Merge branch 'jc/format-patch'
[git/debian.git] / builtin-pack-objects.c
blob448461bc4826da8b34ebf32c876d9b2341f51447
1 #include "builtin.h"
2 #include "cache.h"
3 #include "object.h"
4 #include "blob.h"
5 #include "commit.h"
6 #include "tag.h"
7 #include "tree.h"
8 #include "delta.h"
9 #include "pack.h"
10 #include "csum-file.h"
11 #include "tree-walk.h"
12 #include <sys/time.h>
13 #include <signal.h>
15 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";
17 struct object_entry {
18 unsigned char sha1[20];
19 unsigned long size; /* uncompressed size */
20 unsigned long offset; /* offset into the final pack file;
21 * nonzero if already written.
23 unsigned int depth; /* delta depth */
24 unsigned int delta_limit; /* base adjustment for in-pack delta */
25 unsigned int hash; /* name hint hash */
26 enum object_type type;
27 enum object_type in_pack_type; /* could be delta */
28 unsigned long delta_size; /* delta data size (uncompressed) */
29 struct object_entry *delta; /* delta base object */
30 struct packed_git *in_pack; /* already in pack */
31 unsigned int in_pack_offset;
32 struct object_entry *delta_child; /* deltified objects who bases me */
33 struct object_entry *delta_sibling; /* other deltified objects who
34 * uses the same base as me
36 int preferred_base; /* we do not pack this, but is encouraged to
37 * be used as the base objectto delta huge
38 * objects against.
43 * Objects we are going to pack are collected in objects array (dynamically
44 * expanded). nr_objects & nr_alloc controls this array. They are stored
45 * in the order we see -- typically rev-list --objects order that gives us
46 * nice "minimum seek" order.
48 * sorted-by-sha ans sorted-by-type are arrays of pointers that point at
49 * elements in the objects array. The former is used to build the pack
50 * index (lists object names in the ascending order to help offset lookup),
51 * and the latter is used to group similar things together by try_delta()
52 * heuristics.
55 static unsigned char object_list_sha1[20];
56 static int non_empty;
57 static int no_reuse_delta;
58 static int local;
59 static int incremental;
60 static struct object_entry **sorted_by_sha, **sorted_by_type;
61 static struct object_entry *objects;
62 static int nr_objects, nr_alloc, nr_result;
63 static const char *base_name;
64 static unsigned char pack_file_sha1[20];
65 static int progress = 1;
66 static volatile sig_atomic_t progress_update;
67 static int window = 10;
70 * The object names in objects array are hashed with this hashtable,
71 * to help looking up the entry by object name. Binary search from
72 * sorted_by_sha is also possible but this was easier to code and faster.
73 * This hashtable is built after all the objects are seen.
75 static int *object_ix;
76 static int object_ix_hashsz;
79 * Pack index for existing packs give us easy access to the offsets into
80 * corresponding pack file where each object's data starts, but the entries
81 * do not store the size of the compressed representation (uncompressed
82 * size is easily available by examining the pack entry header). We build
83 * a hashtable of existing packs (pack_revindex), and keep reverse index
84 * here -- pack index file is sorted by object name mapping to offset; this
85 * pack_revindex[].revindex array is an ordered list of offsets, so if you
86 * know the offset of an object, next offset is where its packed
87 * representation ends.
89 struct pack_revindex {
90 struct packed_git *p;
91 unsigned long *revindex;
92 } *pack_revindex = NULL;
93 static int pack_revindex_hashsz;
96 * stats
98 static int written;
99 static int written_delta;
100 static int reused;
101 static int reused_delta;
103 static int pack_revindex_ix(struct packed_git *p)
105 unsigned long ui = (unsigned long)p;
106 int i;
108 ui = ui ^ (ui >> 16); /* defeat structure alignment */
109 i = (int)(ui % pack_revindex_hashsz);
110 while (pack_revindex[i].p) {
111 if (pack_revindex[i].p == p)
112 return i;
113 if (++i == pack_revindex_hashsz)
114 i = 0;
116 return -1 - i;
119 static void prepare_pack_ix(void)
121 int num;
122 struct packed_git *p;
123 for (num = 0, p = packed_git; p; p = p->next)
124 num++;
125 if (!num)
126 return;
127 pack_revindex_hashsz = num * 11;
128 pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
129 for (p = packed_git; p; p = p->next) {
130 num = pack_revindex_ix(p);
131 num = - 1 - num;
132 pack_revindex[num].p = p;
134 /* revindex elements are lazily initialized */
137 static int cmp_offset(const void *a_, const void *b_)
139 unsigned long a = *(unsigned long *) a_;
140 unsigned long b = *(unsigned long *) b_;
141 if (a < b)
142 return -1;
143 else if (a == b)
144 return 0;
145 else
146 return 1;
150 * Ordered list of offsets of objects in the pack.
152 static void prepare_pack_revindex(struct pack_revindex *rix)
154 struct packed_git *p = rix->p;
155 int num_ent = num_packed_objects(p);
156 int i;
157 void *index = p->index_base + 256;
159 rix->revindex = xmalloc(sizeof(unsigned long) * (num_ent + 1));
160 for (i = 0; i < num_ent; i++) {
161 unsigned int hl = *((unsigned int *)((char *) index + 24*i));
162 rix->revindex[i] = ntohl(hl);
164 /* This knows the pack format -- the 20-byte trailer
165 * follows immediately after the last object data.
167 rix->revindex[num_ent] = p->pack_size - 20;
168 qsort(rix->revindex, num_ent, sizeof(unsigned long), cmp_offset);
171 static unsigned long find_packed_object_size(struct packed_git *p,
172 unsigned long ofs)
174 int num;
175 int lo, hi;
176 struct pack_revindex *rix;
177 unsigned long *revindex;
178 num = pack_revindex_ix(p);
179 if (num < 0)
180 die("internal error: pack revindex uninitialized");
181 rix = &pack_revindex[num];
182 if (!rix->revindex)
183 prepare_pack_revindex(rix);
184 revindex = rix->revindex;
185 lo = 0;
186 hi = num_packed_objects(p) + 1;
187 do {
188 int mi = (lo + hi) / 2;
189 if (revindex[mi] == ofs) {
190 return revindex[mi+1] - ofs;
192 else if (ofs < revindex[mi])
193 hi = mi;
194 else
195 lo = mi + 1;
196 } while (lo < hi);
197 die("internal error: pack revindex corrupt");
200 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
202 unsigned long othersize, delta_size;
203 char type[10];
204 void *otherbuf = read_sha1_file(entry->delta->sha1, type, &othersize);
205 void *delta_buf;
207 if (!otherbuf)
208 die("unable to read %s", sha1_to_hex(entry->delta->sha1));
209 delta_buf = diff_delta(otherbuf, othersize,
210 buf, size, &delta_size, 0);
211 if (!delta_buf || delta_size != entry->delta_size)
212 die("delta size changed");
213 free(buf);
214 free(otherbuf);
215 return delta_buf;
219 * The per-object header is a pretty dense thing, which is
220 * - first byte: low four bits are "size", then three bits of "type",
221 * and the high bit is "size continues".
222 * - each byte afterwards: low seven bits are size continuation,
223 * with the high bit being "size continues"
225 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
227 int n = 1;
228 unsigned char c;
230 if (type < OBJ_COMMIT || type > OBJ_DELTA)
231 die("bad type %d", type);
233 c = (type << 4) | (size & 15);
234 size >>= 4;
235 while (size) {
236 *hdr++ = c | 0x80;
237 c = size & 0x7f;
238 size >>= 7;
239 n++;
241 *hdr = c;
242 return n;
245 static unsigned long write_object(struct sha1file *f,
246 struct object_entry *entry)
248 unsigned long size;
249 char type[10];
250 void *buf;
251 unsigned char header[10];
252 unsigned hdrlen, datalen;
253 enum object_type obj_type;
254 int to_reuse = 0;
256 if (entry->preferred_base)
257 return 0;
259 obj_type = entry->type;
260 if (! entry->in_pack)
261 to_reuse = 0; /* can't reuse what we don't have */
262 else if (obj_type == OBJ_DELTA)
263 to_reuse = 1; /* check_object() decided it for us */
264 else if (obj_type != entry->in_pack_type)
265 to_reuse = 0; /* pack has delta which is unusable */
266 else if (entry->delta)
267 to_reuse = 0; /* we want to pack afresh */
268 else
269 to_reuse = 1; /* we have it in-pack undeltified,
270 * and we do not need to deltify it.
273 if (!entry->in_pack && !entry->delta) {
274 unsigned char *map;
275 unsigned long mapsize;
276 map = map_sha1_file(entry->sha1, &mapsize);
277 if (map && !legacy_loose_object(map)) {
278 /* We can copy straight into the pack file */
279 sha1write(f, map, mapsize);
280 munmap(map, mapsize);
281 written++;
282 reused++;
283 return mapsize;
285 if (map)
286 munmap(map, mapsize);
289 if (! to_reuse) {
290 buf = read_sha1_file(entry->sha1, type, &size);
291 if (!buf)
292 die("unable to read %s", sha1_to_hex(entry->sha1));
293 if (size != entry->size)
294 die("object %s size inconsistency (%lu vs %lu)",
295 sha1_to_hex(entry->sha1), size, entry->size);
296 if (entry->delta) {
297 buf = delta_against(buf, size, entry);
298 size = entry->delta_size;
299 obj_type = OBJ_DELTA;
302 * The object header is a byte of 'type' followed by zero or
303 * more bytes of length. For deltas, the 20 bytes of delta
304 * sha1 follows that.
306 hdrlen = encode_header(obj_type, size, header);
307 sha1write(f, header, hdrlen);
309 if (entry->delta) {
310 sha1write(f, entry->delta, 20);
311 hdrlen += 20;
313 datalen = sha1write_compressed(f, buf, size);
314 free(buf);
316 else {
317 struct packed_git *p = entry->in_pack;
318 use_packed_git(p);
320 datalen = find_packed_object_size(p, entry->in_pack_offset);
321 buf = (char *) p->pack_base + entry->in_pack_offset;
322 sha1write(f, buf, datalen);
323 unuse_packed_git(p);
324 hdrlen = 0; /* not really */
325 if (obj_type == OBJ_DELTA)
326 reused_delta++;
327 reused++;
329 if (obj_type == OBJ_DELTA)
330 written_delta++;
331 written++;
332 return hdrlen + datalen;
335 static unsigned long write_one(struct sha1file *f,
336 struct object_entry *e,
337 unsigned long offset)
339 if (e->offset)
340 /* offset starts from header size and cannot be zero
341 * if it is written already.
343 return offset;
344 e->offset = offset;
345 offset += write_object(f, e);
346 /* if we are deltified, write out its base object. */
347 if (e->delta)
348 offset = write_one(f, e->delta, offset);
349 return offset;
352 static void write_pack_file(void)
354 int i;
355 struct sha1file *f;
356 unsigned long offset;
357 struct pack_header hdr;
358 unsigned last_percent = 999;
359 int do_progress = 0;
361 if (!base_name)
362 f = sha1fd(1, "<stdout>");
363 else {
364 f = sha1create("%s-%s.%s", base_name,
365 sha1_to_hex(object_list_sha1), "pack");
366 do_progress = progress;
368 if (do_progress)
369 fprintf(stderr, "Writing %d objects.\n", nr_result);
371 hdr.hdr_signature = htonl(PACK_SIGNATURE);
372 hdr.hdr_version = htonl(PACK_VERSION);
373 hdr.hdr_entries = htonl(nr_result);
374 sha1write(f, &hdr, sizeof(hdr));
375 offset = sizeof(hdr);
376 if (!nr_result)
377 goto done;
378 for (i = 0; i < nr_objects; i++) {
379 offset = write_one(f, objects + i, offset);
380 if (do_progress) {
381 unsigned percent = written * 100 / nr_result;
382 if (progress_update || percent != last_percent) {
383 fprintf(stderr, "%4u%% (%u/%u) done\r",
384 percent, written, nr_result);
385 progress_update = 0;
386 last_percent = percent;
390 if (do_progress)
391 fputc('\n', stderr);
392 done:
393 sha1close(f, pack_file_sha1, 1);
396 static void write_index_file(void)
398 int i;
399 struct sha1file *f = sha1create("%s-%s.%s", base_name,
400 sha1_to_hex(object_list_sha1), "idx");
401 struct object_entry **list = sorted_by_sha;
402 struct object_entry **last = list + nr_result;
403 unsigned int array[256];
406 * Write the first-level table (the list is sorted,
407 * but we use a 256-entry lookup to be able to avoid
408 * having to do eight extra binary search iterations).
410 for (i = 0; i < 256; i++) {
411 struct object_entry **next = list;
412 while (next < last) {
413 struct object_entry *entry = *next;
414 if (entry->sha1[0] != i)
415 break;
416 next++;
418 array[i] = htonl(next - sorted_by_sha);
419 list = next;
421 sha1write(f, array, 256 * sizeof(int));
424 * Write the actual SHA1 entries..
426 list = sorted_by_sha;
427 for (i = 0; i < nr_result; i++) {
428 struct object_entry *entry = *list++;
429 unsigned int offset = htonl(entry->offset);
430 sha1write(f, &offset, 4);
431 sha1write(f, entry->sha1, 20);
433 sha1write(f, pack_file_sha1, 20);
434 sha1close(f, NULL, 1);
437 static int locate_object_entry_hash(const unsigned char *sha1)
439 int i;
440 unsigned int ui;
441 memcpy(&ui, sha1, sizeof(unsigned int));
442 i = ui % object_ix_hashsz;
443 while (0 < object_ix[i]) {
444 if (!memcmp(sha1, objects[object_ix[i]-1].sha1, 20))
445 return i;
446 if (++i == object_ix_hashsz)
447 i = 0;
449 return -1 - i;
452 static struct object_entry *locate_object_entry(const unsigned char *sha1)
454 int i;
456 if (!object_ix_hashsz)
457 return NULL;
459 i = locate_object_entry_hash(sha1);
460 if (0 <= i)
461 return &objects[object_ix[i]-1];
462 return NULL;
465 static void rehash_objects(void)
467 int i;
468 struct object_entry *oe;
470 object_ix_hashsz = nr_objects * 3;
471 if (object_ix_hashsz < 1024)
472 object_ix_hashsz = 1024;
473 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
474 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
475 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
476 int ix = locate_object_entry_hash(oe->sha1);
477 if (0 <= ix)
478 continue;
479 ix = -1 - ix;
480 object_ix[ix] = i + 1;
484 static unsigned name_hash(const char *name)
486 unsigned char c;
487 unsigned hash = 0;
490 * This effectively just creates a sortable number from the
491 * last sixteen non-whitespace characters. Last characters
492 * count "most", so things that end in ".c" sort together.
494 while ((c = *name++) != 0) {
495 if (isspace(c))
496 continue;
497 hash = (hash >> 2) + (c << 24);
499 return hash;
502 static int add_object_entry(const unsigned char *sha1, unsigned hash, int exclude)
504 unsigned int idx = nr_objects;
505 struct object_entry *entry;
506 struct packed_git *p;
507 unsigned int found_offset = 0;
508 struct packed_git *found_pack = NULL;
509 int ix, status = 0;
511 if (!exclude) {
512 for (p = packed_git; p; p = p->next) {
513 struct pack_entry e;
514 if (find_pack_entry_one(sha1, &e, p)) {
515 if (incremental)
516 return 0;
517 if (local && !p->pack_local)
518 return 0;
519 if (!found_pack) {
520 found_offset = e.offset;
521 found_pack = e.p;
526 if ((entry = locate_object_entry(sha1)) != NULL)
527 goto already_added;
529 if (idx >= nr_alloc) {
530 unsigned int needed = (idx + 1024) * 3 / 2;
531 objects = xrealloc(objects, needed * sizeof(*entry));
532 nr_alloc = needed;
534 entry = objects + idx;
535 nr_objects = idx + 1;
536 memset(entry, 0, sizeof(*entry));
537 memcpy(entry->sha1, sha1, 20);
538 entry->hash = hash;
540 if (object_ix_hashsz * 3 <= nr_objects * 4)
541 rehash_objects();
542 else {
543 ix = locate_object_entry_hash(entry->sha1);
544 if (0 <= ix)
545 die("internal error in object hashing.");
546 object_ix[-1 - ix] = idx + 1;
548 status = 1;
550 already_added:
551 if (progress_update) {
552 fprintf(stderr, "Counting objects...%d\r", nr_objects);
553 progress_update = 0;
555 if (exclude)
556 entry->preferred_base = 1;
557 else {
558 if (found_pack) {
559 entry->in_pack = found_pack;
560 entry->in_pack_offset = found_offset;
563 return status;
566 struct pbase_tree_cache {
567 unsigned char sha1[20];
568 int ref;
569 int temporary;
570 void *tree_data;
571 unsigned long tree_size;
574 static struct pbase_tree_cache *(pbase_tree_cache[256]);
575 static int pbase_tree_cache_ix(const unsigned char *sha1)
577 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
579 static int pbase_tree_cache_ix_incr(int ix)
581 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
584 static struct pbase_tree {
585 struct pbase_tree *next;
586 /* This is a phony "cache" entry; we are not
587 * going to evict it nor find it through _get()
588 * mechanism -- this is for the toplevel node that
589 * would almost always change with any commit.
591 struct pbase_tree_cache pcache;
592 } *pbase_tree;
594 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
596 struct pbase_tree_cache *ent, *nent;
597 void *data;
598 unsigned long size;
599 char type[20];
600 int neigh;
601 int my_ix = pbase_tree_cache_ix(sha1);
602 int available_ix = -1;
604 /* pbase-tree-cache acts as a limited hashtable.
605 * your object will be found at your index or within a few
606 * slots after that slot if it is cached.
608 for (neigh = 0; neigh < 8; neigh++) {
609 ent = pbase_tree_cache[my_ix];
610 if (ent && !memcmp(ent->sha1, sha1, 20)) {
611 ent->ref++;
612 return ent;
614 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
615 ((0 <= available_ix) &&
616 (!ent && pbase_tree_cache[available_ix])))
617 available_ix = my_ix;
618 if (!ent)
619 break;
620 my_ix = pbase_tree_cache_ix_incr(my_ix);
623 /* Did not find one. Either we got a bogus request or
624 * we need to read and perhaps cache.
626 data = read_sha1_file(sha1, type, &size);
627 if (!data)
628 return NULL;
629 if (strcmp(type, tree_type)) {
630 free(data);
631 return NULL;
634 /* We need to either cache or return a throwaway copy */
636 if (available_ix < 0)
637 ent = NULL;
638 else {
639 ent = pbase_tree_cache[available_ix];
640 my_ix = available_ix;
643 if (!ent) {
644 nent = xmalloc(sizeof(*nent));
645 nent->temporary = (available_ix < 0);
647 else {
648 /* evict and reuse */
649 free(ent->tree_data);
650 nent = ent;
652 memcpy(nent->sha1, sha1, 20);
653 nent->tree_data = data;
654 nent->tree_size = size;
655 nent->ref = 1;
656 if (!nent->temporary)
657 pbase_tree_cache[my_ix] = nent;
658 return nent;
661 static void pbase_tree_put(struct pbase_tree_cache *cache)
663 if (!cache->temporary) {
664 cache->ref--;
665 return;
667 free(cache->tree_data);
668 free(cache);
671 static int name_cmp_len(const char *name)
673 int i;
674 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
676 return i;
679 static void add_pbase_object(struct tree_desc *tree,
680 const char *name,
681 int cmplen,
682 const char *fullname)
684 struct name_entry entry;
686 while (tree_entry(tree,&entry)) {
687 unsigned long size;
688 char type[20];
690 if (entry.pathlen != cmplen ||
691 memcmp(entry.path, name, cmplen) ||
692 !has_sha1_file(entry.sha1) ||
693 sha1_object_info(entry.sha1, type, &size))
694 continue;
695 if (name[cmplen] != '/') {
696 unsigned hash = name_hash(fullname);
697 add_object_entry(entry.sha1, hash, 1);
698 return;
700 if (!strcmp(type, tree_type)) {
701 struct tree_desc sub;
702 struct pbase_tree_cache *tree;
703 const char *down = name+cmplen+1;
704 int downlen = name_cmp_len(down);
706 tree = pbase_tree_get(entry.sha1);
707 if (!tree)
708 return;
709 sub.buf = tree->tree_data;
710 sub.size = tree->tree_size;
712 add_pbase_object(&sub, down, downlen, fullname);
713 pbase_tree_put(tree);
718 static unsigned *done_pbase_paths;
719 static int done_pbase_paths_num;
720 static int done_pbase_paths_alloc;
721 static int done_pbase_path_pos(unsigned hash)
723 int lo = 0;
724 int hi = done_pbase_paths_num;
725 while (lo < hi) {
726 int mi = (hi + lo) / 2;
727 if (done_pbase_paths[mi] == hash)
728 return mi;
729 if (done_pbase_paths[mi] < hash)
730 hi = mi;
731 else
732 lo = mi + 1;
734 return -lo-1;
737 static int check_pbase_path(unsigned hash)
739 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
740 if (0 <= pos)
741 return 1;
742 pos = -pos - 1;
743 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
744 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
745 done_pbase_paths = xrealloc(done_pbase_paths,
746 done_pbase_paths_alloc *
747 sizeof(unsigned));
749 done_pbase_paths_num++;
750 if (pos < done_pbase_paths_num)
751 memmove(done_pbase_paths + pos + 1,
752 done_pbase_paths + pos,
753 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
754 done_pbase_paths[pos] = hash;
755 return 0;
758 static void add_preferred_base_object(char *name, unsigned hash)
760 struct pbase_tree *it;
761 int cmplen = name_cmp_len(name);
763 if (check_pbase_path(hash))
764 return;
766 for (it = pbase_tree; it; it = it->next) {
767 if (cmplen == 0) {
768 hash = name_hash("");
769 add_object_entry(it->pcache.sha1, hash, 1);
771 else {
772 struct tree_desc tree;
773 tree.buf = it->pcache.tree_data;
774 tree.size = it->pcache.tree_size;
775 add_pbase_object(&tree, name, cmplen, name);
780 static void add_preferred_base(unsigned char *sha1)
782 struct pbase_tree *it;
783 void *data;
784 unsigned long size;
785 unsigned char tree_sha1[20];
787 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
788 if (!data)
789 return;
791 for (it = pbase_tree; it; it = it->next) {
792 if (!memcmp(it->pcache.sha1, tree_sha1, 20)) {
793 free(data);
794 return;
798 it = xcalloc(1, sizeof(*it));
799 it->next = pbase_tree;
800 pbase_tree = it;
802 memcpy(it->pcache.sha1, tree_sha1, 20);
803 it->pcache.tree_data = data;
804 it->pcache.tree_size = size;
807 static void check_object(struct object_entry *entry)
809 char type[20];
811 if (entry->in_pack && !entry->preferred_base) {
812 unsigned char base[20];
813 unsigned long size;
814 struct object_entry *base_entry;
816 /* We want in_pack_type even if we do not reuse delta.
817 * There is no point not reusing non-delta representations.
819 check_reuse_pack_delta(entry->in_pack,
820 entry->in_pack_offset,
821 base, &size,
822 &entry->in_pack_type);
824 /* Check if it is delta, and the base is also an object
825 * we are going to pack. If so we will reuse the existing
826 * delta.
828 if (!no_reuse_delta &&
829 entry->in_pack_type == OBJ_DELTA &&
830 (base_entry = locate_object_entry(base)) &&
831 (!base_entry->preferred_base)) {
833 /* Depth value does not matter - find_deltas()
834 * will never consider reused delta as the
835 * base object to deltify other objects
836 * against, in order to avoid circular deltas.
839 /* uncompressed size of the delta data */
840 entry->size = entry->delta_size = size;
841 entry->delta = base_entry;
842 entry->type = OBJ_DELTA;
844 entry->delta_sibling = base_entry->delta_child;
845 base_entry->delta_child = entry;
847 return;
849 /* Otherwise we would do the usual */
852 if (sha1_object_info(entry->sha1, type, &entry->size))
853 die("unable to get type of object %s",
854 sha1_to_hex(entry->sha1));
856 if (!strcmp(type, commit_type)) {
857 entry->type = OBJ_COMMIT;
858 } else if (!strcmp(type, tree_type)) {
859 entry->type = OBJ_TREE;
860 } else if (!strcmp(type, blob_type)) {
861 entry->type = OBJ_BLOB;
862 } else if (!strcmp(type, tag_type)) {
863 entry->type = OBJ_TAG;
864 } else
865 die("unable to pack object %s of type %s",
866 sha1_to_hex(entry->sha1), type);
869 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
871 struct object_entry *child = me->delta_child;
872 unsigned int m = n;
873 while (child) {
874 unsigned int c = check_delta_limit(child, n + 1);
875 if (m < c)
876 m = c;
877 child = child->delta_sibling;
879 return m;
882 static void get_object_details(void)
884 int i;
885 struct object_entry *entry;
887 prepare_pack_ix();
888 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
889 check_object(entry);
891 if (nr_objects == nr_result) {
893 * Depth of objects that depend on the entry -- this
894 * is subtracted from depth-max to break too deep
895 * delta chain because of delta data reusing.
896 * However, we loosen this restriction when we know we
897 * are creating a thin pack -- it will have to be
898 * expanded on the other end anyway, so do not
899 * artificially cut the delta chain and let it go as
900 * deep as it wants.
902 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
903 if (!entry->delta && entry->delta_child)
904 entry->delta_limit =
905 check_delta_limit(entry, 1);
909 typedef int (*entry_sort_t)(const struct object_entry *, const struct object_entry *);
911 static entry_sort_t current_sort;
913 static int sort_comparator(const void *_a, const void *_b)
915 struct object_entry *a = *(struct object_entry **)_a;
916 struct object_entry *b = *(struct object_entry **)_b;
917 return current_sort(a,b);
920 static struct object_entry **create_sorted_list(entry_sort_t sort)
922 struct object_entry **list = xmalloc(nr_objects * sizeof(struct object_entry *));
923 int i;
925 for (i = 0; i < nr_objects; i++)
926 list[i] = objects + i;
927 current_sort = sort;
928 qsort(list, nr_objects, sizeof(struct object_entry *), sort_comparator);
929 return list;
932 static int sha1_sort(const struct object_entry *a, const struct object_entry *b)
934 return memcmp(a->sha1, b->sha1, 20);
937 static struct object_entry **create_final_object_list(void)
939 struct object_entry **list;
940 int i, j;
942 for (i = nr_result = 0; i < nr_objects; i++)
943 if (!objects[i].preferred_base)
944 nr_result++;
945 list = xmalloc(nr_result * sizeof(struct object_entry *));
946 for (i = j = 0; i < nr_objects; i++) {
947 if (!objects[i].preferred_base)
948 list[j++] = objects + i;
950 current_sort = sha1_sort;
951 qsort(list, nr_result, sizeof(struct object_entry *), sort_comparator);
952 return list;
955 static int type_size_sort(const struct object_entry *a, const struct object_entry *b)
957 if (a->type < b->type)
958 return -1;
959 if (a->type > b->type)
960 return 1;
961 if (a->hash < b->hash)
962 return -1;
963 if (a->hash > b->hash)
964 return 1;
965 if (a->preferred_base < b->preferred_base)
966 return -1;
967 if (a->preferred_base > b->preferred_base)
968 return 1;
969 if (a->size < b->size)
970 return -1;
971 if (a->size > b->size)
972 return 1;
973 return a < b ? -1 : (a > b);
976 struct unpacked {
977 struct object_entry *entry;
978 void *data;
979 struct delta_index *index;
983 * We search for deltas _backwards_ in a list sorted by type and
984 * by size, so that we see progressively smaller and smaller files.
985 * That's because we prefer deltas to be from the bigger file
986 * to the smaller - deletes are potentially cheaper, but perhaps
987 * more importantly, the bigger file is likely the more recent
988 * one.
990 static int try_delta(struct unpacked *trg, struct unpacked *src,
991 unsigned max_depth)
993 struct object_entry *trg_entry = trg->entry;
994 struct object_entry *src_entry = src->entry;
995 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
996 char type[10];
997 void *delta_buf;
999 /* Don't bother doing diffs between different types */
1000 if (trg_entry->type != src_entry->type)
1001 return -1;
1003 /* We do not compute delta to *create* objects we are not
1004 * going to pack.
1006 if (trg_entry->preferred_base)
1007 return -1;
1010 * We do not bother to try a delta that we discarded
1011 * on an earlier try, but only when reusing delta data.
1013 if (!no_reuse_delta && trg_entry->in_pack &&
1014 trg_entry->in_pack == src_entry->in_pack)
1015 return 0;
1018 * If the current object is at pack edge, take the depth the
1019 * objects that depend on the current object into account --
1020 * otherwise they would become too deep.
1022 if (trg_entry->delta_child) {
1023 if (max_depth <= trg_entry->delta_limit)
1024 return 0;
1025 max_depth -= trg_entry->delta_limit;
1027 if (src_entry->depth >= max_depth)
1028 return 0;
1030 /* Now some size filtering heuristics. */
1031 trg_size = trg_entry->size;
1032 max_size = trg_size/2 - 20;
1033 max_size = max_size * (max_depth - src_entry->depth) / max_depth;
1034 if (max_size == 0)
1035 return 0;
1036 if (trg_entry->delta && trg_entry->delta_size <= max_size)
1037 max_size = trg_entry->delta_size-1;
1038 src_size = src_entry->size;
1039 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1040 if (sizediff >= max_size)
1041 return 0;
1043 /* Load data if not already done */
1044 if (!trg->data) {
1045 trg->data = read_sha1_file(trg_entry->sha1, type, &sz);
1046 if (sz != trg_size)
1047 die("object %s inconsistent object length (%lu vs %lu)",
1048 sha1_to_hex(trg_entry->sha1), sz, trg_size);
1050 if (!src->data) {
1051 src->data = read_sha1_file(src_entry->sha1, type, &sz);
1052 if (sz != src_size)
1053 die("object %s inconsistent object length (%lu vs %lu)",
1054 sha1_to_hex(src_entry->sha1), sz, src_size);
1056 if (!src->index) {
1057 src->index = create_delta_index(src->data, src_size);
1058 if (!src->index)
1059 die("out of memory");
1062 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1063 if (!delta_buf)
1064 return 0;
1066 trg_entry->delta = src_entry;
1067 trg_entry->delta_size = delta_size;
1068 trg_entry->depth = src_entry->depth + 1;
1069 free(delta_buf);
1070 return 1;
1073 static void progress_interval(int signum)
1075 progress_update = 1;
1078 static void find_deltas(struct object_entry **list, int window, int depth)
1080 int i, idx;
1081 unsigned int array_size = window * sizeof(struct unpacked);
1082 struct unpacked *array = xmalloc(array_size);
1083 unsigned processed = 0;
1084 unsigned last_percent = 999;
1086 memset(array, 0, array_size);
1087 i = nr_objects;
1088 idx = 0;
1089 if (progress)
1090 fprintf(stderr, "Deltifying %d objects.\n", nr_result);
1092 while (--i >= 0) {
1093 struct object_entry *entry = list[i];
1094 struct unpacked *n = array + idx;
1095 int j;
1097 if (!entry->preferred_base)
1098 processed++;
1100 if (progress) {
1101 unsigned percent = processed * 100 / nr_result;
1102 if (percent != last_percent || progress_update) {
1103 fprintf(stderr, "%4u%% (%u/%u) done\r",
1104 percent, processed, nr_result);
1105 progress_update = 0;
1106 last_percent = percent;
1110 if (entry->delta)
1111 /* This happens if we decided to reuse existing
1112 * delta from a pack. "!no_reuse_delta &&" is implied.
1114 continue;
1116 if (entry->size < 50)
1117 continue;
1118 free_delta_index(n->index);
1119 n->index = NULL;
1120 free(n->data);
1121 n->data = NULL;
1122 n->entry = entry;
1124 j = window;
1125 while (--j > 0) {
1126 unsigned int other_idx = idx + j;
1127 struct unpacked *m;
1128 if (other_idx >= window)
1129 other_idx -= window;
1130 m = array + other_idx;
1131 if (!m->entry)
1132 break;
1133 if (try_delta(n, m, depth) < 0)
1134 break;
1136 /* if we made n a delta, and if n is already at max
1137 * depth, leaving it in the window is pointless. we
1138 * should evict it first.
1140 if (entry->delta && depth <= entry->depth)
1141 continue;
1143 idx++;
1144 if (idx >= window)
1145 idx = 0;
1148 if (progress)
1149 fputc('\n', stderr);
1151 for (i = 0; i < window; ++i) {
1152 free_delta_index(array[i].index);
1153 free(array[i].data);
1155 free(array);
1158 static void prepare_pack(int window, int depth)
1160 get_object_details();
1161 sorted_by_type = create_sorted_list(type_size_sort);
1162 if (window && depth)
1163 find_deltas(sorted_by_type, window+1, depth);
1166 static int reuse_cached_pack(unsigned char *sha1, int pack_to_stdout)
1168 static const char cache[] = "pack-cache/pack-%s.%s";
1169 char *cached_pack, *cached_idx;
1170 int ifd, ofd, ifd_ix = -1;
1172 cached_pack = git_path(cache, sha1_to_hex(sha1), "pack");
1173 ifd = open(cached_pack, O_RDONLY);
1174 if (ifd < 0)
1175 return 0;
1177 if (!pack_to_stdout) {
1178 cached_idx = git_path(cache, sha1_to_hex(sha1), "idx");
1179 ifd_ix = open(cached_idx, O_RDONLY);
1180 if (ifd_ix < 0) {
1181 close(ifd);
1182 return 0;
1186 if (progress)
1187 fprintf(stderr, "Reusing %d objects pack %s\n", nr_objects,
1188 sha1_to_hex(sha1));
1190 if (pack_to_stdout) {
1191 if (copy_fd(ifd, 1))
1192 exit(1);
1193 close(ifd);
1195 else {
1196 char name[PATH_MAX];
1197 snprintf(name, sizeof(name),
1198 "%s-%s.%s", base_name, sha1_to_hex(sha1), "pack");
1199 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1200 if (ofd < 0)
1201 die("unable to open %s (%s)", name, strerror(errno));
1202 if (copy_fd(ifd, ofd))
1203 exit(1);
1204 close(ifd);
1206 snprintf(name, sizeof(name),
1207 "%s-%s.%s", base_name, sha1_to_hex(sha1), "idx");
1208 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1209 if (ofd < 0)
1210 die("unable to open %s (%s)", name, strerror(errno));
1211 if (copy_fd(ifd_ix, ofd))
1212 exit(1);
1213 close(ifd_ix);
1214 puts(sha1_to_hex(sha1));
1217 return 1;
1220 static void setup_progress_signal(void)
1222 struct sigaction sa;
1223 struct itimerval v;
1225 memset(&sa, 0, sizeof(sa));
1226 sa.sa_handler = progress_interval;
1227 sigemptyset(&sa.sa_mask);
1228 sa.sa_flags = SA_RESTART;
1229 sigaction(SIGALRM, &sa, NULL);
1231 v.it_interval.tv_sec = 1;
1232 v.it_interval.tv_usec = 0;
1233 v.it_value = v.it_interval;
1234 setitimer(ITIMER_REAL, &v, NULL);
1237 static int git_pack_config(const char *k, const char *v)
1239 if(!strcmp(k, "pack.window")) {
1240 window = git_config_int(k, v);
1241 return 0;
1243 return git_default_config(k, v);
1246 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1248 SHA_CTX ctx;
1249 char line[40 + 1 + PATH_MAX + 2];
1250 int depth = 10, pack_to_stdout = 0;
1251 struct object_entry **list;
1252 int num_preferred_base = 0;
1253 int i;
1255 git_config(git_pack_config);
1257 progress = isatty(2);
1258 for (i = 1; i < argc; i++) {
1259 const char *arg = argv[i];
1261 if (*arg == '-') {
1262 if (!strcmp("--non-empty", arg)) {
1263 non_empty = 1;
1264 continue;
1266 if (!strcmp("--local", arg)) {
1267 local = 1;
1268 continue;
1270 if (!strcmp("--progress", arg)) {
1271 progress = 1;
1272 continue;
1274 if (!strcmp("--incremental", arg)) {
1275 incremental = 1;
1276 continue;
1278 if (!strncmp("--window=", arg, 9)) {
1279 char *end;
1280 window = strtoul(arg+9, &end, 0);
1281 if (!arg[9] || *end)
1282 usage(pack_usage);
1283 continue;
1285 if (!strncmp("--depth=", arg, 8)) {
1286 char *end;
1287 depth = strtoul(arg+8, &end, 0);
1288 if (!arg[8] || *end)
1289 usage(pack_usage);
1290 continue;
1292 if (!strcmp("--progress", arg)) {
1293 progress = 1;
1294 continue;
1296 if (!strcmp("-q", arg)) {
1297 progress = 0;
1298 continue;
1300 if (!strcmp("--no-reuse-delta", arg)) {
1301 no_reuse_delta = 1;
1302 continue;
1304 if (!strcmp("--stdout", arg)) {
1305 pack_to_stdout = 1;
1306 continue;
1308 usage(pack_usage);
1310 if (base_name)
1311 usage(pack_usage);
1312 base_name = arg;
1315 if (pack_to_stdout != !base_name)
1316 usage(pack_usage);
1318 prepare_packed_git();
1320 if (progress) {
1321 fprintf(stderr, "Generating pack...\n");
1322 setup_progress_signal();
1325 for (;;) {
1326 unsigned char sha1[20];
1327 unsigned hash;
1329 if (!fgets(line, sizeof(line), stdin)) {
1330 if (feof(stdin))
1331 break;
1332 if (!ferror(stdin))
1333 die("fgets returned NULL, not EOF, not error!");
1334 if (errno != EINTR)
1335 die("fgets: %s", strerror(errno));
1336 clearerr(stdin);
1337 continue;
1340 if (line[0] == '-') {
1341 if (get_sha1_hex(line+1, sha1))
1342 die("expected edge sha1, got garbage:\n %s",
1343 line+1);
1344 if (num_preferred_base++ < window)
1345 add_preferred_base(sha1);
1346 continue;
1348 if (get_sha1_hex(line, sha1))
1349 die("expected sha1, got garbage:\n %s", line);
1350 hash = name_hash(line+41);
1351 add_preferred_base_object(line+41, hash);
1352 add_object_entry(sha1, hash, 0);
1354 if (progress)
1355 fprintf(stderr, "Done counting %d objects.\n", nr_objects);
1356 sorted_by_sha = create_final_object_list();
1357 if (non_empty && !nr_result)
1358 return 0;
1360 SHA1_Init(&ctx);
1361 list = sorted_by_sha;
1362 for (i = 0; i < nr_result; i++) {
1363 struct object_entry *entry = *list++;
1364 SHA1_Update(&ctx, entry->sha1, 20);
1366 SHA1_Final(object_list_sha1, &ctx);
1367 if (progress && (nr_objects != nr_result))
1368 fprintf(stderr, "Result has %d objects.\n", nr_result);
1370 if (reuse_cached_pack(object_list_sha1, pack_to_stdout))
1372 else {
1373 if (nr_result)
1374 prepare_pack(window, depth);
1375 if (progress && pack_to_stdout) {
1376 /* the other end usually displays progress itself */
1377 struct itimerval v = {{0,},};
1378 setitimer(ITIMER_REAL, &v, NULL);
1379 signal(SIGALRM, SIG_IGN );
1380 progress_update = 0;
1382 write_pack_file();
1383 if (!pack_to_stdout) {
1384 write_index_file();
1385 puts(sha1_to_hex(object_list_sha1));
1388 if (progress)
1389 fprintf(stderr, "Total %d, written %d (delta %d), reused %d (delta %d)\n",
1390 nr_result, written, written_delta, reused, reused_delta);
1391 return 0;