Added Packing Heursitics IRC writeup.
[git/raj.git] / pack-objects.c
blob9346392150633a8f240e4f51086a069c091631fe
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 long hl = *((long *)(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 object_ix = 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));
509 if (0) { /* debug */
510 n = name + strlen(name);
511 if (n != name && n[-1] == '\n')
512 n--;
513 while (name <= --n)
514 fputc(*n, stderr);
515 for (p = path; p; p = p->up) {
516 fputc('/', stderr);
517 n = p->elem + p->len;
518 while (p->elem <= --n)
519 fputc(*n, stderr);
521 fprintf(stderr, "\t%08x\n", hash);
523 return hash;
526 static int add_object_entry(const unsigned char *sha1, unsigned hash, int exclude)
528 unsigned int idx = nr_objects;
529 struct object_entry *entry;
530 struct packed_git *p;
531 unsigned int found_offset = 0;
532 struct packed_git *found_pack = NULL;
533 int ix, status = 0;
535 if (!exclude) {
536 for (p = packed_git; p; p = p->next) {
537 struct pack_entry e;
538 if (find_pack_entry_one(sha1, &e, p)) {
539 if (incremental)
540 return 0;
541 if (local && !p->pack_local)
542 return 0;
543 if (!found_pack) {
544 found_offset = e.offset;
545 found_pack = e.p;
550 if ((entry = locate_object_entry(sha1)) != NULL)
551 goto already_added;
553 if (idx >= nr_alloc) {
554 unsigned int needed = (idx + 1024) * 3 / 2;
555 objects = xrealloc(objects, needed * sizeof(*entry));
556 nr_alloc = needed;
558 entry = objects + idx;
559 nr_objects = idx + 1;
560 memset(entry, 0, sizeof(*entry));
561 memcpy(entry->sha1, sha1, 20);
562 entry->hash = hash;
564 if (object_ix_hashsz * 3 <= nr_objects * 4)
565 rehash_objects();
566 else {
567 ix = locate_object_entry_hash(entry->sha1);
568 if (0 <= ix)
569 die("internal error in object hashing.");
570 object_ix[-1 - ix] = idx + 1;
572 status = 1;
574 already_added:
575 if (progress_update) {
576 fprintf(stderr, "Counting objects...%d\r", nr_objects);
577 progress_update = 0;
579 if (exclude)
580 entry->preferred_base = 1;
581 else {
582 if (found_pack) {
583 entry->in_pack = found_pack;
584 entry->in_pack_offset = found_offset;
587 return status;
590 static void add_pbase_tree(struct tree_desc *tree, struct name_path *up)
592 while (tree->size) {
593 const unsigned char *sha1;
594 const char *name;
595 unsigned mode, hash;
596 unsigned long size;
597 char type[20];
599 sha1 = tree_entry_extract(tree, &name, &mode);
600 update_tree_entry(tree);
601 if (!has_sha1_file(sha1))
602 continue;
603 if (sha1_object_info(sha1, type, &size))
604 continue;
606 hash = name_hash(up, name);
607 if (!add_object_entry(sha1, hash, 1))
608 continue;
610 if (!strcmp(type, tree_type)) {
611 struct tree_desc sub;
612 void *elem;
613 struct name_path me;
615 elem = read_sha1_file(sha1, type, &sub.size);
616 sub.buf = elem;
617 if (sub.buf) {
618 me.up = up;
619 me.elem = name;
620 me.len = strlen(name);
621 add_pbase_tree(&sub, &me);
622 free(elem);
628 static void add_preferred_base(unsigned char *sha1)
630 struct tree_desc tree;
631 void *elem;
633 elem = read_object_with_reference(sha1, tree_type, &tree.size, NULL);
634 tree.buf = elem;
635 if (!tree.buf)
636 return;
637 if (add_object_entry(sha1, name_hash(NULL, ""), 1))
638 add_pbase_tree(&tree, NULL);
639 free(elem);
642 static void check_object(struct object_entry *entry)
644 char type[20];
646 if (entry->in_pack && !entry->preferred_base) {
647 unsigned char base[20];
648 unsigned long size;
649 struct object_entry *base_entry;
651 /* We want in_pack_type even if we do not reuse delta.
652 * There is no point not reusing non-delta representations.
654 check_reuse_pack_delta(entry->in_pack,
655 entry->in_pack_offset,
656 base, &size,
657 &entry->in_pack_type);
659 /* Check if it is delta, and the base is also an object
660 * we are going to pack. If so we will reuse the existing
661 * delta.
663 if (!no_reuse_delta &&
664 entry->in_pack_type == OBJ_DELTA &&
665 (base_entry = locate_object_entry(base)) &&
666 (!base_entry->preferred_base)) {
668 /* Depth value does not matter - find_deltas()
669 * will never consider reused delta as the
670 * base object to deltify other objects
671 * against, in order to avoid circular deltas.
674 /* uncompressed size of the delta data */
675 entry->size = entry->delta_size = size;
676 entry->delta = base_entry;
677 entry->type = OBJ_DELTA;
679 entry->delta_sibling = base_entry->delta_child;
680 base_entry->delta_child = entry;
682 return;
684 /* Otherwise we would do the usual */
687 if (sha1_object_info(entry->sha1, type, &entry->size))
688 die("unable to get type of object %s",
689 sha1_to_hex(entry->sha1));
691 if (!strcmp(type, commit_type)) {
692 entry->type = OBJ_COMMIT;
693 } else if (!strcmp(type, tree_type)) {
694 entry->type = OBJ_TREE;
695 } else if (!strcmp(type, blob_type)) {
696 entry->type = OBJ_BLOB;
697 } else if (!strcmp(type, tag_type)) {
698 entry->type = OBJ_TAG;
699 } else
700 die("unable to pack object %s of type %s",
701 sha1_to_hex(entry->sha1), type);
704 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
706 struct object_entry *child = me->delta_child;
707 unsigned int m = n;
708 while (child) {
709 unsigned int c = check_delta_limit(child, n + 1);
710 if (m < c)
711 m = c;
712 child = child->delta_sibling;
714 return m;
717 static void get_object_details(void)
719 int i;
720 struct object_entry *entry;
722 prepare_pack_ix();
723 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
724 check_object(entry);
726 if (nr_objects == nr_result) {
728 * Depth of objects that depend on the entry -- this
729 * is subtracted from depth-max to break too deep
730 * delta chain because of delta data reusing.
731 * However, we loosen this restriction when we know we
732 * are creating a thin pack -- it will have to be
733 * expanded on the other end anyway, so do not
734 * artificially cut the delta chain and let it go as
735 * deep as it wants.
737 for (i = 0, entry = objects; i < nr_objects; i++, entry++)
738 if (!entry->delta && entry->delta_child)
739 entry->delta_limit =
740 check_delta_limit(entry, 1);
744 typedef int (*entry_sort_t)(const struct object_entry *, const struct object_entry *);
746 static entry_sort_t current_sort;
748 static int sort_comparator(const void *_a, const void *_b)
750 struct object_entry *a = *(struct object_entry **)_a;
751 struct object_entry *b = *(struct object_entry **)_b;
752 return current_sort(a,b);
755 static struct object_entry **create_sorted_list(entry_sort_t sort)
757 struct object_entry **list = xmalloc(nr_objects * sizeof(struct object_entry *));
758 int i;
760 for (i = 0; i < nr_objects; i++)
761 list[i] = objects + i;
762 current_sort = sort;
763 qsort(list, nr_objects, sizeof(struct object_entry *), sort_comparator);
764 return list;
767 static int sha1_sort(const struct object_entry *a, const struct object_entry *b)
769 return memcmp(a->sha1, b->sha1, 20);
772 static struct object_entry **create_final_object_list(void)
774 struct object_entry **list;
775 int i, j;
777 for (i = nr_result = 0; i < nr_objects; i++)
778 if (!objects[i].preferred_base)
779 nr_result++;
780 list = xmalloc(nr_result * sizeof(struct object_entry *));
781 for (i = j = 0; i < nr_objects; i++) {
782 if (!objects[i].preferred_base)
783 list[j++] = objects + i;
785 current_sort = sha1_sort;
786 qsort(list, nr_result, sizeof(struct object_entry *), sort_comparator);
787 return list;
790 static int type_size_sort(const struct object_entry *a, const struct object_entry *b)
792 if (a->type < b->type)
793 return -1;
794 if (a->type > b->type)
795 return 1;
796 if (a->hash < b->hash)
797 return -1;
798 if (a->hash > b->hash)
799 return 1;
800 if (a->preferred_base < b->preferred_base)
801 return -1;
802 if (a->preferred_base > b->preferred_base)
803 return 1;
804 if (a->size < b->size)
805 return -1;
806 if (a->size > b->size)
807 return 1;
808 return a < b ? -1 : (a > b);
811 struct unpacked {
812 struct object_entry *entry;
813 void *data;
817 * We search for deltas _backwards_ in a list sorted by type and
818 * by size, so that we see progressively smaller and smaller files.
819 * That's because we prefer deltas to be from the bigger file
820 * to the smaller - deletes are potentially cheaper, but perhaps
821 * more importantly, the bigger file is likely the more recent
822 * one.
824 static int try_delta(struct unpacked *cur, struct unpacked *old, unsigned max_depth)
826 struct object_entry *cur_entry = cur->entry;
827 struct object_entry *old_entry = old->entry;
828 unsigned long size, oldsize, delta_size, sizediff;
829 long max_size;
830 void *delta_buf;
832 /* Don't bother doing diffs between different types */
833 if (cur_entry->type != old_entry->type)
834 return -1;
836 /* We do not compute delta to *create* objects we are not
837 * going to pack.
839 if (cur_entry->preferred_base)
840 return -1;
842 /* If the current object is at pack edge, take the depth the
843 * objects that depend on the current object into account --
844 * otherwise they would become too deep.
846 if (cur_entry->delta_child) {
847 if (max_depth <= cur_entry->delta_limit)
848 return 0;
849 max_depth -= cur_entry->delta_limit;
852 size = cur_entry->size;
853 oldsize = old_entry->size;
854 sizediff = oldsize > size ? oldsize - size : size - oldsize;
856 if (size < 50)
857 return -1;
858 if (old_entry->depth >= max_depth)
859 return 0;
862 * NOTE!
864 * We always delta from the bigger to the smaller, since that's
865 * more space-efficient (deletes don't have to say _what_ they
866 * delete).
868 max_size = size / 2 - 20;
869 if (cur_entry->delta)
870 max_size = cur_entry->delta_size-1;
871 if (sizediff >= max_size)
872 return -1;
873 delta_buf = diff_delta(old->data, oldsize,
874 cur->data, size, &delta_size, max_size);
875 if (!delta_buf)
876 return 0;
877 cur_entry->delta = old_entry;
878 cur_entry->delta_size = delta_size;
879 cur_entry->depth = old_entry->depth + 1;
880 free(delta_buf);
881 return 0;
884 static void progress_interval(int signum)
886 progress_update = 1;
889 static void find_deltas(struct object_entry **list, int window, int depth)
891 int i, idx;
892 unsigned int array_size = window * sizeof(struct unpacked);
893 struct unpacked *array = xmalloc(array_size);
894 unsigned processed = 0;
895 unsigned last_percent = 999;
897 memset(array, 0, array_size);
898 i = nr_objects;
899 idx = 0;
900 if (progress)
901 fprintf(stderr, "Deltifying %d objects.\n", nr_result);
903 while (--i >= 0) {
904 struct object_entry *entry = list[i];
905 struct unpacked *n = array + idx;
906 unsigned long size;
907 char type[10];
908 int j;
910 if (!entry->preferred_base)
911 processed++;
913 if (progress) {
914 unsigned percent = processed * 100 / nr_result;
915 if (percent != last_percent || progress_update) {
916 fprintf(stderr, "%4u%% (%u/%u) done\r",
917 percent, processed, nr_result);
918 progress_update = 0;
919 last_percent = percent;
923 if (entry->delta)
924 /* This happens if we decided to reuse existing
925 * delta from a pack. "!no_reuse_delta &&" is implied.
927 continue;
929 free(n->data);
930 n->entry = entry;
931 n->data = read_sha1_file(entry->sha1, type, &size);
932 if (size != entry->size)
933 die("object %s inconsistent object length (%lu vs %lu)", sha1_to_hex(entry->sha1), size, entry->size);
935 j = window;
936 while (--j > 0) {
937 unsigned int other_idx = idx + j;
938 struct unpacked *m;
939 if (other_idx >= window)
940 other_idx -= window;
941 m = array + other_idx;
942 if (!m->entry)
943 break;
944 if (try_delta(n, m, depth) < 0)
945 break;
947 #if 0
948 /* if we made n a delta, and if n is already at max
949 * depth, leaving it in the window is pointless. we
950 * should evict it first.
951 * ... in theory only; somehow this makes things worse.
953 if (entry->delta && depth <= entry->depth)
954 continue;
955 #endif
956 idx++;
957 if (idx >= window)
958 idx = 0;
961 if (progress)
962 fputc('\n', stderr);
964 for (i = 0; i < window; ++i)
965 free(array[i].data);
966 free(array);
969 static void prepare_pack(int window, int depth)
971 get_object_details();
972 sorted_by_type = create_sorted_list(type_size_sort);
973 if (window && depth)
974 find_deltas(sorted_by_type, window+1, depth);
977 static int reuse_cached_pack(unsigned char *sha1, int pack_to_stdout)
979 static const char cache[] = "pack-cache/pack-%s.%s";
980 char *cached_pack, *cached_idx;
981 int ifd, ofd, ifd_ix = -1;
983 cached_pack = git_path(cache, sha1_to_hex(sha1), "pack");
984 ifd = open(cached_pack, O_RDONLY);
985 if (ifd < 0)
986 return 0;
988 if (!pack_to_stdout) {
989 cached_idx = git_path(cache, sha1_to_hex(sha1), "idx");
990 ifd_ix = open(cached_idx, O_RDONLY);
991 if (ifd_ix < 0) {
992 close(ifd);
993 return 0;
997 if (progress)
998 fprintf(stderr, "Reusing %d objects pack %s\n", nr_objects,
999 sha1_to_hex(sha1));
1001 if (pack_to_stdout) {
1002 if (copy_fd(ifd, 1))
1003 exit(1);
1004 close(ifd);
1006 else {
1007 char name[PATH_MAX];
1008 snprintf(name, sizeof(name),
1009 "%s-%s.%s", base_name, sha1_to_hex(sha1), "pack");
1010 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1011 if (ofd < 0)
1012 die("unable to open %s (%s)", name, strerror(errno));
1013 if (copy_fd(ifd, ofd))
1014 exit(1);
1015 close(ifd);
1017 snprintf(name, sizeof(name),
1018 "%s-%s.%s", base_name, sha1_to_hex(sha1), "idx");
1019 ofd = open(name, O_CREAT | O_EXCL | O_WRONLY, 0666);
1020 if (ofd < 0)
1021 die("unable to open %s (%s)", name, strerror(errno));
1022 if (copy_fd(ifd_ix, ofd))
1023 exit(1);
1024 close(ifd_ix);
1025 puts(sha1_to_hex(sha1));
1028 return 1;
1031 static void setup_progress_signal(void)
1033 struct sigaction sa;
1034 struct itimerval v;
1036 memset(&sa, 0, sizeof(sa));
1037 sa.sa_handler = progress_interval;
1038 sigemptyset(&sa.sa_mask);
1039 sa.sa_flags = SA_RESTART;
1040 sigaction(SIGALRM, &sa, NULL);
1042 v.it_interval.tv_sec = 1;
1043 v.it_interval.tv_usec = 0;
1044 v.it_value = v.it_interval;
1045 setitimer(ITIMER_REAL, &v, NULL);
1048 int main(int argc, char **argv)
1050 SHA_CTX ctx;
1051 char line[PATH_MAX + 20];
1052 int window = 10, depth = 10, pack_to_stdout = 0;
1053 struct object_entry **list;
1054 int i;
1056 setup_git_directory();
1058 for (i = 1; i < argc; i++) {
1059 const char *arg = argv[i];
1061 if (*arg == '-') {
1062 if (!strcmp("--non-empty", arg)) {
1063 non_empty = 1;
1064 continue;
1066 if (!strcmp("--local", arg)) {
1067 local = 1;
1068 continue;
1070 if (!strcmp("--incremental", arg)) {
1071 incremental = 1;
1072 continue;
1074 if (!strncmp("--window=", arg, 9)) {
1075 char *end;
1076 window = strtoul(arg+9, &end, 0);
1077 if (!arg[9] || *end)
1078 usage(pack_usage);
1079 continue;
1081 if (!strncmp("--depth=", arg, 8)) {
1082 char *end;
1083 depth = strtoul(arg+8, &end, 0);
1084 if (!arg[8] || *end)
1085 usage(pack_usage);
1086 continue;
1088 if (!strcmp("-q", arg)) {
1089 progress = 0;
1090 continue;
1092 if (!strcmp("--no-reuse-delta", arg)) {
1093 no_reuse_delta = 1;
1094 continue;
1096 if (!strcmp("--stdout", arg)) {
1097 pack_to_stdout = 1;
1098 continue;
1100 usage(pack_usage);
1102 if (base_name)
1103 usage(pack_usage);
1104 base_name = arg;
1107 if (pack_to_stdout != !base_name)
1108 usage(pack_usage);
1110 prepare_packed_git();
1112 if (progress) {
1113 fprintf(stderr, "Generating pack...\n");
1114 setup_progress_signal();
1117 for (;;) {
1118 unsigned char sha1[20];
1120 if (!fgets(line, sizeof(line), stdin)) {
1121 if (feof(stdin))
1122 break;
1123 if (!ferror(stdin))
1124 die("fgets returned NULL, not EOF, not error!");
1125 if (errno != EINTR)
1126 die("fgets: %s", strerror(errno));
1127 clearerr(stdin);
1128 continue;
1131 if (line[0] == '-') {
1132 if (get_sha1_hex(line+1, sha1))
1133 die("expected edge sha1, got garbage:\n %s",
1134 line+1);
1135 add_preferred_base(sha1);
1136 continue;
1138 if (get_sha1_hex(line, sha1))
1139 die("expected sha1, got garbage:\n %s", line);
1140 add_object_entry(sha1, name_hash(NULL, line+41), 0);
1142 if (progress)
1143 fprintf(stderr, "Done counting %d objects.\n", nr_objects);
1144 sorted_by_sha = create_final_object_list();
1145 if (non_empty && !nr_result)
1146 return 0;
1148 SHA1_Init(&ctx);
1149 list = sorted_by_sha;
1150 for (i = 0; i < nr_result; i++) {
1151 struct object_entry *entry = *list++;
1152 SHA1_Update(&ctx, entry->sha1, 20);
1154 SHA1_Final(object_list_sha1, &ctx);
1155 if (progress && (nr_objects != nr_result))
1156 fprintf(stderr, "Result has %d objects.\n", nr_result);
1158 if (reuse_cached_pack(object_list_sha1, pack_to_stdout))
1160 else {
1161 if (nr_result)
1162 prepare_pack(window, depth);
1163 if (progress && pack_to_stdout) {
1164 /* the other end usually displays progress itself */
1165 struct itimerval v = {{0,},};
1166 setitimer(ITIMER_REAL, &v, NULL);
1167 signal(SIGALRM, SIG_IGN );
1168 progress_update = 0;
1170 write_pack_file();
1171 if (!pack_to_stdout) {
1172 write_index_file();
1173 puts(sha1_to_hex(object_list_sha1));
1176 if (progress)
1177 fprintf(stderr, "Total %d, written %d (delta %d), reused %d (delta %d)\n",
1178 nr_result, written, written_delta, reused, reused_delta);
1179 return 0;