Merge branch 'jn/no-g-plus-s-on-bsd'
[git] / builtin / pack-objects.c
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
19 #include "thread-utils.h"
20
21 static const char pack_usage[] =
22   "git pack-objects [ -q | --progress | --all-progress ]\n"
23   "        [--all-progress-implied]\n"
24   "        [--max-pack-size=<n>] [--local] [--incremental]\n"
25   "        [--window=<n>] [--window-memory=<n>] [--depth=<n>]\n"
26   "        [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset]\n"
27   "        [--threads=<n>] [--non-empty] [--revs [--unpacked | --all]]\n"
28   "        [--reflog] [--stdout | base-name] [--include-tag]\n"
29   "        [--keep-unreachable | --unpack-unreachable]\n"
30   "        [< ref-list | < object-list]";
31
32 struct object_entry {
33         struct pack_idx_entry idx;
34         unsigned long size;     /* uncompressed size */
35         struct packed_git *in_pack;     /* already in pack */
36         off_t in_pack_offset;
37         struct object_entry *delta;     /* delta base object */
38         struct object_entry *delta_child; /* deltified objects who bases me */
39         struct object_entry *delta_sibling; /* other deltified objects who
40                                              * uses the same base as me
41                                              */
42         void *delta_data;       /* cached delta (uncompressed) */
43         unsigned long delta_size;       /* delta data size (uncompressed) */
44         unsigned long z_delta_size;     /* delta data size (compressed) */
45         unsigned int hash;      /* name hint hash */
46         enum object_type type;
47         enum object_type in_pack_type;  /* could be delta */
48         unsigned char in_pack_header_size;
49         unsigned char preferred_base; /* we do not pack this, but is available
50                                        * to be used as the base object to delta
51                                        * objects against.
52                                        */
53         unsigned char no_try_delta;
54         unsigned char tagged; /* near the very tip of refs */
55         unsigned char filled; /* assigned write-order */
56 };
57
58 /*
59  * Objects we are going to pack are collected in objects array (dynamically
60  * expanded).  nr_objects & nr_alloc controls this array.  They are stored
61  * in the order we see -- typically rev-list --objects order that gives us
62  * nice "minimum seek" order.
63  */
64 static struct object_entry *objects;
65 static struct pack_idx_entry **written_list;
66 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
67
68 static int non_empty;
69 static int reuse_delta = 1, reuse_object = 1;
70 static int keep_unreachable, unpack_unreachable, include_tag;
71 static int local;
72 static int incremental;
73 static int ignore_packed_keep;
74 static int allow_ofs_delta;
75 static struct pack_idx_option pack_idx_opts;
76 static const char *base_name;
77 static int progress = 1;
78 static int window = 10;
79 static unsigned long pack_size_limit, pack_size_limit_cfg;
80 static int depth = 50;
81 static int delta_search_threads;
82 static int pack_to_stdout;
83 static int num_preferred_base;
84 static struct progress *progress_state;
85 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
86 static int pack_compression_seen;
87
88 static unsigned long delta_cache_size = 0;
89 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
90 static unsigned long cache_max_small_delta_size = 1000;
91
92 static unsigned long window_memory_limit = 0;
93
94 /*
95  * The object names in objects array are hashed with this hashtable,
96  * to help looking up the entry by object name.
97  * This hashtable is built after all the objects are seen.
98  */
99 static int *object_ix;
100 static int object_ix_hashsz;
101 static struct object_entry *locate_object_entry(const unsigned char *sha1);
102
103 /*
104  * stats
105  */
106 static uint32_t written, written_delta;
107 static uint32_t reused, reused_delta;
108
109
110 static void *get_delta(struct object_entry *entry)
111 {
112         unsigned long size, base_size, delta_size;
113         void *buf, *base_buf, *delta_buf;
114         enum object_type type;
115
116         buf = read_sha1_file(entry->idx.sha1, &type, &size);
117         if (!buf)
118                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
119         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
120         if (!base_buf)
121                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
122         delta_buf = diff_delta(base_buf, base_size,
123                                buf, size, &delta_size, 0);
124         if (!delta_buf || delta_size != entry->delta_size)
125                 die("delta size changed");
126         free(buf);
127         free(base_buf);
128         return delta_buf;
129 }
130
131 static unsigned long do_compress(void **pptr, unsigned long size)
132 {
133         git_zstream stream;
134         void *in, *out;
135         unsigned long maxsize;
136
137         memset(&stream, 0, sizeof(stream));
138         git_deflate_init(&stream, pack_compression_level);
139         maxsize = git_deflate_bound(&stream, size);
140
141         in = *pptr;
142         out = xmalloc(maxsize);
143         *pptr = out;
144
145         stream.next_in = in;
146         stream.avail_in = size;
147         stream.next_out = out;
148         stream.avail_out = maxsize;
149         while (git_deflate(&stream, Z_FINISH) == Z_OK)
150                 ; /* nothing */
151         git_deflate_end(&stream);
152
153         free(in);
154         return stream.total_out;
155 }
156
157 /*
158  * we are going to reuse the existing object data as is.  make
159  * sure it is not corrupt.
160  */
161 static int check_pack_inflate(struct packed_git *p,
162                 struct pack_window **w_curs,
163                 off_t offset,
164                 off_t len,
165                 unsigned long expect)
166 {
167         git_zstream stream;
168         unsigned char fakebuf[4096], *in;
169         int st;
170
171         memset(&stream, 0, sizeof(stream));
172         git_inflate_init(&stream);
173         do {
174                 in = use_pack(p, w_curs, offset, &stream.avail_in);
175                 stream.next_in = in;
176                 stream.next_out = fakebuf;
177                 stream.avail_out = sizeof(fakebuf);
178                 st = git_inflate(&stream, Z_FINISH);
179                 offset += stream.next_in - in;
180         } while (st == Z_OK || st == Z_BUF_ERROR);
181         git_inflate_end(&stream);
182         return (st == Z_STREAM_END &&
183                 stream.total_out == expect &&
184                 stream.total_in == len) ? 0 : -1;
185 }
186
187 static void copy_pack_data(struct sha1file *f,
188                 struct packed_git *p,
189                 struct pack_window **w_curs,
190                 off_t offset,
191                 off_t len)
192 {
193         unsigned char *in;
194         unsigned long avail;
195
196         while (len) {
197                 in = use_pack(p, w_curs, offset, &avail);
198                 if (avail > len)
199                         avail = (unsigned long)len;
200                 sha1write(f, in, avail);
201                 offset += avail;
202                 len -= avail;
203         }
204 }
205
206 /* Return 0 if we will bust the pack-size limit */
207 static unsigned long write_object(struct sha1file *f,
208                                   struct object_entry *entry,
209                                   off_t write_offset)
210 {
211         unsigned long size, limit, datalen;
212         void *buf;
213         unsigned char header[10], dheader[10];
214         unsigned hdrlen;
215         enum object_type type;
216         int usable_delta, to_reuse;
217
218         if (!pack_to_stdout)
219                 crc32_begin(f);
220
221         type = entry->type;
222
223         /* apply size limit if limited packsize and not first object */
224         if (!pack_size_limit || !nr_written)
225                 limit = 0;
226         else if (pack_size_limit <= write_offset)
227                 /*
228                  * the earlier object did not fit the limit; avoid
229                  * mistaking this with unlimited (i.e. limit = 0).
230                  */
231                 limit = 1;
232         else
233                 limit = pack_size_limit - write_offset;
234
235         if (!entry->delta)
236                 usable_delta = 0;       /* no delta */
237         else if (!pack_size_limit)
238                usable_delta = 1;        /* unlimited packfile */
239         else if (entry->delta->idx.offset == (off_t)-1)
240                 usable_delta = 0;       /* base was written to another pack */
241         else if (entry->delta->idx.offset)
242                 usable_delta = 1;       /* base already exists in this pack */
243         else
244                 usable_delta = 0;       /* base could end up in another pack */
245
246         if (!reuse_object)
247                 to_reuse = 0;   /* explicit */
248         else if (!entry->in_pack)
249                 to_reuse = 0;   /* can't reuse what we don't have */
250         else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
251                                 /* check_object() decided it for us ... */
252                 to_reuse = usable_delta;
253                                 /* ... but pack split may override that */
254         else if (type != entry->in_pack_type)
255                 to_reuse = 0;   /* pack has delta which is unusable */
256         else if (entry->delta)
257                 to_reuse = 0;   /* we want to pack afresh */
258         else
259                 to_reuse = 1;   /* we have it in-pack undeltified,
260                                  * and we do not need to deltify it.
261                                  */
262
263         if (!to_reuse) {
264                 no_reuse:
265                 if (!usable_delta) {
266                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
267                         if (!buf)
268                                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
269                         /*
270                          * make sure no cached delta data remains from a
271                          * previous attempt before a pack split occurred.
272                          */
273                         free(entry->delta_data);
274                         entry->delta_data = NULL;
275                         entry->z_delta_size = 0;
276                 } else if (entry->delta_data) {
277                         size = entry->delta_size;
278                         buf = entry->delta_data;
279                         entry->delta_data = NULL;
280                         type = (allow_ofs_delta && entry->delta->idx.offset) ?
281                                 OBJ_OFS_DELTA : OBJ_REF_DELTA;
282                 } else {
283                         buf = get_delta(entry);
284                         size = entry->delta_size;
285                         type = (allow_ofs_delta && entry->delta->idx.offset) ?
286                                 OBJ_OFS_DELTA : OBJ_REF_DELTA;
287                 }
288
289                 if (entry->z_delta_size)
290                         datalen = entry->z_delta_size;
291                 else
292                         datalen = do_compress(&buf, size);
293
294                 /*
295                  * The object header is a byte of 'type' followed by zero or
296                  * more bytes of length.
297                  */
298                 hdrlen = encode_in_pack_object_header(type, size, header);
299
300                 if (type == OBJ_OFS_DELTA) {
301                         /*
302                          * Deltas with relative base contain an additional
303                          * encoding of the relative offset for the delta
304                          * base from this object's position in the pack.
305                          */
306                         off_t ofs = entry->idx.offset - entry->delta->idx.offset;
307                         unsigned pos = sizeof(dheader) - 1;
308                         dheader[pos] = ofs & 127;
309                         while (ofs >>= 7)
310                                 dheader[--pos] = 128 | (--ofs & 127);
311                         if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
312                                 free(buf);
313                                 return 0;
314                         }
315                         sha1write(f, header, hdrlen);
316                         sha1write(f, dheader + pos, sizeof(dheader) - pos);
317                         hdrlen += sizeof(dheader) - pos;
318                 } else if (type == OBJ_REF_DELTA) {
319                         /*
320                          * Deltas with a base reference contain
321                          * an additional 20 bytes for the base sha1.
322                          */
323                         if (limit && hdrlen + 20 + datalen + 20 >= limit) {
324                                 free(buf);
325                                 return 0;
326                         }
327                         sha1write(f, header, hdrlen);
328                         sha1write(f, entry->delta->idx.sha1, 20);
329                         hdrlen += 20;
330                 } else {
331                         if (limit && hdrlen + datalen + 20 >= limit) {
332                                 free(buf);
333                                 return 0;
334                         }
335                         sha1write(f, header, hdrlen);
336                 }
337                 sha1write(f, buf, datalen);
338                 free(buf);
339         }
340         else {
341                 struct packed_git *p = entry->in_pack;
342                 struct pack_window *w_curs = NULL;
343                 struct revindex_entry *revidx;
344                 off_t offset;
345
346                 if (entry->delta)
347                         type = (allow_ofs_delta && entry->delta->idx.offset) ?
348                                 OBJ_OFS_DELTA : OBJ_REF_DELTA;
349                 hdrlen = encode_in_pack_object_header(type, entry->size, header);
350
351                 offset = entry->in_pack_offset;
352                 revidx = find_pack_revindex(p, offset);
353                 datalen = revidx[1].offset - offset;
354                 if (!pack_to_stdout && p->index_version > 1 &&
355                     check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
356                         error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
357                         unuse_pack(&w_curs);
358                         goto no_reuse;
359                 }
360
361                 offset += entry->in_pack_header_size;
362                 datalen -= entry->in_pack_header_size;
363                 if (!pack_to_stdout && p->index_version == 1 &&
364                     check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
365                         error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
366                         unuse_pack(&w_curs);
367                         goto no_reuse;
368                 }
369
370                 if (type == OBJ_OFS_DELTA) {
371                         off_t ofs = entry->idx.offset - entry->delta->idx.offset;
372                         unsigned pos = sizeof(dheader) - 1;
373                         dheader[pos] = ofs & 127;
374                         while (ofs >>= 7)
375                                 dheader[--pos] = 128 | (--ofs & 127);
376                         if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
377                                 unuse_pack(&w_curs);
378                                 return 0;
379                         }
380                         sha1write(f, header, hdrlen);
381                         sha1write(f, dheader + pos, sizeof(dheader) - pos);
382                         hdrlen += sizeof(dheader) - pos;
383                         reused_delta++;
384                 } else if (type == OBJ_REF_DELTA) {
385                         if (limit && hdrlen + 20 + datalen + 20 >= limit) {
386                                 unuse_pack(&w_curs);
387                                 return 0;
388                         }
389                         sha1write(f, header, hdrlen);
390                         sha1write(f, entry->delta->idx.sha1, 20);
391                         hdrlen += 20;
392                         reused_delta++;
393                 } else {
394                         if (limit && hdrlen + datalen + 20 >= limit) {
395                                 unuse_pack(&w_curs);
396                                 return 0;
397                         }
398                         sha1write(f, header, hdrlen);
399                 }
400                 copy_pack_data(f, p, &w_curs, offset, datalen);
401                 unuse_pack(&w_curs);
402                 reused++;
403         }
404         if (usable_delta)
405                 written_delta++;
406         written++;
407         if (!pack_to_stdout)
408                 entry->idx.crc32 = crc32_end(f);
409         return hdrlen + datalen;
410 }
411
412 static int write_one(struct sha1file *f,
413                                struct object_entry *e,
414                                off_t *offset)
415 {
416         unsigned long size;
417
418         /* offset is non zero if object is written already. */
419         if (e->idx.offset || e->preferred_base)
420                 return -1;
421
422         /* if we are deltified, write out base object first. */
423         if (e->delta && !write_one(f, e->delta, offset))
424                 return 0;
425
426         e->idx.offset = *offset;
427         size = write_object(f, e, *offset);
428         if (!size) {
429                 e->idx.offset = 0;
430                 return 0;
431         }
432         written_list[nr_written++] = &e->idx;
433
434         /* make sure off_t is sufficiently large not to wrap */
435         if (signed_add_overflows(*offset, size))
436                 die("pack too large for current definition of off_t");
437         *offset += size;
438         return 1;
439 }
440
441 static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
442                        void *cb_data)
443 {
444         unsigned char peeled[20];
445         struct object_entry *entry = locate_object_entry(sha1);
446
447         if (entry)
448                 entry->tagged = 1;
449         if (!peel_ref(path, peeled)) {
450                 entry = locate_object_entry(peeled);
451                 if (entry)
452                         entry->tagged = 1;
453         }
454         return 0;
455 }
456
457 static void add_to_write_order(struct object_entry **wo,
458                                int *endp,
459                                struct object_entry *e)
460 {
461         if (e->filled)
462                 return;
463         wo[(*endp)++] = e;
464         e->filled = 1;
465 }
466
467 static void add_descendants_to_write_order(struct object_entry **wo,
468                                            int *endp,
469                                            struct object_entry *e)
470 {
471         struct object_entry *child;
472
473         for (child = e->delta_child; child; child = child->delta_sibling)
474                 add_to_write_order(wo, endp, child);
475         for (child = e->delta_child; child; child = child->delta_sibling)
476                 add_descendants_to_write_order(wo, endp, child);
477 }
478
479 static void add_family_to_write_order(struct object_entry **wo,
480                                       int *endp,
481                                       struct object_entry *e)
482 {
483         struct object_entry *root;
484
485         for (root = e; root->delta; root = root->delta)
486                 ; /* nothing */
487         add_to_write_order(wo, endp, root);
488         add_descendants_to_write_order(wo, endp, root);
489 }
490
491 static struct object_entry **compute_write_order(void)
492 {
493         int i, wo_end;
494
495         struct object_entry **wo = xmalloc(nr_objects * sizeof(*wo));
496
497         for (i = 0; i < nr_objects; i++) {
498                 objects[i].tagged = 0;
499                 objects[i].filled = 0;
500                 objects[i].delta_child = NULL;
501                 objects[i].delta_sibling = NULL;
502         }
503
504         /*
505          * Fully connect delta_child/delta_sibling network.
506          * Make sure delta_sibling is sorted in the original
507          * recency order.
508          */
509         for (i = nr_objects - 1; 0 <= i; i--) {
510                 struct object_entry *e = &objects[i];
511                 if (!e->delta)
512                         continue;
513                 /* Mark me as the first child */
514                 e->delta_sibling = e->delta->delta_child;
515                 e->delta->delta_child = e;
516         }
517
518         /*
519          * Mark objects that are at the tip of tags.
520          */
521         for_each_tag_ref(mark_tagged, NULL);
522
523         /*
524          * Give the commits in the original recency order until
525          * we see a tagged tip.
526          */
527         for (i = wo_end = 0; i < nr_objects; i++) {
528                 if (objects[i].tagged)
529                         break;
530                 add_to_write_order(wo, &wo_end, &objects[i]);
531         }
532
533         /*
534          * Then fill all the tagged tips.
535          */
536         for (; i < nr_objects; i++) {
537                 if (objects[i].tagged)
538                         add_to_write_order(wo, &wo_end, &objects[i]);
539         }
540
541         /*
542          * And then all remaining commits and tags.
543          */
544         for (i = 0; i < nr_objects; i++) {
545                 if (objects[i].type != OBJ_COMMIT &&
546                     objects[i].type != OBJ_TAG)
547                         continue;
548                 add_to_write_order(wo, &wo_end, &objects[i]);
549         }
550
551         /*
552          * And then all the trees.
553          */
554         for (i = 0; i < nr_objects; i++) {
555                 if (objects[i].type != OBJ_TREE)
556                         continue;
557                 add_to_write_order(wo, &wo_end, &objects[i]);
558         }
559
560         /*
561          * Finally all the rest in really tight order
562          */
563         for (i = 0; i < nr_objects; i++)
564                 add_family_to_write_order(wo, &wo_end, &objects[i]);
565
566         return wo;
567 }
568
569 static void write_pack_file(void)
570 {
571         uint32_t i = 0, j;
572         struct sha1file *f;
573         off_t offset;
574         struct pack_header hdr;
575         uint32_t nr_remaining = nr_result;
576         time_t last_mtime = 0;
577         struct object_entry **write_order;
578
579         if (progress > pack_to_stdout)
580                 progress_state = start_progress("Writing objects", nr_result);
581         written_list = xmalloc(nr_objects * sizeof(*written_list));
582         write_order = compute_write_order();
583
584         do {
585                 unsigned char sha1[20];
586                 char *pack_tmp_name = NULL;
587
588                 if (pack_to_stdout) {
589                         f = sha1fd_throughput(1, "<stdout>", progress_state);
590                 } else {
591                         char tmpname[PATH_MAX];
592                         int fd;
593                         fd = odb_mkstemp(tmpname, sizeof(tmpname),
594                                          "pack/tmp_pack_XXXXXX");
595                         pack_tmp_name = xstrdup(tmpname);
596                         f = sha1fd(fd, pack_tmp_name);
597                 }
598
599                 hdr.hdr_signature = htonl(PACK_SIGNATURE);
600                 hdr.hdr_version = htonl(PACK_VERSION);
601                 hdr.hdr_entries = htonl(nr_remaining);
602                 sha1write(f, &hdr, sizeof(hdr));
603                 offset = sizeof(hdr);
604                 nr_written = 0;
605                 for (; i < nr_objects; i++) {
606                         struct object_entry *e = write_order[i];
607                         if (!write_one(f, e, &offset))
608                                 break;
609                         display_progress(progress_state, written);
610                 }
611
612                 /*
613                  * Did we write the wrong # entries in the header?
614                  * If so, rewrite it like in fast-import
615                  */
616                 if (pack_to_stdout) {
617                         sha1close(f, sha1, CSUM_CLOSE);
618                 } else if (nr_written == nr_remaining) {
619                         sha1close(f, sha1, CSUM_FSYNC);
620                 } else {
621                         int fd = sha1close(f, sha1, 0);
622                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
623                                                  nr_written, sha1, offset);
624                         close(fd);
625                 }
626
627                 if (!pack_to_stdout) {
628                         struct stat st;
629                         const char *idx_tmp_name;
630                         char tmpname[PATH_MAX];
631
632                         idx_tmp_name = write_idx_file(NULL, written_list, nr_written,
633                                                       &pack_idx_opts, sha1);
634
635                         snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
636                                  base_name, sha1_to_hex(sha1));
637                         free_pack_by_name(tmpname);
638                         if (adjust_shared_perm(pack_tmp_name))
639                                 die_errno("unable to make temporary pack file readable");
640                         if (rename(pack_tmp_name, tmpname))
641                                 die_errno("unable to rename temporary pack file");
642
643                         /*
644                          * Packs are runtime accessed in their mtime
645                          * order since newer packs are more likely to contain
646                          * younger objects.  So if we are creating multiple
647                          * packs then we should modify the mtime of later ones
648                          * to preserve this property.
649                          */
650                         if (stat(tmpname, &st) < 0) {
651                                 warning("failed to stat %s: %s",
652                                         tmpname, strerror(errno));
653                         } else if (!last_mtime) {
654                                 last_mtime = st.st_mtime;
655                         } else {
656                                 struct utimbuf utb;
657                                 utb.actime = st.st_atime;
658                                 utb.modtime = --last_mtime;
659                                 if (utime(tmpname, &utb) < 0)
660                                         warning("failed utime() on %s: %s",
661                                                 tmpname, strerror(errno));
662                         }
663
664                         snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
665                                  base_name, sha1_to_hex(sha1));
666                         if (adjust_shared_perm(idx_tmp_name))
667                                 die_errno("unable to make temporary index file readable");
668                         if (rename(idx_tmp_name, tmpname))
669                                 die_errno("unable to rename temporary index file");
670
671                         free((void *) idx_tmp_name);
672                         free(pack_tmp_name);
673                         puts(sha1_to_hex(sha1));
674                 }
675
676                 /* mark written objects as written to previous pack */
677                 for (j = 0; j < nr_written; j++) {
678                         written_list[j]->offset = (off_t)-1;
679                 }
680                 nr_remaining -= nr_written;
681         } while (nr_remaining && i < nr_objects);
682
683         free(written_list);
684         free(write_order);
685         stop_progress(&progress_state);
686         if (written != nr_result)
687                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
688                         written, nr_result);
689 }
690
691 static int locate_object_entry_hash(const unsigned char *sha1)
692 {
693         int i;
694         unsigned int ui;
695         memcpy(&ui, sha1, sizeof(unsigned int));
696         i = ui % object_ix_hashsz;
697         while (0 < object_ix[i]) {
698                 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
699                         return i;
700                 if (++i == object_ix_hashsz)
701                         i = 0;
702         }
703         return -1 - i;
704 }
705
706 static struct object_entry *locate_object_entry(const unsigned char *sha1)
707 {
708         int i;
709
710         if (!object_ix_hashsz)
711                 return NULL;
712
713         i = locate_object_entry_hash(sha1);
714         if (0 <= i)
715                 return &objects[object_ix[i]-1];
716         return NULL;
717 }
718
719 static void rehash_objects(void)
720 {
721         uint32_t i;
722         struct object_entry *oe;
723
724         object_ix_hashsz = nr_objects * 3;
725         if (object_ix_hashsz < 1024)
726                 object_ix_hashsz = 1024;
727         object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
728         memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
729         for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
730                 int ix = locate_object_entry_hash(oe->idx.sha1);
731                 if (0 <= ix)
732                         continue;
733                 ix = -1 - ix;
734                 object_ix[ix] = i + 1;
735         }
736 }
737
738 static unsigned name_hash(const char *name)
739 {
740         unsigned c, hash = 0;
741
742         if (!name)
743                 return 0;
744
745         /*
746          * This effectively just creates a sortable number from the
747          * last sixteen non-whitespace characters. Last characters
748          * count "most", so things that end in ".c" sort together.
749          */
750         while ((c = *name++) != 0) {
751                 if (isspace(c))
752                         continue;
753                 hash = (hash >> 2) + (c << 24);
754         }
755         return hash;
756 }
757
758 static void setup_delta_attr_check(struct git_attr_check *check)
759 {
760         static struct git_attr *attr_delta;
761
762         if (!attr_delta)
763                 attr_delta = git_attr("delta");
764
765         check[0].attr = attr_delta;
766 }
767
768 static int no_try_delta(const char *path)
769 {
770         struct git_attr_check check[1];
771
772         setup_delta_attr_check(check);
773         if (git_check_attr(path, ARRAY_SIZE(check), check))
774                 return 0;
775         if (ATTR_FALSE(check->value))
776                 return 1;
777         return 0;
778 }
779
780 static int add_object_entry(const unsigned char *sha1, enum object_type type,
781                             const char *name, int exclude)
782 {
783         struct object_entry *entry;
784         struct packed_git *p, *found_pack = NULL;
785         off_t found_offset = 0;
786         int ix;
787         unsigned hash = name_hash(name);
788
789         ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
790         if (ix >= 0) {
791                 if (exclude) {
792                         entry = objects + object_ix[ix] - 1;
793                         if (!entry->preferred_base)
794                                 nr_result--;
795                         entry->preferred_base = 1;
796                 }
797                 return 0;
798         }
799
800         if (!exclude && local && has_loose_object_nonlocal(sha1))
801                 return 0;
802
803         for (p = packed_git; p; p = p->next) {
804                 off_t offset = find_pack_entry_one(sha1, p);
805                 if (offset) {
806                         if (!found_pack) {
807                                 found_offset = offset;
808                                 found_pack = p;
809                         }
810                         if (exclude)
811                                 break;
812                         if (incremental)
813                                 return 0;
814                         if (local && !p->pack_local)
815                                 return 0;
816                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
817                                 return 0;
818                 }
819         }
820
821         if (nr_objects >= nr_alloc) {
822                 nr_alloc = (nr_alloc  + 1024) * 3 / 2;
823                 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
824         }
825
826         entry = objects + nr_objects++;
827         memset(entry, 0, sizeof(*entry));
828         hashcpy(entry->idx.sha1, sha1);
829         entry->hash = hash;
830         if (type)
831                 entry->type = type;
832         if (exclude)
833                 entry->preferred_base = 1;
834         else
835                 nr_result++;
836         if (found_pack) {
837                 entry->in_pack = found_pack;
838                 entry->in_pack_offset = found_offset;
839         }
840
841         if (object_ix_hashsz * 3 <= nr_objects * 4)
842                 rehash_objects();
843         else
844                 object_ix[-1 - ix] = nr_objects;
845
846         display_progress(progress_state, nr_objects);
847
848         if (name && no_try_delta(name))
849                 entry->no_try_delta = 1;
850
851         return 1;
852 }
853
854 struct pbase_tree_cache {
855         unsigned char sha1[20];
856         int ref;
857         int temporary;
858         void *tree_data;
859         unsigned long tree_size;
860 };
861
862 static struct pbase_tree_cache *(pbase_tree_cache[256]);
863 static int pbase_tree_cache_ix(const unsigned char *sha1)
864 {
865         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
866 }
867 static int pbase_tree_cache_ix_incr(int ix)
868 {
869         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
870 }
871
872 static struct pbase_tree {
873         struct pbase_tree *next;
874         /* This is a phony "cache" entry; we are not
875          * going to evict it nor find it through _get()
876          * mechanism -- this is for the toplevel node that
877          * would almost always change with any commit.
878          */
879         struct pbase_tree_cache pcache;
880 } *pbase_tree;
881
882 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
883 {
884         struct pbase_tree_cache *ent, *nent;
885         void *data;
886         unsigned long size;
887         enum object_type type;
888         int neigh;
889         int my_ix = pbase_tree_cache_ix(sha1);
890         int available_ix = -1;
891
892         /* pbase-tree-cache acts as a limited hashtable.
893          * your object will be found at your index or within a few
894          * slots after that slot if it is cached.
895          */
896         for (neigh = 0; neigh < 8; neigh++) {
897                 ent = pbase_tree_cache[my_ix];
898                 if (ent && !hashcmp(ent->sha1, sha1)) {
899                         ent->ref++;
900                         return ent;
901                 }
902                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
903                          ((0 <= available_ix) &&
904                           (!ent && pbase_tree_cache[available_ix])))
905                         available_ix = my_ix;
906                 if (!ent)
907                         break;
908                 my_ix = pbase_tree_cache_ix_incr(my_ix);
909         }
910
911         /* Did not find one.  Either we got a bogus request or
912          * we need to read and perhaps cache.
913          */
914         data = read_sha1_file(sha1, &type, &size);
915         if (!data)
916                 return NULL;
917         if (type != OBJ_TREE) {
918                 free(data);
919                 return NULL;
920         }
921
922         /* We need to either cache or return a throwaway copy */
923
924         if (available_ix < 0)
925                 ent = NULL;
926         else {
927                 ent = pbase_tree_cache[available_ix];
928                 my_ix = available_ix;
929         }
930
931         if (!ent) {
932                 nent = xmalloc(sizeof(*nent));
933                 nent->temporary = (available_ix < 0);
934         }
935         else {
936                 /* evict and reuse */
937                 free(ent->tree_data);
938                 nent = ent;
939         }
940         hashcpy(nent->sha1, sha1);
941         nent->tree_data = data;
942         nent->tree_size = size;
943         nent->ref = 1;
944         if (!nent->temporary)
945                 pbase_tree_cache[my_ix] = nent;
946         return nent;
947 }
948
949 static void pbase_tree_put(struct pbase_tree_cache *cache)
950 {
951         if (!cache->temporary) {
952                 cache->ref--;
953                 return;
954         }
955         free(cache->tree_data);
956         free(cache);
957 }
958
959 static int name_cmp_len(const char *name)
960 {
961         int i;
962         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
963                 ;
964         return i;
965 }
966
967 static void add_pbase_object(struct tree_desc *tree,
968                              const char *name,
969                              int cmplen,
970                              const char *fullname)
971 {
972         struct name_entry entry;
973         int cmp;
974
975         while (tree_entry(tree,&entry)) {
976                 if (S_ISGITLINK(entry.mode))
977                         continue;
978                 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
979                       memcmp(name, entry.path, cmplen);
980                 if (cmp > 0)
981                         continue;
982                 if (cmp < 0)
983                         return;
984                 if (name[cmplen] != '/') {
985                         add_object_entry(entry.sha1,
986                                          object_type(entry.mode),
987                                          fullname, 1);
988                         return;
989                 }
990                 if (S_ISDIR(entry.mode)) {
991                         struct tree_desc sub;
992                         struct pbase_tree_cache *tree;
993                         const char *down = name+cmplen+1;
994                         int downlen = name_cmp_len(down);
995
996                         tree = pbase_tree_get(entry.sha1);
997                         if (!tree)
998                                 return;
999                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1000
1001                         add_pbase_object(&sub, down, downlen, fullname);
1002                         pbase_tree_put(tree);
1003                 }
1004         }
1005 }
1006
1007 static unsigned *done_pbase_paths;
1008 static int done_pbase_paths_num;
1009 static int done_pbase_paths_alloc;
1010 static int done_pbase_path_pos(unsigned hash)
1011 {
1012         int lo = 0;
1013         int hi = done_pbase_paths_num;
1014         while (lo < hi) {
1015                 int mi = (hi + lo) / 2;
1016                 if (done_pbase_paths[mi] == hash)
1017                         return mi;
1018                 if (done_pbase_paths[mi] < hash)
1019                         hi = mi;
1020                 else
1021                         lo = mi + 1;
1022         }
1023         return -lo-1;
1024 }
1025
1026 static int check_pbase_path(unsigned hash)
1027 {
1028         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1029         if (0 <= pos)
1030                 return 1;
1031         pos = -pos - 1;
1032         if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1033                 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1034                 done_pbase_paths = xrealloc(done_pbase_paths,
1035                                             done_pbase_paths_alloc *
1036                                             sizeof(unsigned));
1037         }
1038         done_pbase_paths_num++;
1039         if (pos < done_pbase_paths_num)
1040                 memmove(done_pbase_paths + pos + 1,
1041                         done_pbase_paths + pos,
1042                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1043         done_pbase_paths[pos] = hash;
1044         return 0;
1045 }
1046
1047 static void add_preferred_base_object(const char *name)
1048 {
1049         struct pbase_tree *it;
1050         int cmplen;
1051         unsigned hash = name_hash(name);
1052
1053         if (!num_preferred_base || check_pbase_path(hash))
1054                 return;
1055
1056         cmplen = name_cmp_len(name);
1057         for (it = pbase_tree; it; it = it->next) {
1058                 if (cmplen == 0) {
1059                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1060                 }
1061                 else {
1062                         struct tree_desc tree;
1063                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1064                         add_pbase_object(&tree, name, cmplen, name);
1065                 }
1066         }
1067 }
1068
1069 static void add_preferred_base(unsigned char *sha1)
1070 {
1071         struct pbase_tree *it;
1072         void *data;
1073         unsigned long size;
1074         unsigned char tree_sha1[20];
1075
1076         if (window <= num_preferred_base++)
1077                 return;
1078
1079         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1080         if (!data)
1081                 return;
1082
1083         for (it = pbase_tree; it; it = it->next) {
1084                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1085                         free(data);
1086                         return;
1087                 }
1088         }
1089
1090         it = xcalloc(1, sizeof(*it));
1091         it->next = pbase_tree;
1092         pbase_tree = it;
1093
1094         hashcpy(it->pcache.sha1, tree_sha1);
1095         it->pcache.tree_data = data;
1096         it->pcache.tree_size = size;
1097 }
1098
1099 static void cleanup_preferred_base(void)
1100 {
1101         struct pbase_tree *it;
1102         unsigned i;
1103
1104         it = pbase_tree;
1105         pbase_tree = NULL;
1106         while (it) {
1107                 struct pbase_tree *this = it;
1108                 it = this->next;
1109                 free(this->pcache.tree_data);
1110                 free(this);
1111         }
1112
1113         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1114                 if (!pbase_tree_cache[i])
1115                         continue;
1116                 free(pbase_tree_cache[i]->tree_data);
1117                 free(pbase_tree_cache[i]);
1118                 pbase_tree_cache[i] = NULL;
1119         }
1120
1121         free(done_pbase_paths);
1122         done_pbase_paths = NULL;
1123         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1124 }
1125
1126 static void check_object(struct object_entry *entry)
1127 {
1128         if (entry->in_pack) {
1129                 struct packed_git *p = entry->in_pack;
1130                 struct pack_window *w_curs = NULL;
1131                 const unsigned char *base_ref = NULL;
1132                 struct object_entry *base_entry;
1133                 unsigned long used, used_0;
1134                 unsigned long avail;
1135                 off_t ofs;
1136                 unsigned char *buf, c;
1137
1138                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1139
1140                 /*
1141                  * We want in_pack_type even if we do not reuse delta
1142                  * since non-delta representations could still be reused.
1143                  */
1144                 used = unpack_object_header_buffer(buf, avail,
1145                                                    &entry->in_pack_type,
1146                                                    &entry->size);
1147                 if (used == 0)
1148                         goto give_up;
1149
1150                 /*
1151                  * Determine if this is a delta and if so whether we can
1152                  * reuse it or not.  Otherwise let's find out as cheaply as
1153                  * possible what the actual type and size for this object is.
1154                  */
1155                 switch (entry->in_pack_type) {
1156                 default:
1157                         /* Not a delta hence we've already got all we need. */
1158                         entry->type = entry->in_pack_type;
1159                         entry->in_pack_header_size = used;
1160                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1161                                 goto give_up;
1162                         unuse_pack(&w_curs);
1163                         return;
1164                 case OBJ_REF_DELTA:
1165                         if (reuse_delta && !entry->preferred_base)
1166                                 base_ref = use_pack(p, &w_curs,
1167                                                 entry->in_pack_offset + used, NULL);
1168                         entry->in_pack_header_size = used + 20;
1169                         break;
1170                 case OBJ_OFS_DELTA:
1171                         buf = use_pack(p, &w_curs,
1172                                        entry->in_pack_offset + used, NULL);
1173                         used_0 = 0;
1174                         c = buf[used_0++];
1175                         ofs = c & 127;
1176                         while (c & 128) {
1177                                 ofs += 1;
1178                                 if (!ofs || MSB(ofs, 7)) {
1179                                         error("delta base offset overflow in pack for %s",
1180                                               sha1_to_hex(entry->idx.sha1));
1181                                         goto give_up;
1182                                 }
1183                                 c = buf[used_0++];
1184                                 ofs = (ofs << 7) + (c & 127);
1185                         }
1186                         ofs = entry->in_pack_offset - ofs;
1187                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1188                                 error("delta base offset out of bound for %s",
1189                                       sha1_to_hex(entry->idx.sha1));
1190                                 goto give_up;
1191                         }
1192                         if (reuse_delta && !entry->preferred_base) {
1193                                 struct revindex_entry *revidx;
1194                                 revidx = find_pack_revindex(p, ofs);
1195                                 if (!revidx)
1196                                         goto give_up;
1197                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1198                         }
1199                         entry->in_pack_header_size = used + used_0;
1200                         break;
1201                 }
1202
1203                 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1204                         /*
1205                          * If base_ref was set above that means we wish to
1206                          * reuse delta data, and we even found that base
1207                          * in the list of objects we want to pack. Goodie!
1208                          *
1209                          * Depth value does not matter - find_deltas() will
1210                          * never consider reused delta as the base object to
1211                          * deltify other objects against, in order to avoid
1212                          * circular deltas.
1213                          */
1214                         entry->type = entry->in_pack_type;
1215                         entry->delta = base_entry;
1216                         entry->delta_size = entry->size;
1217                         entry->delta_sibling = base_entry->delta_child;
1218                         base_entry->delta_child = entry;
1219                         unuse_pack(&w_curs);
1220                         return;
1221                 }
1222
1223                 if (entry->type) {
1224                         /*
1225                          * This must be a delta and we already know what the
1226                          * final object type is.  Let's extract the actual
1227                          * object size from the delta header.
1228                          */
1229                         entry->size = get_size_from_delta(p, &w_curs,
1230                                         entry->in_pack_offset + entry->in_pack_header_size);
1231                         if (entry->size == 0)
1232                                 goto give_up;
1233                         unuse_pack(&w_curs);
1234                         return;
1235                 }
1236
1237                 /*
1238                  * No choice but to fall back to the recursive delta walk
1239                  * with sha1_object_info() to find about the object type
1240                  * at this point...
1241                  */
1242                 give_up:
1243                 unuse_pack(&w_curs);
1244         }
1245
1246         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1247         /*
1248          * The error condition is checked in prepare_pack().  This is
1249          * to permit a missing preferred base object to be ignored
1250          * as a preferred base.  Doing so can result in a larger
1251          * pack file, but the transfer will still take place.
1252          */
1253 }
1254
1255 static int pack_offset_sort(const void *_a, const void *_b)
1256 {
1257         const struct object_entry *a = *(struct object_entry **)_a;
1258         const struct object_entry *b = *(struct object_entry **)_b;
1259
1260         /* avoid filesystem trashing with loose objects */
1261         if (!a->in_pack && !b->in_pack)
1262                 return hashcmp(a->idx.sha1, b->idx.sha1);
1263
1264         if (a->in_pack < b->in_pack)
1265                 return -1;
1266         if (a->in_pack > b->in_pack)
1267                 return 1;
1268         return a->in_pack_offset < b->in_pack_offset ? -1 :
1269                         (a->in_pack_offset > b->in_pack_offset);
1270 }
1271
1272 static void get_object_details(void)
1273 {
1274         uint32_t i;
1275         struct object_entry **sorted_by_offset;
1276
1277         sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1278         for (i = 0; i < nr_objects; i++)
1279                 sorted_by_offset[i] = objects + i;
1280         qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1281
1282         for (i = 0; i < nr_objects; i++) {
1283                 struct object_entry *entry = sorted_by_offset[i];
1284                 check_object(entry);
1285                 if (big_file_threshold <= entry->size)
1286                         entry->no_try_delta = 1;
1287         }
1288
1289         free(sorted_by_offset);
1290 }
1291
1292 /*
1293  * We search for deltas in a list sorted by type, by filename hash, and then
1294  * by size, so that we see progressively smaller and smaller files.
1295  * That's because we prefer deltas to be from the bigger file
1296  * to the smaller -- deletes are potentially cheaper, but perhaps
1297  * more importantly, the bigger file is likely the more recent
1298  * one.  The deepest deltas are therefore the oldest objects which are
1299  * less susceptible to be accessed often.
1300  */
1301 static int type_size_sort(const void *_a, const void *_b)
1302 {
1303         const struct object_entry *a = *(struct object_entry **)_a;
1304         const struct object_entry *b = *(struct object_entry **)_b;
1305
1306         if (a->type > b->type)
1307                 return -1;
1308         if (a->type < b->type)
1309                 return 1;
1310         if (a->hash > b->hash)
1311                 return -1;
1312         if (a->hash < b->hash)
1313                 return 1;
1314         if (a->preferred_base > b->preferred_base)
1315                 return -1;
1316         if (a->preferred_base < b->preferred_base)
1317                 return 1;
1318         if (a->size > b->size)
1319                 return -1;
1320         if (a->size < b->size)
1321                 return 1;
1322         return a < b ? -1 : (a > b);  /* newest first */
1323 }
1324
1325 struct unpacked {
1326         struct object_entry *entry;
1327         void *data;
1328         struct delta_index *index;
1329         unsigned depth;
1330 };
1331
1332 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1333                            unsigned long delta_size)
1334 {
1335         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1336                 return 0;
1337
1338         if (delta_size < cache_max_small_delta_size)
1339                 return 1;
1340
1341         /* cache delta, if objects are large enough compared to delta size */
1342         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1343                 return 1;
1344
1345         return 0;
1346 }
1347
1348 #ifndef NO_PTHREADS
1349
1350 static pthread_mutex_t read_mutex;
1351 #define read_lock()             pthread_mutex_lock(&read_mutex)
1352 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1353
1354 static pthread_mutex_t cache_mutex;
1355 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1356 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1357
1358 static pthread_mutex_t progress_mutex;
1359 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1360 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1361
1362 #else
1363
1364 #define read_lock()             (void)0
1365 #define read_unlock()           (void)0
1366 #define cache_lock()            (void)0
1367 #define cache_unlock()          (void)0
1368 #define progress_lock()         (void)0
1369 #define progress_unlock()       (void)0
1370
1371 #endif
1372
1373 static int try_delta(struct unpacked *trg, struct unpacked *src,
1374                      unsigned max_depth, unsigned long *mem_usage)
1375 {
1376         struct object_entry *trg_entry = trg->entry;
1377         struct object_entry *src_entry = src->entry;
1378         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1379         unsigned ref_depth;
1380         enum object_type type;
1381         void *delta_buf;
1382
1383         /* Don't bother doing diffs between different types */
1384         if (trg_entry->type != src_entry->type)
1385                 return -1;
1386
1387         /*
1388          * We do not bother to try a delta that we discarded
1389          * on an earlier try, but only when reusing delta data.
1390          */
1391         if (reuse_delta && trg_entry->in_pack &&
1392             trg_entry->in_pack == src_entry->in_pack &&
1393             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1394             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1395                 return 0;
1396
1397         /* Let's not bust the allowed depth. */
1398         if (src->depth >= max_depth)
1399                 return 0;
1400
1401         /* Now some size filtering heuristics. */
1402         trg_size = trg_entry->size;
1403         if (!trg_entry->delta) {
1404                 max_size = trg_size/2 - 20;
1405                 ref_depth = 1;
1406         } else {
1407                 max_size = trg_entry->delta_size;
1408                 ref_depth = trg->depth;
1409         }
1410         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1411                                                 (max_depth - ref_depth + 1);
1412         if (max_size == 0)
1413                 return 0;
1414         src_size = src_entry->size;
1415         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1416         if (sizediff >= max_size)
1417                 return 0;
1418         if (trg_size < src_size / 32)
1419                 return 0;
1420
1421         /* Load data if not already done */
1422         if (!trg->data) {
1423                 read_lock();
1424                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1425                 read_unlock();
1426                 if (!trg->data)
1427                         die("object %s cannot be read",
1428                             sha1_to_hex(trg_entry->idx.sha1));
1429                 if (sz != trg_size)
1430                         die("object %s inconsistent object length (%lu vs %lu)",
1431                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1432                 *mem_usage += sz;
1433         }
1434         if (!src->data) {
1435                 read_lock();
1436                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1437                 read_unlock();
1438                 if (!src->data) {
1439                         if (src_entry->preferred_base) {
1440                                 static int warned = 0;
1441                                 if (!warned++)
1442                                         warning("object %s cannot be read",
1443                                                 sha1_to_hex(src_entry->idx.sha1));
1444                                 /*
1445                                  * Those objects are not included in the
1446                                  * resulting pack.  Be resilient and ignore
1447                                  * them if they can't be read, in case the
1448                                  * pack could be created nevertheless.
1449                                  */
1450                                 return 0;
1451                         }
1452                         die("object %s cannot be read",
1453                             sha1_to_hex(src_entry->idx.sha1));
1454                 }
1455                 if (sz != src_size)
1456                         die("object %s inconsistent object length (%lu vs %lu)",
1457                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1458                 *mem_usage += sz;
1459         }
1460         if (!src->index) {
1461                 src->index = create_delta_index(src->data, src_size);
1462                 if (!src->index) {
1463                         static int warned = 0;
1464                         if (!warned++)
1465                                 warning("suboptimal pack - out of memory");
1466                         return 0;
1467                 }
1468                 *mem_usage += sizeof_delta_index(src->index);
1469         }
1470
1471         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1472         if (!delta_buf)
1473                 return 0;
1474
1475         if (trg_entry->delta) {
1476                 /* Prefer only shallower same-sized deltas. */
1477                 if (delta_size == trg_entry->delta_size &&
1478                     src->depth + 1 >= trg->depth) {
1479                         free(delta_buf);
1480                         return 0;
1481                 }
1482         }
1483
1484         /*
1485          * Handle memory allocation outside of the cache
1486          * accounting lock.  Compiler will optimize the strangeness
1487          * away when NO_PTHREADS is defined.
1488          */
1489         free(trg_entry->delta_data);
1490         cache_lock();
1491         if (trg_entry->delta_data) {
1492                 delta_cache_size -= trg_entry->delta_size;
1493                 trg_entry->delta_data = NULL;
1494         }
1495         if (delta_cacheable(src_size, trg_size, delta_size)) {
1496                 delta_cache_size += delta_size;
1497                 cache_unlock();
1498                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1499         } else {
1500                 cache_unlock();
1501                 free(delta_buf);
1502         }
1503
1504         trg_entry->delta = src_entry;
1505         trg_entry->delta_size = delta_size;
1506         trg->depth = src->depth + 1;
1507
1508         return 1;
1509 }
1510
1511 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1512 {
1513         struct object_entry *child = me->delta_child;
1514         unsigned int m = n;
1515         while (child) {
1516                 unsigned int c = check_delta_limit(child, n + 1);
1517                 if (m < c)
1518                         m = c;
1519                 child = child->delta_sibling;
1520         }
1521         return m;
1522 }
1523
1524 static unsigned long free_unpacked(struct unpacked *n)
1525 {
1526         unsigned long freed_mem = sizeof_delta_index(n->index);
1527         free_delta_index(n->index);
1528         n->index = NULL;
1529         if (n->data) {
1530                 freed_mem += n->entry->size;
1531                 free(n->data);
1532                 n->data = NULL;
1533         }
1534         n->entry = NULL;
1535         n->depth = 0;
1536         return freed_mem;
1537 }
1538
1539 static void find_deltas(struct object_entry **list, unsigned *list_size,
1540                         int window, int depth, unsigned *processed)
1541 {
1542         uint32_t i, idx = 0, count = 0;
1543         struct unpacked *array;
1544         unsigned long mem_usage = 0;
1545
1546         array = xcalloc(window, sizeof(struct unpacked));
1547
1548         for (;;) {
1549                 struct object_entry *entry;
1550                 struct unpacked *n = array + idx;
1551                 int j, max_depth, best_base = -1;
1552
1553                 progress_lock();
1554                 if (!*list_size) {
1555                         progress_unlock();
1556                         break;
1557                 }
1558                 entry = *list++;
1559                 (*list_size)--;
1560                 if (!entry->preferred_base) {
1561                         (*processed)++;
1562                         display_progress(progress_state, *processed);
1563                 }
1564                 progress_unlock();
1565
1566                 mem_usage -= free_unpacked(n);
1567                 n->entry = entry;
1568
1569                 while (window_memory_limit &&
1570                        mem_usage > window_memory_limit &&
1571                        count > 1) {
1572                         uint32_t tail = (idx + window - count) % window;
1573                         mem_usage -= free_unpacked(array + tail);
1574                         count--;
1575                 }
1576
1577                 /* We do not compute delta to *create* objects we are not
1578                  * going to pack.
1579                  */
1580                 if (entry->preferred_base)
1581                         goto next;
1582
1583                 /*
1584                  * If the current object is at pack edge, take the depth the
1585                  * objects that depend on the current object into account
1586                  * otherwise they would become too deep.
1587                  */
1588                 max_depth = depth;
1589                 if (entry->delta_child) {
1590                         max_depth -= check_delta_limit(entry, 0);
1591                         if (max_depth <= 0)
1592                                 goto next;
1593                 }
1594
1595                 j = window;
1596                 while (--j > 0) {
1597                         int ret;
1598                         uint32_t other_idx = idx + j;
1599                         struct unpacked *m;
1600                         if (other_idx >= window)
1601                                 other_idx -= window;
1602                         m = array + other_idx;
1603                         if (!m->entry)
1604                                 break;
1605                         ret = try_delta(n, m, max_depth, &mem_usage);
1606                         if (ret < 0)
1607                                 break;
1608                         else if (ret > 0)
1609                                 best_base = other_idx;
1610                 }
1611
1612                 /*
1613                  * If we decided to cache the delta data, then it is best
1614                  * to compress it right away.  First because we have to do
1615                  * it anyway, and doing it here while we're threaded will
1616                  * save a lot of time in the non threaded write phase,
1617                  * as well as allow for caching more deltas within
1618                  * the same cache size limit.
1619                  * ...
1620                  * But only if not writing to stdout, since in that case
1621                  * the network is most likely throttling writes anyway,
1622                  * and therefore it is best to go to the write phase ASAP
1623                  * instead, as we can afford spending more time compressing
1624                  * between writes at that moment.
1625                  */
1626                 if (entry->delta_data && !pack_to_stdout) {
1627                         entry->z_delta_size = do_compress(&entry->delta_data,
1628                                                           entry->delta_size);
1629                         cache_lock();
1630                         delta_cache_size -= entry->delta_size;
1631                         delta_cache_size += entry->z_delta_size;
1632                         cache_unlock();
1633                 }
1634
1635                 /* if we made n a delta, and if n is already at max
1636                  * depth, leaving it in the window is pointless.  we
1637                  * should evict it first.
1638                  */
1639                 if (entry->delta && max_depth <= n->depth)
1640                         continue;
1641
1642                 /*
1643                  * Move the best delta base up in the window, after the
1644                  * currently deltified object, to keep it longer.  It will
1645                  * be the first base object to be attempted next.
1646                  */
1647                 if (entry->delta) {
1648                         struct unpacked swap = array[best_base];
1649                         int dist = (window + idx - best_base) % window;
1650                         int dst = best_base;
1651                         while (dist--) {
1652                                 int src = (dst + 1) % window;
1653                                 array[dst] = array[src];
1654                                 dst = src;
1655                         }
1656                         array[dst] = swap;
1657                 }
1658
1659                 next:
1660                 idx++;
1661                 if (count + 1 < window)
1662                         count++;
1663                 if (idx >= window)
1664                         idx = 0;
1665         }
1666
1667         for (i = 0; i < window; ++i) {
1668                 free_delta_index(array[i].index);
1669                 free(array[i].data);
1670         }
1671         free(array);
1672 }
1673
1674 #ifndef NO_PTHREADS
1675
1676 static void try_to_free_from_threads(size_t size)
1677 {
1678         read_lock();
1679         release_pack_memory(size, -1);
1680         read_unlock();
1681 }
1682
1683 static try_to_free_t old_try_to_free_routine;
1684
1685 /*
1686  * The main thread waits on the condition that (at least) one of the workers
1687  * has stopped working (which is indicated in the .working member of
1688  * struct thread_params).
1689  * When a work thread has completed its work, it sets .working to 0 and
1690  * signals the main thread and waits on the condition that .data_ready
1691  * becomes 1.
1692  */
1693
1694 struct thread_params {
1695         pthread_t thread;
1696         struct object_entry **list;
1697         unsigned list_size;
1698         unsigned remaining;
1699         int window;
1700         int depth;
1701         int working;
1702         int data_ready;
1703         pthread_mutex_t mutex;
1704         pthread_cond_t cond;
1705         unsigned *processed;
1706 };
1707
1708 static pthread_cond_t progress_cond;
1709
1710 /*
1711  * Mutex and conditional variable can't be statically-initialized on Windows.
1712  */
1713 static void init_threaded_search(void)
1714 {
1715         init_recursive_mutex(&read_mutex);
1716         pthread_mutex_init(&cache_mutex, NULL);
1717         pthread_mutex_init(&progress_mutex, NULL);
1718         pthread_cond_init(&progress_cond, NULL);
1719         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1720 }
1721
1722 static void cleanup_threaded_search(void)
1723 {
1724         set_try_to_free_routine(old_try_to_free_routine);
1725         pthread_cond_destroy(&progress_cond);
1726         pthread_mutex_destroy(&read_mutex);
1727         pthread_mutex_destroy(&cache_mutex);
1728         pthread_mutex_destroy(&progress_mutex);
1729 }
1730
1731 static void *threaded_find_deltas(void *arg)
1732 {
1733         struct thread_params *me = arg;
1734
1735         while (me->remaining) {
1736                 find_deltas(me->list, &me->remaining,
1737                             me->window, me->depth, me->processed);
1738
1739                 progress_lock();
1740                 me->working = 0;
1741                 pthread_cond_signal(&progress_cond);
1742                 progress_unlock();
1743
1744                 /*
1745                  * We must not set ->data_ready before we wait on the
1746                  * condition because the main thread may have set it to 1
1747                  * before we get here. In order to be sure that new
1748                  * work is available if we see 1 in ->data_ready, it
1749                  * was initialized to 0 before this thread was spawned
1750                  * and we reset it to 0 right away.
1751                  */
1752                 pthread_mutex_lock(&me->mutex);
1753                 while (!me->data_ready)
1754                         pthread_cond_wait(&me->cond, &me->mutex);
1755                 me->data_ready = 0;
1756                 pthread_mutex_unlock(&me->mutex);
1757         }
1758         /* leave ->working 1 so that this doesn't get more work assigned */
1759         return NULL;
1760 }
1761
1762 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1763                            int window, int depth, unsigned *processed)
1764 {
1765         struct thread_params *p;
1766         int i, ret, active_threads = 0;
1767
1768         init_threaded_search();
1769
1770         if (!delta_search_threads)      /* --threads=0 means autodetect */
1771                 delta_search_threads = online_cpus();
1772         if (delta_search_threads <= 1) {
1773                 find_deltas(list, &list_size, window, depth, processed);
1774                 cleanup_threaded_search();
1775                 return;
1776         }
1777         if (progress > pack_to_stdout)
1778                 fprintf(stderr, "Delta compression using up to %d threads.\n",
1779                                 delta_search_threads);
1780         p = xcalloc(delta_search_threads, sizeof(*p));
1781
1782         /* Partition the work amongst work threads. */
1783         for (i = 0; i < delta_search_threads; i++) {
1784                 unsigned sub_size = list_size / (delta_search_threads - i);
1785
1786                 /* don't use too small segments or no deltas will be found */
1787                 if (sub_size < 2*window && i+1 < delta_search_threads)
1788                         sub_size = 0;
1789
1790                 p[i].window = window;
1791                 p[i].depth = depth;
1792                 p[i].processed = processed;
1793                 p[i].working = 1;
1794                 p[i].data_ready = 0;
1795
1796                 /* try to split chunks on "path" boundaries */
1797                 while (sub_size && sub_size < list_size &&
1798                        list[sub_size]->hash &&
1799                        list[sub_size]->hash == list[sub_size-1]->hash)
1800                         sub_size++;
1801
1802                 p[i].list = list;
1803                 p[i].list_size = sub_size;
1804                 p[i].remaining = sub_size;
1805
1806                 list += sub_size;
1807                 list_size -= sub_size;
1808         }
1809
1810         /* Start work threads. */
1811         for (i = 0; i < delta_search_threads; i++) {
1812                 if (!p[i].list_size)
1813                         continue;
1814                 pthread_mutex_init(&p[i].mutex, NULL);
1815                 pthread_cond_init(&p[i].cond, NULL);
1816                 ret = pthread_create(&p[i].thread, NULL,
1817                                      threaded_find_deltas, &p[i]);
1818                 if (ret)
1819                         die("unable to create thread: %s", strerror(ret));
1820                 active_threads++;
1821         }
1822
1823         /*
1824          * Now let's wait for work completion.  Each time a thread is done
1825          * with its work, we steal half of the remaining work from the
1826          * thread with the largest number of unprocessed objects and give
1827          * it to that newly idle thread.  This ensure good load balancing
1828          * until the remaining object list segments are simply too short
1829          * to be worth splitting anymore.
1830          */
1831         while (active_threads) {
1832                 struct thread_params *target = NULL;
1833                 struct thread_params *victim = NULL;
1834                 unsigned sub_size = 0;
1835
1836                 progress_lock();
1837                 for (;;) {
1838                         for (i = 0; !target && i < delta_search_threads; i++)
1839                                 if (!p[i].working)
1840                                         target = &p[i];
1841                         if (target)
1842                                 break;
1843                         pthread_cond_wait(&progress_cond, &progress_mutex);
1844                 }
1845
1846                 for (i = 0; i < delta_search_threads; i++)
1847                         if (p[i].remaining > 2*window &&
1848                             (!victim || victim->remaining < p[i].remaining))
1849                                 victim = &p[i];
1850                 if (victim) {
1851                         sub_size = victim->remaining / 2;
1852                         list = victim->list + victim->list_size - sub_size;
1853                         while (sub_size && list[0]->hash &&
1854                                list[0]->hash == list[-1]->hash) {
1855                                 list++;
1856                                 sub_size--;
1857                         }
1858                         if (!sub_size) {
1859                                 /*
1860                                  * It is possible for some "paths" to have
1861                                  * so many objects that no hash boundary
1862                                  * might be found.  Let's just steal the
1863                                  * exact half in that case.
1864                                  */
1865                                 sub_size = victim->remaining / 2;
1866                                 list -= sub_size;
1867                         }
1868                         target->list = list;
1869                         victim->list_size -= sub_size;
1870                         victim->remaining -= sub_size;
1871                 }
1872                 target->list_size = sub_size;
1873                 target->remaining = sub_size;
1874                 target->working = 1;
1875                 progress_unlock();
1876
1877                 pthread_mutex_lock(&target->mutex);
1878                 target->data_ready = 1;
1879                 pthread_cond_signal(&target->cond);
1880                 pthread_mutex_unlock(&target->mutex);
1881
1882                 if (!sub_size) {
1883                         pthread_join(target->thread, NULL);
1884                         pthread_cond_destroy(&target->cond);
1885                         pthread_mutex_destroy(&target->mutex);
1886                         active_threads--;
1887                 }
1888         }
1889         cleanup_threaded_search();
1890         free(p);
1891 }
1892
1893 #else
1894 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
1895 #endif
1896
1897 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1898 {
1899         unsigned char peeled[20];
1900
1901         if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1902             !peel_ref(path, peeled)        && /* peelable? */
1903             !is_null_sha1(peeled)          && /* annotated tag? */
1904             locate_object_entry(peeled))      /* object packed? */
1905                 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1906         return 0;
1907 }
1908
1909 static void prepare_pack(int window, int depth)
1910 {
1911         struct object_entry **delta_list;
1912         uint32_t i, nr_deltas;
1913         unsigned n;
1914
1915         get_object_details();
1916
1917         /*
1918          * If we're locally repacking then we need to be doubly careful
1919          * from now on in order to make sure no stealth corruption gets
1920          * propagated to the new pack.  Clients receiving streamed packs
1921          * should validate everything they get anyway so no need to incur
1922          * the additional cost here in that case.
1923          */
1924         if (!pack_to_stdout)
1925                 do_check_packed_object_crc = 1;
1926
1927         if (!nr_objects || !window || !depth)
1928                 return;
1929
1930         delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1931         nr_deltas = n = 0;
1932
1933         for (i = 0; i < nr_objects; i++) {
1934                 struct object_entry *entry = objects + i;
1935
1936                 if (entry->delta)
1937                         /* This happens if we decided to reuse existing
1938                          * delta from a pack.  "reuse_delta &&" is implied.
1939                          */
1940                         continue;
1941
1942                 if (entry->size < 50)
1943                         continue;
1944
1945                 if (entry->no_try_delta)
1946                         continue;
1947
1948                 if (!entry->preferred_base) {
1949                         nr_deltas++;
1950                         if (entry->type < 0)
1951                                 die("unable to get type of object %s",
1952                                     sha1_to_hex(entry->idx.sha1));
1953                 } else {
1954                         if (entry->type < 0) {
1955                                 /*
1956                                  * This object is not found, but we
1957                                  * don't have to include it anyway.
1958                                  */
1959                                 continue;
1960                         }
1961                 }
1962
1963                 delta_list[n++] = entry;
1964         }
1965
1966         if (nr_deltas && n > 1) {
1967                 unsigned nr_done = 0;
1968                 if (progress)
1969                         progress_state = start_progress("Compressing objects",
1970                                                         nr_deltas);
1971                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1972                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1973                 stop_progress(&progress_state);
1974                 if (nr_done != nr_deltas)
1975                         die("inconsistency with delta count");
1976         }
1977         free(delta_list);
1978 }
1979
1980 static int git_pack_config(const char *k, const char *v, void *cb)
1981 {
1982         if (!strcmp(k, "pack.window")) {
1983                 window = git_config_int(k, v);
1984                 return 0;
1985         }
1986         if (!strcmp(k, "pack.windowmemory")) {
1987                 window_memory_limit = git_config_ulong(k, v);
1988                 return 0;
1989         }
1990         if (!strcmp(k, "pack.depth")) {
1991                 depth = git_config_int(k, v);
1992                 return 0;
1993         }
1994         if (!strcmp(k, "pack.compression")) {
1995                 int level = git_config_int(k, v);
1996                 if (level == -1)
1997                         level = Z_DEFAULT_COMPRESSION;
1998                 else if (level < 0 || level > Z_BEST_COMPRESSION)
1999                         die("bad pack compression level %d", level);
2000                 pack_compression_level = level;
2001                 pack_compression_seen = 1;
2002                 return 0;
2003         }
2004         if (!strcmp(k, "pack.deltacachesize")) {
2005                 max_delta_cache_size = git_config_int(k, v);
2006                 return 0;
2007         }
2008         if (!strcmp(k, "pack.deltacachelimit")) {
2009                 cache_max_small_delta_size = git_config_int(k, v);
2010                 return 0;
2011         }
2012         if (!strcmp(k, "pack.threads")) {
2013                 delta_search_threads = git_config_int(k, v);
2014                 if (delta_search_threads < 0)
2015                         die("invalid number of threads specified (%d)",
2016                             delta_search_threads);
2017 #ifdef NO_PTHREADS
2018                 if (delta_search_threads != 1)
2019                         warning("no threads support, ignoring %s", k);
2020 #endif
2021                 return 0;
2022         }
2023         if (!strcmp(k, "pack.indexversion")) {
2024                 pack_idx_opts.version = git_config_int(k, v);
2025                 if (pack_idx_opts.version > 2)
2026                         die("bad pack.indexversion=%"PRIu32,
2027                             pack_idx_opts.version);
2028                 return 0;
2029         }
2030         if (!strcmp(k, "pack.packsizelimit")) {
2031                 pack_size_limit_cfg = git_config_ulong(k, v);
2032                 return 0;
2033         }
2034         return git_default_config(k, v, cb);
2035 }
2036
2037 static void read_object_list_from_stdin(void)
2038 {
2039         char line[40 + 1 + PATH_MAX + 2];
2040         unsigned char sha1[20];
2041
2042         for (;;) {
2043                 if (!fgets(line, sizeof(line), stdin)) {
2044                         if (feof(stdin))
2045                                 break;
2046                         if (!ferror(stdin))
2047                                 die("fgets returned NULL, not EOF, not error!");
2048                         if (errno != EINTR)
2049                                 die_errno("fgets");
2050                         clearerr(stdin);
2051                         continue;
2052                 }
2053                 if (line[0] == '-') {
2054                         if (get_sha1_hex(line+1, sha1))
2055                                 die("expected edge sha1, got garbage:\n %s",
2056                                     line);
2057                         add_preferred_base(sha1);
2058                         continue;
2059                 }
2060                 if (get_sha1_hex(line, sha1))
2061                         die("expected sha1, got garbage:\n %s", line);
2062
2063                 add_preferred_base_object(line+41);
2064                 add_object_entry(sha1, 0, line+41, 0);
2065         }
2066 }
2067
2068 #define OBJECT_ADDED (1u<<20)
2069
2070 static void show_commit(struct commit *commit, void *data)
2071 {
2072         add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2073         commit->object.flags |= OBJECT_ADDED;
2074 }
2075
2076 static void show_object(struct object *obj,
2077                         const struct name_path *path, const char *last,
2078                         void *data)
2079 {
2080         char *name = path_name(path, last);
2081
2082         add_preferred_base_object(name);
2083         add_object_entry(obj->sha1, obj->type, name, 0);
2084         obj->flags |= OBJECT_ADDED;
2085
2086         /*
2087          * We will have generated the hash from the name,
2088          * but not saved a pointer to it - we can free it
2089          */
2090         free((char *)name);
2091 }
2092
2093 static void show_edge(struct commit *commit)
2094 {
2095         add_preferred_base(commit->object.sha1);
2096 }
2097
2098 struct in_pack_object {
2099         off_t offset;
2100         struct object *object;
2101 };
2102
2103 struct in_pack {
2104         int alloc;
2105         int nr;
2106         struct in_pack_object *array;
2107 };
2108
2109 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2110 {
2111         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2112         in_pack->array[in_pack->nr].object = object;
2113         in_pack->nr++;
2114 }
2115
2116 /*
2117  * Compare the objects in the offset order, in order to emulate the
2118  * "git rev-list --objects" output that produced the pack originally.
2119  */
2120 static int ofscmp(const void *a_, const void *b_)
2121 {
2122         struct in_pack_object *a = (struct in_pack_object *)a_;
2123         struct in_pack_object *b = (struct in_pack_object *)b_;
2124
2125         if (a->offset < b->offset)
2126                 return -1;
2127         else if (a->offset > b->offset)
2128                 return 1;
2129         else
2130                 return hashcmp(a->object->sha1, b->object->sha1);
2131 }
2132
2133 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2134 {
2135         struct packed_git *p;
2136         struct in_pack in_pack;
2137         uint32_t i;
2138
2139         memset(&in_pack, 0, sizeof(in_pack));
2140
2141         for (p = packed_git; p; p = p->next) {
2142                 const unsigned char *sha1;
2143                 struct object *o;
2144
2145                 if (!p->pack_local || p->pack_keep)
2146                         continue;
2147                 if (open_pack_index(p))
2148                         die("cannot open pack index");
2149
2150                 ALLOC_GROW(in_pack.array,
2151                            in_pack.nr + p->num_objects,
2152                            in_pack.alloc);
2153
2154                 for (i = 0; i < p->num_objects; i++) {
2155                         sha1 = nth_packed_object_sha1(p, i);
2156                         o = lookup_unknown_object(sha1);
2157                         if (!(o->flags & OBJECT_ADDED))
2158                                 mark_in_pack_object(o, p, &in_pack);
2159                         o->flags |= OBJECT_ADDED;
2160                 }
2161         }
2162
2163         if (in_pack.nr) {
2164                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2165                       ofscmp);
2166                 for (i = 0; i < in_pack.nr; i++) {
2167                         struct object *o = in_pack.array[i].object;
2168                         add_object_entry(o->sha1, o->type, "", 0);
2169                 }
2170         }
2171         free(in_pack.array);
2172 }
2173
2174 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2175 {
2176         static struct packed_git *last_found = (void *)1;
2177         struct packed_git *p;
2178
2179         p = (last_found != (void *)1) ? last_found : packed_git;
2180
2181         while (p) {
2182                 if ((!p->pack_local || p->pack_keep) &&
2183                         find_pack_entry_one(sha1, p)) {
2184                         last_found = p;
2185                         return 1;
2186                 }
2187                 if (p == last_found)
2188                         p = packed_git;
2189                 else
2190                         p = p->next;
2191                 if (p == last_found)
2192                         p = p->next;
2193         }
2194         return 0;
2195 }
2196
2197 static void loosen_unused_packed_objects(struct rev_info *revs)
2198 {
2199         struct packed_git *p;
2200         uint32_t i;
2201         const unsigned char *sha1;
2202
2203         for (p = packed_git; p; p = p->next) {
2204                 if (!p->pack_local || p->pack_keep)
2205                         continue;
2206
2207                 if (open_pack_index(p))
2208                         die("cannot open pack index");
2209
2210                 for (i = 0; i < p->num_objects; i++) {
2211                         sha1 = nth_packed_object_sha1(p, i);
2212                         if (!locate_object_entry(sha1) &&
2213                                 !has_sha1_pack_kept_or_nonlocal(sha1))
2214                                 if (force_object_loose(sha1, p->mtime))
2215                                         die("unable to force loose object");
2216                 }
2217         }
2218 }
2219
2220 static void get_object_list(int ac, const char **av)
2221 {
2222         struct rev_info revs;
2223         char line[1000];
2224         int flags = 0;
2225
2226         init_revisions(&revs, NULL);
2227         save_commit_buffer = 0;
2228         setup_revisions(ac, av, &revs, NULL);
2229
2230         while (fgets(line, sizeof(line), stdin) != NULL) {
2231                 int len = strlen(line);
2232                 if (len && line[len - 1] == '\n')
2233                         line[--len] = 0;
2234                 if (!len)
2235                         break;
2236                 if (*line == '-') {
2237                         if (!strcmp(line, "--not")) {
2238                                 flags ^= UNINTERESTING;
2239                                 continue;
2240                         }
2241                         die("not a rev '%s'", line);
2242                 }
2243                 if (handle_revision_arg(line, &revs, flags, 1))
2244                         die("bad revision '%s'", line);
2245         }
2246
2247         if (prepare_revision_walk(&revs))
2248                 die("revision walk setup failed");
2249         mark_edges_uninteresting(revs.commits, &revs, show_edge);
2250         traverse_commit_list(&revs, show_commit, show_object, NULL);
2251
2252         if (keep_unreachable)
2253                 add_objects_in_unpacked_packs(&revs);
2254         if (unpack_unreachable)
2255                 loosen_unused_packed_objects(&revs);
2256 }
2257
2258 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2259 {
2260         int use_internal_rev_list = 0;
2261         int thin = 0;
2262         int all_progress_implied = 0;
2263         uint32_t i;
2264         const char **rp_av;
2265         int rp_ac_alloc = 64;
2266         int rp_ac;
2267
2268         read_replace_refs = 0;
2269
2270         rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2271
2272         rp_av[0] = "pack-objects";
2273         rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2274         rp_ac = 2;
2275
2276         reset_pack_idx_option(&pack_idx_opts);
2277         git_config(git_pack_config, NULL);
2278         if (!pack_compression_seen && core_compression_seen)
2279                 pack_compression_level = core_compression_level;
2280
2281         progress = isatty(2);
2282         for (i = 1; i < argc; i++) {
2283                 const char *arg = argv[i];
2284
2285                 if (*arg != '-')
2286                         break;
2287
2288                 if (!strcmp("--non-empty", arg)) {
2289                         non_empty = 1;
2290                         continue;
2291                 }
2292                 if (!strcmp("--local", arg)) {
2293                         local = 1;
2294                         continue;
2295                 }
2296                 if (!strcmp("--incremental", arg)) {
2297                         incremental = 1;
2298                         continue;
2299                 }
2300                 if (!strcmp("--honor-pack-keep", arg)) {
2301                         ignore_packed_keep = 1;
2302                         continue;
2303                 }
2304                 if (!prefixcmp(arg, "--compression=")) {
2305                         char *end;
2306                         int level = strtoul(arg+14, &end, 0);
2307                         if (!arg[14] || *end)
2308                                 usage(pack_usage);
2309                         if (level == -1)
2310                                 level = Z_DEFAULT_COMPRESSION;
2311                         else if (level < 0 || level > Z_BEST_COMPRESSION)
2312                                 die("bad pack compression level %d", level);
2313                         pack_compression_level = level;
2314                         continue;
2315                 }
2316                 if (!prefixcmp(arg, "--max-pack-size=")) {
2317                         pack_size_limit_cfg = 0;
2318                         if (!git_parse_ulong(arg+16, &pack_size_limit))
2319                                 usage(pack_usage);
2320                         continue;
2321                 }
2322                 if (!prefixcmp(arg, "--window=")) {
2323                         char *end;
2324                         window = strtoul(arg+9, &end, 0);
2325                         if (!arg[9] || *end)
2326                                 usage(pack_usage);
2327                         continue;
2328                 }
2329                 if (!prefixcmp(arg, "--window-memory=")) {
2330                         if (!git_parse_ulong(arg+16, &window_memory_limit))
2331                                 usage(pack_usage);
2332                         continue;
2333                 }
2334                 if (!prefixcmp(arg, "--threads=")) {
2335                         char *end;
2336                         delta_search_threads = strtoul(arg+10, &end, 0);
2337                         if (!arg[10] || *end || delta_search_threads < 0)
2338                                 usage(pack_usage);
2339 #ifdef NO_PTHREADS
2340                         if (delta_search_threads != 1)
2341                                 warning("no threads support, "
2342                                         "ignoring %s", arg);
2343 #endif
2344                         continue;
2345                 }
2346                 if (!prefixcmp(arg, "--depth=")) {
2347                         char *end;
2348                         depth = strtoul(arg+8, &end, 0);
2349                         if (!arg[8] || *end)
2350                                 usage(pack_usage);
2351                         continue;
2352                 }
2353                 if (!strcmp("--progress", arg)) {
2354                         progress = 1;
2355                         continue;
2356                 }
2357                 if (!strcmp("--all-progress", arg)) {
2358                         progress = 2;
2359                         continue;
2360                 }
2361                 if (!strcmp("--all-progress-implied", arg)) {
2362                         all_progress_implied = 1;
2363                         continue;
2364                 }
2365                 if (!strcmp("-q", arg)) {
2366                         progress = 0;
2367                         continue;
2368                 }
2369                 if (!strcmp("--no-reuse-delta", arg)) {
2370                         reuse_delta = 0;
2371                         continue;
2372                 }
2373                 if (!strcmp("--no-reuse-object", arg)) {
2374                         reuse_object = reuse_delta = 0;
2375                         continue;
2376                 }
2377                 if (!strcmp("--delta-base-offset", arg)) {
2378                         allow_ofs_delta = 1;
2379                         continue;
2380                 }
2381                 if (!strcmp("--stdout", arg)) {
2382                         pack_to_stdout = 1;
2383                         continue;
2384                 }
2385                 if (!strcmp("--revs", arg)) {
2386                         use_internal_rev_list = 1;
2387                         continue;
2388                 }
2389                 if (!strcmp("--keep-unreachable", arg)) {
2390                         keep_unreachable = 1;
2391                         continue;
2392                 }
2393                 if (!strcmp("--unpack-unreachable", arg)) {
2394                         unpack_unreachable = 1;
2395                         continue;
2396                 }
2397                 if (!strcmp("--include-tag", arg)) {
2398                         include_tag = 1;
2399                         continue;
2400                 }
2401                 if (!strcmp("--unpacked", arg) ||
2402                     !strcmp("--reflog", arg) ||
2403                     !strcmp("--all", arg)) {
2404                         use_internal_rev_list = 1;
2405                         if (rp_ac >= rp_ac_alloc - 1) {
2406                                 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2407                                 rp_av = xrealloc(rp_av,
2408                                                  rp_ac_alloc * sizeof(*rp_av));
2409                         }
2410                         rp_av[rp_ac++] = arg;
2411                         continue;
2412                 }
2413                 if (!strcmp("--thin", arg)) {
2414                         use_internal_rev_list = 1;
2415                         thin = 1;
2416                         rp_av[1] = "--objects-edge";
2417                         continue;
2418                 }
2419                 if (!prefixcmp(arg, "--index-version=")) {
2420                         char *c;
2421                         pack_idx_opts.version = strtoul(arg + 16, &c, 10);
2422                         if (pack_idx_opts.version > 2)
2423                                 die("bad %s", arg);
2424                         if (*c == ',')
2425                                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2426                         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2427                                 die("bad %s", arg);
2428                         continue;
2429                 }
2430                 if (!strcmp(arg, "--keep-true-parents")) {
2431                         grafts_replace_parents = 0;
2432                         continue;
2433                 }
2434                 usage(pack_usage);
2435         }
2436
2437         /* Traditionally "pack-objects [options] base extra" failed;
2438          * we would however want to take refs parameter that would
2439          * have been given to upstream rev-list ourselves, which means
2440          * we somehow want to say what the base name is.  So the
2441          * syntax would be:
2442          *
2443          * pack-objects [options] base <refs...>
2444          *
2445          * in other words, we would treat the first non-option as the
2446          * base_name and send everything else to the internal revision
2447          * walker.
2448          */
2449
2450         if (!pack_to_stdout)
2451                 base_name = argv[i++];
2452
2453         if (pack_to_stdout != !base_name)
2454                 usage(pack_usage);
2455
2456         if (!pack_to_stdout && !pack_size_limit)
2457                 pack_size_limit = pack_size_limit_cfg;
2458         if (pack_to_stdout && pack_size_limit)
2459                 die("--max-pack-size cannot be used to build a pack for transfer.");
2460         if (pack_size_limit && pack_size_limit < 1024*1024) {
2461                 warning("minimum pack size limit is 1 MiB");
2462                 pack_size_limit = 1024*1024;
2463         }
2464
2465         if (!pack_to_stdout && thin)
2466                 die("--thin cannot be used to build an indexable pack.");
2467
2468         if (keep_unreachable && unpack_unreachable)
2469                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2470
2471         if (progress && all_progress_implied)
2472                 progress = 2;
2473
2474         prepare_packed_git();
2475
2476         if (progress)
2477                 progress_state = start_progress("Counting objects", 0);
2478         if (!use_internal_rev_list)
2479                 read_object_list_from_stdin();
2480         else {
2481                 rp_av[rp_ac] = NULL;
2482                 get_object_list(rp_ac, rp_av);
2483         }
2484         cleanup_preferred_base();
2485         if (include_tag && nr_result)
2486                 for_each_ref(add_ref_tag, NULL);
2487         stop_progress(&progress_state);
2488
2489         if (non_empty && !nr_result)
2490                 return 0;
2491         if (nr_result)
2492                 prepare_pack(window, depth);
2493         write_pack_file();
2494         if (progress)
2495                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2496                         " reused %"PRIu32" (delta %"PRIu32")\n",
2497                         written, written_delta, reused, reused_delta);
2498         return 0;
2499 }