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 "rabinpoly.h"
12 #include "gsimm.h"
13 #include <sys/time.h>
14 #include <signal.h>
16 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";
18 struct object_entry {
19 unsigned char sha1[20];
20 unsigned long size; /* uncompressed size */
21 unsigned long offset; /* offset into the final pack file;
22 * nonzero if already written.
23 */
24 unsigned int depth; /* delta depth */
25 unsigned int delta_limit; /* base adjustment for in-pack delta */
26 unsigned int hash; /* name hint hash */
27 enum object_type type;
28 enum object_type in_pack_type; /* could be delta */
29 unsigned long delta_size; /* delta data size (uncompressed) */
30 struct object_entry *delta; /* delta base object */
31 struct packed_git *in_pack; /* already in pack */
32 unsigned int in_pack_offset;
33 struct object_entry *delta_child; /* delitified objects who bases me */
34 struct object_entry *delta_sibling; /* other deltified objects who
35 * uses the same base as me
36 */
37 int preferred_base; /* we do not pack this, but is encouraged to
38 * be used as the base objectto delta huge
39 * objects against.
40 */
41 };
43 /*
44 * Objects we are going to pack are colected in objects array (dynamically
45 * expanded). nr_objects & nr_alloc controls this array. They are stored
46 * in the order we see -- typically rev-list --objects order that gives us
47 * nice "minimum seek" order.
48 *
49 * sorted-by-sha ans sorted-by-type are arrays of pointers that point at
50 * elements in the objects array. The former is used to build the pack
51 * index (lists object names in the ascending order to help offset lookup),
52 * and the latter is used to group similar things together by try_delta()
53 * heuristics.
54 */
56 static unsigned char object_list_sha1[20];
57 static int non_empty = 0;
58 static int no_reuse_delta = 0;
59 static int local = 0;
60 static int incremental = 0;
61 static struct object_entry **sorted_by_sha, **sorted_by_type;
62 static struct object_entry *objects = NULL;
63 static int nr_objects = 0, nr_alloc = 0, nr_result = 0;
64 static const char *base_name;
65 static unsigned char pack_file_sha1[20];
66 static int progress = 1;
67 static volatile sig_atomic_t progress_update = 0;
69 /*
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.
74 */
75 static int *object_ix = NULL;
76 static int object_ix_hashsz = 0;
78 /*
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.
88 */
89 struct pack_revindex {
90 struct packed_git *p;
91 unsigned long *revindex;
92 } *pack_revindex = NULL;
93 static int pack_revindex_hashsz = 0;
95 /*
96 * stats
97 */
98 static int written = 0;
99 static int written_delta = 0;
100 static int reused = 0;
101 static int reused_delta = 0;
103 static int pack_revindex_ix(struct packed_git *p)
104 {
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;
115 }
116 return -1 - i;
117 }
119 static void prepare_pack_ix(void)
120 {
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;
133 }
134 /* revindex elements are lazily initialized */
135 }
137 static int cmp_offset(const void *a_, const void *b_)
138 {
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;
147 }
149 /*
150 * Ordered list of offsets of objects in the pack.
151 */
152 static void prepare_pack_revindex(struct pack_revindex *rix)
153 {
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 long hl = *((long *)(index + 24 * i));
162 rix->revindex[i] = ntohl(hl);
163 }
164 /* This knows the pack format -- the 20-byte trailer
165 * follows immediately after the last object data.
166 */
167 rix->revindex[num_ent] = p->pack_size - 20;
168 qsort(rix->revindex, num_ent, sizeof(unsigned long), cmp_offset);
169 }
171 static unsigned long find_packed_object_size(struct packed_git *p,
172 unsigned long ofs)
173 {
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;
191 }
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");
198 }
200 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
201 {
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;
216 }
218 /*
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"
224 */
225 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
226 {
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++;
240 }
241 *hdr = c;
242 return n;
243 }
245 static unsigned long write_object(struct sha1file *f,
246 struct object_entry *entry)
247 {
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.
271 */
273 if (! to_reuse) {
274 buf = read_sha1_file(entry->sha1, type, &size);
275 if (!buf)
276 die("unable to read %s", sha1_to_hex(entry->sha1));
277 if (size != entry->size)
278 die("object %s size inconsistency (%lu vs %lu)",
279 sha1_to_hex(entry->sha1), size, entry->size);
280 if (entry->delta) {
281 buf = delta_against(buf, size, entry);
282 size = entry->delta_size;
283 obj_type = OBJ_DELTA;
284 }
285 /*
286 * The object header is a byte of 'type' followed by zero or
287 * more bytes of length. For deltas, the 20 bytes of delta
288 * sha1 follows that.
289 */
290 hdrlen = encode_header(obj_type, size, header);
291 sha1write(f, header, hdrlen);
293 if (entry->delta) {
294 sha1write(f, entry->delta, 20);
295 hdrlen += 20;
296 }
297 datalen = sha1write_compressed(f, buf, size);
298 free(buf);
299 }
300 else {
301 struct packed_git *p = entry->in_pack;
302 use_packed_git(p);
304 datalen = find_packed_object_size(p, entry->in_pack_offset);
305 buf = p->pack_base + entry->in_pack_offset;
306 sha1write(f, buf, datalen);
307 unuse_packed_git(p);
308 hdrlen = 0; /* not really */
309 if (obj_type == OBJ_DELTA)
310 reused_delta++;
311 reused++;
312 }
313 if (obj_type == OBJ_DELTA)
314 written_delta++;
315 written++;
316 return hdrlen + datalen;
317 }
319 static unsigned long write_one(struct sha1file *f,
320 struct object_entry *e,
321 unsigned long offset)
322 {
323 if (e->offset)
324 /* offset starts from header size and cannot be zero
325 * if it is written already.
326 */
327 return offset;
328 e->offset = offset;
329 offset += write_object(f, e);
330 /* if we are deltified, write out its base object. */
331 if (e->delta)
332 offset = write_one(f, e->delta, offset);
333 return offset;
334 }
336 static void write_pack_file(void)
337 {
338 int i;
339 struct sha1file *f;
340 unsigned long offset;
341 struct pack_header hdr;
342 unsigned last_percent = 999;
343 int do_progress = 0;
345 if (!base_name)
346 f = sha1fd(1, "<stdout>");
347 else {
348 f = sha1create("%s-%s.%s", base_name,
349 sha1_to_hex(object_list_sha1), "pack");
350 do_progress = progress;
351 }
352 if (do_progress)
353 fprintf(stderr, "Writing %d objects.\n", nr_result);
355 hdr.hdr_signature = htonl(PACK_SIGNATURE);
356 hdr.hdr_version = htonl(PACK_VERSION);
357 hdr.hdr_entries = htonl(nr_result);
358 sha1write(f, &hdr, sizeof(hdr));
359 offset = sizeof(hdr);
360 if (!nr_result)
361 goto done;
362 for (i = 0; i < nr_objects; i++) {
363 offset = write_one(f, objects + i, offset);
364 if (do_progress) {
365 unsigned percent = written * 100 / nr_result;
366 if (progress_update || percent != last_percent) {
367 fprintf(stderr, "%4u%% (%u/%u) done\r",
368 percent, written, nr_result);
369 progress_update = 0;
370 last_percent = percent;
371 }
372 }
373 }
374 if (do_progress)
375 fputc('\n', stderr);
376 done:
377 sha1close(f, pack_file_sha1, 1);
378 }
380 static void write_index_file(void)
381 {
382 int i;
383 struct sha1file *f = sha1create("%s-%s.%s", base_name,
384 sha1_to_hex(object_list_sha1), "idx");
385 struct object_entry **list = sorted_by_sha;
386 struct object_entry **last = list + nr_result;
387 unsigned int array[256];
389 /*
390 * Write the first-level table (the list is sorted,
391 * but we use a 256-entry lookup to be able to avoid
392 * having to do eight extra binary search iterations).
393 */
394 for (i = 0; i < 256; i++) {
395 struct object_entry **next = list;
396 while (next < last) {
397 struct object_entry *entry = *next;
398 if (entry->sha1[0] != i)
399 break;
400 next++;
401 }
402 array[i] = htonl(next - sorted_by_sha);
403 list = next;
404 }
405 sha1write(f, array, 256 * sizeof(int));
407 /*
408 * Write the actual SHA1 entries..
409 */
410 list = sorted_by_sha;
411 for (i = 0; i < nr_result; i++) {
412 struct object_entry *entry = *list++;
413 unsigned int offset = htonl(entry->offset);
414 sha1write(f, &offset, 4);
415 sha1write(f, entry->sha1, 20);
416 }
417 sha1write(f, pack_file_sha1, 20);
418 sha1close(f, NULL, 1);
419 }
421 static int locate_object_entry_hash(const unsigned char *sha1)
422 {
423 int i;
424 unsigned int ui;
425 memcpy(&ui, sha1, sizeof(unsigned int));
426 i = ui % object_ix_hashsz;
427 while (0 < object_ix[i]) {
428 if (!memcmp(sha1, objects[object_ix[i]-1].sha1, 20))
429 return i;
430 if (++i == object_ix_hashsz)
431 i = 0;
432 }
433 return -1 - i;
434 }
436 static struct object_entry *locate_object_entry(const unsigned char *sha1)
437 {
438 int i;
440 if (!object_ix_hashsz)
441 return NULL;
443 i = locate_object_entry_hash(sha1);
444 if (0 <= i)
445 return &objects[object_ix[i]-1];
446 return NULL;
447 }
449 static void rehash_objects(void)
450 {
451 int i;
452 struct object_entry *oe;
454 object_ix_hashsz = nr_objects * 3;
455 if (object_ix_hashsz < 1024)
456 object_ix_hashsz = 1024;
457 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
458 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
459 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
460 int ix = locate_object_entry_hash(oe->sha1);
461 if (0 <= ix)
462 continue;
463 ix = -1 - ix;
464 object_ix[ix] = i + 1;
465 }
466 }
468 struct name_path {
469 struct name_path *up;
470 const char *elem;
471 int len;
472 };
474 #define DIRBITS 12
476 static unsigned name_hash(struct name_path *path, const char *name)
477 {
478 struct name_path *p = path;
479 const char *n = name + strlen(name);
480 unsigned hash = 0, name_hash = 0, name_done = 0;
482 if (n != name && n[-1] == '\n')
483 n--;
484 while (name <= --n) {
485 unsigned char c = *n;
486 if (c == '/' && !name_done) {
487 name_hash = hash;
488 name_done = 1;
489 hash = 0;
490 }
491 hash = hash * 11 + c;
492 }
493 if (!name_done) {
494 name_hash = hash;
495 hash = 0;
496 }
497 for (p = path; p; p = p->up) {
498 hash = hash * 11 + '/';
499 n = p->elem + p->len;
500 while (p->elem <= --n) {
501 unsigned char c = *n;
502 hash = hash * 11 + c;
503 }
504 }
505 /*
506 * Make sure "Makefile" and "t/Makefile" are hashed separately
507 * but close enough.
508 */
509 hash = (name_hash<<DIRBITS) | (hash & ((1U<<DIRBITS )-1));
510 return hash;
511 }
513 static int add_object_entry(const unsigned char *sha1, unsigned hash, int exclude)
514 {
515 unsigned int idx = nr_objects;
516 struct object_entry *entry;
517 struct packed_git *p;
518 unsigned int found_offset = 0;
519 struct packed_git *found_pack = NULL;
520 int ix, status = 0;
522 if (!exclude) {
523 for (p = packed_git; p; p = p->next) {
524 struct pack_entry e;
525 if (find_pack_entry_one(sha1, &e, p)) {
526 if (incremental)
527 return 0;
528 if (local && !p->pack_local)
529 return 0;
530 if (!found_pack) {
531 found_offset = e.offset;
532 found_pack = e.p;
533 }
534 }
535 }
536 }
537 if ((entry = locate_object_entry(sha1)) != NULL)
538 goto already_added;
540 if (idx >= nr_alloc) {
541 unsigned int needed = (idx + 1024) * 3 / 2;
542 objects = xrealloc(objects, needed * sizeof(*entry));
543 nr_alloc = needed;
544 }
545 entry = objects + idx;
546 nr_objects = idx + 1;
547 memset(entry, 0, sizeof(*entry));
548 memcpy(entry->sha1, sha1, 20);
549 entry->hash = hash;
551 if (object_ix_hashsz * 3 <= nr_objects * 4)
552 rehash_objects();
553 else {
554 ix = locate_object_entry_hash(entry->sha1);
555 if (0 <= ix)
556 die("internal error in object hashing.");
557 object_ix[-1 - ix] = idx + 1;
558 }
559 status = 1;
561 already_added:
562 if (progress_update) {
563 fprintf(stderr, "Counting objects...%d\r", nr_objects);
564 progress_update = 0;
565 }
566 if (exclude)
567 entry->preferred_base = 1;
568 else {
569 if (found_pack) {
570 entry->in_pack = found_pack;
571 entry->in_pack_offset = found_offset;
572 }
573 }
574 return status;
575 }
577 struct pbase_tree_cache {
578 unsigned char sha1[20];
579 int ref;
580 int temporary;
581 void *tree_data;
582 unsigned long tree_size;
583 };
585 static struct pbase_tree_cache *(pbase_tree_cache[256]);
586 static int pbase_tree_cache_ix(const unsigned char *sha1)
587 {
588 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
589 }
590 static int pbase_tree_cache_ix_incr(int ix)
591 {
592 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
593 }
595 static struct pbase_tree {
596 struct pbase_tree *next;
597 /* This is a phony "cache" entry; we are not
598 * going to evict it nor find it through _get()
599 * mechanism -- this is for the toplevel node that
600 * would almost always change with any commit.
601 */
602 struct pbase_tree_cache pcache;
603 } *pbase_tree;
605 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
606 {
607 struct pbase_tree_cache *ent, *nent;
608 void *data;
609 unsigned long size;
610 char type[20];
611 int neigh;
612 int my_ix = pbase_tree_cache_ix(sha1);
613 int available_ix = -1;
615 /* pbase-tree-cache acts as a limited hashtable.
616 * your object will be found at your index or within a few
617 * slots after that slot if it is cached.
618 */
619 for (neigh = 0; neigh < 8; neigh++) {
620 ent = pbase_tree_cache[my_ix];
621 if (ent && !memcmp(ent->sha1, sha1, 20)) {
622 ent->ref++;
623 return ent;
624 }
625 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
626 ((0 <= available_ix) &&
627 (!ent && pbase_tree_cache[available_ix])))
628 available_ix = my_ix;
629 if (!ent)
630 break;
631 my_ix = pbase_tree_cache_ix_incr(my_ix);
632 }
634 /* Did not find one. Either we got a bogus request or
635 * we need to read and perhaps cache.
636 */
637 data = read_sha1_file(sha1, type, &size);
638 if (!data)
639 return NULL;
640 if (strcmp(type, tree_type)) {
641 free(data);
642 return NULL;
643 }
645 /* We need to either cache or return a throwaway copy */
647 if (available_ix < 0)
648 ent = NULL;
649 else {
650 ent = pbase_tree_cache[available_ix];
651 my_ix = available_ix;
652 }
654 if (!ent) {
655 nent = xmalloc(sizeof(*nent));
656 nent->temporary = (available_ix < 0);
657 }
658 else {
659 /* evict and reuse */
660 free(ent->tree_data);
661 nent = ent;
662 }
663 memcpy(nent->sha1, sha1, 20);
664 nent->tree_data = data;
665 nent->tree_size = size;
666 nent->ref = 1;
667 if (!nent->temporary)
668 pbase_tree_cache[my_ix] = nent;
669 return nent;
670 }
672 static void pbase_tree_put(struct pbase_tree_cache *cache)
673 {
674 if (!cache->temporary) {
675 cache->ref--;
676 return;
677 }
678 free(cache->tree_data);
679 free(cache);
680 }
682 static int name_cmp_len(const char *name)
683 {
684 int i;
685 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
686 ;
687 return i;
688 }
690 static void add_pbase_object(struct tree_desc *tree,
691 struct name_path *up,
692 const char *name,
693 int cmplen)
694 {
695 while (tree->size) {
696 const unsigned char *sha1;
697 const char *entry_name;
698 int entry_len;
699 unsigned mode;
700 unsigned long size;
701 char type[20];
703 sha1 = tree_entry_extract(tree, &entry_name, &mode);
704 update_tree_entry(tree);
705 entry_len = strlen(entry_name);
706 if (entry_len != cmplen ||
707 memcmp(entry_name, name, cmplen) ||
708 !has_sha1_file(sha1) ||
709 sha1_object_info(sha1, type, &size))
710 continue;
711 if (name[cmplen] != '/') {
712 unsigned hash = name_hash(up, name);
713 add_object_entry(sha1, hash, 1);
714 return;
715 }
716 if (!strcmp(type, tree_type)) {
717 struct tree_desc sub;
718 struct name_path me;
719 struct pbase_tree_cache *tree;
720 const char *down = name+cmplen+1;
721 int downlen = name_cmp_len(down);
723 tree = pbase_tree_get(sha1);
724 if (!tree)
725 return;
726 sub.buf = tree->tree_data;
727 sub.size = tree->tree_size;
729 me.up = up;
730 me.elem = entry_name;
731 me.len = entry_len;
732 add_pbase_object(&sub, &me, down, downlen);
733 pbase_tree_put(tree);
734 }
735 }
736 }
738 static unsigned *done_pbase_paths;
739 static int done_pbase_paths_num;
740 static int done_pbase_paths_alloc;
741 static int done_pbase_path_pos(unsigned hash)
742 {
743 int lo = 0;
744 int hi = done_pbase_paths_num;
745 while (lo < hi) {
746 int mi = (hi + lo) / 2;
747 if (done_pbase_paths[mi] == hash)
748 return mi;
749 if (done_pbase_paths[mi] < hash)
750 hi = mi;
751 else
752 lo = mi + 1;
753 }
754 return -lo-1;
755 }
757 static int check_pbase_path(unsigned hash)
758 {
759 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
760 if (0 <= pos)
761 return 1;
762 pos = -pos - 1;
763 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
764 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
765 done_pbase_paths = xrealloc(done_pbase_paths,
766 done_pbase_paths_alloc *
767 sizeof(unsigned));
768 }
769 done_pbase_paths_num++;
770 if (pos < done_pbase_paths_num)
771 memmove(done_pbase_paths + pos + 1,
772 done_pbase_paths + pos,
773 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
774 done_pbase_paths[pos] = hash;
775 return 0;
776 }
778 static void add_preferred_base_object(char *name, unsigned hash)
779 {
780 struct pbase_tree *it;
781 int cmplen = name_cmp_len(name);
783 if (check_pbase_path(hash))
784 return;
786 for (it = pbase_tree; it; it = it->next) {
787 if (cmplen == 0) {
788 hash = name_hash(NULL, "");
789 add_object_entry(it->pcache.sha1, hash, 1);
790 }
791 else {
792 struct tree_desc tree;
793 tree.buf = it->pcache.tree_data;
794 tree.size = it->pcache.tree_size;
795 add_pbase_object(&tree, NULL, name, cmplen);
796 }
797 }
798 }
800 static void add_preferred_base(unsigned char *sha1)
801 {
802 struct pbase_tree *it;
803 void *data;
804 unsigned long size;
805 unsigned char tree_sha1[20];
807 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
808 if (!data)
809 return;
811 for (it = pbase_tree; it; it = it->next) {
812 if (!memcmp(it->pcache.sha1, tree_sha1, 20)) {
813 free(data);
814 return;
815 }
816 }
818 it = xcalloc(1, sizeof(*it));
819 it->next = pbase_tree;
820 pbase_tree = it;
822 memcpy(it->pcache.sha1, tree_sha1, 20);
823 it->pcache.tree_data = data;
824 it->pcache.tree_size = size;
825 }
827 static void check_object(struct object_entry *entry)
828 {
829 char type[20];
831 if (entry->in_pack && !entry->preferred_base) {
832 unsigned char base[20];
833 unsigned long size;
834 struct object_entry *base_entry;
836 /* We want in_pack_type even if we do not reuse delta.
837 * There is no point not reusing non-delta representations.
838 */
839 check_reuse_pack_delta(entry->in_pack,
840 entry->in_pack_offset,
841 base, &size,
842 &entry->in_pack_type);
844 /* Check if it is delta, and the base is also an object
845 * we are going to pack. If so we will reuse the existing
846 * delta.
847 */
848 if (!no_reuse_delta &&
849 entry->in_pack_type == OBJ_DELTA &&
850 (base_entry = locate_object_entry(base)) &&
851 (!base_entry->preferred_base)) {
853 /* Depth value does not matter - find_deltas()
854 * will never consider reused delta as the
855 * base object to deltify other objects
856 * against, in order to avoid circular deltas.
857 */
859 /* uncompressed size of the delta data */
860 entry->size = entry->delta_size = size;
861 entry->delta = base_entry;
862 entry->type = OBJ_DELTA;
864 entry->delta_sibling = base_entry->delta_child;
865 base_entry->delta_child = entry;
867 return;
868 }
869 /* Otherwise we would do the usual */
870 }
872 if (sha1_object_info(entry->sha1, type, &entry->size))
873 die("unable to get type of object %s",
874 sha1_to_hex(entry->sha1));
876 if (!strcmp(type, commit_type)) {
877 entry->type = OBJ_COMMIT;
878 } else if (!strcmp(type, tree_type)) {
879 entry->type = OBJ_TREE;
880 } else if (!strcmp(type, blob_type)) {
881 entry->type = OBJ_BLOB;
882 } else if (!strcmp(type, tag_type)) {
883 entry->type = OBJ_TAG;
884 } else
885 die("unable to pack object %s of type %s",
886 sha1_to_hex(entry->sha1), type);
887 }
889 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
890 {
891 struct object_entry *child = me->delta_child;
892 unsigned int m = n;
893 while (child) {
894 unsigned int c = check_delta_limit(child, n + 1);
895 if (m < c)
896 m = c;
897 child = child->delta_sibling;
898 }
899 return m;
900 }
902 static void get_object_details(void)
903 {
904 int i;
905 struct object_entry *entry;
907 prepare_pack_ix();
908 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
909 check_object(entry);
911 if (nr_objects == nr_result) {
912 /*
913 * Depth of objects that depend on the entry -- this
914 * is subtracted from depth-max to break too deep
915 * delta chain because of delta data reusing.
916 * However, we loosen this restriction when we know we
917 * are creating a thin pack -- it will have to be
918 * expanded on the other end anyway, so do not
919 * artificially cut the delta chain and let it go as
920 * deep as it wants.
921 */
922 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
923 if (!entry->delta && entry->delta_child)
924 entry->delta_limit =
925 check_delta_limit(entry, 1);
926 }
927 }
929 typedef int (*entry_sort_t)(const struct object_entry *, const struct object_entry *);
931 static entry_sort_t current_sort;
933 static int sort_comparator(const void *_a, const void *_b)
934 {
935 struct object_entry *a = *(struct object_entry **)_a;
936 struct object_entry *b = *(struct object_entry **)_b;
937 return current_sort(a,b);
938 }
940 static struct object_entry **create_sorted_list(entry_sort_t sort)
941 {
942 struct object_entry **list = xmalloc(nr_objects * sizeof(struct object_entry *));
943 int i;
945 for (i = 0; i < nr_objects; i++)
946 list[i] = objects + i;
947 current_sort = sort;
948 qsort(list, nr_objects, sizeof(struct object_entry *), sort_comparator);
949 return list;
950 }
952 static int sha1_sort(const struct object_entry *a, const struct object_entry *b)
953 {
954 return memcmp(a->sha1, b->sha1, 20);
955 }
957 static struct object_entry **create_final_object_list(void)
958 {
959 struct object_entry **list;
960 int i, j;
962 for (i = nr_result = 0; i < nr_objects; i++)
963 if (!objects[i].preferred_base)
964 nr_result++;
965 list = xmalloc(nr_result * sizeof(struct object_entry *));
966 for (i = j = 0; i < nr_objects; i++) {
967 if (!objects[i].preferred_base)
968 list[j++] = objects + i;
969 }
970 current_sort = sha1_sort;
971 qsort(list, nr_result, sizeof(struct object_entry *), sort_comparator);
972 return list;
973 }
975 static int type_size_sort(const struct object_entry *a, const struct object_entry *b)
976 {
977 if (a->type < b->type)
978 return -1;
979 if (a->type > b->type)
980 return 1;
981 if (a->hash < b->hash)
982 return -1;
983 if (a->hash > b->hash)
984 return 1;
985 if (a->preferred_base < b->preferred_base)
986 return -1;
987 if (a->preferred_base > b->preferred_base)
988 return 1;
989 if (a->size < b->size)
990 return -1;
991 if (a->size > b->size)
992 return 1;
993 return a < b ? -1 : (a > b);
994 }
996 struct unpacked {
997 struct object_entry *entry;
998 unsigned char fingerprint[MD_LENGTH];
999 void *data;
1000 };
1002 /*
1003 * We search for deltas _backwards_ in a list sorted by type and
1004 * by size, so that we see progressively smaller and smaller files.
1005 * That's because we prefer deltas to be from the bigger file
1006 * to the smaller - deletes are potentially cheaper, but perhaps
1007 * more importantly, the bigger file is likely the more recent
1008 * one.
1009 */
1010 static int try_delta(struct unpacked *cur, struct unpacked *old, unsigned max_depth)
1011 {
1012 struct object_entry *cur_entry = cur->entry;
1013 struct object_entry *old_entry = old->entry;
1014 unsigned long size, oldsize, delta_size, sizediff;
1015 long max_size;
1016 void *delta_buf;
1018 /* Don't bother doing diffs between different types */
1019 if (cur_entry->type != old_entry->type)
1020 return -1;
1022 /* We do not compute delta to *create* objects we are not
1023 * going to pack.
1024 */
1025 if (cur_entry->preferred_base)
1026 return -1;
1028 /* If the current object is at pack edge, take the depth the
1029 * objects that depend on the current object into account --
1030 * otherwise they would become too deep.
1031 */
1032 if (cur_entry->delta_child) {
1033 if (max_depth <= cur_entry->delta_limit)
1034 return 0;
1035 max_depth -= cur_entry->delta_limit;
1036 }
1038 size = cur_entry->size;
1039 oldsize = old_entry->size;
1040 sizediff = oldsize > size ? oldsize - size : size - oldsize;
1042 if (size < 50)
1043 return -1;
1044 if (old_entry->depth >= max_depth)
1045 return 0;
1047 if (gb_simm_score(cur->fingerprint, old->fingerprint) < 0.4)
1048 return 0;
1050 /*
1051 * NOTE!
1052 *
1053 * We always delta from the bigger to the smaller, since that's
1054 * more space-efficient (deletes don't have to say _what_ they
1055 * delete).
1056 */
1057 max_size = size / 2 - 20;
1058 if (cur_entry->delta)
1059 max_size = cur_entry->delta_size-1;
1060 if (sizediff >= max_size)
1061 return -1;
1062 delta_buf = diff_delta(old->data, oldsize,
1063 cur->data, size, &delta_size, max_size);
1064 if (!delta_buf)
1065 return 0;
1066 cur_entry->delta = old_entry;
1067 cur_entry->delta_size = delta_size;
1068 cur_entry->depth = old_entry->depth + 1;
1069 free(delta_buf);
1070 return 0;
1071 }
1073 static void progress_interval(int signum)
1074 {
1075 progress_update = 1;
1076 }
1078 static void find_deltas(struct object_entry **list, int window, int depth)
1079 {
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 rabin_reset ();
1087 memset(array, 0, array_size);
1088 i = nr_objects;
1089 idx = 0;
1090 if (progress)
1091 fprintf(stderr, "Deltifying %d objects.\n", nr_result);
1093 while (--i >= 0) {
1094 struct object_entry *entry = list[i];
1095 struct unpacked *n = array + idx;
1096 unsigned long size;
1097 char type[10];
1098 int j;
1100 if (!entry->preferred_base)
1101 processed++;
1103 if (progress) {
1104 unsigned percent = processed * 100 / nr_result;
1105 if (percent != last_percent || progress_update) {
1106 fprintf(stderr, "%4u%% (%u/%u) done\r",
1107 percent, processed, nr_result);
1108 progress_update = 0;
1109 last_percent = percent;
1110 }
1111 }
1113 if (entry->delta)
1114 /* This happens if we decided to reuse existing
1115 * delta from a pack. "!no_reuse_delta &&" is implied.
1116 */
1117 continue;
1119 free(n->data);
1120 n->entry = entry;
1121 n->data = read_sha1_file(entry->sha1, type, &size);
1122 if (size != entry->size)
1123 die("object %s inconsistent object length (%lu vs %lu)", sha1_to_hex(entry->sha1), size, entry->size);
1125 gb_simm_process(n->data, size, n->fingerprint);
1127 j = window;
1128 while (--j > 0) {
1129 unsigned int other_idx = idx + j;
1130 struct unpacked *m;
1131 if (other_idx >= window)
1132 other_idx -= window;
1133 m = array + other_idx;
1134 if (!m->entry)
1135 break;
1137 if (try_delta(n, m, depth) < 0)
1138 break;
1139 }
1140 #if 0
1141 /* if we made n a delta, and if n is already at max
1142 * depth, leaving it in the window is pointless. we
1143 * should evict it first.
1144 * ... in theory only; somehow this makes things worse.
1145 */
1146 if (entry->delta && depth <= entry->depth)
1147 continue;
1148 #endif
1149 idx++;
1150 if (idx >= window)
1151 idx = 0;
1152 }
1154 if (progress)
1155 fputc('\n', stderr);
1157 for (i = 0; i < window; ++i)
1158 free(array[i].data);
1159 free(array);
1160 }
1162 static void prepare_pack(int window, int depth)
1163 {
1164 get_object_details();
1165 sorted_by_type = create_sorted_list(type_size_sort);
1166 if (window && depth)
1167 find_deltas(sorted_by_type, window+1, depth);
1168 }
1170 static int reuse_cached_pack(unsigned char *sha1, int pack_to_stdout)
1171 {
1172 static const char cache[] = "pack-cache/pack-%s.%s";
1173 char *cached_pack, *cached_idx;
1174 int ifd, ofd, ifd_ix = -1;
1176 cached_pack = git_path(cache, sha1_to_hex(sha1), "pack");
1177 ifd = open(cached_pack, O_RDONLY);
1178 if (ifd < 0)
1179 return 0;
1181 if (!pack_to_stdout) {
1182 cached_idx = git_path(cache, sha1_to_hex(sha1), "idx");
1183 ifd_ix = open(cached_idx, O_RDONLY);
1184 if (ifd_ix < 0) {
1185 close(ifd);
1186 return 0;
1187 }
1188 }
1190 if (progress)
1191 fprintf(stderr, "Reusing %d objects pack %s\n", nr_objects,
1192 sha1_to_hex(sha1));
1194 if (pack_to_stdout) {
1195 if (copy_fd(ifd, 1))
1196 exit(1);
1197 close(ifd);
1198 }
1199 else {
1200 char name[PATH_MAX];
1201 snprintf(name, sizeof(name),
1202 "%s-%s.%s", base_name, sha1_to_hex(sha1), "pack");
1203 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1204 if (ofd < 0)
1205 die("unable to open %s (%s)", name, strerror(errno));
1206 if (copy_fd(ifd, ofd))
1207 exit(1);
1208 close(ifd);
1210 snprintf(name, sizeof(name),
1211 "%s-%s.%s", base_name, sha1_to_hex(sha1), "idx");
1212 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1213 if (ofd < 0)
1214 die("unable to open %s (%s)", name, strerror(errno));
1215 if (copy_fd(ifd_ix, ofd))
1216 exit(1);
1217 close(ifd_ix);
1218 puts(sha1_to_hex(sha1));
1219 }
1221 return 1;
1222 }
1224 static void setup_progress_signal(void)
1225 {
1226 struct sigaction sa;
1227 struct itimerval v;
1229 memset(&sa, 0, sizeof(sa));
1230 sa.sa_handler = progress_interval;
1231 sigemptyset(&sa.sa_mask);
1232 sa.sa_flags = SA_RESTART;
1233 sigaction(SIGALRM, &sa, NULL);
1235 v.it_interval.tv_sec = 1;
1236 v.it_interval.tv_usec = 0;
1237 v.it_value = v.it_interval;
1238 setitimer(ITIMER_REAL, &v, NULL);
1239 }
1241 int main(int argc, char **argv)
1242 {
1243 SHA_CTX ctx;
1244 char line[PATH_MAX + 20];
1245 int window = 10, depth = 10, pack_to_stdout = 0;
1246 struct object_entry **list;
1247 int num_preferred_base = 0;
1248 int i;
1250 setup_git_directory();
1252 for (i = 1; i < argc; i++) {
1253 const char *arg = argv[i];
1255 if (*arg == '-') {
1256 if (!strcmp("--non-empty", arg)) {
1257 non_empty = 1;
1258 continue;
1259 }
1260 if (!strcmp("--local", arg)) {
1261 local = 1;
1262 continue;
1263 }
1264 if (!strcmp("--incremental", arg)) {
1265 incremental = 1;
1266 continue;
1267 }
1268 if (!strncmp("--window=", arg, 9)) {
1269 char *end;
1270 window = strtoul(arg+9, &end, 0);
1271 if (!arg[9] || *end)
1272 usage(pack_usage);
1273 continue;
1274 }
1275 if (!strncmp("--depth=", arg, 8)) {
1276 char *end;
1277 depth = strtoul(arg+8, &end, 0);
1278 if (!arg[8] || *end)
1279 usage(pack_usage);
1280 continue;
1281 }
1282 if (!strcmp("-q", arg)) {
1283 progress = 0;
1284 continue;
1285 }
1286 if (!strcmp("--no-reuse-delta", arg)) {
1287 no_reuse_delta = 1;
1288 continue;
1289 }
1290 if (!strcmp("--stdout", arg)) {
1291 pack_to_stdout = 1;
1292 continue;
1293 }
1294 usage(pack_usage);
1295 }
1296 if (base_name)
1297 usage(pack_usage);
1298 base_name = arg;
1299 }
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();
1309 }
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;
1324 }
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;
1333 }
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);
1339 }
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);
1351 }
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))
1357 ;
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;
1367 }
1368 write_pack_file();
1369 if (!pack_to_stdout) {
1370 write_index_file();
1371 puts(sha1_to_hex(object_list_sha1));
1372 }
1373 }
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;
1378 }