Documentation/git-reset:
[git/jrn.git] / builtin-pack-objects.c
blobd3efeff03f89cc0f4d0da463ddf878c28effb31e
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "csum-file.h"
12 #include "tree-walk.h"
13 #include "diff.h"
14 #include "revision.h"
15 #include "list-objects.h"
16 #include "progress.h"
18 #ifdef THREADED_DELTA_SEARCH
19 #include <pthread.h>
20 #endif
22 static const char pack_usage[] = "\
23 git-pack-objects [{ -q | --progress | --all-progress }] \n\
24 [--max-pack-size=N] [--local] [--incremental] \n\
25 [--window=N] [--window-memory=N] [--depth=N] \n\
26 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
27 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
28 [--stdout | base-name] [--keep-unreachable] [<ref-list | <object-list]";
30 struct object_entry {
31 struct pack_idx_entry idx;
32 unsigned long size; /* uncompressed size */
33 struct packed_git *in_pack; /* already in pack */
34 off_t in_pack_offset;
35 struct object_entry *delta; /* delta base object */
36 struct object_entry *delta_child; /* deltified objects who bases me */
37 struct object_entry *delta_sibling; /* other deltified objects who
38 * uses the same base as me
40 void *delta_data; /* cached delta (uncompressed) */
41 unsigned long delta_size; /* delta data size (uncompressed) */
42 unsigned int hash; /* name hint hash */
43 enum object_type type;
44 enum object_type in_pack_type; /* could be delta */
45 unsigned char in_pack_header_size;
46 unsigned char preferred_base; /* we do not pack this, but is available
47 * to be used as the base object to delta
48 * objects against.
50 unsigned char no_try_delta;
54 * Objects we are going to pack are collected in objects array (dynamically
55 * expanded). nr_objects & nr_alloc controls this array. They are stored
56 * in the order we see -- typically rev-list --objects order that gives us
57 * nice "minimum seek" order.
59 static struct object_entry *objects;
60 static struct pack_idx_entry **written_list;
61 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
63 static int non_empty;
64 static int no_reuse_delta, no_reuse_object, keep_unreachable;
65 static int local;
66 static int incremental;
67 static int allow_ofs_delta;
68 static const char *base_name;
69 static int progress = 1;
70 static int window = 10;
71 static uint32_t pack_size_limit;
72 static int depth = 50;
73 static int delta_search_threads = 1;
74 static int pack_to_stdout;
75 static int num_preferred_base;
76 static struct progress *progress_state;
77 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
78 static int pack_compression_seen;
80 static unsigned long delta_cache_size = 0;
81 static unsigned long max_delta_cache_size = 0;
82 static unsigned long cache_max_small_delta_size = 1000;
84 static unsigned long window_memory_limit = 0;
87 * The object names in objects array are hashed with this hashtable,
88 * to help looking up the entry by object name.
89 * This hashtable is built after all the objects are seen.
91 static int *object_ix;
92 static int object_ix_hashsz;
95 * Pack index for existing packs give us easy access to the offsets into
96 * corresponding pack file where each object's data starts, but the entries
97 * do not store the size of the compressed representation (uncompressed
98 * size is easily available by examining the pack entry header). It is
99 * also rather expensive to find the sha1 for an object given its offset.
101 * We build a hashtable of existing packs (pack_revindex), and keep reverse
102 * index here -- pack index file is sorted by object name mapping to offset;
103 * this pack_revindex[].revindex array is a list of offset/index_nr pairs
104 * ordered by offset, so if you know the offset of an object, next offset
105 * is where its packed representation ends and the index_nr can be used to
106 * get the object sha1 from the main index.
108 struct revindex_entry {
109 off_t offset;
110 unsigned int nr;
112 struct pack_revindex {
113 struct packed_git *p;
114 struct revindex_entry *revindex;
116 static struct pack_revindex *pack_revindex;
117 static int pack_revindex_hashsz;
120 * stats
122 static uint32_t written, written_delta;
123 static uint32_t reused, reused_delta;
125 static int pack_revindex_ix(struct packed_git *p)
127 unsigned long ui = (unsigned long)p;
128 int i;
130 ui = ui ^ (ui >> 16); /* defeat structure alignment */
131 i = (int)(ui % pack_revindex_hashsz);
132 while (pack_revindex[i].p) {
133 if (pack_revindex[i].p == p)
134 return i;
135 if (++i == pack_revindex_hashsz)
136 i = 0;
138 return -1 - i;
141 static void prepare_pack_ix(void)
143 int num;
144 struct packed_git *p;
145 for (num = 0, p = packed_git; p; p = p->next)
146 num++;
147 if (!num)
148 return;
149 pack_revindex_hashsz = num * 11;
150 pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
151 for (p = packed_git; p; p = p->next) {
152 num = pack_revindex_ix(p);
153 num = - 1 - num;
154 pack_revindex[num].p = p;
156 /* revindex elements are lazily initialized */
159 static int cmp_offset(const void *a_, const void *b_)
161 const struct revindex_entry *a = a_;
162 const struct revindex_entry *b = b_;
163 return (a->offset < b->offset) ? -1 : (a->offset > b->offset) ? 1 : 0;
167 * Ordered list of offsets of objects in the pack.
169 static void prepare_pack_revindex(struct pack_revindex *rix)
171 struct packed_git *p = rix->p;
172 int num_ent = p->num_objects;
173 int i;
174 const char *index = p->index_data;
176 rix->revindex = xmalloc(sizeof(*rix->revindex) * (num_ent + 1));
177 index += 4 * 256;
179 if (p->index_version > 1) {
180 const uint32_t *off_32 =
181 (uint32_t *)(index + 8 + p->num_objects * (20 + 4));
182 const uint32_t *off_64 = off_32 + p->num_objects;
183 for (i = 0; i < num_ent; i++) {
184 uint32_t off = ntohl(*off_32++);
185 if (!(off & 0x80000000)) {
186 rix->revindex[i].offset = off;
187 } else {
188 rix->revindex[i].offset =
189 ((uint64_t)ntohl(*off_64++)) << 32;
190 rix->revindex[i].offset |=
191 ntohl(*off_64++);
193 rix->revindex[i].nr = i;
195 } else {
196 for (i = 0; i < num_ent; i++) {
197 uint32_t hl = *((uint32_t *)(index + 24 * i));
198 rix->revindex[i].offset = ntohl(hl);
199 rix->revindex[i].nr = i;
203 /* This knows the pack format -- the 20-byte trailer
204 * follows immediately after the last object data.
206 rix->revindex[num_ent].offset = p->pack_size - 20;
207 rix->revindex[num_ent].nr = -1;
208 qsort(rix->revindex, num_ent, sizeof(*rix->revindex), cmp_offset);
211 static struct revindex_entry * find_packed_object(struct packed_git *p,
212 off_t ofs)
214 int num;
215 int lo, hi;
216 struct pack_revindex *rix;
217 struct revindex_entry *revindex;
218 num = pack_revindex_ix(p);
219 if (num < 0)
220 die("internal error: pack revindex uninitialized");
221 rix = &pack_revindex[num];
222 if (!rix->revindex)
223 prepare_pack_revindex(rix);
224 revindex = rix->revindex;
225 lo = 0;
226 hi = p->num_objects + 1;
227 do {
228 int mi = (lo + hi) / 2;
229 if (revindex[mi].offset == ofs) {
230 return revindex + mi;
232 else if (ofs < revindex[mi].offset)
233 hi = mi;
234 else
235 lo = mi + 1;
236 } while (lo < hi);
237 die("internal error: pack revindex corrupt");
240 static const unsigned char *find_packed_object_name(struct packed_git *p,
241 off_t ofs)
243 struct revindex_entry *entry = find_packed_object(p, ofs);
244 return nth_packed_object_sha1(p, entry->nr);
247 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
249 unsigned long othersize, delta_size;
250 enum object_type type;
251 void *otherbuf = read_sha1_file(entry->delta->idx.sha1, &type, &othersize);
252 void *delta_buf;
254 if (!otherbuf)
255 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
256 delta_buf = diff_delta(otherbuf, othersize,
257 buf, size, &delta_size, 0);
258 if (!delta_buf || delta_size != entry->delta_size)
259 die("delta size changed");
260 free(buf);
261 free(otherbuf);
262 return delta_buf;
266 * The per-object header is a pretty dense thing, which is
267 * - first byte: low four bits are "size", then three bits of "type",
268 * and the high bit is "size continues".
269 * - each byte afterwards: low seven bits are size continuation,
270 * with the high bit being "size continues"
272 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
274 int n = 1;
275 unsigned char c;
277 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
278 die("bad type %d", type);
280 c = (type << 4) | (size & 15);
281 size >>= 4;
282 while (size) {
283 *hdr++ = c | 0x80;
284 c = size & 0x7f;
285 size >>= 7;
286 n++;
288 *hdr = c;
289 return n;
293 * we are going to reuse the existing object data as is. make
294 * sure it is not corrupt.
296 static int check_pack_inflate(struct packed_git *p,
297 struct pack_window **w_curs,
298 off_t offset,
299 off_t len,
300 unsigned long expect)
302 z_stream stream;
303 unsigned char fakebuf[4096], *in;
304 int st;
306 memset(&stream, 0, sizeof(stream));
307 inflateInit(&stream);
308 do {
309 in = use_pack(p, w_curs, offset, &stream.avail_in);
310 stream.next_in = in;
311 stream.next_out = fakebuf;
312 stream.avail_out = sizeof(fakebuf);
313 st = inflate(&stream, Z_FINISH);
314 offset += stream.next_in - in;
315 } while (st == Z_OK || st == Z_BUF_ERROR);
316 inflateEnd(&stream);
317 return (st == Z_STREAM_END &&
318 stream.total_out == expect &&
319 stream.total_in == len) ? 0 : -1;
322 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
323 off_t offset, off_t len, unsigned int nr)
325 const uint32_t *index_crc;
326 uint32_t data_crc = crc32(0, Z_NULL, 0);
328 do {
329 unsigned int avail;
330 void *data = use_pack(p, w_curs, offset, &avail);
331 if (avail > len)
332 avail = len;
333 data_crc = crc32(data_crc, data, avail);
334 offset += avail;
335 len -= avail;
336 } while (len);
338 index_crc = p->index_data;
339 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
341 return data_crc != ntohl(*index_crc);
344 static void copy_pack_data(struct sha1file *f,
345 struct packed_git *p,
346 struct pack_window **w_curs,
347 off_t offset,
348 off_t len)
350 unsigned char *in;
351 unsigned int avail;
353 while (len) {
354 in = use_pack(p, w_curs, offset, &avail);
355 if (avail > len)
356 avail = (unsigned int)len;
357 sha1write(f, in, avail);
358 offset += avail;
359 len -= avail;
363 static unsigned long write_object(struct sha1file *f,
364 struct object_entry *entry,
365 off_t write_offset)
367 unsigned long size;
368 enum object_type type;
369 void *buf;
370 unsigned char header[10];
371 unsigned char dheader[10];
372 unsigned hdrlen;
373 off_t datalen;
374 enum object_type obj_type;
375 int to_reuse = 0;
376 /* write limit if limited packsize and not first object */
377 unsigned long limit = pack_size_limit && nr_written ?
378 pack_size_limit - write_offset : 0;
379 /* no if no delta */
380 int usable_delta = !entry->delta ? 0 :
381 /* yes if unlimited packfile */
382 !pack_size_limit ? 1 :
383 /* no if base written to previous pack */
384 entry->delta->idx.offset == (off_t)-1 ? 0 :
385 /* otherwise double-check written to this
386 * pack, like we do below
388 entry->delta->idx.offset ? 1 : 0;
390 if (!pack_to_stdout)
391 crc32_begin(f);
393 obj_type = entry->type;
394 if (no_reuse_object)
395 to_reuse = 0; /* explicit */
396 else if (!entry->in_pack)
397 to_reuse = 0; /* can't reuse what we don't have */
398 else if (obj_type == OBJ_REF_DELTA || obj_type == OBJ_OFS_DELTA)
399 /* check_object() decided it for us ... */
400 to_reuse = usable_delta;
401 /* ... but pack split may override that */
402 else if (obj_type != entry->in_pack_type)
403 to_reuse = 0; /* pack has delta which is unusable */
404 else if (entry->delta)
405 to_reuse = 0; /* we want to pack afresh */
406 else
407 to_reuse = 1; /* we have it in-pack undeltified,
408 * and we do not need to deltify it.
411 if (!to_reuse) {
412 z_stream stream;
413 unsigned long maxsize;
414 void *out;
415 if (!usable_delta) {
416 buf = read_sha1_file(entry->idx.sha1, &obj_type, &size);
417 if (!buf)
418 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
419 } else if (entry->delta_data) {
420 size = entry->delta_size;
421 buf = entry->delta_data;
422 entry->delta_data = NULL;
423 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
424 OBJ_OFS_DELTA : OBJ_REF_DELTA;
425 } else {
426 buf = read_sha1_file(entry->idx.sha1, &type, &size);
427 if (!buf)
428 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
429 buf = delta_against(buf, size, entry);
430 size = entry->delta_size;
431 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
432 OBJ_OFS_DELTA : OBJ_REF_DELTA;
434 /* compress the data to store and put compressed length in datalen */
435 memset(&stream, 0, sizeof(stream));
436 deflateInit(&stream, pack_compression_level);
437 maxsize = deflateBound(&stream, size);
438 out = xmalloc(maxsize);
439 /* Compress it */
440 stream.next_in = buf;
441 stream.avail_in = size;
442 stream.next_out = out;
443 stream.avail_out = maxsize;
444 while (deflate(&stream, Z_FINISH) == Z_OK)
445 /* nothing */;
446 deflateEnd(&stream);
447 datalen = stream.total_out;
450 * The object header is a byte of 'type' followed by zero or
451 * more bytes of length.
453 hdrlen = encode_header(obj_type, size, header);
455 if (obj_type == OBJ_OFS_DELTA) {
457 * Deltas with relative base contain an additional
458 * encoding of the relative offset for the delta
459 * base from this object's position in the pack.
461 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
462 unsigned pos = sizeof(dheader) - 1;
463 dheader[pos] = ofs & 127;
464 while (ofs >>= 7)
465 dheader[--pos] = 128 | (--ofs & 127);
466 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
467 free(out);
468 free(buf);
469 return 0;
471 sha1write(f, header, hdrlen);
472 sha1write(f, dheader + pos, sizeof(dheader) - pos);
473 hdrlen += sizeof(dheader) - pos;
474 } else if (obj_type == OBJ_REF_DELTA) {
476 * Deltas with a base reference contain
477 * an additional 20 bytes for the base sha1.
479 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
480 free(out);
481 free(buf);
482 return 0;
484 sha1write(f, header, hdrlen);
485 sha1write(f, entry->delta->idx.sha1, 20);
486 hdrlen += 20;
487 } else {
488 if (limit && hdrlen + datalen + 20 >= limit) {
489 free(out);
490 free(buf);
491 return 0;
493 sha1write(f, header, hdrlen);
495 sha1write(f, out, datalen);
496 free(out);
497 free(buf);
499 else {
500 struct packed_git *p = entry->in_pack;
501 struct pack_window *w_curs = NULL;
502 struct revindex_entry *revidx;
503 off_t offset;
505 if (entry->delta) {
506 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
507 OBJ_OFS_DELTA : OBJ_REF_DELTA;
508 reused_delta++;
510 hdrlen = encode_header(obj_type, entry->size, header);
511 offset = entry->in_pack_offset;
512 revidx = find_packed_object(p, offset);
513 datalen = revidx[1].offset - offset;
514 if (!pack_to_stdout && p->index_version > 1 &&
515 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
516 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
517 offset += entry->in_pack_header_size;
518 datalen -= entry->in_pack_header_size;
519 if (obj_type == OBJ_OFS_DELTA) {
520 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
521 unsigned pos = sizeof(dheader) - 1;
522 dheader[pos] = ofs & 127;
523 while (ofs >>= 7)
524 dheader[--pos] = 128 | (--ofs & 127);
525 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
526 return 0;
527 sha1write(f, header, hdrlen);
528 sha1write(f, dheader + pos, sizeof(dheader) - pos);
529 hdrlen += sizeof(dheader) - pos;
530 } else if (obj_type == OBJ_REF_DELTA) {
531 if (limit && hdrlen + 20 + datalen + 20 >= limit)
532 return 0;
533 sha1write(f, header, hdrlen);
534 sha1write(f, entry->delta->idx.sha1, 20);
535 hdrlen += 20;
536 } else {
537 if (limit && hdrlen + datalen + 20 >= limit)
538 return 0;
539 sha1write(f, header, hdrlen);
542 if (!pack_to_stdout && p->index_version == 1 &&
543 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
544 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
545 copy_pack_data(f, p, &w_curs, offset, datalen);
546 unuse_pack(&w_curs);
547 reused++;
549 if (usable_delta)
550 written_delta++;
551 written++;
552 if (!pack_to_stdout)
553 entry->idx.crc32 = crc32_end(f);
554 return hdrlen + datalen;
557 static off_t write_one(struct sha1file *f,
558 struct object_entry *e,
559 off_t offset)
561 unsigned long size;
563 /* offset is non zero if object is written already. */
564 if (e->idx.offset || e->preferred_base)
565 return offset;
567 /* if we are deltified, write out base object first. */
568 if (e->delta) {
569 offset = write_one(f, e->delta, offset);
570 if (!offset)
571 return 0;
574 e->idx.offset = offset;
575 size = write_object(f, e, offset);
576 if (!size) {
577 e->idx.offset = 0;
578 return 0;
580 written_list[nr_written++] = &e->idx;
582 /* make sure off_t is sufficiently large not to wrap */
583 if (offset > offset + size)
584 die("pack too large for current definition of off_t");
585 return offset + size;
588 /* forward declaration for write_pack_file */
589 static int adjust_perm(const char *path, mode_t mode);
591 static void write_pack_file(void)
593 uint32_t i = 0, j;
594 struct sha1file *f;
595 off_t offset, offset_one, last_obj_offset = 0;
596 struct pack_header hdr;
597 int do_progress = progress >> pack_to_stdout;
598 uint32_t nr_remaining = nr_result;
600 if (do_progress)
601 progress_state = start_progress("Writing objects", nr_result);
602 written_list = xmalloc(nr_objects * sizeof(*written_list));
604 do {
605 unsigned char sha1[20];
606 char *pack_tmp_name = NULL;
608 if (pack_to_stdout) {
609 f = sha1fd_throughput(1, "<stdout>", progress_state);
610 } else {
611 char tmpname[PATH_MAX];
612 int fd;
613 snprintf(tmpname, sizeof(tmpname),
614 "%s/tmp_pack_XXXXXX", get_object_directory());
615 fd = xmkstemp(tmpname);
616 pack_tmp_name = xstrdup(tmpname);
617 f = sha1fd(fd, pack_tmp_name);
620 hdr.hdr_signature = htonl(PACK_SIGNATURE);
621 hdr.hdr_version = htonl(PACK_VERSION);
622 hdr.hdr_entries = htonl(nr_remaining);
623 sha1write(f, &hdr, sizeof(hdr));
624 offset = sizeof(hdr);
625 nr_written = 0;
626 for (; i < nr_objects; i++) {
627 last_obj_offset = offset;
628 offset_one = write_one(f, objects + i, offset);
629 if (!offset_one)
630 break;
631 offset = offset_one;
632 display_progress(progress_state, written);
636 * Did we write the wrong # entries in the header?
637 * If so, rewrite it like in fast-import
639 if (pack_to_stdout || nr_written == nr_remaining) {
640 sha1close(f, sha1, 1);
641 } else {
642 int fd = sha1close(f, NULL, 0);
643 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
644 close(fd);
647 if (!pack_to_stdout) {
648 mode_t mode = umask(0);
649 char *idx_tmp_name, tmpname[PATH_MAX];
651 umask(mode);
652 mode = 0444 & ~mode;
654 idx_tmp_name = write_idx_file(NULL, written_list,
655 nr_written, sha1);
656 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
657 base_name, sha1_to_hex(sha1));
658 if (adjust_perm(pack_tmp_name, mode))
659 die("unable to make temporary pack file readable: %s",
660 strerror(errno));
661 if (rename(pack_tmp_name, tmpname))
662 die("unable to rename temporary pack file: %s",
663 strerror(errno));
664 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
665 base_name, sha1_to_hex(sha1));
666 if (adjust_perm(idx_tmp_name, mode))
667 die("unable to make temporary index file readable: %s",
668 strerror(errno));
669 if (rename(idx_tmp_name, tmpname))
670 die("unable to rename temporary index file: %s",
671 strerror(errno));
672 free(idx_tmp_name);
673 free(pack_tmp_name);
674 puts(sha1_to_hex(sha1));
677 /* mark written objects as written to previous pack */
678 for (j = 0; j < nr_written; j++) {
679 written_list[j]->offset = (off_t)-1;
681 nr_remaining -= nr_written;
682 } while (nr_remaining && i < nr_objects);
684 free(written_list);
685 stop_progress(&progress_state);
686 if (written != nr_result)
687 die("wrote %u objects while expecting %u", written, nr_result);
689 * We have scanned through [0 ... i). Since we have written
690 * the correct number of objects, the remaining [i ... nr_objects)
691 * items must be either already written (due to out-of-order delta base)
692 * or a preferred base. Count those which are neither and complain if any.
694 for (j = 0; i < nr_objects; i++) {
695 struct object_entry *e = objects + i;
696 j += !e->idx.offset && !e->preferred_base;
698 if (j)
699 die("wrote %u objects as expected but %u unwritten", written, j);
702 static int locate_object_entry_hash(const unsigned char *sha1)
704 int i;
705 unsigned int ui;
706 memcpy(&ui, sha1, sizeof(unsigned int));
707 i = ui % object_ix_hashsz;
708 while (0 < object_ix[i]) {
709 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
710 return i;
711 if (++i == object_ix_hashsz)
712 i = 0;
714 return -1 - i;
717 static struct object_entry *locate_object_entry(const unsigned char *sha1)
719 int i;
721 if (!object_ix_hashsz)
722 return NULL;
724 i = locate_object_entry_hash(sha1);
725 if (0 <= i)
726 return &objects[object_ix[i]-1];
727 return NULL;
730 static void rehash_objects(void)
732 uint32_t i;
733 struct object_entry *oe;
735 object_ix_hashsz = nr_objects * 3;
736 if (object_ix_hashsz < 1024)
737 object_ix_hashsz = 1024;
738 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
739 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
740 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
741 int ix = locate_object_entry_hash(oe->idx.sha1);
742 if (0 <= ix)
743 continue;
744 ix = -1 - ix;
745 object_ix[ix] = i + 1;
749 static unsigned name_hash(const char *name)
751 unsigned char c;
752 unsigned hash = 0;
754 if (!name)
755 return 0;
758 * This effectively just creates a sortable number from the
759 * last sixteen non-whitespace characters. Last characters
760 * count "most", so things that end in ".c" sort together.
762 while ((c = *name++) != 0) {
763 if (isspace(c))
764 continue;
765 hash = (hash >> 2) + (c << 24);
767 return hash;
770 static void setup_delta_attr_check(struct git_attr_check *check)
772 static struct git_attr *attr_delta;
774 if (!attr_delta)
775 attr_delta = git_attr("delta", 5);
777 check[0].attr = attr_delta;
780 static int no_try_delta(const char *path)
782 struct git_attr_check check[1];
784 setup_delta_attr_check(check);
785 if (git_checkattr(path, ARRAY_SIZE(check), check))
786 return 0;
787 if (ATTR_FALSE(check->value))
788 return 1;
789 return 0;
792 static int add_object_entry(const unsigned char *sha1, enum object_type type,
793 const char *name, int exclude)
795 struct object_entry *entry;
796 struct packed_git *p, *found_pack = NULL;
797 off_t found_offset = 0;
798 int ix;
799 unsigned hash = name_hash(name);
801 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
802 if (ix >= 0) {
803 if (exclude) {
804 entry = objects + object_ix[ix] - 1;
805 if (!entry->preferred_base)
806 nr_result--;
807 entry->preferred_base = 1;
809 return 0;
812 for (p = packed_git; p; p = p->next) {
813 off_t offset = find_pack_entry_one(sha1, p);
814 if (offset) {
815 if (!found_pack) {
816 found_offset = offset;
817 found_pack = p;
819 if (exclude)
820 break;
821 if (incremental)
822 return 0;
823 if (local && !p->pack_local)
824 return 0;
828 if (nr_objects >= nr_alloc) {
829 nr_alloc = (nr_alloc + 1024) * 3 / 2;
830 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
833 entry = objects + nr_objects++;
834 memset(entry, 0, sizeof(*entry));
835 hashcpy(entry->idx.sha1, sha1);
836 entry->hash = hash;
837 if (type)
838 entry->type = type;
839 if (exclude)
840 entry->preferred_base = 1;
841 else
842 nr_result++;
843 if (found_pack) {
844 entry->in_pack = found_pack;
845 entry->in_pack_offset = found_offset;
848 if (object_ix_hashsz * 3 <= nr_objects * 4)
849 rehash_objects();
850 else
851 object_ix[-1 - ix] = nr_objects;
853 display_progress(progress_state, nr_objects);
855 if (name && no_try_delta(name))
856 entry->no_try_delta = 1;
858 return 1;
861 struct pbase_tree_cache {
862 unsigned char sha1[20];
863 int ref;
864 int temporary;
865 void *tree_data;
866 unsigned long tree_size;
869 static struct pbase_tree_cache *(pbase_tree_cache[256]);
870 static int pbase_tree_cache_ix(const unsigned char *sha1)
872 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
874 static int pbase_tree_cache_ix_incr(int ix)
876 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
879 static struct pbase_tree {
880 struct pbase_tree *next;
881 /* This is a phony "cache" entry; we are not
882 * going to evict it nor find it through _get()
883 * mechanism -- this is for the toplevel node that
884 * would almost always change with any commit.
886 struct pbase_tree_cache pcache;
887 } *pbase_tree;
889 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
891 struct pbase_tree_cache *ent, *nent;
892 void *data;
893 unsigned long size;
894 enum object_type type;
895 int neigh;
896 int my_ix = pbase_tree_cache_ix(sha1);
897 int available_ix = -1;
899 /* pbase-tree-cache acts as a limited hashtable.
900 * your object will be found at your index or within a few
901 * slots after that slot if it is cached.
903 for (neigh = 0; neigh < 8; neigh++) {
904 ent = pbase_tree_cache[my_ix];
905 if (ent && !hashcmp(ent->sha1, sha1)) {
906 ent->ref++;
907 return ent;
909 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
910 ((0 <= available_ix) &&
911 (!ent && pbase_tree_cache[available_ix])))
912 available_ix = my_ix;
913 if (!ent)
914 break;
915 my_ix = pbase_tree_cache_ix_incr(my_ix);
918 /* Did not find one. Either we got a bogus request or
919 * we need to read and perhaps cache.
921 data = read_sha1_file(sha1, &type, &size);
922 if (!data)
923 return NULL;
924 if (type != OBJ_TREE) {
925 free(data);
926 return NULL;
929 /* We need to either cache or return a throwaway copy */
931 if (available_ix < 0)
932 ent = NULL;
933 else {
934 ent = pbase_tree_cache[available_ix];
935 my_ix = available_ix;
938 if (!ent) {
939 nent = xmalloc(sizeof(*nent));
940 nent->temporary = (available_ix < 0);
942 else {
943 /* evict and reuse */
944 free(ent->tree_data);
945 nent = ent;
947 hashcpy(nent->sha1, sha1);
948 nent->tree_data = data;
949 nent->tree_size = size;
950 nent->ref = 1;
951 if (!nent->temporary)
952 pbase_tree_cache[my_ix] = nent;
953 return nent;
956 static void pbase_tree_put(struct pbase_tree_cache *cache)
958 if (!cache->temporary) {
959 cache->ref--;
960 return;
962 free(cache->tree_data);
963 free(cache);
966 static int name_cmp_len(const char *name)
968 int i;
969 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
971 return i;
974 static void add_pbase_object(struct tree_desc *tree,
975 const char *name,
976 int cmplen,
977 const char *fullname)
979 struct name_entry entry;
980 int cmp;
982 while (tree_entry(tree,&entry)) {
983 if (S_ISGITLINK(entry.mode))
984 continue;
985 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
986 memcmp(name, entry.path, cmplen);
987 if (cmp > 0)
988 continue;
989 if (cmp < 0)
990 return;
991 if (name[cmplen] != '/') {
992 add_object_entry(entry.sha1,
993 object_type(entry.mode),
994 fullname, 1);
995 return;
997 if (S_ISDIR(entry.mode)) {
998 struct tree_desc sub;
999 struct pbase_tree_cache *tree;
1000 const char *down = name+cmplen+1;
1001 int downlen = name_cmp_len(down);
1003 tree = pbase_tree_get(entry.sha1);
1004 if (!tree)
1005 return;
1006 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1008 add_pbase_object(&sub, down, downlen, fullname);
1009 pbase_tree_put(tree);
1014 static unsigned *done_pbase_paths;
1015 static int done_pbase_paths_num;
1016 static int done_pbase_paths_alloc;
1017 static int done_pbase_path_pos(unsigned hash)
1019 int lo = 0;
1020 int hi = done_pbase_paths_num;
1021 while (lo < hi) {
1022 int mi = (hi + lo) / 2;
1023 if (done_pbase_paths[mi] == hash)
1024 return mi;
1025 if (done_pbase_paths[mi] < hash)
1026 hi = mi;
1027 else
1028 lo = mi + 1;
1030 return -lo-1;
1033 static int check_pbase_path(unsigned hash)
1035 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1036 if (0 <= pos)
1037 return 1;
1038 pos = -pos - 1;
1039 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1040 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1041 done_pbase_paths = xrealloc(done_pbase_paths,
1042 done_pbase_paths_alloc *
1043 sizeof(unsigned));
1045 done_pbase_paths_num++;
1046 if (pos < done_pbase_paths_num)
1047 memmove(done_pbase_paths + pos + 1,
1048 done_pbase_paths + pos,
1049 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1050 done_pbase_paths[pos] = hash;
1051 return 0;
1054 static void add_preferred_base_object(const char *name)
1056 struct pbase_tree *it;
1057 int cmplen;
1058 unsigned hash = name_hash(name);
1060 if (!num_preferred_base || check_pbase_path(hash))
1061 return;
1063 cmplen = name_cmp_len(name);
1064 for (it = pbase_tree; it; it = it->next) {
1065 if (cmplen == 0) {
1066 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1068 else {
1069 struct tree_desc tree;
1070 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1071 add_pbase_object(&tree, name, cmplen, name);
1076 static void add_preferred_base(unsigned char *sha1)
1078 struct pbase_tree *it;
1079 void *data;
1080 unsigned long size;
1081 unsigned char tree_sha1[20];
1083 if (window <= num_preferred_base++)
1084 return;
1086 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1087 if (!data)
1088 return;
1090 for (it = pbase_tree; it; it = it->next) {
1091 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1092 free(data);
1093 return;
1097 it = xcalloc(1, sizeof(*it));
1098 it->next = pbase_tree;
1099 pbase_tree = it;
1101 hashcpy(it->pcache.sha1, tree_sha1);
1102 it->pcache.tree_data = data;
1103 it->pcache.tree_size = size;
1106 static void check_object(struct object_entry *entry)
1108 if (entry->in_pack) {
1109 struct packed_git *p = entry->in_pack;
1110 struct pack_window *w_curs = NULL;
1111 const unsigned char *base_ref = NULL;
1112 struct object_entry *base_entry;
1113 unsigned long used, used_0;
1114 unsigned int avail;
1115 off_t ofs;
1116 unsigned char *buf, c;
1118 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1121 * We want in_pack_type even if we do not reuse delta
1122 * since non-delta representations could still be reused.
1124 used = unpack_object_header_gently(buf, avail,
1125 &entry->in_pack_type,
1126 &entry->size);
1129 * Determine if this is a delta and if so whether we can
1130 * reuse it or not. Otherwise let's find out as cheaply as
1131 * possible what the actual type and size for this object is.
1133 switch (entry->in_pack_type) {
1134 default:
1135 /* Not a delta hence we've already got all we need. */
1136 entry->type = entry->in_pack_type;
1137 entry->in_pack_header_size = used;
1138 unuse_pack(&w_curs);
1139 return;
1140 case OBJ_REF_DELTA:
1141 if (!no_reuse_delta && !entry->preferred_base)
1142 base_ref = use_pack(p, &w_curs,
1143 entry->in_pack_offset + used, NULL);
1144 entry->in_pack_header_size = used + 20;
1145 break;
1146 case OBJ_OFS_DELTA:
1147 buf = use_pack(p, &w_curs,
1148 entry->in_pack_offset + used, NULL);
1149 used_0 = 0;
1150 c = buf[used_0++];
1151 ofs = c & 127;
1152 while (c & 128) {
1153 ofs += 1;
1154 if (!ofs || MSB(ofs, 7))
1155 die("delta base offset overflow in pack for %s",
1156 sha1_to_hex(entry->idx.sha1));
1157 c = buf[used_0++];
1158 ofs = (ofs << 7) + (c & 127);
1160 if (ofs >= entry->in_pack_offset)
1161 die("delta base offset out of bound for %s",
1162 sha1_to_hex(entry->idx.sha1));
1163 ofs = entry->in_pack_offset - ofs;
1164 if (!no_reuse_delta && !entry->preferred_base)
1165 base_ref = find_packed_object_name(p, ofs);
1166 entry->in_pack_header_size = used + used_0;
1167 break;
1170 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1172 * If base_ref was set above that means we wish to
1173 * reuse delta data, and we even found that base
1174 * in the list of objects we want to pack. Goodie!
1176 * Depth value does not matter - find_deltas() will
1177 * never consider reused delta as the base object to
1178 * deltify other objects against, in order to avoid
1179 * circular deltas.
1181 entry->type = entry->in_pack_type;
1182 entry->delta = base_entry;
1183 entry->delta_sibling = base_entry->delta_child;
1184 base_entry->delta_child = entry;
1185 unuse_pack(&w_curs);
1186 return;
1189 if (entry->type) {
1191 * This must be a delta and we already know what the
1192 * final object type is. Let's extract the actual
1193 * object size from the delta header.
1195 entry->size = get_size_from_delta(p, &w_curs,
1196 entry->in_pack_offset + entry->in_pack_header_size);
1197 unuse_pack(&w_curs);
1198 return;
1202 * No choice but to fall back to the recursive delta walk
1203 * with sha1_object_info() to find about the object type
1204 * at this point...
1206 unuse_pack(&w_curs);
1209 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1210 if (entry->type < 0)
1211 die("unable to get type of object %s",
1212 sha1_to_hex(entry->idx.sha1));
1215 static int pack_offset_sort(const void *_a, const void *_b)
1217 const struct object_entry *a = *(struct object_entry **)_a;
1218 const struct object_entry *b = *(struct object_entry **)_b;
1220 /* avoid filesystem trashing with loose objects */
1221 if (!a->in_pack && !b->in_pack)
1222 return hashcmp(a->idx.sha1, b->idx.sha1);
1224 if (a->in_pack < b->in_pack)
1225 return -1;
1226 if (a->in_pack > b->in_pack)
1227 return 1;
1228 return a->in_pack_offset < b->in_pack_offset ? -1 :
1229 (a->in_pack_offset > b->in_pack_offset);
1232 static void get_object_details(void)
1234 uint32_t i;
1235 struct object_entry **sorted_by_offset;
1237 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1238 for (i = 0; i < nr_objects; i++)
1239 sorted_by_offset[i] = objects + i;
1240 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1242 prepare_pack_ix();
1243 for (i = 0; i < nr_objects; i++)
1244 check_object(sorted_by_offset[i]);
1245 free(sorted_by_offset);
1249 * We search for deltas in a list sorted by type, by filename hash, and then
1250 * by size, so that we see progressively smaller and smaller files.
1251 * That's because we prefer deltas to be from the bigger file
1252 * to the smaller -- deletes are potentially cheaper, but perhaps
1253 * more importantly, the bigger file is likely the more recent
1254 * one. The deepest deltas are therefore the oldest objects which are
1255 * less susceptible to be accessed often.
1257 static int type_size_sort(const void *_a, const void *_b)
1259 const struct object_entry *a = *(struct object_entry **)_a;
1260 const struct object_entry *b = *(struct object_entry **)_b;
1262 if (a->type > b->type)
1263 return -1;
1264 if (a->type < b->type)
1265 return 1;
1266 if (a->hash > b->hash)
1267 return -1;
1268 if (a->hash < b->hash)
1269 return 1;
1270 if (a->preferred_base > b->preferred_base)
1271 return -1;
1272 if (a->preferred_base < b->preferred_base)
1273 return 1;
1274 if (a->size > b->size)
1275 return -1;
1276 if (a->size < b->size)
1277 return 1;
1278 return a < b ? -1 : (a > b); /* newest first */
1281 struct unpacked {
1282 struct object_entry *entry;
1283 void *data;
1284 struct delta_index *index;
1285 unsigned depth;
1288 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1289 unsigned long delta_size)
1291 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1292 return 0;
1294 if (delta_size < cache_max_small_delta_size)
1295 return 1;
1297 /* cache delta, if objects are large enough compared to delta size */
1298 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1299 return 1;
1301 return 0;
1304 #ifdef THREADED_DELTA_SEARCH
1306 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1307 #define read_lock() pthread_mutex_lock(&read_mutex)
1308 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1310 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1311 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1312 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1314 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1315 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1316 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1318 #else
1320 #define read_lock() (void)0
1321 #define read_unlock() (void)0
1322 #define cache_lock() (void)0
1323 #define cache_unlock() (void)0
1324 #define progress_lock() (void)0
1325 #define progress_unlock() (void)0
1327 #endif
1329 static int try_delta(struct unpacked *trg, struct unpacked *src,
1330 unsigned max_depth, unsigned long *mem_usage)
1332 struct object_entry *trg_entry = trg->entry;
1333 struct object_entry *src_entry = src->entry;
1334 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1335 unsigned ref_depth;
1336 enum object_type type;
1337 void *delta_buf;
1339 /* Don't bother doing diffs between different types */
1340 if (trg_entry->type != src_entry->type)
1341 return -1;
1344 * We do not bother to try a delta that we discarded
1345 * on an earlier try, but only when reusing delta data.
1347 if (!no_reuse_delta && trg_entry->in_pack &&
1348 trg_entry->in_pack == src_entry->in_pack &&
1349 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1350 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1351 return 0;
1353 /* Let's not bust the allowed depth. */
1354 if (src->depth >= max_depth)
1355 return 0;
1357 /* Now some size filtering heuristics. */
1358 trg_size = trg_entry->size;
1359 if (!trg_entry->delta) {
1360 max_size = trg_size/2 - 20;
1361 ref_depth = 1;
1362 } else {
1363 max_size = trg_entry->delta_size;
1364 ref_depth = trg->depth;
1366 max_size = max_size * (max_depth - src->depth) /
1367 (max_depth - ref_depth + 1);
1368 if (max_size == 0)
1369 return 0;
1370 src_size = src_entry->size;
1371 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1372 if (sizediff >= max_size)
1373 return 0;
1374 if (trg_size < src_size / 32)
1375 return 0;
1377 /* Load data if not already done */
1378 if (!trg->data) {
1379 read_lock();
1380 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1381 read_unlock();
1382 if (!trg->data)
1383 die("object %s cannot be read",
1384 sha1_to_hex(trg_entry->idx.sha1));
1385 if (sz != trg_size)
1386 die("object %s inconsistent object length (%lu vs %lu)",
1387 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1388 *mem_usage += sz;
1390 if (!src->data) {
1391 read_lock();
1392 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1393 read_unlock();
1394 if (!src->data)
1395 die("object %s cannot be read",
1396 sha1_to_hex(src_entry->idx.sha1));
1397 if (sz != src_size)
1398 die("object %s inconsistent object length (%lu vs %lu)",
1399 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1400 *mem_usage += sz;
1402 if (!src->index) {
1403 src->index = create_delta_index(src->data, src_size);
1404 if (!src->index) {
1405 static int warned = 0;
1406 if (!warned++)
1407 warning("suboptimal pack - out of memory");
1408 return 0;
1410 *mem_usage += sizeof_delta_index(src->index);
1413 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1414 if (!delta_buf)
1415 return 0;
1417 if (trg_entry->delta) {
1418 /* Prefer only shallower same-sized deltas. */
1419 if (delta_size == trg_entry->delta_size &&
1420 src->depth + 1 >= trg->depth) {
1421 free(delta_buf);
1422 return 0;
1427 * Handle memory allocation outside of the cache
1428 * accounting lock. Compiler will optimize the strangeness
1429 * away when THREADED_DELTA_SEARCH is not defined.
1431 if (trg_entry->delta_data)
1432 free(trg_entry->delta_data);
1433 cache_lock();
1434 if (trg_entry->delta_data) {
1435 delta_cache_size -= trg_entry->delta_size;
1436 trg_entry->delta_data = NULL;
1438 if (delta_cacheable(src_size, trg_size, delta_size)) {
1439 delta_cache_size += delta_size;
1440 cache_unlock();
1441 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1442 } else {
1443 cache_unlock();
1444 free(delta_buf);
1447 trg_entry->delta = src_entry;
1448 trg_entry->delta_size = delta_size;
1449 trg->depth = src->depth + 1;
1451 return 1;
1454 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1456 struct object_entry *child = me->delta_child;
1457 unsigned int m = n;
1458 while (child) {
1459 unsigned int c = check_delta_limit(child, n + 1);
1460 if (m < c)
1461 m = c;
1462 child = child->delta_sibling;
1464 return m;
1467 static unsigned long free_unpacked(struct unpacked *n)
1469 unsigned long freed_mem = sizeof_delta_index(n->index);
1470 free_delta_index(n->index);
1471 n->index = NULL;
1472 if (n->data) {
1473 freed_mem += n->entry->size;
1474 free(n->data);
1475 n->data = NULL;
1477 n->entry = NULL;
1478 n->depth = 0;
1479 return freed_mem;
1482 static void find_deltas(struct object_entry **list, unsigned *list_size,
1483 int window, int depth, unsigned *processed)
1485 uint32_t i, idx = 0, count = 0;
1486 unsigned int array_size = window * sizeof(struct unpacked);
1487 struct unpacked *array;
1488 unsigned long mem_usage = 0;
1490 array = xmalloc(array_size);
1491 memset(array, 0, array_size);
1493 for (;;) {
1494 struct object_entry *entry = *list++;
1495 struct unpacked *n = array + idx;
1496 int j, max_depth, best_base = -1;
1498 progress_lock();
1499 if (!*list_size) {
1500 progress_unlock();
1501 break;
1503 (*list_size)--;
1504 if (!entry->preferred_base) {
1505 (*processed)++;
1506 display_progress(progress_state, *processed);
1508 progress_unlock();
1510 mem_usage -= free_unpacked(n);
1511 n->entry = entry;
1513 while (window_memory_limit &&
1514 mem_usage > window_memory_limit &&
1515 count > 1) {
1516 uint32_t tail = (idx + window - count) % window;
1517 mem_usage -= free_unpacked(array + tail);
1518 count--;
1521 /* We do not compute delta to *create* objects we are not
1522 * going to pack.
1524 if (entry->preferred_base)
1525 goto next;
1528 * If the current object is at pack edge, take the depth the
1529 * objects that depend on the current object into account
1530 * otherwise they would become too deep.
1532 max_depth = depth;
1533 if (entry->delta_child) {
1534 max_depth -= check_delta_limit(entry, 0);
1535 if (max_depth <= 0)
1536 goto next;
1539 j = window;
1540 while (--j > 0) {
1541 int ret;
1542 uint32_t other_idx = idx + j;
1543 struct unpacked *m;
1544 if (other_idx >= window)
1545 other_idx -= window;
1546 m = array + other_idx;
1547 if (!m->entry)
1548 break;
1549 ret = try_delta(n, m, max_depth, &mem_usage);
1550 if (ret < 0)
1551 break;
1552 else if (ret > 0)
1553 best_base = other_idx;
1556 /* if we made n a delta, and if n is already at max
1557 * depth, leaving it in the window is pointless. we
1558 * should evict it first.
1560 if (entry->delta && depth <= n->depth)
1561 continue;
1564 * Move the best delta base up in the window, after the
1565 * currently deltified object, to keep it longer. It will
1566 * be the first base object to be attempted next.
1568 if (entry->delta) {
1569 struct unpacked swap = array[best_base];
1570 int dist = (window + idx - best_base) % window;
1571 int dst = best_base;
1572 while (dist--) {
1573 int src = (dst + 1) % window;
1574 array[dst] = array[src];
1575 dst = src;
1577 array[dst] = swap;
1580 next:
1581 idx++;
1582 if (count + 1 < window)
1583 count++;
1584 if (idx >= window)
1585 idx = 0;
1588 for (i = 0; i < window; ++i) {
1589 free_delta_index(array[i].index);
1590 free(array[i].data);
1592 free(array);
1595 #ifdef THREADED_DELTA_SEARCH
1598 * The main thread waits on the condition that (at least) one of the workers
1599 * has stopped working (which is indicated in the .working member of
1600 * struct thread_params).
1601 * When a work thread has completed its work, it sets .working to 0 and
1602 * signals the main thread and waits on the condition that .data_ready
1603 * becomes 1.
1606 struct thread_params {
1607 pthread_t thread;
1608 struct object_entry **list;
1609 unsigned list_size;
1610 unsigned remaining;
1611 int window;
1612 int depth;
1613 int working;
1614 int data_ready;
1615 pthread_mutex_t mutex;
1616 pthread_cond_t cond;
1617 unsigned *processed;
1620 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1622 static void *threaded_find_deltas(void *arg)
1624 struct thread_params *me = arg;
1626 while (me->remaining) {
1627 find_deltas(me->list, &me->remaining,
1628 me->window, me->depth, me->processed);
1630 progress_lock();
1631 me->working = 0;
1632 pthread_cond_signal(&progress_cond);
1633 progress_unlock();
1636 * We must not set ->data_ready before we wait on the
1637 * condition because the main thread may have set it to 1
1638 * before we get here. In order to be sure that new
1639 * work is available if we see 1 in ->data_ready, it
1640 * was initialized to 0 before this thread was spawned
1641 * and we reset it to 0 right away.
1643 pthread_mutex_lock(&me->mutex);
1644 while (!me->data_ready)
1645 pthread_cond_wait(&me->cond, &me->mutex);
1646 me->data_ready = 0;
1647 pthread_mutex_unlock(&me->mutex);
1649 /* leave ->working 1 so that this doesn't get more work assigned */
1650 return NULL;
1653 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1654 int window, int depth, unsigned *processed)
1656 struct thread_params p[delta_search_threads];
1657 int i, ret, active_threads = 0;
1659 if (delta_search_threads <= 1) {
1660 find_deltas(list, &list_size, window, depth, processed);
1661 return;
1664 /* Partition the work amongst work threads. */
1665 for (i = 0; i < delta_search_threads; i++) {
1666 unsigned sub_size = list_size / (delta_search_threads - i);
1668 p[i].window = window;
1669 p[i].depth = depth;
1670 p[i].processed = processed;
1671 p[i].working = 1;
1672 p[i].data_ready = 0;
1674 /* try to split chunks on "path" boundaries */
1675 while (sub_size && sub_size < list_size &&
1676 list[sub_size]->hash &&
1677 list[sub_size]->hash == list[sub_size-1]->hash)
1678 sub_size++;
1680 p[i].list = list;
1681 p[i].list_size = sub_size;
1682 p[i].remaining = sub_size;
1684 list += sub_size;
1685 list_size -= sub_size;
1688 /* Start work threads. */
1689 for (i = 0; i < delta_search_threads; i++) {
1690 if (!p[i].list_size)
1691 continue;
1692 pthread_mutex_init(&p[i].mutex, NULL);
1693 pthread_cond_init(&p[i].cond, NULL);
1694 ret = pthread_create(&p[i].thread, NULL,
1695 threaded_find_deltas, &p[i]);
1696 if (ret)
1697 die("unable to create thread: %s", strerror(ret));
1698 active_threads++;
1702 * Now let's wait for work completion. Each time a thread is done
1703 * with its work, we steal half of the remaining work from the
1704 * thread with the largest number of unprocessed objects and give
1705 * it to that newly idle thread. This ensure good load balancing
1706 * until the remaining object list segments are simply too short
1707 * to be worth splitting anymore.
1709 while (active_threads) {
1710 struct thread_params *target = NULL;
1711 struct thread_params *victim = NULL;
1712 unsigned sub_size = 0;
1714 progress_lock();
1715 for (;;) {
1716 for (i = 0; !target && i < delta_search_threads; i++)
1717 if (!p[i].working)
1718 target = &p[i];
1719 if (target)
1720 break;
1721 pthread_cond_wait(&progress_cond, &progress_mutex);
1724 for (i = 0; i < delta_search_threads; i++)
1725 if (p[i].remaining > 2*window &&
1726 (!victim || victim->remaining < p[i].remaining))
1727 victim = &p[i];
1728 if (victim) {
1729 sub_size = victim->remaining / 2;
1730 list = victim->list + victim->list_size - sub_size;
1731 while (sub_size && list[0]->hash &&
1732 list[0]->hash == list[-1]->hash) {
1733 list++;
1734 sub_size--;
1736 if (!sub_size) {
1738 * It is possible for some "paths" to have
1739 * so many objects that no hash boundary
1740 * might be found. Let's just steal the
1741 * exact half in that case.
1743 sub_size = victim->remaining / 2;
1744 list -= sub_size;
1746 target->list = list;
1747 victim->list_size -= sub_size;
1748 victim->remaining -= sub_size;
1750 target->list_size = sub_size;
1751 target->remaining = sub_size;
1752 target->working = 1;
1753 progress_unlock();
1755 pthread_mutex_lock(&target->mutex);
1756 target->data_ready = 1;
1757 pthread_cond_signal(&target->cond);
1758 pthread_mutex_unlock(&target->mutex);
1760 if (!sub_size) {
1761 pthread_join(target->thread, NULL);
1762 pthread_cond_destroy(&target->cond);
1763 pthread_mutex_destroy(&target->mutex);
1764 active_threads--;
1769 #else
1770 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1771 #endif
1773 static void prepare_pack(int window, int depth)
1775 struct object_entry **delta_list;
1776 uint32_t i, n, nr_deltas;
1778 get_object_details();
1780 if (!nr_objects || !window || !depth)
1781 return;
1783 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1784 nr_deltas = n = 0;
1786 for (i = 0; i < nr_objects; i++) {
1787 struct object_entry *entry = objects + i;
1789 if (entry->delta)
1790 /* This happens if we decided to reuse existing
1791 * delta from a pack. "!no_reuse_delta &&" is implied.
1793 continue;
1795 if (entry->size < 50)
1796 continue;
1798 if (entry->no_try_delta)
1799 continue;
1801 if (!entry->preferred_base)
1802 nr_deltas++;
1804 delta_list[n++] = entry;
1807 if (nr_deltas && n > 1) {
1808 unsigned nr_done = 0;
1809 if (progress)
1810 progress_state = start_progress("Compressing objects",
1811 nr_deltas);
1812 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1813 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1814 stop_progress(&progress_state);
1815 if (nr_done != nr_deltas)
1816 die("inconsistency with delta count");
1818 free(delta_list);
1821 static int git_pack_config(const char *k, const char *v)
1823 if(!strcmp(k, "pack.window")) {
1824 window = git_config_int(k, v);
1825 return 0;
1827 if (!strcmp(k, "pack.windowmemory")) {
1828 window_memory_limit = git_config_ulong(k, v);
1829 return 0;
1831 if (!strcmp(k, "pack.depth")) {
1832 depth = git_config_int(k, v);
1833 return 0;
1835 if (!strcmp(k, "pack.compression")) {
1836 int level = git_config_int(k, v);
1837 if (level == -1)
1838 level = Z_DEFAULT_COMPRESSION;
1839 else if (level < 0 || level > Z_BEST_COMPRESSION)
1840 die("bad pack compression level %d", level);
1841 pack_compression_level = level;
1842 pack_compression_seen = 1;
1843 return 0;
1845 if (!strcmp(k, "pack.deltacachesize")) {
1846 max_delta_cache_size = git_config_int(k, v);
1847 return 0;
1849 if (!strcmp(k, "pack.deltacachelimit")) {
1850 cache_max_small_delta_size = git_config_int(k, v);
1851 return 0;
1853 if (!strcmp(k, "pack.threads")) {
1854 delta_search_threads = git_config_int(k, v);
1855 if (delta_search_threads < 1)
1856 die("invalid number of threads specified (%d)",
1857 delta_search_threads);
1858 #ifndef THREADED_DELTA_SEARCH
1859 if (delta_search_threads > 1)
1860 warning("no threads support, ignoring %s", k);
1861 #endif
1862 return 0;
1864 if (!strcmp(k, "pack.indexversion")) {
1865 pack_idx_default_version = git_config_int(k, v);
1866 if (pack_idx_default_version > 2)
1867 die("bad pack.indexversion=%d", pack_idx_default_version);
1868 return 0;
1870 return git_default_config(k, v);
1873 static void read_object_list_from_stdin(void)
1875 char line[40 + 1 + PATH_MAX + 2];
1876 unsigned char sha1[20];
1878 for (;;) {
1879 if (!fgets(line, sizeof(line), stdin)) {
1880 if (feof(stdin))
1881 break;
1882 if (!ferror(stdin))
1883 die("fgets returned NULL, not EOF, not error!");
1884 if (errno != EINTR)
1885 die("fgets: %s", strerror(errno));
1886 clearerr(stdin);
1887 continue;
1889 if (line[0] == '-') {
1890 if (get_sha1_hex(line+1, sha1))
1891 die("expected edge sha1, got garbage:\n %s",
1892 line);
1893 add_preferred_base(sha1);
1894 continue;
1896 if (get_sha1_hex(line, sha1))
1897 die("expected sha1, got garbage:\n %s", line);
1899 add_preferred_base_object(line+41);
1900 add_object_entry(sha1, 0, line+41, 0);
1904 #define OBJECT_ADDED (1u<<20)
1906 static void show_commit(struct commit *commit)
1908 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1909 commit->object.flags |= OBJECT_ADDED;
1912 static void show_object(struct object_array_entry *p)
1914 add_preferred_base_object(p->name);
1915 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1916 p->item->flags |= OBJECT_ADDED;
1919 static void show_edge(struct commit *commit)
1921 add_preferred_base(commit->object.sha1);
1924 struct in_pack_object {
1925 off_t offset;
1926 struct object *object;
1929 struct in_pack {
1930 int alloc;
1931 int nr;
1932 struct in_pack_object *array;
1935 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1937 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1938 in_pack->array[in_pack->nr].object = object;
1939 in_pack->nr++;
1943 * Compare the objects in the offset order, in order to emulate the
1944 * "git-rev-list --objects" output that produced the pack originally.
1946 static int ofscmp(const void *a_, const void *b_)
1948 struct in_pack_object *a = (struct in_pack_object *)a_;
1949 struct in_pack_object *b = (struct in_pack_object *)b_;
1951 if (a->offset < b->offset)
1952 return -1;
1953 else if (a->offset > b->offset)
1954 return 1;
1955 else
1956 return hashcmp(a->object->sha1, b->object->sha1);
1959 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1961 struct packed_git *p;
1962 struct in_pack in_pack;
1963 uint32_t i;
1965 memset(&in_pack, 0, sizeof(in_pack));
1967 for (p = packed_git; p; p = p->next) {
1968 const unsigned char *sha1;
1969 struct object *o;
1971 for (i = 0; i < revs->num_ignore_packed; i++) {
1972 if (matches_pack_name(p, revs->ignore_packed[i]))
1973 break;
1975 if (revs->num_ignore_packed <= i)
1976 continue;
1977 if (open_pack_index(p))
1978 die("cannot open pack index");
1980 ALLOC_GROW(in_pack.array,
1981 in_pack.nr + p->num_objects,
1982 in_pack.alloc);
1984 for (i = 0; i < p->num_objects; i++) {
1985 sha1 = nth_packed_object_sha1(p, i);
1986 o = lookup_unknown_object(sha1);
1987 if (!(o->flags & OBJECT_ADDED))
1988 mark_in_pack_object(o, p, &in_pack);
1989 o->flags |= OBJECT_ADDED;
1993 if (in_pack.nr) {
1994 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1995 ofscmp);
1996 for (i = 0; i < in_pack.nr; i++) {
1997 struct object *o = in_pack.array[i].object;
1998 add_object_entry(o->sha1, o->type, "", 0);
2001 free(in_pack.array);
2004 static void get_object_list(int ac, const char **av)
2006 struct rev_info revs;
2007 char line[1000];
2008 int flags = 0;
2010 init_revisions(&revs, NULL);
2011 save_commit_buffer = 0;
2012 track_object_refs = 0;
2013 setup_revisions(ac, av, &revs, NULL);
2015 while (fgets(line, sizeof(line), stdin) != NULL) {
2016 int len = strlen(line);
2017 if (len && line[len - 1] == '\n')
2018 line[--len] = 0;
2019 if (!len)
2020 break;
2021 if (*line == '-') {
2022 if (!strcmp(line, "--not")) {
2023 flags ^= UNINTERESTING;
2024 continue;
2026 die("not a rev '%s'", line);
2028 if (handle_revision_arg(line, &revs, flags, 1))
2029 die("bad revision '%s'", line);
2032 prepare_revision_walk(&revs);
2033 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2034 traverse_commit_list(&revs, show_commit, show_object);
2036 if (keep_unreachable)
2037 add_objects_in_unpacked_packs(&revs);
2040 static int adjust_perm(const char *path, mode_t mode)
2042 if (chmod(path, mode))
2043 return -1;
2044 return adjust_shared_perm(path);
2047 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2049 int use_internal_rev_list = 0;
2050 int thin = 0;
2051 uint32_t i;
2052 const char **rp_av;
2053 int rp_ac_alloc = 64;
2054 int rp_ac;
2056 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2058 rp_av[0] = "pack-objects";
2059 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2060 rp_ac = 2;
2062 git_config(git_pack_config);
2063 if (!pack_compression_seen && core_compression_seen)
2064 pack_compression_level = core_compression_level;
2066 progress = isatty(2);
2067 for (i = 1; i < argc; i++) {
2068 const char *arg = argv[i];
2070 if (*arg != '-')
2071 break;
2073 if (!strcmp("--non-empty", arg)) {
2074 non_empty = 1;
2075 continue;
2077 if (!strcmp("--local", arg)) {
2078 local = 1;
2079 continue;
2081 if (!strcmp("--incremental", arg)) {
2082 incremental = 1;
2083 continue;
2085 if (!prefixcmp(arg, "--compression=")) {
2086 char *end;
2087 int level = strtoul(arg+14, &end, 0);
2088 if (!arg[14] || *end)
2089 usage(pack_usage);
2090 if (level == -1)
2091 level = Z_DEFAULT_COMPRESSION;
2092 else if (level < 0 || level > Z_BEST_COMPRESSION)
2093 die("bad pack compression level %d", level);
2094 pack_compression_level = level;
2095 continue;
2097 if (!prefixcmp(arg, "--max-pack-size=")) {
2098 char *end;
2099 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2100 if (!arg[16] || *end)
2101 usage(pack_usage);
2102 continue;
2104 if (!prefixcmp(arg, "--window=")) {
2105 char *end;
2106 window = strtoul(arg+9, &end, 0);
2107 if (!arg[9] || *end)
2108 usage(pack_usage);
2109 continue;
2111 if (!prefixcmp(arg, "--window-memory=")) {
2112 if (!git_parse_ulong(arg+16, &window_memory_limit))
2113 usage(pack_usage);
2114 continue;
2116 if (!prefixcmp(arg, "--threads=")) {
2117 char *end;
2118 delta_search_threads = strtoul(arg+10, &end, 0);
2119 if (!arg[10] || *end || delta_search_threads < 1)
2120 usage(pack_usage);
2121 #ifndef THREADED_DELTA_SEARCH
2122 if (delta_search_threads > 1)
2123 warning("no threads support, "
2124 "ignoring %s", arg);
2125 #endif
2126 continue;
2128 if (!prefixcmp(arg, "--depth=")) {
2129 char *end;
2130 depth = strtoul(arg+8, &end, 0);
2131 if (!arg[8] || *end)
2132 usage(pack_usage);
2133 continue;
2135 if (!strcmp("--progress", arg)) {
2136 progress = 1;
2137 continue;
2139 if (!strcmp("--all-progress", arg)) {
2140 progress = 2;
2141 continue;
2143 if (!strcmp("-q", arg)) {
2144 progress = 0;
2145 continue;
2147 if (!strcmp("--no-reuse-delta", arg)) {
2148 no_reuse_delta = 1;
2149 continue;
2151 if (!strcmp("--no-reuse-object", arg)) {
2152 no_reuse_object = no_reuse_delta = 1;
2153 continue;
2155 if (!strcmp("--delta-base-offset", arg)) {
2156 allow_ofs_delta = 1;
2157 continue;
2159 if (!strcmp("--stdout", arg)) {
2160 pack_to_stdout = 1;
2161 continue;
2163 if (!strcmp("--revs", arg)) {
2164 use_internal_rev_list = 1;
2165 continue;
2167 if (!strcmp("--keep-unreachable", arg)) {
2168 keep_unreachable = 1;
2169 continue;
2171 if (!strcmp("--unpacked", arg) ||
2172 !prefixcmp(arg, "--unpacked=") ||
2173 !strcmp("--reflog", arg) ||
2174 !strcmp("--all", arg)) {
2175 use_internal_rev_list = 1;
2176 if (rp_ac >= rp_ac_alloc - 1) {
2177 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2178 rp_av = xrealloc(rp_av,
2179 rp_ac_alloc * sizeof(*rp_av));
2181 rp_av[rp_ac++] = arg;
2182 continue;
2184 if (!strcmp("--thin", arg)) {
2185 use_internal_rev_list = 1;
2186 thin = 1;
2187 rp_av[1] = "--objects-edge";
2188 continue;
2190 if (!prefixcmp(arg, "--index-version=")) {
2191 char *c;
2192 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2193 if (pack_idx_default_version > 2)
2194 die("bad %s", arg);
2195 if (*c == ',')
2196 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2197 if (*c || pack_idx_off32_limit & 0x80000000)
2198 die("bad %s", arg);
2199 continue;
2201 usage(pack_usage);
2204 /* Traditionally "pack-objects [options] base extra" failed;
2205 * we would however want to take refs parameter that would
2206 * have been given to upstream rev-list ourselves, which means
2207 * we somehow want to say what the base name is. So the
2208 * syntax would be:
2210 * pack-objects [options] base <refs...>
2212 * in other words, we would treat the first non-option as the
2213 * base_name and send everything else to the internal revision
2214 * walker.
2217 if (!pack_to_stdout)
2218 base_name = argv[i++];
2220 if (pack_to_stdout != !base_name)
2221 usage(pack_usage);
2223 if (pack_to_stdout && pack_size_limit)
2224 die("--max-pack-size cannot be used to build a pack for transfer.");
2226 if (!pack_to_stdout && thin)
2227 die("--thin cannot be used to build an indexable pack.");
2229 prepare_packed_git();
2231 if (progress)
2232 progress_state = start_progress("Counting objects", 0);
2233 if (!use_internal_rev_list)
2234 read_object_list_from_stdin();
2235 else {
2236 rp_av[rp_ac] = NULL;
2237 get_object_list(rp_ac, rp_av);
2239 stop_progress(&progress_state);
2241 if (non_empty && !nr_result)
2242 return 0;
2243 if (nr_result)
2244 prepare_pack(window, depth);
2245 write_pack_file();
2246 if (progress)
2247 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2248 written, written_delta, reused, reused_delta);
2249 return 0;