Merge branch 'jn/test-prereq-perl-doc'
[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 "streaming.h"
20 #include "thread-utils.h"
21
22 static const char *pack_usage[] = {
23         N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
24         N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
25         NULL
26 };
27
28 struct object_entry {
29         struct pack_idx_entry idx;
30         unsigned long size;     /* uncompressed size */
31         struct packed_git *in_pack;     /* already in pack */
32         off_t in_pack_offset;
33         struct object_entry *delta;     /* delta base object */
34         struct object_entry *delta_child; /* deltified objects who bases me */
35         struct object_entry *delta_sibling; /* other deltified objects who
36                                              * uses the same base as me
37                                              */
38         void *delta_data;       /* cached delta (uncompressed) */
39         unsigned long delta_size;       /* delta data size (uncompressed) */
40         unsigned long z_delta_size;     /* delta data size (compressed) */
41         enum object_type type;
42         enum object_type in_pack_type;  /* could be delta */
43         uint32_t hash;                  /* name hint hash */
44         unsigned char in_pack_header_size;
45         unsigned preferred_base:1; /*
46                                     * we do not pack this, but is available
47                                     * to be used as the base object to delta
48                                     * objects against.
49                                     */
50         unsigned no_try_delta:1;
51         unsigned tagged:1; /* near the very tip of refs */
52         unsigned filled:1; /* assigned write-order */
53 };
54
55 /*
56  * Objects we are going to pack are collected in objects array (dynamically
57  * expanded).  nr_objects & nr_alloc controls this array.  They are stored
58  * in the order we see -- typically rev-list --objects order that gives us
59  * nice "minimum seek" order.
60  */
61 static struct object_entry *objects;
62 static struct pack_idx_entry **written_list;
63 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
64
65 static int non_empty;
66 static int reuse_delta = 1, reuse_object = 1;
67 static int keep_unreachable, unpack_unreachable, include_tag;
68 static unsigned long unpack_unreachable_expiration;
69 static int local;
70 static int incremental;
71 static int ignore_packed_keep;
72 static int allow_ofs_delta;
73 static struct pack_idx_option pack_idx_opts;
74 static const char *base_name;
75 static int progress = 1;
76 static int window = 10;
77 static unsigned long pack_size_limit;
78 static int depth = 50;
79 static int delta_search_threads;
80 static int pack_to_stdout;
81 static int num_preferred_base;
82 static struct progress *progress_state;
83 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
84 static int pack_compression_seen;
85
86 static unsigned long delta_cache_size = 0;
87 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
88 static unsigned long cache_max_small_delta_size = 1000;
89
90 static unsigned long window_memory_limit = 0;
91
92 /*
93  * The object names in objects array are hashed with this hashtable,
94  * to help looking up the entry by object name.
95  * This hashtable is built after all the objects are seen.
96  */
97 static int *object_ix;
98 static int object_ix_hashsz;
99 static struct object_entry *locate_object_entry(const unsigned char *sha1);
100
101 /*
102  * stats
103  */
104 static uint32_t written, written_delta;
105 static uint32_t reused, reused_delta;
106
107
108 static void *get_delta(struct object_entry *entry)
109 {
110         unsigned long size, base_size, delta_size;
111         void *buf, *base_buf, *delta_buf;
112         enum object_type type;
113
114         buf = read_sha1_file(entry->idx.sha1, &type, &size);
115         if (!buf)
116                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
117         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
118         if (!base_buf)
119                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
120         delta_buf = diff_delta(base_buf, base_size,
121                                buf, size, &delta_size, 0);
122         if (!delta_buf || delta_size != entry->delta_size)
123                 die("delta size changed");
124         free(buf);
125         free(base_buf);
126         return delta_buf;
127 }
128
129 static unsigned long do_compress(void **pptr, unsigned long size)
130 {
131         git_zstream stream;
132         void *in, *out;
133         unsigned long maxsize;
134
135         memset(&stream, 0, sizeof(stream));
136         git_deflate_init(&stream, pack_compression_level);
137         maxsize = git_deflate_bound(&stream, size);
138
139         in = *pptr;
140         out = xmalloc(maxsize);
141         *pptr = out;
142
143         stream.next_in = in;
144         stream.avail_in = size;
145         stream.next_out = out;
146         stream.avail_out = maxsize;
147         while (git_deflate(&stream, Z_FINISH) == Z_OK)
148                 ; /* nothing */
149         git_deflate_end(&stream);
150
151         free(in);
152         return stream.total_out;
153 }
154
155 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
156                                            const unsigned char *sha1)
157 {
158         git_zstream stream;
159         unsigned char ibuf[1024 * 16];
160         unsigned char obuf[1024 * 16];
161         unsigned long olen = 0;
162
163         memset(&stream, 0, sizeof(stream));
164         git_deflate_init(&stream, pack_compression_level);
165
166         for (;;) {
167                 ssize_t readlen;
168                 int zret = Z_OK;
169                 readlen = read_istream(st, ibuf, sizeof(ibuf));
170                 if (readlen == -1)
171                         die(_("unable to read %s"), sha1_to_hex(sha1));
172
173                 stream.next_in = ibuf;
174                 stream.avail_in = readlen;
175                 while ((stream.avail_in || readlen == 0) &&
176                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
177                         stream.next_out = obuf;
178                         stream.avail_out = sizeof(obuf);
179                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
180                         sha1write(f, obuf, stream.next_out - obuf);
181                         olen += stream.next_out - obuf;
182                 }
183                 if (stream.avail_in)
184                         die(_("deflate error (%d)"), zret);
185                 if (readlen == 0) {
186                         if (zret != Z_STREAM_END)
187                                 die(_("deflate error (%d)"), zret);
188                         break;
189                 }
190         }
191         git_deflate_end(&stream);
192         return olen;
193 }
194
195 /*
196  * we are going to reuse the existing object data as is.  make
197  * sure it is not corrupt.
198  */
199 static int check_pack_inflate(struct packed_git *p,
200                 struct pack_window **w_curs,
201                 off_t offset,
202                 off_t len,
203                 unsigned long expect)
204 {
205         git_zstream stream;
206         unsigned char fakebuf[4096], *in;
207         int st;
208
209         memset(&stream, 0, sizeof(stream));
210         git_inflate_init(&stream);
211         do {
212                 in = use_pack(p, w_curs, offset, &stream.avail_in);
213                 stream.next_in = in;
214                 stream.next_out = fakebuf;
215                 stream.avail_out = sizeof(fakebuf);
216                 st = git_inflate(&stream, Z_FINISH);
217                 offset += stream.next_in - in;
218         } while (st == Z_OK || st == Z_BUF_ERROR);
219         git_inflate_end(&stream);
220         return (st == Z_STREAM_END &&
221                 stream.total_out == expect &&
222                 stream.total_in == len) ? 0 : -1;
223 }
224
225 static void copy_pack_data(struct sha1file *f,
226                 struct packed_git *p,
227                 struct pack_window **w_curs,
228                 off_t offset,
229                 off_t len)
230 {
231         unsigned char *in;
232         unsigned long avail;
233
234         while (len) {
235                 in = use_pack(p, w_curs, offset, &avail);
236                 if (avail > len)
237                         avail = (unsigned long)len;
238                 sha1write(f, in, avail);
239                 offset += avail;
240                 len -= avail;
241         }
242 }
243
244 /* Return 0 if we will bust the pack-size limit */
245 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
246                                            unsigned long limit, int usable_delta)
247 {
248         unsigned long size, datalen;
249         unsigned char header[10], dheader[10];
250         unsigned hdrlen;
251         enum object_type type;
252         void *buf;
253         struct git_istream *st = NULL;
254
255         if (!usable_delta) {
256                 if (entry->type == OBJ_BLOB &&
257                     entry->size > big_file_threshold &&
258                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
259                         buf = NULL;
260                 else {
261                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
262                         if (!buf)
263                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
264                 }
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 (st) /* large blob case, just assume we don't compress well */
286                 datalen = size;
287         else if (entry->z_delta_size)
288                 datalen = entry->z_delta_size;
289         else
290                 datalen = do_compress(&buf, size);
291
292         /*
293          * The object header is a byte of 'type' followed by zero or
294          * more bytes of length.
295          */
296         hdrlen = encode_in_pack_object_header(type, size, header);
297
298         if (type == OBJ_OFS_DELTA) {
299                 /*
300                  * Deltas with relative base contain an additional
301                  * encoding of the relative offset for the delta
302                  * base from this object's position in the pack.
303                  */
304                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
305                 unsigned pos = sizeof(dheader) - 1;
306                 dheader[pos] = ofs & 127;
307                 while (ofs >>= 7)
308                         dheader[--pos] = 128 | (--ofs & 127);
309                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
310                         if (st)
311                                 close_istream(st);
312                         free(buf);
313                         return 0;
314                 }
315                 sha1write(f, header, hdrlen);
316                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
317                 hdrlen += sizeof(dheader) - pos;
318         } else if (type == OBJ_REF_DELTA) {
319                 /*
320                  * Deltas with a base reference contain
321                  * an additional 20 bytes for the base sha1.
322                  */
323                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
324                         if (st)
325                                 close_istream(st);
326                         free(buf);
327                         return 0;
328                 }
329                 sha1write(f, header, hdrlen);
330                 sha1write(f, entry->delta->idx.sha1, 20);
331                 hdrlen += 20;
332         } else {
333                 if (limit && hdrlen + datalen + 20 >= limit) {
334                         if (st)
335                                 close_istream(st);
336                         free(buf);
337                         return 0;
338                 }
339                 sha1write(f, header, hdrlen);
340         }
341         if (st) {
342                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
343                 close_istream(st);
344         } else {
345                 sha1write(f, buf, datalen);
346                 free(buf);
347         }
348
349         return hdrlen + datalen;
350 }
351
352 /* Return 0 if we will bust the pack-size limit */
353 static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
354                                         unsigned long limit, int usable_delta)
355 {
356         struct packed_git *p = entry->in_pack;
357         struct pack_window *w_curs = NULL;
358         struct revindex_entry *revidx;
359         off_t offset;
360         enum object_type type = entry->type;
361         unsigned long datalen;
362         unsigned char header[10], dheader[10];
363         unsigned hdrlen;
364
365         if (entry->delta)
366                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
367                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
368         hdrlen = encode_in_pack_object_header(type, entry->size, header);
369
370         offset = entry->in_pack_offset;
371         revidx = find_pack_revindex(p, offset);
372         datalen = revidx[1].offset - offset;
373         if (!pack_to_stdout && p->index_version > 1 &&
374             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
375                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
376                 unuse_pack(&w_curs);
377                 return write_no_reuse_object(f, entry, limit, usable_delta);
378         }
379
380         offset += entry->in_pack_header_size;
381         datalen -= entry->in_pack_header_size;
382
383         if (!pack_to_stdout && p->index_version == 1 &&
384             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
385                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
386                 unuse_pack(&w_curs);
387                 return write_no_reuse_object(f, entry, limit, usable_delta);
388         }
389
390         if (type == OBJ_OFS_DELTA) {
391                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
392                 unsigned pos = sizeof(dheader) - 1;
393                 dheader[pos] = ofs & 127;
394                 while (ofs >>= 7)
395                         dheader[--pos] = 128 | (--ofs & 127);
396                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
397                         unuse_pack(&w_curs);
398                         return 0;
399                 }
400                 sha1write(f, header, hdrlen);
401                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
402                 hdrlen += sizeof(dheader) - pos;
403                 reused_delta++;
404         } else if (type == OBJ_REF_DELTA) {
405                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
406                         unuse_pack(&w_curs);
407                         return 0;
408                 }
409                 sha1write(f, header, hdrlen);
410                 sha1write(f, entry->delta->idx.sha1, 20);
411                 hdrlen += 20;
412                 reused_delta++;
413         } else {
414                 if (limit && hdrlen + datalen + 20 >= limit) {
415                         unuse_pack(&w_curs);
416                         return 0;
417                 }
418                 sha1write(f, header, hdrlen);
419         }
420         copy_pack_data(f, p, &w_curs, offset, datalen);
421         unuse_pack(&w_curs);
422         reused++;
423         return hdrlen + datalen;
424 }
425
426 /* Return 0 if we will bust the pack-size limit */
427 static unsigned long write_object(struct sha1file *f,
428                                   struct object_entry *entry,
429                                   off_t write_offset)
430 {
431         unsigned long limit, len;
432         int usable_delta, to_reuse;
433
434         if (!pack_to_stdout)
435                 crc32_begin(f);
436
437         /* apply size limit if limited packsize and not first object */
438         if (!pack_size_limit || !nr_written)
439                 limit = 0;
440         else if (pack_size_limit <= write_offset)
441                 /*
442                  * the earlier object did not fit the limit; avoid
443                  * mistaking this with unlimited (i.e. limit = 0).
444                  */
445                 limit = 1;
446         else
447                 limit = pack_size_limit - write_offset;
448
449         if (!entry->delta)
450                 usable_delta = 0;       /* no delta */
451         else if (!pack_size_limit)
452                usable_delta = 1;        /* unlimited packfile */
453         else if (entry->delta->idx.offset == (off_t)-1)
454                 usable_delta = 0;       /* base was written to another pack */
455         else if (entry->delta->idx.offset)
456                 usable_delta = 1;       /* base already exists in this pack */
457         else
458                 usable_delta = 0;       /* base could end up in another pack */
459
460         if (!reuse_object)
461                 to_reuse = 0;   /* explicit */
462         else if (!entry->in_pack)
463                 to_reuse = 0;   /* can't reuse what we don't have */
464         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
465                                 /* check_object() decided it for us ... */
466                 to_reuse = usable_delta;
467                                 /* ... but pack split may override that */
468         else if (entry->type != entry->in_pack_type)
469                 to_reuse = 0;   /* pack has delta which is unusable */
470         else if (entry->delta)
471                 to_reuse = 0;   /* we want to pack afresh */
472         else
473                 to_reuse = 1;   /* we have it in-pack undeltified,
474                                  * and we do not need to deltify it.
475                                  */
476
477         if (!to_reuse)
478                 len = write_no_reuse_object(f, entry, limit, usable_delta);
479         else
480                 len = write_reuse_object(f, entry, limit, usable_delta);
481         if (!len)
482                 return 0;
483
484         if (usable_delta)
485                 written_delta++;
486         written++;
487         if (!pack_to_stdout)
488                 entry->idx.crc32 = crc32_end(f);
489         return len;
490 }
491
492 enum write_one_status {
493         WRITE_ONE_SKIP = -1, /* already written */
494         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
495         WRITE_ONE_WRITTEN = 1, /* normal */
496         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
497 };
498
499 static enum write_one_status write_one(struct sha1file *f,
500                                        struct object_entry *e,
501                                        off_t *offset)
502 {
503         unsigned long size;
504         int recursing;
505
506         /*
507          * we set offset to 1 (which is an impossible value) to mark
508          * the fact that this object is involved in "write its base
509          * first before writing a deltified object" recursion.
510          */
511         recursing = (e->idx.offset == 1);
512         if (recursing) {
513                 warning("recursive delta detected for object %s",
514                         sha1_to_hex(e->idx.sha1));
515                 return WRITE_ONE_RECURSIVE;
516         } else if (e->idx.offset || e->preferred_base) {
517                 /* offset is non zero if object is written already. */
518                 return WRITE_ONE_SKIP;
519         }
520
521         /* if we are deltified, write out base object first. */
522         if (e->delta) {
523                 e->idx.offset = 1; /* now recurse */
524                 switch (write_one(f, e->delta, offset)) {
525                 case WRITE_ONE_RECURSIVE:
526                         /* we cannot depend on this one */
527                         e->delta = NULL;
528                         break;
529                 default:
530                         break;
531                 case WRITE_ONE_BREAK:
532                         e->idx.offset = recursing;
533                         return WRITE_ONE_BREAK;
534                 }
535         }
536
537         e->idx.offset = *offset;
538         size = write_object(f, e, *offset);
539         if (!size) {
540                 e->idx.offset = recursing;
541                 return WRITE_ONE_BREAK;
542         }
543         written_list[nr_written++] = &e->idx;
544
545         /* make sure off_t is sufficiently large not to wrap */
546         if (signed_add_overflows(*offset, size))
547                 die("pack too large for current definition of off_t");
548         *offset += size;
549         return WRITE_ONE_WRITTEN;
550 }
551
552 static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
553                        void *cb_data)
554 {
555         unsigned char peeled[20];
556         struct object_entry *entry = locate_object_entry(sha1);
557
558         if (entry)
559                 entry->tagged = 1;
560         if (!peel_ref(path, peeled)) {
561                 entry = locate_object_entry(peeled);
562                 if (entry)
563                         entry->tagged = 1;
564         }
565         return 0;
566 }
567
568 static inline void add_to_write_order(struct object_entry **wo,
569                                unsigned int *endp,
570                                struct object_entry *e)
571 {
572         if (e->filled)
573                 return;
574         wo[(*endp)++] = e;
575         e->filled = 1;
576 }
577
578 static void add_descendants_to_write_order(struct object_entry **wo,
579                                            unsigned int *endp,
580                                            struct object_entry *e)
581 {
582         int add_to_order = 1;
583         while (e) {
584                 if (add_to_order) {
585                         struct object_entry *s;
586                         /* add this node... */
587                         add_to_write_order(wo, endp, e);
588                         /* all its siblings... */
589                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
590                                 add_to_write_order(wo, endp, s);
591                         }
592                 }
593                 /* drop down a level to add left subtree nodes if possible */
594                 if (e->delta_child) {
595                         add_to_order = 1;
596                         e = e->delta_child;
597                 } else {
598                         add_to_order = 0;
599                         /* our sibling might have some children, it is next */
600                         if (e->delta_sibling) {
601                                 e = e->delta_sibling;
602                                 continue;
603                         }
604                         /* go back to our parent node */
605                         e = e->delta;
606                         while (e && !e->delta_sibling) {
607                                 /* we're on the right side of a subtree, keep
608                                  * going up until we can go right again */
609                                 e = e->delta;
610                         }
611                         if (!e) {
612                                 /* done- we hit our original root node */
613                                 return;
614                         }
615                         /* pass it off to sibling at this level */
616                         e = e->delta_sibling;
617                 }
618         };
619 }
620
621 static void add_family_to_write_order(struct object_entry **wo,
622                                       unsigned int *endp,
623                                       struct object_entry *e)
624 {
625         struct object_entry *root;
626
627         for (root = e; root->delta; root = root->delta)
628                 ; /* nothing */
629         add_descendants_to_write_order(wo, endp, root);
630 }
631
632 static struct object_entry **compute_write_order(void)
633 {
634         unsigned int i, wo_end, last_untagged;
635
636         struct object_entry **wo = xmalloc(nr_objects * sizeof(*wo));
637
638         for (i = 0; i < nr_objects; i++) {
639                 objects[i].tagged = 0;
640                 objects[i].filled = 0;
641                 objects[i].delta_child = NULL;
642                 objects[i].delta_sibling = NULL;
643         }
644
645         /*
646          * Fully connect delta_child/delta_sibling network.
647          * Make sure delta_sibling is sorted in the original
648          * recency order.
649          */
650         for (i = nr_objects; i > 0;) {
651                 struct object_entry *e = &objects[--i];
652                 if (!e->delta)
653                         continue;
654                 /* Mark me as the first child */
655                 e->delta_sibling = e->delta->delta_child;
656                 e->delta->delta_child = e;
657         }
658
659         /*
660          * Mark objects that are at the tip of tags.
661          */
662         for_each_tag_ref(mark_tagged, NULL);
663
664         /*
665          * Give the objects in the original recency order until
666          * we see a tagged tip.
667          */
668         for (i = wo_end = 0; i < nr_objects; i++) {
669                 if (objects[i].tagged)
670                         break;
671                 add_to_write_order(wo, &wo_end, &objects[i]);
672         }
673         last_untagged = i;
674
675         /*
676          * Then fill all the tagged tips.
677          */
678         for (; i < nr_objects; i++) {
679                 if (objects[i].tagged)
680                         add_to_write_order(wo, &wo_end, &objects[i]);
681         }
682
683         /*
684          * And then all remaining commits and tags.
685          */
686         for (i = last_untagged; i < nr_objects; i++) {
687                 if (objects[i].type != OBJ_COMMIT &&
688                     objects[i].type != OBJ_TAG)
689                         continue;
690                 add_to_write_order(wo, &wo_end, &objects[i]);
691         }
692
693         /*
694          * And then all the trees.
695          */
696         for (i = last_untagged; i < nr_objects; i++) {
697                 if (objects[i].type != OBJ_TREE)
698                         continue;
699                 add_to_write_order(wo, &wo_end, &objects[i]);
700         }
701
702         /*
703          * Finally all the rest in really tight order
704          */
705         for (i = last_untagged; i < nr_objects; i++) {
706                 if (!objects[i].filled)
707                         add_family_to_write_order(wo, &wo_end, &objects[i]);
708         }
709
710         if (wo_end != nr_objects)
711                 die("ordered %u objects, expected %"PRIu32, wo_end, nr_objects);
712
713         return wo;
714 }
715
716 static void write_pack_file(void)
717 {
718         uint32_t i = 0, j;
719         struct sha1file *f;
720         off_t offset;
721         uint32_t nr_remaining = nr_result;
722         time_t last_mtime = 0;
723         struct object_entry **write_order;
724
725         if (progress > pack_to_stdout)
726                 progress_state = start_progress("Writing objects", nr_result);
727         written_list = xmalloc(nr_objects * sizeof(*written_list));
728         write_order = compute_write_order();
729
730         do {
731                 unsigned char sha1[20];
732                 char *pack_tmp_name = NULL;
733
734                 if (pack_to_stdout)
735                         f = sha1fd_throughput(1, "<stdout>", progress_state);
736                 else
737                         f = create_tmp_packfile(&pack_tmp_name);
738
739                 offset = write_pack_header(f, nr_remaining);
740                 if (!offset)
741                         die_errno("unable to write pack header");
742                 nr_written = 0;
743                 for (; i < nr_objects; i++) {
744                         struct object_entry *e = write_order[i];
745                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
746                                 break;
747                         display_progress(progress_state, written);
748                 }
749
750                 /*
751                  * Did we write the wrong # entries in the header?
752                  * If so, rewrite it like in fast-import
753                  */
754                 if (pack_to_stdout) {
755                         sha1close(f, sha1, CSUM_CLOSE);
756                 } else if (nr_written == nr_remaining) {
757                         sha1close(f, sha1, CSUM_FSYNC);
758                 } else {
759                         int fd = sha1close(f, sha1, 0);
760                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
761                                                  nr_written, sha1, offset);
762                         close(fd);
763                 }
764
765                 if (!pack_to_stdout) {
766                         struct stat st;
767                         char tmpname[PATH_MAX];
768
769                         /*
770                          * Packs are runtime accessed in their mtime
771                          * order since newer packs are more likely to contain
772                          * younger objects.  So if we are creating multiple
773                          * packs then we should modify the mtime of later ones
774                          * to preserve this property.
775                          */
776                         if (stat(pack_tmp_name, &st) < 0) {
777                                 warning("failed to stat %s: %s",
778                                         pack_tmp_name, strerror(errno));
779                         } else if (!last_mtime) {
780                                 last_mtime = st.st_mtime;
781                         } else {
782                                 struct utimbuf utb;
783                                 utb.actime = st.st_atime;
784                                 utb.modtime = --last_mtime;
785                                 if (utime(pack_tmp_name, &utb) < 0)
786                                         warning("failed utime() on %s: %s",
787                                                 tmpname, strerror(errno));
788                         }
789
790                         /* Enough space for "-<sha-1>.pack"? */
791                         if (sizeof(tmpname) <= strlen(base_name) + 50)
792                                 die("pack base name '%s' too long", base_name);
793                         snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
794                         finish_tmp_packfile(tmpname, pack_tmp_name,
795                                             written_list, nr_written,
796                                             &pack_idx_opts, sha1);
797                         free(pack_tmp_name);
798                         puts(sha1_to_hex(sha1));
799                 }
800
801                 /* mark written objects as written to previous pack */
802                 for (j = 0; j < nr_written; j++) {
803                         written_list[j]->offset = (off_t)-1;
804                 }
805                 nr_remaining -= nr_written;
806         } while (nr_remaining && i < nr_objects);
807
808         free(written_list);
809         free(write_order);
810         stop_progress(&progress_state);
811         if (written != nr_result)
812                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
813                         written, nr_result);
814 }
815
816 static int locate_object_entry_hash(const unsigned char *sha1)
817 {
818         int i;
819         unsigned int ui;
820         memcpy(&ui, sha1, sizeof(unsigned int));
821         i = ui % object_ix_hashsz;
822         while (0 < object_ix[i]) {
823                 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
824                         return i;
825                 if (++i == object_ix_hashsz)
826                         i = 0;
827         }
828         return -1 - i;
829 }
830
831 static struct object_entry *locate_object_entry(const unsigned char *sha1)
832 {
833         int i;
834
835         if (!object_ix_hashsz)
836                 return NULL;
837
838         i = locate_object_entry_hash(sha1);
839         if (0 <= i)
840                 return &objects[object_ix[i]-1];
841         return NULL;
842 }
843
844 static void rehash_objects(void)
845 {
846         uint32_t i;
847         struct object_entry *oe;
848
849         object_ix_hashsz = nr_objects * 3;
850         if (object_ix_hashsz < 1024)
851                 object_ix_hashsz = 1024;
852         object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
853         memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
854         for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
855                 int ix = locate_object_entry_hash(oe->idx.sha1);
856                 if (0 <= ix)
857                         continue;
858                 ix = -1 - ix;
859                 object_ix[ix] = i + 1;
860         }
861 }
862
863 static uint32_t name_hash(const char *name)
864 {
865         uint32_t c, hash = 0;
866
867         if (!name)
868                 return 0;
869
870         /*
871          * This effectively just creates a sortable number from the
872          * last sixteen non-whitespace characters. Last characters
873          * count "most", so things that end in ".c" sort together.
874          */
875         while ((c = *name++) != 0) {
876                 if (isspace(c))
877                         continue;
878                 hash = (hash >> 2) + (c << 24);
879         }
880         return hash;
881 }
882
883 static void setup_delta_attr_check(struct git_attr_check *check)
884 {
885         static struct git_attr *attr_delta;
886
887         if (!attr_delta)
888                 attr_delta = git_attr("delta");
889
890         check[0].attr = attr_delta;
891 }
892
893 static int no_try_delta(const char *path)
894 {
895         struct git_attr_check check[1];
896
897         setup_delta_attr_check(check);
898         if (git_check_attr(path, ARRAY_SIZE(check), check))
899                 return 0;
900         if (ATTR_FALSE(check->value))
901                 return 1;
902         return 0;
903 }
904
905 static int add_object_entry(const unsigned char *sha1, enum object_type type,
906                             const char *name, int exclude)
907 {
908         struct object_entry *entry;
909         struct packed_git *p, *found_pack = NULL;
910         off_t found_offset = 0;
911         int ix;
912         uint32_t hash = name_hash(name);
913
914         ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
915         if (ix >= 0) {
916                 if (exclude) {
917                         entry = objects + object_ix[ix] - 1;
918                         if (!entry->preferred_base)
919                                 nr_result--;
920                         entry->preferred_base = 1;
921                 }
922                 return 0;
923         }
924
925         if (!exclude && local && has_loose_object_nonlocal(sha1))
926                 return 0;
927
928         for (p = packed_git; p; p = p->next) {
929                 off_t offset = find_pack_entry_one(sha1, p);
930                 if (offset) {
931                         if (!found_pack) {
932                                 if (!is_pack_valid(p)) {
933                                         warning("packfile %s cannot be accessed", p->pack_name);
934                                         continue;
935                                 }
936                                 found_offset = offset;
937                                 found_pack = p;
938                         }
939                         if (exclude)
940                                 break;
941                         if (incremental)
942                                 return 0;
943                         if (local && !p->pack_local)
944                                 return 0;
945                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
946                                 return 0;
947                 }
948         }
949
950         if (nr_objects >= nr_alloc) {
951                 nr_alloc = (nr_alloc  + 1024) * 3 / 2;
952                 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
953         }
954
955         entry = objects + nr_objects++;
956         memset(entry, 0, sizeof(*entry));
957         hashcpy(entry->idx.sha1, sha1);
958         entry->hash = hash;
959         if (type)
960                 entry->type = type;
961         if (exclude)
962                 entry->preferred_base = 1;
963         else
964                 nr_result++;
965         if (found_pack) {
966                 entry->in_pack = found_pack;
967                 entry->in_pack_offset = found_offset;
968         }
969
970         if (object_ix_hashsz * 3 <= nr_objects * 4)
971                 rehash_objects();
972         else
973                 object_ix[-1 - ix] = nr_objects;
974
975         display_progress(progress_state, nr_objects);
976
977         if (name && no_try_delta(name))
978                 entry->no_try_delta = 1;
979
980         return 1;
981 }
982
983 struct pbase_tree_cache {
984         unsigned char sha1[20];
985         int ref;
986         int temporary;
987         void *tree_data;
988         unsigned long tree_size;
989 };
990
991 static struct pbase_tree_cache *(pbase_tree_cache[256]);
992 static int pbase_tree_cache_ix(const unsigned char *sha1)
993 {
994         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
995 }
996 static int pbase_tree_cache_ix_incr(int ix)
997 {
998         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
999 }
1000
1001 static struct pbase_tree {
1002         struct pbase_tree *next;
1003         /* This is a phony "cache" entry; we are not
1004          * going to evict it nor find it through _get()
1005          * mechanism -- this is for the toplevel node that
1006          * would almost always change with any commit.
1007          */
1008         struct pbase_tree_cache pcache;
1009 } *pbase_tree;
1010
1011 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1012 {
1013         struct pbase_tree_cache *ent, *nent;
1014         void *data;
1015         unsigned long size;
1016         enum object_type type;
1017         int neigh;
1018         int my_ix = pbase_tree_cache_ix(sha1);
1019         int available_ix = -1;
1020
1021         /* pbase-tree-cache acts as a limited hashtable.
1022          * your object will be found at your index or within a few
1023          * slots after that slot if it is cached.
1024          */
1025         for (neigh = 0; neigh < 8; neigh++) {
1026                 ent = pbase_tree_cache[my_ix];
1027                 if (ent && !hashcmp(ent->sha1, sha1)) {
1028                         ent->ref++;
1029                         return ent;
1030                 }
1031                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1032                          ((0 <= available_ix) &&
1033                           (!ent && pbase_tree_cache[available_ix])))
1034                         available_ix = my_ix;
1035                 if (!ent)
1036                         break;
1037                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1038         }
1039
1040         /* Did not find one.  Either we got a bogus request or
1041          * we need to read and perhaps cache.
1042          */
1043         data = read_sha1_file(sha1, &type, &size);
1044         if (!data)
1045                 return NULL;
1046         if (type != OBJ_TREE) {
1047                 free(data);
1048                 return NULL;
1049         }
1050
1051         /* We need to either cache or return a throwaway copy */
1052
1053         if (available_ix < 0)
1054                 ent = NULL;
1055         else {
1056                 ent = pbase_tree_cache[available_ix];
1057                 my_ix = available_ix;
1058         }
1059
1060         if (!ent) {
1061                 nent = xmalloc(sizeof(*nent));
1062                 nent->temporary = (available_ix < 0);
1063         }
1064         else {
1065                 /* evict and reuse */
1066                 free(ent->tree_data);
1067                 nent = ent;
1068         }
1069         hashcpy(nent->sha1, sha1);
1070         nent->tree_data = data;
1071         nent->tree_size = size;
1072         nent->ref = 1;
1073         if (!nent->temporary)
1074                 pbase_tree_cache[my_ix] = nent;
1075         return nent;
1076 }
1077
1078 static void pbase_tree_put(struct pbase_tree_cache *cache)
1079 {
1080         if (!cache->temporary) {
1081                 cache->ref--;
1082                 return;
1083         }
1084         free(cache->tree_data);
1085         free(cache);
1086 }
1087
1088 static int name_cmp_len(const char *name)
1089 {
1090         int i;
1091         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1092                 ;
1093         return i;
1094 }
1095
1096 static void add_pbase_object(struct tree_desc *tree,
1097                              const char *name,
1098                              int cmplen,
1099                              const char *fullname)
1100 {
1101         struct name_entry entry;
1102         int cmp;
1103
1104         while (tree_entry(tree,&entry)) {
1105                 if (S_ISGITLINK(entry.mode))
1106                         continue;
1107                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1108                       memcmp(name, entry.path, cmplen);
1109                 if (cmp > 0)
1110                         continue;
1111                 if (cmp < 0)
1112                         return;
1113                 if (name[cmplen] != '/') {
1114                         add_object_entry(entry.sha1,
1115                                          object_type(entry.mode),
1116                                          fullname, 1);
1117                         return;
1118                 }
1119                 if (S_ISDIR(entry.mode)) {
1120                         struct tree_desc sub;
1121                         struct pbase_tree_cache *tree;
1122                         const char *down = name+cmplen+1;
1123                         int downlen = name_cmp_len(down);
1124
1125                         tree = pbase_tree_get(entry.sha1);
1126                         if (!tree)
1127                                 return;
1128                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1129
1130                         add_pbase_object(&sub, down, downlen, fullname);
1131                         pbase_tree_put(tree);
1132                 }
1133         }
1134 }
1135
1136 static unsigned *done_pbase_paths;
1137 static int done_pbase_paths_num;
1138 static int done_pbase_paths_alloc;
1139 static int done_pbase_path_pos(unsigned hash)
1140 {
1141         int lo = 0;
1142         int hi = done_pbase_paths_num;
1143         while (lo < hi) {
1144                 int mi = (hi + lo) / 2;
1145                 if (done_pbase_paths[mi] == hash)
1146                         return mi;
1147                 if (done_pbase_paths[mi] < hash)
1148                         hi = mi;
1149                 else
1150                         lo = mi + 1;
1151         }
1152         return -lo-1;
1153 }
1154
1155 static int check_pbase_path(unsigned hash)
1156 {
1157         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1158         if (0 <= pos)
1159                 return 1;
1160         pos = -pos - 1;
1161         if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1162                 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1163                 done_pbase_paths = xrealloc(done_pbase_paths,
1164                                             done_pbase_paths_alloc *
1165                                             sizeof(unsigned));
1166         }
1167         done_pbase_paths_num++;
1168         if (pos < done_pbase_paths_num)
1169                 memmove(done_pbase_paths + pos + 1,
1170                         done_pbase_paths + pos,
1171                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1172         done_pbase_paths[pos] = hash;
1173         return 0;
1174 }
1175
1176 static void add_preferred_base_object(const char *name)
1177 {
1178         struct pbase_tree *it;
1179         int cmplen;
1180         unsigned hash = name_hash(name);
1181
1182         if (!num_preferred_base || check_pbase_path(hash))
1183                 return;
1184
1185         cmplen = name_cmp_len(name);
1186         for (it = pbase_tree; it; it = it->next) {
1187                 if (cmplen == 0) {
1188                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1189                 }
1190                 else {
1191                         struct tree_desc tree;
1192                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1193                         add_pbase_object(&tree, name, cmplen, name);
1194                 }
1195         }
1196 }
1197
1198 static void add_preferred_base(unsigned char *sha1)
1199 {
1200         struct pbase_tree *it;
1201         void *data;
1202         unsigned long size;
1203         unsigned char tree_sha1[20];
1204
1205         if (window <= num_preferred_base++)
1206                 return;
1207
1208         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1209         if (!data)
1210                 return;
1211
1212         for (it = pbase_tree; it; it = it->next) {
1213                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1214                         free(data);
1215                         return;
1216                 }
1217         }
1218
1219         it = xcalloc(1, sizeof(*it));
1220         it->next = pbase_tree;
1221         pbase_tree = it;
1222
1223         hashcpy(it->pcache.sha1, tree_sha1);
1224         it->pcache.tree_data = data;
1225         it->pcache.tree_size = size;
1226 }
1227
1228 static void cleanup_preferred_base(void)
1229 {
1230         struct pbase_tree *it;
1231         unsigned i;
1232
1233         it = pbase_tree;
1234         pbase_tree = NULL;
1235         while (it) {
1236                 struct pbase_tree *this = it;
1237                 it = this->next;
1238                 free(this->pcache.tree_data);
1239                 free(this);
1240         }
1241
1242         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1243                 if (!pbase_tree_cache[i])
1244                         continue;
1245                 free(pbase_tree_cache[i]->tree_data);
1246                 free(pbase_tree_cache[i]);
1247                 pbase_tree_cache[i] = NULL;
1248         }
1249
1250         free(done_pbase_paths);
1251         done_pbase_paths = NULL;
1252         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1253 }
1254
1255 static void check_object(struct object_entry *entry)
1256 {
1257         if (entry->in_pack) {
1258                 struct packed_git *p = entry->in_pack;
1259                 struct pack_window *w_curs = NULL;
1260                 const unsigned char *base_ref = NULL;
1261                 struct object_entry *base_entry;
1262                 unsigned long used, used_0;
1263                 unsigned long avail;
1264                 off_t ofs;
1265                 unsigned char *buf, c;
1266
1267                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1268
1269                 /*
1270                  * We want in_pack_type even if we do not reuse delta
1271                  * since non-delta representations could still be reused.
1272                  */
1273                 used = unpack_object_header_buffer(buf, avail,
1274                                                    &entry->in_pack_type,
1275                                                    &entry->size);
1276                 if (used == 0)
1277                         goto give_up;
1278
1279                 /*
1280                  * Determine if this is a delta and if so whether we can
1281                  * reuse it or not.  Otherwise let's find out as cheaply as
1282                  * possible what the actual type and size for this object is.
1283                  */
1284                 switch (entry->in_pack_type) {
1285                 default:
1286                         /* Not a delta hence we've already got all we need. */
1287                         entry->type = entry->in_pack_type;
1288                         entry->in_pack_header_size = used;
1289                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1290                                 goto give_up;
1291                         unuse_pack(&w_curs);
1292                         return;
1293                 case OBJ_REF_DELTA:
1294                         if (reuse_delta && !entry->preferred_base)
1295                                 base_ref = use_pack(p, &w_curs,
1296                                                 entry->in_pack_offset + used, NULL);
1297                         entry->in_pack_header_size = used + 20;
1298                         break;
1299                 case OBJ_OFS_DELTA:
1300                         buf = use_pack(p, &w_curs,
1301                                        entry->in_pack_offset + used, NULL);
1302                         used_0 = 0;
1303                         c = buf[used_0++];
1304                         ofs = c & 127;
1305                         while (c & 128) {
1306                                 ofs += 1;
1307                                 if (!ofs || MSB(ofs, 7)) {
1308                                         error("delta base offset overflow in pack for %s",
1309                                               sha1_to_hex(entry->idx.sha1));
1310                                         goto give_up;
1311                                 }
1312                                 c = buf[used_0++];
1313                                 ofs = (ofs << 7) + (c & 127);
1314                         }
1315                         ofs = entry->in_pack_offset - ofs;
1316                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1317                                 error("delta base offset out of bound for %s",
1318                                       sha1_to_hex(entry->idx.sha1));
1319                                 goto give_up;
1320                         }
1321                         if (reuse_delta && !entry->preferred_base) {
1322                                 struct revindex_entry *revidx;
1323                                 revidx = find_pack_revindex(p, ofs);
1324                                 if (!revidx)
1325                                         goto give_up;
1326                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1327                         }
1328                         entry->in_pack_header_size = used + used_0;
1329                         break;
1330                 }
1331
1332                 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1333                         /*
1334                          * If base_ref was set above that means we wish to
1335                          * reuse delta data, and we even found that base
1336                          * in the list of objects we want to pack. Goodie!
1337                          *
1338                          * Depth value does not matter - find_deltas() will
1339                          * never consider reused delta as the base object to
1340                          * deltify other objects against, in order to avoid
1341                          * circular deltas.
1342                          */
1343                         entry->type = entry->in_pack_type;
1344                         entry->delta = base_entry;
1345                         entry->delta_size = entry->size;
1346                         entry->delta_sibling = base_entry->delta_child;
1347                         base_entry->delta_child = entry;
1348                         unuse_pack(&w_curs);
1349                         return;
1350                 }
1351
1352                 if (entry->type) {
1353                         /*
1354                          * This must be a delta and we already know what the
1355                          * final object type is.  Let's extract the actual
1356                          * object size from the delta header.
1357                          */
1358                         entry->size = get_size_from_delta(p, &w_curs,
1359                                         entry->in_pack_offset + entry->in_pack_header_size);
1360                         if (entry->size == 0)
1361                                 goto give_up;
1362                         unuse_pack(&w_curs);
1363                         return;
1364                 }
1365
1366                 /*
1367                  * No choice but to fall back to the recursive delta walk
1368                  * with sha1_object_info() to find about the object type
1369                  * at this point...
1370                  */
1371                 give_up:
1372                 unuse_pack(&w_curs);
1373         }
1374
1375         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1376         /*
1377          * The error condition is checked in prepare_pack().  This is
1378          * to permit a missing preferred base object to be ignored
1379          * as a preferred base.  Doing so can result in a larger
1380          * pack file, but the transfer will still take place.
1381          */
1382 }
1383
1384 static int pack_offset_sort(const void *_a, const void *_b)
1385 {
1386         const struct object_entry *a = *(struct object_entry **)_a;
1387         const struct object_entry *b = *(struct object_entry **)_b;
1388
1389         /* avoid filesystem trashing with loose objects */
1390         if (!a->in_pack && !b->in_pack)
1391                 return hashcmp(a->idx.sha1, b->idx.sha1);
1392
1393         if (a->in_pack < b->in_pack)
1394                 return -1;
1395         if (a->in_pack > b->in_pack)
1396                 return 1;
1397         return a->in_pack_offset < b->in_pack_offset ? -1 :
1398                         (a->in_pack_offset > b->in_pack_offset);
1399 }
1400
1401 static void get_object_details(void)
1402 {
1403         uint32_t i;
1404         struct object_entry **sorted_by_offset;
1405
1406         sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1407         for (i = 0; i < nr_objects; i++)
1408                 sorted_by_offset[i] = objects + i;
1409         qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1410
1411         for (i = 0; i < nr_objects; i++) {
1412                 struct object_entry *entry = sorted_by_offset[i];
1413                 check_object(entry);
1414                 if (big_file_threshold < entry->size)
1415                         entry->no_try_delta = 1;
1416         }
1417
1418         free(sorted_by_offset);
1419 }
1420
1421 /*
1422  * We search for deltas in a list sorted by type, by filename hash, and then
1423  * by size, so that we see progressively smaller and smaller files.
1424  * That's because we prefer deltas to be from the bigger file
1425  * to the smaller -- deletes are potentially cheaper, but perhaps
1426  * more importantly, the bigger file is likely the more recent
1427  * one.  The deepest deltas are therefore the oldest objects which are
1428  * less susceptible to be accessed often.
1429  */
1430 static int type_size_sort(const void *_a, const void *_b)
1431 {
1432         const struct object_entry *a = *(struct object_entry **)_a;
1433         const struct object_entry *b = *(struct object_entry **)_b;
1434
1435         if (a->type > b->type)
1436                 return -1;
1437         if (a->type < b->type)
1438                 return 1;
1439         if (a->hash > b->hash)
1440                 return -1;
1441         if (a->hash < b->hash)
1442                 return 1;
1443         if (a->preferred_base > b->preferred_base)
1444                 return -1;
1445         if (a->preferred_base < b->preferred_base)
1446                 return 1;
1447         if (a->size > b->size)
1448                 return -1;
1449         if (a->size < b->size)
1450                 return 1;
1451         return a < b ? -1 : (a > b);  /* newest first */
1452 }
1453
1454 struct unpacked {
1455         struct object_entry *entry;
1456         void *data;
1457         struct delta_index *index;
1458         unsigned depth;
1459 };
1460
1461 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1462                            unsigned long delta_size)
1463 {
1464         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1465                 return 0;
1466
1467         if (delta_size < cache_max_small_delta_size)
1468                 return 1;
1469
1470         /* cache delta, if objects are large enough compared to delta size */
1471         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1472                 return 1;
1473
1474         return 0;
1475 }
1476
1477 #ifndef NO_PTHREADS
1478
1479 static pthread_mutex_t read_mutex;
1480 #define read_lock()             pthread_mutex_lock(&read_mutex)
1481 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1482
1483 static pthread_mutex_t cache_mutex;
1484 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1485 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1486
1487 static pthread_mutex_t progress_mutex;
1488 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1489 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1490
1491 #else
1492
1493 #define read_lock()             (void)0
1494 #define read_unlock()           (void)0
1495 #define cache_lock()            (void)0
1496 #define cache_unlock()          (void)0
1497 #define progress_lock()         (void)0
1498 #define progress_unlock()       (void)0
1499
1500 #endif
1501
1502 static int try_delta(struct unpacked *trg, struct unpacked *src,
1503                      unsigned max_depth, unsigned long *mem_usage)
1504 {
1505         struct object_entry *trg_entry = trg->entry;
1506         struct object_entry *src_entry = src->entry;
1507         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1508         unsigned ref_depth;
1509         enum object_type type;
1510         void *delta_buf;
1511
1512         /* Don't bother doing diffs between different types */
1513         if (trg_entry->type != src_entry->type)
1514                 return -1;
1515
1516         /*
1517          * We do not bother to try a delta that we discarded on an
1518          * earlier try, but only when reusing delta data.  Note that
1519          * src_entry that is marked as the preferred_base should always
1520          * be considered, as even if we produce a suboptimal delta against
1521          * it, we will still save the transfer cost, as we already know
1522          * the other side has it and we won't send src_entry at all.
1523          */
1524         if (reuse_delta && trg_entry->in_pack &&
1525             trg_entry->in_pack == src_entry->in_pack &&
1526             !src_entry->preferred_base &&
1527             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1528             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1529                 return 0;
1530
1531         /* Let's not bust the allowed depth. */
1532         if (src->depth >= max_depth)
1533                 return 0;
1534
1535         /* Now some size filtering heuristics. */
1536         trg_size = trg_entry->size;
1537         if (!trg_entry->delta) {
1538                 max_size = trg_size/2 - 20;
1539                 ref_depth = 1;
1540         } else {
1541                 max_size = trg_entry->delta_size;
1542                 ref_depth = trg->depth;
1543         }
1544         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1545                                                 (max_depth - ref_depth + 1);
1546         if (max_size == 0)
1547                 return 0;
1548         src_size = src_entry->size;
1549         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1550         if (sizediff >= max_size)
1551                 return 0;
1552         if (trg_size < src_size / 32)
1553                 return 0;
1554
1555         /* Load data if not already done */
1556         if (!trg->data) {
1557                 read_lock();
1558                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1559                 read_unlock();
1560                 if (!trg->data)
1561                         die("object %s cannot be read",
1562                             sha1_to_hex(trg_entry->idx.sha1));
1563                 if (sz != trg_size)
1564                         die("object %s inconsistent object length (%lu vs %lu)",
1565                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1566                 *mem_usage += sz;
1567         }
1568         if (!src->data) {
1569                 read_lock();
1570                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1571                 read_unlock();
1572                 if (!src->data) {
1573                         if (src_entry->preferred_base) {
1574                                 static int warned = 0;
1575                                 if (!warned++)
1576                                         warning("object %s cannot be read",
1577                                                 sha1_to_hex(src_entry->idx.sha1));
1578                                 /*
1579                                  * Those objects are not included in the
1580                                  * resulting pack.  Be resilient and ignore
1581                                  * them if they can't be read, in case the
1582                                  * pack could be created nevertheless.
1583                                  */
1584                                 return 0;
1585                         }
1586                         die("object %s cannot be read",
1587                             sha1_to_hex(src_entry->idx.sha1));
1588                 }
1589                 if (sz != src_size)
1590                         die("object %s inconsistent object length (%lu vs %lu)",
1591                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1592                 *mem_usage += sz;
1593         }
1594         if (!src->index) {
1595                 src->index = create_delta_index(src->data, src_size);
1596                 if (!src->index) {
1597                         static int warned = 0;
1598                         if (!warned++)
1599                                 warning("suboptimal pack - out of memory");
1600                         return 0;
1601                 }
1602                 *mem_usage += sizeof_delta_index(src->index);
1603         }
1604
1605         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1606         if (!delta_buf)
1607                 return 0;
1608
1609         if (trg_entry->delta) {
1610                 /* Prefer only shallower same-sized deltas. */
1611                 if (delta_size == trg_entry->delta_size &&
1612                     src->depth + 1 >= trg->depth) {
1613                         free(delta_buf);
1614                         return 0;
1615                 }
1616         }
1617
1618         /*
1619          * Handle memory allocation outside of the cache
1620          * accounting lock.  Compiler will optimize the strangeness
1621          * away when NO_PTHREADS is defined.
1622          */
1623         free(trg_entry->delta_data);
1624         cache_lock();
1625         if (trg_entry->delta_data) {
1626                 delta_cache_size -= trg_entry->delta_size;
1627                 trg_entry->delta_data = NULL;
1628         }
1629         if (delta_cacheable(src_size, trg_size, delta_size)) {
1630                 delta_cache_size += delta_size;
1631                 cache_unlock();
1632                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1633         } else {
1634                 cache_unlock();
1635                 free(delta_buf);
1636         }
1637
1638         trg_entry->delta = src_entry;
1639         trg_entry->delta_size = delta_size;
1640         trg->depth = src->depth + 1;
1641
1642         return 1;
1643 }
1644
1645 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1646 {
1647         struct object_entry *child = me->delta_child;
1648         unsigned int m = n;
1649         while (child) {
1650                 unsigned int c = check_delta_limit(child, n + 1);
1651                 if (m < c)
1652                         m = c;
1653                 child = child->delta_sibling;
1654         }
1655         return m;
1656 }
1657
1658 static unsigned long free_unpacked(struct unpacked *n)
1659 {
1660         unsigned long freed_mem = sizeof_delta_index(n->index);
1661         free_delta_index(n->index);
1662         n->index = NULL;
1663         if (n->data) {
1664                 freed_mem += n->entry->size;
1665                 free(n->data);
1666                 n->data = NULL;
1667         }
1668         n->entry = NULL;
1669         n->depth = 0;
1670         return freed_mem;
1671 }
1672
1673 static void find_deltas(struct object_entry **list, unsigned *list_size,
1674                         int window, int depth, unsigned *processed)
1675 {
1676         uint32_t i, idx = 0, count = 0;
1677         struct unpacked *array;
1678         unsigned long mem_usage = 0;
1679
1680         array = xcalloc(window, sizeof(struct unpacked));
1681
1682         for (;;) {
1683                 struct object_entry *entry;
1684                 struct unpacked *n = array + idx;
1685                 int j, max_depth, best_base = -1;
1686
1687                 progress_lock();
1688                 if (!*list_size) {
1689                         progress_unlock();
1690                         break;
1691                 }
1692                 entry = *list++;
1693                 (*list_size)--;
1694                 if (!entry->preferred_base) {
1695                         (*processed)++;
1696                         display_progress(progress_state, *processed);
1697                 }
1698                 progress_unlock();
1699
1700                 mem_usage -= free_unpacked(n);
1701                 n->entry = entry;
1702
1703                 while (window_memory_limit &&
1704                        mem_usage > window_memory_limit &&
1705                        count > 1) {
1706                         uint32_t tail = (idx + window - count) % window;
1707                         mem_usage -= free_unpacked(array + tail);
1708                         count--;
1709                 }
1710
1711                 /* We do not compute delta to *create* objects we are not
1712                  * going to pack.
1713                  */
1714                 if (entry->preferred_base)
1715                         goto next;
1716
1717                 /*
1718                  * If the current object is at pack edge, take the depth the
1719                  * objects that depend on the current object into account
1720                  * otherwise they would become too deep.
1721                  */
1722                 max_depth = depth;
1723                 if (entry->delta_child) {
1724                         max_depth -= check_delta_limit(entry, 0);
1725                         if (max_depth <= 0)
1726                                 goto next;
1727                 }
1728
1729                 j = window;
1730                 while (--j > 0) {
1731                         int ret;
1732                         uint32_t other_idx = idx + j;
1733                         struct unpacked *m;
1734                         if (other_idx >= window)
1735                                 other_idx -= window;
1736                         m = array + other_idx;
1737                         if (!m->entry)
1738                                 break;
1739                         ret = try_delta(n, m, max_depth, &mem_usage);
1740                         if (ret < 0)
1741                                 break;
1742                         else if (ret > 0)
1743                                 best_base = other_idx;
1744                 }
1745
1746                 /*
1747                  * If we decided to cache the delta data, then it is best
1748                  * to compress it right away.  First because we have to do
1749                  * it anyway, and doing it here while we're threaded will
1750                  * save a lot of time in the non threaded write phase,
1751                  * as well as allow for caching more deltas within
1752                  * the same cache size limit.
1753                  * ...
1754                  * But only if not writing to stdout, since in that case
1755                  * the network is most likely throttling writes anyway,
1756                  * and therefore it is best to go to the write phase ASAP
1757                  * instead, as we can afford spending more time compressing
1758                  * between writes at that moment.
1759                  */
1760                 if (entry->delta_data && !pack_to_stdout) {
1761                         entry->z_delta_size = do_compress(&entry->delta_data,
1762                                                           entry->delta_size);
1763                         cache_lock();
1764                         delta_cache_size -= entry->delta_size;
1765                         delta_cache_size += entry->z_delta_size;
1766                         cache_unlock();
1767                 }
1768
1769                 /* if we made n a delta, and if n is already at max
1770                  * depth, leaving it in the window is pointless.  we
1771                  * should evict it first.
1772                  */
1773                 if (entry->delta && max_depth <= n->depth)
1774                         continue;
1775
1776                 /*
1777                  * Move the best delta base up in the window, after the
1778                  * currently deltified object, to keep it longer.  It will
1779                  * be the first base object to be attempted next.
1780                  */
1781                 if (entry->delta) {
1782                         struct unpacked swap = array[best_base];
1783                         int dist = (window + idx - best_base) % window;
1784                         int dst = best_base;
1785                         while (dist--) {
1786                                 int src = (dst + 1) % window;
1787                                 array[dst] = array[src];
1788                                 dst = src;
1789                         }
1790                         array[dst] = swap;
1791                 }
1792
1793                 next:
1794                 idx++;
1795                 if (count + 1 < window)
1796                         count++;
1797                 if (idx >= window)
1798                         idx = 0;
1799         }
1800
1801         for (i = 0; i < window; ++i) {
1802                 free_delta_index(array[i].index);
1803                 free(array[i].data);
1804         }
1805         free(array);
1806 }
1807
1808 #ifndef NO_PTHREADS
1809
1810 static void try_to_free_from_threads(size_t size)
1811 {
1812         read_lock();
1813         release_pack_memory(size);
1814         read_unlock();
1815 }
1816
1817 static try_to_free_t old_try_to_free_routine;
1818
1819 /*
1820  * The main thread waits on the condition that (at least) one of the workers
1821  * has stopped working (which is indicated in the .working member of
1822  * struct thread_params).
1823  * When a work thread has completed its work, it sets .working to 0 and
1824  * signals the main thread and waits on the condition that .data_ready
1825  * becomes 1.
1826  */
1827
1828 struct thread_params {
1829         pthread_t thread;
1830         struct object_entry **list;
1831         unsigned list_size;
1832         unsigned remaining;
1833         int window;
1834         int depth;
1835         int working;
1836         int data_ready;
1837         pthread_mutex_t mutex;
1838         pthread_cond_t cond;
1839         unsigned *processed;
1840 };
1841
1842 static pthread_cond_t progress_cond;
1843
1844 /*
1845  * Mutex and conditional variable can't be statically-initialized on Windows.
1846  */
1847 static void init_threaded_search(void)
1848 {
1849         init_recursive_mutex(&read_mutex);
1850         pthread_mutex_init(&cache_mutex, NULL);
1851         pthread_mutex_init(&progress_mutex, NULL);
1852         pthread_cond_init(&progress_cond, NULL);
1853         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1854 }
1855
1856 static void cleanup_threaded_search(void)
1857 {
1858         set_try_to_free_routine(old_try_to_free_routine);
1859         pthread_cond_destroy(&progress_cond);
1860         pthread_mutex_destroy(&read_mutex);
1861         pthread_mutex_destroy(&cache_mutex);
1862         pthread_mutex_destroy(&progress_mutex);
1863 }
1864
1865 static void *threaded_find_deltas(void *arg)
1866 {
1867         struct thread_params *me = arg;
1868
1869         while (me->remaining) {
1870                 find_deltas(me->list, &me->remaining,
1871                             me->window, me->depth, me->processed);
1872
1873                 progress_lock();
1874                 me->working = 0;
1875                 pthread_cond_signal(&progress_cond);
1876                 progress_unlock();
1877
1878                 /*
1879                  * We must not set ->data_ready before we wait on the
1880                  * condition because the main thread may have set it to 1
1881                  * before we get here. In order to be sure that new
1882                  * work is available if we see 1 in ->data_ready, it
1883                  * was initialized to 0 before this thread was spawned
1884                  * and we reset it to 0 right away.
1885                  */
1886                 pthread_mutex_lock(&me->mutex);
1887                 while (!me->data_ready)
1888                         pthread_cond_wait(&me->cond, &me->mutex);
1889                 me->data_ready = 0;
1890                 pthread_mutex_unlock(&me->mutex);
1891         }
1892         /* leave ->working 1 so that this doesn't get more work assigned */
1893         return NULL;
1894 }
1895
1896 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1897                            int window, int depth, unsigned *processed)
1898 {
1899         struct thread_params *p;
1900         int i, ret, active_threads = 0;
1901
1902         init_threaded_search();
1903
1904         if (!delta_search_threads)      /* --threads=0 means autodetect */
1905                 delta_search_threads = online_cpus();
1906         if (delta_search_threads <= 1) {
1907                 find_deltas(list, &list_size, window, depth, processed);
1908                 cleanup_threaded_search();
1909                 return;
1910         }
1911         if (progress > pack_to_stdout)
1912                 fprintf(stderr, "Delta compression using up to %d threads.\n",
1913                                 delta_search_threads);
1914         p = xcalloc(delta_search_threads, sizeof(*p));
1915
1916         /* Partition the work amongst work threads. */
1917         for (i = 0; i < delta_search_threads; i++) {
1918                 unsigned sub_size = list_size / (delta_search_threads - i);
1919
1920                 /* don't use too small segments or no deltas will be found */
1921                 if (sub_size < 2*window && i+1 < delta_search_threads)
1922                         sub_size = 0;
1923
1924                 p[i].window = window;
1925                 p[i].depth = depth;
1926                 p[i].processed = processed;
1927                 p[i].working = 1;
1928                 p[i].data_ready = 0;
1929
1930                 /* try to split chunks on "path" boundaries */
1931                 while (sub_size && sub_size < list_size &&
1932                        list[sub_size]->hash &&
1933                        list[sub_size]->hash == list[sub_size-1]->hash)
1934                         sub_size++;
1935
1936                 p[i].list = list;
1937                 p[i].list_size = sub_size;
1938                 p[i].remaining = sub_size;
1939
1940                 list += sub_size;
1941                 list_size -= sub_size;
1942         }
1943
1944         /* Start work threads. */
1945         for (i = 0; i < delta_search_threads; i++) {
1946                 if (!p[i].list_size)
1947                         continue;
1948                 pthread_mutex_init(&p[i].mutex, NULL);
1949                 pthread_cond_init(&p[i].cond, NULL);
1950                 ret = pthread_create(&p[i].thread, NULL,
1951                                      threaded_find_deltas, &p[i]);
1952                 if (ret)
1953                         die("unable to create thread: %s", strerror(ret));
1954                 active_threads++;
1955         }
1956
1957         /*
1958          * Now let's wait for work completion.  Each time a thread is done
1959          * with its work, we steal half of the remaining work from the
1960          * thread with the largest number of unprocessed objects and give
1961          * it to that newly idle thread.  This ensure good load balancing
1962          * until the remaining object list segments are simply too short
1963          * to be worth splitting anymore.
1964          */
1965         while (active_threads) {
1966                 struct thread_params *target = NULL;
1967                 struct thread_params *victim = NULL;
1968                 unsigned sub_size = 0;
1969
1970                 progress_lock();
1971                 for (;;) {
1972                         for (i = 0; !target && i < delta_search_threads; i++)
1973                                 if (!p[i].working)
1974                                         target = &p[i];
1975                         if (target)
1976                                 break;
1977                         pthread_cond_wait(&progress_cond, &progress_mutex);
1978                 }
1979
1980                 for (i = 0; i < delta_search_threads; i++)
1981                         if (p[i].remaining > 2*window &&
1982                             (!victim || victim->remaining < p[i].remaining))
1983                                 victim = &p[i];
1984                 if (victim) {
1985                         sub_size = victim->remaining / 2;
1986                         list = victim->list + victim->list_size - sub_size;
1987                         while (sub_size && list[0]->hash &&
1988                                list[0]->hash == list[-1]->hash) {
1989                                 list++;
1990                                 sub_size--;
1991                         }
1992                         if (!sub_size) {
1993                                 /*
1994                                  * It is possible for some "paths" to have
1995                                  * so many objects that no hash boundary
1996                                  * might be found.  Let's just steal the
1997                                  * exact half in that case.
1998                                  */
1999                                 sub_size = victim->remaining / 2;
2000                                 list -= sub_size;
2001                         }
2002                         target->list = list;
2003                         victim->list_size -= sub_size;
2004                         victim->remaining -= sub_size;
2005                 }
2006                 target->list_size = sub_size;
2007                 target->remaining = sub_size;
2008                 target->working = 1;
2009                 progress_unlock();
2010
2011                 pthread_mutex_lock(&target->mutex);
2012                 target->data_ready = 1;
2013                 pthread_cond_signal(&target->cond);
2014                 pthread_mutex_unlock(&target->mutex);
2015
2016                 if (!sub_size) {
2017                         pthread_join(target->thread, NULL);
2018                         pthread_cond_destroy(&target->cond);
2019                         pthread_mutex_destroy(&target->mutex);
2020                         active_threads--;
2021                 }
2022         }
2023         cleanup_threaded_search();
2024         free(p);
2025 }
2026
2027 #else
2028 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2029 #endif
2030
2031 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2032 {
2033         unsigned char peeled[20];
2034
2035         if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
2036             !peel_ref(path, peeled)        && /* peelable? */
2037             locate_object_entry(peeled))      /* object packed? */
2038                 add_object_entry(sha1, OBJ_TAG, NULL, 0);
2039         return 0;
2040 }
2041
2042 static void prepare_pack(int window, int depth)
2043 {
2044         struct object_entry **delta_list;
2045         uint32_t i, nr_deltas;
2046         unsigned n;
2047
2048         get_object_details();
2049
2050         /*
2051          * If we're locally repacking then we need to be doubly careful
2052          * from now on in order to make sure no stealth corruption gets
2053          * propagated to the new pack.  Clients receiving streamed packs
2054          * should validate everything they get anyway so no need to incur
2055          * the additional cost here in that case.
2056          */
2057         if (!pack_to_stdout)
2058                 do_check_packed_object_crc = 1;
2059
2060         if (!nr_objects || !window || !depth)
2061                 return;
2062
2063         delta_list = xmalloc(nr_objects * sizeof(*delta_list));
2064         nr_deltas = n = 0;
2065
2066         for (i = 0; i < nr_objects; i++) {
2067                 struct object_entry *entry = objects + i;
2068
2069                 if (entry->delta)
2070                         /* This happens if we decided to reuse existing
2071                          * delta from a pack.  "reuse_delta &&" is implied.
2072                          */
2073                         continue;
2074
2075                 if (entry->size < 50)
2076                         continue;
2077
2078                 if (entry->no_try_delta)
2079                         continue;
2080
2081                 if (!entry->preferred_base) {
2082                         nr_deltas++;
2083                         if (entry->type < 0)
2084                                 die("unable to get type of object %s",
2085                                     sha1_to_hex(entry->idx.sha1));
2086                 } else {
2087                         if (entry->type < 0) {
2088                                 /*
2089                                  * This object is not found, but we
2090                                  * don't have to include it anyway.
2091                                  */
2092                                 continue;
2093                         }
2094                 }
2095
2096                 delta_list[n++] = entry;
2097         }
2098
2099         if (nr_deltas && n > 1) {
2100                 unsigned nr_done = 0;
2101                 if (progress)
2102                         progress_state = start_progress("Compressing objects",
2103                                                         nr_deltas);
2104                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2105                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2106                 stop_progress(&progress_state);
2107                 if (nr_done != nr_deltas)
2108                         die("inconsistency with delta count");
2109         }
2110         free(delta_list);
2111 }
2112
2113 static int git_pack_config(const char *k, const char *v, void *cb)
2114 {
2115         if (!strcmp(k, "pack.window")) {
2116                 window = git_config_int(k, v);
2117                 return 0;
2118         }
2119         if (!strcmp(k, "pack.windowmemory")) {
2120                 window_memory_limit = git_config_ulong(k, v);
2121                 return 0;
2122         }
2123         if (!strcmp(k, "pack.depth")) {
2124                 depth = git_config_int(k, v);
2125                 return 0;
2126         }
2127         if (!strcmp(k, "pack.compression")) {
2128                 int level = git_config_int(k, v);
2129                 if (level == -1)
2130                         level = Z_DEFAULT_COMPRESSION;
2131                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2132                         die("bad pack compression level %d", level);
2133                 pack_compression_level = level;
2134                 pack_compression_seen = 1;
2135                 return 0;
2136         }
2137         if (!strcmp(k, "pack.deltacachesize")) {
2138                 max_delta_cache_size = git_config_int(k, v);
2139                 return 0;
2140         }
2141         if (!strcmp(k, "pack.deltacachelimit")) {
2142                 cache_max_small_delta_size = git_config_int(k, v);
2143                 return 0;
2144         }
2145         if (!strcmp(k, "pack.threads")) {
2146                 delta_search_threads = git_config_int(k, v);
2147                 if (delta_search_threads < 0)
2148                         die("invalid number of threads specified (%d)",
2149                             delta_search_threads);
2150 #ifdef NO_PTHREADS
2151                 if (delta_search_threads != 1)
2152                         warning("no threads support, ignoring %s", k);
2153 #endif
2154                 return 0;
2155         }
2156         if (!strcmp(k, "pack.indexversion")) {
2157                 pack_idx_opts.version = git_config_int(k, v);
2158                 if (pack_idx_opts.version > 2)
2159                         die("bad pack.indexversion=%"PRIu32,
2160                             pack_idx_opts.version);
2161                 return 0;
2162         }
2163         return git_default_config(k, v, cb);
2164 }
2165
2166 static void read_object_list_from_stdin(void)
2167 {
2168         char line[40 + 1 + PATH_MAX + 2];
2169         unsigned char sha1[20];
2170
2171         for (;;) {
2172                 if (!fgets(line, sizeof(line), stdin)) {
2173                         if (feof(stdin))
2174                                 break;
2175                         if (!ferror(stdin))
2176                                 die("fgets returned NULL, not EOF, not error!");
2177                         if (errno != EINTR)
2178                                 die_errno("fgets");
2179                         clearerr(stdin);
2180                         continue;
2181                 }
2182                 if (line[0] == '-') {
2183                         if (get_sha1_hex(line+1, sha1))
2184                                 die("expected edge sha1, got garbage:\n %s",
2185                                     line);
2186                         add_preferred_base(sha1);
2187                         continue;
2188                 }
2189                 if (get_sha1_hex(line, sha1))
2190                         die("expected sha1, got garbage:\n %s", line);
2191
2192                 add_preferred_base_object(line+41);
2193                 add_object_entry(sha1, 0, line+41, 0);
2194         }
2195 }
2196
2197 #define OBJECT_ADDED (1u<<20)
2198
2199 static void show_commit(struct commit *commit, void *data)
2200 {
2201         add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2202         commit->object.flags |= OBJECT_ADDED;
2203 }
2204
2205 static void show_object(struct object *obj,
2206                         const struct name_path *path, const char *last,
2207                         void *data)
2208 {
2209         char *name = path_name(path, last);
2210
2211         add_preferred_base_object(name);
2212         add_object_entry(obj->sha1, obj->type, name, 0);
2213         obj->flags |= OBJECT_ADDED;
2214
2215         /*
2216          * We will have generated the hash from the name,
2217          * but not saved a pointer to it - we can free it
2218          */
2219         free((char *)name);
2220 }
2221
2222 static void show_edge(struct commit *commit)
2223 {
2224         add_preferred_base(commit->object.sha1);
2225 }
2226
2227 struct in_pack_object {
2228         off_t offset;
2229         struct object *object;
2230 };
2231
2232 struct in_pack {
2233         int alloc;
2234         int nr;
2235         struct in_pack_object *array;
2236 };
2237
2238 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2239 {
2240         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2241         in_pack->array[in_pack->nr].object = object;
2242         in_pack->nr++;
2243 }
2244
2245 /*
2246  * Compare the objects in the offset order, in order to emulate the
2247  * "git rev-list --objects" output that produced the pack originally.
2248  */
2249 static int ofscmp(const void *a_, const void *b_)
2250 {
2251         struct in_pack_object *a = (struct in_pack_object *)a_;
2252         struct in_pack_object *b = (struct in_pack_object *)b_;
2253
2254         if (a->offset < b->offset)
2255                 return -1;
2256         else if (a->offset > b->offset)
2257                 return 1;
2258         else
2259                 return hashcmp(a->object->sha1, b->object->sha1);
2260 }
2261
2262 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2263 {
2264         struct packed_git *p;
2265         struct in_pack in_pack;
2266         uint32_t i;
2267
2268         memset(&in_pack, 0, sizeof(in_pack));
2269
2270         for (p = packed_git; p; p = p->next) {
2271                 const unsigned char *sha1;
2272                 struct object *o;
2273
2274                 if (!p->pack_local || p->pack_keep)
2275                         continue;
2276                 if (open_pack_index(p))
2277                         die("cannot open pack index");
2278
2279                 ALLOC_GROW(in_pack.array,
2280                            in_pack.nr + p->num_objects,
2281                            in_pack.alloc);
2282
2283                 for (i = 0; i < p->num_objects; i++) {
2284                         sha1 = nth_packed_object_sha1(p, i);
2285                         o = lookup_unknown_object(sha1);
2286                         if (!(o->flags & OBJECT_ADDED))
2287                                 mark_in_pack_object(o, p, &in_pack);
2288                         o->flags |= OBJECT_ADDED;
2289                 }
2290         }
2291
2292         if (in_pack.nr) {
2293                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2294                       ofscmp);
2295                 for (i = 0; i < in_pack.nr; i++) {
2296                         struct object *o = in_pack.array[i].object;
2297                         add_object_entry(o->sha1, o->type, "", 0);
2298                 }
2299         }
2300         free(in_pack.array);
2301 }
2302
2303 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2304 {
2305         static struct packed_git *last_found = (void *)1;
2306         struct packed_git *p;
2307
2308         p = (last_found != (void *)1) ? last_found : packed_git;
2309
2310         while (p) {
2311                 if ((!p->pack_local || p->pack_keep) &&
2312                         find_pack_entry_one(sha1, p)) {
2313                         last_found = p;
2314                         return 1;
2315                 }
2316                 if (p == last_found)
2317                         p = packed_git;
2318                 else
2319                         p = p->next;
2320                 if (p == last_found)
2321                         p = p->next;
2322         }
2323         return 0;
2324 }
2325
2326 static void loosen_unused_packed_objects(struct rev_info *revs)
2327 {
2328         struct packed_git *p;
2329         uint32_t i;
2330         const unsigned char *sha1;
2331
2332         for (p = packed_git; p; p = p->next) {
2333                 if (!p->pack_local || p->pack_keep)
2334                         continue;
2335
2336                 if (unpack_unreachable_expiration &&
2337                     p->mtime < unpack_unreachable_expiration)
2338                         continue;
2339
2340                 if (open_pack_index(p))
2341                         die("cannot open pack index");
2342
2343                 for (i = 0; i < p->num_objects; i++) {
2344                         sha1 = nth_packed_object_sha1(p, i);
2345                         if (!locate_object_entry(sha1) &&
2346                                 !has_sha1_pack_kept_or_nonlocal(sha1))
2347                                 if (force_object_loose(sha1, p->mtime))
2348                                         die("unable to force loose object");
2349                 }
2350         }
2351 }
2352
2353 static void get_object_list(int ac, const char **av)
2354 {
2355         struct rev_info revs;
2356         char line[1000];
2357         int flags = 0;
2358
2359         init_revisions(&revs, NULL);
2360         save_commit_buffer = 0;
2361         setup_revisions(ac, av, &revs, NULL);
2362
2363         while (fgets(line, sizeof(line), stdin) != NULL) {
2364                 int len = strlen(line);
2365                 if (len && line[len - 1] == '\n')
2366                         line[--len] = 0;
2367                 if (!len)
2368                         break;
2369                 if (*line == '-') {
2370                         if (!strcmp(line, "--not")) {
2371                                 flags ^= UNINTERESTING;
2372                                 continue;
2373                         }
2374                         die("not a rev '%s'", line);
2375                 }
2376                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2377                         die("bad revision '%s'", line);
2378         }
2379
2380         if (prepare_revision_walk(&revs))
2381                 die("revision walk setup failed");
2382         mark_edges_uninteresting(&revs, show_edge);
2383         traverse_commit_list(&revs, show_commit, show_object, NULL);
2384
2385         if (keep_unreachable)
2386                 add_objects_in_unpacked_packs(&revs);
2387         if (unpack_unreachable)
2388                 loosen_unused_packed_objects(&revs);
2389 }
2390
2391 static int option_parse_index_version(const struct option *opt,
2392                                       const char *arg, int unset)
2393 {
2394         char *c;
2395         const char *val = arg;
2396         pack_idx_opts.version = strtoul(val, &c, 10);
2397         if (pack_idx_opts.version > 2)
2398                 die(_("unsupported index version %s"), val);
2399         if (*c == ',' && c[1])
2400                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2401         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2402                 die(_("bad index version '%s'"), val);
2403         return 0;
2404 }
2405
2406 static int option_parse_unpack_unreachable(const struct option *opt,
2407                                            const char *arg, int unset)
2408 {
2409         if (unset) {
2410                 unpack_unreachable = 0;
2411                 unpack_unreachable_expiration = 0;
2412         }
2413         else {
2414                 unpack_unreachable = 1;
2415                 if (arg)
2416                         unpack_unreachable_expiration = approxidate(arg);
2417         }
2418         return 0;
2419 }
2420
2421 static int option_parse_ulong(const struct option *opt,
2422                               const char *arg, int unset)
2423 {
2424         if (unset)
2425                 die(_("option %s does not accept negative form"),
2426                     opt->long_name);
2427
2428         if (!git_parse_ulong(arg, opt->value))
2429                 die(_("unable to parse value '%s' for option %s"),
2430                     arg, opt->long_name);
2431         return 0;
2432 }
2433
2434 #define OPT_ULONG(s, l, v, h) \
2435         { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2436           PARSE_OPT_NONEG, option_parse_ulong }
2437
2438 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2439 {
2440         int use_internal_rev_list = 0;
2441         int thin = 0;
2442         int all_progress_implied = 0;
2443         const char *rp_av[6];
2444         int rp_ac = 0;
2445         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2446         struct option pack_objects_options[] = {
2447                 OPT_SET_INT('q', "quiet", &progress,
2448                             N_("do not show progress meter"), 0),
2449                 OPT_SET_INT(0, "progress", &progress,
2450                             N_("show progress meter"), 1),
2451                 OPT_SET_INT(0, "all-progress", &progress,
2452                             N_("show progress meter during object writing phase"), 2),
2453                 OPT_BOOL(0, "all-progress-implied",
2454                          &all_progress_implied,
2455                          N_("similar to --all-progress when progress meter is shown")),
2456                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2457                   N_("write the pack index file in the specified idx format version"),
2458                   0, option_parse_index_version },
2459                 OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2460                           N_("maximum size of each output pack file")),
2461                 OPT_BOOL(0, "local", &local,
2462                          N_("ignore borrowed objects from alternate object store")),
2463                 OPT_BOOL(0, "incremental", &incremental,
2464                          N_("ignore packed objects")),
2465                 OPT_INTEGER(0, "window", &window,
2466                             N_("limit pack window by objects")),
2467                 OPT_ULONG(0, "window-memory", &window_memory_limit,
2468                           N_("limit pack window by memory in addition to object limit")),
2469                 OPT_INTEGER(0, "depth", &depth,
2470                             N_("maximum length of delta chain allowed in the resulting pack")),
2471                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2472                          N_("reuse existing deltas")),
2473                 OPT_BOOL(0, "reuse-object", &reuse_object,
2474                          N_("reuse existing objects")),
2475                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2476                          N_("use OFS_DELTA objects")),
2477                 OPT_INTEGER(0, "threads", &delta_search_threads,
2478                             N_("use threads when searching for best delta matches")),
2479                 OPT_BOOL(0, "non-empty", &non_empty,
2480                          N_("do not create an empty pack output")),
2481                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2482                          N_("read revision arguments from standard input")),
2483                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2484                   N_("limit the objects to those that are not yet packed"),
2485                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2486                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2487                   N_("include objects reachable from any reference"),
2488                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2489                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2490                   N_("include objects referred by reflog entries"),
2491                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2492                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2493                          N_("output pack to stdout")),
2494                 OPT_BOOL(0, "include-tag", &include_tag,
2495                          N_("include tag objects that refer to objects to be packed")),
2496                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2497                          N_("keep unreachable objects")),
2498                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2499                   N_("unpack unreachable objects newer than <time>"),
2500                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2501                 OPT_BOOL(0, "thin", &thin,
2502                          N_("create thin packs")),
2503                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2504                          N_("ignore packs that have companion .keep file")),
2505                 OPT_INTEGER(0, "compression", &pack_compression_level,
2506                             N_("pack compression level")),
2507                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2508                             N_("do not hide commits by grafts"), 0),
2509                 OPT_END(),
2510         };
2511
2512         read_replace_refs = 0;
2513
2514         reset_pack_idx_option(&pack_idx_opts);
2515         git_config(git_pack_config, NULL);
2516         if (!pack_compression_seen && core_compression_seen)
2517                 pack_compression_level = core_compression_level;
2518
2519         progress = isatty(2);
2520         argc = parse_options(argc, argv, prefix, pack_objects_options,
2521                              pack_usage, 0);
2522
2523         if (argc) {
2524                 base_name = argv[0];
2525                 argc--;
2526         }
2527         if (pack_to_stdout != !base_name || argc)
2528                 usage_with_options(pack_usage, pack_objects_options);
2529
2530         rp_av[rp_ac++] = "pack-objects";
2531         if (thin) {
2532                 use_internal_rev_list = 1;
2533                 rp_av[rp_ac++] = "--objects-edge";
2534         } else
2535                 rp_av[rp_ac++] = "--objects";
2536
2537         if (rev_list_all) {
2538                 use_internal_rev_list = 1;
2539                 rp_av[rp_ac++] = "--all";
2540         }
2541         if (rev_list_reflog) {
2542                 use_internal_rev_list = 1;
2543                 rp_av[rp_ac++] = "--reflog";
2544         }
2545         if (rev_list_unpacked) {
2546                 use_internal_rev_list = 1;
2547                 rp_av[rp_ac++] = "--unpacked";
2548         }
2549
2550         if (!reuse_object)
2551                 reuse_delta = 0;
2552         if (pack_compression_level == -1)
2553                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2554         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2555                 die("bad pack compression level %d", pack_compression_level);
2556 #ifdef NO_PTHREADS
2557         if (delta_search_threads != 1)
2558                 warning("no threads support, ignoring --threads");
2559 #endif
2560         if (!pack_to_stdout && !pack_size_limit)
2561                 pack_size_limit = pack_size_limit_cfg;
2562         if (pack_to_stdout && pack_size_limit)
2563                 die("--max-pack-size cannot be used to build a pack for transfer.");
2564         if (pack_size_limit && pack_size_limit < 1024*1024) {
2565                 warning("minimum pack size limit is 1 MiB");
2566                 pack_size_limit = 1024*1024;
2567         }
2568
2569         if (!pack_to_stdout && thin)
2570                 die("--thin cannot be used to build an indexable pack.");
2571
2572         if (keep_unreachable && unpack_unreachable)
2573                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2574
2575         if (progress && all_progress_implied)
2576                 progress = 2;
2577
2578         prepare_packed_git();
2579
2580         if (progress)
2581                 progress_state = start_progress("Counting objects", 0);
2582         if (!use_internal_rev_list)
2583                 read_object_list_from_stdin();
2584         else {
2585                 rp_av[rp_ac] = NULL;
2586                 get_object_list(rp_ac, rp_av);
2587         }
2588         cleanup_preferred_base();
2589         if (include_tag && nr_result)
2590                 for_each_ref(add_ref_tag, NULL);
2591         stop_progress(&progress_state);
2592
2593         if (non_empty && !nr_result)
2594                 return 0;
2595         if (nr_result)
2596                 prepare_pack(window, depth);
2597         write_pack_file();
2598         if (progress)
2599                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2600                         " reused %"PRIu32" (delta %"PRIu32")\n",
2601                         written, written_delta, reused, reused_delta);
2602         return 0;
2603 }