builtin-commit: add --date option
[git/vmiklos.git] / builtin-pack-objects.c
blob4429d53a1e82b81f0a82e34bf2d6ae463aeeadee
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 " [--all-progress-implied]\n"
28 " [--max-pack-size=N] [--local] [--incremental]\n"
29 " [--window=N] [--window-memory=N] [--depth=N]\n"
30 " [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset]\n"
31 " [--threads=N] [--non-empty] [--revs [--unpacked | --all]*]\n"
32 " [--reflog] [--stdout | base-name] [--include-tag]\n"
33 " [--keep-unreachable | --unpack-unreachable \n"
34 " [<ref-list | <object-list]";
36 struct object_entry {
37 struct pack_idx_entry idx;
38 unsigned long size; /* uncompressed size */
39 struct packed_git *in_pack; /* already in pack */
40 off_t in_pack_offset;
41 struct object_entry *delta; /* delta base object */
42 struct object_entry *delta_child; /* deltified objects who bases me */
43 struct object_entry *delta_sibling; /* other deltified objects who
44 * uses the same base as me
46 void *delta_data; /* cached delta (uncompressed) */
47 unsigned long delta_size; /* delta data size (uncompressed) */
48 unsigned long z_delta_size; /* delta data size (compressed) */
49 unsigned int hash; /* name hint hash */
50 enum object_type type;
51 enum object_type in_pack_type; /* could be delta */
52 unsigned char in_pack_header_size;
53 unsigned char preferred_base; /* we do not pack this, but is available
54 * to be used as the base object to delta
55 * objects against.
57 unsigned char no_try_delta;
61 * Objects we are going to pack are collected in objects array (dynamically
62 * expanded). nr_objects & nr_alloc controls this array. They are stored
63 * in the order we see -- typically rev-list --objects order that gives us
64 * nice "minimum seek" order.
66 static struct object_entry *objects;
67 static struct pack_idx_entry **written_list;
68 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
70 static int non_empty;
71 static int reuse_delta = 1, reuse_object = 1;
72 static int keep_unreachable, unpack_unreachable, include_tag;
73 static int local;
74 static int incremental;
75 static int ignore_packed_keep;
76 static int allow_ofs_delta;
77 static const char *base_name;
78 static int progress = 1;
79 static int window = 10;
80 static uint32_t pack_size_limit, pack_size_limit_cfg;
81 static int depth = 50;
82 static int delta_search_threads;
83 static int pack_to_stdout;
84 static int num_preferred_base;
85 static struct progress *progress_state;
86 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
87 static int pack_compression_seen;
89 static unsigned long delta_cache_size = 0;
90 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
91 static unsigned long cache_max_small_delta_size = 1000;
93 static unsigned long window_memory_limit = 0;
96 * The object names in objects array are hashed with this hashtable,
97 * to help looking up the entry by object name.
98 * This hashtable is built after all the objects are seen.
100 static int *object_ix;
101 static int object_ix_hashsz;
104 * stats
106 static uint32_t written, written_delta;
107 static uint32_t reused, reused_delta;
110 static void *get_delta(struct object_entry *entry)
112 unsigned long size, base_size, delta_size;
113 void *buf, *base_buf, *delta_buf;
114 enum object_type type;
116 buf = read_sha1_file(entry->idx.sha1, &type, &size);
117 if (!buf)
118 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
119 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
120 if (!base_buf)
121 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
122 delta_buf = diff_delta(base_buf, base_size,
123 buf, size, &delta_size, 0);
124 if (!delta_buf || delta_size != entry->delta_size)
125 die("delta size changed");
126 free(buf);
127 free(base_buf);
128 return delta_buf;
131 static unsigned long do_compress(void **pptr, unsigned long size)
133 z_stream stream;
134 void *in, *out;
135 unsigned long maxsize;
137 memset(&stream, 0, sizeof(stream));
138 deflateInit(&stream, pack_compression_level);
139 maxsize = deflateBound(&stream, size);
141 in = *pptr;
142 out = xmalloc(maxsize);
143 *pptr = out;
145 stream.next_in = in;
146 stream.avail_in = size;
147 stream.next_out = out;
148 stream.avail_out = maxsize;
149 while (deflate(&stream, Z_FINISH) == Z_OK)
150 ; /* nothing */
151 deflateEnd(&stream);
153 free(in);
154 return stream.total_out;
158 * The per-object header is a pretty dense thing, which is
159 * - first byte: low four bits are "size", then three bits of "type",
160 * and the high bit is "size continues".
161 * - each byte afterwards: low seven bits are size continuation,
162 * with the high bit being "size continues"
164 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
166 int n = 1;
167 unsigned char c;
169 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
170 die("bad type %d", type);
172 c = (type << 4) | (size & 15);
173 size >>= 4;
174 while (size) {
175 *hdr++ = c | 0x80;
176 c = size & 0x7f;
177 size >>= 7;
178 n++;
180 *hdr = c;
181 return n;
185 * we are going to reuse the existing object data as is. make
186 * sure it is not corrupt.
188 static int check_pack_inflate(struct packed_git *p,
189 struct pack_window **w_curs,
190 off_t offset,
191 off_t len,
192 unsigned long expect)
194 z_stream stream;
195 unsigned char fakebuf[4096], *in;
196 int st;
198 memset(&stream, 0, sizeof(stream));
199 git_inflate_init(&stream);
200 do {
201 in = use_pack(p, w_curs, offset, &stream.avail_in);
202 stream.next_in = in;
203 stream.next_out = fakebuf;
204 stream.avail_out = sizeof(fakebuf);
205 st = git_inflate(&stream, Z_FINISH);
206 offset += stream.next_in - in;
207 } while (st == Z_OK || st == Z_BUF_ERROR);
208 git_inflate_end(&stream);
209 return (st == Z_STREAM_END &&
210 stream.total_out == expect &&
211 stream.total_in == len) ? 0 : -1;
214 static void copy_pack_data(struct sha1file *f,
215 struct packed_git *p,
216 struct pack_window **w_curs,
217 off_t offset,
218 off_t len)
220 unsigned char *in;
221 unsigned int avail;
223 while (len) {
224 in = use_pack(p, w_curs, offset, &avail);
225 if (avail > len)
226 avail = (unsigned int)len;
227 sha1write(f, in, avail);
228 offset += avail;
229 len -= avail;
233 static unsigned long write_object(struct sha1file *f,
234 struct object_entry *entry,
235 off_t write_offset)
237 unsigned long size, limit, datalen;
238 void *buf;
239 unsigned char header[10], dheader[10];
240 unsigned hdrlen;
241 enum object_type type;
242 int usable_delta, to_reuse;
244 if (!pack_to_stdout)
245 crc32_begin(f);
247 type = entry->type;
249 /* write limit if limited packsize and not first object */
250 if (!pack_size_limit || !nr_written)
251 limit = 0;
252 else if (pack_size_limit <= write_offset)
254 * the earlier object did not fit the limit; avoid
255 * mistaking this with unlimited (i.e. limit = 0).
257 limit = 1;
258 else
259 limit = pack_size_limit - write_offset;
261 if (!entry->delta)
262 usable_delta = 0; /* no delta */
263 else if (!pack_size_limit)
264 usable_delta = 1; /* unlimited packfile */
265 else if (entry->delta->idx.offset == (off_t)-1)
266 usable_delta = 0; /* base was written to another pack */
267 else if (entry->delta->idx.offset)
268 usable_delta = 1; /* base already exists in this pack */
269 else
270 usable_delta = 0; /* base could end up in another pack */
272 if (!reuse_object)
273 to_reuse = 0; /* explicit */
274 else if (!entry->in_pack)
275 to_reuse = 0; /* can't reuse what we don't have */
276 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
277 /* check_object() decided it for us ... */
278 to_reuse = usable_delta;
279 /* ... but pack split may override that */
280 else if (type != entry->in_pack_type)
281 to_reuse = 0; /* pack has delta which is unusable */
282 else if (entry->delta)
283 to_reuse = 0; /* we want to pack afresh */
284 else
285 to_reuse = 1; /* we have it in-pack undeltified,
286 * and we do not need to deltify it.
289 if (!to_reuse) {
290 no_reuse:
291 if (!usable_delta) {
292 buf = read_sha1_file(entry->idx.sha1, &type, &size);
293 if (!buf)
294 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
296 * make sure no cached delta data remains from a
297 * previous attempt before a pack split occurred.
299 free(entry->delta_data);
300 entry->delta_data = NULL;
301 entry->z_delta_size = 0;
302 } else if (entry->delta_data) {
303 size = entry->delta_size;
304 buf = entry->delta_data;
305 entry->delta_data = NULL;
306 type = (allow_ofs_delta && entry->delta->idx.offset) ?
307 OBJ_OFS_DELTA : OBJ_REF_DELTA;
308 } else {
309 buf = get_delta(entry);
310 size = entry->delta_size;
311 type = (allow_ofs_delta && entry->delta->idx.offset) ?
312 OBJ_OFS_DELTA : OBJ_REF_DELTA;
315 if (entry->z_delta_size)
316 datalen = entry->z_delta_size;
317 else
318 datalen = do_compress(&buf, size);
321 * The object header is a byte of 'type' followed by zero or
322 * more bytes of length.
324 hdrlen = encode_header(type, size, header);
326 if (type == OBJ_OFS_DELTA) {
328 * Deltas with relative base contain an additional
329 * encoding of the relative offset for the delta
330 * base from this object's position in the pack.
332 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
333 unsigned pos = sizeof(dheader) - 1;
334 dheader[pos] = ofs & 127;
335 while (ofs >>= 7)
336 dheader[--pos] = 128 | (--ofs & 127);
337 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
338 free(buf);
339 return 0;
341 sha1write(f, header, hdrlen);
342 sha1write(f, dheader + pos, sizeof(dheader) - pos);
343 hdrlen += sizeof(dheader) - pos;
344 } else if (type == OBJ_REF_DELTA) {
346 * Deltas with a base reference contain
347 * an additional 20 bytes for the base sha1.
349 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
350 free(buf);
351 return 0;
353 sha1write(f, header, hdrlen);
354 sha1write(f, entry->delta->idx.sha1, 20);
355 hdrlen += 20;
356 } else {
357 if (limit && hdrlen + datalen + 20 >= limit) {
358 free(buf);
359 return 0;
361 sha1write(f, header, hdrlen);
363 sha1write(f, buf, datalen);
364 free(buf);
366 else {
367 struct packed_git *p = entry->in_pack;
368 struct pack_window *w_curs = NULL;
369 struct revindex_entry *revidx;
370 off_t offset;
372 if (entry->delta)
373 type = (allow_ofs_delta && entry->delta->idx.offset) ?
374 OBJ_OFS_DELTA : OBJ_REF_DELTA;
375 hdrlen = encode_header(type, entry->size, header);
377 offset = entry->in_pack_offset;
378 revidx = find_pack_revindex(p, offset);
379 datalen = revidx[1].offset - offset;
380 if (!pack_to_stdout && p->index_version > 1 &&
381 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
382 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
383 unuse_pack(&w_curs);
384 goto no_reuse;
387 offset += entry->in_pack_header_size;
388 datalen -= entry->in_pack_header_size;
389 if (!pack_to_stdout && p->index_version == 1 &&
390 check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
391 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
392 unuse_pack(&w_curs);
393 goto no_reuse;
396 if (type == OBJ_OFS_DELTA) {
397 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
398 unsigned pos = sizeof(dheader) - 1;
399 dheader[pos] = ofs & 127;
400 while (ofs >>= 7)
401 dheader[--pos] = 128 | (--ofs & 127);
402 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
403 unuse_pack(&w_curs);
404 return 0;
406 sha1write(f, header, hdrlen);
407 sha1write(f, dheader + pos, sizeof(dheader) - pos);
408 hdrlen += sizeof(dheader) - pos;
409 reused_delta++;
410 } else if (type == OBJ_REF_DELTA) {
411 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
412 unuse_pack(&w_curs);
413 return 0;
415 sha1write(f, header, hdrlen);
416 sha1write(f, entry->delta->idx.sha1, 20);
417 hdrlen += 20;
418 reused_delta++;
419 } else {
420 if (limit && hdrlen + datalen + 20 >= limit) {
421 unuse_pack(&w_curs);
422 return 0;
424 sha1write(f, header, hdrlen);
426 copy_pack_data(f, p, &w_curs, offset, datalen);
427 unuse_pack(&w_curs);
428 reused++;
430 if (usable_delta)
431 written_delta++;
432 written++;
433 if (!pack_to_stdout)
434 entry->idx.crc32 = crc32_end(f);
435 return hdrlen + datalen;
438 static int write_one(struct sha1file *f,
439 struct object_entry *e,
440 off_t *offset)
442 unsigned long size;
444 /* offset is non zero if object is written already. */
445 if (e->idx.offset || e->preferred_base)
446 return 1;
448 /* if we are deltified, write out base object first. */
449 if (e->delta && !write_one(f, e->delta, offset))
450 return 0;
452 e->idx.offset = *offset;
453 size = write_object(f, e, *offset);
454 if (!size) {
455 e->idx.offset = 0;
456 return 0;
458 written_list[nr_written++] = &e->idx;
460 /* make sure off_t is sufficiently large not to wrap */
461 if (*offset > *offset + size)
462 die("pack too large for current definition of off_t");
463 *offset += size;
464 return 1;
467 /* forward declaration for write_pack_file */
468 static int adjust_perm(const char *path, mode_t mode);
470 static void write_pack_file(void)
472 uint32_t i = 0, j;
473 struct sha1file *f;
474 off_t offset;
475 struct pack_header hdr;
476 uint32_t nr_remaining = nr_result;
477 time_t last_mtime = 0;
479 if (progress > pack_to_stdout)
480 progress_state = start_progress("Writing objects", nr_result);
481 written_list = xmalloc(nr_objects * sizeof(*written_list));
483 do {
484 unsigned char sha1[20];
485 char *pack_tmp_name = NULL;
487 if (pack_to_stdout) {
488 f = sha1fd_throughput(1, "<stdout>", progress_state);
489 } else {
490 char tmpname[PATH_MAX];
491 int fd;
492 fd = odb_mkstemp(tmpname, sizeof(tmpname),
493 "pack/tmp_pack_XXXXXX");
494 pack_tmp_name = xstrdup(tmpname);
495 f = sha1fd(fd, pack_tmp_name);
498 hdr.hdr_signature = htonl(PACK_SIGNATURE);
499 hdr.hdr_version = htonl(PACK_VERSION);
500 hdr.hdr_entries = htonl(nr_remaining);
501 sha1write(f, &hdr, sizeof(hdr));
502 offset = sizeof(hdr);
503 nr_written = 0;
504 for (; i < nr_objects; i++) {
505 if (!write_one(f, objects + i, &offset))
506 break;
507 display_progress(progress_state, written);
511 * Did we write the wrong # entries in the header?
512 * If so, rewrite it like in fast-import
514 if (pack_to_stdout) {
515 sha1close(f, sha1, CSUM_CLOSE);
516 } else if (nr_written == nr_remaining) {
517 sha1close(f, sha1, CSUM_FSYNC);
518 } else {
519 int fd = sha1close(f, sha1, 0);
520 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
521 nr_written, sha1, offset);
522 close(fd);
525 if (!pack_to_stdout) {
526 mode_t mode = umask(0);
527 struct stat st;
528 char *idx_tmp_name, tmpname[PATH_MAX];
530 umask(mode);
531 mode = 0444 & ~mode;
533 idx_tmp_name = write_idx_file(NULL, written_list,
534 nr_written, sha1);
536 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
537 base_name, sha1_to_hex(sha1));
538 free_pack_by_name(tmpname);
539 if (adjust_perm(pack_tmp_name, mode))
540 die_errno("unable to make temporary pack file readable");
541 if (rename(pack_tmp_name, tmpname))
542 die_errno("unable to rename temporary pack file");
545 * Packs are runtime accessed in their mtime
546 * order since newer packs are more likely to contain
547 * younger objects. So if we are creating multiple
548 * packs then we should modify the mtime of later ones
549 * to preserve this property.
551 if (stat(tmpname, &st) < 0) {
552 warning("failed to stat %s: %s",
553 tmpname, strerror(errno));
554 } else if (!last_mtime) {
555 last_mtime = st.st_mtime;
556 } else {
557 struct utimbuf utb;
558 utb.actime = st.st_atime;
559 utb.modtime = --last_mtime;
560 if (utime(tmpname, &utb) < 0)
561 warning("failed utime() on %s: %s",
562 tmpname, strerror(errno));
565 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
566 base_name, sha1_to_hex(sha1));
567 if (adjust_perm(idx_tmp_name, mode))
568 die_errno("unable to make temporary index file readable");
569 if (rename(idx_tmp_name, tmpname))
570 die_errno("unable to rename temporary index file");
572 free(idx_tmp_name);
573 free(pack_tmp_name);
574 puts(sha1_to_hex(sha1));
577 /* mark written objects as written to previous pack */
578 for (j = 0; j < nr_written; j++) {
579 written_list[j]->offset = (off_t)-1;
581 nr_remaining -= nr_written;
582 } while (nr_remaining && i < nr_objects);
584 free(written_list);
585 stop_progress(&progress_state);
586 if (written != nr_result)
587 die("wrote %"PRIu32" objects while expecting %"PRIu32,
588 written, nr_result);
590 * We have scanned through [0 ... i). Since we have written
591 * the correct number of objects, the remaining [i ... nr_objects)
592 * items must be either already written (due to out-of-order delta base)
593 * or a preferred base. Count those which are neither and complain if any.
595 for (j = 0; i < nr_objects; i++) {
596 struct object_entry *e = objects + i;
597 j += !e->idx.offset && !e->preferred_base;
599 if (j)
600 die("wrote %"PRIu32" objects as expected but %"PRIu32
601 " unwritten", written, j);
604 static int locate_object_entry_hash(const unsigned char *sha1)
606 int i;
607 unsigned int ui;
608 memcpy(&ui, sha1, sizeof(unsigned int));
609 i = ui % object_ix_hashsz;
610 while (0 < object_ix[i]) {
611 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
612 return i;
613 if (++i == object_ix_hashsz)
614 i = 0;
616 return -1 - i;
619 static struct object_entry *locate_object_entry(const unsigned char *sha1)
621 int i;
623 if (!object_ix_hashsz)
624 return NULL;
626 i = locate_object_entry_hash(sha1);
627 if (0 <= i)
628 return &objects[object_ix[i]-1];
629 return NULL;
632 static void rehash_objects(void)
634 uint32_t i;
635 struct object_entry *oe;
637 object_ix_hashsz = nr_objects * 3;
638 if (object_ix_hashsz < 1024)
639 object_ix_hashsz = 1024;
640 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
641 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
642 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
643 int ix = locate_object_entry_hash(oe->idx.sha1);
644 if (0 <= ix)
645 continue;
646 ix = -1 - ix;
647 object_ix[ix] = i + 1;
651 static unsigned name_hash(const char *name)
653 unsigned c, hash = 0;
655 if (!name)
656 return 0;
659 * This effectively just creates a sortable number from the
660 * last sixteen non-whitespace characters. Last characters
661 * count "most", so things that end in ".c" sort together.
663 while ((c = *name++) != 0) {
664 if (isspace(c))
665 continue;
666 hash = (hash >> 2) + (c << 24);
668 return hash;
671 static void setup_delta_attr_check(struct git_attr_check *check)
673 static struct git_attr *attr_delta;
675 if (!attr_delta)
676 attr_delta = git_attr("delta", 5);
678 check[0].attr = attr_delta;
681 static int no_try_delta(const char *path)
683 struct git_attr_check check[1];
685 setup_delta_attr_check(check);
686 if (git_checkattr(path, ARRAY_SIZE(check), check))
687 return 0;
688 if (ATTR_FALSE(check->value))
689 return 1;
690 return 0;
693 static int add_object_entry(const unsigned char *sha1, enum object_type type,
694 const char *name, int exclude)
696 struct object_entry *entry;
697 struct packed_git *p, *found_pack = NULL;
698 off_t found_offset = 0;
699 int ix;
700 unsigned hash = name_hash(name);
702 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
703 if (ix >= 0) {
704 if (exclude) {
705 entry = objects + object_ix[ix] - 1;
706 if (!entry->preferred_base)
707 nr_result--;
708 entry->preferred_base = 1;
710 return 0;
713 if (!exclude && local && has_loose_object_nonlocal(sha1))
714 return 0;
716 for (p = packed_git; p; p = p->next) {
717 off_t offset = find_pack_entry_one(sha1, p);
718 if (offset) {
719 if (!found_pack) {
720 found_offset = offset;
721 found_pack = p;
723 if (exclude)
724 break;
725 if (incremental)
726 return 0;
727 if (local && !p->pack_local)
728 return 0;
729 if (ignore_packed_keep && p->pack_local && p->pack_keep)
730 return 0;
734 if (nr_objects >= nr_alloc) {
735 nr_alloc = (nr_alloc + 1024) * 3 / 2;
736 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
739 entry = objects + nr_objects++;
740 memset(entry, 0, sizeof(*entry));
741 hashcpy(entry->idx.sha1, sha1);
742 entry->hash = hash;
743 if (type)
744 entry->type = type;
745 if (exclude)
746 entry->preferred_base = 1;
747 else
748 nr_result++;
749 if (found_pack) {
750 entry->in_pack = found_pack;
751 entry->in_pack_offset = found_offset;
754 if (object_ix_hashsz * 3 <= nr_objects * 4)
755 rehash_objects();
756 else
757 object_ix[-1 - ix] = nr_objects;
759 display_progress(progress_state, nr_objects);
761 if (name && no_try_delta(name))
762 entry->no_try_delta = 1;
764 return 1;
767 struct pbase_tree_cache {
768 unsigned char sha1[20];
769 int ref;
770 int temporary;
771 void *tree_data;
772 unsigned long tree_size;
775 static struct pbase_tree_cache *(pbase_tree_cache[256]);
776 static int pbase_tree_cache_ix(const unsigned char *sha1)
778 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
780 static int pbase_tree_cache_ix_incr(int ix)
782 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
785 static struct pbase_tree {
786 struct pbase_tree *next;
787 /* This is a phony "cache" entry; we are not
788 * going to evict it nor find it through _get()
789 * mechanism -- this is for the toplevel node that
790 * would almost always change with any commit.
792 struct pbase_tree_cache pcache;
793 } *pbase_tree;
795 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
797 struct pbase_tree_cache *ent, *nent;
798 void *data;
799 unsigned long size;
800 enum object_type type;
801 int neigh;
802 int my_ix = pbase_tree_cache_ix(sha1);
803 int available_ix = -1;
805 /* pbase-tree-cache acts as a limited hashtable.
806 * your object will be found at your index or within a few
807 * slots after that slot if it is cached.
809 for (neigh = 0; neigh < 8; neigh++) {
810 ent = pbase_tree_cache[my_ix];
811 if (ent && !hashcmp(ent->sha1, sha1)) {
812 ent->ref++;
813 return ent;
815 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
816 ((0 <= available_ix) &&
817 (!ent && pbase_tree_cache[available_ix])))
818 available_ix = my_ix;
819 if (!ent)
820 break;
821 my_ix = pbase_tree_cache_ix_incr(my_ix);
824 /* Did not find one. Either we got a bogus request or
825 * we need to read and perhaps cache.
827 data = read_sha1_file(sha1, &type, &size);
828 if (!data)
829 return NULL;
830 if (type != OBJ_TREE) {
831 free(data);
832 return NULL;
835 /* We need to either cache or return a throwaway copy */
837 if (available_ix < 0)
838 ent = NULL;
839 else {
840 ent = pbase_tree_cache[available_ix];
841 my_ix = available_ix;
844 if (!ent) {
845 nent = xmalloc(sizeof(*nent));
846 nent->temporary = (available_ix < 0);
848 else {
849 /* evict and reuse */
850 free(ent->tree_data);
851 nent = ent;
853 hashcpy(nent->sha1, sha1);
854 nent->tree_data = data;
855 nent->tree_size = size;
856 nent->ref = 1;
857 if (!nent->temporary)
858 pbase_tree_cache[my_ix] = nent;
859 return nent;
862 static void pbase_tree_put(struct pbase_tree_cache *cache)
864 if (!cache->temporary) {
865 cache->ref--;
866 return;
868 free(cache->tree_data);
869 free(cache);
872 static int name_cmp_len(const char *name)
874 int i;
875 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
877 return i;
880 static void add_pbase_object(struct tree_desc *tree,
881 const char *name,
882 int cmplen,
883 const char *fullname)
885 struct name_entry entry;
886 int cmp;
888 while (tree_entry(tree,&entry)) {
889 if (S_ISGITLINK(entry.mode))
890 continue;
891 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
892 memcmp(name, entry.path, cmplen);
893 if (cmp > 0)
894 continue;
895 if (cmp < 0)
896 return;
897 if (name[cmplen] != '/') {
898 add_object_entry(entry.sha1,
899 object_type(entry.mode),
900 fullname, 1);
901 return;
903 if (S_ISDIR(entry.mode)) {
904 struct tree_desc sub;
905 struct pbase_tree_cache *tree;
906 const char *down = name+cmplen+1;
907 int downlen = name_cmp_len(down);
909 tree = pbase_tree_get(entry.sha1);
910 if (!tree)
911 return;
912 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
914 add_pbase_object(&sub, down, downlen, fullname);
915 pbase_tree_put(tree);
920 static unsigned *done_pbase_paths;
921 static int done_pbase_paths_num;
922 static int done_pbase_paths_alloc;
923 static int done_pbase_path_pos(unsigned hash)
925 int lo = 0;
926 int hi = done_pbase_paths_num;
927 while (lo < hi) {
928 int mi = (hi + lo) / 2;
929 if (done_pbase_paths[mi] == hash)
930 return mi;
931 if (done_pbase_paths[mi] < hash)
932 hi = mi;
933 else
934 lo = mi + 1;
936 return -lo-1;
939 static int check_pbase_path(unsigned hash)
941 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
942 if (0 <= pos)
943 return 1;
944 pos = -pos - 1;
945 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
946 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
947 done_pbase_paths = xrealloc(done_pbase_paths,
948 done_pbase_paths_alloc *
949 sizeof(unsigned));
951 done_pbase_paths_num++;
952 if (pos < done_pbase_paths_num)
953 memmove(done_pbase_paths + pos + 1,
954 done_pbase_paths + pos,
955 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
956 done_pbase_paths[pos] = hash;
957 return 0;
960 static void add_preferred_base_object(const char *name)
962 struct pbase_tree *it;
963 int cmplen;
964 unsigned hash = name_hash(name);
966 if (!num_preferred_base || check_pbase_path(hash))
967 return;
969 cmplen = name_cmp_len(name);
970 for (it = pbase_tree; it; it = it->next) {
971 if (cmplen == 0) {
972 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
974 else {
975 struct tree_desc tree;
976 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
977 add_pbase_object(&tree, name, cmplen, name);
982 static void add_preferred_base(unsigned char *sha1)
984 struct pbase_tree *it;
985 void *data;
986 unsigned long size;
987 unsigned char tree_sha1[20];
989 if (window <= num_preferred_base++)
990 return;
992 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
993 if (!data)
994 return;
996 for (it = pbase_tree; it; it = it->next) {
997 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
998 free(data);
999 return;
1003 it = xcalloc(1, sizeof(*it));
1004 it->next = pbase_tree;
1005 pbase_tree = it;
1007 hashcpy(it->pcache.sha1, tree_sha1);
1008 it->pcache.tree_data = data;
1009 it->pcache.tree_size = size;
1012 static void cleanup_preferred_base(void)
1014 struct pbase_tree *it;
1015 unsigned i;
1017 it = pbase_tree;
1018 pbase_tree = NULL;
1019 while (it) {
1020 struct pbase_tree *this = it;
1021 it = this->next;
1022 free(this->pcache.tree_data);
1023 free(this);
1026 for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1027 if (!pbase_tree_cache[i])
1028 continue;
1029 free(pbase_tree_cache[i]->tree_data);
1030 free(pbase_tree_cache[i]);
1031 pbase_tree_cache[i] = NULL;
1034 free(done_pbase_paths);
1035 done_pbase_paths = NULL;
1036 done_pbase_paths_num = done_pbase_paths_alloc = 0;
1039 static void check_object(struct object_entry *entry)
1041 if (entry->in_pack) {
1042 struct packed_git *p = entry->in_pack;
1043 struct pack_window *w_curs = NULL;
1044 const unsigned char *base_ref = NULL;
1045 struct object_entry *base_entry;
1046 unsigned long used, used_0;
1047 unsigned int avail;
1048 off_t ofs;
1049 unsigned char *buf, c;
1051 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1054 * We want in_pack_type even if we do not reuse delta
1055 * since non-delta representations could still be reused.
1057 used = unpack_object_header_buffer(buf, avail,
1058 &entry->in_pack_type,
1059 &entry->size);
1060 if (used == 0)
1061 goto give_up;
1064 * Determine if this is a delta and if so whether we can
1065 * reuse it or not. Otherwise let's find out as cheaply as
1066 * possible what the actual type and size for this object is.
1068 switch (entry->in_pack_type) {
1069 default:
1070 /* Not a delta hence we've already got all we need. */
1071 entry->type = entry->in_pack_type;
1072 entry->in_pack_header_size = used;
1073 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1074 goto give_up;
1075 unuse_pack(&w_curs);
1076 return;
1077 case OBJ_REF_DELTA:
1078 if (reuse_delta && !entry->preferred_base)
1079 base_ref = use_pack(p, &w_curs,
1080 entry->in_pack_offset + used, NULL);
1081 entry->in_pack_header_size = used + 20;
1082 break;
1083 case OBJ_OFS_DELTA:
1084 buf = use_pack(p, &w_curs,
1085 entry->in_pack_offset + used, NULL);
1086 used_0 = 0;
1087 c = buf[used_0++];
1088 ofs = c & 127;
1089 while (c & 128) {
1090 ofs += 1;
1091 if (!ofs || MSB(ofs, 7)) {
1092 error("delta base offset overflow in pack for %s",
1093 sha1_to_hex(entry->idx.sha1));
1094 goto give_up;
1096 c = buf[used_0++];
1097 ofs = (ofs << 7) + (c & 127);
1099 ofs = entry->in_pack_offset - ofs;
1100 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1101 error("delta base offset out of bound for %s",
1102 sha1_to_hex(entry->idx.sha1));
1103 goto give_up;
1105 if (reuse_delta && !entry->preferred_base) {
1106 struct revindex_entry *revidx;
1107 revidx = find_pack_revindex(p, ofs);
1108 if (!revidx)
1109 goto give_up;
1110 base_ref = nth_packed_object_sha1(p, revidx->nr);
1112 entry->in_pack_header_size = used + used_0;
1113 break;
1116 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1118 * If base_ref was set above that means we wish to
1119 * reuse delta data, and we even found that base
1120 * in the list of objects we want to pack. Goodie!
1122 * Depth value does not matter - find_deltas() will
1123 * never consider reused delta as the base object to
1124 * deltify other objects against, in order to avoid
1125 * circular deltas.
1127 entry->type = entry->in_pack_type;
1128 entry->delta = base_entry;
1129 entry->delta_size = entry->size;
1130 entry->delta_sibling = base_entry->delta_child;
1131 base_entry->delta_child = entry;
1132 unuse_pack(&w_curs);
1133 return;
1136 if (entry->type) {
1138 * This must be a delta and we already know what the
1139 * final object type is. Let's extract the actual
1140 * object size from the delta header.
1142 entry->size = get_size_from_delta(p, &w_curs,
1143 entry->in_pack_offset + entry->in_pack_header_size);
1144 if (entry->size == 0)
1145 goto give_up;
1146 unuse_pack(&w_curs);
1147 return;
1151 * No choice but to fall back to the recursive delta walk
1152 * with sha1_object_info() to find about the object type
1153 * at this point...
1155 give_up:
1156 unuse_pack(&w_curs);
1159 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1161 * The error condition is checked in prepare_pack(). This is
1162 * to permit a missing preferred base object to be ignored
1163 * as a preferred base. Doing so can result in a larger
1164 * pack file, but the transfer will still take place.
1168 static int pack_offset_sort(const void *_a, const void *_b)
1170 const struct object_entry *a = *(struct object_entry **)_a;
1171 const struct object_entry *b = *(struct object_entry **)_b;
1173 /* avoid filesystem trashing with loose objects */
1174 if (!a->in_pack && !b->in_pack)
1175 return hashcmp(a->idx.sha1, b->idx.sha1);
1177 if (a->in_pack < b->in_pack)
1178 return -1;
1179 if (a->in_pack > b->in_pack)
1180 return 1;
1181 return a->in_pack_offset < b->in_pack_offset ? -1 :
1182 (a->in_pack_offset > b->in_pack_offset);
1185 static void get_object_details(void)
1187 uint32_t i;
1188 struct object_entry **sorted_by_offset;
1190 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1191 for (i = 0; i < nr_objects; i++)
1192 sorted_by_offset[i] = objects + i;
1193 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1195 for (i = 0; i < nr_objects; i++)
1196 check_object(sorted_by_offset[i]);
1198 free(sorted_by_offset);
1202 * We search for deltas in a list sorted by type, by filename hash, and then
1203 * by size, so that we see progressively smaller and smaller files.
1204 * That's because we prefer deltas to be from the bigger file
1205 * to the smaller -- deletes are potentially cheaper, but perhaps
1206 * more importantly, the bigger file is likely the more recent
1207 * one. The deepest deltas are therefore the oldest objects which are
1208 * less susceptible to be accessed often.
1210 static int type_size_sort(const void *_a, const void *_b)
1212 const struct object_entry *a = *(struct object_entry **)_a;
1213 const struct object_entry *b = *(struct object_entry **)_b;
1215 if (a->type > b->type)
1216 return -1;
1217 if (a->type < b->type)
1218 return 1;
1219 if (a->hash > b->hash)
1220 return -1;
1221 if (a->hash < b->hash)
1222 return 1;
1223 if (a->preferred_base > b->preferred_base)
1224 return -1;
1225 if (a->preferred_base < b->preferred_base)
1226 return 1;
1227 if (a->size > b->size)
1228 return -1;
1229 if (a->size < b->size)
1230 return 1;
1231 return a < b ? -1 : (a > b); /* newest first */
1234 struct unpacked {
1235 struct object_entry *entry;
1236 void *data;
1237 struct delta_index *index;
1238 unsigned depth;
1241 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1242 unsigned long delta_size)
1244 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1245 return 0;
1247 if (delta_size < cache_max_small_delta_size)
1248 return 1;
1250 /* cache delta, if objects are large enough compared to delta size */
1251 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1252 return 1;
1254 return 0;
1257 #ifdef THREADED_DELTA_SEARCH
1259 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1260 #define read_lock() pthread_mutex_lock(&read_mutex)
1261 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1263 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1264 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1265 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1267 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1268 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1269 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1271 #else
1273 #define read_lock() (void)0
1274 #define read_unlock() (void)0
1275 #define cache_lock() (void)0
1276 #define cache_unlock() (void)0
1277 #define progress_lock() (void)0
1278 #define progress_unlock() (void)0
1280 #endif
1282 static int try_delta(struct unpacked *trg, struct unpacked *src,
1283 unsigned max_depth, unsigned long *mem_usage)
1285 struct object_entry *trg_entry = trg->entry;
1286 struct object_entry *src_entry = src->entry;
1287 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1288 unsigned ref_depth;
1289 enum object_type type;
1290 void *delta_buf;
1292 /* Don't bother doing diffs between different types */
1293 if (trg_entry->type != src_entry->type)
1294 return -1;
1297 * We do not bother to try a delta that we discarded
1298 * on an earlier try, but only when reusing delta data.
1300 if (reuse_delta && trg_entry->in_pack &&
1301 trg_entry->in_pack == src_entry->in_pack &&
1302 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1303 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1304 return 0;
1306 /* Let's not bust the allowed depth. */
1307 if (src->depth >= max_depth)
1308 return 0;
1310 /* Now some size filtering heuristics. */
1311 trg_size = trg_entry->size;
1312 if (!trg_entry->delta) {
1313 max_size = trg_size/2 - 20;
1314 ref_depth = 1;
1315 } else {
1316 max_size = trg_entry->delta_size;
1317 ref_depth = trg->depth;
1319 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1320 (max_depth - ref_depth + 1);
1321 if (max_size == 0)
1322 return 0;
1323 src_size = src_entry->size;
1324 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1325 if (sizediff >= max_size)
1326 return 0;
1327 if (trg_size < src_size / 32)
1328 return 0;
1330 /* Load data if not already done */
1331 if (!trg->data) {
1332 read_lock();
1333 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1334 read_unlock();
1335 if (!trg->data)
1336 die("object %s cannot be read",
1337 sha1_to_hex(trg_entry->idx.sha1));
1338 if (sz != trg_size)
1339 die("object %s inconsistent object length (%lu vs %lu)",
1340 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1341 *mem_usage += sz;
1343 if (!src->data) {
1344 read_lock();
1345 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1346 read_unlock();
1347 if (!src->data)
1348 die("object %s cannot be read",
1349 sha1_to_hex(src_entry->idx.sha1));
1350 if (sz != src_size)
1351 die("object %s inconsistent object length (%lu vs %lu)",
1352 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1353 *mem_usage += sz;
1355 if (!src->index) {
1356 src->index = create_delta_index(src->data, src_size);
1357 if (!src->index) {
1358 static int warned = 0;
1359 if (!warned++)
1360 warning("suboptimal pack - out of memory");
1361 return 0;
1363 *mem_usage += sizeof_delta_index(src->index);
1366 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1367 if (!delta_buf)
1368 return 0;
1370 if (trg_entry->delta) {
1371 /* Prefer only shallower same-sized deltas. */
1372 if (delta_size == trg_entry->delta_size &&
1373 src->depth + 1 >= trg->depth) {
1374 free(delta_buf);
1375 return 0;
1380 * Handle memory allocation outside of the cache
1381 * accounting lock. Compiler will optimize the strangeness
1382 * away when THREADED_DELTA_SEARCH is not defined.
1384 free(trg_entry->delta_data);
1385 cache_lock();
1386 if (trg_entry->delta_data) {
1387 delta_cache_size -= trg_entry->delta_size;
1388 trg_entry->delta_data = NULL;
1390 if (delta_cacheable(src_size, trg_size, delta_size)) {
1391 delta_cache_size += delta_size;
1392 cache_unlock();
1393 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1394 } else {
1395 cache_unlock();
1396 free(delta_buf);
1399 trg_entry->delta = src_entry;
1400 trg_entry->delta_size = delta_size;
1401 trg->depth = src->depth + 1;
1403 return 1;
1406 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1408 struct object_entry *child = me->delta_child;
1409 unsigned int m = n;
1410 while (child) {
1411 unsigned int c = check_delta_limit(child, n + 1);
1412 if (m < c)
1413 m = c;
1414 child = child->delta_sibling;
1416 return m;
1419 static unsigned long free_unpacked(struct unpacked *n)
1421 unsigned long freed_mem = sizeof_delta_index(n->index);
1422 free_delta_index(n->index);
1423 n->index = NULL;
1424 if (n->data) {
1425 freed_mem += n->entry->size;
1426 free(n->data);
1427 n->data = NULL;
1429 n->entry = NULL;
1430 n->depth = 0;
1431 return freed_mem;
1434 static void find_deltas(struct object_entry **list, unsigned *list_size,
1435 int window, int depth, unsigned *processed)
1437 uint32_t i, idx = 0, count = 0;
1438 struct unpacked *array;
1439 unsigned long mem_usage = 0;
1441 array = xcalloc(window, sizeof(struct unpacked));
1443 for (;;) {
1444 struct object_entry *entry;
1445 struct unpacked *n = array + idx;
1446 int j, max_depth, best_base = -1;
1448 progress_lock();
1449 if (!*list_size) {
1450 progress_unlock();
1451 break;
1453 entry = *list++;
1454 (*list_size)--;
1455 if (!entry->preferred_base) {
1456 (*processed)++;
1457 display_progress(progress_state, *processed);
1459 progress_unlock();
1461 mem_usage -= free_unpacked(n);
1462 n->entry = entry;
1464 while (window_memory_limit &&
1465 mem_usage > window_memory_limit &&
1466 count > 1) {
1467 uint32_t tail = (idx + window - count) % window;
1468 mem_usage -= free_unpacked(array + tail);
1469 count--;
1472 /* We do not compute delta to *create* objects we are not
1473 * going to pack.
1475 if (entry->preferred_base)
1476 goto next;
1479 * If the current object is at pack edge, take the depth the
1480 * objects that depend on the current object into account
1481 * otherwise they would become too deep.
1483 max_depth = depth;
1484 if (entry->delta_child) {
1485 max_depth -= check_delta_limit(entry, 0);
1486 if (max_depth <= 0)
1487 goto next;
1490 j = window;
1491 while (--j > 0) {
1492 int ret;
1493 uint32_t other_idx = idx + j;
1494 struct unpacked *m;
1495 if (other_idx >= window)
1496 other_idx -= window;
1497 m = array + other_idx;
1498 if (!m->entry)
1499 break;
1500 ret = try_delta(n, m, max_depth, &mem_usage);
1501 if (ret < 0)
1502 break;
1503 else if (ret > 0)
1504 best_base = other_idx;
1508 * If we decided to cache the delta data, then it is best
1509 * to compress it right away. First because we have to do
1510 * it anyway, and doing it here while we're threaded will
1511 * save a lot of time in the non threaded write phase,
1512 * as well as allow for caching more deltas within
1513 * the same cache size limit.
1514 * ...
1515 * But only if not writing to stdout, since in that case
1516 * the network is most likely throttling writes anyway,
1517 * and therefore it is best to go to the write phase ASAP
1518 * instead, as we can afford spending more time compressing
1519 * between writes at that moment.
1521 if (entry->delta_data && !pack_to_stdout) {
1522 entry->z_delta_size = do_compress(&entry->delta_data,
1523 entry->delta_size);
1524 cache_lock();
1525 delta_cache_size -= entry->delta_size;
1526 delta_cache_size += entry->z_delta_size;
1527 cache_unlock();
1530 /* if we made n a delta, and if n is already at max
1531 * depth, leaving it in the window is pointless. we
1532 * should evict it first.
1534 if (entry->delta && max_depth <= n->depth)
1535 continue;
1538 * Move the best delta base up in the window, after the
1539 * currently deltified object, to keep it longer. It will
1540 * be the first base object to be attempted next.
1542 if (entry->delta) {
1543 struct unpacked swap = array[best_base];
1544 int dist = (window + idx - best_base) % window;
1545 int dst = best_base;
1546 while (dist--) {
1547 int src = (dst + 1) % window;
1548 array[dst] = array[src];
1549 dst = src;
1551 array[dst] = swap;
1554 next:
1555 idx++;
1556 if (count + 1 < window)
1557 count++;
1558 if (idx >= window)
1559 idx = 0;
1562 for (i = 0; i < window; ++i) {
1563 free_delta_index(array[i].index);
1564 free(array[i].data);
1566 free(array);
1569 #ifdef THREADED_DELTA_SEARCH
1572 * The main thread waits on the condition that (at least) one of the workers
1573 * has stopped working (which is indicated in the .working member of
1574 * struct thread_params).
1575 * When a work thread has completed its work, it sets .working to 0 and
1576 * signals the main thread and waits on the condition that .data_ready
1577 * becomes 1.
1580 struct thread_params {
1581 pthread_t thread;
1582 struct object_entry **list;
1583 unsigned list_size;
1584 unsigned remaining;
1585 int window;
1586 int depth;
1587 int working;
1588 int data_ready;
1589 pthread_mutex_t mutex;
1590 pthread_cond_t cond;
1591 unsigned *processed;
1594 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1596 static void *threaded_find_deltas(void *arg)
1598 struct thread_params *me = arg;
1600 while (me->remaining) {
1601 find_deltas(me->list, &me->remaining,
1602 me->window, me->depth, me->processed);
1604 progress_lock();
1605 me->working = 0;
1606 pthread_cond_signal(&progress_cond);
1607 progress_unlock();
1610 * We must not set ->data_ready before we wait on the
1611 * condition because the main thread may have set it to 1
1612 * before we get here. In order to be sure that new
1613 * work is available if we see 1 in ->data_ready, it
1614 * was initialized to 0 before this thread was spawned
1615 * and we reset it to 0 right away.
1617 pthread_mutex_lock(&me->mutex);
1618 while (!me->data_ready)
1619 pthread_cond_wait(&me->cond, &me->mutex);
1620 me->data_ready = 0;
1621 pthread_mutex_unlock(&me->mutex);
1623 /* leave ->working 1 so that this doesn't get more work assigned */
1624 return NULL;
1627 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1628 int window, int depth, unsigned *processed)
1630 struct thread_params *p;
1631 int i, ret, active_threads = 0;
1633 if (!delta_search_threads) /* --threads=0 means autodetect */
1634 delta_search_threads = online_cpus();
1635 if (delta_search_threads <= 1) {
1636 find_deltas(list, &list_size, window, depth, processed);
1637 return;
1639 if (progress > pack_to_stdout)
1640 fprintf(stderr, "Delta compression using up to %d threads.\n",
1641 delta_search_threads);
1642 p = xcalloc(delta_search_threads, sizeof(*p));
1644 /* Partition the work amongst work threads. */
1645 for (i = 0; i < delta_search_threads; i++) {
1646 unsigned sub_size = list_size / (delta_search_threads - i);
1648 /* don't use too small segments or no deltas will be found */
1649 if (sub_size < 2*window && i+1 < delta_search_threads)
1650 sub_size = 0;
1652 p[i].window = window;
1653 p[i].depth = depth;
1654 p[i].processed = processed;
1655 p[i].working = 1;
1656 p[i].data_ready = 0;
1658 /* try to split chunks on "path" boundaries */
1659 while (sub_size && sub_size < list_size &&
1660 list[sub_size]->hash &&
1661 list[sub_size]->hash == list[sub_size-1]->hash)
1662 sub_size++;
1664 p[i].list = list;
1665 p[i].list_size = sub_size;
1666 p[i].remaining = sub_size;
1668 list += sub_size;
1669 list_size -= sub_size;
1672 /* Start work threads. */
1673 for (i = 0; i < delta_search_threads; i++) {
1674 if (!p[i].list_size)
1675 continue;
1676 pthread_mutex_init(&p[i].mutex, NULL);
1677 pthread_cond_init(&p[i].cond, NULL);
1678 ret = pthread_create(&p[i].thread, NULL,
1679 threaded_find_deltas, &p[i]);
1680 if (ret)
1681 die("unable to create thread: %s", strerror(ret));
1682 active_threads++;
1686 * Now let's wait for work completion. Each time a thread is done
1687 * with its work, we steal half of the remaining work from the
1688 * thread with the largest number of unprocessed objects and give
1689 * it to that newly idle thread. This ensure good load balancing
1690 * until the remaining object list segments are simply too short
1691 * to be worth splitting anymore.
1693 while (active_threads) {
1694 struct thread_params *target = NULL;
1695 struct thread_params *victim = NULL;
1696 unsigned sub_size = 0;
1698 progress_lock();
1699 for (;;) {
1700 for (i = 0; !target && i < delta_search_threads; i++)
1701 if (!p[i].working)
1702 target = &p[i];
1703 if (target)
1704 break;
1705 pthread_cond_wait(&progress_cond, &progress_mutex);
1708 for (i = 0; i < delta_search_threads; i++)
1709 if (p[i].remaining > 2*window &&
1710 (!victim || victim->remaining < p[i].remaining))
1711 victim = &p[i];
1712 if (victim) {
1713 sub_size = victim->remaining / 2;
1714 list = victim->list + victim->list_size - sub_size;
1715 while (sub_size && list[0]->hash &&
1716 list[0]->hash == list[-1]->hash) {
1717 list++;
1718 sub_size--;
1720 if (!sub_size) {
1722 * It is possible for some "paths" to have
1723 * so many objects that no hash boundary
1724 * might be found. Let's just steal the
1725 * exact half in that case.
1727 sub_size = victim->remaining / 2;
1728 list -= sub_size;
1730 target->list = list;
1731 victim->list_size -= sub_size;
1732 victim->remaining -= sub_size;
1734 target->list_size = sub_size;
1735 target->remaining = sub_size;
1736 target->working = 1;
1737 progress_unlock();
1739 pthread_mutex_lock(&target->mutex);
1740 target->data_ready = 1;
1741 pthread_cond_signal(&target->cond);
1742 pthread_mutex_unlock(&target->mutex);
1744 if (!sub_size) {
1745 pthread_join(target->thread, NULL);
1746 pthread_cond_destroy(&target->cond);
1747 pthread_mutex_destroy(&target->mutex);
1748 active_threads--;
1751 free(p);
1754 #else
1755 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1756 #endif
1758 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1760 unsigned char peeled[20];
1762 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1763 !peel_ref(path, peeled) && /* peelable? */
1764 !is_null_sha1(peeled) && /* annotated tag? */
1765 locate_object_entry(peeled)) /* object packed? */
1766 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1767 return 0;
1770 static void prepare_pack(int window, int depth)
1772 struct object_entry **delta_list;
1773 uint32_t i, nr_deltas;
1774 unsigned n;
1776 get_object_details();
1779 * If we're locally repacking then we need to be doubly careful
1780 * from now on in order to make sure no stealth corruption gets
1781 * propagated to the new pack. Clients receiving streamed packs
1782 * should validate everything they get anyway so no need to incur
1783 * the additional cost here in that case.
1785 if (!pack_to_stdout)
1786 do_check_packed_object_crc = 1;
1788 if (!nr_objects || !window || !depth)
1789 return;
1791 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1792 nr_deltas = n = 0;
1794 for (i = 0; i < nr_objects; i++) {
1795 struct object_entry *entry = objects + i;
1797 if (entry->delta)
1798 /* This happens if we decided to reuse existing
1799 * delta from a pack. "reuse_delta &&" is implied.
1801 continue;
1803 if (entry->size < 50)
1804 continue;
1806 if (entry->no_try_delta)
1807 continue;
1809 if (!entry->preferred_base) {
1810 nr_deltas++;
1811 if (entry->type < 0)
1812 die("unable to get type of object %s",
1813 sha1_to_hex(entry->idx.sha1));
1814 } else {
1815 if (entry->type < 0) {
1817 * This object is not found, but we
1818 * don't have to include it anyway.
1820 continue;
1824 delta_list[n++] = entry;
1827 if (nr_deltas && n > 1) {
1828 unsigned nr_done = 0;
1829 if (progress)
1830 progress_state = start_progress("Compressing objects",
1831 nr_deltas);
1832 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1833 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1834 stop_progress(&progress_state);
1835 if (nr_done != nr_deltas)
1836 die("inconsistency with delta count");
1838 free(delta_list);
1841 static int git_pack_config(const char *k, const char *v, void *cb)
1843 if (!strcmp(k, "pack.window")) {
1844 window = git_config_int(k, v);
1845 return 0;
1847 if (!strcmp(k, "pack.windowmemory")) {
1848 window_memory_limit = git_config_ulong(k, v);
1849 return 0;
1851 if (!strcmp(k, "pack.depth")) {
1852 depth = git_config_int(k, v);
1853 return 0;
1855 if (!strcmp(k, "pack.compression")) {
1856 int level = git_config_int(k, v);
1857 if (level == -1)
1858 level = Z_DEFAULT_COMPRESSION;
1859 else if (level < 0 || level > Z_BEST_COMPRESSION)
1860 die("bad pack compression level %d", level);
1861 pack_compression_level = level;
1862 pack_compression_seen = 1;
1863 return 0;
1865 if (!strcmp(k, "pack.deltacachesize")) {
1866 max_delta_cache_size = git_config_int(k, v);
1867 return 0;
1869 if (!strcmp(k, "pack.deltacachelimit")) {
1870 cache_max_small_delta_size = git_config_int(k, v);
1871 return 0;
1873 if (!strcmp(k, "pack.threads")) {
1874 delta_search_threads = git_config_int(k, v);
1875 if (delta_search_threads < 0)
1876 die("invalid number of threads specified (%d)",
1877 delta_search_threads);
1878 #ifndef THREADED_DELTA_SEARCH
1879 if (delta_search_threads != 1)
1880 warning("no threads support, ignoring %s", k);
1881 #endif
1882 return 0;
1884 if (!strcmp(k, "pack.indexversion")) {
1885 pack_idx_default_version = git_config_int(k, v);
1886 if (pack_idx_default_version > 2)
1887 die("bad pack.indexversion=%"PRIu32,
1888 pack_idx_default_version);
1889 return 0;
1891 if (!strcmp(k, "pack.packsizelimit")) {
1892 pack_size_limit_cfg = git_config_ulong(k, v);
1893 return 0;
1895 return git_default_config(k, v, cb);
1898 static void read_object_list_from_stdin(void)
1900 char line[40 + 1 + PATH_MAX + 2];
1901 unsigned char sha1[20];
1903 for (;;) {
1904 if (!fgets(line, sizeof(line), stdin)) {
1905 if (feof(stdin))
1906 break;
1907 if (!ferror(stdin))
1908 die("fgets returned NULL, not EOF, not error!");
1909 if (errno != EINTR)
1910 die_errno("fgets");
1911 clearerr(stdin);
1912 continue;
1914 if (line[0] == '-') {
1915 if (get_sha1_hex(line+1, sha1))
1916 die("expected edge sha1, got garbage:\n %s",
1917 line);
1918 add_preferred_base(sha1);
1919 continue;
1921 if (get_sha1_hex(line, sha1))
1922 die("expected sha1, got garbage:\n %s", line);
1924 add_preferred_base_object(line+41);
1925 add_object_entry(sha1, 0, line+41, 0);
1929 #define OBJECT_ADDED (1u<<20)
1931 static void show_commit(struct commit *commit, void *data)
1933 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1934 commit->object.flags |= OBJECT_ADDED;
1937 static void show_object(struct object *obj, const struct name_path *path, const char *last)
1939 char *name = path_name(path, last);
1941 add_preferred_base_object(name);
1942 add_object_entry(obj->sha1, obj->type, name, 0);
1943 obj->flags |= OBJECT_ADDED;
1946 * We will have generated the hash from the name,
1947 * but not saved a pointer to it - we can free it
1949 free((char *)name);
1952 static void show_edge(struct commit *commit)
1954 add_preferred_base(commit->object.sha1);
1957 struct in_pack_object {
1958 off_t offset;
1959 struct object *object;
1962 struct in_pack {
1963 int alloc;
1964 int nr;
1965 struct in_pack_object *array;
1968 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1970 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1971 in_pack->array[in_pack->nr].object = object;
1972 in_pack->nr++;
1976 * Compare the objects in the offset order, in order to emulate the
1977 * "git rev-list --objects" output that produced the pack originally.
1979 static int ofscmp(const void *a_, const void *b_)
1981 struct in_pack_object *a = (struct in_pack_object *)a_;
1982 struct in_pack_object *b = (struct in_pack_object *)b_;
1984 if (a->offset < b->offset)
1985 return -1;
1986 else if (a->offset > b->offset)
1987 return 1;
1988 else
1989 return hashcmp(a->object->sha1, b->object->sha1);
1992 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1994 struct packed_git *p;
1995 struct in_pack in_pack;
1996 uint32_t i;
1998 memset(&in_pack, 0, sizeof(in_pack));
2000 for (p = packed_git; p; p = p->next) {
2001 const unsigned char *sha1;
2002 struct object *o;
2004 if (!p->pack_local || p->pack_keep)
2005 continue;
2006 if (open_pack_index(p))
2007 die("cannot open pack index");
2009 ALLOC_GROW(in_pack.array,
2010 in_pack.nr + p->num_objects,
2011 in_pack.alloc);
2013 for (i = 0; i < p->num_objects; i++) {
2014 sha1 = nth_packed_object_sha1(p, i);
2015 o = lookup_unknown_object(sha1);
2016 if (!(o->flags & OBJECT_ADDED))
2017 mark_in_pack_object(o, p, &in_pack);
2018 o->flags |= OBJECT_ADDED;
2022 if (in_pack.nr) {
2023 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2024 ofscmp);
2025 for (i = 0; i < in_pack.nr; i++) {
2026 struct object *o = in_pack.array[i].object;
2027 add_object_entry(o->sha1, o->type, "", 0);
2030 free(in_pack.array);
2033 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2035 static struct packed_git *last_found = (void *)1;
2036 struct packed_git *p;
2038 p = (last_found != (void *)1) ? last_found : packed_git;
2040 while (p) {
2041 if ((!p->pack_local || p->pack_keep) &&
2042 find_pack_entry_one(sha1, p)) {
2043 last_found = p;
2044 return 1;
2046 if (p == last_found)
2047 p = packed_git;
2048 else
2049 p = p->next;
2050 if (p == last_found)
2051 p = p->next;
2053 return 0;
2056 static void loosen_unused_packed_objects(struct rev_info *revs)
2058 struct packed_git *p;
2059 uint32_t i;
2060 const unsigned char *sha1;
2062 for (p = packed_git; p; p = p->next) {
2063 if (!p->pack_local || p->pack_keep)
2064 continue;
2066 if (open_pack_index(p))
2067 die("cannot open pack index");
2069 for (i = 0; i < p->num_objects; i++) {
2070 sha1 = nth_packed_object_sha1(p, i);
2071 if (!locate_object_entry(sha1) &&
2072 !has_sha1_pack_kept_or_nonlocal(sha1))
2073 if (force_object_loose(sha1, p->mtime))
2074 die("unable to force loose object");
2079 static void get_object_list(int ac, const char **av)
2081 struct rev_info revs;
2082 char line[1000];
2083 int flags = 0;
2085 init_revisions(&revs, NULL);
2086 save_commit_buffer = 0;
2087 setup_revisions(ac, av, &revs, NULL);
2089 while (fgets(line, sizeof(line), stdin) != NULL) {
2090 int len = strlen(line);
2091 if (len && line[len - 1] == '\n')
2092 line[--len] = 0;
2093 if (!len)
2094 break;
2095 if (*line == '-') {
2096 if (!strcmp(line, "--not")) {
2097 flags ^= UNINTERESTING;
2098 continue;
2100 die("not a rev '%s'", line);
2102 if (handle_revision_arg(line, &revs, flags, 1))
2103 die("bad revision '%s'", line);
2106 if (prepare_revision_walk(&revs))
2107 die("revision walk setup failed");
2108 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2109 traverse_commit_list(&revs, show_commit, show_object, NULL);
2111 if (keep_unreachable)
2112 add_objects_in_unpacked_packs(&revs);
2113 if (unpack_unreachable)
2114 loosen_unused_packed_objects(&revs);
2117 static int adjust_perm(const char *path, mode_t mode)
2119 if (chmod(path, mode))
2120 return -1;
2121 return adjust_shared_perm(path);
2124 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2126 int use_internal_rev_list = 0;
2127 int thin = 0;
2128 int all_progress_implied = 0;
2129 uint32_t i;
2130 const char **rp_av;
2131 int rp_ac_alloc = 64;
2132 int rp_ac;
2134 read_replace_refs = 0;
2136 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2138 rp_av[0] = "pack-objects";
2139 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2140 rp_ac = 2;
2142 git_config(git_pack_config, NULL);
2143 if (!pack_compression_seen && core_compression_seen)
2144 pack_compression_level = core_compression_level;
2146 progress = isatty(2);
2147 for (i = 1; i < argc; i++) {
2148 const char *arg = argv[i];
2150 if (*arg != '-')
2151 break;
2153 if (!strcmp("--non-empty", arg)) {
2154 non_empty = 1;
2155 continue;
2157 if (!strcmp("--local", arg)) {
2158 local = 1;
2159 continue;
2161 if (!strcmp("--incremental", arg)) {
2162 incremental = 1;
2163 continue;
2165 if (!strcmp("--honor-pack-keep", arg)) {
2166 ignore_packed_keep = 1;
2167 continue;
2169 if (!prefixcmp(arg, "--compression=")) {
2170 char *end;
2171 int level = strtoul(arg+14, &end, 0);
2172 if (!arg[14] || *end)
2173 usage(pack_usage);
2174 if (level == -1)
2175 level = Z_DEFAULT_COMPRESSION;
2176 else if (level < 0 || level > Z_BEST_COMPRESSION)
2177 die("bad pack compression level %d", level);
2178 pack_compression_level = level;
2179 continue;
2181 if (!prefixcmp(arg, "--max-pack-size=")) {
2182 char *end;
2183 pack_size_limit_cfg = 0;
2184 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2185 if (!arg[16] || *end)
2186 usage(pack_usage);
2187 continue;
2189 if (!prefixcmp(arg, "--window=")) {
2190 char *end;
2191 window = strtoul(arg+9, &end, 0);
2192 if (!arg[9] || *end)
2193 usage(pack_usage);
2194 continue;
2196 if (!prefixcmp(arg, "--window-memory=")) {
2197 if (!git_parse_ulong(arg+16, &window_memory_limit))
2198 usage(pack_usage);
2199 continue;
2201 if (!prefixcmp(arg, "--threads=")) {
2202 char *end;
2203 delta_search_threads = strtoul(arg+10, &end, 0);
2204 if (!arg[10] || *end || delta_search_threads < 0)
2205 usage(pack_usage);
2206 #ifndef THREADED_DELTA_SEARCH
2207 if (delta_search_threads != 1)
2208 warning("no threads support, "
2209 "ignoring %s", arg);
2210 #endif
2211 continue;
2213 if (!prefixcmp(arg, "--depth=")) {
2214 char *end;
2215 depth = strtoul(arg+8, &end, 0);
2216 if (!arg[8] || *end)
2217 usage(pack_usage);
2218 continue;
2220 if (!strcmp("--progress", arg)) {
2221 progress = 1;
2222 continue;
2224 if (!strcmp("--all-progress", arg)) {
2225 progress = 2;
2226 continue;
2228 if (!strcmp("--all-progress-implied", arg)) {
2229 all_progress_implied = 1;
2230 continue;
2232 if (!strcmp("-q", arg)) {
2233 progress = 0;
2234 continue;
2236 if (!strcmp("--no-reuse-delta", arg)) {
2237 reuse_delta = 0;
2238 continue;
2240 if (!strcmp("--no-reuse-object", arg)) {
2241 reuse_object = reuse_delta = 0;
2242 continue;
2244 if (!strcmp("--delta-base-offset", arg)) {
2245 allow_ofs_delta = 1;
2246 continue;
2248 if (!strcmp("--stdout", arg)) {
2249 pack_to_stdout = 1;
2250 continue;
2252 if (!strcmp("--revs", arg)) {
2253 use_internal_rev_list = 1;
2254 continue;
2256 if (!strcmp("--keep-unreachable", arg)) {
2257 keep_unreachable = 1;
2258 continue;
2260 if (!strcmp("--unpack-unreachable", arg)) {
2261 unpack_unreachable = 1;
2262 continue;
2264 if (!strcmp("--include-tag", arg)) {
2265 include_tag = 1;
2266 continue;
2268 if (!strcmp("--unpacked", arg) ||
2269 !strcmp("--reflog", arg) ||
2270 !strcmp("--all", arg)) {
2271 use_internal_rev_list = 1;
2272 if (rp_ac >= rp_ac_alloc - 1) {
2273 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2274 rp_av = xrealloc(rp_av,
2275 rp_ac_alloc * sizeof(*rp_av));
2277 rp_av[rp_ac++] = arg;
2278 continue;
2280 if (!strcmp("--thin", arg)) {
2281 use_internal_rev_list = 1;
2282 thin = 1;
2283 rp_av[1] = "--objects-edge";
2284 continue;
2286 if (!prefixcmp(arg, "--index-version=")) {
2287 char *c;
2288 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2289 if (pack_idx_default_version > 2)
2290 die("bad %s", arg);
2291 if (*c == ',')
2292 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2293 if (*c || pack_idx_off32_limit & 0x80000000)
2294 die("bad %s", arg);
2295 continue;
2297 if (!strcmp(arg, "--keep-true-parents")) {
2298 grafts_replace_parents = 0;
2299 continue;
2301 usage(pack_usage);
2304 /* Traditionally "pack-objects [options] base extra" failed;
2305 * we would however want to take refs parameter that would
2306 * have been given to upstream rev-list ourselves, which means
2307 * we somehow want to say what the base name is. So the
2308 * syntax would be:
2310 * pack-objects [options] base <refs...>
2312 * in other words, we would treat the first non-option as the
2313 * base_name and send everything else to the internal revision
2314 * walker.
2317 if (!pack_to_stdout)
2318 base_name = argv[i++];
2320 if (pack_to_stdout != !base_name)
2321 usage(pack_usage);
2323 if (!pack_to_stdout && !pack_size_limit)
2324 pack_size_limit = pack_size_limit_cfg;
2326 if (pack_to_stdout && pack_size_limit)
2327 die("--max-pack-size cannot be used to build a pack for transfer.");
2329 if (!pack_to_stdout && thin)
2330 die("--thin cannot be used to build an indexable pack.");
2332 if (keep_unreachable && unpack_unreachable)
2333 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2335 if (progress && all_progress_implied)
2336 progress = 2;
2338 prepare_packed_git();
2340 if (progress)
2341 progress_state = start_progress("Counting objects", 0);
2342 if (!use_internal_rev_list)
2343 read_object_list_from_stdin();
2344 else {
2345 rp_av[rp_ac] = NULL;
2346 get_object_list(rp_ac, rp_av);
2348 cleanup_preferred_base();
2349 if (include_tag && nr_result)
2350 for_each_ref(add_ref_tag, NULL);
2351 stop_progress(&progress_state);
2353 if (non_empty && !nr_result)
2354 return 0;
2355 if (nr_result)
2356 prepare_pack(window, depth);
2357 write_pack_file();
2358 if (progress)
2359 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2360 " reused %"PRIu32" (delta %"PRIu32")\n",
2361 written, written_delta, reused, reused_delta);
2362 return 0;