Add selftests verifying concatenation of multiple notes for the same commit
[git/jnareb-git.git] / builtin-pack-objects.c
blob02f9246cdb29903e9d7aafe47604f6882a3e7f86
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]*]\n"
31 " [--reflog] [--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 = 256 * 1024 * 1024;
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 cleanup_preferred_base(void)
1013 struct pbase_tree *it;
1014 unsigned i;
1016 it = pbase_tree;
1017 pbase_tree = NULL;
1018 while (it) {
1019 struct pbase_tree *this = it;
1020 it = this->next;
1021 free(this->pcache.tree_data);
1022 free(this);
1025 for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1026 if (!pbase_tree_cache[i])
1027 continue;
1028 free(pbase_tree_cache[i]->tree_data);
1029 free(pbase_tree_cache[i]);
1030 pbase_tree_cache[i] = NULL;
1033 free(done_pbase_paths);
1034 done_pbase_paths = NULL;
1035 done_pbase_paths_num = done_pbase_paths_alloc = 0;
1038 static void check_object(struct object_entry *entry)
1040 if (entry->in_pack) {
1041 struct packed_git *p = entry->in_pack;
1042 struct pack_window *w_curs = NULL;
1043 const unsigned char *base_ref = NULL;
1044 struct object_entry *base_entry;
1045 unsigned long used, used_0;
1046 unsigned int avail;
1047 off_t ofs;
1048 unsigned char *buf, c;
1050 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1053 * We want in_pack_type even if we do not reuse delta
1054 * since non-delta representations could still be reused.
1056 used = unpack_object_header_buffer(buf, avail,
1057 &entry->in_pack_type,
1058 &entry->size);
1059 if (used == 0)
1060 goto give_up;
1063 * Determine if this is a delta and if so whether we can
1064 * reuse it or not. Otherwise let's find out as cheaply as
1065 * possible what the actual type and size for this object is.
1067 switch (entry->in_pack_type) {
1068 default:
1069 /* Not a delta hence we've already got all we need. */
1070 entry->type = entry->in_pack_type;
1071 entry->in_pack_header_size = used;
1072 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1073 goto give_up;
1074 unuse_pack(&w_curs);
1075 return;
1076 case OBJ_REF_DELTA:
1077 if (reuse_delta && !entry->preferred_base)
1078 base_ref = use_pack(p, &w_curs,
1079 entry->in_pack_offset + used, NULL);
1080 entry->in_pack_header_size = used + 20;
1081 break;
1082 case OBJ_OFS_DELTA:
1083 buf = use_pack(p, &w_curs,
1084 entry->in_pack_offset + used, NULL);
1085 used_0 = 0;
1086 c = buf[used_0++];
1087 ofs = c & 127;
1088 while (c & 128) {
1089 ofs += 1;
1090 if (!ofs || MSB(ofs, 7)) {
1091 error("delta base offset overflow in pack for %s",
1092 sha1_to_hex(entry->idx.sha1));
1093 goto give_up;
1095 c = buf[used_0++];
1096 ofs = (ofs << 7) + (c & 127);
1098 ofs = entry->in_pack_offset - ofs;
1099 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1100 error("delta base offset out of bound for %s",
1101 sha1_to_hex(entry->idx.sha1));
1102 goto give_up;
1104 if (reuse_delta && !entry->preferred_base) {
1105 struct revindex_entry *revidx;
1106 revidx = find_pack_revindex(p, ofs);
1107 if (!revidx)
1108 goto give_up;
1109 base_ref = nth_packed_object_sha1(p, revidx->nr);
1111 entry->in_pack_header_size = used + used_0;
1112 break;
1115 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1117 * If base_ref was set above that means we wish to
1118 * reuse delta data, and we even found that base
1119 * in the list of objects we want to pack. Goodie!
1121 * Depth value does not matter - find_deltas() will
1122 * never consider reused delta as the base object to
1123 * deltify other objects against, in order to avoid
1124 * circular deltas.
1126 entry->type = entry->in_pack_type;
1127 entry->delta = base_entry;
1128 entry->delta_size = entry->size;
1129 entry->delta_sibling = base_entry->delta_child;
1130 base_entry->delta_child = entry;
1131 unuse_pack(&w_curs);
1132 return;
1135 if (entry->type) {
1137 * This must be a delta and we already know what the
1138 * final object type is. Let's extract the actual
1139 * object size from the delta header.
1141 entry->size = get_size_from_delta(p, &w_curs,
1142 entry->in_pack_offset + entry->in_pack_header_size);
1143 if (entry->size == 0)
1144 goto give_up;
1145 unuse_pack(&w_curs);
1146 return;
1150 * No choice but to fall back to the recursive delta walk
1151 * with sha1_object_info() to find about the object type
1152 * at this point...
1154 give_up:
1155 unuse_pack(&w_curs);
1158 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1160 * The error condition is checked in prepare_pack(). This is
1161 * to permit a missing preferred base object to be ignored
1162 * as a preferred base. Doing so can result in a larger
1163 * pack file, but the transfer will still take place.
1167 static int pack_offset_sort(const void *_a, const void *_b)
1169 const struct object_entry *a = *(struct object_entry **)_a;
1170 const struct object_entry *b = *(struct object_entry **)_b;
1172 /* avoid filesystem trashing with loose objects */
1173 if (!a->in_pack && !b->in_pack)
1174 return hashcmp(a->idx.sha1, b->idx.sha1);
1176 if (a->in_pack < b->in_pack)
1177 return -1;
1178 if (a->in_pack > b->in_pack)
1179 return 1;
1180 return a->in_pack_offset < b->in_pack_offset ? -1 :
1181 (a->in_pack_offset > b->in_pack_offset);
1184 static void get_object_details(void)
1186 uint32_t i;
1187 struct object_entry **sorted_by_offset;
1189 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1190 for (i = 0; i < nr_objects; i++)
1191 sorted_by_offset[i] = objects + i;
1192 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1194 for (i = 0; i < nr_objects; i++)
1195 check_object(sorted_by_offset[i]);
1197 free(sorted_by_offset);
1201 * We search for deltas in a list sorted by type, by filename hash, and then
1202 * by size, so that we see progressively smaller and smaller files.
1203 * That's because we prefer deltas to be from the bigger file
1204 * to the smaller -- deletes are potentially cheaper, but perhaps
1205 * more importantly, the bigger file is likely the more recent
1206 * one. The deepest deltas are therefore the oldest objects which are
1207 * less susceptible to be accessed often.
1209 static int type_size_sort(const void *_a, const void *_b)
1211 const struct object_entry *a = *(struct object_entry **)_a;
1212 const struct object_entry *b = *(struct object_entry **)_b;
1214 if (a->type > b->type)
1215 return -1;
1216 if (a->type < b->type)
1217 return 1;
1218 if (a->hash > b->hash)
1219 return -1;
1220 if (a->hash < b->hash)
1221 return 1;
1222 if (a->preferred_base > b->preferred_base)
1223 return -1;
1224 if (a->preferred_base < b->preferred_base)
1225 return 1;
1226 if (a->size > b->size)
1227 return -1;
1228 if (a->size < b->size)
1229 return 1;
1230 return a < b ? -1 : (a > b); /* newest first */
1233 struct unpacked {
1234 struct object_entry *entry;
1235 void *data;
1236 struct delta_index *index;
1237 unsigned depth;
1240 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1241 unsigned long delta_size)
1243 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1244 return 0;
1246 if (delta_size < cache_max_small_delta_size)
1247 return 1;
1249 /* cache delta, if objects are large enough compared to delta size */
1250 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1251 return 1;
1253 return 0;
1256 #ifdef THREADED_DELTA_SEARCH
1258 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1259 #define read_lock() pthread_mutex_lock(&read_mutex)
1260 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1262 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1263 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1264 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1266 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1267 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1268 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1270 #else
1272 #define read_lock() (void)0
1273 #define read_unlock() (void)0
1274 #define cache_lock() (void)0
1275 #define cache_unlock() (void)0
1276 #define progress_lock() (void)0
1277 #define progress_unlock() (void)0
1279 #endif
1281 static int try_delta(struct unpacked *trg, struct unpacked *src,
1282 unsigned max_depth, unsigned long *mem_usage)
1284 struct object_entry *trg_entry = trg->entry;
1285 struct object_entry *src_entry = src->entry;
1286 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1287 unsigned ref_depth;
1288 enum object_type type;
1289 void *delta_buf;
1291 /* Don't bother doing diffs between different types */
1292 if (trg_entry->type != src_entry->type)
1293 return -1;
1296 * We do not bother to try a delta that we discarded
1297 * on an earlier try, but only when reusing delta data.
1299 if (reuse_delta && trg_entry->in_pack &&
1300 trg_entry->in_pack == src_entry->in_pack &&
1301 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1302 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1303 return 0;
1305 /* Let's not bust the allowed depth. */
1306 if (src->depth >= max_depth)
1307 return 0;
1309 /* Now some size filtering heuristics. */
1310 trg_size = trg_entry->size;
1311 if (!trg_entry->delta) {
1312 max_size = trg_size/2 - 20;
1313 ref_depth = 1;
1314 } else {
1315 max_size = trg_entry->delta_size;
1316 ref_depth = trg->depth;
1318 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1319 (max_depth - ref_depth + 1);
1320 if (max_size == 0)
1321 return 0;
1322 src_size = src_entry->size;
1323 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1324 if (sizediff >= max_size)
1325 return 0;
1326 if (trg_size < src_size / 32)
1327 return 0;
1329 /* Load data if not already done */
1330 if (!trg->data) {
1331 read_lock();
1332 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1333 read_unlock();
1334 if (!trg->data)
1335 die("object %s cannot be read",
1336 sha1_to_hex(trg_entry->idx.sha1));
1337 if (sz != trg_size)
1338 die("object %s inconsistent object length (%lu vs %lu)",
1339 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1340 *mem_usage += sz;
1342 if (!src->data) {
1343 read_lock();
1344 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1345 read_unlock();
1346 if (!src->data)
1347 die("object %s cannot be read",
1348 sha1_to_hex(src_entry->idx.sha1));
1349 if (sz != src_size)
1350 die("object %s inconsistent object length (%lu vs %lu)",
1351 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1352 *mem_usage += sz;
1354 if (!src->index) {
1355 src->index = create_delta_index(src->data, src_size);
1356 if (!src->index) {
1357 static int warned = 0;
1358 if (!warned++)
1359 warning("suboptimal pack - out of memory");
1360 return 0;
1362 *mem_usage += sizeof_delta_index(src->index);
1365 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1366 if (!delta_buf)
1367 return 0;
1369 if (trg_entry->delta) {
1370 /* Prefer only shallower same-sized deltas. */
1371 if (delta_size == trg_entry->delta_size &&
1372 src->depth + 1 >= trg->depth) {
1373 free(delta_buf);
1374 return 0;
1379 * Handle memory allocation outside of the cache
1380 * accounting lock. Compiler will optimize the strangeness
1381 * away when THREADED_DELTA_SEARCH is not defined.
1383 free(trg_entry->delta_data);
1384 cache_lock();
1385 if (trg_entry->delta_data) {
1386 delta_cache_size -= trg_entry->delta_size;
1387 trg_entry->delta_data = NULL;
1389 if (delta_cacheable(src_size, trg_size, delta_size)) {
1390 delta_cache_size += delta_size;
1391 cache_unlock();
1392 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1393 } else {
1394 cache_unlock();
1395 free(delta_buf);
1398 trg_entry->delta = src_entry;
1399 trg_entry->delta_size = delta_size;
1400 trg->depth = src->depth + 1;
1402 return 1;
1405 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1407 struct object_entry *child = me->delta_child;
1408 unsigned int m = n;
1409 while (child) {
1410 unsigned int c = check_delta_limit(child, n + 1);
1411 if (m < c)
1412 m = c;
1413 child = child->delta_sibling;
1415 return m;
1418 static unsigned long free_unpacked(struct unpacked *n)
1420 unsigned long freed_mem = sizeof_delta_index(n->index);
1421 free_delta_index(n->index);
1422 n->index = NULL;
1423 if (n->data) {
1424 freed_mem += n->entry->size;
1425 free(n->data);
1426 n->data = NULL;
1428 n->entry = NULL;
1429 n->depth = 0;
1430 return freed_mem;
1433 static void find_deltas(struct object_entry **list, unsigned *list_size,
1434 int window, int depth, unsigned *processed)
1436 uint32_t i, idx = 0, count = 0;
1437 struct unpacked *array;
1438 unsigned long mem_usage = 0;
1440 array = xcalloc(window, sizeof(struct unpacked));
1442 for (;;) {
1443 struct object_entry *entry;
1444 struct unpacked *n = array + idx;
1445 int j, max_depth, best_base = -1;
1447 progress_lock();
1448 if (!*list_size) {
1449 progress_unlock();
1450 break;
1452 entry = *list++;
1453 (*list_size)--;
1454 if (!entry->preferred_base) {
1455 (*processed)++;
1456 display_progress(progress_state, *processed);
1458 progress_unlock();
1460 mem_usage -= free_unpacked(n);
1461 n->entry = entry;
1463 while (window_memory_limit &&
1464 mem_usage > window_memory_limit &&
1465 count > 1) {
1466 uint32_t tail = (idx + window - count) % window;
1467 mem_usage -= free_unpacked(array + tail);
1468 count--;
1471 /* We do not compute delta to *create* objects we are not
1472 * going to pack.
1474 if (entry->preferred_base)
1475 goto next;
1478 * If the current object is at pack edge, take the depth the
1479 * objects that depend on the current object into account
1480 * otherwise they would become too deep.
1482 max_depth = depth;
1483 if (entry->delta_child) {
1484 max_depth -= check_delta_limit(entry, 0);
1485 if (max_depth <= 0)
1486 goto next;
1489 j = window;
1490 while (--j > 0) {
1491 int ret;
1492 uint32_t other_idx = idx + j;
1493 struct unpacked *m;
1494 if (other_idx >= window)
1495 other_idx -= window;
1496 m = array + other_idx;
1497 if (!m->entry)
1498 break;
1499 ret = try_delta(n, m, max_depth, &mem_usage);
1500 if (ret < 0)
1501 break;
1502 else if (ret > 0)
1503 best_base = other_idx;
1507 * If we decided to cache the delta data, then it is best
1508 * to compress it right away. First because we have to do
1509 * it anyway, and doing it here while we're threaded will
1510 * save a lot of time in the non threaded write phase,
1511 * as well as allow for caching more deltas within
1512 * the same cache size limit.
1513 * ...
1514 * But only if not writing to stdout, since in that case
1515 * the network is most likely throttling writes anyway,
1516 * and therefore it is best to go to the write phase ASAP
1517 * instead, as we can afford spending more time compressing
1518 * between writes at that moment.
1520 if (entry->delta_data && !pack_to_stdout) {
1521 entry->z_delta_size = do_compress(&entry->delta_data,
1522 entry->delta_size);
1523 cache_lock();
1524 delta_cache_size -= entry->delta_size;
1525 delta_cache_size += entry->z_delta_size;
1526 cache_unlock();
1529 /* if we made n a delta, and if n is already at max
1530 * depth, leaving it in the window is pointless. we
1531 * should evict it first.
1533 if (entry->delta && max_depth <= n->depth)
1534 continue;
1537 * Move the best delta base up in the window, after the
1538 * currently deltified object, to keep it longer. It will
1539 * be the first base object to be attempted next.
1541 if (entry->delta) {
1542 struct unpacked swap = array[best_base];
1543 int dist = (window + idx - best_base) % window;
1544 int dst = best_base;
1545 while (dist--) {
1546 int src = (dst + 1) % window;
1547 array[dst] = array[src];
1548 dst = src;
1550 array[dst] = swap;
1553 next:
1554 idx++;
1555 if (count + 1 < window)
1556 count++;
1557 if (idx >= window)
1558 idx = 0;
1561 for (i = 0; i < window; ++i) {
1562 free_delta_index(array[i].index);
1563 free(array[i].data);
1565 free(array);
1568 #ifdef THREADED_DELTA_SEARCH
1571 * The main thread waits on the condition that (at least) one of the workers
1572 * has stopped working (which is indicated in the .working member of
1573 * struct thread_params).
1574 * When a work thread has completed its work, it sets .working to 0 and
1575 * signals the main thread and waits on the condition that .data_ready
1576 * becomes 1.
1579 struct thread_params {
1580 pthread_t thread;
1581 struct object_entry **list;
1582 unsigned list_size;
1583 unsigned remaining;
1584 int window;
1585 int depth;
1586 int working;
1587 int data_ready;
1588 pthread_mutex_t mutex;
1589 pthread_cond_t cond;
1590 unsigned *processed;
1593 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1595 static void *threaded_find_deltas(void *arg)
1597 struct thread_params *me = arg;
1599 while (me->remaining) {
1600 find_deltas(me->list, &me->remaining,
1601 me->window, me->depth, me->processed);
1603 progress_lock();
1604 me->working = 0;
1605 pthread_cond_signal(&progress_cond);
1606 progress_unlock();
1609 * We must not set ->data_ready before we wait on the
1610 * condition because the main thread may have set it to 1
1611 * before we get here. In order to be sure that new
1612 * work is available if we see 1 in ->data_ready, it
1613 * was initialized to 0 before this thread was spawned
1614 * and we reset it to 0 right away.
1616 pthread_mutex_lock(&me->mutex);
1617 while (!me->data_ready)
1618 pthread_cond_wait(&me->cond, &me->mutex);
1619 me->data_ready = 0;
1620 pthread_mutex_unlock(&me->mutex);
1622 /* leave ->working 1 so that this doesn't get more work assigned */
1623 return NULL;
1626 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1627 int window, int depth, unsigned *processed)
1629 struct thread_params *p;
1630 int i, ret, active_threads = 0;
1632 if (delta_search_threads <= 1) {
1633 find_deltas(list, &list_size, window, depth, processed);
1634 return;
1636 if (progress > pack_to_stdout)
1637 fprintf(stderr, "Delta compression using up to %d threads.\n",
1638 delta_search_threads);
1639 p = xcalloc(delta_search_threads, sizeof(*p));
1641 /* Partition the work amongst work threads. */
1642 for (i = 0; i < delta_search_threads; i++) {
1643 unsigned sub_size = list_size / (delta_search_threads - i);
1645 /* don't use too small segments or no deltas will be found */
1646 if (sub_size < 2*window && i+1 < delta_search_threads)
1647 sub_size = 0;
1649 p[i].window = window;
1650 p[i].depth = depth;
1651 p[i].processed = processed;
1652 p[i].working = 1;
1653 p[i].data_ready = 0;
1655 /* try to split chunks on "path" boundaries */
1656 while (sub_size && sub_size < list_size &&
1657 list[sub_size]->hash &&
1658 list[sub_size]->hash == list[sub_size-1]->hash)
1659 sub_size++;
1661 p[i].list = list;
1662 p[i].list_size = sub_size;
1663 p[i].remaining = sub_size;
1665 list += sub_size;
1666 list_size -= sub_size;
1669 /* Start work threads. */
1670 for (i = 0; i < delta_search_threads; i++) {
1671 if (!p[i].list_size)
1672 continue;
1673 pthread_mutex_init(&p[i].mutex, NULL);
1674 pthread_cond_init(&p[i].cond, NULL);
1675 ret = pthread_create(&p[i].thread, NULL,
1676 threaded_find_deltas, &p[i]);
1677 if (ret)
1678 die("unable to create thread: %s", strerror(ret));
1679 active_threads++;
1683 * Now let's wait for work completion. Each time a thread is done
1684 * with its work, we steal half of the remaining work from the
1685 * thread with the largest number of unprocessed objects and give
1686 * it to that newly idle thread. This ensure good load balancing
1687 * until the remaining object list segments are simply too short
1688 * to be worth splitting anymore.
1690 while (active_threads) {
1691 struct thread_params *target = NULL;
1692 struct thread_params *victim = NULL;
1693 unsigned sub_size = 0;
1695 progress_lock();
1696 for (;;) {
1697 for (i = 0; !target && i < delta_search_threads; i++)
1698 if (!p[i].working)
1699 target = &p[i];
1700 if (target)
1701 break;
1702 pthread_cond_wait(&progress_cond, &progress_mutex);
1705 for (i = 0; i < delta_search_threads; i++)
1706 if (p[i].remaining > 2*window &&
1707 (!victim || victim->remaining < p[i].remaining))
1708 victim = &p[i];
1709 if (victim) {
1710 sub_size = victim->remaining / 2;
1711 list = victim->list + victim->list_size - sub_size;
1712 while (sub_size && list[0]->hash &&
1713 list[0]->hash == list[-1]->hash) {
1714 list++;
1715 sub_size--;
1717 if (!sub_size) {
1719 * It is possible for some "paths" to have
1720 * so many objects that no hash boundary
1721 * might be found. Let's just steal the
1722 * exact half in that case.
1724 sub_size = victim->remaining / 2;
1725 list -= sub_size;
1727 target->list = list;
1728 victim->list_size -= sub_size;
1729 victim->remaining -= sub_size;
1731 target->list_size = sub_size;
1732 target->remaining = sub_size;
1733 target->working = 1;
1734 progress_unlock();
1736 pthread_mutex_lock(&target->mutex);
1737 target->data_ready = 1;
1738 pthread_cond_signal(&target->cond);
1739 pthread_mutex_unlock(&target->mutex);
1741 if (!sub_size) {
1742 pthread_join(target->thread, NULL);
1743 pthread_cond_destroy(&target->cond);
1744 pthread_mutex_destroy(&target->mutex);
1745 active_threads--;
1748 free(p);
1751 #else
1752 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1753 #endif
1755 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1757 unsigned char peeled[20];
1759 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1760 !peel_ref(path, peeled) && /* peelable? */
1761 !is_null_sha1(peeled) && /* annotated tag? */
1762 locate_object_entry(peeled)) /* object packed? */
1763 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1764 return 0;
1767 static void prepare_pack(int window, int depth)
1769 struct object_entry **delta_list;
1770 uint32_t i, nr_deltas;
1771 unsigned n;
1773 get_object_details();
1776 * If we're locally repacking then we need to be doubly careful
1777 * from now on in order to make sure no stealth corruption gets
1778 * propagated to the new pack. Clients receiving streamed packs
1779 * should validate everything they get anyway so no need to incur
1780 * the additional cost here in that case.
1782 if (!pack_to_stdout)
1783 do_check_packed_object_crc = 1;
1785 if (!nr_objects || !window || !depth)
1786 return;
1788 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1789 nr_deltas = n = 0;
1791 for (i = 0; i < nr_objects; i++) {
1792 struct object_entry *entry = objects + i;
1794 if (entry->delta)
1795 /* This happens if we decided to reuse existing
1796 * delta from a pack. "reuse_delta &&" is implied.
1798 continue;
1800 if (entry->size < 50)
1801 continue;
1803 if (entry->no_try_delta)
1804 continue;
1806 if (!entry->preferred_base) {
1807 nr_deltas++;
1808 if (entry->type < 0)
1809 die("unable to get type of object %s",
1810 sha1_to_hex(entry->idx.sha1));
1811 } else {
1812 if (entry->type < 0) {
1814 * This object is not found, but we
1815 * don't have to include it anyway.
1817 continue;
1821 delta_list[n++] = entry;
1824 if (nr_deltas && n > 1) {
1825 unsigned nr_done = 0;
1826 if (progress)
1827 progress_state = start_progress("Compressing objects",
1828 nr_deltas);
1829 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1830 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1831 stop_progress(&progress_state);
1832 if (nr_done != nr_deltas)
1833 die("inconsistency with delta count");
1835 free(delta_list);
1838 static int git_pack_config(const char *k, const char *v, void *cb)
1840 if (!strcmp(k, "pack.window")) {
1841 window = git_config_int(k, v);
1842 return 0;
1844 if (!strcmp(k, "pack.windowmemory")) {
1845 window_memory_limit = git_config_ulong(k, v);
1846 return 0;
1848 if (!strcmp(k, "pack.depth")) {
1849 depth = git_config_int(k, v);
1850 return 0;
1852 if (!strcmp(k, "pack.compression")) {
1853 int level = git_config_int(k, v);
1854 if (level == -1)
1855 level = Z_DEFAULT_COMPRESSION;
1856 else if (level < 0 || level > Z_BEST_COMPRESSION)
1857 die("bad pack compression level %d", level);
1858 pack_compression_level = level;
1859 pack_compression_seen = 1;
1860 return 0;
1862 if (!strcmp(k, "pack.deltacachesize")) {
1863 max_delta_cache_size = git_config_int(k, v);
1864 return 0;
1866 if (!strcmp(k, "pack.deltacachelimit")) {
1867 cache_max_small_delta_size = git_config_int(k, v);
1868 return 0;
1870 if (!strcmp(k, "pack.threads")) {
1871 delta_search_threads = git_config_int(k, v);
1872 if (delta_search_threads < 0)
1873 die("invalid number of threads specified (%d)",
1874 delta_search_threads);
1875 #ifndef THREADED_DELTA_SEARCH
1876 if (delta_search_threads != 1)
1877 warning("no threads support, ignoring %s", k);
1878 #endif
1879 return 0;
1881 if (!strcmp(k, "pack.indexversion")) {
1882 pack_idx_default_version = git_config_int(k, v);
1883 if (pack_idx_default_version > 2)
1884 die("bad pack.indexversion=%"PRIu32,
1885 pack_idx_default_version);
1886 return 0;
1888 if (!strcmp(k, "pack.packsizelimit")) {
1889 pack_size_limit_cfg = git_config_ulong(k, v);
1890 return 0;
1892 return git_default_config(k, v, cb);
1895 static void read_object_list_from_stdin(void)
1897 char line[40 + 1 + PATH_MAX + 2];
1898 unsigned char sha1[20];
1900 for (;;) {
1901 if (!fgets(line, sizeof(line), stdin)) {
1902 if (feof(stdin))
1903 break;
1904 if (!ferror(stdin))
1905 die("fgets returned NULL, not EOF, not error!");
1906 if (errno != EINTR)
1907 die_errno("fgets");
1908 clearerr(stdin);
1909 continue;
1911 if (line[0] == '-') {
1912 if (get_sha1_hex(line+1, sha1))
1913 die("expected edge sha1, got garbage:\n %s",
1914 line);
1915 add_preferred_base(sha1);
1916 continue;
1918 if (get_sha1_hex(line, sha1))
1919 die("expected sha1, got garbage:\n %s", line);
1921 add_preferred_base_object(line+41);
1922 add_object_entry(sha1, 0, line+41, 0);
1926 #define OBJECT_ADDED (1u<<20)
1928 static void show_commit(struct commit *commit, void *data)
1930 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1931 commit->object.flags |= OBJECT_ADDED;
1934 static void show_object(struct object *obj, const struct name_path *path, const char *last)
1936 char *name = path_name(path, last);
1938 add_preferred_base_object(name);
1939 add_object_entry(obj->sha1, obj->type, name, 0);
1940 obj->flags |= OBJECT_ADDED;
1943 * We will have generated the hash from the name,
1944 * but not saved a pointer to it - we can free it
1946 free((char *)name);
1949 static void show_edge(struct commit *commit)
1951 add_preferred_base(commit->object.sha1);
1954 struct in_pack_object {
1955 off_t offset;
1956 struct object *object;
1959 struct in_pack {
1960 int alloc;
1961 int nr;
1962 struct in_pack_object *array;
1965 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1967 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1968 in_pack->array[in_pack->nr].object = object;
1969 in_pack->nr++;
1973 * Compare the objects in the offset order, in order to emulate the
1974 * "git rev-list --objects" output that produced the pack originally.
1976 static int ofscmp(const void *a_, const void *b_)
1978 struct in_pack_object *a = (struct in_pack_object *)a_;
1979 struct in_pack_object *b = (struct in_pack_object *)b_;
1981 if (a->offset < b->offset)
1982 return -1;
1983 else if (a->offset > b->offset)
1984 return 1;
1985 else
1986 return hashcmp(a->object->sha1, b->object->sha1);
1989 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1991 struct packed_git *p;
1992 struct in_pack in_pack;
1993 uint32_t i;
1995 memset(&in_pack, 0, sizeof(in_pack));
1997 for (p = packed_git; p; p = p->next) {
1998 const unsigned char *sha1;
1999 struct object *o;
2001 if (!p->pack_local || p->pack_keep)
2002 continue;
2003 if (open_pack_index(p))
2004 die("cannot open pack index");
2006 ALLOC_GROW(in_pack.array,
2007 in_pack.nr + p->num_objects,
2008 in_pack.alloc);
2010 for (i = 0; i < p->num_objects; i++) {
2011 sha1 = nth_packed_object_sha1(p, i);
2012 o = lookup_unknown_object(sha1);
2013 if (!(o->flags & OBJECT_ADDED))
2014 mark_in_pack_object(o, p, &in_pack);
2015 o->flags |= OBJECT_ADDED;
2019 if (in_pack.nr) {
2020 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2021 ofscmp);
2022 for (i = 0; i < in_pack.nr; i++) {
2023 struct object *o = in_pack.array[i].object;
2024 add_object_entry(o->sha1, o->type, "", 0);
2027 free(in_pack.array);
2030 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2032 static struct packed_git *last_found = (void *)1;
2033 struct packed_git *p;
2035 p = (last_found != (void *)1) ? last_found : packed_git;
2037 while (p) {
2038 if ((!p->pack_local || p->pack_keep) &&
2039 find_pack_entry_one(sha1, p)) {
2040 last_found = p;
2041 return 1;
2043 if (p == last_found)
2044 p = packed_git;
2045 else
2046 p = p->next;
2047 if (p == last_found)
2048 p = p->next;
2050 return 0;
2053 static void loosen_unused_packed_objects(struct rev_info *revs)
2055 struct packed_git *p;
2056 uint32_t i;
2057 const unsigned char *sha1;
2059 for (p = packed_git; p; p = p->next) {
2060 if (!p->pack_local || p->pack_keep)
2061 continue;
2063 if (open_pack_index(p))
2064 die("cannot open pack index");
2066 for (i = 0; i < p->num_objects; i++) {
2067 sha1 = nth_packed_object_sha1(p, i);
2068 if (!locate_object_entry(sha1) &&
2069 !has_sha1_pack_kept_or_nonlocal(sha1))
2070 if (force_object_loose(sha1, p->mtime))
2071 die("unable to force loose object");
2076 static void get_object_list(int ac, const char **av)
2078 struct rev_info revs;
2079 char line[1000];
2080 int flags = 0;
2082 init_revisions(&revs, NULL);
2083 save_commit_buffer = 0;
2084 setup_revisions(ac, av, &revs, NULL);
2086 while (fgets(line, sizeof(line), stdin) != NULL) {
2087 int len = strlen(line);
2088 if (len && line[len - 1] == '\n')
2089 line[--len] = 0;
2090 if (!len)
2091 break;
2092 if (*line == '-') {
2093 if (!strcmp(line, "--not")) {
2094 flags ^= UNINTERESTING;
2095 continue;
2097 die("not a rev '%s'", line);
2099 if (handle_revision_arg(line, &revs, flags, 1))
2100 die("bad revision '%s'", line);
2103 if (prepare_revision_walk(&revs))
2104 die("revision walk setup failed");
2105 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2106 traverse_commit_list(&revs, show_commit, show_object, NULL);
2108 if (keep_unreachable)
2109 add_objects_in_unpacked_packs(&revs);
2110 if (unpack_unreachable)
2111 loosen_unused_packed_objects(&revs);
2114 static int adjust_perm(const char *path, mode_t mode)
2116 if (chmod(path, mode))
2117 return -1;
2118 return adjust_shared_perm(path);
2121 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2123 int use_internal_rev_list = 0;
2124 int thin = 0;
2125 uint32_t i;
2126 const char **rp_av;
2127 int rp_ac_alloc = 64;
2128 int rp_ac;
2130 read_replace_refs = 0;
2132 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2134 rp_av[0] = "pack-objects";
2135 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2136 rp_ac = 2;
2138 git_config(git_pack_config, NULL);
2139 if (!pack_compression_seen && core_compression_seen)
2140 pack_compression_level = core_compression_level;
2142 progress = isatty(2);
2143 for (i = 1; i < argc; i++) {
2144 const char *arg = argv[i];
2146 if (*arg != '-')
2147 break;
2149 if (!strcmp("--non-empty", arg)) {
2150 non_empty = 1;
2151 continue;
2153 if (!strcmp("--local", arg)) {
2154 local = 1;
2155 continue;
2157 if (!strcmp("--incremental", arg)) {
2158 incremental = 1;
2159 continue;
2161 if (!strcmp("--honor-pack-keep", arg)) {
2162 ignore_packed_keep = 1;
2163 continue;
2165 if (!prefixcmp(arg, "--compression=")) {
2166 char *end;
2167 int level = strtoul(arg+14, &end, 0);
2168 if (!arg[14] || *end)
2169 usage(pack_usage);
2170 if (level == -1)
2171 level = Z_DEFAULT_COMPRESSION;
2172 else if (level < 0 || level > Z_BEST_COMPRESSION)
2173 die("bad pack compression level %d", level);
2174 pack_compression_level = level;
2175 continue;
2177 if (!prefixcmp(arg, "--max-pack-size=")) {
2178 char *end;
2179 pack_size_limit_cfg = 0;
2180 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2181 if (!arg[16] || *end)
2182 usage(pack_usage);
2183 continue;
2185 if (!prefixcmp(arg, "--window=")) {
2186 char *end;
2187 window = strtoul(arg+9, &end, 0);
2188 if (!arg[9] || *end)
2189 usage(pack_usage);
2190 continue;
2192 if (!prefixcmp(arg, "--window-memory=")) {
2193 if (!git_parse_ulong(arg+16, &window_memory_limit))
2194 usage(pack_usage);
2195 continue;
2197 if (!prefixcmp(arg, "--threads=")) {
2198 char *end;
2199 delta_search_threads = strtoul(arg+10, &end, 0);
2200 if (!arg[10] || *end || delta_search_threads < 0)
2201 usage(pack_usage);
2202 #ifndef THREADED_DELTA_SEARCH
2203 if (delta_search_threads != 1)
2204 warning("no threads support, "
2205 "ignoring %s", arg);
2206 #endif
2207 continue;
2209 if (!prefixcmp(arg, "--depth=")) {
2210 char *end;
2211 depth = strtoul(arg+8, &end, 0);
2212 if (!arg[8] || *end)
2213 usage(pack_usage);
2214 continue;
2216 if (!strcmp("--progress", arg)) {
2217 progress = 1;
2218 continue;
2220 if (!strcmp("--all-progress", arg)) {
2221 progress = 2;
2222 continue;
2224 if (!strcmp("-q", arg)) {
2225 progress = 0;
2226 continue;
2228 if (!strcmp("--no-reuse-delta", arg)) {
2229 reuse_delta = 0;
2230 continue;
2232 if (!strcmp("--no-reuse-object", arg)) {
2233 reuse_object = reuse_delta = 0;
2234 continue;
2236 if (!strcmp("--delta-base-offset", arg)) {
2237 allow_ofs_delta = 1;
2238 continue;
2240 if (!strcmp("--stdout", arg)) {
2241 pack_to_stdout = 1;
2242 continue;
2244 if (!strcmp("--revs", arg)) {
2245 use_internal_rev_list = 1;
2246 continue;
2248 if (!strcmp("--keep-unreachable", arg)) {
2249 keep_unreachable = 1;
2250 continue;
2252 if (!strcmp("--unpack-unreachable", arg)) {
2253 unpack_unreachable = 1;
2254 continue;
2256 if (!strcmp("--include-tag", arg)) {
2257 include_tag = 1;
2258 continue;
2260 if (!strcmp("--unpacked", arg) ||
2261 !strcmp("--reflog", arg) ||
2262 !strcmp("--all", arg)) {
2263 use_internal_rev_list = 1;
2264 if (rp_ac >= rp_ac_alloc - 1) {
2265 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2266 rp_av = xrealloc(rp_av,
2267 rp_ac_alloc * sizeof(*rp_av));
2269 rp_av[rp_ac++] = arg;
2270 continue;
2272 if (!strcmp("--thin", arg)) {
2273 use_internal_rev_list = 1;
2274 thin = 1;
2275 rp_av[1] = "--objects-edge";
2276 continue;
2278 if (!prefixcmp(arg, "--index-version=")) {
2279 char *c;
2280 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2281 if (pack_idx_default_version > 2)
2282 die("bad %s", arg);
2283 if (*c == ',')
2284 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2285 if (*c || pack_idx_off32_limit & 0x80000000)
2286 die("bad %s", arg);
2287 continue;
2289 if (!strcmp(arg, "--keep-true-parents")) {
2290 grafts_replace_parents = 0;
2291 continue;
2293 usage(pack_usage);
2296 /* Traditionally "pack-objects [options] base extra" failed;
2297 * we would however want to take refs parameter that would
2298 * have been given to upstream rev-list ourselves, which means
2299 * we somehow want to say what the base name is. So the
2300 * syntax would be:
2302 * pack-objects [options] base <refs...>
2304 * in other words, we would treat the first non-option as the
2305 * base_name and send everything else to the internal revision
2306 * walker.
2309 if (!pack_to_stdout)
2310 base_name = argv[i++];
2312 if (pack_to_stdout != !base_name)
2313 usage(pack_usage);
2315 if (!pack_to_stdout && !pack_size_limit)
2316 pack_size_limit = pack_size_limit_cfg;
2318 if (pack_to_stdout && pack_size_limit)
2319 die("--max-pack-size cannot be used to build a pack for transfer.");
2321 if (!pack_to_stdout && thin)
2322 die("--thin cannot be used to build an indexable pack.");
2324 if (keep_unreachable && unpack_unreachable)
2325 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2327 #ifdef THREADED_DELTA_SEARCH
2328 if (!delta_search_threads) /* --threads=0 means autodetect */
2329 delta_search_threads = online_cpus();
2330 #endif
2332 prepare_packed_git();
2334 if (progress)
2335 progress_state = start_progress("Counting objects", 0);
2336 if (!use_internal_rev_list)
2337 read_object_list_from_stdin();
2338 else {
2339 rp_av[rp_ac] = NULL;
2340 get_object_list(rp_ac, rp_av);
2342 cleanup_preferred_base();
2343 if (include_tag && nr_result)
2344 for_each_ref(add_ref_tag, NULL);
2345 stop_progress(&progress_state);
2347 if (non_empty && !nr_result)
2348 return 0;
2349 if (nr_result)
2350 prepare_pack(window, depth);
2351 write_pack_file();
2352 if (progress)
2353 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2354 " reused %"PRIu32" (delta %"PRIu32")\n",
2355 written, written_delta, reused, reused_delta);
2356 return 0;