         unsigned char pack_hash[GIT_MAX_RAWSZ];
1707         unsigned foreign_nr = 1;        /* zero is a "good" value, assume bad */
1708         int report_end_of_input = 0;
1709         int hash_algo = 0;
1710
1711         /*
1712          * index-pack never needs to fetch missing objects except when
1713          * REF_DELTA bases are missing (which are explicitly handled). It only
1714          * accesses the repo to do hash collision checks and to check which
1715          * REF_DELTA bases need to be fetched.
1716          */
1717         fetch_if_missing = 0;
1718
1719         if (argc == 2 && !strcmp(argv[1], "-h"))
1720                 usage(index_pack_usage);
1721
1722         read_replace_refs = 0;
1723         fsck_options.walk = mark_link;
1724
1725         reset_pack_idx_option(&opts);
1726         git_config(git_index_pack_config, &opts);
1727         if (prefix && chdir(prefix))
1728                 die(_("Cannot come back to cwd"));
1729
1730         for (i = 1; i < argc; i++) {
1731                 const char *arg = argv[i];
1732
1733                 if (*arg == '-') {
1734                         if (!strcmp(arg, "--stdin")) {
1735                                 from_stdin = 1;
1736                         } else if (!strcmp(arg, "--fix-thin")) {
1737                                 fix_thin_pack = 1;
1738                         } else if (skip_to_optional_arg(arg, "--strict", &arg)) {
1739                                 strict = 1;
1740                                 do_fsck_object = 1;
1741                                 fsck_set_msg_types(&fsck_options, arg);
1742                         } else if (!strcmp(arg, "--check-self-contained-and-connected")) {
1743                                 strict = 1;
1744                                 check_self_contained_and_connected = 1;
1745                         } else if (!strcmp(arg, "--fsck-objects")) {
1746                                 do_fsck_object = 1;
1747                         } else if (!strcmp(arg, "--verify")) {
1748                                 verify = 1;
1749                         } else if (!strcmp(arg, "--verify-stat")) {
1750                                 verify = 1;
1751                                 show_stat = 1;
1752                         } else if (!strcmp(arg, "--verify-stat-only")) {
1753                                 verify = 1;
1754                                 show_stat = 1;
1755                                 stat_only = 1;
1756                         } else if (skip_to_optional_arg(arg, "--keep", &keep_msg)) {
1757                                 ; /* nothing to do */
1758                         } else if (skip_to_optional_arg(arg, "--promisor", &promisor_msg)) {
1759                                 ; /* already parsed */
1760                         } else if (starts_with(arg, "--threads=")) {
1761                                 char *end;
1762                                 nr_threads = strtoul(arg+10, &end, 0);
1763                                 if (!arg[10] || *end || nr_threads < 0)
1764                                         usage(index_pack_usage);
1765                                 if (!HAVE_THREADS && nr_threads != 1) {
1766                                         warning(_("no threads support, ignoring %s"), arg);
1767                                         nr_threads = 1;
1768                                 }
1769                         } else if (starts_with(arg, "--pack_header=")) {
1770                                 struct pack_header *hdr;
1771                                 char *c;
1772
1773                                 hdr = (struct pack_header *)input_buffer;
1774                                 hdr->hdr_signature = htonl(PACK_SIGNATURE);
1775                                 hdr->hdr_version = htonl(strtoul(arg + 14, &c, 10));
1776                                 if (*c != ',')
1777                                         die(_("bad %s"), arg);
1778                                 hdr->hdr_entries = htonl(strtoul(c + 1, &c, 10));
1779                                 if (*c)
1780                                         die(_("bad %s"), arg);
1781                                 input_len = sizeof(*hdr);
1782                         } else if (!strcmp(arg, "-v")) {
1783                                 verbose = 1;
1784                         } else if (!strcmp(arg, "--show-resolving-progress")) {
1785                                 show_resolving_progress = 1;
1786                         } else if (!strcmp(arg, "--report-end-of-input")) {
1787                                 report_end_of_input = 1;
1788                         } else if (!strcmp(arg, "-o")) {
1789                                 if (index_name || (i+1) >= argc)
1790                                         usage(index_pack_usage);
1791                                 index_name = argv[++i];
1792                         } else if (starts_with(arg, "--index-version=")) {
1793                                 char *c;
1794                                 opts.version = strtoul(arg + 16, &c, 10);
1795                                 if (opts.version > 2)
1796                                         die(_("bad %s"), arg);
1797                                 if (*c == ',')
1798                                         opts.off32_limit = strtoul(c+1, &c, 0);
1799                                 if (*c || opts.off32_limit & 0x80000000)
1800                                         die(_("bad %s"), arg);
1801                         } else if (skip_prefix(arg, "--max-input-size=", &arg)) {
1802                                 max_input_size = strtoumax(arg, NULL, 10);
1803                         } else if (skip_prefix(arg, "--object-format=", &arg)) {
1804                                 hash_algo = hash_algo_by_name(arg);
1805                                 if (hash_algo == GIT_HASH_UNKNOWN)
1806                                         die(_("unknown hash algorithm '%s'"), arg);
1807                                 repo_set_hash_algo(the_repository, hash_algo);
1808                         } else
1809                                 usage(index_pack_usage);
1810                         continue;
1811                 }
1812
1813                 if (pack_name)
1814                         usage(index_pack_usage);
1815                 pack_name = arg;
1816         }
1817
1818         if (!pack_name && !from_stdin)
1819                 usage(index_pack_usage);
1820         if (fix_thin_pack && !from_stdin)
1821                 die(_("--fix-thin cannot be used without --stdin"));
1822         if (from_stdin && !startup_info->have_repository)
1823                 die(_("--stdin requires a git repository"));
1824         if (from_stdin && hash_algo)
1825                 die(_("--object-format cannot be used with --stdin"));
1826         if (!index_name && pack_name)
1827                 index_name = derive_filename(pack_name, "idx", &index_name_buf);
1828
1829         if (verify) {
1830                 if (!index_name)
1831                         die(_("--verify with no packfile name given"));
1832                 read_idx_option(&opts, index_name);
1833                 opts.flags |= WRITE_IDX_VERIFY | WRITE_IDX_STRICT;
1834         }
1835         if (strict)
1836                 opts.flags |= WRITE_IDX_STRICT;
1837
1838         if (HAVE_THREADS && !nr_threads) {
1839                 nr_threads = online_cpus();
1840                 /*
1841                  * Experiments show that going above 20 threads doesn't help,
1842                  * no matter how many cores you have. Below that, we tend to
1843                  * max at half the number of online_cpus(), presumably because
1844                  * half of those are hyperthreads rather than full cores. We'll
1845                  * never reduce the level below "3", though, to match a
1846                  * historical value that nobody complained about.
1847                  */
1848                 if (nr_threads < 4)
1849                         ; /* too few cores to consider capping */
1850                 else if (nr_threads < 6)
1851                         nr_threads = 3; /* historic cap */
1852                 else if (nr_threads < 40)
1853                         nr_threads /= 2;
1854                 else
1855                         nr_threads = 20; /* hard cap */
1856         }
1857
1858         curr_pack = open_pack_file(pack_name);
1859         parse_pack_header();
1860         objects = xcalloc(st_add(nr_objects, 1), sizeof(struct object_entry));
1861         if (show_stat)
1862                 obj_stat = xcalloc(st_add(nr_objects, 1), sizeof(struct object_stat));
1863         ofs_deltas = xcalloc(nr_objects, sizeof(struct ofs_delta_entry));
1864         parse_pack_objects(pack_hash);
1865         if (report_end_of_input)
1866                 write_in_full(2, "\0", 1);
1867         resolve_deltas();
1868         conclude_pack(fix_thin_pack, curr_pack, pack_hash);
1869         free(ofs_deltas);
1870         free(ref_deltas);
1871         if (strict)
1872                 foreign_nr = check_objects();
1873
1874         if (show_stat)
1875                 show_pack_info(stat_only);
1876
1877         ALLOC_ARRAY(idx_objects, nr_objects);
1878         for (i = 0; i < nr_objects; i++)
1879                 idx_objects[i] = &objects[i].idx;
1880         curr_index = write_idx_file(index_name, idx_objects, nr_objects, &opts, pack_hash);
1881         free(idx_objects);
1882
1883         if (!verify)
1884                 final(pack_name, curr_pack,
1885                       index_name, curr_index,
1886                       keep_msg, promisor_msg,
1887                       pack_hash);
1888         else
1889                 close(input_fd);
1890
1891         if (do_fsck_object && fsck_finish(&fsck_options))
1892                 die(_("fsck error in pack objects"));
1893
1894         free(objects);
1895         strbuf_release(&index_name_buf);
1896         if (pack_name == NULL)
1897                 free((void *) curr_pack);
1898         if (index_name == NULL)
1899                 free((void *) curr_index);
1900
1901         /*
1902          * Let the caller know this pack is not self contained
1903          */
1904         if (check_self_contained_and_connected && foreign_nr)
1905                 return 1;
1906
1907         return 0;
1908 }