Merge branch 'jc/dirwalk-n-cache-tree' into jc/cache-tree
[git/jrn.git] / pack-objects.c
blob77284cfdbaef108676c95bd4d9771e40080fd406
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; /* delitified 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 colected 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;
68 * The object names in objects array are hashed with this hashtable,
69 * to help looking up the entry by object name. Binary search from
70 * sorted_by_sha is also possible but this was easier to code and faster.
71 * This hashtable is built after all the objects are seen.
73 static int *object_ix = NULL;
74 static int object_ix_hashsz = 0;
77 * Pack index for existing packs give us easy access to the offsets into
78 * corresponding pack file where each object's data starts, but the entries
79 * do not store the size of the compressed representation (uncompressed
80 * size is easily available by examining the pack entry header). We build
81 * a hashtable of existing packs (pack_revindex), and keep reverse index
82 * here -- pack index file is sorted by object name mapping to offset; this
83 * pack_revindex[].revindex array is an ordered list of offsets, so if you
84 * know the offset of an object, next offset is where its packed
85 * representation ends.
87 struct pack_revindex {
88 struct packed_git *p;
89 unsigned long *revindex;
90 } *pack_revindex = NULL;
91 static int pack_revindex_hashsz = 0;
94 * stats
96 static int written = 0;
97 static int written_delta = 0;
98 static int reused = 0;
99 static int reused_delta = 0;
101 static int pack_revindex_ix(struct packed_git *p)
103 unsigned long ui = (unsigned long)p;
104 int i;
106 ui = ui ^ (ui >> 16); /* defeat structure alignment */
107 i = (int)(ui % pack_revindex_hashsz);
108 while (pack_revindex[i].p) {
109 if (pack_revindex[i].p == p)
110 return i;
111 if (++i == pack_revindex_hashsz)
112 i = 0;
114 return -1 - i;
117 static void prepare_pack_ix(void)
119 int num;
120 struct packed_git *p;
121 for (num = 0, p = packed_git; p; p = p->next)
122 num++;
123 if (!num)
124 return;
125 pack_revindex_hashsz = num * 11;
126 pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
127 for (p = packed_git; p; p = p->next) {
128 num = pack_revindex_ix(p);
129 num = - 1 - num;
130 pack_revindex[num].p = p;
132 /* revindex elements are lazily initialized */
135 static int cmp_offset(const void *a_, const void *b_)
137 unsigned long a = *(unsigned long *) a_;
138 unsigned long b = *(unsigned long *) b_;
139 if (a < b)
140 return -1;
141 else if (a == b)
142 return 0;
143 else
144 return 1;
148 * Ordered list of offsets of objects in the pack.
150 static void prepare_pack_revindex(struct pack_revindex *rix)
152 struct packed_git *p = rix->p;
153 int num_ent = num_packed_objects(p);
154 int i;
155 void *index = p->index_base + 256;
157 rix->revindex = xmalloc(sizeof(unsigned long) * (num_ent + 1));
158 for (i = 0; i < num_ent; i++) {
159 unsigned int hl = *((unsigned int *)(index + 24 * i));
160 rix->revindex[i] = ntohl(hl);
162 /* This knows the pack format -- the 20-byte trailer
163 * follows immediately after the last object data.
165 rix->revindex[num_ent] = p->pack_size - 20;
166 qsort(rix->revindex, num_ent, sizeof(unsigned long), cmp_offset);
169 static unsigned long find_packed_object_size(struct packed_git *p,
170 unsigned long ofs)
172 int num;
173 int lo, hi;
174 struct pack_revindex *rix;
175 unsigned long *revindex;
176 num = pack_revindex_ix(p);
177 if (num < 0)
178 die("internal error: pack revindex uninitialized");
179 rix = &pack_revindex[num];
180 if (!rix->revindex)
181 prepare_pack_revindex(rix);
182 revindex = rix->revindex;
183 lo = 0;
184 hi = num_packed_objects(p) + 1;
185 do {
186 int mi = (lo + hi) / 2;
187 if (revindex[mi] == ofs) {
188 return revindex[mi+1] - ofs;
190 else if (ofs < revindex[mi])
191 hi = mi;
192 else
193 lo = mi + 1;
194 } while (lo < hi);
195 die("internal error: pack revindex corrupt");
198 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
200 unsigned long othersize, delta_size;
201 char type[10];
202 void *otherbuf = read_sha1_file(entry->delta->sha1, type, &othersize);
203 void *delta_buf;
205 if (!otherbuf)
206 die("unable to read %s", sha1_to_hex(entry->delta->sha1));
207 delta_buf = diff_delta(otherbuf, othersize,
208 buf, size, &delta_size, 0);
209 if (!delta_buf || delta_size != entry->delta_size)
210 die("delta size changed");
211 free(buf);
212 free(otherbuf);
213 return delta_buf;
217 * The per-object header is a pretty dense thing, which is
218 * - first byte: low four bits are "size", then three bits of "type",
219 * and the high bit is "size continues".
220 * - each byte afterwards: low seven bits are size continuation,
221 * with the high bit being "size continues"
223 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
225 int n = 1;
226 unsigned char c;
228 if (type < OBJ_COMMIT || type > OBJ_DELTA)
229 die("bad type %d", type);
231 c = (type << 4) | (size & 15);
232 size >>= 4;
233 while (size) {
234 *hdr++ = c | 0x80;
235 c = size & 0x7f;
236 size >>= 7;
237 n++;
239 *hdr = c;
240 return n;
243 static unsigned long write_object(struct sha1file *f,
244 struct object_entry *entry)
246 unsigned long size;
247 char type[10];
248 void *buf;
249 unsigned char header[10];
250 unsigned hdrlen, datalen;
251 enum object_type obj_type;
252 int to_reuse = 0;
254 if (entry->preferred_base)
255 return 0;
257 obj_type = entry->type;
258 if (! entry->in_pack)
259 to_reuse = 0; /* can't reuse what we don't have */
260 else if (obj_type == OBJ_DELTA)
261 to_reuse = 1; /* check_object() decided it for us */
262 else if (obj_type != entry->in_pack_type)
263 to_reuse = 0; /* pack has delta which is unusable */
264 else if (entry->delta)
265 to_reuse = 0; /* we want to pack afresh */
266 else
267 to_reuse = 1; /* we have it in-pack undeltified,
268 * and we do not need to deltify it.
271 if (! to_reuse) {
272 buf = read_sha1_file(entry->sha1, type, &size);
273 if (!buf)
274 die("unable to read %s", sha1_to_hex(entry->sha1));
275 if (size != entry->size)
276 die("object %s size inconsistency (%lu vs %lu)",
277 sha1_to_hex(entry->sha1), size, entry->size);
278 if (entry->delta) {
279 buf = delta_against(buf, size, entry);
280 size = entry->delta_size;
281 obj_type = OBJ_DELTA;
284 * The object header is a byte of 'type' followed by zero or
285 * more bytes of length. For deltas, the 20 bytes of delta
286 * sha1 follows that.
288 hdrlen = encode_header(obj_type, size, header);
289 sha1write(f, header, hdrlen);
291 if (entry->delta) {
292 sha1write(f, entry->delta, 20);
293 hdrlen += 20;
295 datalen = sha1write_compressed(f, buf, size);
296 free(buf);
298 else {
299 struct packed_git *p = entry->in_pack;
300 use_packed_git(p);
302 datalen = find_packed_object_size(p, entry->in_pack_offset);
303 buf = p->pack_base + entry->in_pack_offset;
304 sha1write(f, buf, datalen);
305 unuse_packed_git(p);
306 hdrlen = 0; /* not really */
307 if (obj_type == OBJ_DELTA)
308 reused_delta++;
309 reused++;
311 if (obj_type == OBJ_DELTA)
312 written_delta++;
313 written++;
314 return hdrlen + datalen;
317 static unsigned long write_one(struct sha1file *f,
318 struct object_entry *e,
319 unsigned long offset)
321 if (e->offset)
322 /* offset starts from header size and cannot be zero
323 * if it is written already.
325 return offset;
326 e->offset = offset;
327 offset += write_object(f, e);
328 /* if we are deltified, write out its base object. */
329 if (e->delta)
330 offset = write_one(f, e->delta, offset);
331 return offset;
334 static void write_pack_file(void)
336 int i;
337 struct sha1file *f;
338 unsigned long offset;
339 struct pack_header hdr;
340 unsigned last_percent = 999;
341 int do_progress = 0;
343 if (!base_name)
344 f = sha1fd(1, "<stdout>");
345 else {
346 f = sha1create("%s-%s.%s", base_name,
347 sha1_to_hex(object_list_sha1), "pack");
348 do_progress = progress;
350 if (do_progress)
351 fprintf(stderr, "Writing %d objects.\n", nr_result);
353 hdr.hdr_signature = htonl(PACK_SIGNATURE);
354 hdr.hdr_version = htonl(PACK_VERSION);
355 hdr.hdr_entries = htonl(nr_result);
356 sha1write(f, &hdr, sizeof(hdr));
357 offset = sizeof(hdr);
358 if (!nr_result)
359 goto done;
360 for (i = 0; i < nr_objects; i++) {
361 offset = write_one(f, objects + i, offset);
362 if (do_progress) {
363 unsigned percent = written * 100 / nr_result;
364 if (progress_update || percent != last_percent) {
365 fprintf(stderr, "%4u%% (%u/%u) done\r",
366 percent, written, nr_result);
367 progress_update = 0;
368 last_percent = percent;
372 if (do_progress)
373 fputc('\n', stderr);
374 done:
375 sha1close(f, pack_file_sha1, 1);
378 static void write_index_file(void)
380 int i;
381 struct sha1file *f = sha1create("%s-%s.%s", base_name,
382 sha1_to_hex(object_list_sha1), "idx");
383 struct object_entry **list = sorted_by_sha;
384 struct object_entry **last = list + nr_result;
385 unsigned int array[256];
388 * Write the first-level table (the list is sorted,
389 * but we use a 256-entry lookup to be able to avoid
390 * having to do eight extra binary search iterations).
392 for (i = 0; i < 256; i++) {
393 struct object_entry **next = list;
394 while (next < last) {
395 struct object_entry *entry = *next;
396 if (entry->sha1[0] != i)
397 break;
398 next++;
400 array[i] = htonl(next - sorted_by_sha);
401 list = next;
403 sha1write(f, array, 256 * sizeof(int));
406 * Write the actual SHA1 entries..
408 list = sorted_by_sha;
409 for (i = 0; i < nr_result; i++) {
410 struct object_entry *entry = *list++;
411 unsigned int offset = htonl(entry->offset);
412 sha1write(f, &offset, 4);
413 sha1write(f, entry->sha1, 20);
415 sha1write(f, pack_file_sha1, 20);
416 sha1close(f, NULL, 1);
419 static int locate_object_entry_hash(const unsigned char *sha1)
421 int i;
422 unsigned int ui;
423 memcpy(&ui, sha1, sizeof(unsigned int));
424 i = ui % object_ix_hashsz;
425 while (0 < object_ix[i]) {
426 if (!memcmp(sha1, objects[object_ix[i]-1].sha1, 20))
427 return i;
428 if (++i == object_ix_hashsz)
429 i = 0;
431 return -1 - i;
434 static struct object_entry *locate_object_entry(const unsigned char *sha1)
436 int i;
438 if (!object_ix_hashsz)
439 return NULL;
441 i = locate_object_entry_hash(sha1);
442 if (0 <= i)
443 return &objects[object_ix[i]-1];
444 return NULL;
447 static void rehash_objects(void)
449 int i;
450 struct object_entry *oe;
452 object_ix_hashsz = nr_objects * 3;
453 if (object_ix_hashsz < 1024)
454 object_ix_hashsz = 1024;
455 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
456 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
457 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
458 int ix = locate_object_entry_hash(oe->sha1);
459 if (0 <= ix)
460 continue;
461 ix = -1 - ix;
462 object_ix[ix] = i + 1;
466 struct name_path {
467 struct name_path *up;
468 const char *elem;
469 int len;
472 #define DIRBITS 12
474 static unsigned name_hash(struct name_path *path, const char *name)
476 struct name_path *p = path;
477 const char *n = name + strlen(name);
478 unsigned hash = 0, name_hash = 0, name_done = 0;
480 if (n != name && n[-1] == '\n')
481 n--;
482 while (name <= --n) {
483 unsigned char c = *n;
484 if (c == '/' && !name_done) {
485 name_hash = hash;
486 name_done = 1;
487 hash = 0;
489 hash = hash * 11 + c;
491 if (!name_done) {
492 name_hash = hash;
493 hash = 0;
495 for (p = path; p; p = p->up) {
496 hash = hash * 11 + '/';
497 n = p->elem + p->len;
498 while (p->elem <= --n) {
499 unsigned char c = *n;
500 hash = hash * 11 + c;
504 * Make sure "Makefile" and "t/Makefile" are hashed separately
505 * but close enough.
507 hash = (name_hash<<DIRBITS) | (hash & ((1U<<DIRBITS )-1));
508 return hash;
511 static int add_object_entry(const unsigned char *sha1, unsigned hash, int exclude)
513 unsigned int idx = nr_objects;
514 struct object_entry *entry;
515 struct packed_git *p;
516 unsigned int found_offset = 0;
517 struct packed_git *found_pack = NULL;
518 int ix, status = 0;
520 if (!exclude) {
521 for (p = packed_git; p; p = p->next) {
522 struct pack_entry e;
523 if (find_pack_entry_one(sha1, &e, p)) {
524 if (incremental)
525 return 0;
526 if (local && !p->pack_local)
527 return 0;
528 if (!found_pack) {
529 found_offset = e.offset;
530 found_pack = e.p;
535 if ((entry = locate_object_entry(sha1)) != NULL)
536 goto already_added;
538 if (idx >= nr_alloc) {
539 unsigned int needed = (idx + 1024) * 3 / 2;
540 objects = xrealloc(objects, needed * sizeof(*entry));
541 nr_alloc = needed;
543 entry = objects + idx;
544 nr_objects = idx + 1;
545 memset(entry, 0, sizeof(*entry));
546 memcpy(entry->sha1, sha1, 20);
547 entry->hash = hash;
549 if (object_ix_hashsz * 3 <= nr_objects * 4)
550 rehash_objects();
551 else {
552 ix = locate_object_entry_hash(entry->sha1);
553 if (0 <= ix)
554 die("internal error in object hashing.");
555 object_ix[-1 - ix] = idx + 1;
557 status = 1;
559 already_added:
560 if (progress_update) {
561 fprintf(stderr, "Counting objects...%d\r", nr_objects);
562 progress_update = 0;
564 if (exclude)
565 entry->preferred_base = 1;
566 else {
567 if (found_pack) {
568 entry->in_pack = found_pack;
569 entry->in_pack_offset = found_offset;
572 return status;
575 struct pbase_tree_cache {
576 unsigned char sha1[20];
577 int ref;
578 int temporary;
579 void *tree_data;
580 unsigned long tree_size;
583 static struct pbase_tree_cache *(pbase_tree_cache[256]);
584 static int pbase_tree_cache_ix(const unsigned char *sha1)
586 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
588 static int pbase_tree_cache_ix_incr(int ix)
590 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
593 static struct pbase_tree {
594 struct pbase_tree *next;
595 /* This is a phony "cache" entry; we are not
596 * going to evict it nor find it through _get()
597 * mechanism -- this is for the toplevel node that
598 * would almost always change with any commit.
600 struct pbase_tree_cache pcache;
601 } *pbase_tree;
603 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
605 struct pbase_tree_cache *ent, *nent;
606 void *data;
607 unsigned long size;
608 char type[20];
609 int neigh;
610 int my_ix = pbase_tree_cache_ix(sha1);
611 int available_ix = -1;
613 /* pbase-tree-cache acts as a limited hashtable.
614 * your object will be found at your index or within a few
615 * slots after that slot if it is cached.
617 for (neigh = 0; neigh < 8; neigh++) {
618 ent = pbase_tree_cache[my_ix];
619 if (ent && !memcmp(ent->sha1, sha1, 20)) {
620 ent->ref++;
621 return ent;
623 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
624 ((0 <= available_ix) &&
625 (!ent && pbase_tree_cache[available_ix])))
626 available_ix = my_ix;
627 if (!ent)
628 break;
629 my_ix = pbase_tree_cache_ix_incr(my_ix);
632 /* Did not find one. Either we got a bogus request or
633 * we need to read and perhaps cache.
635 data = read_sha1_file(sha1, type, &size);
636 if (!data)
637 return NULL;
638 if (strcmp(type, tree_type)) {
639 free(data);
640 return NULL;
643 /* We need to either cache or return a throwaway copy */
645 if (available_ix < 0)
646 ent = NULL;
647 else {
648 ent = pbase_tree_cache[available_ix];
649 my_ix = available_ix;
652 if (!ent) {
653 nent = xmalloc(sizeof(*nent));
654 nent->temporary = (available_ix < 0);
656 else {
657 /* evict and reuse */
658 free(ent->tree_data);
659 nent = ent;
661 memcpy(nent->sha1, sha1, 20);
662 nent->tree_data = data;
663 nent->tree_size = size;
664 nent->ref = 1;
665 if (!nent->temporary)
666 pbase_tree_cache[my_ix] = nent;
667 return nent;
670 static void pbase_tree_put(struct pbase_tree_cache *cache)
672 if (!cache->temporary) {
673 cache->ref--;
674 return;
676 free(cache->tree_data);
677 free(cache);
680 static int name_cmp_len(const char *name)
682 int i;
683 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
685 return i;
688 static void add_pbase_object(struct tree_desc *tree,
689 struct name_path *up,
690 const char *name,
691 int cmplen)
693 while (tree->size) {
694 const unsigned char *sha1;
695 const char *entry_name;
696 int entry_len;
697 unsigned mode;
698 unsigned long size;
699 char type[20];
701 sha1 = tree_entry_extract(tree, &entry_name, &mode);
702 update_tree_entry(tree);
703 entry_len = strlen(entry_name);
704 if (entry_len != cmplen ||
705 memcmp(entry_name, name, cmplen) ||
706 !has_sha1_file(sha1) ||
707 sha1_object_info(sha1, type, &size))
708 continue;
709 if (name[cmplen] != '/') {
710 unsigned hash = name_hash(up, name);
711 add_object_entry(sha1, hash, 1);
712 return;
714 if (!strcmp(type, tree_type)) {
715 struct tree_desc sub;
716 struct name_path me;
717 struct pbase_tree_cache *tree;
718 const char *down = name+cmplen+1;
719 int downlen = name_cmp_len(down);
721 tree = pbase_tree_get(sha1);
722 if (!tree)
723 return;
724 sub.buf = tree->tree_data;
725 sub.size = tree->tree_size;
727 me.up = up;
728 me.elem = entry_name;
729 me.len = entry_len;
730 add_pbase_object(&sub, &me, down, downlen);
731 pbase_tree_put(tree);
736 static unsigned *done_pbase_paths;
737 static int done_pbase_paths_num;
738 static int done_pbase_paths_alloc;
739 static int done_pbase_path_pos(unsigned hash)
741 int lo = 0;
742 int hi = done_pbase_paths_num;
743 while (lo < hi) {
744 int mi = (hi + lo) / 2;
745 if (done_pbase_paths[mi] == hash)
746 return mi;
747 if (done_pbase_paths[mi] < hash)
748 hi = mi;
749 else
750 lo = mi + 1;
752 return -lo-1;
755 static int check_pbase_path(unsigned hash)
757 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
758 if (0 <= pos)
759 return 1;
760 pos = -pos - 1;
761 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
762 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
763 done_pbase_paths = xrealloc(done_pbase_paths,
764 done_pbase_paths_alloc *
765 sizeof(unsigned));
767 done_pbase_paths_num++;
768 if (pos < done_pbase_paths_num)
769 memmove(done_pbase_paths + pos + 1,
770 done_pbase_paths + pos,
771 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
772 done_pbase_paths[pos] = hash;
773 return 0;
776 static void add_preferred_base_object(char *name, unsigned hash)
778 struct pbase_tree *it;
779 int cmplen = name_cmp_len(name);
781 if (check_pbase_path(hash))
782 return;
784 for (it = pbase_tree; it; it = it->next) {
785 if (cmplen == 0) {
786 hash = name_hash(NULL, "");
787 add_object_entry(it->pcache.sha1, hash, 1);
789 else {
790 struct tree_desc tree;
791 tree.buf = it->pcache.tree_data;
792 tree.size = it->pcache.tree_size;
793 add_pbase_object(&tree, NULL, name, cmplen);
798 static void add_preferred_base(unsigned char *sha1)
800 struct pbase_tree *it;
801 void *data;
802 unsigned long size;
803 unsigned char tree_sha1[20];
805 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
806 if (!data)
807 return;
809 for (it = pbase_tree; it; it = it->next) {
810 if (!memcmp(it->pcache.sha1, tree_sha1, 20)) {
811 free(data);
812 return;
816 it = xcalloc(1, sizeof(*it));
817 it->next = pbase_tree;
818 pbase_tree = it;
820 memcpy(it->pcache.sha1, tree_sha1, 20);
821 it->pcache.tree_data = data;
822 it->pcache.tree_size = size;
825 static void check_object(struct object_entry *entry)
827 char type[20];
829 if (entry->in_pack && !entry->preferred_base) {
830 unsigned char base[20];
831 unsigned long size;
832 struct object_entry *base_entry;
834 /* We want in_pack_type even if we do not reuse delta.
835 * There is no point not reusing non-delta representations.
837 check_reuse_pack_delta(entry->in_pack,
838 entry->in_pack_offset,
839 base, &size,
840 &entry->in_pack_type);
842 /* Check if it is delta, and the base is also an object
843 * we are going to pack. If so we will reuse the existing
844 * delta.
846 if (!no_reuse_delta &&
847 entry->in_pack_type == OBJ_DELTA &&
848 (base_entry = locate_object_entry(base)) &&
849 (!base_entry->preferred_base)) {
851 /* Depth value does not matter - find_deltas()
852 * will never consider reused delta as the
853 * base object to deltify other objects
854 * against, in order to avoid circular deltas.
857 /* uncompressed size of the delta data */
858 entry->size = entry->delta_size = size;
859 entry->delta = base_entry;
860 entry->type = OBJ_DELTA;
862 entry->delta_sibling = base_entry->delta_child;
863 base_entry->delta_child = entry;
865 return;
867 /* Otherwise we would do the usual */
870 if (sha1_object_info(entry->sha1, type, &entry->size))
871 die("unable to get type of object %s",
872 sha1_to_hex(entry->sha1));
874 if (!strcmp(type, commit_type)) {
875 entry->type = OBJ_COMMIT;
876 } else if (!strcmp(type, tree_type)) {
877 entry->type = OBJ_TREE;
878 } else if (!strcmp(type, blob_type)) {
879 entry->type = OBJ_BLOB;
880 } else if (!strcmp(type, tag_type)) {
881 entry->type = OBJ_TAG;
882 } else
883 die("unable to pack object %s of type %s",
884 sha1_to_hex(entry->sha1), type);
887 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
889 struct object_entry *child = me->delta_child;
890 unsigned int m = n;
891 while (child) {
892 unsigned int c = check_delta_limit(child, n + 1);
893 if (m < c)
894 m = c;
895 child = child->delta_sibling;
897 return m;
900 static void get_object_details(void)
902 int i;
903 struct object_entry *entry;
905 prepare_pack_ix();
906 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
907 check_object(entry);
909 if (nr_objects == nr_result) {
911 * Depth of objects that depend on the entry -- this
912 * is subtracted from depth-max to break too deep
913 * delta chain because of delta data reusing.
914 * However, we loosen this restriction when we know we
915 * are creating a thin pack -- it will have to be
916 * expanded on the other end anyway, so do not
917 * artificially cut the delta chain and let it go as
918 * deep as it wants.
920 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
921 if (!entry->delta && entry->delta_child)
922 entry->delta_limit =
923 check_delta_limit(entry, 1);
927 typedef int (*entry_sort_t)(const struct object_entry *, const struct object_entry *);
929 static entry_sort_t current_sort;
931 static int sort_comparator(const void *_a, const void *_b)
933 struct object_entry *a = *(struct object_entry **)_a;
934 struct object_entry *b = *(struct object_entry **)_b;
935 return current_sort(a,b);
938 static struct object_entry **create_sorted_list(entry_sort_t sort)
940 struct object_entry **list = xmalloc(nr_objects * sizeof(struct object_entry *));
941 int i;
943 for (i = 0; i < nr_objects; i++)
944 list[i] = objects + i;
945 current_sort = sort;
946 qsort(list, nr_objects, sizeof(struct object_entry *), sort_comparator);
947 return list;
950 static int sha1_sort(const struct object_entry *a, const struct object_entry *b)
952 return memcmp(a->sha1, b->sha1, 20);
955 static struct object_entry **create_final_object_list(void)
957 struct object_entry **list;
958 int i, j;
960 for (i = nr_result = 0; i < nr_objects; i++)
961 if (!objects[i].preferred_base)
962 nr_result++;
963 list = xmalloc(nr_result * sizeof(struct object_entry *));
964 for (i = j = 0; i < nr_objects; i++) {
965 if (!objects[i].preferred_base)
966 list[j++] = objects + i;
968 current_sort = sha1_sort;
969 qsort(list, nr_result, sizeof(struct object_entry *), sort_comparator);
970 return list;
973 static int type_size_sort(const struct object_entry *a, const struct object_entry *b)
975 if (a->type < b->type)
976 return -1;
977 if (a->type > b->type)
978 return 1;
979 if (a->hash < b->hash)
980 return -1;
981 if (a->hash > b->hash)
982 return 1;
983 if (a->preferred_base < b->preferred_base)
984 return -1;
985 if (a->preferred_base > b->preferred_base)
986 return 1;
987 if (a->size < b->size)
988 return -1;
989 if (a->size > b->size)
990 return 1;
991 return a < b ? -1 : (a > b);
994 struct unpacked {
995 struct object_entry *entry;
996 void *data;
997 struct delta_index *index;
1001 * We search for deltas _backwards_ in a list sorted by type and
1002 * by size, so that we see progressively smaller and smaller files.
1003 * That's because we prefer deltas to be from the bigger file
1004 * to the smaller - deletes are potentially cheaper, but perhaps
1005 * more importantly, the bigger file is likely the more recent
1006 * one.
1008 static int try_delta(struct unpacked *trg, struct unpacked *src,
1009 struct delta_index *src_index, unsigned max_depth)
1011 struct object_entry *trg_entry = trg->entry;
1012 struct object_entry *src_entry = src->entry;
1013 unsigned long size, src_size, delta_size, sizediff, max_size;
1014 void *delta_buf;
1016 /* Don't bother doing diffs between different types */
1017 if (trg_entry->type != src_entry->type)
1018 return -1;
1020 /* We do not compute delta to *create* objects we are not
1021 * going to pack.
1023 if (trg_entry->preferred_base)
1024 return -1;
1027 * If the current object is at pack edge, take the depth the
1028 * objects that depend on the current object into account --
1029 * otherwise they would become too deep.
1031 if (trg_entry->delta_child) {
1032 if (max_depth <= trg_entry->delta_limit)
1033 return 0;
1034 max_depth -= trg_entry->delta_limit;
1036 if (src_entry->depth >= max_depth)
1037 return 0;
1039 /* Now some size filtering heuristics. */
1040 size = trg_entry->size;
1041 max_size = size/2 - 20;
1042 max_size = max_size * (max_depth - src_entry->depth) / max_depth;
1043 if (max_size == 0)
1044 return 0;
1045 if (trg_entry->delta && trg_entry->delta_size <= max_size)
1046 max_size = trg_entry->delta_size-1;
1047 src_size = src_entry->size;
1048 sizediff = src_size < size ? size - src_size : 0;
1049 if (sizediff >= max_size)
1050 return 0;
1052 delta_buf = create_delta(src_index, trg->data, size, &delta_size, max_size);
1053 if (!delta_buf)
1054 return 0;
1056 trg_entry->delta = src_entry;
1057 trg_entry->delta_size = delta_size;
1058 trg_entry->depth = src_entry->depth + 1;
1059 free(delta_buf);
1060 return 1;
1063 static void progress_interval(int signum)
1065 progress_update = 1;
1068 static void find_deltas(struct object_entry **list, int window, int depth)
1070 int i, idx;
1071 unsigned int array_size = window * sizeof(struct unpacked);
1072 struct unpacked *array = xmalloc(array_size);
1073 unsigned processed = 0;
1074 unsigned last_percent = 999;
1076 memset(array, 0, array_size);
1077 i = nr_objects;
1078 idx = 0;
1079 if (progress)
1080 fprintf(stderr, "Deltifying %d objects.\n", nr_result);
1082 while (--i >= 0) {
1083 struct object_entry *entry = list[i];
1084 struct unpacked *n = array + idx;
1085 unsigned long size;
1086 char type[10];
1087 int j;
1089 if (!entry->preferred_base)
1090 processed++;
1092 if (progress) {
1093 unsigned percent = processed * 100 / nr_result;
1094 if (percent != last_percent || progress_update) {
1095 fprintf(stderr, "%4u%% (%u/%u) done\r",
1096 percent, processed, nr_result);
1097 progress_update = 0;
1098 last_percent = percent;
1102 if (entry->delta)
1103 /* This happens if we decided to reuse existing
1104 * delta from a pack. "!no_reuse_delta &&" is implied.
1106 continue;
1108 if (entry->size < 50)
1109 continue;
1110 free_delta_index(n->index);
1111 n->index = NULL;
1112 free(n->data);
1113 n->entry = entry;
1114 n->data = read_sha1_file(entry->sha1, type, &size);
1115 if (size != entry->size)
1116 die("object %s inconsistent object length (%lu vs %lu)",
1117 sha1_to_hex(entry->sha1), size, entry->size);
1119 j = window;
1120 while (--j > 0) {
1121 unsigned int other_idx = idx + j;
1122 struct unpacked *m;
1123 if (other_idx >= window)
1124 other_idx -= window;
1125 m = array + other_idx;
1126 if (!m->entry)
1127 break;
1128 if (try_delta(n, m, m->index, depth) < 0)
1129 break;
1131 /* if we made n a delta, and if n is already at max
1132 * depth, leaving it in the window is pointless. we
1133 * should evict it first.
1135 if (entry->delta && depth <= entry->depth)
1136 continue;
1138 n->index = create_delta_index(n->data, size);
1139 if (!n->index)
1140 die("out of memory");
1142 idx++;
1143 if (idx >= window)
1144 idx = 0;
1147 if (progress)
1148 fputc('\n', stderr);
1150 for (i = 0; i < window; ++i) {
1151 free_delta_index(array[i].index);
1152 free(array[i].data);
1154 free(array);
1157 static void prepare_pack(int window, int depth)
1159 get_object_details();
1160 sorted_by_type = create_sorted_list(type_size_sort);
1161 if (window && depth)
1162 find_deltas(sorted_by_type, window+1, depth);
1165 static int reuse_cached_pack(unsigned char *sha1, int pack_to_stdout)
1167 static const char cache[] = "pack-cache/pack-%s.%s";
1168 char *cached_pack, *cached_idx;
1169 int ifd, ofd, ifd_ix = -1;
1171 cached_pack = git_path(cache, sha1_to_hex(sha1), "pack");
1172 ifd = open(cached_pack, O_RDONLY);
1173 if (ifd < 0)
1174 return 0;
1176 if (!pack_to_stdout) {
1177 cached_idx = git_path(cache, sha1_to_hex(sha1), "idx");
1178 ifd_ix = open(cached_idx, O_RDONLY);
1179 if (ifd_ix < 0) {
1180 close(ifd);
1181 return 0;
1185 if (progress)
1186 fprintf(stderr, "Reusing %d objects pack %s\n", nr_objects,
1187 sha1_to_hex(sha1));
1189 if (pack_to_stdout) {
1190 if (copy_fd(ifd, 1))
1191 exit(1);
1192 close(ifd);
1194 else {
1195 char name[PATH_MAX];
1196 snprintf(name, sizeof(name),
1197 "%s-%s.%s", base_name, sha1_to_hex(sha1), "pack");
1198 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1199 if (ofd < 0)
1200 die("unable to open %s (%s)", name, strerror(errno));
1201 if (copy_fd(ifd, ofd))
1202 exit(1);
1203 close(ifd);
1205 snprintf(name, sizeof(name),
1206 "%s-%s.%s", base_name, sha1_to_hex(sha1), "idx");
1207 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1208 if (ofd < 0)
1209 die("unable to open %s (%s)", name, strerror(errno));
1210 if (copy_fd(ifd_ix, ofd))
1211 exit(1);
1212 close(ifd_ix);
1213 puts(sha1_to_hex(sha1));
1216 return 1;
1219 static void setup_progress_signal(void)
1221 struct sigaction sa;
1222 struct itimerval v;
1224 memset(&sa, 0, sizeof(sa));
1225 sa.sa_handler = progress_interval;
1226 sigemptyset(&sa.sa_mask);
1227 sa.sa_flags = SA_RESTART;
1228 sigaction(SIGALRM, &sa, NULL);
1230 v.it_interval.tv_sec = 1;
1231 v.it_interval.tv_usec = 0;
1232 v.it_value = v.it_interval;
1233 setitimer(ITIMER_REAL, &v, NULL);
1236 int main(int argc, char **argv)
1238 SHA_CTX ctx;
1239 char line[40 + 1 + PATH_MAX + 2];
1240 int window = 10, depth = 10, pack_to_stdout = 0;
1241 struct object_entry **list;
1242 int num_preferred_base = 0;
1243 int i;
1245 setup_git_directory();
1247 progress = isatty(2);
1248 for (i = 1; i < argc; i++) {
1249 const char *arg = argv[i];
1251 if (*arg == '-') {
1252 if (!strcmp("--non-empty", arg)) {
1253 non_empty = 1;
1254 continue;
1256 if (!strcmp("--local", arg)) {
1257 local = 1;
1258 continue;
1260 if (!strcmp("--incremental", arg)) {
1261 incremental = 1;
1262 continue;
1264 if (!strncmp("--window=", arg, 9)) {
1265 char *end;
1266 window = strtoul(arg+9, &end, 0);
1267 if (!arg[9] || *end)
1268 usage(pack_usage);
1269 continue;
1271 if (!strncmp("--depth=", arg, 8)) {
1272 char *end;
1273 depth = strtoul(arg+8, &end, 0);
1274 if (!arg[8] || *end)
1275 usage(pack_usage);
1276 continue;
1278 if (!strcmp("--progress", arg)) {
1279 progress = 1;
1280 continue;
1282 if (!strcmp("-q", arg)) {
1283 progress = 0;
1284 continue;
1286 if (!strcmp("--no-reuse-delta", arg)) {
1287 no_reuse_delta = 1;
1288 continue;
1290 if (!strcmp("--stdout", arg)) {
1291 pack_to_stdout = 1;
1292 continue;
1294 usage(pack_usage);
1296 if (base_name)
1297 usage(pack_usage);
1298 base_name = arg;
1301 if (pack_to_stdout != !base_name)
1302 usage(pack_usage);
1304 prepare_packed_git();
1306 if (progress) {
1307 fprintf(stderr, "Generating pack...\n");
1308 setup_progress_signal();
1311 for (;;) {
1312 unsigned char sha1[20];
1313 unsigned hash;
1315 if (!fgets(line, sizeof(line), stdin)) {
1316 if (feof(stdin))
1317 break;
1318 if (!ferror(stdin))
1319 die("fgets returned NULL, not EOF, not error!");
1320 if (errno != EINTR)
1321 die("fgets: %s", strerror(errno));
1322 clearerr(stdin);
1323 continue;
1326 if (line[0] == '-') {
1327 if (get_sha1_hex(line+1, sha1))
1328 die("expected edge sha1, got garbage:\n %s",
1329 line+1);
1330 if (num_preferred_base++ < window)
1331 add_preferred_base(sha1);
1332 continue;
1334 if (get_sha1_hex(line, sha1))
1335 die("expected sha1, got garbage:\n %s", line);
1336 hash = name_hash(NULL, line+41);
1337 add_preferred_base_object(line+41, hash);
1338 add_object_entry(sha1, hash, 0);
1340 if (progress)
1341 fprintf(stderr, "Done counting %d objects.\n", nr_objects);
1342 sorted_by_sha = create_final_object_list();
1343 if (non_empty && !nr_result)
1344 return 0;
1346 SHA1_Init(&ctx);
1347 list = sorted_by_sha;
1348 for (i = 0; i < nr_result; i++) {
1349 struct object_entry *entry = *list++;
1350 SHA1_Update(&ctx, entry->sha1, 20);
1352 SHA1_Final(object_list_sha1, &ctx);
1353 if (progress && (nr_objects != nr_result))
1354 fprintf(stderr, "Result has %d objects.\n", nr_result);
1356 if (reuse_cached_pack(object_list_sha1, pack_to_stdout))
1358 else {
1359 if (nr_result)
1360 prepare_pack(window, depth);
1361 if (progress && pack_to_stdout) {
1362 /* the other end usually displays progress itself */
1363 struct itimerval v = {{0,},};
1364 setitimer(ITIMER_REAL, &v, NULL);
1365 signal(SIGALRM, SIG_IGN );
1366 progress_update = 0;
1368 write_pack_file();
1369 if (!pack_to_stdout) {
1370 write_index_file();
1371 puts(sha1_to_hex(object_list_sha1));
1374 if (progress)
1375 fprintf(stderr, "Total %d, written %d (delta %d), reused %d (delta %d)\n",
1376 nr_result, written, written_delta, reused, reused_delta);
1377 return 0;