git repack: keep commits hidden by a graft
[git/dscho.git] / builtin-pack-objects.c
blob77bff85884e249f894e755e9e42a19a0e2a82f84
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 "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
20 #ifdef THREADED_DELTA_SEARCH
21 #include "thread-utils.h"
22 #include <pthread.h>
23 #endif
25 static const char pack_usage[] = "\
26 git pack-objects [{ -q | --progress | --all-progress }] \n\
27 [--max-pack-size=N] [--local] [--incremental] \n\
28 [--window=N] [--window-memory=N] [--depth=N] \n\
29 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
30 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
31 [--stdout | base-name] [--include-tag] \n\
32 [--keep-unreachable | --unpack-unreachable] \n\
33 [<ref-list | <object-list]";
35 struct object_entry {
36 struct pack_idx_entry idx;
37 unsigned long size; /* uncompressed size */
38 struct packed_git *in_pack; /* already in pack */
39 off_t in_pack_offset;
40 struct object_entry *delta; /* delta base object */
41 struct object_entry *delta_child; /* deltified objects who bases me */
42 struct object_entry *delta_sibling; /* other deltified objects who
43 * uses the same base as me
45 void *delta_data; /* cached delta (uncompressed) */
46 unsigned long delta_size; /* delta data size (uncompressed) */
47 unsigned long z_delta_size; /* delta data size (compressed) */
48 unsigned int hash; /* name hint hash */
49 enum object_type type;
50 enum object_type in_pack_type; /* could be delta */
51 unsigned char in_pack_header_size;
52 unsigned char preferred_base; /* we do not pack this, but is available
53 * to be used as the base object to delta
54 * objects against.
56 unsigned char no_try_delta;
60 * Objects we are going to pack are collected in objects array (dynamically
61 * expanded). nr_objects & nr_alloc controls this array. They are stored
62 * in the order we see -- typically rev-list --objects order that gives us
63 * nice "minimum seek" order.
65 static struct object_entry *objects;
66 static struct pack_idx_entry **written_list;
67 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
69 static int non_empty;
70 static int reuse_delta = 1, reuse_object = 1;
71 static int keep_unreachable, unpack_unreachable, include_tag;
72 static int local;
73 static int incremental;
74 static int ignore_packed_keep;
75 static int allow_ofs_delta;
76 static const char *base_name;
77 static int progress = 1;
78 static int window = 10;
79 static uint32_t pack_size_limit, pack_size_limit_cfg;
80 static int depth = 50;
81 static int delta_search_threads;
82 static int pack_to_stdout;
83 static int num_preferred_base;
84 static struct progress *progress_state;
85 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
86 static int pack_compression_seen;
88 static unsigned long delta_cache_size = 0;
89 static unsigned long max_delta_cache_size = 0;
90 static unsigned long cache_max_small_delta_size = 1000;
92 static unsigned long window_memory_limit = 0;
95 * The object names in objects array are hashed with this hashtable,
96 * to help looking up the entry by object name.
97 * This hashtable is built after all the objects are seen.
99 static int *object_ix;
100 static int object_ix_hashsz;
103 * stats
105 static uint32_t written, written_delta;
106 static uint32_t reused, reused_delta;
109 static void *get_delta(struct object_entry *entry)
111 unsigned long size, base_size, delta_size;
112 void *buf, *base_buf, *delta_buf;
113 enum object_type type;
115 buf = read_sha1_file(entry->idx.sha1, &type, &size);
116 if (!buf)
117 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
118 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
119 if (!base_buf)
120 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
121 delta_buf = diff_delta(base_buf, base_size,
122 buf, size, &delta_size, 0);
123 if (!delta_buf || delta_size != entry->delta_size)
124 die("delta size changed");
125 free(buf);
126 free(base_buf);
127 return delta_buf;
130 static unsigned long do_compress(void **pptr, unsigned long size)
132 z_stream stream;
133 void *in, *out;
134 unsigned long maxsize;
136 memset(&stream, 0, sizeof(stream));
137 deflateInit(&stream, pack_compression_level);
138 maxsize = deflateBound(&stream, size);
140 in = *pptr;
141 out = xmalloc(maxsize);
142 *pptr = out;
144 stream.next_in = in;
145 stream.avail_in = size;
146 stream.next_out = out;
147 stream.avail_out = maxsize;
148 while (deflate(&stream, Z_FINISH) == Z_OK)
149 ; /* nothing */
150 deflateEnd(&stream);
152 free(in);
153 return stream.total_out;
157 * The per-object header is a pretty dense thing, which is
158 * - first byte: low four bits are "size", then three bits of "type",
159 * and the high bit is "size continues".
160 * - each byte afterwards: low seven bits are size continuation,
161 * with the high bit being "size continues"
163 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
165 int n = 1;
166 unsigned char c;
168 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
169 die("bad type %d", type);
171 c = (type << 4) | (size & 15);
172 size >>= 4;
173 while (size) {
174 *hdr++ = c | 0x80;
175 c = size & 0x7f;
176 size >>= 7;
177 n++;
179 *hdr = c;
180 return n;
184 * we are going to reuse the existing object data as is. make
185 * sure it is not corrupt.
187 static int check_pack_inflate(struct packed_git *p,
188 struct pack_window **w_curs,
189 off_t offset,
190 off_t len,
191 unsigned long expect)
193 z_stream stream;
194 unsigned char fakebuf[4096], *in;
195 int st;
197 memset(&stream, 0, sizeof(stream));
198 git_inflate_init(&stream);
199 do {
200 in = use_pack(p, w_curs, offset, &stream.avail_in);
201 stream.next_in = in;
202 stream.next_out = fakebuf;
203 stream.avail_out = sizeof(fakebuf);
204 st = git_inflate(&stream, Z_FINISH);
205 offset += stream.next_in - in;
206 } while (st == Z_OK || st == Z_BUF_ERROR);
207 git_inflate_end(&stream);
208 return (st == Z_STREAM_END &&
209 stream.total_out == expect &&
210 stream.total_in == len) ? 0 : -1;
213 static void copy_pack_data(struct sha1file *f,
214 struct packed_git *p,
215 struct pack_window **w_curs,
216 off_t offset,
217 off_t len)
219 unsigned char *in;
220 unsigned int avail;
222 while (len) {
223 in = use_pack(p, w_curs, offset, &avail);
224 if (avail > len)
225 avail = (unsigned int)len;
226 sha1write(f, in, avail);
227 offset += avail;
228 len -= avail;
232 static unsigned long write_object(struct sha1file *f,
233 struct object_entry *entry,
234 off_t write_offset)
236 unsigned long size, limit, datalen;
237 void *buf;
238 unsigned char header[10], dheader[10];
239 unsigned hdrlen;
240 enum object_type type;
241 int usable_delta, to_reuse;
243 if (!pack_to_stdout)
244 crc32_begin(f);
246 type = entry->type;
248 /* write limit if limited packsize and not first object */
249 if (!pack_size_limit || !nr_written)
250 limit = 0;
251 else if (pack_size_limit <= write_offset)
253 * the earlier object did not fit the limit; avoid
254 * mistaking this with unlimited (i.e. limit = 0).
256 limit = 1;
257 else
258 limit = pack_size_limit - write_offset;
260 if (!entry->delta)
261 usable_delta = 0; /* no delta */
262 else if (!pack_size_limit)
263 usable_delta = 1; /* unlimited packfile */
264 else if (entry->delta->idx.offset == (off_t)-1)
265 usable_delta = 0; /* base was written to another pack */
266 else if (entry->delta->idx.offset)
267 usable_delta = 1; /* base already exists in this pack */
268 else
269 usable_delta = 0; /* base could end up in another pack */
271 if (!reuse_object)
272 to_reuse = 0; /* explicit */
273 else if (!entry->in_pack)
274 to_reuse = 0; /* can't reuse what we don't have */
275 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
276 /* check_object() decided it for us ... */
277 to_reuse = usable_delta;
278 /* ... but pack split may override that */
279 else if (type != entry->in_pack_type)
280 to_reuse = 0; /* pack has delta which is unusable */
281 else if (entry->delta)
282 to_reuse = 0; /* we want to pack afresh */
283 else
284 to_reuse = 1; /* we have it in-pack undeltified,
285 * and we do not need to deltify it.
288 if (!to_reuse) {
289 no_reuse:
290 if (!usable_delta) {
291 buf = read_sha1_file(entry->idx.sha1, &type, &size);
292 if (!buf)
293 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
295 * make sure no cached delta data remains from a
296 * previous attempt before a pack split occurred.
298 free(entry->delta_data);
299 entry->delta_data = NULL;
300 entry->z_delta_size = 0;
301 } else if (entry->delta_data) {
302 size = entry->delta_size;
303 buf = entry->delta_data;
304 entry->delta_data = NULL;
305 type = (allow_ofs_delta && entry->delta->idx.offset) ?
306 OBJ_OFS_DELTA : OBJ_REF_DELTA;
307 } else {
308 buf = get_delta(entry);
309 size = entry->delta_size;
310 type = (allow_ofs_delta && entry->delta->idx.offset) ?
311 OBJ_OFS_DELTA : OBJ_REF_DELTA;
314 if (entry->z_delta_size)
315 datalen = entry->z_delta_size;
316 else
317 datalen = do_compress(&buf, size);
320 * The object header is a byte of 'type' followed by zero or
321 * more bytes of length.
323 hdrlen = encode_header(type, size, header);
325 if (type == OBJ_OFS_DELTA) {
327 * Deltas with relative base contain an additional
328 * encoding of the relative offset for the delta
329 * base from this object's position in the pack.
331 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
332 unsigned pos = sizeof(dheader) - 1;
333 dheader[pos] = ofs & 127;
334 while (ofs >>= 7)
335 dheader[--pos] = 128 | (--ofs & 127);
336 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
337 free(buf);
338 return 0;
340 sha1write(f, header, hdrlen);
341 sha1write(f, dheader + pos, sizeof(dheader) - pos);
342 hdrlen += sizeof(dheader) - pos;
343 } else if (type == OBJ_REF_DELTA) {
345 * Deltas with a base reference contain
346 * an additional 20 bytes for the base sha1.
348 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
349 free(buf);
350 return 0;
352 sha1write(f, header, hdrlen);
353 sha1write(f, entry->delta->idx.sha1, 20);
354 hdrlen += 20;
355 } else {
356 if (limit && hdrlen + datalen + 20 >= limit) {
357 free(buf);
358 return 0;
360 sha1write(f, header, hdrlen);
362 sha1write(f, buf, datalen);
363 free(buf);
365 else {
366 struct packed_git *p = entry->in_pack;
367 struct pack_window *w_curs = NULL;
368 struct revindex_entry *revidx;
369 off_t offset;
371 if (entry->delta)
372 type = (allow_ofs_delta && entry->delta->idx.offset) ?
373 OBJ_OFS_DELTA : OBJ_REF_DELTA;
374 hdrlen = encode_header(type, entry->size, header);
376 offset = entry->in_pack_offset;
377 revidx = find_pack_revindex(p, offset);
378 datalen = revidx[1].offset - offset;
379 if (!pack_to_stdout && p->index_version > 1 &&
380 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
381 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
382 unuse_pack(&w_curs);
383 goto no_reuse;
386 offset += entry->in_pack_header_size;
387 datalen -= entry->in_pack_header_size;
388 if (!pack_to_stdout && p->index_version == 1 &&
389 check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
390 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
391 unuse_pack(&w_curs);
392 goto no_reuse;
395 if (type == OBJ_OFS_DELTA) {
396 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
397 unsigned pos = sizeof(dheader) - 1;
398 dheader[pos] = ofs & 127;
399 while (ofs >>= 7)
400 dheader[--pos] = 128 | (--ofs & 127);
401 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
402 unuse_pack(&w_curs);
403 return 0;
405 sha1write(f, header, hdrlen);
406 sha1write(f, dheader + pos, sizeof(dheader) - pos);
407 hdrlen += sizeof(dheader) - pos;
408 reused_delta++;
409 } else if (type == OBJ_REF_DELTA) {
410 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
411 unuse_pack(&w_curs);
412 return 0;
414 sha1write(f, header, hdrlen);
415 sha1write(f, entry->delta->idx.sha1, 20);
416 hdrlen += 20;
417 reused_delta++;
418 } else {
419 if (limit && hdrlen + datalen + 20 >= limit) {
420 unuse_pack(&w_curs);
421 return 0;
423 sha1write(f, header, hdrlen);
425 copy_pack_data(f, p, &w_curs, offset, datalen);
426 unuse_pack(&w_curs);
427 reused++;
429 if (usable_delta)
430 written_delta++;
431 written++;
432 if (!pack_to_stdout)
433 entry->idx.crc32 = crc32_end(f);
434 return hdrlen + datalen;
437 static int write_one(struct sha1file *f,
438 struct object_entry *e,
439 off_t *offset)
441 unsigned long size;
443 /* offset is non zero if object is written already. */
444 if (e->idx.offset || e->preferred_base)
445 return 1;
447 /* if we are deltified, write out base object first. */
448 if (e->delta && !write_one(f, e->delta, offset))
449 return 0;
451 e->idx.offset = *offset;
452 size = write_object(f, e, *offset);
453 if (!size) {
454 e->idx.offset = 0;
455 return 0;
457 written_list[nr_written++] = &e->idx;
459 /* make sure off_t is sufficiently large not to wrap */
460 if (*offset > *offset + size)
461 die("pack too large for current definition of off_t");
462 *offset += size;
463 return 1;
466 /* forward declaration for write_pack_file */
467 static int adjust_perm(const char *path, mode_t mode);
469 static void write_pack_file(void)
471 uint32_t i = 0, j;
472 struct sha1file *f;
473 off_t offset;
474 struct pack_header hdr;
475 uint32_t nr_remaining = nr_result;
476 time_t last_mtime = 0;
478 if (progress > pack_to_stdout)
479 progress_state = start_progress("Writing objects", nr_result);
480 written_list = xmalloc(nr_objects * sizeof(*written_list));
482 do {
483 unsigned char sha1[20];
484 char *pack_tmp_name = NULL;
486 if (pack_to_stdout) {
487 f = sha1fd_throughput(1, "<stdout>", progress_state);
488 } else {
489 char tmpname[PATH_MAX];
490 int fd;
491 fd = odb_mkstemp(tmpname, sizeof(tmpname),
492 "pack/tmp_pack_XXXXXX");
493 pack_tmp_name = xstrdup(tmpname);
494 f = sha1fd(fd, pack_tmp_name);
497 hdr.hdr_signature = htonl(PACK_SIGNATURE);
498 hdr.hdr_version = htonl(PACK_VERSION);
499 hdr.hdr_entries = htonl(nr_remaining);
500 sha1write(f, &hdr, sizeof(hdr));
501 offset = sizeof(hdr);
502 nr_written = 0;
503 for (; i < nr_objects; i++) {
504 if (!write_one(f, objects + i, &offset))
505 break;
506 display_progress(progress_state, written);
510 * Did we write the wrong # entries in the header?
511 * If so, rewrite it like in fast-import
513 if (pack_to_stdout) {
514 sha1close(f, sha1, CSUM_CLOSE);
515 } else if (nr_written == nr_remaining) {
516 sha1close(f, sha1, CSUM_FSYNC);
517 } else {
518 int fd = sha1close(f, sha1, 0);
519 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
520 nr_written, sha1, offset);
521 close(fd);
524 if (!pack_to_stdout) {
525 mode_t mode = umask(0);
526 struct stat st;
527 char *idx_tmp_name, tmpname[PATH_MAX];
529 umask(mode);
530 mode = 0444 & ~mode;
532 idx_tmp_name = write_idx_file(NULL, written_list,
533 nr_written, sha1);
535 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
536 base_name, sha1_to_hex(sha1));
537 free_pack_by_name(tmpname);
538 if (adjust_perm(pack_tmp_name, mode))
539 die_errno("unable to make temporary pack file readable");
540 if (rename(pack_tmp_name, tmpname))
541 die_errno("unable to rename temporary pack file");
544 * Packs are runtime accessed in their mtime
545 * order since newer packs are more likely to contain
546 * younger objects. So if we are creating multiple
547 * packs then we should modify the mtime of later ones
548 * to preserve this property.
550 if (stat(tmpname, &st) < 0) {
551 warning("failed to stat %s: %s",
552 tmpname, strerror(errno));
553 } else if (!last_mtime) {
554 last_mtime = st.st_mtime;
555 } else {
556 struct utimbuf utb;
557 utb.actime = st.st_atime;
558 utb.modtime = --last_mtime;
559 if (utime(tmpname, &utb) < 0)
560 warning("failed utime() on %s: %s",
561 tmpname, strerror(errno));
564 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
565 base_name, sha1_to_hex(sha1));
566 if (adjust_perm(idx_tmp_name, mode))
567 die_errno("unable to make temporary index file readable");
568 if (rename(idx_tmp_name, tmpname))
569 die_errno("unable to rename temporary index file");
571 free(idx_tmp_name);
572 free(pack_tmp_name);
573 puts(sha1_to_hex(sha1));
576 /* mark written objects as written to previous pack */
577 for (j = 0; j < nr_written; j++) {
578 written_list[j]->offset = (off_t)-1;
580 nr_remaining -= nr_written;
581 } while (nr_remaining && i < nr_objects);
583 free(written_list);
584 stop_progress(&progress_state);
585 if (written != nr_result)
586 die("wrote %"PRIu32" objects while expecting %"PRIu32,
587 written, nr_result);
589 * We have scanned through [0 ... i). Since we have written
590 * the correct number of objects, the remaining [i ... nr_objects)
591 * items must be either already written (due to out-of-order delta base)
592 * or a preferred base. Count those which are neither and complain if any.
594 for (j = 0; i < nr_objects; i++) {
595 struct object_entry *e = objects + i;
596 j += !e->idx.offset && !e->preferred_base;
598 if (j)
599 die("wrote %"PRIu32" objects as expected but %"PRIu32
600 " unwritten", written, j);
603 static int locate_object_entry_hash(const unsigned char *sha1)
605 int i;
606 unsigned int ui;
607 memcpy(&ui, sha1, sizeof(unsigned int));
608 i = ui % object_ix_hashsz;
609 while (0 < object_ix[i]) {
610 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
611 return i;
612 if (++i == object_ix_hashsz)
613 i = 0;
615 return -1 - i;
618 static struct object_entry *locate_object_entry(const unsigned char *sha1)
620 int i;
622 if (!object_ix_hashsz)
623 return NULL;
625 i = locate_object_entry_hash(sha1);
626 if (0 <= i)
627 return &objects[object_ix[i]-1];
628 return NULL;
631 static void rehash_objects(void)
633 uint32_t i;
634 struct object_entry *oe;
636 object_ix_hashsz = nr_objects * 3;
637 if (object_ix_hashsz < 1024)
638 object_ix_hashsz = 1024;
639 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
640 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
641 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
642 int ix = locate_object_entry_hash(oe->idx.sha1);
643 if (0 <= ix)
644 continue;
645 ix = -1 - ix;
646 object_ix[ix] = i + 1;
650 static unsigned name_hash(const char *name)
652 unsigned c, hash = 0;
654 if (!name)
655 return 0;
658 * This effectively just creates a sortable number from the
659 * last sixteen non-whitespace characters. Last characters
660 * count "most", so things that end in ".c" sort together.
662 while ((c = *name++) != 0) {
663 if (isspace(c))
664 continue;
665 hash = (hash >> 2) + (c << 24);
667 return hash;
670 static void setup_delta_attr_check(struct git_attr_check *check)
672 static struct git_attr *attr_delta;
674 if (!attr_delta)
675 attr_delta = git_attr("delta", 5);
677 check[0].attr = attr_delta;
680 static int no_try_delta(const char *path)
682 struct git_attr_check check[1];
684 setup_delta_attr_check(check);
685 if (git_checkattr(path, ARRAY_SIZE(check), check))
686 return 0;
687 if (ATTR_FALSE(check->value))
688 return 1;
689 return 0;
692 static int add_object_entry(const unsigned char *sha1, enum object_type type,
693 const char *name, int exclude)
695 struct object_entry *entry;
696 struct packed_git *p, *found_pack = NULL;
697 off_t found_offset = 0;
698 int ix;
699 unsigned hash = name_hash(name);
701 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
702 if (ix >= 0) {
703 if (exclude) {
704 entry = objects + object_ix[ix] - 1;
705 if (!entry->preferred_base)
706 nr_result--;
707 entry->preferred_base = 1;
709 return 0;
712 if (!exclude && local && has_loose_object_nonlocal(sha1))
713 return 0;
715 for (p = packed_git; p; p = p->next) {
716 off_t offset = find_pack_entry_one(sha1, p);
717 if (offset) {
718 if (!found_pack) {
719 found_offset = offset;
720 found_pack = p;
722 if (exclude)
723 break;
724 if (incremental)
725 return 0;
726 if (local && !p->pack_local)
727 return 0;
728 if (ignore_packed_keep && p->pack_local && p->pack_keep)
729 return 0;
733 if (nr_objects >= nr_alloc) {
734 nr_alloc = (nr_alloc + 1024) * 3 / 2;
735 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
738 entry = objects + nr_objects++;
739 memset(entry, 0, sizeof(*entry));
740 hashcpy(entry->idx.sha1, sha1);
741 entry->hash = hash;
742 if (type)
743 entry->type = type;
744 if (exclude)
745 entry->preferred_base = 1;
746 else
747 nr_result++;
748 if (found_pack) {
749 entry->in_pack = found_pack;
750 entry->in_pack_offset = found_offset;
753 if (object_ix_hashsz * 3 <= nr_objects * 4)
754 rehash_objects();
755 else
756 object_ix[-1 - ix] = nr_objects;
758 display_progress(progress_state, nr_objects);
760 if (name && no_try_delta(name))
761 entry->no_try_delta = 1;
763 return 1;
766 struct pbase_tree_cache {
767 unsigned char sha1[20];
768 int ref;
769 int temporary;
770 void *tree_data;
771 unsigned long tree_size;
774 static struct pbase_tree_cache *(pbase_tree_cache[256]);
775 static int pbase_tree_cache_ix(const unsigned char *sha1)
777 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
779 static int pbase_tree_cache_ix_incr(int ix)
781 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
784 static struct pbase_tree {
785 struct pbase_tree *next;
786 /* This is a phony "cache" entry; we are not
787 * going to evict it nor find it through _get()
788 * mechanism -- this is for the toplevel node that
789 * would almost always change with any commit.
791 struct pbase_tree_cache pcache;
792 } *pbase_tree;
794 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
796 struct pbase_tree_cache *ent, *nent;
797 void *data;
798 unsigned long size;
799 enum object_type type;
800 int neigh;
801 int my_ix = pbase_tree_cache_ix(sha1);
802 int available_ix = -1;
804 /* pbase-tree-cache acts as a limited hashtable.
805 * your object will be found at your index or within a few
806 * slots after that slot if it is cached.
808 for (neigh = 0; neigh < 8; neigh++) {
809 ent = pbase_tree_cache[my_ix];
810 if (ent && !hashcmp(ent->sha1, sha1)) {
811 ent->ref++;
812 return ent;
814 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
815 ((0 <= available_ix) &&
816 (!ent && pbase_tree_cache[available_ix])))
817 available_ix = my_ix;
818 if (!ent)
819 break;
820 my_ix = pbase_tree_cache_ix_incr(my_ix);
823 /* Did not find one. Either we got a bogus request or
824 * we need to read and perhaps cache.
826 data = read_sha1_file(sha1, &type, &size);
827 if (!data)
828 return NULL;
829 if (type != OBJ_TREE) {
830 free(data);
831 return NULL;
834 /* We need to either cache or return a throwaway copy */
836 if (available_ix < 0)
837 ent = NULL;
838 else {
839 ent = pbase_tree_cache[available_ix];
840 my_ix = available_ix;
843 if (!ent) {
844 nent = xmalloc(sizeof(*nent));
845 nent->temporary = (available_ix < 0);
847 else {
848 /* evict and reuse */
849 free(ent->tree_data);
850 nent = ent;
852 hashcpy(nent->sha1, sha1);
853 nent->tree_data = data;
854 nent->tree_size = size;
855 nent->ref = 1;
856 if (!nent->temporary)
857 pbase_tree_cache[my_ix] = nent;
858 return nent;
861 static void pbase_tree_put(struct pbase_tree_cache *cache)
863 if (!cache->temporary) {
864 cache->ref--;
865 return;
867 free(cache->tree_data);
868 free(cache);
871 static int name_cmp_len(const char *name)
873 int i;
874 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
876 return i;
879 static void add_pbase_object(struct tree_desc *tree,
880 const char *name,
881 int cmplen,
882 const char *fullname)
884 struct name_entry entry;
885 int cmp;
887 while (tree_entry(tree,&entry)) {
888 if (S_ISGITLINK(entry.mode))
889 continue;
890 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
891 memcmp(name, entry.path, cmplen);
892 if (cmp > 0)
893 continue;
894 if (cmp < 0)
895 return;
896 if (name[cmplen] != '/') {
897 add_object_entry(entry.sha1,
898 object_type(entry.mode),
899 fullname, 1);
900 return;
902 if (S_ISDIR(entry.mode)) {
903 struct tree_desc sub;
904 struct pbase_tree_cache *tree;
905 const char *down = name+cmplen+1;
906 int downlen = name_cmp_len(down);
908 tree = pbase_tree_get(entry.sha1);
909 if (!tree)
910 return;
911 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
913 add_pbase_object(&sub, down, downlen, fullname);
914 pbase_tree_put(tree);
919 static unsigned *done_pbase_paths;
920 static int done_pbase_paths_num;
921 static int done_pbase_paths_alloc;
922 static int done_pbase_path_pos(unsigned hash)
924 int lo = 0;
925 int hi = done_pbase_paths_num;
926 while (lo < hi) {
927 int mi = (hi + lo) / 2;
928 if (done_pbase_paths[mi] == hash)
929 return mi;
930 if (done_pbase_paths[mi] < hash)
931 hi = mi;
932 else
933 lo = mi + 1;
935 return -lo-1;
938 static int check_pbase_path(unsigned hash)
940 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
941 if (0 <= pos)
942 return 1;
943 pos = -pos - 1;
944 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
945 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
946 done_pbase_paths = xrealloc(done_pbase_paths,
947 done_pbase_paths_alloc *
948 sizeof(unsigned));
950 done_pbase_paths_num++;
951 if (pos < done_pbase_paths_num)
952 memmove(done_pbase_paths + pos + 1,
953 done_pbase_paths + pos,
954 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
955 done_pbase_paths[pos] = hash;
956 return 0;
959 static void add_preferred_base_object(const char *name)
961 struct pbase_tree *it;
962 int cmplen;
963 unsigned hash = name_hash(name);
965 if (!num_preferred_base || check_pbase_path(hash))
966 return;
968 cmplen = name_cmp_len(name);
969 for (it = pbase_tree; it; it = it->next) {
970 if (cmplen == 0) {
971 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
973 else {
974 struct tree_desc tree;
975 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
976 add_pbase_object(&tree, name, cmplen, name);
981 static void add_preferred_base(unsigned char *sha1)
983 struct pbase_tree *it;
984 void *data;
985 unsigned long size;
986 unsigned char tree_sha1[20];
988 if (window <= num_preferred_base++)
989 return;
991 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
992 if (!data)
993 return;
995 for (it = pbase_tree; it; it = it->next) {
996 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
997 free(data);
998 return;
1002 it = xcalloc(1, sizeof(*it));
1003 it->next = pbase_tree;
1004 pbase_tree = it;
1006 hashcpy(it->pcache.sha1, tree_sha1);
1007 it->pcache.tree_data = data;
1008 it->pcache.tree_size = size;
1011 static void check_object(struct object_entry *entry)
1013 if (entry->in_pack) {
1014 struct packed_git *p = entry->in_pack;
1015 struct pack_window *w_curs = NULL;
1016 const unsigned char *base_ref = NULL;
1017 struct object_entry *base_entry;
1018 unsigned long used, used_0;
1019 unsigned int avail;
1020 off_t ofs;
1021 unsigned char *buf, c;
1023 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1026 * We want in_pack_type even if we do not reuse delta
1027 * since non-delta representations could still be reused.
1029 used = unpack_object_header_buffer(buf, avail,
1030 &entry->in_pack_type,
1031 &entry->size);
1032 if (used == 0)
1033 goto give_up;
1036 * Determine if this is a delta and if so whether we can
1037 * reuse it or not. Otherwise let's find out as cheaply as
1038 * possible what the actual type and size for this object is.
1040 switch (entry->in_pack_type) {
1041 default:
1042 /* Not a delta hence we've already got all we need. */
1043 entry->type = entry->in_pack_type;
1044 entry->in_pack_header_size = used;
1045 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1046 goto give_up;
1047 unuse_pack(&w_curs);
1048 return;
1049 case OBJ_REF_DELTA:
1050 if (reuse_delta && !entry->preferred_base)
1051 base_ref = use_pack(p, &w_curs,
1052 entry->in_pack_offset + used, NULL);
1053 entry->in_pack_header_size = used + 20;
1054 break;
1055 case OBJ_OFS_DELTA:
1056 buf = use_pack(p, &w_curs,
1057 entry->in_pack_offset + used, NULL);
1058 used_0 = 0;
1059 c = buf[used_0++];
1060 ofs = c & 127;
1061 while (c & 128) {
1062 ofs += 1;
1063 if (!ofs || MSB(ofs, 7)) {
1064 error("delta base offset overflow in pack for %s",
1065 sha1_to_hex(entry->idx.sha1));
1066 goto give_up;
1068 c = buf[used_0++];
1069 ofs = (ofs << 7) + (c & 127);
1071 ofs = entry->in_pack_offset - ofs;
1072 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1073 error("delta base offset out of bound for %s",
1074 sha1_to_hex(entry->idx.sha1));
1075 goto give_up;
1077 if (reuse_delta && !entry->preferred_base) {
1078 struct revindex_entry *revidx;
1079 revidx = find_pack_revindex(p, ofs);
1080 if (!revidx)
1081 goto give_up;
1082 base_ref = nth_packed_object_sha1(p, revidx->nr);
1084 entry->in_pack_header_size = used + used_0;
1085 break;
1088 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1090 * If base_ref was set above that means we wish to
1091 * reuse delta data, and we even found that base
1092 * in the list of objects we want to pack. Goodie!
1094 * Depth value does not matter - find_deltas() will
1095 * never consider reused delta as the base object to
1096 * deltify other objects against, in order to avoid
1097 * circular deltas.
1099 entry->type = entry->in_pack_type;
1100 entry->delta = base_entry;
1101 entry->delta_size = entry->size;
1102 entry->delta_sibling = base_entry->delta_child;
1103 base_entry->delta_child = entry;
1104 unuse_pack(&w_curs);
1105 return;
1108 if (entry->type) {
1110 * This must be a delta and we already know what the
1111 * final object type is. Let's extract the actual
1112 * object size from the delta header.
1114 entry->size = get_size_from_delta(p, &w_curs,
1115 entry->in_pack_offset + entry->in_pack_header_size);
1116 if (entry->size == 0)
1117 goto give_up;
1118 unuse_pack(&w_curs);
1119 return;
1123 * No choice but to fall back to the recursive delta walk
1124 * with sha1_object_info() to find about the object type
1125 * at this point...
1127 give_up:
1128 unuse_pack(&w_curs);
1131 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1133 * The error condition is checked in prepare_pack(). This is
1134 * to permit a missing preferred base object to be ignored
1135 * as a preferred base. Doing so can result in a larger
1136 * pack file, but the transfer will still take place.
1140 static int pack_offset_sort(const void *_a, const void *_b)
1142 const struct object_entry *a = *(struct object_entry **)_a;
1143 const struct object_entry *b = *(struct object_entry **)_b;
1145 /* avoid filesystem trashing with loose objects */
1146 if (!a->in_pack && !b->in_pack)
1147 return hashcmp(a->idx.sha1, b->idx.sha1);
1149 if (a->in_pack < b->in_pack)
1150 return -1;
1151 if (a->in_pack > b->in_pack)
1152 return 1;
1153 return a->in_pack_offset < b->in_pack_offset ? -1 :
1154 (a->in_pack_offset > b->in_pack_offset);
1157 static void get_object_details(void)
1159 uint32_t i;
1160 struct object_entry **sorted_by_offset;
1162 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1163 for (i = 0; i < nr_objects; i++)
1164 sorted_by_offset[i] = objects + i;
1165 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1167 for (i = 0; i < nr_objects; i++)
1168 check_object(sorted_by_offset[i]);
1170 free(sorted_by_offset);
1174 * We search for deltas in a list sorted by type, by filename hash, and then
1175 * by size, so that we see progressively smaller and smaller files.
1176 * That's because we prefer deltas to be from the bigger file
1177 * to the smaller -- deletes are potentially cheaper, but perhaps
1178 * more importantly, the bigger file is likely the more recent
1179 * one. The deepest deltas are therefore the oldest objects which are
1180 * less susceptible to be accessed often.
1182 static int type_size_sort(const void *_a, const void *_b)
1184 const struct object_entry *a = *(struct object_entry **)_a;
1185 const struct object_entry *b = *(struct object_entry **)_b;
1187 if (a->type > b->type)
1188 return -1;
1189 if (a->type < b->type)
1190 return 1;
1191 if (a->hash > b->hash)
1192 return -1;
1193 if (a->hash < b->hash)
1194 return 1;
1195 if (a->preferred_base > b->preferred_base)
1196 return -1;
1197 if (a->preferred_base < b->preferred_base)
1198 return 1;
1199 if (a->size > b->size)
1200 return -1;
1201 if (a->size < b->size)
1202 return 1;
1203 return a < b ? -1 : (a > b); /* newest first */
1206 struct unpacked {
1207 struct object_entry *entry;
1208 void *data;
1209 struct delta_index *index;
1210 unsigned depth;
1213 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1214 unsigned long delta_size)
1216 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1217 return 0;
1219 if (delta_size < cache_max_small_delta_size)
1220 return 1;
1222 /* cache delta, if objects are large enough compared to delta size */
1223 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1224 return 1;
1226 return 0;
1229 #ifdef THREADED_DELTA_SEARCH
1231 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1232 #define read_lock() pthread_mutex_lock(&read_mutex)
1233 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1235 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1236 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1237 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1239 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1240 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1241 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1243 #else
1245 #define read_lock() (void)0
1246 #define read_unlock() (void)0
1247 #define cache_lock() (void)0
1248 #define cache_unlock() (void)0
1249 #define progress_lock() (void)0
1250 #define progress_unlock() (void)0
1252 #endif
1254 static int try_delta(struct unpacked *trg, struct unpacked *src,
1255 unsigned max_depth, unsigned long *mem_usage)
1257 struct object_entry *trg_entry = trg->entry;
1258 struct object_entry *src_entry = src->entry;
1259 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1260 unsigned ref_depth;
1261 enum object_type type;
1262 void *delta_buf;
1264 /* Don't bother doing diffs between different types */
1265 if (trg_entry->type != src_entry->type)
1266 return -1;
1269 * We do not bother to try a delta that we discarded
1270 * on an earlier try, but only when reusing delta data.
1272 if (reuse_delta && trg_entry->in_pack &&
1273 trg_entry->in_pack == src_entry->in_pack &&
1274 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1275 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1276 return 0;
1278 /* Let's not bust the allowed depth. */
1279 if (src->depth >= max_depth)
1280 return 0;
1282 /* Now some size filtering heuristics. */
1283 trg_size = trg_entry->size;
1284 if (!trg_entry->delta) {
1285 max_size = trg_size/2 - 20;
1286 ref_depth = 1;
1287 } else {
1288 max_size = trg_entry->delta_size;
1289 ref_depth = trg->depth;
1291 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1292 (max_depth - ref_depth + 1);
1293 if (max_size == 0)
1294 return 0;
1295 src_size = src_entry->size;
1296 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1297 if (sizediff >= max_size)
1298 return 0;
1299 if (trg_size < src_size / 32)
1300 return 0;
1302 /* Load data if not already done */
1303 if (!trg->data) {
1304 read_lock();
1305 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1306 read_unlock();
1307 if (!trg->data)
1308 die("object %s cannot be read",
1309 sha1_to_hex(trg_entry->idx.sha1));
1310 if (sz != trg_size)
1311 die("object %s inconsistent object length (%lu vs %lu)",
1312 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1313 *mem_usage += sz;
1315 if (!src->data) {
1316 read_lock();
1317 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1318 read_unlock();
1319 if (!src->data)
1320 die("object %s cannot be read",
1321 sha1_to_hex(src_entry->idx.sha1));
1322 if (sz != src_size)
1323 die("object %s inconsistent object length (%lu vs %lu)",
1324 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1325 *mem_usage += sz;
1327 if (!src->index) {
1328 src->index = create_delta_index(src->data, src_size);
1329 if (!src->index) {
1330 static int warned = 0;
1331 if (!warned++)
1332 warning("suboptimal pack - out of memory");
1333 return 0;
1335 *mem_usage += sizeof_delta_index(src->index);
1338 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1339 if (!delta_buf)
1340 return 0;
1342 if (trg_entry->delta) {
1343 /* Prefer only shallower same-sized deltas. */
1344 if (delta_size == trg_entry->delta_size &&
1345 src->depth + 1 >= trg->depth) {
1346 free(delta_buf);
1347 return 0;
1352 * Handle memory allocation outside of the cache
1353 * accounting lock. Compiler will optimize the strangeness
1354 * away when THREADED_DELTA_SEARCH is not defined.
1356 free(trg_entry->delta_data);
1357 cache_lock();
1358 if (trg_entry->delta_data) {
1359 delta_cache_size -= trg_entry->delta_size;
1360 trg_entry->delta_data = NULL;
1362 if (delta_cacheable(src_size, trg_size, delta_size)) {
1363 delta_cache_size += delta_size;
1364 cache_unlock();
1365 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1366 } else {
1367 cache_unlock();
1368 free(delta_buf);
1371 trg_entry->delta = src_entry;
1372 trg_entry->delta_size = delta_size;
1373 trg->depth = src->depth + 1;
1375 return 1;
1378 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1380 struct object_entry *child = me->delta_child;
1381 unsigned int m = n;
1382 while (child) {
1383 unsigned int c = check_delta_limit(child, n + 1);
1384 if (m < c)
1385 m = c;
1386 child = child->delta_sibling;
1388 return m;
1391 static unsigned long free_unpacked(struct unpacked *n)
1393 unsigned long freed_mem = sizeof_delta_index(n->index);
1394 free_delta_index(n->index);
1395 n->index = NULL;
1396 if (n->data) {
1397 freed_mem += n->entry->size;
1398 free(n->data);
1399 n->data = NULL;
1401 n->entry = NULL;
1402 n->depth = 0;
1403 return freed_mem;
1406 static void find_deltas(struct object_entry **list, unsigned *list_size,
1407 int window, int depth, unsigned *processed)
1409 uint32_t i, idx = 0, count = 0;
1410 struct unpacked *array;
1411 unsigned long mem_usage = 0;
1413 array = xcalloc(window, sizeof(struct unpacked));
1415 for (;;) {
1416 struct object_entry *entry;
1417 struct unpacked *n = array + idx;
1418 int j, max_depth, best_base = -1;
1420 progress_lock();
1421 if (!*list_size) {
1422 progress_unlock();
1423 break;
1425 entry = *list++;
1426 (*list_size)--;
1427 if (!entry->preferred_base) {
1428 (*processed)++;
1429 display_progress(progress_state, *processed);
1431 progress_unlock();
1433 mem_usage -= free_unpacked(n);
1434 n->entry = entry;
1436 while (window_memory_limit &&
1437 mem_usage > window_memory_limit &&
1438 count > 1) {
1439 uint32_t tail = (idx + window - count) % window;
1440 mem_usage -= free_unpacked(array + tail);
1441 count--;
1444 /* We do not compute delta to *create* objects we are not
1445 * going to pack.
1447 if (entry->preferred_base)
1448 goto next;
1451 * If the current object is at pack edge, take the depth the
1452 * objects that depend on the current object into account
1453 * otherwise they would become too deep.
1455 max_depth = depth;
1456 if (entry->delta_child) {
1457 max_depth -= check_delta_limit(entry, 0);
1458 if (max_depth <= 0)
1459 goto next;
1462 j = window;
1463 while (--j > 0) {
1464 int ret;
1465 uint32_t other_idx = idx + j;
1466 struct unpacked *m;
1467 if (other_idx >= window)
1468 other_idx -= window;
1469 m = array + other_idx;
1470 if (!m->entry)
1471 break;
1472 ret = try_delta(n, m, max_depth, &mem_usage);
1473 if (ret < 0)
1474 break;
1475 else if (ret > 0)
1476 best_base = other_idx;
1480 * If we decided to cache the delta data, then it is best
1481 * to compress it right away. First because we have to do
1482 * it anyway, and doing it here while we're threaded will
1483 * save a lot of time in the non threaded write phase,
1484 * as well as allow for caching more deltas within
1485 * the same cache size limit.
1486 * ...
1487 * But only if not writing to stdout, since in that case
1488 * the network is most likely throttling writes anyway,
1489 * and therefore it is best to go to the write phase ASAP
1490 * instead, as we can afford spending more time compressing
1491 * between writes at that moment.
1493 if (entry->delta_data && !pack_to_stdout) {
1494 entry->z_delta_size = do_compress(&entry->delta_data,
1495 entry->delta_size);
1496 cache_lock();
1497 delta_cache_size -= entry->delta_size;
1498 delta_cache_size += entry->z_delta_size;
1499 cache_unlock();
1502 /* if we made n a delta, and if n is already at max
1503 * depth, leaving it in the window is pointless. we
1504 * should evict it first.
1506 if (entry->delta && max_depth <= n->depth)
1507 continue;
1510 * Move the best delta base up in the window, after the
1511 * currently deltified object, to keep it longer. It will
1512 * be the first base object to be attempted next.
1514 if (entry->delta) {
1515 struct unpacked swap = array[best_base];
1516 int dist = (window + idx - best_base) % window;
1517 int dst = best_base;
1518 while (dist--) {
1519 int src = (dst + 1) % window;
1520 array[dst] = array[src];
1521 dst = src;
1523 array[dst] = swap;
1526 next:
1527 idx++;
1528 if (count + 1 < window)
1529 count++;
1530 if (idx >= window)
1531 idx = 0;
1534 for (i = 0; i < window; ++i) {
1535 free_delta_index(array[i].index);
1536 free(array[i].data);
1538 free(array);
1541 #ifdef THREADED_DELTA_SEARCH
1544 * The main thread waits on the condition that (at least) one of the workers
1545 * has stopped working (which is indicated in the .working member of
1546 * struct thread_params).
1547 * When a work thread has completed its work, it sets .working to 0 and
1548 * signals the main thread and waits on the condition that .data_ready
1549 * becomes 1.
1552 struct thread_params {
1553 pthread_t thread;
1554 struct object_entry **list;
1555 unsigned list_size;
1556 unsigned remaining;
1557 int window;
1558 int depth;
1559 int working;
1560 int data_ready;
1561 pthread_mutex_t mutex;
1562 pthread_cond_t cond;
1563 unsigned *processed;
1566 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1568 static void *threaded_find_deltas(void *arg)
1570 struct thread_params *me = arg;
1572 while (me->remaining) {
1573 find_deltas(me->list, &me->remaining,
1574 me->window, me->depth, me->processed);
1576 progress_lock();
1577 me->working = 0;
1578 pthread_cond_signal(&progress_cond);
1579 progress_unlock();
1582 * We must not set ->data_ready before we wait on the
1583 * condition because the main thread may have set it to 1
1584 * before we get here. In order to be sure that new
1585 * work is available if we see 1 in ->data_ready, it
1586 * was initialized to 0 before this thread was spawned
1587 * and we reset it to 0 right away.
1589 pthread_mutex_lock(&me->mutex);
1590 while (!me->data_ready)
1591 pthread_cond_wait(&me->cond, &me->mutex);
1592 me->data_ready = 0;
1593 pthread_mutex_unlock(&me->mutex);
1595 /* leave ->working 1 so that this doesn't get more work assigned */
1596 return NULL;
1599 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1600 int window, int depth, unsigned *processed)
1602 struct thread_params p[delta_search_threads];
1603 int i, ret, active_threads = 0;
1605 if (delta_search_threads <= 1) {
1606 find_deltas(list, &list_size, window, depth, processed);
1607 return;
1609 if (progress > pack_to_stdout)
1610 fprintf(stderr, "Delta compression using up to %d threads.\n",
1611 delta_search_threads);
1613 /* Partition the work amongst work threads. */
1614 for (i = 0; i < delta_search_threads; i++) {
1615 unsigned sub_size = list_size / (delta_search_threads - i);
1617 /* don't use too small segments or no deltas will be found */
1618 if (sub_size < 2*window && i+1 < delta_search_threads)
1619 sub_size = 0;
1621 p[i].window = window;
1622 p[i].depth = depth;
1623 p[i].processed = processed;
1624 p[i].working = 1;
1625 p[i].data_ready = 0;
1627 /* try to split chunks on "path" boundaries */
1628 while (sub_size && sub_size < list_size &&
1629 list[sub_size]->hash &&
1630 list[sub_size]->hash == list[sub_size-1]->hash)
1631 sub_size++;
1633 p[i].list = list;
1634 p[i].list_size = sub_size;
1635 p[i].remaining = sub_size;
1637 list += sub_size;
1638 list_size -= sub_size;
1641 /* Start work threads. */
1642 for (i = 0; i < delta_search_threads; i++) {
1643 if (!p[i].list_size)
1644 continue;
1645 pthread_mutex_init(&p[i].mutex, NULL);
1646 pthread_cond_init(&p[i].cond, NULL);
1647 ret = pthread_create(&p[i].thread, NULL,
1648 threaded_find_deltas, &p[i]);
1649 if (ret)
1650 die("unable to create thread: %s", strerror(ret));
1651 active_threads++;
1655 * Now let's wait for work completion. Each time a thread is done
1656 * with its work, we steal half of the remaining work from the
1657 * thread with the largest number of unprocessed objects and give
1658 * it to that newly idle thread. This ensure good load balancing
1659 * until the remaining object list segments are simply too short
1660 * to be worth splitting anymore.
1662 while (active_threads) {
1663 struct thread_params *target = NULL;
1664 struct thread_params *victim = NULL;
1665 unsigned sub_size = 0;
1667 progress_lock();
1668 for (;;) {
1669 for (i = 0; !target && i < delta_search_threads; i++)
1670 if (!p[i].working)
1671 target = &p[i];
1672 if (target)
1673 break;
1674 pthread_cond_wait(&progress_cond, &progress_mutex);
1677 for (i = 0; i < delta_search_threads; i++)
1678 if (p[i].remaining > 2*window &&
1679 (!victim || victim->remaining < p[i].remaining))
1680 victim = &p[i];
1681 if (victim) {
1682 sub_size = victim->remaining / 2;
1683 list = victim->list + victim->list_size - sub_size;
1684 while (sub_size && list[0]->hash &&
1685 list[0]->hash == list[-1]->hash) {
1686 list++;
1687 sub_size--;
1689 if (!sub_size) {
1691 * It is possible for some "paths" to have
1692 * so many objects that no hash boundary
1693 * might be found. Let's just steal the
1694 * exact half in that case.
1696 sub_size = victim->remaining / 2;
1697 list -= sub_size;
1699 target->list = list;
1700 victim->list_size -= sub_size;
1701 victim->remaining -= sub_size;
1703 target->list_size = sub_size;
1704 target->remaining = sub_size;
1705 target->working = 1;
1706 progress_unlock();
1708 pthread_mutex_lock(&target->mutex);
1709 target->data_ready = 1;
1710 pthread_cond_signal(&target->cond);
1711 pthread_mutex_unlock(&target->mutex);
1713 if (!sub_size) {
1714 pthread_join(target->thread, NULL);
1715 pthread_cond_destroy(&target->cond);
1716 pthread_mutex_destroy(&target->mutex);
1717 active_threads--;
1722 #else
1723 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1724 #endif
1726 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1728 unsigned char peeled[20];
1730 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1731 !peel_ref(path, peeled) && /* peelable? */
1732 !is_null_sha1(peeled) && /* annotated tag? */
1733 locate_object_entry(peeled)) /* object packed? */
1734 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1735 return 0;
1738 static void prepare_pack(int window, int depth)
1740 struct object_entry **delta_list;
1741 uint32_t i, nr_deltas;
1742 unsigned n;
1744 get_object_details();
1747 * If we're locally repacking then we need to be doubly careful
1748 * from now on in order to make sure no stealth corruption gets
1749 * propagated to the new pack. Clients receiving streamed packs
1750 * should validate everything they get anyway so no need to incur
1751 * the additional cost here in that case.
1753 if (!pack_to_stdout)
1754 do_check_packed_object_crc = 1;
1756 if (!nr_objects || !window || !depth)
1757 return;
1759 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1760 nr_deltas = n = 0;
1762 for (i = 0; i < nr_objects; i++) {
1763 struct object_entry *entry = objects + i;
1765 if (entry->delta)
1766 /* This happens if we decided to reuse existing
1767 * delta from a pack. "reuse_delta &&" is implied.
1769 continue;
1771 if (entry->size < 50)
1772 continue;
1774 if (entry->no_try_delta)
1775 continue;
1777 if (!entry->preferred_base) {
1778 nr_deltas++;
1779 if (entry->type < 0)
1780 die("unable to get type of object %s",
1781 sha1_to_hex(entry->idx.sha1));
1782 } else {
1783 if (entry->type < 0) {
1785 * This object is not found, but we
1786 * don't have to include it anyway.
1788 continue;
1792 delta_list[n++] = entry;
1795 if (nr_deltas && n > 1) {
1796 unsigned nr_done = 0;
1797 if (progress)
1798 progress_state = start_progress("Compressing objects",
1799 nr_deltas);
1800 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1801 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1802 stop_progress(&progress_state);
1803 if (nr_done != nr_deltas)
1804 die("inconsistency with delta count");
1806 free(delta_list);
1809 static int git_pack_config(const char *k, const char *v, void *cb)
1811 if(!strcmp(k, "pack.window")) {
1812 window = git_config_int(k, v);
1813 return 0;
1815 if (!strcmp(k, "pack.windowmemory")) {
1816 window_memory_limit = git_config_ulong(k, v);
1817 return 0;
1819 if (!strcmp(k, "pack.depth")) {
1820 depth = git_config_int(k, v);
1821 return 0;
1823 if (!strcmp(k, "pack.compression")) {
1824 int level = git_config_int(k, v);
1825 if (level == -1)
1826 level = Z_DEFAULT_COMPRESSION;
1827 else if (level < 0 || level > Z_BEST_COMPRESSION)
1828 die("bad pack compression level %d", level);
1829 pack_compression_level = level;
1830 pack_compression_seen = 1;
1831 return 0;
1833 if (!strcmp(k, "pack.deltacachesize")) {
1834 max_delta_cache_size = git_config_int(k, v);
1835 return 0;
1837 if (!strcmp(k, "pack.deltacachelimit")) {
1838 cache_max_small_delta_size = git_config_int(k, v);
1839 return 0;
1841 if (!strcmp(k, "pack.threads")) {
1842 delta_search_threads = git_config_int(k, v);
1843 if (delta_search_threads < 0)
1844 die("invalid number of threads specified (%d)",
1845 delta_search_threads);
1846 #ifndef THREADED_DELTA_SEARCH
1847 if (delta_search_threads != 1)
1848 warning("no threads support, ignoring %s", k);
1849 #endif
1850 return 0;
1852 if (!strcmp(k, "pack.indexversion")) {
1853 pack_idx_default_version = git_config_int(k, v);
1854 if (pack_idx_default_version > 2)
1855 die("bad pack.indexversion=%"PRIu32,
1856 pack_idx_default_version);
1857 return 0;
1859 if (!strcmp(k, "pack.packsizelimit")) {
1860 pack_size_limit_cfg = git_config_ulong(k, v);
1861 return 0;
1863 return git_default_config(k, v, cb);
1866 static void read_object_list_from_stdin(void)
1868 char line[40 + 1 + PATH_MAX + 2];
1869 unsigned char sha1[20];
1871 for (;;) {
1872 if (!fgets(line, sizeof(line), stdin)) {
1873 if (feof(stdin))
1874 break;
1875 if (!ferror(stdin))
1876 die("fgets returned NULL, not EOF, not error!");
1877 if (errno != EINTR)
1878 die_errno("fgets");
1879 clearerr(stdin);
1880 continue;
1882 if (line[0] == '-') {
1883 if (get_sha1_hex(line+1, sha1))
1884 die("expected edge sha1, got garbage:\n %s",
1885 line);
1886 add_preferred_base(sha1);
1887 continue;
1889 if (get_sha1_hex(line, sha1))
1890 die("expected sha1, got garbage:\n %s", line);
1892 add_preferred_base_object(line+41);
1893 add_object_entry(sha1, 0, line+41, 0);
1897 #define OBJECT_ADDED (1u<<20)
1899 static void show_commit(struct commit *commit, void *data)
1901 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1902 commit->object.flags |= OBJECT_ADDED;
1905 static void show_object(struct object *obj, const struct name_path *path, const char *last)
1907 char *name = path_name(path, last);
1909 add_preferred_base_object(name);
1910 add_object_entry(obj->sha1, obj->type, name, 0);
1911 obj->flags |= OBJECT_ADDED;
1914 * We will have generated the hash from the name,
1915 * but not saved a pointer to it - we can free it
1917 free((char *)name);
1920 static void show_edge(struct commit *commit)
1922 add_preferred_base(commit->object.sha1);
1925 struct in_pack_object {
1926 off_t offset;
1927 struct object *object;
1930 struct in_pack {
1931 int alloc;
1932 int nr;
1933 struct in_pack_object *array;
1936 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1938 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1939 in_pack->array[in_pack->nr].object = object;
1940 in_pack->nr++;
1944 * Compare the objects in the offset order, in order to emulate the
1945 * "git rev-list --objects" output that produced the pack originally.
1947 static int ofscmp(const void *a_, const void *b_)
1949 struct in_pack_object *a = (struct in_pack_object *)a_;
1950 struct in_pack_object *b = (struct in_pack_object *)b_;
1952 if (a->offset < b->offset)
1953 return -1;
1954 else if (a->offset > b->offset)
1955 return 1;
1956 else
1957 return hashcmp(a->object->sha1, b->object->sha1);
1960 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1962 struct packed_git *p;
1963 struct in_pack in_pack;
1964 uint32_t i;
1966 memset(&in_pack, 0, sizeof(in_pack));
1968 for (p = packed_git; p; p = p->next) {
1969 const unsigned char *sha1;
1970 struct object *o;
1972 if (!p->pack_local || p->pack_keep)
1973 continue;
1974 if (open_pack_index(p))
1975 die("cannot open pack index");
1977 ALLOC_GROW(in_pack.array,
1978 in_pack.nr + p->num_objects,
1979 in_pack.alloc);
1981 for (i = 0; i < p->num_objects; i++) {
1982 sha1 = nth_packed_object_sha1(p, i);
1983 o = lookup_unknown_object(sha1);
1984 if (!(o->flags & OBJECT_ADDED))
1985 mark_in_pack_object(o, p, &in_pack);
1986 o->flags |= OBJECT_ADDED;
1990 if (in_pack.nr) {
1991 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1992 ofscmp);
1993 for (i = 0; i < in_pack.nr; i++) {
1994 struct object *o = in_pack.array[i].object;
1995 add_object_entry(o->sha1, o->type, "", 0);
1998 free(in_pack.array);
2001 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2003 static struct packed_git *last_found = (void *)1;
2004 struct packed_git *p;
2006 p = (last_found != (void *)1) ? last_found : packed_git;
2008 while (p) {
2009 if ((!p->pack_local || p->pack_keep) &&
2010 find_pack_entry_one(sha1, p)) {
2011 last_found = p;
2012 return 1;
2014 if (p == last_found)
2015 p = packed_git;
2016 else
2017 p = p->next;
2018 if (p == last_found)
2019 p = p->next;
2021 return 0;
2024 static void loosen_unused_packed_objects(struct rev_info *revs)
2026 struct packed_git *p;
2027 uint32_t i;
2028 const unsigned char *sha1;
2030 for (p = packed_git; p; p = p->next) {
2031 if (!p->pack_local || p->pack_keep)
2032 continue;
2034 if (open_pack_index(p))
2035 die("cannot open pack index");
2037 for (i = 0; i < p->num_objects; i++) {
2038 sha1 = nth_packed_object_sha1(p, i);
2039 if (!locate_object_entry(sha1) &&
2040 !has_sha1_pack_kept_or_nonlocal(sha1))
2041 if (force_object_loose(sha1, p->mtime))
2042 die("unable to force loose object");
2047 static void get_object_list(int ac, const char **av)
2049 struct rev_info revs;
2050 char line[1000];
2051 int flags = 0;
2053 init_revisions(&revs, NULL);
2054 save_commit_buffer = 0;
2055 setup_revisions(ac, av, &revs, NULL);
2057 while (fgets(line, sizeof(line), stdin) != NULL) {
2058 int len = strlen(line);
2059 if (len && line[len - 1] == '\n')
2060 line[--len] = 0;
2061 if (!len)
2062 break;
2063 if (*line == '-') {
2064 if (!strcmp(line, "--not")) {
2065 flags ^= UNINTERESTING;
2066 continue;
2068 die("not a rev '%s'", line);
2070 if (handle_revision_arg(line, &revs, flags, 1))
2071 die("bad revision '%s'", line);
2074 if (prepare_revision_walk(&revs))
2075 die("revision walk setup failed");
2076 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2077 traverse_commit_list(&revs, show_commit, show_object, NULL);
2079 if (keep_unreachable)
2080 add_objects_in_unpacked_packs(&revs);
2081 if (unpack_unreachable)
2082 loosen_unused_packed_objects(&revs);
2085 static int adjust_perm(const char *path, mode_t mode)
2087 if (chmod(path, mode))
2088 return -1;
2089 return adjust_shared_perm(path);
2092 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2094 int use_internal_rev_list = 0;
2095 int thin = 0;
2096 uint32_t i;
2097 const char **rp_av;
2098 int rp_ac_alloc = 64;
2099 int rp_ac;
2101 read_replace_refs = 0;
2103 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2105 rp_av[0] = "pack-objects";
2106 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2107 rp_ac = 2;
2109 git_config(git_pack_config, NULL);
2110 if (!pack_compression_seen && core_compression_seen)
2111 pack_compression_level = core_compression_level;
2113 progress = isatty(2);
2114 for (i = 1; i < argc; i++) {
2115 const char *arg = argv[i];
2117 if (*arg != '-')
2118 break;
2120 if (!strcmp("--non-empty", arg)) {
2121 non_empty = 1;
2122 continue;
2124 if (!strcmp("--local", arg)) {
2125 local = 1;
2126 continue;
2128 if (!strcmp("--incremental", arg)) {
2129 incremental = 1;
2130 continue;
2132 if (!strcmp("--honor-pack-keep", arg)) {
2133 ignore_packed_keep = 1;
2134 continue;
2136 if (!prefixcmp(arg, "--compression=")) {
2137 char *end;
2138 int level = strtoul(arg+14, &end, 0);
2139 if (!arg[14] || *end)
2140 usage(pack_usage);
2141 if (level == -1)
2142 level = Z_DEFAULT_COMPRESSION;
2143 else if (level < 0 || level > Z_BEST_COMPRESSION)
2144 die("bad pack compression level %d", level);
2145 pack_compression_level = level;
2146 continue;
2148 if (!prefixcmp(arg, "--max-pack-size=")) {
2149 char *end;
2150 pack_size_limit_cfg = 0;
2151 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2152 if (!arg[16] || *end)
2153 usage(pack_usage);
2154 continue;
2156 if (!prefixcmp(arg, "--window=")) {
2157 char *end;
2158 window = strtoul(arg+9, &end, 0);
2159 if (!arg[9] || *end)
2160 usage(pack_usage);
2161 continue;
2163 if (!prefixcmp(arg, "--window-memory=")) {
2164 if (!git_parse_ulong(arg+16, &window_memory_limit))
2165 usage(pack_usage);
2166 continue;
2168 if (!prefixcmp(arg, "--threads=")) {
2169 char *end;
2170 delta_search_threads = strtoul(arg+10, &end, 0);
2171 if (!arg[10] || *end || delta_search_threads < 0)
2172 usage(pack_usage);
2173 #ifndef THREADED_DELTA_SEARCH
2174 if (delta_search_threads != 1)
2175 warning("no threads support, "
2176 "ignoring %s", arg);
2177 #endif
2178 continue;
2180 if (!prefixcmp(arg, "--depth=")) {
2181 char *end;
2182 depth = strtoul(arg+8, &end, 0);
2183 if (!arg[8] || *end)
2184 usage(pack_usage);
2185 continue;
2187 if (!strcmp("--progress", arg)) {
2188 progress = 1;
2189 continue;
2191 if (!strcmp("--all-progress", arg)) {
2192 progress = 2;
2193 continue;
2195 if (!strcmp("-q", arg)) {
2196 progress = 0;
2197 continue;
2199 if (!strcmp("--no-reuse-delta", arg)) {
2200 reuse_delta = 0;
2201 continue;
2203 if (!strcmp("--no-reuse-object", arg)) {
2204 reuse_object = reuse_delta = 0;
2205 continue;
2207 if (!strcmp("--delta-base-offset", arg)) {
2208 allow_ofs_delta = 1;
2209 continue;
2211 if (!strcmp("--stdout", arg)) {
2212 pack_to_stdout = 1;
2213 continue;
2215 if (!strcmp("--revs", arg)) {
2216 use_internal_rev_list = 1;
2217 continue;
2219 if (!strcmp("--keep-unreachable", arg)) {
2220 keep_unreachable = 1;
2221 continue;
2223 if (!strcmp("--unpack-unreachable", arg)) {
2224 unpack_unreachable = 1;
2225 continue;
2227 if (!strcmp("--include-tag", arg)) {
2228 include_tag = 1;
2229 continue;
2231 if (!strcmp("--unpacked", arg) ||
2232 !strcmp("--reflog", arg) ||
2233 !strcmp("--all", arg)) {
2234 use_internal_rev_list = 1;
2235 if (rp_ac >= rp_ac_alloc - 1) {
2236 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2237 rp_av = xrealloc(rp_av,
2238 rp_ac_alloc * sizeof(*rp_av));
2240 rp_av[rp_ac++] = arg;
2241 continue;
2243 if (!strcmp("--thin", arg)) {
2244 use_internal_rev_list = 1;
2245 thin = 1;
2246 rp_av[1] = "--objects-edge";
2247 continue;
2249 if (!prefixcmp(arg, "--index-version=")) {
2250 char *c;
2251 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2252 if (pack_idx_default_version > 2)
2253 die("bad %s", arg);
2254 if (*c == ',')
2255 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2256 if (*c || pack_idx_off32_limit & 0x80000000)
2257 die("bad %s", arg);
2258 continue;
2260 if (!strcmp(arg, "--grafts-replace-parents=no")) {
2261 grafts_replace_parents = 0;
2262 continue;
2264 usage(pack_usage);
2267 /* Traditionally "pack-objects [options] base extra" failed;
2268 * we would however want to take refs parameter that would
2269 * have been given to upstream rev-list ourselves, which means
2270 * we somehow want to say what the base name is. So the
2271 * syntax would be:
2273 * pack-objects [options] base <refs...>
2275 * in other words, we would treat the first non-option as the
2276 * base_name and send everything else to the internal revision
2277 * walker.
2280 if (!pack_to_stdout)
2281 base_name = argv[i++];
2283 if (pack_to_stdout != !base_name)
2284 usage(pack_usage);
2286 if (!pack_to_stdout && !pack_size_limit)
2287 pack_size_limit = pack_size_limit_cfg;
2289 if (pack_to_stdout && pack_size_limit)
2290 die("--max-pack-size cannot be used to build a pack for transfer.");
2292 if (!pack_to_stdout && thin)
2293 die("--thin cannot be used to build an indexable pack.");
2295 if (keep_unreachable && unpack_unreachable)
2296 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2298 #ifdef THREADED_DELTA_SEARCH
2299 if (!delta_search_threads) /* --threads=0 means autodetect */
2300 delta_search_threads = online_cpus();
2301 #endif
2303 prepare_packed_git();
2305 if (progress)
2306 progress_state = start_progress("Counting objects", 0);
2307 if (!use_internal_rev_list)
2308 read_object_list_from_stdin();
2309 else {
2310 rp_av[rp_ac] = NULL;
2311 get_object_list(rp_ac, rp_av);
2313 if (include_tag && nr_result)
2314 for_each_ref(add_ref_tag, NULL);
2315 stop_progress(&progress_state);
2317 if (non_empty && !nr_result)
2318 return 0;
2319 if (nr_result)
2320 prepare_pack(window, depth);
2321 write_pack_file();
2322 if (progress)
2323 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2324 " reused %"PRIu32" (delta %"PRIu32")\n",
2325 written, written_delta, reused, reused_delta);
2326 return 0;