Imported Upstream version 1.7.10.4
[platform/upstream/git.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 --stdout [options...] [< ref-list | < object-list]",
23         "git pack-objects [options...] base-name [< ref-list | < object-list]",
24         NULL
25 };
26
27 struct object_entry {
28         struct pack_idx_entry idx;
29         unsigned long size;     /* uncompressed size */
30         struct packed_git *in_pack;     /* already in pack */
31         off_t in_pack_offset;
32         struct object_entry *delta;     /* delta base object */
33         struct object_entry *delta_child; /* deltified objects who bases me */
34         struct object_entry *delta_sibling; /* other deltified objects who
35                                              * uses the same base as me
36                                              */
37         void *delta_data;       /* cached delta (uncompressed) */
38         unsigned long delta_size;       /* delta data size (uncompressed) */
39         unsigned long z_delta_size;     /* delta data size (compressed) */
40         unsigned int hash;      /* name hint hash */
41         enum object_type type;
42         enum object_type in_pack_type;  /* could be delta */
43         unsigned char in_pack_header_size;
44         unsigned char preferred_base; /* we do not pack this, but is available
45                                        * to be used as the base object to delta
46                                        * objects against.
47                                        */
48         unsigned char no_try_delta;
49         unsigned char tagged; /* near the very tip of refs */
50         unsigned char filled; /* assigned write-order */
51 };
52
53 /*
54  * Objects we are going to pack are collected in objects array (dynamically
55  * expanded).  nr_objects & nr_alloc controls this array.  They are stored
56  * in the order we see -- typically rev-list --objects order that gives us
57  * nice "minimum seek" order.
58  */
59 static struct object_entry *objects;
60 static struct pack_idx_entry **written_list;
61 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
62
63 static int non_empty;
64 static int reuse_delta = 1, reuse_object = 1;
65 static int keep_unreachable, unpack_unreachable, include_tag;
66 static unsigned long unpack_unreachable_expiration;
67 static int local;
68 static int incremental;
69 static int ignore_packed_keep;
70 static int allow_ofs_delta;
71 static struct pack_idx_option pack_idx_opts;
72 static const char *base_name;
73 static int progress = 1;
74 static int window = 10;
75 static unsigned long pack_size_limit;
76 static int depth = 50;
77 static int delta_search_threads;
78 static int pack_to_stdout;
79 static int num_preferred_base;
80 static struct progress *progress_state;
81 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
82 static int pack_compression_seen;
83
84 static unsigned long delta_cache_size = 0;
85 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
86 static unsigned long cache_max_small_delta_size = 1000;
87
88 static unsigned long window_memory_limit = 0;
89
90 /*
91  * The object names in objects array are hashed with this hashtable,
92  * to help looking up the entry by object name.
93  * This hashtable is built after all the objects are seen.
94  */
95 static int *object_ix;
96 static int object_ix_hashsz;
97 static struct object_entry *locate_object_entry(const unsigned char *sha1);
98
99 /*
100  * stats
101  */
102 static uint32_t written, written_delta;
103 static uint32_t reused, reused_delta;
104
105
106 static void *get_delta(struct object_entry *entry)
107 {
108         unsigned long size, base_size, delta_size;
109         void *buf, *base_buf, *delta_buf;
110         enum object_type type;
111
112         buf = read_sha1_file(entry->idx.sha1, &type, &size);
113         if (!buf)
114                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
115         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
116         if (!base_buf)
117                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
118         delta_buf = diff_delta(base_buf, base_size,
119                                buf, size, &delta_size, 0);
120         if (!delta_buf || delta_size != entry->delta_size)
121                 die("delta size changed");
122         free(buf);
123         free(base_buf);
124         return delta_buf;
125 }
126
127 static unsigned long do_compress(void **pptr, unsigned long size)
128 {
129         git_zstream stream;
130         void *in, *out;
131         unsigned long maxsize;
132
133         memset(&stream, 0, sizeof(stream));
134         git_deflate_init(&stream, pack_compression_level);
135         maxsize = git_deflate_bound(&stream, size);
136
137         in = *pptr;
138         out = xmalloc(maxsize);
139         *pptr = out;
140
141         stream.next_in = in;
142         stream.avail_in = size;
143         stream.next_out = out;
144         stream.avail_out = maxsize;
145         while (git_deflate(&stream, Z_FINISH) == Z_OK)
146                 ; /* nothing */
147         git_deflate_end(&stream);
148
149         free(in);
150         return stream.total_out;
151 }
152
153 /*
154  * we are going to reuse the existing object data as is.  make
155  * sure it is not corrupt.
156  */
157 static int check_pack_inflate(struct packed_git *p,
158                 struct pack_window **w_curs,
159                 off_t offset,
160                 off_t len,
161                 unsigned long expect)
162 {
163         git_zstream stream;
164         unsigned char fakebuf[4096], *in;
165         int st;
166
167         memset(&stream, 0, sizeof(stream));
168         git_inflate_init(&stream);
169         do {
170                 in = use_pack(p, w_curs, offset, &stream.avail_in);
171                 stream.next_in = in;
172                 stream.next_out = fakebuf;
173                 stream.avail_out = sizeof(fakebuf);
174                 st = git_inflate(&stream, Z_FINISH);
175                 offset += stream.next_in - in;
176         } while (st == Z_OK || st == Z_BUF_ERROR);
177         git_inflate_end(&stream);
178         return (st == Z_STREAM_END &&
179                 stream.total_out == expect &&
180                 stream.total_in == len) ? 0 : -1;
181 }
182
183 static void copy_pack_data(struct sha1file *f,
184                 struct packed_git *p,
185                 struct pack_window **w_curs,
186                 off_t offset,
187                 off_t len)
188 {
189         unsigned char *in;
190         unsigned long avail;
191
192         while (len) {
193                 in = use_pack(p, w_curs, offset, &avail);
194                 if (avail > len)
195                         avail = (unsigned long)len;
196                 sha1write(f, in, avail);
197                 offset += avail;
198                 len -= avail;
199         }
200 }
201
202 /* Return 0 if we will bust the pack-size limit */
203 static unsigned long write_object(struct sha1file *f,
204                                   struct object_entry *entry,
205                                   off_t write_offset)
206 {
207         unsigned long size, limit, datalen;
208         void *buf;
209         unsigned char header[10], dheader[10];
210         unsigned hdrlen;
211         enum object_type type;
212         int usable_delta, to_reuse;
213
214         if (!pack_to_stdout)
215                 crc32_begin(f);
216
217         type = entry->type;
218
219         /* apply size limit if limited packsize and not first object */
220         if (!pack_size_limit || !nr_written)
221                 limit = 0;
222         else if (pack_size_limit <= write_offset)
223                 /*
224                  * the earlier object did not fit the limit; avoid
225                  * mistaking this with unlimited (i.e. limit = 0).
226                  */
227                 limit = 1;
228         else
229                 limit = pack_size_limit - write_offset;
230
231         if (!entry->delta)
232                 usable_delta = 0;       /* no delta */
233         else if (!pack_size_limit)
234                usable_delta = 1;        /* unlimited packfile */
235         else if (entry->delta->idx.offset == (off_t)-1)
236                 usable_delta = 0;       /* base was written to another pack */
237         else if (entry->delta->idx.offset)
238                 usable_delta = 1;       /* base already exists in this pack */
239         else
240                 usable_delta = 0;       /* base could end up in another pack */
241
242         if (!reuse_object)
243                 to_reuse = 0;   /* explicit */
244         else if (!entry->in_pack)
245                 to_reuse = 0;   /* can't reuse what we don't have */
246         else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
247                                 /* check_object() decided it for us ... */
248                 to_reuse = usable_delta;
249                                 /* ... but pack split may override that */
250         else if (type != entry->in_pack_type)
251                 to_reuse = 0;   /* pack has delta which is unusable */
252         else if (entry->delta)
253                 to_reuse = 0;   /* we want to pack afresh */
254         else
255                 to_reuse = 1;   /* we have it in-pack undeltified,
256                                  * and we do not need to deltify it.
257                                  */
258
259         if (!to_reuse) {
260                 no_reuse:
261                 if (!usable_delta) {
262                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
263                         if (!buf)
264                                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
265                         /*
266                          * make sure no cached delta data remains from a
267                          * previous attempt before a pack split occurred.
268                          */
269                         free(entry->delta_data);
270                         entry->delta_data = NULL;
271                         entry->z_delta_size = 0;
272                 } else if (entry->delta_data) {
273                         size = entry->delta_size;
274                         buf = entry->delta_data;
275                         entry->delta_data = NULL;
276                         type = (allow_ofs_delta && entry->delta->idx.offset) ?
277                                 OBJ_OFS_DELTA : OBJ_REF_DELTA;
278                 } else {
279                         buf = get_delta(entry);
280                         size = entry->delta_size;
281                         type = (allow_ofs_delta && entry->delta->idx.offset) ?
282                                 OBJ_OFS_DELTA : OBJ_REF_DELTA;
283                 }
284
285                 if (entry->z_delta_size)
286                         datalen = entry->z_delta_size;
287                 else
288                         datalen = do_compress(&buf, size);
289
290                 /*
291                  * The object header is a byte of 'type' followed by zero or
292                  * more bytes of length.
293                  */
294                 hdrlen = encode_in_pack_object_header(type, size, header);
295
296                 if (type == OBJ_OFS_DELTA) {
297                         /*
298                          * Deltas with relative base contain an additional
299                          * encoding of the relative offset for the delta
300                          * base from this object's position in the pack.
301                          */
302                         off_t ofs = entry->idx.offset - entry->delta->idx.offset;
303                         unsigned pos = sizeof(dheader) - 1;
304                         dheader[pos] = ofs & 127;
305                         while (ofs >>= 7)
306                                 dheader[--pos] = 128 | (--ofs & 127);
307                         if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
308                                 free(buf);
309                                 return 0;
310                         }
311                         sha1write(f, header, hdrlen);
312                         sha1write(f, dheader + pos, sizeof(dheader) - pos);
313                         hdrlen += sizeof(dheader) - pos;
314                 } else if (type == OBJ_REF_DELTA) {
315                         /*
316                          * Deltas with a base reference contain
317                          * an additional 20 bytes for the base sha1.
318                          */
319                         if (limit && hdrlen + 20 + datalen + 20 >= limit) {
320                                 free(buf);
321                                 return 0;
322                         }
323                         sha1write(f, header, hdrlen);
324                         sha1write(f, entry->delta->idx.sha1, 20);
325                         hdrlen += 20;
326                 } else {
327                         if (limit && hdrlen + datalen + 20 >= limit) {
328                                 free(buf);
329                                 return 0;
330                         }
331                         sha1write(f, header, hdrlen);
332                 }
333                 sha1write(f, buf, datalen);
334                 free(buf);
335         }
336         else {
337                 struct packed_git *p = entry->in_pack;
338                 struct pack_window *w_curs = NULL;
339                 struct revindex_entry *revidx;
340                 off_t offset;
341
342                 if (entry->delta)
343                         type = (allow_ofs_delta && entry->delta->idx.offset) ?
344                                 OBJ_OFS_DELTA : OBJ_REF_DELTA;
345                 hdrlen = encode_in_pack_object_header(type, entry->size, header);
346
347                 offset = entry->in_pack_offset;
348                 revidx = find_pack_revindex(p, offset);
349                 datalen = revidx[1].offset - offset;
350                 if (!pack_to_stdout && p->index_version > 1 &&
351                     check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
352                         error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
353                         unuse_pack(&w_curs);
354                         goto no_reuse;
355                 }
356
357                 offset += entry->in_pack_header_size;
358                 datalen -= entry->in_pack_header_size;
359                 if (!pack_to_stdout && p->index_version == 1 &&
360                     check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
361                         error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
362                         unuse_pack(&w_curs);
363                         goto no_reuse;
364                 }
365
366                 if (type == OBJ_OFS_DELTA) {
367                         off_t ofs = entry->idx.offset - entry->delta->idx.offset;
368                         unsigned pos = sizeof(dheader) - 1;
369                         dheader[pos] = ofs & 127;
370                         while (ofs >>= 7)
371                                 dheader[--pos] = 128 | (--ofs & 127);
372                         if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
373                                 unuse_pack(&w_curs);
374                                 return 0;
375                         }
376                         sha1write(f, header, hdrlen);
377                         sha1write(f, dheader + pos, sizeof(dheader) - pos);
378                         hdrlen += sizeof(dheader) - pos;
379                         reused_delta++;
380                 } else if (type == OBJ_REF_DELTA) {
381                         if (limit && hdrlen + 20 + datalen + 20 >= limit) {
382                                 unuse_pack(&w_curs);
383                                 return 0;
384                         }
385                         sha1write(f, header, hdrlen);
386                         sha1write(f, entry->delta->idx.sha1, 20);
387                         hdrlen += 20;
388                         reused_delta++;
389                 } else {
390                         if (limit && hdrlen + datalen + 20 >= limit) {
391                                 unuse_pack(&w_curs);
392                                 return 0;
393                         }
394                         sha1write(f, header, hdrlen);
395                 }
396                 copy_pack_data(f, p, &w_curs, offset, datalen);
397                 unuse_pack(&w_curs);
398                 reused++;
399         }
400         if (usable_delta)
401                 written_delta++;
402         written++;
403         if (!pack_to_stdout)
404                 entry->idx.crc32 = crc32_end(f);
405         return hdrlen + datalen;
406 }
407
408 enum write_one_status {
409         WRITE_ONE_SKIP = -1, /* already written */
410         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
411         WRITE_ONE_WRITTEN = 1, /* normal */
412         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
413 };
414
415 static enum write_one_status write_one(struct sha1file *f,
416                                        struct object_entry *e,
417                                        off_t *offset)
418 {
419         unsigned long size;
420         int recursing;
421
422         /*
423          * we set offset to 1 (which is an impossible value) to mark
424          * the fact that this object is involved in "write its base
425          * first before writing a deltified object" recursion.
426          */
427         recursing = (e->idx.offset == 1);
428         if (recursing) {
429                 warning("recursive delta detected for object %s",
430                         sha1_to_hex(e->idx.sha1));
431                 return WRITE_ONE_RECURSIVE;
432         } else if (e->idx.offset || e->preferred_base) {
433                 /* offset is non zero if object is written already. */
434                 return WRITE_ONE_SKIP;
435         }
436
437         /* if we are deltified, write out base object first. */
438         if (e->delta) {
439                 e->idx.offset = 1; /* now recurse */
440                 switch (write_one(f, e->delta, offset)) {
441                 case WRITE_ONE_RECURSIVE:
442                         /* we cannot depend on this one */
443                         e->delta = NULL;
444                         break;
445                 default:
446                         break;
447                 case WRITE_ONE_BREAK:
448                         e->idx.offset = recursing;
449                         return WRITE_ONE_BREAK;
450                 }
451         }
452
453         e->idx.offset = *offset;
454         size = write_object(f, e, *offset);
455         if (!size) {
456                 e->idx.offset = recursing;
457                 return WRITE_ONE_BREAK;
458         }
459         written_list[nr_written++] = &e->idx;
460
461         /* make sure off_t is sufficiently large not to wrap */
462         if (signed_add_overflows(*offset, size))
463                 die("pack too large for current definition of off_t");
464         *offset += size;
465         return WRITE_ONE_WRITTEN;
466 }
467
468 static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
469                        void *cb_data)
470 {
471         unsigned char peeled[20];
472         struct object_entry *entry = locate_object_entry(sha1);
473
474         if (entry)
475                 entry->tagged = 1;
476         if (!peel_ref(path, peeled)) {
477                 entry = locate_object_entry(peeled);
478                 if (entry)
479                         entry->tagged = 1;
480         }
481         return 0;
482 }
483
484 static inline void add_to_write_order(struct object_entry **wo,
485                                unsigned int *endp,
486                                struct object_entry *e)
487 {
488         if (e->filled)
489                 return;
490         wo[(*endp)++] = e;
491         e->filled = 1;
492 }
493
494 static void add_descendants_to_write_order(struct object_entry **wo,
495                                            unsigned int *endp,
496                                            struct object_entry *e)
497 {
498         int add_to_order = 1;
499         while (e) {
500                 if (add_to_order) {
501                         struct object_entry *s;
502                         /* add this node... */
503                         add_to_write_order(wo, endp, e);
504                         /* all its siblings... */
505                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
506                                 add_to_write_order(wo, endp, s);
507                         }
508                 }
509                 /* drop down a level to add left subtree nodes if possible */
510                 if (e->delta_child) {
511                         add_to_order = 1;
512                         e = e->delta_child;
513                 } else {
514                         add_to_order = 0;
515                         /* our sibling might have some children, it is next */
516                         if (e->delta_sibling) {
517                                 e = e->delta_sibling;
518                                 continue;
519                         }
520                         /* go back to our parent node */
521                         e = e->delta;
522                         while (e && !e->delta_sibling) {
523                                 /* we're on the right side of a subtree, keep
524                                  * going up until we can go right again */
525                                 e = e->delta;
526                         }
527                         if (!e) {
528                                 /* done- we hit our original root node */
529                                 return;
530                         }
531                         /* pass it off to sibling at this level */
532                         e = e->delta_sibling;
533                 }
534         };
535 }
536
537 static void add_family_to_write_order(struct object_entry **wo,
538                                       unsigned int *endp,
539                                       struct object_entry *e)
540 {
541         struct object_entry *root;
542
543         for (root = e; root->delta; root = root->delta)
544                 ; /* nothing */
545         add_descendants_to_write_order(wo, endp, root);
546 }
547
548 static struct object_entry **compute_write_order(void)
549 {
550         unsigned int i, wo_end, last_untagged;
551
552         struct object_entry **wo = xmalloc(nr_objects * sizeof(*wo));
553
554         for (i = 0; i < nr_objects; i++) {
555                 objects[i].tagged = 0;
556                 objects[i].filled = 0;
557                 objects[i].delta_child = NULL;
558                 objects[i].delta_sibling = NULL;
559         }
560
561         /*
562          * Fully connect delta_child/delta_sibling network.
563          * Make sure delta_sibling is sorted in the original
564          * recency order.
565          */
566         for (i = nr_objects; i > 0;) {
567                 struct object_entry *e = &objects[--i];
568                 if (!e->delta)
569                         continue;
570                 /* Mark me as the first child */
571                 e->delta_sibling = e->delta->delta_child;
572                 e->delta->delta_child = e;
573         }
574
575         /*
576          * Mark objects that are at the tip of tags.
577          */
578         for_each_tag_ref(mark_tagged, NULL);
579
580         /*
581          * Give the objects in the original recency order until
582          * we see a tagged tip.
583          */
584         for (i = wo_end = 0; i < nr_objects; i++) {
585                 if (objects[i].tagged)
586                         break;
587                 add_to_write_order(wo, &wo_end, &objects[i]);
588         }
589         last_untagged = i;
590
591         /*
592          * Then fill all the tagged tips.
593          */
594         for (; i < nr_objects; i++) {
595                 if (objects[i].tagged)
596                         add_to_write_order(wo, &wo_end, &objects[i]);
597         }
598
599         /*
600          * And then all remaining commits and tags.
601          */
602         for (i = last_untagged; i < nr_objects; i++) {
603                 if (objects[i].type != OBJ_COMMIT &&
604                     objects[i].type != OBJ_TAG)
605                         continue;
606                 add_to_write_order(wo, &wo_end, &objects[i]);
607         }
608
609         /*
610          * And then all the trees.
611          */
612         for (i = last_untagged; i < nr_objects; i++) {
613                 if (objects[i].type != OBJ_TREE)
614                         continue;
615                 add_to_write_order(wo, &wo_end, &objects[i]);
616         }
617
618         /*
619          * Finally all the rest in really tight order
620          */
621         for (i = last_untagged; i < nr_objects; i++) {
622                 if (!objects[i].filled)
623                         add_family_to_write_order(wo, &wo_end, &objects[i]);
624         }
625
626         if (wo_end != nr_objects)
627                 die("ordered %u objects, expected %"PRIu32, wo_end, nr_objects);
628
629         return wo;
630 }
631
632 static void write_pack_file(void)
633 {
634         uint32_t i = 0, j;
635         struct sha1file *f;
636         off_t offset;
637         uint32_t nr_remaining = nr_result;
638         time_t last_mtime = 0;
639         struct object_entry **write_order;
640
641         if (progress > pack_to_stdout)
642                 progress_state = start_progress("Writing objects", nr_result);
643         written_list = xmalloc(nr_objects * sizeof(*written_list));
644         write_order = compute_write_order();
645
646         do {
647                 unsigned char sha1[20];
648                 char *pack_tmp_name = NULL;
649
650                 if (pack_to_stdout)
651                         f = sha1fd_throughput(1, "<stdout>", progress_state);
652                 else
653                         f = create_tmp_packfile(&pack_tmp_name);
654
655                 offset = write_pack_header(f, nr_remaining);
656                 if (!offset)
657                         die_errno("unable to write pack header");
658                 nr_written = 0;
659                 for (; i < nr_objects; i++) {
660                         struct object_entry *e = write_order[i];
661                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
662                                 break;
663                         display_progress(progress_state, written);
664                 }
665
666                 /*
667                  * Did we write the wrong # entries in the header?
668                  * If so, rewrite it like in fast-import
669                  */
670                 if (pack_to_stdout) {
671                         sha1close(f, sha1, CSUM_CLOSE);
672                 } else if (nr_written == nr_remaining) {
673                         sha1close(f, sha1, CSUM_FSYNC);
674                 } else {
675                         int fd = sha1close(f, sha1, 0);
676                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
677                                                  nr_written, sha1, offset);
678                         close(fd);
679                 }
680
681                 if (!pack_to_stdout) {
682                         struct stat st;
683                         char tmpname[PATH_MAX];
684
685                         /*
686                          * Packs are runtime accessed in their mtime
687                          * order since newer packs are more likely to contain
688                          * younger objects.  So if we are creating multiple
689                          * packs then we should modify the mtime of later ones
690                          * to preserve this property.
691                          */
692                         if (stat(pack_tmp_name, &st) < 0) {
693                                 warning("failed to stat %s: %s",
694                                         pack_tmp_name, strerror(errno));
695                         } else if (!last_mtime) {
696                                 last_mtime = st.st_mtime;
697                         } else {
698                                 struct utimbuf utb;
699                                 utb.actime = st.st_atime;
700                                 utb.modtime = --last_mtime;
701                                 if (utime(pack_tmp_name, &utb) < 0)
702                                         warning("failed utime() on %s: %s",
703                                                 tmpname, strerror(errno));
704                         }
705
706                         /* Enough space for "-<sha-1>.pack"? */
707                         if (sizeof(tmpname) <= strlen(base_name) + 50)
708                                 die("pack base name '%s' too long", base_name);
709                         snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
710                         finish_tmp_packfile(tmpname, pack_tmp_name,
711                                             written_list, nr_written,
712                                             &pack_idx_opts, sha1);
713                         free(pack_tmp_name);
714                         puts(sha1_to_hex(sha1));
715                 }
716
717                 /* mark written objects as written to previous pack */
718                 for (j = 0; j < nr_written; j++) {
719                         written_list[j]->offset = (off_t)-1;
720                 }
721                 nr_remaining -= nr_written;
722         } while (nr_remaining && i < nr_objects);
723
724         free(written_list);
725         free(write_order);
726         stop_progress(&progress_state);
727         if (written != nr_result)
728                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
729                         written, nr_result);
730 }
731
732 static int locate_object_entry_hash(const unsigned char *sha1)
733 {
734         int i;
735         unsigned int ui;
736         memcpy(&ui, sha1, sizeof(unsigned int));
737         i = ui % object_ix_hashsz;
738         while (0 < object_ix[i]) {
739                 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
740                         return i;
741                 if (++i == object_ix_hashsz)
742                         i = 0;
743         }
744         return -1 - i;
745 }
746
747 static struct object_entry *locate_object_entry(const unsigned char *sha1)
748 {
749         int i;
750
751         if (!object_ix_hashsz)
752                 return NULL;
753
754         i = locate_object_entry_hash(sha1);
755         if (0 <= i)
756                 return &objects[object_ix[i]-1];
757         return NULL;
758 }
759
760 static void rehash_objects(void)
761 {
762         uint32_t i;
763         struct object_entry *oe;
764
765         object_ix_hashsz = nr_objects * 3;
766         if (object_ix_hashsz < 1024)
767                 object_ix_hashsz = 1024;
768         object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
769         memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
770         for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
771                 int ix = locate_object_entry_hash(oe->idx.sha1);
772                 if (0 <= ix)
773                         continue;
774                 ix = -1 - ix;
775                 object_ix[ix] = i + 1;
776         }
777 }
778
779 static unsigned name_hash(const char *name)
780 {
781         unsigned c, hash = 0;
782
783         if (!name)
784                 return 0;
785
786         /*
787          * This effectively just creates a sortable number from the
788          * last sixteen non-whitespace characters. Last characters
789          * count "most", so things that end in ".c" sort together.
790          */
791         while ((c = *name++) != 0) {
792                 if (isspace(c))
793                         continue;
794                 hash = (hash >> 2) + (c << 24);
795         }
796         return hash;
797 }
798
799 static void setup_delta_attr_check(struct git_attr_check *check)
800 {
801         static struct git_attr *attr_delta;
802
803         if (!attr_delta)
804                 attr_delta = git_attr("delta");
805
806         check[0].attr = attr_delta;
807 }
808
809 static int no_try_delta(const char *path)
810 {
811         struct git_attr_check check[1];
812
813         setup_delta_attr_check(check);
814         if (git_check_attr(path, ARRAY_SIZE(check), check))
815                 return 0;
816         if (ATTR_FALSE(check->value))
817                 return 1;
818         return 0;
819 }
820
821 static int add_object_entry(const unsigned char *sha1, enum object_type type,
822                             const char *name, int exclude)
823 {
824         struct object_entry *entry;
825         struct packed_git *p, *found_pack = NULL;
826         off_t found_offset = 0;
827         int ix;
828         unsigned hash = name_hash(name);
829
830         ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
831         if (ix >= 0) {
832                 if (exclude) {
833                         entry = objects + object_ix[ix] - 1;
834                         if (!entry->preferred_base)
835                                 nr_result--;
836                         entry->preferred_base = 1;
837                 }
838                 return 0;
839         }
840
841         if (!exclude && local && has_loose_object_nonlocal(sha1))
842                 return 0;
843
844         for (p = packed_git; p; p = p->next) {
845                 off_t offset = find_pack_entry_one(sha1, p);
846                 if (offset) {
847                         if (!found_pack) {
848                                 if (!is_pack_valid(p)) {
849                                         warning("packfile %s cannot be accessed", p->pack_name);
850                                         continue;
851                                 }
852                                 found_offset = offset;
853                                 found_pack = p;
854                         }
855                         if (exclude)
856                                 break;
857                         if (incremental)
858                                 return 0;
859                         if (local && !p->pack_local)
860                                 return 0;
861                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
862                                 return 0;
863                 }
864         }
865
866         if (nr_objects >= nr_alloc) {
867                 nr_alloc = (nr_alloc  + 1024) * 3 / 2;
868                 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
869         }
870
871         entry = objects + nr_objects++;
872         memset(entry, 0, sizeof(*entry));
873         hashcpy(entry->idx.sha1, sha1);
874         entry->hash = hash;
875         if (type)
876                 entry->type = type;
877         if (exclude)
878                 entry->preferred_base = 1;
879         else
880                 nr_result++;
881         if (found_pack) {
882                 entry->in_pack = found_pack;
883                 entry->in_pack_offset = found_offset;
884         }
885
886         if (object_ix_hashsz * 3 <= nr_objects * 4)
887                 rehash_objects();
888         else
889                 object_ix[-1 - ix] = nr_objects;
890
891         display_progress(progress_state, nr_objects);
892
893         if (name && no_try_delta(name))
894                 entry->no_try_delta = 1;
895
896         return 1;
897 }
898
899 struct pbase_tree_cache {
900         unsigned char sha1[20];
901         int ref;
902         int temporary;
903         void *tree_data;
904         unsigned long tree_size;
905 };
906
907 static struct pbase_tree_cache *(pbase_tree_cache[256]);
908 static int pbase_tree_cache_ix(const unsigned char *sha1)
909 {
910         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
911 }
912 static int pbase_tree_cache_ix_incr(int ix)
913 {
914         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
915 }
916
917 static struct pbase_tree {
918         struct pbase_tree *next;
919         /* This is a phony "cache" entry; we are not
920          * going to evict it nor find it through _get()
921          * mechanism -- this is for the toplevel node that
922          * would almost always change with any commit.
923          */
924         struct pbase_tree_cache pcache;
925 } *pbase_tree;
926
927 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
928 {
929         struct pbase_tree_cache *ent, *nent;
930         void *data;
931         unsigned long size;
932         enum object_type type;
933         int neigh;
934         int my_ix = pbase_tree_cache_ix(sha1);
935         int available_ix = -1;
936
937         /* pbase-tree-cache acts as a limited hashtable.
938          * your object will be found at your index or within a few
939          * slots after that slot if it is cached.
940          */
941         for (neigh = 0; neigh < 8; neigh++) {
942                 ent = pbase_tree_cache[my_ix];
943                 if (ent && !hashcmp(ent->sha1, sha1)) {
944                         ent->ref++;
945                         return ent;
946                 }
947                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
948                          ((0 <= available_ix) &&
949                           (!ent && pbase_tree_cache[available_ix])))
950                         available_ix = my_ix;
951                 if (!ent)
952                         break;
953                 my_ix = pbase_tree_cache_ix_incr(my_ix);
954         }
955
956         /* Did not find one.  Either we got a bogus request or
957          * we need to read and perhaps cache.
958          */
959         data = read_sha1_file(sha1, &type, &size);
960         if (!data)
961                 return NULL;
962         if (type != OBJ_TREE) {
963                 free(data);
964                 return NULL;
965         }
966
967         /* We need to either cache or return a throwaway copy */
968
969         if (available_ix < 0)
970                 ent = NULL;
971         else {
972                 ent = pbase_tree_cache[available_ix];
973                 my_ix = available_ix;
974         }
975
976         if (!ent) {
977                 nent = xmalloc(sizeof(*nent));
978                 nent->temporary = (available_ix < 0);
979         }
980         else {
981                 /* evict and reuse */
982                 free(ent->tree_data);
983                 nent = ent;
984         }
985         hashcpy(nent->sha1, sha1);
986         nent->tree_data = data;
987         nent->tree_size = size;
988         nent->ref = 1;
989         if (!nent->temporary)
990                 pbase_tree_cache[my_ix] = nent;
991         return nent;
992 }
993
994 static void pbase_tree_put(struct pbase_tree_cache *cache)
995 {
996         if (!cache->temporary) {
997                 cache->ref--;
998                 return;
999         }
1000         free(cache->tree_data);
1001         free(cache);
1002 }
1003
1004 static int name_cmp_len(const char *name)
1005 {
1006         int i;
1007         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1008                 ;
1009         return i;
1010 }
1011
1012 static void add_pbase_object(struct tree_desc *tree,
1013                              const char *name,
1014                              int cmplen,
1015                              const char *fullname)
1016 {
1017         struct name_entry entry;
1018         int cmp;
1019
1020         while (tree_entry(tree,&entry)) {
1021                 if (S_ISGITLINK(entry.mode))
1022                         continue;
1023                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1024                       memcmp(name, entry.path, cmplen);
1025                 if (cmp > 0)
1026                         continue;
1027                 if (cmp < 0)
1028                         return;
1029                 if (name[cmplen] != '/') {
1030                         add_object_entry(entry.sha1,
1031                                          object_type(entry.mode),
1032                                          fullname, 1);
1033                         return;
1034                 }
1035                 if (S_ISDIR(entry.mode)) {
1036                         struct tree_desc sub;
1037                         struct pbase_tree_cache *tree;
1038                         const char *down = name+cmplen+1;
1039                         int downlen = name_cmp_len(down);
1040
1041                         tree = pbase_tree_get(entry.sha1);
1042                         if (!tree)
1043                                 return;
1044                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1045
1046                         add_pbase_object(&sub, down, downlen, fullname);
1047                         pbase_tree_put(tree);
1048                 }
1049         }
1050 }
1051
1052 static unsigned *done_pbase_paths;
1053 static int done_pbase_paths_num;
1054 static int done_pbase_paths_alloc;
1055 static int done_pbase_path_pos(unsigned hash)
1056 {
1057         int lo = 0;
1058         int hi = done_pbase_paths_num;
1059         while (lo < hi) {
1060                 int mi = (hi + lo) / 2;
1061                 if (done_pbase_paths[mi] == hash)
1062                         return mi;
1063                 if (done_pbase_paths[mi] < hash)
1064                         hi = mi;
1065                 else
1066                         lo = mi + 1;
1067         }
1068         return -lo-1;
1069 }
1070
1071 static int check_pbase_path(unsigned hash)
1072 {
1073         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1074         if (0 <= pos)
1075                 return 1;
1076         pos = -pos - 1;
1077         if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1078                 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1079                 done_pbase_paths = xrealloc(done_pbase_paths,
1080                                             done_pbase_paths_alloc *
1081                                             sizeof(unsigned));
1082         }
1083         done_pbase_paths_num++;
1084         if (pos < done_pbase_paths_num)
1085                 memmove(done_pbase_paths + pos + 1,
1086                         done_pbase_paths + pos,
1087                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1088         done_pbase_paths[pos] = hash;
1089         return 0;
1090 }
1091
1092 static void add_preferred_base_object(const char *name)
1093 {
1094         struct pbase_tree *it;
1095         int cmplen;
1096         unsigned hash = name_hash(name);
1097
1098         if (!num_preferred_base || check_pbase_path(hash))
1099                 return;
1100
1101         cmplen = name_cmp_len(name);
1102         for (it = pbase_tree; it; it = it->next) {
1103                 if (cmplen == 0) {
1104                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1105                 }
1106                 else {
1107                         struct tree_desc tree;
1108                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1109                         add_pbase_object(&tree, name, cmplen, name);
1110                 }
1111         }
1112 }
1113
1114 static void add_preferred_base(unsigned char *sha1)
1115 {
1116         struct pbase_tree *it;
1117         void *data;
1118         unsigned long size;
1119         unsigned char tree_sha1[20];
1120
1121         if (window <= num_preferred_base++)
1122                 return;
1123
1124         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1125         if (!data)
1126                 return;
1127
1128         for (it = pbase_tree; it; it = it->next) {
1129                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1130                         free(data);
1131                         return;
1132                 }
1133         }
1134
1135         it = xcalloc(1, sizeof(*it));
1136         it->next = pbase_tree;
1137         pbase_tree = it;
1138
1139         hashcpy(it->pcache.sha1, tree_sha1);
1140         it->pcache.tree_data = data;
1141         it->pcache.tree_size = size;
1142 }
1143
1144 static void cleanup_preferred_base(void)
1145 {
1146         struct pbase_tree *it;
1147         unsigned i;
1148
1149         it = pbase_tree;
1150         pbase_tree = NULL;
1151         while (it) {
1152                 struct pbase_tree *this = it;
1153                 it = this->next;
1154                 free(this->pcache.tree_data);
1155                 free(this);
1156         }
1157
1158         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1159                 if (!pbase_tree_cache[i])
1160                         continue;
1161                 free(pbase_tree_cache[i]->tree_data);
1162                 free(pbase_tree_cache[i]);
1163                 pbase_tree_cache[i] = NULL;
1164         }
1165
1166         free(done_pbase_paths);
1167         done_pbase_paths = NULL;
1168         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1169 }
1170
1171 static void check_object(struct object_entry *entry)
1172 {
1173         if (entry->in_pack) {
1174                 struct packed_git *p = entry->in_pack;
1175                 struct pack_window *w_curs = NULL;
1176                 const unsigned char *base_ref = NULL;
1177                 struct object_entry *base_entry;
1178                 unsigned long used, used_0;
1179                 unsigned long avail;
1180                 off_t ofs;
1181                 unsigned char *buf, c;
1182
1183                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1184
1185                 /*
1186                  * We want in_pack_type even if we do not reuse delta
1187                  * since non-delta representations could still be reused.
1188                  */
1189                 used = unpack_object_header_buffer(buf, avail,
1190                                                    &entry->in_pack_type,
1191                                                    &entry->size);
1192                 if (used == 0)
1193                         goto give_up;
1194
1195                 /*
1196                  * Determine if this is a delta and if so whether we can
1197                  * reuse it or not.  Otherwise let's find out as cheaply as
1198                  * possible what the actual type and size for this object is.
1199                  */
1200                 switch (entry->in_pack_type) {
1201                 default:
1202                         /* Not a delta hence we've already got all we need. */
1203                         entry->type = entry->in_pack_type;
1204                         entry->in_pack_header_size = used;
1205                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1206                                 goto give_up;
1207                         unuse_pack(&w_curs);
1208                         return;
1209                 case OBJ_REF_DELTA:
1210                         if (reuse_delta && !entry->preferred_base)
1211                                 base_ref = use_pack(p, &w_curs,
1212                                                 entry->in_pack_offset + used, NULL);
1213                         entry->in_pack_header_size = used + 20;
1214                         break;
1215                 case OBJ_OFS_DELTA:
1216                         buf = use_pack(p, &w_curs,
1217                                        entry->in_pack_offset + used, NULL);
1218                         used_0 = 0;
1219                         c = buf[used_0++];
1220                         ofs = c & 127;
1221                         while (c & 128) {
1222                                 ofs += 1;
1223                                 if (!ofs || MSB(ofs, 7)) {
1224                                         error("delta base offset overflow in pack for %s",
1225                                               sha1_to_hex(entry->idx.sha1));
1226                                         goto give_up;
1227                                 }
1228                                 c = buf[used_0++];
1229                                 ofs = (ofs << 7) + (c & 127);
1230                         }
1231                         ofs = entry->in_pack_offset - ofs;
1232                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1233                                 error("delta base offset out of bound for %s",
1234                                       sha1_to_hex(entry->idx.sha1));
1235                                 goto give_up;
1236                         }
1237                         if (reuse_delta && !entry->preferred_base) {
1238                                 struct revindex_entry *revidx;
1239                                 revidx = find_pack_revindex(p, ofs);
1240                                 if (!revidx)
1241                                         goto give_up;
1242                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1243                         }
1244                         entry->in_pack_header_size = used + used_0;
1245                         break;
1246                 }
1247
1248                 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1249                         /*
1250                          * If base_ref was set above that means we wish to
1251                          * reuse delta data, and we even found that base
1252                          * in the list of objects we want to pack. Goodie!
1253                          *
1254                          * Depth value does not matter - find_deltas() will
1255                          * never consider reused delta as the base object to
1256                          * deltify other objects against, in order to avoid
1257                          * circular deltas.
1258                          */
1259                         entry->type = entry->in_pack_type;
1260                         entry->delta = base_entry;
1261                         entry->delta_size = entry->size;
1262                         entry->delta_sibling = base_entry->delta_child;
1263                         base_entry->delta_child = entry;
1264                         unuse_pack(&w_curs);
1265                         return;
1266                 }
1267
1268                 if (entry->type) {
1269                         /*
1270                          * This must be a delta and we already know what the
1271                          * final object type is.  Let's extract the actual
1272                          * object size from the delta header.
1273                          */
1274                         entry->size = get_size_from_delta(p, &w_curs,
1275                                         entry->in_pack_offset + entry->in_pack_header_size);
1276                         if (entry->size == 0)
1277                                 goto give_up;
1278                         unuse_pack(&w_curs);
1279                         return;
1280                 }
1281
1282                 /*
1283                  * No choice but to fall back to the recursive delta walk
1284                  * with sha1_object_info() to find about the object type
1285                  * at this point...
1286                  */
1287                 give_up:
1288                 unuse_pack(&w_curs);
1289         }
1290
1291         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1292         /*
1293          * The error condition is checked in prepare_pack().  This is
1294          * to permit a missing preferred base object to be ignored
1295          * as a preferred base.  Doing so can result in a larger
1296          * pack file, but the transfer will still take place.
1297          */
1298 }
1299
1300 static int pack_offset_sort(const void *_a, const void *_b)
1301 {
1302         const struct object_entry *a = *(struct object_entry **)_a;
1303         const struct object_entry *b = *(struct object_entry **)_b;
1304
1305         /* avoid filesystem trashing with loose objects */
1306         if (!a->in_pack && !b->in_pack)
1307                 return hashcmp(a->idx.sha1, b->idx.sha1);
1308
1309         if (a->in_pack < b->in_pack)
1310                 return -1;
1311         if (a->in_pack > b->in_pack)
1312                 return 1;
1313         return a->in_pack_offset < b->in_pack_offset ? -1 :
1314                         (a->in_pack_offset > b->in_pack_offset);
1315 }
1316
1317 static void get_object_details(void)
1318 {
1319         uint32_t i;
1320         struct object_entry **sorted_by_offset;
1321
1322         sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1323         for (i = 0; i < nr_objects; i++)
1324                 sorted_by_offset[i] = objects + i;
1325         qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1326
1327         for (i = 0; i < nr_objects; i++) {
1328                 struct object_entry *entry = sorted_by_offset[i];
1329                 check_object(entry);
1330                 if (big_file_threshold <= entry->size)
1331                         entry->no_try_delta = 1;
1332         }
1333
1334         free(sorted_by_offset);
1335 }
1336
1337 /*
1338  * We search for deltas in a list sorted by type, by filename hash, and then
1339  * by size, so that we see progressively smaller and smaller files.
1340  * That's because we prefer deltas to be from the bigger file
1341  * to the smaller -- deletes are potentially cheaper, but perhaps
1342  * more importantly, the bigger file is likely the more recent
1343  * one.  The deepest deltas are therefore the oldest objects which are
1344  * less susceptible to be accessed often.
1345  */
1346 static int type_size_sort(const void *_a, const void *_b)
1347 {
1348         const struct object_entry *a = *(struct object_entry **)_a;
1349         const struct object_entry *b = *(struct object_entry **)_b;
1350
1351         if (a->type > b->type)
1352                 return -1;
1353         if (a->type < b->type)
1354                 return 1;
1355         if (a->hash > b->hash)
1356                 return -1;
1357         if (a->hash < b->hash)
1358                 return 1;
1359         if (a->preferred_base > b->preferred_base)
1360                 return -1;
1361         if (a->preferred_base < b->preferred_base)
1362                 return 1;
1363         if (a->size > b->size)
1364                 return -1;
1365         if (a->size < b->size)
1366                 return 1;
1367         return a < b ? -1 : (a > b);  /* newest first */
1368 }
1369
1370 struct unpacked {
1371         struct object_entry *entry;
1372         void *data;
1373         struct delta_index *index;
1374         unsigned depth;
1375 };
1376
1377 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1378                            unsigned long delta_size)
1379 {
1380         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1381                 return 0;
1382
1383         if (delta_size < cache_max_small_delta_size)
1384                 return 1;
1385
1386         /* cache delta, if objects are large enough compared to delta size */
1387         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1388                 return 1;
1389
1390         return 0;
1391 }
1392
1393 #ifndef NO_PTHREADS
1394
1395 static pthread_mutex_t read_mutex;
1396 #define read_lock()             pthread_mutex_lock(&read_mutex)
1397 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1398
1399 static pthread_mutex_t cache_mutex;
1400 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1401 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1402
1403 static pthread_mutex_t progress_mutex;
1404 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1405 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1406
1407 #else
1408
1409 #define read_lock()             (void)0
1410 #define read_unlock()           (void)0
1411 #define cache_lock()            (void)0
1412 #define cache_unlock()          (void)0
1413 #define progress_lock()         (void)0
1414 #define progress_unlock()       (void)0
1415
1416 #endif
1417
1418 static int try_delta(struct unpacked *trg, struct unpacked *src,
1419                      unsigned max_depth, unsigned long *mem_usage)
1420 {
1421         struct object_entry *trg_entry = trg->entry;
1422         struct object_entry *src_entry = src->entry;
1423         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1424         unsigned ref_depth;
1425         enum object_type type;
1426         void *delta_buf;
1427
1428         /* Don't bother doing diffs between different types */
1429         if (trg_entry->type != src_entry->type)
1430                 return -1;
1431
1432         /*
1433          * We do not bother to try a delta that we discarded on an
1434          * earlier try, but only when reusing delta data.  Note that
1435          * src_entry that is marked as the preferred_base should always
1436          * be considered, as even if we produce a suboptimal delta against
1437          * it, we will still save the transfer cost, as we already know
1438          * the other side has it and we won't send src_entry at all.
1439          */
1440         if (reuse_delta && trg_entry->in_pack &&
1441             trg_entry->in_pack == src_entry->in_pack &&
1442             !src_entry->preferred_base &&
1443             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1444             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1445                 return 0;
1446
1447         /* Let's not bust the allowed depth. */
1448         if (src->depth >= max_depth)
1449                 return 0;
1450
1451         /* Now some size filtering heuristics. */
1452         trg_size = trg_entry->size;
1453         if (!trg_entry->delta) {
1454                 max_size = trg_size/2 - 20;
1455                 ref_depth = 1;
1456         } else {
1457                 max_size = trg_entry->delta_size;
1458                 ref_depth = trg->depth;
1459         }
1460         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1461                                                 (max_depth - ref_depth + 1);
1462         if (max_size == 0)
1463                 return 0;
1464         src_size = src_entry->size;
1465         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1466         if (sizediff >= max_size)
1467                 return 0;
1468         if (trg_size < src_size / 32)
1469                 return 0;
1470
1471         /* Load data if not already done */
1472         if (!trg->data) {
1473                 read_lock();
1474                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1475                 read_unlock();
1476                 if (!trg->data)
1477                         die("object %s cannot be read",
1478                             sha1_to_hex(trg_entry->idx.sha1));
1479                 if (sz != trg_size)
1480                         die("object %s inconsistent object length (%lu vs %lu)",
1481                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1482                 *mem_usage += sz;
1483         }
1484         if (!src->data) {
1485                 read_lock();
1486                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1487                 read_unlock();
1488                 if (!src->data) {
1489                         if (src_entry->preferred_base) {
1490                                 static int warned = 0;
1491                                 if (!warned++)
1492                                         warning("object %s cannot be read",
1493                                                 sha1_to_hex(src_entry->idx.sha1));
1494                                 /*
1495                                  * Those objects are not included in the
1496                                  * resulting pack.  Be resilient and ignore
1497                                  * them if they can't be read, in case the
1498                                  * pack could be created nevertheless.
1499                                  */
1500                                 return 0;
1501                         }
1502                         die("object %s cannot be read",
1503                             sha1_to_hex(src_entry->idx.sha1));
1504                 }
1505                 if (sz != src_size)
1506                         die("object %s inconsistent object length (%lu vs %lu)",
1507                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1508                 *mem_usage += sz;
1509         }
1510         if (!src->index) {
1511                 src->index = create_delta_index(src->data, src_size);
1512                 if (!src->index) {
1513                         static int warned = 0;
1514                         if (!warned++)
1515                                 warning("suboptimal pack - out of memory");
1516                         return 0;
1517                 }
1518                 *mem_usage += sizeof_delta_index(src->index);
1519         }
1520
1521         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1522         if (!delta_buf)
1523                 return 0;
1524
1525         if (trg_entry->delta) {
1526                 /* Prefer only shallower same-sized deltas. */
1527                 if (delta_size == trg_entry->delta_size &&
1528                     src->depth + 1 >= trg->depth) {
1529                         free(delta_buf);
1530                         return 0;
1531                 }
1532         }
1533
1534         /*
1535          * Handle memory allocation outside of the cache
1536          * accounting lock.  Compiler will optimize the strangeness
1537          * away when NO_PTHREADS is defined.
1538          */
1539         free(trg_entry->delta_data);
1540         cache_lock();
1541         if (trg_entry->delta_data) {
1542                 delta_cache_size -= trg_entry->delta_size;
1543                 trg_entry->delta_data = NULL;
1544         }
1545         if (delta_cacheable(src_size, trg_size, delta_size)) {
1546                 delta_cache_size += delta_size;
1547                 cache_unlock();
1548                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1549         } else {
1550                 cache_unlock();
1551                 free(delta_buf);
1552         }
1553
1554         trg_entry->delta = src_entry;
1555         trg_entry->delta_size = delta_size;
1556         trg->depth = src->depth + 1;
1557
1558         return 1;
1559 }
1560
1561 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1562 {
1563         struct object_entry *child = me->delta_child;
1564         unsigned int m = n;
1565         while (child) {
1566                 unsigned int c = check_delta_limit(child, n + 1);
1567                 if (m < c)
1568                         m = c;
1569                 child = child->delta_sibling;
1570         }
1571         return m;
1572 }
1573
1574 static unsigned long free_unpacked(struct unpacked *n)
1575 {
1576         unsigned long freed_mem = sizeof_delta_index(n->index);
1577         free_delta_index(n->index);
1578         n->index = NULL;
1579         if (n->data) {
1580                 freed_mem += n->entry->size;
1581                 free(n->data);
1582                 n->data = NULL;
1583         }
1584         n->entry = NULL;
1585         n->depth = 0;
1586         return freed_mem;
1587 }
1588
1589 static void find_deltas(struct object_entry **list, unsigned *list_size,
1590                         int window, int depth, unsigned *processed)
1591 {
1592         uint32_t i, idx = 0, count = 0;
1593         struct unpacked *array;
1594         unsigned long mem_usage = 0;
1595
1596         array = xcalloc(window, sizeof(struct unpacked));
1597
1598         for (;;) {
1599                 struct object_entry *entry;
1600                 struct unpacked *n = array + idx;
1601                 int j, max_depth, best_base = -1;
1602
1603                 progress_lock();
1604                 if (!*list_size) {
1605                         progress_unlock();
1606                         break;
1607                 }
1608                 entry = *list++;
1609                 (*list_size)--;
1610                 if (!entry->preferred_base) {
1611                         (*processed)++;
1612                         display_progress(progress_state, *processed);
1613                 }
1614                 progress_unlock();
1615
1616                 mem_usage -= free_unpacked(n);
1617                 n->entry = entry;
1618
1619                 while (window_memory_limit &&
1620                        mem_usage > window_memory_limit &&
1621                        count > 1) {
1622                         uint32_t tail = (idx + window - count) % window;
1623                         mem_usage -= free_unpacked(array + tail);
1624                         count--;
1625                 }
1626
1627                 /* We do not compute delta to *create* objects we are not
1628                  * going to pack.
1629                  */
1630                 if (entry->preferred_base)
1631                         goto next;
1632
1633                 /*
1634                  * If the current object is at pack edge, take the depth the
1635                  * objects that depend on the current object into account
1636                  * otherwise they would become too deep.
1637                  */
1638                 max_depth = depth;
1639                 if (entry->delta_child) {
1640                         max_depth -= check_delta_limit(entry, 0);
1641                         if (max_depth <= 0)
1642                                 goto next;
1643                 }
1644
1645                 j = window;
1646                 while (--j > 0) {
1647                         int ret;
1648                         uint32_t other_idx = idx + j;
1649                         struct unpacked *m;
1650                         if (other_idx >= window)
1651                                 other_idx -= window;
1652                         m = array + other_idx;
1653                         if (!m->entry)
1654                                 break;
1655                         ret = try_delta(n, m, max_depth, &mem_usage);
1656                         if (ret < 0)
1657                                 break;
1658                         else if (ret > 0)
1659                                 best_base = other_idx;
1660                 }
1661
1662                 /*
1663                  * If we decided to cache the delta data, then it is best
1664                  * to compress it right away.  First because we have to do
1665                  * it anyway, and doing it here while we're threaded will
1666                  * save a lot of time in the non threaded write phase,
1667                  * as well as allow for caching more deltas within
1668                  * the same cache size limit.
1669                  * ...
1670                  * But only if not writing to stdout, since in that case
1671                  * the network is most likely throttling writes anyway,
1672                  * and therefore it is best to go to the write phase ASAP
1673                  * instead, as we can afford spending more time compressing
1674                  * between writes at that moment.
1675                  */
1676                 if (entry->delta_data && !pack_to_stdout) {
1677                         entry->z_delta_size = do_compress(&entry->delta_data,
1678                                                           entry->delta_size);
1679                         cache_lock();
1680                         delta_cache_size -= entry->delta_size;
1681                         delta_cache_size += entry->z_delta_size;
1682                         cache_unlock();
1683                 }
1684
1685                 /* if we made n a delta, and if n is already at max
1686                  * depth, leaving it in the window is pointless.  we
1687                  * should evict it first.
1688                  */
1689                 if (entry->delta && max_depth <= n->depth)
1690                         continue;
1691
1692                 /*
1693                  * Move the best delta base up in the window, after the
1694                  * currently deltified object, to keep it longer.  It will
1695                  * be the first base object to be attempted next.
1696                  */
1697                 if (entry->delta) {
1698                         struct unpacked swap = array[best_base];
1699                         int dist = (window + idx - best_base) % window;
1700                         int dst = best_base;
1701                         while (dist--) {
1702                                 int src = (dst + 1) % window;
1703                                 array[dst] = array[src];
1704                                 dst = src;
1705                         }
1706                         array[dst] = swap;
1707                 }
1708
1709                 next:
1710                 idx++;
1711                 if (count + 1 < window)
1712                         count++;
1713                 if (idx >= window)
1714                         idx = 0;
1715         }
1716
1717         for (i = 0; i < window; ++i) {
1718                 free_delta_index(array[i].index);
1719                 free(array[i].data);
1720         }
1721         free(array);
1722 }
1723
1724 #ifndef NO_PTHREADS
1725
1726 static void try_to_free_from_threads(size_t size)
1727 {
1728         read_lock();
1729         release_pack_memory(size, -1);
1730         read_unlock();
1731 }
1732
1733 static try_to_free_t old_try_to_free_routine;
1734
1735 /*
1736  * The main thread waits on the condition that (at least) one of the workers
1737  * has stopped working (which is indicated in the .working member of
1738  * struct thread_params).
1739  * When a work thread has completed its work, it sets .working to 0 and
1740  * signals the main thread and waits on the condition that .data_ready
1741  * becomes 1.
1742  */
1743
1744 struct thread_params {
1745         pthread_t thread;
1746         struct object_entry **list;
1747         unsigned list_size;
1748         unsigned remaining;
1749         int window;
1750         int depth;
1751         int working;
1752         int data_ready;
1753         pthread_mutex_t mutex;
1754         pthread_cond_t cond;
1755         unsigned *processed;
1756 };
1757
1758 static pthread_cond_t progress_cond;
1759
1760 /*
1761  * Mutex and conditional variable can't be statically-initialized on Windows.
1762  */
1763 static void init_threaded_search(void)
1764 {
1765         init_recursive_mutex(&read_mutex);
1766         pthread_mutex_init(&cache_mutex, NULL);
1767         pthread_mutex_init(&progress_mutex, NULL);
1768         pthread_cond_init(&progress_cond, NULL);
1769         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1770 }
1771
1772 static void cleanup_threaded_search(void)
1773 {
1774         set_try_to_free_routine(old_try_to_free_routine);
1775         pthread_cond_destroy(&progress_cond);
1776         pthread_mutex_destroy(&read_mutex);
1777         pthread_mutex_destroy(&cache_mutex);
1778         pthread_mutex_destroy(&progress_mutex);
1779 }
1780
1781 static void *threaded_find_deltas(void *arg)
1782 {
1783         struct thread_params *me = arg;
1784
1785         while (me->remaining) {
1786                 find_deltas(me->list, &me->remaining,
1787                             me->window, me->depth, me->processed);
1788
1789                 progress_lock();
1790                 me->working = 0;
1791                 pthread_cond_signal(&progress_cond);
1792                 progress_unlock();
1793
1794                 /*
1795                  * We must not set ->data_ready before we wait on the
1796                  * condition because the main thread may have set it to 1
1797                  * before we get here. In order to be sure that new
1798                  * work is available if we see 1 in ->data_ready, it
1799                  * was initialized to 0 before this thread was spawned
1800                  * and we reset it to 0 right away.
1801                  */
1802                 pthread_mutex_lock(&me->mutex);
1803                 while (!me->data_ready)
1804                         pthread_cond_wait(&me->cond, &me->mutex);
1805                 me->data_ready = 0;
1806                 pthread_mutex_unlock(&me->mutex);
1807         }
1808         /* leave ->working 1 so that this doesn't get more work assigned */
1809         return NULL;
1810 }
1811
1812 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1813                            int window, int depth, unsigned *processed)
1814 {
1815         struct thread_params *p;
1816         int i, ret, active_threads = 0;
1817
1818         init_threaded_search();
1819
1820         if (!delta_search_threads)      /* --threads=0 means autodetect */
1821                 delta_search_threads = online_cpus();
1822         if (delta_search_threads <= 1) {
1823                 find_deltas(list, &list_size, window, depth, processed);
1824                 cleanup_threaded_search();
1825                 return;
1826         }
1827         if (progress > pack_to_stdout)
1828                 fprintf(stderr, "Delta compression using up to %d threads.\n",
1829                                 delta_search_threads);
1830         p = xcalloc(delta_search_threads, sizeof(*p));
1831
1832         /* Partition the work amongst work threads. */
1833         for (i = 0; i < delta_search_threads; i++) {
1834                 unsigned sub_size = list_size / (delta_search_threads - i);
1835
1836                 /* don't use too small segments or no deltas will be found */
1837                 if (sub_size < 2*window && i+1 < delta_search_threads)
1838                         sub_size = 0;
1839
1840                 p[i].window = window;
1841                 p[i].depth = depth;
1842                 p[i].processed = processed;
1843                 p[i].working = 1;
1844                 p[i].data_ready = 0;
1845
1846                 /* try to split chunks on "path" boundaries */
1847                 while (sub_size && sub_size < list_size &&
1848                        list[sub_size]->hash &&
1849                        list[sub_size]->hash == list[sub_size-1]->hash)
1850                         sub_size++;
1851
1852                 p[i].list = list;
1853                 p[i].list_size = sub_size;
1854                 p[i].remaining = sub_size;
1855
1856                 list += sub_size;
1857                 list_size -= sub_size;
1858         }
1859
1860         /* Start work threads. */
1861         for (i = 0; i < delta_search_threads; i++) {
1862                 if (!p[i].list_size)
1863                         continue;
1864                 pthread_mutex_init(&p[i].mutex, NULL);
1865                 pthread_cond_init(&p[i].cond, NULL);
1866                 ret = pthread_create(&p[i].thread, NULL,
1867                                      threaded_find_deltas, &p[i]);
1868                 if (ret)
1869                         die("unable to create thread: %s", strerror(ret));
1870                 active_threads++;
1871         }
1872
1873         /*
1874          * Now let's wait for work completion.  Each time a thread is done
1875          * with its work, we steal half of the remaining work from the
1876          * thread with the largest number of unprocessed objects and give
1877          * it to that newly idle thread.  This ensure good load balancing
1878          * until the remaining object list segments are simply too short
1879          * to be worth splitting anymore.
1880          */
1881         while (active_threads) {
1882                 struct thread_params *target = NULL;
1883                 struct thread_params *victim = NULL;
1884                 unsigned sub_size = 0;
1885
1886                 progress_lock();
1887                 for (;;) {
1888                         for (i = 0; !target && i < delta_search_threads; i++)
1889                                 if (!p[i].working)
1890                                         target = &p[i];
1891                         if (target)
1892                                 break;
1893                         pthread_cond_wait(&progress_cond, &progress_mutex);
1894                 }
1895
1896                 for (i = 0; i < delta_search_threads; i++)
1897                         if (p[i].remaining > 2*window &&
1898                             (!victim || victim->remaining < p[i].remaining))
1899                                 victim = &p[i];
1900                 if (victim) {
1901                         sub_size = victim->remaining / 2;
1902                         list = victim->list + victim->list_size - sub_size;
1903                         while (sub_size && list[0]->hash &&
1904                                list[0]->hash == list[-1]->hash) {
1905                                 list++;
1906                                 sub_size--;
1907                         }
1908                         if (!sub_size) {
1909                                 /*
1910                                  * It is possible for some "paths" to have
1911                                  * so many objects that no hash boundary
1912                                  * might be found.  Let's just steal the
1913                                  * exact half in that case.
1914                                  */
1915                                 sub_size = victim->remaining / 2;
1916                                 list -= sub_size;
1917                         }
1918                         target->list = list;
1919                         victim->list_size -= sub_size;
1920                         victim->remaining -= sub_size;
1921                 }
1922                 target->list_size = sub_size;
1923                 target->remaining = sub_size;
1924                 target->working = 1;
1925                 progress_unlock();
1926
1927                 pthread_mutex_lock(&target->mutex);
1928                 target->data_ready = 1;
1929                 pthread_cond_signal(&target->cond);
1930                 pthread_mutex_unlock(&target->mutex);
1931
1932                 if (!sub_size) {
1933                         pthread_join(target->thread, NULL);
1934                         pthread_cond_destroy(&target->cond);
1935                         pthread_mutex_destroy(&target->mutex);
1936                         active_threads--;
1937                 }
1938         }
1939         cleanup_threaded_search();
1940         free(p);
1941 }
1942
1943 #else
1944 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
1945 #endif
1946
1947 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1948 {
1949         unsigned char peeled[20];
1950
1951         if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1952             !peel_ref(path, peeled)        && /* peelable? */
1953             !is_null_sha1(peeled)          && /* annotated tag? */
1954             locate_object_entry(peeled))      /* object packed? */
1955                 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1956         return 0;
1957 }
1958
1959 static void prepare_pack(int window, int depth)
1960 {
1961         struct object_entry **delta_list;
1962         uint32_t i, nr_deltas;
1963         unsigned n;
1964
1965         get_object_details();
1966
1967         /*
1968          * If we're locally repacking then we need to be doubly careful
1969          * from now on in order to make sure no stealth corruption gets
1970          * propagated to the new pack.  Clients receiving streamed packs
1971          * should validate everything they get anyway so no need to incur
1972          * the additional cost here in that case.
1973          */
1974         if (!pack_to_stdout)
1975                 do_check_packed_object_crc = 1;
1976
1977         if (!nr_objects || !window || !depth)
1978                 return;
1979
1980         delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1981         nr_deltas = n = 0;
1982
1983         for (i = 0; i < nr_objects; i++) {
1984                 struct object_entry *entry = objects + i;
1985
1986                 if (entry->delta)
1987                         /* This happens if we decided to reuse existing
1988                          * delta from a pack.  "reuse_delta &&" is implied.
1989                          */
1990                         continue;
1991
1992                 if (entry->size < 50)
1993                         continue;
1994
1995                 if (entry->no_try_delta)
1996                         continue;
1997
1998                 if (!entry->preferred_base) {
1999                         nr_deltas++;
2000                         if (entry->type < 0)
2001                                 die("unable to get type of object %s",
2002                                     sha1_to_hex(entry->idx.sha1));
2003                 } else {
2004                         if (entry->type < 0) {
2005                                 /*
2006                                  * This object is not found, but we
2007                                  * don't have to include it anyway.
2008                                  */
2009                                 continue;
2010                         }
2011                 }
2012
2013                 delta_list[n++] = entry;
2014         }
2015
2016         if (nr_deltas && n > 1) {
2017                 unsigned nr_done = 0;
2018                 if (progress)
2019                         progress_state = start_progress("Compressing objects",
2020                                                         nr_deltas);
2021                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2022                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2023                 stop_progress(&progress_state);
2024                 if (nr_done != nr_deltas)
2025                         die("inconsistency with delta count");
2026         }
2027         free(delta_list);
2028 }
2029
2030 static int git_pack_config(const char *k, const char *v, void *cb)
2031 {
2032         if (!strcmp(k, "pack.window")) {
2033                 window = git_config_int(k, v);
2034                 return 0;
2035         }
2036         if (!strcmp(k, "pack.windowmemory")) {
2037                 window_memory_limit = git_config_ulong(k, v);
2038                 return 0;
2039         }
2040         if (!strcmp(k, "pack.depth")) {
2041                 depth = git_config_int(k, v);
2042                 return 0;
2043         }
2044         if (!strcmp(k, "pack.compression")) {
2045                 int level = git_config_int(k, v);
2046                 if (level == -1)
2047                         level = Z_DEFAULT_COMPRESSION;
2048                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2049                         die("bad pack compression level %d", level);
2050                 pack_compression_level = level;
2051                 pack_compression_seen = 1;
2052                 return 0;
2053         }
2054         if (!strcmp(k, "pack.deltacachesize")) {
2055                 max_delta_cache_size = git_config_int(k, v);
2056                 return 0;
2057         }
2058         if (!strcmp(k, "pack.deltacachelimit")) {
2059                 cache_max_small_delta_size = git_config_int(k, v);
2060                 return 0;
2061         }
2062         if (!strcmp(k, "pack.threads")) {
2063                 delta_search_threads = git_config_int(k, v);
2064                 if (delta_search_threads < 0)
2065                         die("invalid number of threads specified (%d)",
2066                             delta_search_threads);
2067 #ifdef NO_PTHREADS
2068                 if (delta_search_threads != 1)
2069                         warning("no threads support, ignoring %s", k);
2070 #endif
2071                 return 0;
2072         }
2073         if (!strcmp(k, "pack.indexversion")) {
2074                 pack_idx_opts.version = git_config_int(k, v);
2075                 if (pack_idx_opts.version > 2)
2076                         die("bad pack.indexversion=%"PRIu32,
2077                             pack_idx_opts.version);
2078                 return 0;
2079         }
2080         return git_default_config(k, v, cb);
2081 }
2082
2083 static void read_object_list_from_stdin(void)
2084 {
2085         char line[40 + 1 + PATH_MAX + 2];
2086         unsigned char sha1[20];
2087
2088         for (;;) {
2089                 if (!fgets(line, sizeof(line), stdin)) {
2090                         if (feof(stdin))
2091                                 break;
2092                         if (!ferror(stdin))
2093                                 die("fgets returned NULL, not EOF, not error!");
2094                         if (errno != EINTR)
2095                                 die_errno("fgets");
2096                         clearerr(stdin);
2097                         continue;
2098                 }
2099                 if (line[0] == '-') {
2100                         if (get_sha1_hex(line+1, sha1))
2101                                 die("expected edge sha1, got garbage:\n %s",
2102                                     line);
2103                         add_preferred_base(sha1);
2104                         continue;
2105                 }
2106                 if (get_sha1_hex(line, sha1))
2107                         die("expected sha1, got garbage:\n %s", line);
2108
2109                 add_preferred_base_object(line+41);
2110                 add_object_entry(sha1, 0, line+41, 0);
2111         }
2112 }
2113
2114 #define OBJECT_ADDED (1u<<20)
2115
2116 static void show_commit(struct commit *commit, void *data)
2117 {
2118         add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2119         commit->object.flags |= OBJECT_ADDED;
2120 }
2121
2122 static void show_object(struct object *obj,
2123                         const struct name_path *path, const char *last,
2124                         void *data)
2125 {
2126         char *name = path_name(path, last);
2127
2128         add_preferred_base_object(name);
2129         add_object_entry(obj->sha1, obj->type, name, 0);
2130         obj->flags |= OBJECT_ADDED;
2131
2132         /*
2133          * We will have generated the hash from the name,
2134          * but not saved a pointer to it - we can free it
2135          */
2136         free((char *)name);
2137 }
2138
2139 static void show_edge(struct commit *commit)
2140 {
2141         add_preferred_base(commit->object.sha1);
2142 }
2143
2144 struct in_pack_object {
2145         off_t offset;
2146         struct object *object;
2147 };
2148
2149 struct in_pack {
2150         int alloc;
2151         int nr;
2152         struct in_pack_object *array;
2153 };
2154
2155 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2156 {
2157         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2158         in_pack->array[in_pack->nr].object = object;
2159         in_pack->nr++;
2160 }
2161
2162 /*
2163  * Compare the objects in the offset order, in order to emulate the
2164  * "git rev-list --objects" output that produced the pack originally.
2165  */
2166 static int ofscmp(const void *a_, const void *b_)
2167 {
2168         struct in_pack_object *a = (struct in_pack_object *)a_;
2169         struct in_pack_object *b = (struct in_pack_object *)b_;
2170
2171         if (a->offset < b->offset)
2172                 return -1;
2173         else if (a->offset > b->offset)
2174                 return 1;
2175         else
2176                 return hashcmp(a->object->sha1, b->object->sha1);
2177 }
2178
2179 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2180 {
2181         struct packed_git *p;
2182         struct in_pack in_pack;
2183         uint32_t i;
2184
2185         memset(&in_pack, 0, sizeof(in_pack));
2186
2187         for (p = packed_git; p; p = p->next) {
2188                 const unsigned char *sha1;
2189                 struct object *o;
2190
2191                 if (!p->pack_local || p->pack_keep)
2192                         continue;
2193                 if (open_pack_index(p))
2194                         die("cannot open pack index");
2195
2196                 ALLOC_GROW(in_pack.array,
2197                            in_pack.nr + p->num_objects,
2198                            in_pack.alloc);
2199
2200                 for (i = 0; i < p->num_objects; i++) {
2201                         sha1 = nth_packed_object_sha1(p, i);
2202                         o = lookup_unknown_object(sha1);
2203                         if (!(o->flags & OBJECT_ADDED))
2204                                 mark_in_pack_object(o, p, &in_pack);
2205                         o->flags |= OBJECT_ADDED;
2206                 }
2207         }
2208
2209         if (in_pack.nr) {
2210                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2211                       ofscmp);
2212                 for (i = 0; i < in_pack.nr; i++) {
2213                         struct object *o = in_pack.array[i].object;
2214                         add_object_entry(o->sha1, o->type, "", 0);
2215                 }
2216         }
2217         free(in_pack.array);
2218 }
2219
2220 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2221 {
2222         static struct packed_git *last_found = (void *)1;
2223         struct packed_git *p;
2224
2225         p = (last_found != (void *)1) ? last_found : packed_git;
2226
2227         while (p) {
2228                 if ((!p->pack_local || p->pack_keep) &&
2229                         find_pack_entry_one(sha1, p)) {
2230                         last_found = p;
2231                         return 1;
2232                 }
2233                 if (p == last_found)
2234                         p = packed_git;
2235                 else
2236                         p = p->next;
2237                 if (p == last_found)
2238                         p = p->next;
2239         }
2240         return 0;
2241 }
2242
2243 static void loosen_unused_packed_objects(struct rev_info *revs)
2244 {
2245         struct packed_git *p;
2246         uint32_t i;
2247         const unsigned char *sha1;
2248
2249         for (p = packed_git; p; p = p->next) {
2250                 if (!p->pack_local || p->pack_keep)
2251                         continue;
2252
2253                 if (unpack_unreachable_expiration &&
2254                     p->mtime < unpack_unreachable_expiration)
2255                         continue;
2256
2257                 if (open_pack_index(p))
2258                         die("cannot open pack index");
2259
2260                 for (i = 0; i < p->num_objects; i++) {
2261                         sha1 = nth_packed_object_sha1(p, i);
2262                         if (!locate_object_entry(sha1) &&
2263                                 !has_sha1_pack_kept_or_nonlocal(sha1))
2264                                 if (force_object_loose(sha1, p->mtime))
2265                                         die("unable to force loose object");
2266                 }
2267         }
2268 }
2269
2270 static void get_object_list(int ac, const char **av)
2271 {
2272         struct rev_info revs;
2273         char line[1000];
2274         int flags = 0;
2275
2276         init_revisions(&revs, NULL);
2277         save_commit_buffer = 0;
2278         setup_revisions(ac, av, &revs, NULL);
2279
2280         while (fgets(line, sizeof(line), stdin) != NULL) {
2281                 int len = strlen(line);
2282                 if (len && line[len - 1] == '\n')
2283                         line[--len] = 0;
2284                 if (!len)
2285                         break;
2286                 if (*line == '-') {
2287                         if (!strcmp(line, "--not")) {
2288                                 flags ^= UNINTERESTING;
2289                                 continue;
2290                         }
2291                         die("not a rev '%s'", line);
2292                 }
2293                 if (handle_revision_arg(line, &revs, flags, 1))
2294                         die("bad revision '%s'", line);
2295         }
2296
2297         if (prepare_revision_walk(&revs))
2298                 die("revision walk setup failed");
2299         mark_edges_uninteresting(revs.commits, &revs, show_edge);
2300         traverse_commit_list(&revs, show_commit, show_object, NULL);
2301
2302         if (keep_unreachable)
2303                 add_objects_in_unpacked_packs(&revs);
2304         if (unpack_unreachable)
2305                 loosen_unused_packed_objects(&revs);
2306 }
2307
2308 static int option_parse_index_version(const struct option *opt,
2309                                       const char *arg, int unset)
2310 {
2311         char *c;
2312         const char *val = arg;
2313         pack_idx_opts.version = strtoul(val, &c, 10);
2314         if (pack_idx_opts.version > 2)
2315                 die(_("unsupported index version %s"), val);
2316         if (*c == ',' && c[1])
2317                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2318         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2319                 die(_("bad index version '%s'"), val);
2320         return 0;
2321 }
2322
2323 static int option_parse_unpack_unreachable(const struct option *opt,
2324                                            const char *arg, int unset)
2325 {
2326         if (unset) {
2327                 unpack_unreachable = 0;
2328                 unpack_unreachable_expiration = 0;
2329         }
2330         else {
2331                 unpack_unreachable = 1;
2332                 if (arg)
2333                         unpack_unreachable_expiration = approxidate(arg);
2334         }
2335         return 0;
2336 }
2337
2338 static int option_parse_ulong(const struct option *opt,
2339                               const char *arg, int unset)
2340 {
2341         if (unset)
2342                 die(_("option %s does not accept negative form"),
2343                     opt->long_name);
2344
2345         if (!git_parse_ulong(arg, opt->value))
2346                 die(_("unable to parse value '%s' for option %s"),
2347                     arg, opt->long_name);
2348         return 0;
2349 }
2350
2351 #define OPT_ULONG(s, l, v, h) \
2352         { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2353           PARSE_OPT_NONEG, option_parse_ulong }
2354
2355 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2356 {
2357         int use_internal_rev_list = 0;
2358         int thin = 0;
2359         int all_progress_implied = 0;
2360         const char *rp_av[6];
2361         int rp_ac = 0;
2362         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2363         struct option pack_objects_options[] = {
2364                 OPT_SET_INT('q', "quiet", &progress,
2365                             "do not show progress meter", 0),
2366                 OPT_SET_INT(0, "progress", &progress,
2367                             "show progress meter", 1),
2368                 OPT_SET_INT(0, "all-progress", &progress,
2369                             "show progress meter during object writing phase", 2),
2370                 OPT_BOOL(0, "all-progress-implied",
2371                          &all_progress_implied,
2372                          "similar to --all-progress when progress meter is shown"),
2373                 { OPTION_CALLBACK, 0, "index-version", NULL, "version[,offset]",
2374                   "write the pack index file in the specified idx format version",
2375                   0, option_parse_index_version },
2376                 OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2377                           "maximum size of each output pack file"),
2378                 OPT_BOOL(0, "local", &local,
2379                          "ignore borrowed objects from alternate object store"),
2380                 OPT_BOOL(0, "incremental", &incremental,
2381                          "ignore packed objects"),
2382                 OPT_INTEGER(0, "window", &window,
2383                             "limit pack window by objects"),
2384                 OPT_ULONG(0, "window-memory", &window_memory_limit,
2385                           "limit pack window by memory in addition to object limit"),
2386                 OPT_INTEGER(0, "depth", &depth,
2387                             "maximum length of delta chain allowed in the resulting pack"),
2388                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2389                          "reuse existing deltas"),
2390                 OPT_BOOL(0, "reuse-object", &reuse_object,
2391                          "reuse existing objects"),
2392                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2393                          "use OFS_DELTA objects"),
2394                 OPT_INTEGER(0, "threads", &delta_search_threads,
2395                             "use threads when searching for best delta matches"),
2396                 OPT_BOOL(0, "non-empty", &non_empty,
2397                          "do not create an empty pack output"),
2398                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2399                          "read revision arguments from standard input"),
2400                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2401                   "limit the objects to those that are not yet packed",
2402                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2403                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2404                   "include objects reachable from any reference",
2405                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2406                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2407                   "include objects referred by reflog entries",
2408                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2409                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2410                          "output pack to stdout"),
2411                 OPT_BOOL(0, "include-tag", &include_tag,
2412                          "include tag objects that refer to objects to be packed"),
2413                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2414                          "keep unreachable objects"),
2415                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, "time",
2416                   "unpack unreachable objects newer than <time>",
2417                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2418                 OPT_BOOL(0, "thin", &thin,
2419                          "create thin packs"),
2420                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2421                          "ignore packs that have companion .keep file"),
2422                 OPT_INTEGER(0, "compression", &pack_compression_level,
2423                             "pack compression level"),
2424                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2425                             "do not hide commits by grafts", 0),
2426                 OPT_END(),
2427         };
2428
2429         read_replace_refs = 0;
2430
2431         reset_pack_idx_option(&pack_idx_opts);
2432         git_config(git_pack_config, NULL);
2433         if (!pack_compression_seen && core_compression_seen)
2434                 pack_compression_level = core_compression_level;
2435
2436         progress = isatty(2);
2437         argc = parse_options(argc, argv, prefix, pack_objects_options,
2438                              pack_usage, 0);
2439
2440         if (argc) {
2441                 base_name = argv[0];
2442                 argc--;
2443         }
2444         if (pack_to_stdout != !base_name || argc)
2445                 usage_with_options(pack_usage, pack_objects_options);
2446
2447         rp_av[rp_ac++] = "pack-objects";
2448         if (thin) {
2449                 use_internal_rev_list = 1;
2450                 rp_av[rp_ac++] = "--objects-edge";
2451         } else
2452                 rp_av[rp_ac++] = "--objects";
2453
2454         if (rev_list_all) {
2455                 use_internal_rev_list = 1;
2456                 rp_av[rp_ac++] = "--all";
2457         }
2458         if (rev_list_reflog) {
2459                 use_internal_rev_list = 1;
2460                 rp_av[rp_ac++] = "--reflog";
2461         }
2462         if (rev_list_unpacked) {
2463                 use_internal_rev_list = 1;
2464                 rp_av[rp_ac++] = "--unpacked";
2465         }
2466
2467         if (!reuse_object)
2468                 reuse_delta = 0;
2469         if (pack_compression_level == -1)
2470                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2471         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2472                 die("bad pack compression level %d", pack_compression_level);
2473 #ifdef NO_PTHREADS
2474         if (delta_search_threads != 1)
2475                 warning("no threads support, ignoring --threads");
2476 #endif
2477         if (!pack_to_stdout && !pack_size_limit)
2478                 pack_size_limit = pack_size_limit_cfg;
2479         if (pack_to_stdout && pack_size_limit)
2480                 die("--max-pack-size cannot be used to build a pack for transfer.");
2481         if (pack_size_limit && pack_size_limit < 1024*1024) {
2482                 warning("minimum pack size limit is 1 MiB");
2483                 pack_size_limit = 1024*1024;
2484         }
2485
2486         if (!pack_to_stdout && thin)
2487                 die("--thin cannot be used to build an indexable pack.");
2488
2489         if (keep_unreachable && unpack_unreachable)
2490                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2491
2492         if (progress && all_progress_implied)
2493                 progress = 2;
2494
2495         prepare_packed_git();
2496
2497         if (progress)
2498                 progress_state = start_progress("Counting objects", 0);
2499         if (!use_internal_rev_list)
2500                 read_object_list_from_stdin();
2501         else {
2502                 rp_av[rp_ac] = NULL;
2503                 get_object_list(rp_ac, rp_av);
2504         }
2505         cleanup_preferred_base();
2506         if (include_tag && nr_result)
2507                 for_each_ref(add_ref_tag, NULL);
2508         stop_progress(&progress_state);
2509
2510         if (non_empty && !nr_result)
2511                 return 0;
2512         if (nr_result)
2513                 prepare_pack(window, depth);
2514         write_pack_file();
2515         if (progress)
2516                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2517                         " reused %"PRIu32" (delta %"PRIu32")\n",
2518                         written, written_delta, reused, reused_delta);
2519         return 0;
2520 }