Merge branches 'jc/rev-list' and 'jc/pack-thin'
[git] / pack-objects.c
1 #include "cache.h"
2 #include "object.h"
3 #include "delta.h"
4 #include "pack.h"
5 #include "csum-file.h"
6 #include "diff.h"
7 #include <sys/time.h>
8 #include <signal.h>
9
10 static const char pack_usage[] = "git-pack-objects [-q] [--no-reuse-delta] [--non-empty] [--local] [--incremental] [--window=N] [--depth=N] {--stdout | base-name} < object-list";
11
12 struct object_entry {
13         unsigned char sha1[20];
14         unsigned long size;     /* uncompressed size */
15         unsigned long offset;   /* offset into the final pack file;
16                                  * nonzero if already written.
17                                  */
18         unsigned int depth;     /* delta depth */
19         unsigned int delta_limit;       /* base adjustment for in-pack delta */
20         unsigned int hash;      /* name hint hash */
21         enum object_type type;
22         enum object_type in_pack_type;  /* could be delta */
23         unsigned long delta_size;       /* delta data size (uncompressed) */
24         struct object_entry *delta;     /* delta base object */
25         struct packed_git *in_pack;     /* already in pack */
26         unsigned int in_pack_offset;
27         struct object_entry *delta_child; /* delitified objects who bases me */
28         struct object_entry *delta_sibling; /* other deltified objects who
29                                              * uses the same base as me
30                                              */
31         int preferred_base;     /* we do not pack this, but is encouraged to
32                                  * be used as the base objectto delta huge
33                                  * objects against.
34                                  */
35         int based_on_preferred; /* current delta candidate is a preferred
36                                  * one, or delta against a preferred one.
37                                  */
38 };
39
40 /*
41  * Objects we are going to pack are colected in objects array (dynamically
42  * expanded).  nr_objects & nr_alloc controls this array.  They are stored
43  * in the order we see -- typically rev-list --objects order that gives us
44  * nice "minimum seek" order.
45  *
46  * sorted-by-sha ans sorted-by-type are arrays of pointers that point at
47  * elements in the objects array.  The former is used to build the pack
48  * index (lists object names in the ascending order to help offset lookup),
49  * and the latter is used to group similar things together by try_delta()
50  * heuristics.
51  */
52
53 static unsigned char object_list_sha1[20];
54 static int non_empty = 0;
55 static int no_reuse_delta = 0;
56 static int local = 0;
57 static int incremental = 0;
58 static struct object_entry **sorted_by_sha, **sorted_by_type;
59 static struct object_entry *objects = NULL;
60 static int nr_objects = 0, nr_alloc = 0, nr_result = 0;
61 static const char *base_name;
62 static unsigned char pack_file_sha1[20];
63 static int progress = 1;
64 static volatile int progress_update = 0;
65
66 /*
67  * The object names in objects array are hashed with this hashtable,
68  * to help looking up the entry by object name.  Binary search from
69  * sorted_by_sha is also possible but this was easier to code and faster.
70  * This hashtable is built after all the objects are seen.
71  */
72 static int *object_ix = NULL;
73 static int object_ix_hashsz = 0;
74
75 /*
76  * Pack index for existing packs give us easy access to the offsets into
77  * corresponding pack file where each object's data starts, but the entries
78  * do not store the size of the compressed representation (uncompressed
79  * size is easily available by examining the pack entry header).  We build
80  * a hashtable of existing packs (pack_revindex), and keep reverse index
81  * here -- pack index file is sorted by object name mapping to offset; this
82  * pack_revindex[].revindex array is an ordered list of offsets, so if you
83  * know the offset of an object, next offset is where its packed
84  * representation ends.
85  */
86 struct pack_revindex {
87         struct packed_git *p;
88         unsigned long *revindex;
89 } *pack_revindex = NULL;
90 static int pack_revindex_hashsz = 0;
91
92 /*
93  * stats
94  */
95 static int written = 0;
96 static int written_delta = 0;
97 static int reused = 0;
98 static int reused_delta = 0;
99
100 static int pack_revindex_ix(struct packed_git *p)
101 {
102         unsigned int ui = (unsigned int) p;
103         int i;
104
105         ui = ui ^ (ui >> 16); /* defeat structure alignment */
106         i = (int)(ui % pack_revindex_hashsz);
107         while (pack_revindex[i].p) {
108                 if (pack_revindex[i].p == p)
109                         return i;
110                 if (++i == pack_revindex_hashsz)
111                         i = 0;
112         }
113         return -1 - i;
114 }
115
116 static void prepare_pack_ix(void)
117 {
118         int num;
119         struct packed_git *p;
120         for (num = 0, p = packed_git; p; p = p->next)
121                 num++;
122         if (!num)
123                 return;
124         pack_revindex_hashsz = num * 11;
125         pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
126         for (p = packed_git; p; p = p->next) {
127                 num = pack_revindex_ix(p);
128                 num = - 1 - num;
129                 pack_revindex[num].p = p;
130         }
131         /* revindex elements are lazily initialized */
132 }
133
134 static int cmp_offset(const void *a_, const void *b_)
135 {
136         unsigned long a = *(unsigned long *) a_;
137         unsigned long b = *(unsigned long *) b_;
138         if (a < b)
139                 return -1;
140         else if (a == b)
141                 return 0;
142         else
143                 return 1;
144 }
145
146 /*
147  * Ordered list of offsets of objects in the pack.
148  */
149 static void prepare_pack_revindex(struct pack_revindex *rix)
150 {
151         struct packed_git *p = rix->p;
152         int num_ent = num_packed_objects(p);
153         int i;
154         void *index = p->index_base + 256;
155
156         rix->revindex = xmalloc(sizeof(unsigned long) * (num_ent + 1));
157         for (i = 0; i < num_ent; i++) {
158                 long hl = *((long *)(index + 24 * i));
159                 rix->revindex[i] = ntohl(hl);
160         }
161         /* This knows the pack format -- the 20-byte trailer
162          * follows immediately after the last object data.
163          */
164         rix->revindex[num_ent] = p->pack_size - 20;
165         qsort(rix->revindex, num_ent, sizeof(unsigned long), cmp_offset);
166 }
167
168 static unsigned long find_packed_object_size(struct packed_git *p,
169                                              unsigned long ofs)
170 {
171         int num;
172         int lo, hi;
173         struct pack_revindex *rix;
174         unsigned long *revindex;
175         num = pack_revindex_ix(p);
176         if (num < 0)
177                 die("internal error: pack revindex uninitialized");
178         rix = &pack_revindex[num];
179         if (!rix->revindex)
180                 prepare_pack_revindex(rix);
181         revindex = rix->revindex;
182         lo = 0;
183         hi = num_packed_objects(p) + 1;
184         do {
185                 int mi = (lo + hi) / 2;
186                 if (revindex[mi] == ofs) {
187                         return revindex[mi+1] - ofs;
188                 }
189                 else if (ofs < revindex[mi])
190                         hi = mi;
191                 else
192                         lo = mi + 1;
193         } while (lo < hi);
194         die("internal error: pack revindex corrupt");
195 }
196
197 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
198 {
199         unsigned long othersize, delta_size;
200         char type[10];
201         void *otherbuf = read_sha1_file(entry->delta->sha1, type, &othersize);
202         void *delta_buf;
203
204         if (!otherbuf)
205                 die("unable to read %s", sha1_to_hex(entry->delta->sha1));
206         delta_buf = diff_delta(otherbuf, othersize,
207                                buf, size, &delta_size, 0);
208         if (!delta_buf || delta_size != entry->delta_size)
209                 die("delta size changed");
210         free(buf);
211         free(otherbuf);
212         return delta_buf;
213 }
214
215 /*
216  * The per-object header is a pretty dense thing, which is
217  *  - first byte: low four bits are "size", then three bits of "type",
218  *    and the high bit is "size continues".
219  *  - each byte afterwards: low seven bits are size continuation,
220  *    with the high bit being "size continues"
221  */
222 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
223 {
224         int n = 1;
225         unsigned char c;
226
227         if (type < OBJ_COMMIT || type > OBJ_DELTA)
228                 die("bad type %d", type);
229
230         c = (type << 4) | (size & 15);
231         size >>= 4;
232         while (size) {
233                 *hdr++ = c | 0x80;
234                 c = size & 0x7f;
235                 size >>= 7;
236                 n++;
237         }
238         *hdr = c;
239         return n;
240 }
241
242 static unsigned long write_object(struct sha1file *f,
243                                   struct object_entry *entry)
244 {
245         unsigned long size;
246         char type[10];
247         void *buf;
248         unsigned char header[10];
249         unsigned hdrlen, datalen;
250         enum object_type obj_type;
251         int to_reuse = 0;
252
253         if (entry->preferred_base)
254                 return 0;
255
256         obj_type = entry->type;
257         if (! entry->in_pack)
258                 to_reuse = 0;   /* can't reuse what we don't have */
259         else if (obj_type == OBJ_DELTA)
260                 to_reuse = 1;   /* check_object() decided it for us */
261         else if (obj_type != entry->in_pack_type)
262                 to_reuse = 0;   /* pack has delta which is unusable */
263         else if (entry->delta)
264                 to_reuse = 0;   /* we want to pack afresh */
265         else
266                 to_reuse = 1;   /* we have it in-pack undeltified,
267                                  * and we do not need to deltify it.
268                                  */
269
270         if (! to_reuse) {
271                 buf = read_sha1_file(entry->sha1, type, &size);
272                 if (!buf)
273                         die("unable to read %s", sha1_to_hex(entry->sha1));
274                 if (size != entry->size)
275                         die("object %s size inconsistency (%lu vs %lu)",
276                             sha1_to_hex(entry->sha1), size, entry->size);
277                 if (entry->delta) {
278                         buf = delta_against(buf, size, entry);
279                         size = entry->delta_size;
280                         obj_type = OBJ_DELTA;
281                 }
282                 /*
283                  * The object header is a byte of 'type' followed by zero or
284                  * more bytes of length.  For deltas, the 20 bytes of delta
285                  * sha1 follows that.
286                  */
287                 hdrlen = encode_header(obj_type, size, header);
288                 sha1write(f, header, hdrlen);
289
290                 if (entry->delta) {
291                         sha1write(f, entry->delta, 20);
292                         hdrlen += 20;
293                 }
294                 datalen = sha1write_compressed(f, buf, size);
295                 free(buf);
296         }
297         else {
298                 struct packed_git *p = entry->in_pack;
299                 use_packed_git(p);
300
301                 datalen = find_packed_object_size(p, entry->in_pack_offset);
302                 buf = p->pack_base + entry->in_pack_offset;
303                 sha1write(f, buf, datalen);
304                 unuse_packed_git(p);
305                 hdrlen = 0; /* not really */
306                 if (obj_type == OBJ_DELTA)
307                         reused_delta++;
308                 reused++;
309         }
310         if (obj_type == OBJ_DELTA)
311                 written_delta++;
312         written++;
313         return hdrlen + datalen;
314 }
315
316 static unsigned long write_one(struct sha1file *f,
317                                struct object_entry *e,
318                                unsigned long offset)
319 {
320         if (e->offset)
321                 /* offset starts from header size and cannot be zero
322                  * if it is written already.
323                  */
324                 return offset;
325         e->offset = offset;
326         offset += write_object(f, e);
327         /* if we are deltified, write out its base object. */
328         if (e->delta)
329                 offset = write_one(f, e->delta, offset);
330         return offset;
331 }
332
333 static void write_pack_file(void)
334 {
335         int i;
336         struct sha1file *f;
337         unsigned long offset;
338         struct pack_header hdr;
339         unsigned last_percent = 999;
340         int do_progress = 0;
341
342         if (!base_name)
343                 f = sha1fd(1, "<stdout>");
344         else {
345                 f = sha1create("%s-%s.%s", base_name,
346                                sha1_to_hex(object_list_sha1), "pack");
347                 do_progress = progress;
348         }
349         if (do_progress)
350                 fprintf(stderr, "Writing %d objects.\n", nr_result);
351
352         hdr.hdr_signature = htonl(PACK_SIGNATURE);
353         hdr.hdr_version = htonl(PACK_VERSION);
354         hdr.hdr_entries = htonl(nr_result);
355         sha1write(f, &hdr, sizeof(hdr));
356         offset = sizeof(hdr);
357         if (!nr_result)
358                 goto done;
359         for (i = 0; i < nr_objects; i++) {
360                 offset = write_one(f, objects + i, offset);
361                 if (do_progress) {
362                         unsigned percent = written * 100 / nr_result;
363                         if (progress_update || percent != last_percent) {
364                                 fprintf(stderr, "%4u%% (%u/%u) done\r",
365                                         percent, written, nr_result);
366                                 progress_update = 0;
367                                 last_percent = percent;
368                         }
369                 }
370         }
371         if (do_progress)
372                 fputc('\n', stderr);
373  done:
374         sha1close(f, pack_file_sha1, 1);
375 }
376
377 static void write_index_file(void)
378 {
379         int i;
380         struct sha1file *f = sha1create("%s-%s.%s", base_name,
381                                         sha1_to_hex(object_list_sha1), "idx");
382         struct object_entry **list = sorted_by_sha;
383         struct object_entry **last = list + nr_result;
384         unsigned int array[256];
385
386         /*
387          * Write the first-level table (the list is sorted,
388          * but we use a 256-entry lookup to be able to avoid
389          * having to do eight extra binary search iterations).
390          */
391         for (i = 0; i < 256; i++) {
392                 struct object_entry **next = list;
393                 while (next < last) {
394                         struct object_entry *entry = *next;
395                         if (entry->sha1[0] != i)
396                                 break;
397                         next++;
398                 }
399                 array[i] = htonl(next - sorted_by_sha);
400                 list = next;
401         }
402         sha1write(f, array, 256 * sizeof(int));
403
404         /*
405          * Write the actual SHA1 entries..
406          */
407         list = sorted_by_sha;
408         for (i = 0; i < nr_result; i++) {
409                 struct object_entry *entry = *list++;
410                 unsigned int offset = htonl(entry->offset);
411                 sha1write(f, &offset, 4);
412                 sha1write(f, entry->sha1, 20);
413         }
414         sha1write(f, pack_file_sha1, 20);
415         sha1close(f, NULL, 1);
416 }
417
418 static int locate_object_entry_hash(const unsigned char *sha1)
419 {
420         int i;
421         unsigned int ui;
422         memcpy(&ui, sha1, sizeof(unsigned int));
423         i = ui % object_ix_hashsz;
424         while (0 < object_ix[i]) {
425                 if (!memcmp(sha1, objects[object_ix[i]-1].sha1, 20))
426                         return i;
427                 if (++i == object_ix_hashsz)
428                         i = 0;
429         }
430         return -1 - i;
431 }
432
433 static struct object_entry *locate_object_entry(const unsigned char *sha1)
434 {
435         int i;
436
437         if (!object_ix_hashsz)
438                 return NULL;
439
440         i = locate_object_entry_hash(sha1);
441         if (0 <= i)
442                 return &objects[object_ix[i]-1];
443         return NULL;
444 }
445
446 static void rehash_objects(void)
447 {
448         int i;
449         struct object_entry *oe;
450
451         object_ix_hashsz = nr_objects * 3;
452         if (object_ix_hashsz < 1024)
453                 object_ix_hashsz = 1024;
454         object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
455         object_ix = memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
456         for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
457                 int ix = locate_object_entry_hash(oe->sha1);
458                 if (0 <= ix)
459                         continue;
460                 ix = -1 - ix;
461                 object_ix[ix] = i + 1;
462         }
463 }
464
465 struct name_path {
466         struct name_path *up;
467         const char *elem;
468         int len;
469 };
470
471 #define DIRBITS 12
472
473 static unsigned name_hash(struct name_path *path, const char *name)
474 {
475         struct name_path *p = path;
476         const char *n = name + strlen(name);
477         unsigned hash = 0, name_hash = 0, name_done = 0;
478
479         if (n != name && n[-1] == '\n')
480                 n--;
481         while (name <= --n) {
482                 unsigned char c = *n;
483                 if (c == '/' && !name_done) {
484                         name_hash = hash;
485                         name_done = 1;
486                         hash = 0;
487                 }
488                 hash = hash * 11 + c;
489         }
490         if (!name_done) {
491                 name_hash = hash;
492                 hash = 0;
493         }
494         for (p = path; p; p = p->up) {
495                 hash = hash * 11 + '/';
496                 n = p->elem + p->len;
497                 while (p->elem <= --n) {
498                         unsigned char c = *n;
499                         hash = hash * 11 + c;
500                 }
501         }
502         /*
503          * Make sure "Makefile" and "t/Makefile" are hashed separately
504          * but close enough.
505          */
506         hash = (name_hash<<DIRBITS) | (hash & ((1U<<DIRBITS )-1));
507
508         if (0) { /* debug */
509                 n = name + strlen(name);
510                 if (n != name && n[-1] == '\n')
511                         n--;
512                 while (name <= --n)
513                         fputc(*n, stderr);
514                 for (p = path; p; p = p->up) {
515                         fputc('/', stderr);
516                         n = p->elem + p->len;
517                         while (p->elem <= --n)
518                                 fputc(*n, stderr);
519                 }
520                 fprintf(stderr, "\t%08x\n", hash);
521         }
522         return hash;
523 }
524
525 static int add_object_entry(const unsigned char *sha1, unsigned hash, int exclude)
526 {
527         unsigned int idx = nr_objects;
528         struct object_entry *entry;
529         struct packed_git *p;
530         unsigned int found_offset = 0;
531         struct packed_git *found_pack = NULL;
532         int ix, status = 0;
533
534         if (!exclude) {
535                 for (p = packed_git; p; p = p->next) {
536                         struct pack_entry e;
537                         if (find_pack_entry_one(sha1, &e, p)) {
538                                 if (incremental)
539                                         return 0;
540                                 if (local && !p->pack_local)
541                                         return 0;
542                                 if (!found_pack) {
543                                         found_offset = e.offset;
544                                         found_pack = e.p;
545                                 }
546                         }
547                 }
548         }
549         if ((entry = locate_object_entry(sha1)) != NULL)
550                 goto already_added;
551
552         if (idx >= nr_alloc) {
553                 unsigned int needed = (idx + 1024) * 3 / 2;
554                 objects = xrealloc(objects, needed * sizeof(*entry));
555                 nr_alloc = needed;
556         }
557         entry = objects + idx;
558         nr_objects = idx + 1;
559         memset(entry, 0, sizeof(*entry));
560         memcpy(entry->sha1, sha1, 20);
561         entry->hash = hash;
562
563         if (object_ix_hashsz * 3 <= nr_objects * 4)
564                 rehash_objects();
565         else {
566                 ix = locate_object_entry_hash(entry->sha1);
567                 if (0 <= ix)
568                         die("internal error in object hashing.");
569                 object_ix[-1 - ix] = idx + 1;
570         }
571         status = 1;
572
573  already_added:
574         if (progress_update) {
575                 fprintf(stderr, "Counting objects...%d\r", nr_objects);
576                 progress_update = 0;
577         }
578         if (exclude)
579                 entry->preferred_base = 1;
580         else {
581                 if (found_pack) {
582                         entry->in_pack = found_pack;
583                         entry->in_pack_offset = found_offset;
584                 }
585         }
586         return status;
587 }
588
589 static void add_pbase_tree(struct tree_desc *tree, struct name_path *up)
590 {
591         while (tree->size) {
592                 const unsigned char *sha1;
593                 const char *name;
594                 unsigned mode, hash;
595                 unsigned long size;
596                 char type[20];
597
598                 sha1 = tree_entry_extract(tree, &name, &mode);
599                 update_tree_entry(tree);
600                 if (!has_sha1_file(sha1))
601                         continue;
602                 if (sha1_object_info(sha1, type, &size))
603                         continue;
604
605                 hash = name_hash(up, name);
606                 if (!add_object_entry(sha1, hash, 1))
607                         continue;
608
609                 if (!strcmp(type, "tree")) {
610                         struct tree_desc sub;
611                         void *elem;
612                         struct name_path me;
613
614                         elem = read_sha1_file(sha1, type, &sub.size);
615                         sub.buf = elem;
616                         if (sub.buf) {
617                                 me.up = up;
618                                 me.elem = name;
619                                 me.len = strlen(name);
620                                 add_pbase_tree(&sub, &me);
621                                 free(elem);
622                         }
623                 }
624         }
625 }
626
627 static void add_preferred_base(unsigned char *sha1)
628 {
629         struct tree_desc tree;
630         void *elem;
631
632         elem = read_object_with_reference(sha1, "tree", &tree.size, NULL);
633         tree.buf = elem;
634         if (!tree.buf)
635                 return;
636         if (add_object_entry(sha1, name_hash(NULL, ""), 1))
637                 add_pbase_tree(&tree, NULL);
638         free(elem);
639 }
640
641 static void check_object(struct object_entry *entry)
642 {
643         char type[20];
644
645         if (entry->in_pack && !entry->preferred_base) {
646                 unsigned char base[20];
647                 unsigned long size;
648                 struct object_entry *base_entry;
649
650                 /* We want in_pack_type even if we do not reuse delta.
651                  * There is no point not reusing non-delta representations.
652                  */
653                 check_reuse_pack_delta(entry->in_pack,
654                                        entry->in_pack_offset,
655                                        base, &size,
656                                        &entry->in_pack_type);
657
658                 /* Check if it is delta, and the base is also an object
659                  * we are going to pack.  If so we will reuse the existing
660                  * delta.
661                  */
662                 if (!no_reuse_delta &&
663                     entry->in_pack_type == OBJ_DELTA &&
664                     (base_entry = locate_object_entry(base)) &&
665                     (!base_entry->preferred_base)) {
666
667                         /* Depth value does not matter - find_deltas()
668                          * will never consider reused delta as the
669                          * base object to deltify other objects
670                          * against, in order to avoid circular deltas.
671                          */
672
673                         /* uncompressed size of the delta data */
674                         entry->size = entry->delta_size = size;
675                         entry->delta = base_entry;
676                         entry->type = OBJ_DELTA;
677
678                         entry->delta_sibling = base_entry->delta_child;
679                         base_entry->delta_child = entry;
680
681                         return;
682                 }
683                 /* Otherwise we would do the usual */
684         }
685
686         if (sha1_object_info(entry->sha1, type, &entry->size))
687                 die("unable to get type of object %s",
688                     sha1_to_hex(entry->sha1));
689
690         if (!strcmp(type, "commit")) {
691                 entry->type = OBJ_COMMIT;
692         } else if (!strcmp(type, "tree")) {
693                 entry->type = OBJ_TREE;
694         } else if (!strcmp(type, "blob")) {
695                 entry->type = OBJ_BLOB;
696         } else if (!strcmp(type, "tag")) {
697                 entry->type = OBJ_TAG;
698         } else
699                 die("unable to pack object %s of type %s",
700                     sha1_to_hex(entry->sha1), type);
701 }
702
703 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
704 {
705         struct object_entry *child = me->delta_child;
706         unsigned int m = n;
707         while (child) {
708                 unsigned int c = check_delta_limit(child, n + 1);
709                 if (m < c)
710                         m = c;
711                 child = child->delta_sibling;
712         }
713         return m;
714 }
715
716 static void get_object_details(void)
717 {
718         int i;
719         struct object_entry *entry;
720
721         prepare_pack_ix();
722         for (i = 0, entry = objects; i < nr_objects; i++, entry++)
723                 check_object(entry);
724
725         if (nr_objects == nr_result) {
726                 /*
727                  * Depth of objects that depend on the entry -- this
728                  * is subtracted from depth-max to break too deep
729                  * delta chain because of delta data reusing.
730                  * However, we loosen this restriction when we know we
731                  * are creating a thin pack -- it will have to be
732                  * expanded on the other end anyway, so do not
733                  * artificially cut the delta chain and let it go as
734                  * deep as it wants.
735                  */
736                 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
737                         if (!entry->delta && entry->delta_child)
738                                 entry->delta_limit =
739                                         check_delta_limit(entry, 1);
740         }
741 }
742
743 typedef int (*entry_sort_t)(const struct object_entry *, const struct object_entry *);
744
745 static entry_sort_t current_sort;
746
747 static int sort_comparator(const void *_a, const void *_b)
748 {
749         struct object_entry *a = *(struct object_entry **)_a;
750         struct object_entry *b = *(struct object_entry **)_b;
751         return current_sort(a,b);
752 }
753
754 static struct object_entry **create_sorted_list(entry_sort_t sort)
755 {
756         struct object_entry **list = xmalloc(nr_objects * sizeof(struct object_entry *));
757         int i;
758
759         for (i = 0; i < nr_objects; i++)
760                 list[i] = objects + i;
761         current_sort = sort;
762         qsort(list, nr_objects, sizeof(struct object_entry *), sort_comparator);
763         return list;
764 }
765
766 static int sha1_sort(const struct object_entry *a, const struct object_entry *b)
767 {
768         return memcmp(a->sha1, b->sha1, 20);
769 }
770
771 static struct object_entry **create_final_object_list()
772 {
773         struct object_entry **list;
774         int i, j;
775
776         for (i = nr_result = 0; i < nr_objects; i++)
777                 if (!objects[i].preferred_base)
778                         nr_result++;
779         list = xmalloc(nr_result * sizeof(struct object_entry *));
780         for (i = j = 0; i < nr_objects; i++) {
781                 if (!objects[i].preferred_base)
782                         list[j++] = objects + i;
783         }
784         current_sort = sha1_sort;
785         qsort(list, nr_result, sizeof(struct object_entry *), sort_comparator);
786         return list;
787 }
788
789 static int type_size_sort(const struct object_entry *a, const struct object_entry *b)
790 {
791         if (a->type < b->type)
792                 return -1;
793         if (a->type > b->type)
794                 return 1;
795         if (a->hash < b->hash)
796                 return -1;
797         if (a->hash > b->hash)
798                 return 1;
799         if (a->preferred_base < b->preferred_base)
800                 return -1;
801         if (a->preferred_base > b->preferred_base)
802                 return 1;
803         if (a->size < b->size)
804                 return -1;
805         if (a->size > b->size)
806                 return 1;
807         return a < b ? -1 : (a > b);
808 }
809
810 struct unpacked {
811         struct object_entry *entry;
812         void *data;
813 };
814
815 /*
816  * We search for deltas _backwards_ in a list sorted by type and
817  * by size, so that we see progressively smaller and smaller files.
818  * That's because we prefer deltas to be from the bigger file
819  * to the smaller - deletes are potentially cheaper, but perhaps
820  * more importantly, the bigger file is likely the more recent
821  * one.
822  */
823 static int try_delta(struct unpacked *cur, struct unpacked *old, unsigned max_depth)
824 {
825         struct object_entry *cur_entry = cur->entry;
826         struct object_entry *old_entry = old->entry;
827         int old_preferred = (old_entry->preferred_base ||
828                              old_entry->based_on_preferred);
829         unsigned long size, oldsize, delta_size, sizediff;
830         long max_size;
831         void *delta_buf;
832
833         /* Don't bother doing diffs between different types */
834         if (cur_entry->type != old_entry->type)
835                 return -1;
836
837         /* We do not compute delta to *create* objects we are not
838          * going to pack.
839          */
840         if (cur_entry->preferred_base)
841                 return -1;
842
843         /* If the current object is at pack edge, take the depth the
844          * objects that depend on the current object into account --
845          * otherwise they would become too deep.
846          */
847         if (cur_entry->delta_child) {
848                 if (max_depth <= cur_entry->delta_limit)
849                         return 0;
850                 max_depth -= cur_entry->delta_limit;
851         }
852
853         size = cur_entry->size;
854         if (size < 50)
855                 return -1;
856         oldsize = old_entry->size;
857         sizediff = oldsize > size ? oldsize - size : size - oldsize;
858         if (sizediff > size / 8)
859                 return -1;
860         if (old_entry->depth >= max_depth)
861                 return 0;
862
863         /*
864          * NOTE!
865          *
866          * We always delta from the bigger to the smaller, since that's
867          * more space-efficient (deletes don't have to say _what_ they
868          * delete).
869          */
870         max_size = size / 2 - 20;
871         if (cur_entry->delta) {
872                 if (cur_entry->based_on_preferred) {
873                         if (old_preferred)
874                                 max_size = cur_entry->delta_size-1;
875                         else
876                                 /* trying with non-preferred one when we
877                                  * already have a delta based on preferred
878                                  * one is pointless.
879                                  */
880                                 return -1;
881                 }
882                 else if (!old_preferred)
883                         max_size = cur_entry->delta_size-1;
884                 else
885                         /* otherwise...  even if delta with a
886                          * preferred one produces a bigger result than
887                          * what we currently have, which is based on a
888                          * non-preferred one, it is OK.
889                          */
890                         ;
891         }
892         if (sizediff >= max_size)
893                 return -1;
894         delta_buf = diff_delta(old->data, oldsize,
895                                cur->data, size, &delta_size, max_size);
896         if (!delta_buf)
897                 return 0;
898         cur_entry->delta = old_entry;
899         cur_entry->delta_size = delta_size;
900         cur_entry->depth = old_entry->depth + 1;
901         cur_entry->based_on_preferred = old_preferred;
902         free(delta_buf);
903         return 0;
904 }
905
906 static void progress_interval(int signum)
907 {
908         signal(SIGALRM, progress_interval);
909         progress_update = 1;
910 }
911
912 static void find_deltas(struct object_entry **list, int window, int depth)
913 {
914         int i, idx;
915         unsigned int array_size = window * sizeof(struct unpacked);
916         struct unpacked *array = xmalloc(array_size);
917         unsigned processed = 0;
918         unsigned last_percent = 999;
919
920         memset(array, 0, array_size);
921         i = nr_objects;
922         idx = 0;
923         if (progress)
924                 fprintf(stderr, "Deltifying %d objects.\n", nr_result);
925
926         while (--i >= 0) {
927                 struct object_entry *entry = list[i];
928                 struct unpacked *n = array + idx;
929                 unsigned long size;
930                 char type[10];
931                 int j;
932
933                 if (!entry->preferred_base)
934                         processed++;
935
936                 if (progress) {
937                         unsigned percent = processed * 100 / nr_result;
938                         if (percent != last_percent || progress_update) {
939                                 fprintf(stderr, "%4u%% (%u/%u) done\r",
940                                         percent, processed, nr_result);
941                                 progress_update = 0;
942                                 last_percent = percent;
943                         }
944                 }
945
946                 if (entry->delta)
947                         /* This happens if we decided to reuse existing
948                          * delta from a pack.  "!no_reuse_delta &&" is implied.
949                          */
950                         continue;
951
952                 free(n->data);
953                 n->entry = entry;
954                 n->data = read_sha1_file(entry->sha1, type, &size);
955                 if (size != entry->size)
956                         die("object %s inconsistent object length (%lu vs %lu)", sha1_to_hex(entry->sha1), size, entry->size);
957
958                 j = window;
959                 while (--j > 0) {
960                         unsigned int other_idx = idx + j;
961                         struct unpacked *m;
962                         if (other_idx >= window)
963                                 other_idx -= window;
964                         m = array + other_idx;
965                         if (!m->entry)
966                                 break;
967                         if (try_delta(n, m, depth) < 0)
968                                 break;
969                 }
970                 idx++;
971                 if (idx >= window)
972                         idx = 0;
973         }
974
975         if (progress)
976                 fputc('\n', stderr);
977
978         for (i = 0; i < window; ++i)
979                 free(array[i].data);
980         free(array);
981 }
982
983 static void prepare_pack(int window, int depth)
984 {
985         get_object_details();
986         sorted_by_type = create_sorted_list(type_size_sort);
987         if (window && depth)
988                 find_deltas(sorted_by_type, window+1, depth);
989 }
990
991 static int reuse_cached_pack(unsigned char *sha1, int pack_to_stdout)
992 {
993         static const char cache[] = "pack-cache/pack-%s.%s";
994         char *cached_pack, *cached_idx;
995         int ifd, ofd, ifd_ix = -1;
996
997         cached_pack = git_path(cache, sha1_to_hex(sha1), "pack");
998         ifd = open(cached_pack, O_RDONLY);
999         if (ifd < 0)
1000                 return 0;
1001
1002         if (!pack_to_stdout) {
1003                 cached_idx = git_path(cache, sha1_to_hex(sha1), "idx");
1004                 ifd_ix = open(cached_idx, O_RDONLY);
1005                 if (ifd_ix < 0) {
1006                         close(ifd);
1007                         return 0;
1008                 }
1009         }
1010
1011         if (progress)
1012                 fprintf(stderr, "Reusing %d objects pack %s\n", nr_objects,
1013                         sha1_to_hex(sha1));
1014
1015         if (pack_to_stdout) {
1016                 if (copy_fd(ifd, 1))
1017                         exit(1);
1018                 close(ifd);
1019         }
1020         else {
1021                 char name[PATH_MAX];
1022                 snprintf(name, sizeof(name),
1023                          "%s-%s.%s", base_name, sha1_to_hex(sha1), "pack");
1024                 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1025                 if (ofd < 0)
1026                         die("unable to open %s (%s)", name, strerror(errno));
1027                 if (copy_fd(ifd, ofd))
1028                         exit(1);
1029                 close(ifd);
1030
1031                 snprintf(name, sizeof(name),
1032                          "%s-%s.%s", base_name, sha1_to_hex(sha1), "idx");
1033                 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1034                 if (ofd < 0)
1035                         die("unable to open %s (%s)", name, strerror(errno));
1036                 if (copy_fd(ifd_ix, ofd))
1037                         exit(1);
1038                 close(ifd_ix);
1039                 puts(sha1_to_hex(sha1));
1040         }
1041
1042         return 1;
1043 }
1044
1045 int main(int argc, char **argv)
1046 {
1047         SHA_CTX ctx;
1048         char line[PATH_MAX + 20];
1049         int window = 10, depth = 10, pack_to_stdout = 0;
1050         struct object_entry **list;
1051         int i;
1052
1053         setup_git_directory();
1054
1055         for (i = 1; i < argc; i++) {
1056                 const char *arg = argv[i];
1057
1058                 if (*arg == '-') {
1059                         if (!strcmp("--non-empty", arg)) {
1060                                 non_empty = 1;
1061                                 continue;
1062                         }
1063                         if (!strcmp("--local", arg)) {
1064                                 local = 1;
1065                                 continue;
1066                         }
1067                         if (!strcmp("--incremental", arg)) {
1068                                 incremental = 1;
1069                                 continue;
1070                         }
1071                         if (!strncmp("--window=", arg, 9)) {
1072                                 char *end;
1073                                 window = strtoul(arg+9, &end, 0);
1074                                 if (!arg[9] || *end)
1075                                         usage(pack_usage);
1076                                 continue;
1077                         }
1078                         if (!strncmp("--depth=", arg, 8)) {
1079                                 char *end;
1080                                 depth = strtoul(arg+8, &end, 0);
1081                                 if (!arg[8] || *end)
1082                                         usage(pack_usage);
1083                                 continue;
1084                         }
1085                         if (!strcmp("-q", arg)) {
1086                                 progress = 0;
1087                                 continue;
1088                         }
1089                         if (!strcmp("--no-reuse-delta", arg)) {
1090                                 no_reuse_delta = 1;
1091                                 continue;
1092                         }
1093                         if (!strcmp("--stdout", arg)) {
1094                                 pack_to_stdout = 1;
1095                                 continue;
1096                         }
1097                         usage(pack_usage);
1098                 }
1099                 if (base_name)
1100                         usage(pack_usage);
1101                 base_name = arg;
1102         }
1103
1104         if (pack_to_stdout != !base_name)
1105                 usage(pack_usage);
1106
1107         prepare_packed_git();
1108
1109         if (progress) {
1110                 struct itimerval v;
1111                 v.it_interval.tv_sec = 1;
1112                 v.it_interval.tv_usec = 0;
1113                 v.it_value = v.it_interval;
1114                 signal(SIGALRM, progress_interval);
1115                 setitimer(ITIMER_REAL, &v, NULL);
1116                 fprintf(stderr, "Generating pack...\n");
1117         }
1118
1119         while (fgets(line, sizeof(line), stdin) != NULL) {
1120                 unsigned char sha1[20];
1121
1122                 if (line[0] == '-') {
1123                         if (get_sha1_hex(line+1, sha1))
1124                                 die("expected edge sha1, got garbage:\n %s",
1125                                     line+1);
1126                         add_preferred_base(sha1);
1127                         continue;
1128                 }
1129                 if (get_sha1_hex(line, sha1))
1130                         die("expected sha1, got garbage:\n %s", line);
1131                 add_object_entry(sha1, name_hash(NULL, line+41), 0);
1132         }
1133         if (progress)
1134                 fprintf(stderr, "Done counting %d objects.\n", nr_objects);
1135         sorted_by_sha = create_final_object_list();
1136         if (non_empty && !nr_result)
1137                 return 0;
1138
1139         SHA1_Init(&ctx);
1140         list = sorted_by_sha;
1141         for (i = 0; i < nr_result; i++) {
1142                 struct object_entry *entry = *list++;
1143                 SHA1_Update(&ctx, entry->sha1, 20);
1144         }
1145         SHA1_Final(object_list_sha1, &ctx);
1146         if (progress && (nr_objects != nr_result))
1147                 fprintf(stderr, "Result has %d objects.\n", nr_result);
1148
1149         if (reuse_cached_pack(object_list_sha1, pack_to_stdout))
1150                 ;
1151         else {
1152                 if (nr_result)
1153                         prepare_pack(window, depth);
1154                 if (progress && pack_to_stdout) {
1155                         /* the other end usually displays progress itself */
1156                         struct itimerval v = {{0,},};
1157                         setitimer(ITIMER_REAL, &v, NULL);
1158                         signal(SIGALRM, SIG_IGN );
1159                         progress_update = 0;
1160                 }
1161                 write_pack_file();
1162                 if (!pack_to_stdout) {
1163                         write_index_file();
1164                         puts(sha1_to_hex(object_list_sha1));
1165                 }
1166         }
1167         if (progress)
1168                 fprintf(stderr, "Total %d, written %d (delta %d), reused %d (delta %d)\n",
1169                         nr_result, written, written_delta, reused, reused_delta);
1170         return 0;
1171 }