pack-objects: check pack.window for default window size
[git/peff.git] / pack-objects.c
blob861c7f08ff1b68eee141442411b4706a1bcc2518
1 #include "cache.h"
2 #include "object.h"
3 #include "blob.h"
4 #include "commit.h"
5 #include "tag.h"
6 #include "tree.h"
7 #include "delta.h"
8 #include "pack.h"
9 #include "csum-file.h"
10 #include "tree-walk.h"
11 #include <sys/time.h>
12 #include <signal.h>
14 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";
16 struct object_entry {
17 unsigned char sha1[20];
18 unsigned long size; /* uncompressed size */
19 unsigned long offset; /* offset into the final pack file;
20 * nonzero if already written.
22 unsigned int depth; /* delta depth */
23 unsigned int delta_limit; /* base adjustment for in-pack delta */
24 unsigned int hash; /* name hint hash */
25 enum object_type type;
26 enum object_type in_pack_type; /* could be delta */
27 unsigned long delta_size; /* delta data size (uncompressed) */
28 struct object_entry *delta; /* delta base object */
29 struct packed_git *in_pack; /* already in pack */
30 unsigned int in_pack_offset;
31 struct object_entry *delta_child; /* deltified objects who bases me */
32 struct object_entry *delta_sibling; /* other deltified objects who
33 * uses the same base as me
35 int preferred_base; /* we do not pack this, but is encouraged to
36 * be used as the base objectto delta huge
37 * objects against.
42 * Objects we are going to pack are collected in objects array (dynamically
43 * expanded). nr_objects & nr_alloc controls this array. They are stored
44 * in the order we see -- typically rev-list --objects order that gives us
45 * nice "minimum seek" order.
47 * sorted-by-sha ans sorted-by-type are arrays of pointers that point at
48 * elements in the objects array. The former is used to build the pack
49 * index (lists object names in the ascending order to help offset lookup),
50 * and the latter is used to group similar things together by try_delta()
51 * heuristics.
54 static unsigned char object_list_sha1[20];
55 static int non_empty = 0;
56 static int no_reuse_delta = 0;
57 static int local = 0;
58 static int incremental = 0;
59 static struct object_entry **sorted_by_sha, **sorted_by_type;
60 static struct object_entry *objects = NULL;
61 static int nr_objects = 0, nr_alloc = 0, nr_result = 0;
62 static const char *base_name;
63 static unsigned char pack_file_sha1[20];
64 static int progress = 1;
65 static volatile sig_atomic_t progress_update = 0;
66 static int window = 10;
69 * The object names in objects array are hashed with this hashtable,
70 * to help looking up the entry by object name. Binary search from
71 * sorted_by_sha is also possible but this was easier to code and faster.
72 * This hashtable is built after all the objects are seen.
74 static int *object_ix = NULL;
75 static int object_ix_hashsz = 0;
78 * Pack index for existing packs give us easy access to the offsets into
79 * corresponding pack file where each object's data starts, but the entries
80 * do not store the size of the compressed representation (uncompressed
81 * size is easily available by examining the pack entry header). We build
82 * a hashtable of existing packs (pack_revindex), and keep reverse index
83 * here -- pack index file is sorted by object name mapping to offset; this
84 * pack_revindex[].revindex array is an ordered list of offsets, so if you
85 * know the offset of an object, next offset is where its packed
86 * representation ends.
88 struct pack_revindex {
89 struct packed_git *p;
90 unsigned long *revindex;
91 } *pack_revindex = NULL;
92 static int pack_revindex_hashsz = 0;
95 * stats
97 static int written = 0;
98 static int written_delta = 0;
99 static int reused = 0;
100 static int reused_delta = 0;
102 static int pack_revindex_ix(struct packed_git *p)
104 unsigned long ui = (unsigned long)p;
105 int i;
107 ui = ui ^ (ui >> 16); /* defeat structure alignment */
108 i = (int)(ui % pack_revindex_hashsz);
109 while (pack_revindex[i].p) {
110 if (pack_revindex[i].p == p)
111 return i;
112 if (++i == pack_revindex_hashsz)
113 i = 0;
115 return -1 - i;
118 static void prepare_pack_ix(void)
120 int num;
121 struct packed_git *p;
122 for (num = 0, p = packed_git; p; p = p->next)
123 num++;
124 if (!num)
125 return;
126 pack_revindex_hashsz = num * 11;
127 pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
128 for (p = packed_git; p; p = p->next) {
129 num = pack_revindex_ix(p);
130 num = - 1 - num;
131 pack_revindex[num].p = p;
133 /* revindex elements are lazily initialized */
136 static int cmp_offset(const void *a_, const void *b_)
138 unsigned long a = *(unsigned long *) a_;
139 unsigned long b = *(unsigned long *) b_;
140 if (a < b)
141 return -1;
142 else if (a == b)
143 return 0;
144 else
145 return 1;
149 * Ordered list of offsets of objects in the pack.
151 static void prepare_pack_revindex(struct pack_revindex *rix)
153 struct packed_git *p = rix->p;
154 int num_ent = num_packed_objects(p);
155 int i;
156 void *index = p->index_base + 256;
158 rix->revindex = xmalloc(sizeof(unsigned long) * (num_ent + 1));
159 for (i = 0; i < num_ent; i++) {
160 unsigned int hl = *((unsigned int *)((char *) index + 24*i));
161 rix->revindex[i] = ntohl(hl);
163 /* This knows the pack format -- the 20-byte trailer
164 * follows immediately after the last object data.
166 rix->revindex[num_ent] = p->pack_size - 20;
167 qsort(rix->revindex, num_ent, sizeof(unsigned long), cmp_offset);
170 static unsigned long find_packed_object_size(struct packed_git *p,
171 unsigned long ofs)
173 int num;
174 int lo, hi;
175 struct pack_revindex *rix;
176 unsigned long *revindex;
177 num = pack_revindex_ix(p);
178 if (num < 0)
179 die("internal error: pack revindex uninitialized");
180 rix = &pack_revindex[num];
181 if (!rix->revindex)
182 prepare_pack_revindex(rix);
183 revindex = rix->revindex;
184 lo = 0;
185 hi = num_packed_objects(p) + 1;
186 do {
187 int mi = (lo + hi) / 2;
188 if (revindex[mi] == ofs) {
189 return revindex[mi+1] - ofs;
191 else if (ofs < revindex[mi])
192 hi = mi;
193 else
194 lo = mi + 1;
195 } while (lo < hi);
196 die("internal error: pack revindex corrupt");
199 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
201 unsigned long othersize, delta_size;
202 char type[10];
203 void *otherbuf = read_sha1_file(entry->delta->sha1, type, &othersize);
204 void *delta_buf;
206 if (!otherbuf)
207 die("unable to read %s", sha1_to_hex(entry->delta->sha1));
208 delta_buf = diff_delta(otherbuf, othersize,
209 buf, size, &delta_size, 0);
210 if (!delta_buf || delta_size != entry->delta_size)
211 die("delta size changed");
212 free(buf);
213 free(otherbuf);
214 return delta_buf;
218 * The per-object header is a pretty dense thing, which is
219 * - first byte: low four bits are "size", then three bits of "type",
220 * and the high bit is "size continues".
221 * - each byte afterwards: low seven bits are size continuation,
222 * with the high bit being "size continues"
224 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
226 int n = 1;
227 unsigned char c;
229 if (type < OBJ_COMMIT || type > OBJ_DELTA)
230 die("bad type %d", type);
232 c = (type << 4) | (size & 15);
233 size >>= 4;
234 while (size) {
235 *hdr++ = c | 0x80;
236 c = size & 0x7f;
237 size >>= 7;
238 n++;
240 *hdr = c;
241 return n;
244 static unsigned long write_object(struct sha1file *f,
245 struct object_entry *entry)
247 unsigned long size;
248 char type[10];
249 void *buf;
250 unsigned char header[10];
251 unsigned hdrlen, datalen;
252 enum object_type obj_type;
253 int to_reuse = 0;
255 if (entry->preferred_base)
256 return 0;
258 obj_type = entry->type;
259 if (! entry->in_pack)
260 to_reuse = 0; /* can't reuse what we don't have */
261 else if (obj_type == OBJ_DELTA)
262 to_reuse = 1; /* check_object() decided it for us */
263 else if (obj_type != entry->in_pack_type)
264 to_reuse = 0; /* pack has delta which is unusable */
265 else if (entry->delta)
266 to_reuse = 0; /* we want to pack afresh */
267 else
268 to_reuse = 1; /* we have it in-pack undeltified,
269 * and we do not need to deltify it.
272 if (! to_reuse) {
273 buf = read_sha1_file(entry->sha1, type, &size);
274 if (!buf)
275 die("unable to read %s", sha1_to_hex(entry->sha1));
276 if (size != entry->size)
277 die("object %s size inconsistency (%lu vs %lu)",
278 sha1_to_hex(entry->sha1), size, entry->size);
279 if (entry->delta) {
280 buf = delta_against(buf, size, entry);
281 size = entry->delta_size;
282 obj_type = OBJ_DELTA;
285 * The object header is a byte of 'type' followed by zero or
286 * more bytes of length. For deltas, the 20 bytes of delta
287 * sha1 follows that.
289 hdrlen = encode_header(obj_type, size, header);
290 sha1write(f, header, hdrlen);
292 if (entry->delta) {
293 sha1write(f, entry->delta, 20);
294 hdrlen += 20;
296 datalen = sha1write_compressed(f, buf, size);
297 free(buf);
299 else {
300 struct packed_git *p = entry->in_pack;
301 use_packed_git(p);
303 datalen = find_packed_object_size(p, entry->in_pack_offset);
304 buf = (char *) p->pack_base + entry->in_pack_offset;
305 sha1write(f, buf, datalen);
306 unuse_packed_git(p);
307 hdrlen = 0; /* not really */
308 if (obj_type == OBJ_DELTA)
309 reused_delta++;
310 reused++;
312 if (obj_type == OBJ_DELTA)
313 written_delta++;
314 written++;
315 return hdrlen + datalen;
318 static unsigned long write_one(struct sha1file *f,
319 struct object_entry *e,
320 unsigned long offset)
322 if (e->offset)
323 /* offset starts from header size and cannot be zero
324 * if it is written already.
326 return offset;
327 e->offset = offset;
328 offset += write_object(f, e);
329 /* if we are deltified, write out its base object. */
330 if (e->delta)
331 offset = write_one(f, e->delta, offset);
332 return offset;
335 static void write_pack_file(void)
337 int i;
338 struct sha1file *f;
339 unsigned long offset;
340 struct pack_header hdr;
341 unsigned last_percent = 999;
342 int do_progress = 0;
344 if (!base_name)
345 f = sha1fd(1, "<stdout>");
346 else {
347 f = sha1create("%s-%s.%s", base_name,
348 sha1_to_hex(object_list_sha1), "pack");
349 do_progress = progress;
351 if (do_progress)
352 fprintf(stderr, "Writing %d objects.\n", nr_result);
354 hdr.hdr_signature = htonl(PACK_SIGNATURE);
355 hdr.hdr_version = htonl(PACK_VERSION);
356 hdr.hdr_entries = htonl(nr_result);
357 sha1write(f, &hdr, sizeof(hdr));
358 offset = sizeof(hdr);
359 if (!nr_result)
360 goto done;
361 for (i = 0; i < nr_objects; i++) {
362 offset = write_one(f, objects + i, offset);
363 if (do_progress) {
364 unsigned percent = written * 100 / nr_result;
365 if (progress_update || percent != last_percent) {
366 fprintf(stderr, "%4u%% (%u/%u) done\r",
367 percent, written, nr_result);
368 progress_update = 0;
369 last_percent = percent;
373 if (do_progress)
374 fputc('\n', stderr);
375 done:
376 sha1close(f, pack_file_sha1, 1);
379 static void write_index_file(void)
381 int i;
382 struct sha1file *f = sha1create("%s-%s.%s", base_name,
383 sha1_to_hex(object_list_sha1), "idx");
384 struct object_entry **list = sorted_by_sha;
385 struct object_entry **last = list + nr_result;
386 unsigned int array[256];
389 * Write the first-level table (the list is sorted,
390 * but we use a 256-entry lookup to be able to avoid
391 * having to do eight extra binary search iterations).
393 for (i = 0; i < 256; i++) {
394 struct object_entry **next = list;
395 while (next < last) {
396 struct object_entry *entry = *next;
397 if (entry->sha1[0] != i)
398 break;
399 next++;
401 array[i] = htonl(next - sorted_by_sha);
402 list = next;
404 sha1write(f, array, 256 * sizeof(int));
407 * Write the actual SHA1 entries..
409 list = sorted_by_sha;
410 for (i = 0; i < nr_result; i++) {
411 struct object_entry *entry = *list++;
412 unsigned int offset = htonl(entry->offset);
413 sha1write(f, &offset, 4);
414 sha1write(f, entry->sha1, 20);
416 sha1write(f, pack_file_sha1, 20);
417 sha1close(f, NULL, 1);
420 static int locate_object_entry_hash(const unsigned char *sha1)
422 int i;
423 unsigned int ui;
424 memcpy(&ui, sha1, sizeof(unsigned int));
425 i = ui % object_ix_hashsz;
426 while (0 < object_ix[i]) {
427 if (!memcmp(sha1, objects[object_ix[i]-1].sha1, 20))
428 return i;
429 if (++i == object_ix_hashsz)
430 i = 0;
432 return -1 - i;
435 static struct object_entry *locate_object_entry(const unsigned char *sha1)
437 int i;
439 if (!object_ix_hashsz)
440 return NULL;
442 i = locate_object_entry_hash(sha1);
443 if (0 <= i)
444 return &objects[object_ix[i]-1];
445 return NULL;
448 static void rehash_objects(void)
450 int i;
451 struct object_entry *oe;
453 object_ix_hashsz = nr_objects * 3;
454 if (object_ix_hashsz < 1024)
455 object_ix_hashsz = 1024;
456 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
457 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
458 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
459 int ix = locate_object_entry_hash(oe->sha1);
460 if (0 <= ix)
461 continue;
462 ix = -1 - ix;
463 object_ix[ix] = i + 1;
467 static unsigned name_hash(const char *name)
469 unsigned char c;
470 unsigned hash = 0;
473 * This effectively just creates a sortable number from the
474 * last sixteen non-whitespace characters. Last characters
475 * count "most", so things that end in ".c" sort together.
477 while ((c = *name++) != 0) {
478 if (isspace(c))
479 continue;
480 hash = (hash >> 2) + (c << 24);
482 return hash;
485 static int add_object_entry(const unsigned char *sha1, unsigned hash, int exclude)
487 unsigned int idx = nr_objects;
488 struct object_entry *entry;
489 struct packed_git *p;
490 unsigned int found_offset = 0;
491 struct packed_git *found_pack = NULL;
492 int ix, status = 0;
494 if (!exclude) {
495 for (p = packed_git; p; p = p->next) {
496 struct pack_entry e;
497 if (find_pack_entry_one(sha1, &e, p)) {
498 if (incremental)
499 return 0;
500 if (local && !p->pack_local)
501 return 0;
502 if (!found_pack) {
503 found_offset = e.offset;
504 found_pack = e.p;
509 if ((entry = locate_object_entry(sha1)) != NULL)
510 goto already_added;
512 if (idx >= nr_alloc) {
513 unsigned int needed = (idx + 1024) * 3 / 2;
514 objects = xrealloc(objects, needed * sizeof(*entry));
515 nr_alloc = needed;
517 entry = objects + idx;
518 nr_objects = idx + 1;
519 memset(entry, 0, sizeof(*entry));
520 memcpy(entry->sha1, sha1, 20);
521 entry->hash = hash;
523 if (object_ix_hashsz * 3 <= nr_objects * 4)
524 rehash_objects();
525 else {
526 ix = locate_object_entry_hash(entry->sha1);
527 if (0 <= ix)
528 die("internal error in object hashing.");
529 object_ix[-1 - ix] = idx + 1;
531 status = 1;
533 already_added:
534 if (progress_update) {
535 fprintf(stderr, "Counting objects...%d\r", nr_objects);
536 progress_update = 0;
538 if (exclude)
539 entry->preferred_base = 1;
540 else {
541 if (found_pack) {
542 entry->in_pack = found_pack;
543 entry->in_pack_offset = found_offset;
546 return status;
549 struct pbase_tree_cache {
550 unsigned char sha1[20];
551 int ref;
552 int temporary;
553 void *tree_data;
554 unsigned long tree_size;
557 static struct pbase_tree_cache *(pbase_tree_cache[256]);
558 static int pbase_tree_cache_ix(const unsigned char *sha1)
560 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
562 static int pbase_tree_cache_ix_incr(int ix)
564 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
567 static struct pbase_tree {
568 struct pbase_tree *next;
569 /* This is a phony "cache" entry; we are not
570 * going to evict it nor find it through _get()
571 * mechanism -- this is for the toplevel node that
572 * would almost always change with any commit.
574 struct pbase_tree_cache pcache;
575 } *pbase_tree;
577 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
579 struct pbase_tree_cache *ent, *nent;
580 void *data;
581 unsigned long size;
582 char type[20];
583 int neigh;
584 int my_ix = pbase_tree_cache_ix(sha1);
585 int available_ix = -1;
587 /* pbase-tree-cache acts as a limited hashtable.
588 * your object will be found at your index or within a few
589 * slots after that slot if it is cached.
591 for (neigh = 0; neigh < 8; neigh++) {
592 ent = pbase_tree_cache[my_ix];
593 if (ent && !memcmp(ent->sha1, sha1, 20)) {
594 ent->ref++;
595 return ent;
597 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
598 ((0 <= available_ix) &&
599 (!ent && pbase_tree_cache[available_ix])))
600 available_ix = my_ix;
601 if (!ent)
602 break;
603 my_ix = pbase_tree_cache_ix_incr(my_ix);
606 /* Did not find one. Either we got a bogus request or
607 * we need to read and perhaps cache.
609 data = read_sha1_file(sha1, type, &size);
610 if (!data)
611 return NULL;
612 if (strcmp(type, tree_type)) {
613 free(data);
614 return NULL;
617 /* We need to either cache or return a throwaway copy */
619 if (available_ix < 0)
620 ent = NULL;
621 else {
622 ent = pbase_tree_cache[available_ix];
623 my_ix = available_ix;
626 if (!ent) {
627 nent = xmalloc(sizeof(*nent));
628 nent->temporary = (available_ix < 0);
630 else {
631 /* evict and reuse */
632 free(ent->tree_data);
633 nent = ent;
635 memcpy(nent->sha1, sha1, 20);
636 nent->tree_data = data;
637 nent->tree_size = size;
638 nent->ref = 1;
639 if (!nent->temporary)
640 pbase_tree_cache[my_ix] = nent;
641 return nent;
644 static void pbase_tree_put(struct pbase_tree_cache *cache)
646 if (!cache->temporary) {
647 cache->ref--;
648 return;
650 free(cache->tree_data);
651 free(cache);
654 static int name_cmp_len(const char *name)
656 int i;
657 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
659 return i;
662 static void add_pbase_object(struct tree_desc *tree,
663 const char *name,
664 int cmplen,
665 const char *fullname)
667 struct name_entry entry;
669 while (tree_entry(tree,&entry)) {
670 unsigned long size;
671 char type[20];
673 if (entry.pathlen != cmplen ||
674 memcmp(entry.path, name, cmplen) ||
675 !has_sha1_file(entry.sha1) ||
676 sha1_object_info(entry.sha1, type, &size))
677 continue;
678 if (name[cmplen] != '/') {
679 unsigned hash = name_hash(fullname);
680 add_object_entry(entry.sha1, hash, 1);
681 return;
683 if (!strcmp(type, tree_type)) {
684 struct tree_desc sub;
685 struct pbase_tree_cache *tree;
686 const char *down = name+cmplen+1;
687 int downlen = name_cmp_len(down);
689 tree = pbase_tree_get(entry.sha1);
690 if (!tree)
691 return;
692 sub.buf = tree->tree_data;
693 sub.size = tree->tree_size;
695 add_pbase_object(&sub, down, downlen, fullname);
696 pbase_tree_put(tree);
701 static unsigned *done_pbase_paths;
702 static int done_pbase_paths_num;
703 static int done_pbase_paths_alloc;
704 static int done_pbase_path_pos(unsigned hash)
706 int lo = 0;
707 int hi = done_pbase_paths_num;
708 while (lo < hi) {
709 int mi = (hi + lo) / 2;
710 if (done_pbase_paths[mi] == hash)
711 return mi;
712 if (done_pbase_paths[mi] < hash)
713 hi = mi;
714 else
715 lo = mi + 1;
717 return -lo-1;
720 static int check_pbase_path(unsigned hash)
722 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
723 if (0 <= pos)
724 return 1;
725 pos = -pos - 1;
726 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
727 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
728 done_pbase_paths = xrealloc(done_pbase_paths,
729 done_pbase_paths_alloc *
730 sizeof(unsigned));
732 done_pbase_paths_num++;
733 if (pos < done_pbase_paths_num)
734 memmove(done_pbase_paths + pos + 1,
735 done_pbase_paths + pos,
736 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
737 done_pbase_paths[pos] = hash;
738 return 0;
741 static void add_preferred_base_object(char *name, unsigned hash)
743 struct pbase_tree *it;
744 int cmplen = name_cmp_len(name);
746 if (check_pbase_path(hash))
747 return;
749 for (it = pbase_tree; it; it = it->next) {
750 if (cmplen == 0) {
751 hash = name_hash("");
752 add_object_entry(it->pcache.sha1, hash, 1);
754 else {
755 struct tree_desc tree;
756 tree.buf = it->pcache.tree_data;
757 tree.size = it->pcache.tree_size;
758 add_pbase_object(&tree, name, cmplen, name);
763 static void add_preferred_base(unsigned char *sha1)
765 struct pbase_tree *it;
766 void *data;
767 unsigned long size;
768 unsigned char tree_sha1[20];
770 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
771 if (!data)
772 return;
774 for (it = pbase_tree; it; it = it->next) {
775 if (!memcmp(it->pcache.sha1, tree_sha1, 20)) {
776 free(data);
777 return;
781 it = xcalloc(1, sizeof(*it));
782 it->next = pbase_tree;
783 pbase_tree = it;
785 memcpy(it->pcache.sha1, tree_sha1, 20);
786 it->pcache.tree_data = data;
787 it->pcache.tree_size = size;
790 static void check_object(struct object_entry *entry)
792 char type[20];
794 if (entry->in_pack && !entry->preferred_base) {
795 unsigned char base[20];
796 unsigned long size;
797 struct object_entry *base_entry;
799 /* We want in_pack_type even if we do not reuse delta.
800 * There is no point not reusing non-delta representations.
802 check_reuse_pack_delta(entry->in_pack,
803 entry->in_pack_offset,
804 base, &size,
805 &entry->in_pack_type);
807 /* Check if it is delta, and the base is also an object
808 * we are going to pack. If so we will reuse the existing
809 * delta.
811 if (!no_reuse_delta &&
812 entry->in_pack_type == OBJ_DELTA &&
813 (base_entry = locate_object_entry(base)) &&
814 (!base_entry->preferred_base)) {
816 /* Depth value does not matter - find_deltas()
817 * will never consider reused delta as the
818 * base object to deltify other objects
819 * against, in order to avoid circular deltas.
822 /* uncompressed size of the delta data */
823 entry->size = entry->delta_size = size;
824 entry->delta = base_entry;
825 entry->type = OBJ_DELTA;
827 entry->delta_sibling = base_entry->delta_child;
828 base_entry->delta_child = entry;
830 return;
832 /* Otherwise we would do the usual */
835 if (sha1_object_info(entry->sha1, type, &entry->size))
836 die("unable to get type of object %s",
837 sha1_to_hex(entry->sha1));
839 if (!strcmp(type, commit_type)) {
840 entry->type = OBJ_COMMIT;
841 } else if (!strcmp(type, tree_type)) {
842 entry->type = OBJ_TREE;
843 } else if (!strcmp(type, blob_type)) {
844 entry->type = OBJ_BLOB;
845 } else if (!strcmp(type, tag_type)) {
846 entry->type = OBJ_TAG;
847 } else
848 die("unable to pack object %s of type %s",
849 sha1_to_hex(entry->sha1), type);
852 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
854 struct object_entry *child = me->delta_child;
855 unsigned int m = n;
856 while (child) {
857 unsigned int c = check_delta_limit(child, n + 1);
858 if (m < c)
859 m = c;
860 child = child->delta_sibling;
862 return m;
865 static void get_object_details(void)
867 int i;
868 struct object_entry *entry;
870 prepare_pack_ix();
871 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
872 check_object(entry);
874 if (nr_objects == nr_result) {
876 * Depth of objects that depend on the entry -- this
877 * is subtracted from depth-max to break too deep
878 * delta chain because of delta data reusing.
879 * However, we loosen this restriction when we know we
880 * are creating a thin pack -- it will have to be
881 * expanded on the other end anyway, so do not
882 * artificially cut the delta chain and let it go as
883 * deep as it wants.
885 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
886 if (!entry->delta && entry->delta_child)
887 entry->delta_limit =
888 check_delta_limit(entry, 1);
892 typedef int (*entry_sort_t)(const struct object_entry *, const struct object_entry *);
894 static entry_sort_t current_sort;
896 static int sort_comparator(const void *_a, const void *_b)
898 struct object_entry *a = *(struct object_entry **)_a;
899 struct object_entry *b = *(struct object_entry **)_b;
900 return current_sort(a,b);
903 static struct object_entry **create_sorted_list(entry_sort_t sort)
905 struct object_entry **list = xmalloc(nr_objects * sizeof(struct object_entry *));
906 int i;
908 for (i = 0; i < nr_objects; i++)
909 list[i] = objects + i;
910 current_sort = sort;
911 qsort(list, nr_objects, sizeof(struct object_entry *), sort_comparator);
912 return list;
915 static int sha1_sort(const struct object_entry *a, const struct object_entry *b)
917 return memcmp(a->sha1, b->sha1, 20);
920 static struct object_entry **create_final_object_list(void)
922 struct object_entry **list;
923 int i, j;
925 for (i = nr_result = 0; i < nr_objects; i++)
926 if (!objects[i].preferred_base)
927 nr_result++;
928 list = xmalloc(nr_result * sizeof(struct object_entry *));
929 for (i = j = 0; i < nr_objects; i++) {
930 if (!objects[i].preferred_base)
931 list[j++] = objects + i;
933 current_sort = sha1_sort;
934 qsort(list, nr_result, sizeof(struct object_entry *), sort_comparator);
935 return list;
938 static int type_size_sort(const struct object_entry *a, const struct object_entry *b)
940 if (a->type < b->type)
941 return -1;
942 if (a->type > b->type)
943 return 1;
944 if (a->hash < b->hash)
945 return -1;
946 if (a->hash > b->hash)
947 return 1;
948 if (a->preferred_base < b->preferred_base)
949 return -1;
950 if (a->preferred_base > b->preferred_base)
951 return 1;
952 if (a->size < b->size)
953 return -1;
954 if (a->size > b->size)
955 return 1;
956 return a < b ? -1 : (a > b);
959 struct unpacked {
960 struct object_entry *entry;
961 void *data;
962 struct delta_index *index;
966 * We search for deltas _backwards_ in a list sorted by type and
967 * by size, so that we see progressively smaller and smaller files.
968 * That's because we prefer deltas to be from the bigger file
969 * to the smaller - deletes are potentially cheaper, but perhaps
970 * more importantly, the bigger file is likely the more recent
971 * one.
973 static int try_delta(struct unpacked *trg, struct unpacked *src,
974 unsigned max_depth)
976 struct object_entry *trg_entry = trg->entry;
977 struct object_entry *src_entry = src->entry;
978 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
979 char type[10];
980 void *delta_buf;
982 /* Don't bother doing diffs between different types */
983 if (trg_entry->type != src_entry->type)
984 return -1;
986 /* We do not compute delta to *create* objects we are not
987 * going to pack.
989 if (trg_entry->preferred_base)
990 return -1;
993 * We do not bother to try a delta that we discarded
994 * on an earlier try, but only when reusing delta data.
996 if (!no_reuse_delta && trg_entry->in_pack &&
997 trg_entry->in_pack == src_entry->in_pack)
998 return 0;
1001 * If the current object is at pack edge, take the depth the
1002 * objects that depend on the current object into account --
1003 * otherwise they would become too deep.
1005 if (trg_entry->delta_child) {
1006 if (max_depth <= trg_entry->delta_limit)
1007 return 0;
1008 max_depth -= trg_entry->delta_limit;
1010 if (src_entry->depth >= max_depth)
1011 return 0;
1013 /* Now some size filtering heuristics. */
1014 trg_size = trg_entry->size;
1015 max_size = trg_size/2 - 20;
1016 max_size = max_size * (max_depth - src_entry->depth) / max_depth;
1017 if (max_size == 0)
1018 return 0;
1019 if (trg_entry->delta && trg_entry->delta_size <= max_size)
1020 max_size = trg_entry->delta_size-1;
1021 src_size = src_entry->size;
1022 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1023 if (sizediff >= max_size)
1024 return 0;
1026 /* Load data if not already done */
1027 if (!trg->data) {
1028 trg->data = read_sha1_file(trg_entry->sha1, type, &sz);
1029 if (sz != trg_size)
1030 die("object %s inconsistent object length (%lu vs %lu)",
1031 sha1_to_hex(trg_entry->sha1), sz, trg_size);
1033 if (!src->data) {
1034 src->data = read_sha1_file(src_entry->sha1, type, &sz);
1035 if (sz != src_size)
1036 die("object %s inconsistent object length (%lu vs %lu)",
1037 sha1_to_hex(src_entry->sha1), sz, src_size);
1039 if (!src->index) {
1040 src->index = create_delta_index(src->data, src_size);
1041 if (!src->index)
1042 die("out of memory");
1045 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1046 if (!delta_buf)
1047 return 0;
1049 trg_entry->delta = src_entry;
1050 trg_entry->delta_size = delta_size;
1051 trg_entry->depth = src_entry->depth + 1;
1052 free(delta_buf);
1053 return 1;
1056 static void progress_interval(int signum)
1058 progress_update = 1;
1061 static void find_deltas(struct object_entry **list, int window, int depth)
1063 int i, idx;
1064 unsigned int array_size = window * sizeof(struct unpacked);
1065 struct unpacked *array = xmalloc(array_size);
1066 unsigned processed = 0;
1067 unsigned last_percent = 999;
1069 memset(array, 0, array_size);
1070 i = nr_objects;
1071 idx = 0;
1072 if (progress)
1073 fprintf(stderr, "Deltifying %d objects.\n", nr_result);
1075 while (--i >= 0) {
1076 struct object_entry *entry = list[i];
1077 struct unpacked *n = array + idx;
1078 int j;
1080 if (!entry->preferred_base)
1081 processed++;
1083 if (progress) {
1084 unsigned percent = processed * 100 / nr_result;
1085 if (percent != last_percent || progress_update) {
1086 fprintf(stderr, "%4u%% (%u/%u) done\r",
1087 percent, processed, nr_result);
1088 progress_update = 0;
1089 last_percent = percent;
1093 if (entry->delta)
1094 /* This happens if we decided to reuse existing
1095 * delta from a pack. "!no_reuse_delta &&" is implied.
1097 continue;
1099 if (entry->size < 50)
1100 continue;
1101 free_delta_index(n->index);
1102 n->index = NULL;
1103 free(n->data);
1104 n->data = NULL;
1105 n->entry = entry;
1107 j = window;
1108 while (--j > 0) {
1109 unsigned int other_idx = idx + j;
1110 struct unpacked *m;
1111 if (other_idx >= window)
1112 other_idx -= window;
1113 m = array + other_idx;
1114 if (!m->entry)
1115 break;
1116 if (try_delta(n, m, depth) < 0)
1117 break;
1119 /* if we made n a delta, and if n is already at max
1120 * depth, leaving it in the window is pointless. we
1121 * should evict it first.
1123 if (entry->delta && depth <= entry->depth)
1124 continue;
1126 idx++;
1127 if (idx >= window)
1128 idx = 0;
1131 if (progress)
1132 fputc('\n', stderr);
1134 for (i = 0; i < window; ++i) {
1135 free_delta_index(array[i].index);
1136 free(array[i].data);
1138 free(array);
1141 static void prepare_pack(int window, int depth)
1143 get_object_details();
1144 sorted_by_type = create_sorted_list(type_size_sort);
1145 if (window && depth)
1146 find_deltas(sorted_by_type, window+1, depth);
1149 static int reuse_cached_pack(unsigned char *sha1, int pack_to_stdout)
1151 static const char cache[] = "pack-cache/pack-%s.%s";
1152 char *cached_pack, *cached_idx;
1153 int ifd, ofd, ifd_ix = -1;
1155 cached_pack = git_path(cache, sha1_to_hex(sha1), "pack");
1156 ifd = open(cached_pack, O_RDONLY);
1157 if (ifd < 0)
1158 return 0;
1160 if (!pack_to_stdout) {
1161 cached_idx = git_path(cache, sha1_to_hex(sha1), "idx");
1162 ifd_ix = open(cached_idx, O_RDONLY);
1163 if (ifd_ix < 0) {
1164 close(ifd);
1165 return 0;
1169 if (progress)
1170 fprintf(stderr, "Reusing %d objects pack %s\n", nr_objects,
1171 sha1_to_hex(sha1));
1173 if (pack_to_stdout) {
1174 if (copy_fd(ifd, 1))
1175 exit(1);
1176 close(ifd);
1178 else {
1179 char name[PATH_MAX];
1180 snprintf(name, sizeof(name),
1181 "%s-%s.%s", base_name, sha1_to_hex(sha1), "pack");
1182 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1183 if (ofd < 0)
1184 die("unable to open %s (%s)", name, strerror(errno));
1185 if (copy_fd(ifd, ofd))
1186 exit(1);
1187 close(ifd);
1189 snprintf(name, sizeof(name),
1190 "%s-%s.%s", base_name, sha1_to_hex(sha1), "idx");
1191 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1192 if (ofd < 0)
1193 die("unable to open %s (%s)", name, strerror(errno));
1194 if (copy_fd(ifd_ix, ofd))
1195 exit(1);
1196 close(ifd_ix);
1197 puts(sha1_to_hex(sha1));
1200 return 1;
1203 static void setup_progress_signal(void)
1205 struct sigaction sa;
1206 struct itimerval v;
1208 memset(&sa, 0, sizeof(sa));
1209 sa.sa_handler = progress_interval;
1210 sigemptyset(&sa.sa_mask);
1211 sa.sa_flags = SA_RESTART;
1212 sigaction(SIGALRM, &sa, NULL);
1214 v.it_interval.tv_sec = 1;
1215 v.it_interval.tv_usec = 0;
1216 v.it_value = v.it_interval;
1217 setitimer(ITIMER_REAL, &v, NULL);
1220 static int git_pack_config(const char *k, const char *v)
1222 if(!strcmp(k, "pack.window")) {
1223 window = git_config_int(k, v);
1224 return 0;
1226 return git_default_config(k, v);
1229 int main(int argc, char **argv)
1231 SHA_CTX ctx;
1232 char line[40 + 1 + PATH_MAX + 2];
1233 int depth = 10, pack_to_stdout = 0;
1234 struct object_entry **list;
1235 int num_preferred_base = 0;
1236 int i;
1238 setup_git_directory();
1239 git_config(git_pack_config);
1241 progress = isatty(2);
1242 for (i = 1; i < argc; i++) {
1243 const char *arg = argv[i];
1245 if (*arg == '-') {
1246 if (!strcmp("--non-empty", arg)) {
1247 non_empty = 1;
1248 continue;
1250 if (!strcmp("--local", arg)) {
1251 local = 1;
1252 continue;
1254 if (!strcmp("--progress", arg)) {
1255 progress = 1;
1256 continue;
1258 if (!strcmp("--incremental", arg)) {
1259 incremental = 1;
1260 continue;
1262 if (!strncmp("--window=", arg, 9)) {
1263 char *end;
1264 window = strtoul(arg+9, &end, 0);
1265 if (!arg[9] || *end)
1266 usage(pack_usage);
1267 continue;
1269 if (!strncmp("--depth=", arg, 8)) {
1270 char *end;
1271 depth = strtoul(arg+8, &end, 0);
1272 if (!arg[8] || *end)
1273 usage(pack_usage);
1274 continue;
1276 if (!strcmp("--progress", arg)) {
1277 progress = 1;
1278 continue;
1280 if (!strcmp("-q", arg)) {
1281 progress = 0;
1282 continue;
1284 if (!strcmp("--no-reuse-delta", arg)) {
1285 no_reuse_delta = 1;
1286 continue;
1288 if (!strcmp("--stdout", arg)) {
1289 pack_to_stdout = 1;
1290 continue;
1292 usage(pack_usage);
1294 if (base_name)
1295 usage(pack_usage);
1296 base_name = arg;
1299 if (pack_to_stdout != !base_name)
1300 usage(pack_usage);
1302 prepare_packed_git();
1304 if (progress) {
1305 fprintf(stderr, "Generating pack...\n");
1306 setup_progress_signal();
1309 for (;;) {
1310 unsigned char sha1[20];
1311 unsigned hash;
1313 if (!fgets(line, sizeof(line), stdin)) {
1314 if (feof(stdin))
1315 break;
1316 if (!ferror(stdin))
1317 die("fgets returned NULL, not EOF, not error!");
1318 if (errno != EINTR)
1319 die("fgets: %s", strerror(errno));
1320 clearerr(stdin);
1321 continue;
1324 if (line[0] == '-') {
1325 if (get_sha1_hex(line+1, sha1))
1326 die("expected edge sha1, got garbage:\n %s",
1327 line+1);
1328 if (num_preferred_base++ < window)
1329 add_preferred_base(sha1);
1330 continue;
1332 if (get_sha1_hex(line, sha1))
1333 die("expected sha1, got garbage:\n %s", line);
1334 hash = name_hash(line+41);
1335 add_preferred_base_object(line+41, hash);
1336 add_object_entry(sha1, hash, 0);
1338 if (progress)
1339 fprintf(stderr, "Done counting %d objects.\n", nr_objects);
1340 sorted_by_sha = create_final_object_list();
1341 if (non_empty && !nr_result)
1342 return 0;
1344 SHA1_Init(&ctx);
1345 list = sorted_by_sha;
1346 for (i = 0; i < nr_result; i++) {
1347 struct object_entry *entry = *list++;
1348 SHA1_Update(&ctx, entry->sha1, 20);
1350 SHA1_Final(object_list_sha1, &ctx);
1351 if (progress && (nr_objects != nr_result))
1352 fprintf(stderr, "Result has %d objects.\n", nr_result);
1354 if (reuse_cached_pack(object_list_sha1, pack_to_stdout))
1356 else {
1357 if (nr_result)
1358 prepare_pack(window, depth);
1359 if (progress && pack_to_stdout) {
1360 /* the other end usually displays progress itself */
1361 struct itimerval v = {{0,},};
1362 setitimer(ITIMER_REAL, &v, NULL);
1363 signal(SIGALRM, SIG_IGN );
1364 progress_update = 0;
1366 write_pack_file();
1367 if (!pack_to_stdout) {
1368 write_index_file();
1369 puts(sha1_to_hex(object_list_sha1));
1372 if (progress)
1373 fprintf(stderr, "Total %d, written %d (delta %d), reused %d (delta %d)\n",
1374 nr_result, written, written_delta, reused, reused_delta);
1375 return 0;