Let check_preimage() use memset() to initialize "struct checkout"
[git/debian.git] / builtin-pack-objects.c
blob539e75d56f7a33fdb1971b51ee0681c43e9662b0
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 #ifndef NO_PTHREADS
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 unsigned long 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 /* apply size 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 static void write_pack_file(void)
469 uint32_t i = 0, j;
470 struct sha1file *f;
471 off_t offset;
472 struct pack_header hdr;
473 uint32_t nr_remaining = nr_result;
474 time_t last_mtime = 0;
476 if (progress > pack_to_stdout)
477 progress_state = start_progress("Writing objects", nr_result);
478 written_list = xmalloc(nr_objects * sizeof(*written_list));
480 do {
481 unsigned char sha1[20];
482 char *pack_tmp_name = NULL;
484 if (pack_to_stdout) {
485 f = sha1fd_throughput(1, "<stdout>", progress_state);
486 } else {
487 char tmpname[PATH_MAX];
488 int fd;
489 fd = odb_mkstemp(tmpname, sizeof(tmpname),
490 "pack/tmp_pack_XXXXXX");
491 pack_tmp_name = xstrdup(tmpname);
492 f = sha1fd(fd, pack_tmp_name);
495 hdr.hdr_signature = htonl(PACK_SIGNATURE);
496 hdr.hdr_version = htonl(PACK_VERSION);
497 hdr.hdr_entries = htonl(nr_remaining);
498 sha1write(f, &hdr, sizeof(hdr));
499 offset = sizeof(hdr);
500 nr_written = 0;
501 for (; i < nr_objects; i++) {
502 if (!write_one(f, objects + i, &offset))
503 break;
504 display_progress(progress_state, written);
508 * Did we write the wrong # entries in the header?
509 * If so, rewrite it like in fast-import
511 if (pack_to_stdout) {
512 sha1close(f, sha1, CSUM_CLOSE);
513 } else if (nr_written == nr_remaining) {
514 sha1close(f, sha1, CSUM_FSYNC);
515 } else {
516 int fd = sha1close(f, sha1, 0);
517 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
518 nr_written, sha1, offset);
519 close(fd);
522 if (!pack_to_stdout) {
523 struct stat st;
524 const char *idx_tmp_name;
525 char tmpname[PATH_MAX];
527 idx_tmp_name = write_idx_file(NULL, written_list,
528 nr_written, sha1);
530 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
531 base_name, sha1_to_hex(sha1));
532 free_pack_by_name(tmpname);
533 if (adjust_shared_perm(pack_tmp_name))
534 die_errno("unable to make temporary pack file readable");
535 if (rename(pack_tmp_name, tmpname))
536 die_errno("unable to rename temporary pack file");
539 * Packs are runtime accessed in their mtime
540 * order since newer packs are more likely to contain
541 * younger objects. So if we are creating multiple
542 * packs then we should modify the mtime of later ones
543 * to preserve this property.
545 if (stat(tmpname, &st) < 0) {
546 warning("failed to stat %s: %s",
547 tmpname, strerror(errno));
548 } else if (!last_mtime) {
549 last_mtime = st.st_mtime;
550 } else {
551 struct utimbuf utb;
552 utb.actime = st.st_atime;
553 utb.modtime = --last_mtime;
554 if (utime(tmpname, &utb) < 0)
555 warning("failed utime() on %s: %s",
556 tmpname, strerror(errno));
559 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
560 base_name, sha1_to_hex(sha1));
561 if (adjust_shared_perm(idx_tmp_name))
562 die_errno("unable to make temporary index file readable");
563 if (rename(idx_tmp_name, tmpname))
564 die_errno("unable to rename temporary index file");
566 free((void *) idx_tmp_name);
567 free(pack_tmp_name);
568 puts(sha1_to_hex(sha1));
571 /* mark written objects as written to previous pack */
572 for (j = 0; j < nr_written; j++) {
573 written_list[j]->offset = (off_t)-1;
575 nr_remaining -= nr_written;
576 } while (nr_remaining && i < nr_objects);
578 free(written_list);
579 stop_progress(&progress_state);
580 if (written != nr_result)
581 die("wrote %"PRIu32" objects while expecting %"PRIu32,
582 written, nr_result);
585 static int locate_object_entry_hash(const unsigned char *sha1)
587 int i;
588 unsigned int ui;
589 memcpy(&ui, sha1, sizeof(unsigned int));
590 i = ui % object_ix_hashsz;
591 while (0 < object_ix[i]) {
592 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
593 return i;
594 if (++i == object_ix_hashsz)
595 i = 0;
597 return -1 - i;
600 static struct object_entry *locate_object_entry(const unsigned char *sha1)
602 int i;
604 if (!object_ix_hashsz)
605 return NULL;
607 i = locate_object_entry_hash(sha1);
608 if (0 <= i)
609 return &objects[object_ix[i]-1];
610 return NULL;
613 static void rehash_objects(void)
615 uint32_t i;
616 struct object_entry *oe;
618 object_ix_hashsz = nr_objects * 3;
619 if (object_ix_hashsz < 1024)
620 object_ix_hashsz = 1024;
621 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
622 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
623 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
624 int ix = locate_object_entry_hash(oe->idx.sha1);
625 if (0 <= ix)
626 continue;
627 ix = -1 - ix;
628 object_ix[ix] = i + 1;
632 static unsigned name_hash(const char *name)
634 unsigned c, hash = 0;
636 if (!name)
637 return 0;
640 * This effectively just creates a sortable number from the
641 * last sixteen non-whitespace characters. Last characters
642 * count "most", so things that end in ".c" sort together.
644 while ((c = *name++) != 0) {
645 if (isspace(c))
646 continue;
647 hash = (hash >> 2) + (c << 24);
649 return hash;
652 static void setup_delta_attr_check(struct git_attr_check *check)
654 static struct git_attr *attr_delta;
656 if (!attr_delta)
657 attr_delta = git_attr("delta");
659 check[0].attr = attr_delta;
662 static int no_try_delta(const char *path)
664 struct git_attr_check check[1];
666 setup_delta_attr_check(check);
667 if (git_checkattr(path, ARRAY_SIZE(check), check))
668 return 0;
669 if (ATTR_FALSE(check->value))
670 return 1;
671 return 0;
674 static int add_object_entry(const unsigned char *sha1, enum object_type type,
675 const char *name, int exclude)
677 struct object_entry *entry;
678 struct packed_git *p, *found_pack = NULL;
679 off_t found_offset = 0;
680 int ix;
681 unsigned hash = name_hash(name);
683 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
684 if (ix >= 0) {
685 if (exclude) {
686 entry = objects + object_ix[ix] - 1;
687 if (!entry->preferred_base)
688 nr_result--;
689 entry->preferred_base = 1;
691 return 0;
694 if (!exclude && local && has_loose_object_nonlocal(sha1))
695 return 0;
697 for (p = packed_git; p; p = p->next) {
698 off_t offset = find_pack_entry_one(sha1, p);
699 if (offset) {
700 if (!found_pack) {
701 found_offset = offset;
702 found_pack = p;
704 if (exclude)
705 break;
706 if (incremental)
707 return 0;
708 if (local && !p->pack_local)
709 return 0;
710 if (ignore_packed_keep && p->pack_local && p->pack_keep)
711 return 0;
715 if (nr_objects >= nr_alloc) {
716 nr_alloc = (nr_alloc + 1024) * 3 / 2;
717 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
720 entry = objects + nr_objects++;
721 memset(entry, 0, sizeof(*entry));
722 hashcpy(entry->idx.sha1, sha1);
723 entry->hash = hash;
724 if (type)
725 entry->type = type;
726 if (exclude)
727 entry->preferred_base = 1;
728 else
729 nr_result++;
730 if (found_pack) {
731 entry->in_pack = found_pack;
732 entry->in_pack_offset = found_offset;
735 if (object_ix_hashsz * 3 <= nr_objects * 4)
736 rehash_objects();
737 else
738 object_ix[-1 - ix] = nr_objects;
740 display_progress(progress_state, nr_objects);
742 if (name && no_try_delta(name))
743 entry->no_try_delta = 1;
745 return 1;
748 struct pbase_tree_cache {
749 unsigned char sha1[20];
750 int ref;
751 int temporary;
752 void *tree_data;
753 unsigned long tree_size;
756 static struct pbase_tree_cache *(pbase_tree_cache[256]);
757 static int pbase_tree_cache_ix(const unsigned char *sha1)
759 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
761 static int pbase_tree_cache_ix_incr(int ix)
763 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
766 static struct pbase_tree {
767 struct pbase_tree *next;
768 /* This is a phony "cache" entry; we are not
769 * going to evict it nor find it through _get()
770 * mechanism -- this is for the toplevel node that
771 * would almost always change with any commit.
773 struct pbase_tree_cache pcache;
774 } *pbase_tree;
776 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
778 struct pbase_tree_cache *ent, *nent;
779 void *data;
780 unsigned long size;
781 enum object_type type;
782 int neigh;
783 int my_ix = pbase_tree_cache_ix(sha1);
784 int available_ix = -1;
786 /* pbase-tree-cache acts as a limited hashtable.
787 * your object will be found at your index or within a few
788 * slots after that slot if it is cached.
790 for (neigh = 0; neigh < 8; neigh++) {
791 ent = pbase_tree_cache[my_ix];
792 if (ent && !hashcmp(ent->sha1, sha1)) {
793 ent->ref++;
794 return ent;
796 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
797 ((0 <= available_ix) &&
798 (!ent && pbase_tree_cache[available_ix])))
799 available_ix = my_ix;
800 if (!ent)
801 break;
802 my_ix = pbase_tree_cache_ix_incr(my_ix);
805 /* Did not find one. Either we got a bogus request or
806 * we need to read and perhaps cache.
808 data = read_sha1_file(sha1, &type, &size);
809 if (!data)
810 return NULL;
811 if (type != OBJ_TREE) {
812 free(data);
813 return NULL;
816 /* We need to either cache or return a throwaway copy */
818 if (available_ix < 0)
819 ent = NULL;
820 else {
821 ent = pbase_tree_cache[available_ix];
822 my_ix = available_ix;
825 if (!ent) {
826 nent = xmalloc(sizeof(*nent));
827 nent->temporary = (available_ix < 0);
829 else {
830 /* evict and reuse */
831 free(ent->tree_data);
832 nent = ent;
834 hashcpy(nent->sha1, sha1);
835 nent->tree_data = data;
836 nent->tree_size = size;
837 nent->ref = 1;
838 if (!nent->temporary)
839 pbase_tree_cache[my_ix] = nent;
840 return nent;
843 static void pbase_tree_put(struct pbase_tree_cache *cache)
845 if (!cache->temporary) {
846 cache->ref--;
847 return;
849 free(cache->tree_data);
850 free(cache);
853 static int name_cmp_len(const char *name)
855 int i;
856 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
858 return i;
861 static void add_pbase_object(struct tree_desc *tree,
862 const char *name,
863 int cmplen,
864 const char *fullname)
866 struct name_entry entry;
867 int cmp;
869 while (tree_entry(tree,&entry)) {
870 if (S_ISGITLINK(entry.mode))
871 continue;
872 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
873 memcmp(name, entry.path, cmplen);
874 if (cmp > 0)
875 continue;
876 if (cmp < 0)
877 return;
878 if (name[cmplen] != '/') {
879 add_object_entry(entry.sha1,
880 object_type(entry.mode),
881 fullname, 1);
882 return;
884 if (S_ISDIR(entry.mode)) {
885 struct tree_desc sub;
886 struct pbase_tree_cache *tree;
887 const char *down = name+cmplen+1;
888 int downlen = name_cmp_len(down);
890 tree = pbase_tree_get(entry.sha1);
891 if (!tree)
892 return;
893 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
895 add_pbase_object(&sub, down, downlen, fullname);
896 pbase_tree_put(tree);
901 static unsigned *done_pbase_paths;
902 static int done_pbase_paths_num;
903 static int done_pbase_paths_alloc;
904 static int done_pbase_path_pos(unsigned hash)
906 int lo = 0;
907 int hi = done_pbase_paths_num;
908 while (lo < hi) {
909 int mi = (hi + lo) / 2;
910 if (done_pbase_paths[mi] == hash)
911 return mi;
912 if (done_pbase_paths[mi] < hash)
913 hi = mi;
914 else
915 lo = mi + 1;
917 return -lo-1;
920 static int check_pbase_path(unsigned hash)
922 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
923 if (0 <= pos)
924 return 1;
925 pos = -pos - 1;
926 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
927 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
928 done_pbase_paths = xrealloc(done_pbase_paths,
929 done_pbase_paths_alloc *
930 sizeof(unsigned));
932 done_pbase_paths_num++;
933 if (pos < done_pbase_paths_num)
934 memmove(done_pbase_paths + pos + 1,
935 done_pbase_paths + pos,
936 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
937 done_pbase_paths[pos] = hash;
938 return 0;
941 static void add_preferred_base_object(const char *name)
943 struct pbase_tree *it;
944 int cmplen;
945 unsigned hash = name_hash(name);
947 if (!num_preferred_base || check_pbase_path(hash))
948 return;
950 cmplen = name_cmp_len(name);
951 for (it = pbase_tree; it; it = it->next) {
952 if (cmplen == 0) {
953 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
955 else {
956 struct tree_desc tree;
957 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
958 add_pbase_object(&tree, name, cmplen, name);
963 static void add_preferred_base(unsigned char *sha1)
965 struct pbase_tree *it;
966 void *data;
967 unsigned long size;
968 unsigned char tree_sha1[20];
970 if (window <= num_preferred_base++)
971 return;
973 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
974 if (!data)
975 return;
977 for (it = pbase_tree; it; it = it->next) {
978 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
979 free(data);
980 return;
984 it = xcalloc(1, sizeof(*it));
985 it->next = pbase_tree;
986 pbase_tree = it;
988 hashcpy(it->pcache.sha1, tree_sha1);
989 it->pcache.tree_data = data;
990 it->pcache.tree_size = size;
993 static void cleanup_preferred_base(void)
995 struct pbase_tree *it;
996 unsigned i;
998 it = pbase_tree;
999 pbase_tree = NULL;
1000 while (it) {
1001 struct pbase_tree *this = it;
1002 it = this->next;
1003 free(this->pcache.tree_data);
1004 free(this);
1007 for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1008 if (!pbase_tree_cache[i])
1009 continue;
1010 free(pbase_tree_cache[i]->tree_data);
1011 free(pbase_tree_cache[i]);
1012 pbase_tree_cache[i] = NULL;
1015 free(done_pbase_paths);
1016 done_pbase_paths = NULL;
1017 done_pbase_paths_num = done_pbase_paths_alloc = 0;
1020 static void check_object(struct object_entry *entry)
1022 if (entry->in_pack) {
1023 struct packed_git *p = entry->in_pack;
1024 struct pack_window *w_curs = NULL;
1025 const unsigned char *base_ref = NULL;
1026 struct object_entry *base_entry;
1027 unsigned long used, used_0;
1028 unsigned int avail;
1029 off_t ofs;
1030 unsigned char *buf, c;
1032 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1035 * We want in_pack_type even if we do not reuse delta
1036 * since non-delta representations could still be reused.
1038 used = unpack_object_header_buffer(buf, avail,
1039 &entry->in_pack_type,
1040 &entry->size);
1041 if (used == 0)
1042 goto give_up;
1045 * Determine if this is a delta and if so whether we can
1046 * reuse it or not. Otherwise let's find out as cheaply as
1047 * possible what the actual type and size for this object is.
1049 switch (entry->in_pack_type) {
1050 default:
1051 /* Not a delta hence we've already got all we need. */
1052 entry->type = entry->in_pack_type;
1053 entry->in_pack_header_size = used;
1054 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1055 goto give_up;
1056 unuse_pack(&w_curs);
1057 return;
1058 case OBJ_REF_DELTA:
1059 if (reuse_delta && !entry->preferred_base)
1060 base_ref = use_pack(p, &w_curs,
1061 entry->in_pack_offset + used, NULL);
1062 entry->in_pack_header_size = used + 20;
1063 break;
1064 case OBJ_OFS_DELTA:
1065 buf = use_pack(p, &w_curs,
1066 entry->in_pack_offset + used, NULL);
1067 used_0 = 0;
1068 c = buf[used_0++];
1069 ofs = c & 127;
1070 while (c & 128) {
1071 ofs += 1;
1072 if (!ofs || MSB(ofs, 7)) {
1073 error("delta base offset overflow in pack for %s",
1074 sha1_to_hex(entry->idx.sha1));
1075 goto give_up;
1077 c = buf[used_0++];
1078 ofs = (ofs << 7) + (c & 127);
1080 ofs = entry->in_pack_offset - ofs;
1081 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1082 error("delta base offset out of bound for %s",
1083 sha1_to_hex(entry->idx.sha1));
1084 goto give_up;
1086 if (reuse_delta && !entry->preferred_base) {
1087 struct revindex_entry *revidx;
1088 revidx = find_pack_revindex(p, ofs);
1089 if (!revidx)
1090 goto give_up;
1091 base_ref = nth_packed_object_sha1(p, revidx->nr);
1093 entry->in_pack_header_size = used + used_0;
1094 break;
1097 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1099 * If base_ref was set above that means we wish to
1100 * reuse delta data, and we even found that base
1101 * in the list of objects we want to pack. Goodie!
1103 * Depth value does not matter - find_deltas() will
1104 * never consider reused delta as the base object to
1105 * deltify other objects against, in order to avoid
1106 * circular deltas.
1108 entry->type = entry->in_pack_type;
1109 entry->delta = base_entry;
1110 entry->delta_size = entry->size;
1111 entry->delta_sibling = base_entry->delta_child;
1112 base_entry->delta_child = entry;
1113 unuse_pack(&w_curs);
1114 return;
1117 if (entry->type) {
1119 * This must be a delta and we already know what the
1120 * final object type is. Let's extract the actual
1121 * object size from the delta header.
1123 entry->size = get_size_from_delta(p, &w_curs,
1124 entry->in_pack_offset + entry->in_pack_header_size);
1125 if (entry->size == 0)
1126 goto give_up;
1127 unuse_pack(&w_curs);
1128 return;
1132 * No choice but to fall back to the recursive delta walk
1133 * with sha1_object_info() to find about the object type
1134 * at this point...
1136 give_up:
1137 unuse_pack(&w_curs);
1140 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1142 * The error condition is checked in prepare_pack(). This is
1143 * to permit a missing preferred base object to be ignored
1144 * as a preferred base. Doing so can result in a larger
1145 * pack file, but the transfer will still take place.
1149 static int pack_offset_sort(const void *_a, const void *_b)
1151 const struct object_entry *a = *(struct object_entry **)_a;
1152 const struct object_entry *b = *(struct object_entry **)_b;
1154 /* avoid filesystem trashing with loose objects */
1155 if (!a->in_pack && !b->in_pack)
1156 return hashcmp(a->idx.sha1, b->idx.sha1);
1158 if (a->in_pack < b->in_pack)
1159 return -1;
1160 if (a->in_pack > b->in_pack)
1161 return 1;
1162 return a->in_pack_offset < b->in_pack_offset ? -1 :
1163 (a->in_pack_offset > b->in_pack_offset);
1166 static void get_object_details(void)
1168 uint32_t i;
1169 struct object_entry **sorted_by_offset;
1171 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1172 for (i = 0; i < nr_objects; i++)
1173 sorted_by_offset[i] = objects + i;
1174 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1176 for (i = 0; i < nr_objects; i++)
1177 check_object(sorted_by_offset[i]);
1179 free(sorted_by_offset);
1183 * We search for deltas in a list sorted by type, by filename hash, and then
1184 * by size, so that we see progressively smaller and smaller files.
1185 * That's because we prefer deltas to be from the bigger file
1186 * to the smaller -- deletes are potentially cheaper, but perhaps
1187 * more importantly, the bigger file is likely the more recent
1188 * one. The deepest deltas are therefore the oldest objects which are
1189 * less susceptible to be accessed often.
1191 static int type_size_sort(const void *_a, const void *_b)
1193 const struct object_entry *a = *(struct object_entry **)_a;
1194 const struct object_entry *b = *(struct object_entry **)_b;
1196 if (a->type > b->type)
1197 return -1;
1198 if (a->type < b->type)
1199 return 1;
1200 if (a->hash > b->hash)
1201 return -1;
1202 if (a->hash < b->hash)
1203 return 1;
1204 if (a->preferred_base > b->preferred_base)
1205 return -1;
1206 if (a->preferred_base < b->preferred_base)
1207 return 1;
1208 if (a->size > b->size)
1209 return -1;
1210 if (a->size < b->size)
1211 return 1;
1212 return a < b ? -1 : (a > b); /* newest first */
1215 struct unpacked {
1216 struct object_entry *entry;
1217 void *data;
1218 struct delta_index *index;
1219 unsigned depth;
1222 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1223 unsigned long delta_size)
1225 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1226 return 0;
1228 if (delta_size < cache_max_small_delta_size)
1229 return 1;
1231 /* cache delta, if objects are large enough compared to delta size */
1232 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1233 return 1;
1235 return 0;
1238 #ifndef NO_PTHREADS
1240 static pthread_mutex_t read_mutex;
1241 #define read_lock() pthread_mutex_lock(&read_mutex)
1242 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1244 static pthread_mutex_t cache_mutex;
1245 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1246 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1248 static pthread_mutex_t progress_mutex;
1249 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1250 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1252 #else
1254 #define read_lock() (void)0
1255 #define read_unlock() (void)0
1256 #define cache_lock() (void)0
1257 #define cache_unlock() (void)0
1258 #define progress_lock() (void)0
1259 #define progress_unlock() (void)0
1261 #endif
1263 static int try_delta(struct unpacked *trg, struct unpacked *src,
1264 unsigned max_depth, unsigned long *mem_usage)
1266 struct object_entry *trg_entry = trg->entry;
1267 struct object_entry *src_entry = src->entry;
1268 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1269 unsigned ref_depth;
1270 enum object_type type;
1271 void *delta_buf;
1273 /* Don't bother doing diffs between different types */
1274 if (trg_entry->type != src_entry->type)
1275 return -1;
1278 * We do not bother to try a delta that we discarded
1279 * on an earlier try, but only when reusing delta data.
1281 if (reuse_delta && trg_entry->in_pack &&
1282 trg_entry->in_pack == src_entry->in_pack &&
1283 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1284 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1285 return 0;
1287 /* Let's not bust the allowed depth. */
1288 if (src->depth >= max_depth)
1289 return 0;
1291 /* Now some size filtering heuristics. */
1292 trg_size = trg_entry->size;
1293 if (!trg_entry->delta) {
1294 max_size = trg_size/2 - 20;
1295 ref_depth = 1;
1296 } else {
1297 max_size = trg_entry->delta_size;
1298 ref_depth = trg->depth;
1300 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1301 (max_depth - ref_depth + 1);
1302 if (max_size == 0)
1303 return 0;
1304 src_size = src_entry->size;
1305 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1306 if (sizediff >= max_size)
1307 return 0;
1308 if (trg_size < src_size / 32)
1309 return 0;
1311 /* Load data if not already done */
1312 if (!trg->data) {
1313 read_lock();
1314 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1315 read_unlock();
1316 if (!trg->data)
1317 die("object %s cannot be read",
1318 sha1_to_hex(trg_entry->idx.sha1));
1319 if (sz != trg_size)
1320 die("object %s inconsistent object length (%lu vs %lu)",
1321 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1322 *mem_usage += sz;
1324 if (!src->data) {
1325 read_lock();
1326 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1327 read_unlock();
1328 if (!src->data)
1329 die("object %s cannot be read",
1330 sha1_to_hex(src_entry->idx.sha1));
1331 if (sz != src_size)
1332 die("object %s inconsistent object length (%lu vs %lu)",
1333 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1334 *mem_usage += sz;
1336 if (!src->index) {
1337 src->index = create_delta_index(src->data, src_size);
1338 if (!src->index) {
1339 static int warned = 0;
1340 if (!warned++)
1341 warning("suboptimal pack - out of memory");
1342 return 0;
1344 *mem_usage += sizeof_delta_index(src->index);
1347 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1348 if (!delta_buf)
1349 return 0;
1351 if (trg_entry->delta) {
1352 /* Prefer only shallower same-sized deltas. */
1353 if (delta_size == trg_entry->delta_size &&
1354 src->depth + 1 >= trg->depth) {
1355 free(delta_buf);
1356 return 0;
1361 * Handle memory allocation outside of the cache
1362 * accounting lock. Compiler will optimize the strangeness
1363 * away when NO_PTHREADS is defined.
1365 free(trg_entry->delta_data);
1366 cache_lock();
1367 if (trg_entry->delta_data) {
1368 delta_cache_size -= trg_entry->delta_size;
1369 trg_entry->delta_data = NULL;
1371 if (delta_cacheable(src_size, trg_size, delta_size)) {
1372 delta_cache_size += delta_size;
1373 cache_unlock();
1374 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1375 } else {
1376 cache_unlock();
1377 free(delta_buf);
1380 trg_entry->delta = src_entry;
1381 trg_entry->delta_size = delta_size;
1382 trg->depth = src->depth + 1;
1384 return 1;
1387 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1389 struct object_entry *child = me->delta_child;
1390 unsigned int m = n;
1391 while (child) {
1392 unsigned int c = check_delta_limit(child, n + 1);
1393 if (m < c)
1394 m = c;
1395 child = child->delta_sibling;
1397 return m;
1400 static unsigned long free_unpacked(struct unpacked *n)
1402 unsigned long freed_mem = sizeof_delta_index(n->index);
1403 free_delta_index(n->index);
1404 n->index = NULL;
1405 if (n->data) {
1406 freed_mem += n->entry->size;
1407 free(n->data);
1408 n->data = NULL;
1410 n->entry = NULL;
1411 n->depth = 0;
1412 return freed_mem;
1415 static void find_deltas(struct object_entry **list, unsigned *list_size,
1416 int window, int depth, unsigned *processed)
1418 uint32_t i, idx = 0, count = 0;
1419 struct unpacked *array;
1420 unsigned long mem_usage = 0;
1422 array = xcalloc(window, sizeof(struct unpacked));
1424 for (;;) {
1425 struct object_entry *entry;
1426 struct unpacked *n = array + idx;
1427 int j, max_depth, best_base = -1;
1429 progress_lock();
1430 if (!*list_size) {
1431 progress_unlock();
1432 break;
1434 entry = *list++;
1435 (*list_size)--;
1436 if (!entry->preferred_base) {
1437 (*processed)++;
1438 display_progress(progress_state, *processed);
1440 progress_unlock();
1442 mem_usage -= free_unpacked(n);
1443 n->entry = entry;
1445 while (window_memory_limit &&
1446 mem_usage > window_memory_limit &&
1447 count > 1) {
1448 uint32_t tail = (idx + window - count) % window;
1449 mem_usage -= free_unpacked(array + tail);
1450 count--;
1453 /* We do not compute delta to *create* objects we are not
1454 * going to pack.
1456 if (entry->preferred_base)
1457 goto next;
1460 * If the current object is at pack edge, take the depth the
1461 * objects that depend on the current object into account
1462 * otherwise they would become too deep.
1464 max_depth = depth;
1465 if (entry->delta_child) {
1466 max_depth -= check_delta_limit(entry, 0);
1467 if (max_depth <= 0)
1468 goto next;
1471 j = window;
1472 while (--j > 0) {
1473 int ret;
1474 uint32_t other_idx = idx + j;
1475 struct unpacked *m;
1476 if (other_idx >= window)
1477 other_idx -= window;
1478 m = array + other_idx;
1479 if (!m->entry)
1480 break;
1481 ret = try_delta(n, m, max_depth, &mem_usage);
1482 if (ret < 0)
1483 break;
1484 else if (ret > 0)
1485 best_base = other_idx;
1489 * If we decided to cache the delta data, then it is best
1490 * to compress it right away. First because we have to do
1491 * it anyway, and doing it here while we're threaded will
1492 * save a lot of time in the non threaded write phase,
1493 * as well as allow for caching more deltas within
1494 * the same cache size limit.
1495 * ...
1496 * But only if not writing to stdout, since in that case
1497 * the network is most likely throttling writes anyway,
1498 * and therefore it is best to go to the write phase ASAP
1499 * instead, as we can afford spending more time compressing
1500 * between writes at that moment.
1502 if (entry->delta_data && !pack_to_stdout) {
1503 entry->z_delta_size = do_compress(&entry->delta_data,
1504 entry->delta_size);
1505 cache_lock();
1506 delta_cache_size -= entry->delta_size;
1507 delta_cache_size += entry->z_delta_size;
1508 cache_unlock();
1511 /* if we made n a delta, and if n is already at max
1512 * depth, leaving it in the window is pointless. we
1513 * should evict it first.
1515 if (entry->delta && max_depth <= n->depth)
1516 continue;
1519 * Move the best delta base up in the window, after the
1520 * currently deltified object, to keep it longer. It will
1521 * be the first base object to be attempted next.
1523 if (entry->delta) {
1524 struct unpacked swap = array[best_base];
1525 int dist = (window + idx - best_base) % window;
1526 int dst = best_base;
1527 while (dist--) {
1528 int src = (dst + 1) % window;
1529 array[dst] = array[src];
1530 dst = src;
1532 array[dst] = swap;
1535 next:
1536 idx++;
1537 if (count + 1 < window)
1538 count++;
1539 if (idx >= window)
1540 idx = 0;
1543 for (i = 0; i < window; ++i) {
1544 free_delta_index(array[i].index);
1545 free(array[i].data);
1547 free(array);
1550 #ifndef NO_PTHREADS
1553 * The main thread waits on the condition that (at least) one of the workers
1554 * has stopped working (which is indicated in the .working member of
1555 * struct thread_params).
1556 * When a work thread has completed its work, it sets .working to 0 and
1557 * signals the main thread and waits on the condition that .data_ready
1558 * becomes 1.
1561 struct thread_params {
1562 pthread_t thread;
1563 struct object_entry **list;
1564 unsigned list_size;
1565 unsigned remaining;
1566 int window;
1567 int depth;
1568 int working;
1569 int data_ready;
1570 pthread_mutex_t mutex;
1571 pthread_cond_t cond;
1572 unsigned *processed;
1575 static pthread_cond_t progress_cond;
1578 * Mutex and conditional variable can't be statically-initialized on Windows.
1580 static void init_threaded_search(void)
1582 pthread_mutex_init(&read_mutex, NULL);
1583 pthread_mutex_init(&cache_mutex, NULL);
1584 pthread_mutex_init(&progress_mutex, NULL);
1585 pthread_cond_init(&progress_cond, NULL);
1588 static void cleanup_threaded_search(void)
1590 pthread_cond_destroy(&progress_cond);
1591 pthread_mutex_destroy(&read_mutex);
1592 pthread_mutex_destroy(&cache_mutex);
1593 pthread_mutex_destroy(&progress_mutex);
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 init_threaded_search();
1635 if (!delta_search_threads) /* --threads=0 means autodetect */
1636 delta_search_threads = online_cpus();
1637 if (delta_search_threads <= 1) {
1638 find_deltas(list, &list_size, window, depth, processed);
1639 cleanup_threaded_search();
1640 return;
1642 if (progress > pack_to_stdout)
1643 fprintf(stderr, "Delta compression using up to %d threads.\n",
1644 delta_search_threads);
1645 p = xcalloc(delta_search_threads, sizeof(*p));
1647 /* Partition the work amongst work threads. */
1648 for (i = 0; i < delta_search_threads; i++) {
1649 unsigned sub_size = list_size / (delta_search_threads - i);
1651 /* don't use too small segments or no deltas will be found */
1652 if (sub_size < 2*window && i+1 < delta_search_threads)
1653 sub_size = 0;
1655 p[i].window = window;
1656 p[i].depth = depth;
1657 p[i].processed = processed;
1658 p[i].working = 1;
1659 p[i].data_ready = 0;
1661 /* try to split chunks on "path" boundaries */
1662 while (sub_size && sub_size < list_size &&
1663 list[sub_size]->hash &&
1664 list[sub_size]->hash == list[sub_size-1]->hash)
1665 sub_size++;
1667 p[i].list = list;
1668 p[i].list_size = sub_size;
1669 p[i].remaining = sub_size;
1671 list += sub_size;
1672 list_size -= sub_size;
1675 /* Start work threads. */
1676 for (i = 0; i < delta_search_threads; i++) {
1677 if (!p[i].list_size)
1678 continue;
1679 pthread_mutex_init(&p[i].mutex, NULL);
1680 pthread_cond_init(&p[i].cond, NULL);
1681 ret = pthread_create(&p[i].thread, NULL,
1682 threaded_find_deltas, &p[i]);
1683 if (ret)
1684 die("unable to create thread: %s", strerror(ret));
1685 active_threads++;
1689 * Now let's wait for work completion. Each time a thread is done
1690 * with its work, we steal half of the remaining work from the
1691 * thread with the largest number of unprocessed objects and give
1692 * it to that newly idle thread. This ensure good load balancing
1693 * until the remaining object list segments are simply too short
1694 * to be worth splitting anymore.
1696 while (active_threads) {
1697 struct thread_params *target = NULL;
1698 struct thread_params *victim = NULL;
1699 unsigned sub_size = 0;
1701 progress_lock();
1702 for (;;) {
1703 for (i = 0; !target && i < delta_search_threads; i++)
1704 if (!p[i].working)
1705 target = &p[i];
1706 if (target)
1707 break;
1708 pthread_cond_wait(&progress_cond, &progress_mutex);
1711 for (i = 0; i < delta_search_threads; i++)
1712 if (p[i].remaining > 2*window &&
1713 (!victim || victim->remaining < p[i].remaining))
1714 victim = &p[i];
1715 if (victim) {
1716 sub_size = victim->remaining / 2;
1717 list = victim->list + victim->list_size - sub_size;
1718 while (sub_size && list[0]->hash &&
1719 list[0]->hash == list[-1]->hash) {
1720 list++;
1721 sub_size--;
1723 if (!sub_size) {
1725 * It is possible for some "paths" to have
1726 * so many objects that no hash boundary
1727 * might be found. Let's just steal the
1728 * exact half in that case.
1730 sub_size = victim->remaining / 2;
1731 list -= sub_size;
1733 target->list = list;
1734 victim->list_size -= sub_size;
1735 victim->remaining -= sub_size;
1737 target->list_size = sub_size;
1738 target->remaining = sub_size;
1739 target->working = 1;
1740 progress_unlock();
1742 pthread_mutex_lock(&target->mutex);
1743 target->data_ready = 1;
1744 pthread_cond_signal(&target->cond);
1745 pthread_mutex_unlock(&target->mutex);
1747 if (!sub_size) {
1748 pthread_join(target->thread, NULL);
1749 pthread_cond_destroy(&target->cond);
1750 pthread_mutex_destroy(&target->mutex);
1751 active_threads--;
1754 cleanup_threaded_search();
1755 free(p);
1758 #else
1759 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1760 #endif
1762 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1764 unsigned char peeled[20];
1766 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1767 !peel_ref(path, peeled) && /* peelable? */
1768 !is_null_sha1(peeled) && /* annotated tag? */
1769 locate_object_entry(peeled)) /* object packed? */
1770 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1771 return 0;
1774 static void prepare_pack(int window, int depth)
1776 struct object_entry **delta_list;
1777 uint32_t i, nr_deltas;
1778 unsigned n;
1780 get_object_details();
1783 * If we're locally repacking then we need to be doubly careful
1784 * from now on in order to make sure no stealth corruption gets
1785 * propagated to the new pack. Clients receiving streamed packs
1786 * should validate everything they get anyway so no need to incur
1787 * the additional cost here in that case.
1789 if (!pack_to_stdout)
1790 do_check_packed_object_crc = 1;
1792 if (!nr_objects || !window || !depth)
1793 return;
1795 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1796 nr_deltas = n = 0;
1798 for (i = 0; i < nr_objects; i++) {
1799 struct object_entry *entry = objects + i;
1801 if (entry->delta)
1802 /* This happens if we decided to reuse existing
1803 * delta from a pack. "reuse_delta &&" is implied.
1805 continue;
1807 if (entry->size < 50)
1808 continue;
1810 if (entry->no_try_delta)
1811 continue;
1813 if (!entry->preferred_base) {
1814 nr_deltas++;
1815 if (entry->type < 0)
1816 die("unable to get type of object %s",
1817 sha1_to_hex(entry->idx.sha1));
1818 } else {
1819 if (entry->type < 0) {
1821 * This object is not found, but we
1822 * don't have to include it anyway.
1824 continue;
1828 delta_list[n++] = entry;
1831 if (nr_deltas && n > 1) {
1832 unsigned nr_done = 0;
1833 if (progress)
1834 progress_state = start_progress("Compressing objects",
1835 nr_deltas);
1836 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1837 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1838 stop_progress(&progress_state);
1839 if (nr_done != nr_deltas)
1840 die("inconsistency with delta count");
1842 free(delta_list);
1845 static int git_pack_config(const char *k, const char *v, void *cb)
1847 if (!strcmp(k, "pack.window")) {
1848 window = git_config_int(k, v);
1849 return 0;
1851 if (!strcmp(k, "pack.windowmemory")) {
1852 window_memory_limit = git_config_ulong(k, v);
1853 return 0;
1855 if (!strcmp(k, "pack.depth")) {
1856 depth = git_config_int(k, v);
1857 return 0;
1859 if (!strcmp(k, "pack.compression")) {
1860 int level = git_config_int(k, v);
1861 if (level == -1)
1862 level = Z_DEFAULT_COMPRESSION;
1863 else if (level < 0 || level > Z_BEST_COMPRESSION)
1864 die("bad pack compression level %d", level);
1865 pack_compression_level = level;
1866 pack_compression_seen = 1;
1867 return 0;
1869 if (!strcmp(k, "pack.deltacachesize")) {
1870 max_delta_cache_size = git_config_int(k, v);
1871 return 0;
1873 if (!strcmp(k, "pack.deltacachelimit")) {
1874 cache_max_small_delta_size = git_config_int(k, v);
1875 return 0;
1877 if (!strcmp(k, "pack.threads")) {
1878 delta_search_threads = git_config_int(k, v);
1879 if (delta_search_threads < 0)
1880 die("invalid number of threads specified (%d)",
1881 delta_search_threads);
1882 #ifdef NO_PTHREADS
1883 if (delta_search_threads != 1)
1884 warning("no threads support, ignoring %s", k);
1885 #endif
1886 return 0;
1888 if (!strcmp(k, "pack.indexversion")) {
1889 pack_idx_default_version = git_config_int(k, v);
1890 if (pack_idx_default_version > 2)
1891 die("bad pack.indexversion=%"PRIu32,
1892 pack_idx_default_version);
1893 return 0;
1895 if (!strcmp(k, "pack.packsizelimit")) {
1896 pack_size_limit_cfg = git_config_ulong(k, v);
1897 return 0;
1899 return git_default_config(k, v, cb);
1902 static void read_object_list_from_stdin(void)
1904 char line[40 + 1 + PATH_MAX + 2];
1905 unsigned char sha1[20];
1907 for (;;) {
1908 if (!fgets(line, sizeof(line), stdin)) {
1909 if (feof(stdin))
1910 break;
1911 if (!ferror(stdin))
1912 die("fgets returned NULL, not EOF, not error!");
1913 if (errno != EINTR)
1914 die_errno("fgets");
1915 clearerr(stdin);
1916 continue;
1918 if (line[0] == '-') {
1919 if (get_sha1_hex(line+1, sha1))
1920 die("expected edge sha1, got garbage:\n %s",
1921 line);
1922 add_preferred_base(sha1);
1923 continue;
1925 if (get_sha1_hex(line, sha1))
1926 die("expected sha1, got garbage:\n %s", line);
1928 add_preferred_base_object(line+41);
1929 add_object_entry(sha1, 0, line+41, 0);
1933 #define OBJECT_ADDED (1u<<20)
1935 static void show_commit(struct commit *commit, void *data)
1937 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1938 commit->object.flags |= OBJECT_ADDED;
1941 static void show_object(struct object *obj, const struct name_path *path, const char *last)
1943 char *name = path_name(path, last);
1945 add_preferred_base_object(name);
1946 add_object_entry(obj->sha1, obj->type, name, 0);
1947 obj->flags |= OBJECT_ADDED;
1950 * We will have generated the hash from the name,
1951 * but not saved a pointer to it - we can free it
1953 free((char *)name);
1956 static void show_edge(struct commit *commit)
1958 add_preferred_base(commit->object.sha1);
1961 struct in_pack_object {
1962 off_t offset;
1963 struct object *object;
1966 struct in_pack {
1967 int alloc;
1968 int nr;
1969 struct in_pack_object *array;
1972 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1974 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1975 in_pack->array[in_pack->nr].object = object;
1976 in_pack->nr++;
1980 * Compare the objects in the offset order, in order to emulate the
1981 * "git rev-list --objects" output that produced the pack originally.
1983 static int ofscmp(const void *a_, const void *b_)
1985 struct in_pack_object *a = (struct in_pack_object *)a_;
1986 struct in_pack_object *b = (struct in_pack_object *)b_;
1988 if (a->offset < b->offset)
1989 return -1;
1990 else if (a->offset > b->offset)
1991 return 1;
1992 else
1993 return hashcmp(a->object->sha1, b->object->sha1);
1996 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1998 struct packed_git *p;
1999 struct in_pack in_pack;
2000 uint32_t i;
2002 memset(&in_pack, 0, sizeof(in_pack));
2004 for (p = packed_git; p; p = p->next) {
2005 const unsigned char *sha1;
2006 struct object *o;
2008 if (!p->pack_local || p->pack_keep)
2009 continue;
2010 if (open_pack_index(p))
2011 die("cannot open pack index");
2013 ALLOC_GROW(in_pack.array,
2014 in_pack.nr + p->num_objects,
2015 in_pack.alloc);
2017 for (i = 0; i < p->num_objects; i++) {
2018 sha1 = nth_packed_object_sha1(p, i);
2019 o = lookup_unknown_object(sha1);
2020 if (!(o->flags & OBJECT_ADDED))
2021 mark_in_pack_object(o, p, &in_pack);
2022 o->flags |= OBJECT_ADDED;
2026 if (in_pack.nr) {
2027 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2028 ofscmp);
2029 for (i = 0; i < in_pack.nr; i++) {
2030 struct object *o = in_pack.array[i].object;
2031 add_object_entry(o->sha1, o->type, "", 0);
2034 free(in_pack.array);
2037 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2039 static struct packed_git *last_found = (void *)1;
2040 struct packed_git *p;
2042 p = (last_found != (void *)1) ? last_found : packed_git;
2044 while (p) {
2045 if ((!p->pack_local || p->pack_keep) &&
2046 find_pack_entry_one(sha1, p)) {
2047 last_found = p;
2048 return 1;
2050 if (p == last_found)
2051 p = packed_git;
2052 else
2053 p = p->next;
2054 if (p == last_found)
2055 p = p->next;
2057 return 0;
2060 static void loosen_unused_packed_objects(struct rev_info *revs)
2062 struct packed_git *p;
2063 uint32_t i;
2064 const unsigned char *sha1;
2066 for (p = packed_git; p; p = p->next) {
2067 if (!p->pack_local || p->pack_keep)
2068 continue;
2070 if (open_pack_index(p))
2071 die("cannot open pack index");
2073 for (i = 0; i < p->num_objects; i++) {
2074 sha1 = nth_packed_object_sha1(p, i);
2075 if (!locate_object_entry(sha1) &&
2076 !has_sha1_pack_kept_or_nonlocal(sha1))
2077 if (force_object_loose(sha1, p->mtime))
2078 die("unable to force loose object");
2083 static void get_object_list(int ac, const char **av)
2085 struct rev_info revs;
2086 char line[1000];
2087 int flags = 0;
2089 init_revisions(&revs, NULL);
2090 save_commit_buffer = 0;
2091 setup_revisions(ac, av, &revs, NULL);
2093 while (fgets(line, sizeof(line), stdin) != NULL) {
2094 int len = strlen(line);
2095 if (len && line[len - 1] == '\n')
2096 line[--len] = 0;
2097 if (!len)
2098 break;
2099 if (*line == '-') {
2100 if (!strcmp(line, "--not")) {
2101 flags ^= UNINTERESTING;
2102 continue;
2104 die("not a rev '%s'", line);
2106 if (handle_revision_arg(line, &revs, flags, 1))
2107 die("bad revision '%s'", line);
2110 if (prepare_revision_walk(&revs))
2111 die("revision walk setup failed");
2112 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2113 traverse_commit_list(&revs, show_commit, show_object, NULL);
2115 if (keep_unreachable)
2116 add_objects_in_unpacked_packs(&revs);
2117 if (unpack_unreachable)
2118 loosen_unused_packed_objects(&revs);
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 int all_progress_implied = 0;
2126 uint32_t i;
2127 const char **rp_av;
2128 int rp_ac_alloc = 64;
2129 int rp_ac;
2131 read_replace_refs = 0;
2133 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2135 rp_av[0] = "pack-objects";
2136 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2137 rp_ac = 2;
2139 git_config(git_pack_config, NULL);
2140 if (!pack_compression_seen && core_compression_seen)
2141 pack_compression_level = core_compression_level;
2143 progress = isatty(2);
2144 for (i = 1; i < argc; i++) {
2145 const char *arg = argv[i];
2147 if (*arg != '-')
2148 break;
2150 if (!strcmp("--non-empty", arg)) {
2151 non_empty = 1;
2152 continue;
2154 if (!strcmp("--local", arg)) {
2155 local = 1;
2156 continue;
2158 if (!strcmp("--incremental", arg)) {
2159 incremental = 1;
2160 continue;
2162 if (!strcmp("--honor-pack-keep", arg)) {
2163 ignore_packed_keep = 1;
2164 continue;
2166 if (!prefixcmp(arg, "--compression=")) {
2167 char *end;
2168 int level = strtoul(arg+14, &end, 0);
2169 if (!arg[14] || *end)
2170 usage(pack_usage);
2171 if (level == -1)
2172 level = Z_DEFAULT_COMPRESSION;
2173 else if (level < 0 || level > Z_BEST_COMPRESSION)
2174 die("bad pack compression level %d", level);
2175 pack_compression_level = level;
2176 continue;
2178 if (!prefixcmp(arg, "--max-pack-size=")) {
2179 pack_size_limit_cfg = 0;
2180 if (!git_parse_ulong(arg+16, &pack_size_limit))
2181 usage(pack_usage);
2182 continue;
2184 if (!prefixcmp(arg, "--window=")) {
2185 char *end;
2186 window = strtoul(arg+9, &end, 0);
2187 if (!arg[9] || *end)
2188 usage(pack_usage);
2189 continue;
2191 if (!prefixcmp(arg, "--window-memory=")) {
2192 if (!git_parse_ulong(arg+16, &window_memory_limit))
2193 usage(pack_usage);
2194 continue;
2196 if (!prefixcmp(arg, "--threads=")) {
2197 char *end;
2198 delta_search_threads = strtoul(arg+10, &end, 0);
2199 if (!arg[10] || *end || delta_search_threads < 0)
2200 usage(pack_usage);
2201 #ifdef NO_PTHREADS
2202 if (delta_search_threads != 1)
2203 warning("no threads support, "
2204 "ignoring %s", arg);
2205 #endif
2206 continue;
2208 if (!prefixcmp(arg, "--depth=")) {
2209 char *end;
2210 depth = strtoul(arg+8, &end, 0);
2211 if (!arg[8] || *end)
2212 usage(pack_usage);
2213 continue;
2215 if (!strcmp("--progress", arg)) {
2216 progress = 1;
2217 continue;
2219 if (!strcmp("--all-progress", arg)) {
2220 progress = 2;
2221 continue;
2223 if (!strcmp("--all-progress-implied", arg)) {
2224 all_progress_implied = 1;
2225 continue;
2227 if (!strcmp("-q", arg)) {
2228 progress = 0;
2229 continue;
2231 if (!strcmp("--no-reuse-delta", arg)) {
2232 reuse_delta = 0;
2233 continue;
2235 if (!strcmp("--no-reuse-object", arg)) {
2236 reuse_object = reuse_delta = 0;
2237 continue;
2239 if (!strcmp("--delta-base-offset", arg)) {
2240 allow_ofs_delta = 1;
2241 continue;
2243 if (!strcmp("--stdout", arg)) {
2244 pack_to_stdout = 1;
2245 continue;
2247 if (!strcmp("--revs", arg)) {
2248 use_internal_rev_list = 1;
2249 continue;
2251 if (!strcmp("--keep-unreachable", arg)) {
2252 keep_unreachable = 1;
2253 continue;
2255 if (!strcmp("--unpack-unreachable", arg)) {
2256 unpack_unreachable = 1;
2257 continue;
2259 if (!strcmp("--include-tag", arg)) {
2260 include_tag = 1;
2261 continue;
2263 if (!strcmp("--unpacked", arg) ||
2264 !strcmp("--reflog", arg) ||
2265 !strcmp("--all", arg)) {
2266 use_internal_rev_list = 1;
2267 if (rp_ac >= rp_ac_alloc - 1) {
2268 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2269 rp_av = xrealloc(rp_av,
2270 rp_ac_alloc * sizeof(*rp_av));
2272 rp_av[rp_ac++] = arg;
2273 continue;
2275 if (!strcmp("--thin", arg)) {
2276 use_internal_rev_list = 1;
2277 thin = 1;
2278 rp_av[1] = "--objects-edge";
2279 continue;
2281 if (!prefixcmp(arg, "--index-version=")) {
2282 char *c;
2283 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2284 if (pack_idx_default_version > 2)
2285 die("bad %s", arg);
2286 if (*c == ',')
2287 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2288 if (*c || pack_idx_off32_limit & 0x80000000)
2289 die("bad %s", arg);
2290 continue;
2292 if (!strcmp(arg, "--keep-true-parents")) {
2293 grafts_replace_parents = 0;
2294 continue;
2296 usage(pack_usage);
2299 /* Traditionally "pack-objects [options] base extra" failed;
2300 * we would however want to take refs parameter that would
2301 * have been given to upstream rev-list ourselves, which means
2302 * we somehow want to say what the base name is. So the
2303 * syntax would be:
2305 * pack-objects [options] base <refs...>
2307 * in other words, we would treat the first non-option as the
2308 * base_name and send everything else to the internal revision
2309 * walker.
2312 if (!pack_to_stdout)
2313 base_name = argv[i++];
2315 if (pack_to_stdout != !base_name)
2316 usage(pack_usage);
2318 if (!pack_to_stdout && !pack_size_limit)
2319 pack_size_limit = pack_size_limit_cfg;
2320 if (pack_to_stdout && pack_size_limit)
2321 die("--max-pack-size cannot be used to build a pack for transfer.");
2322 if (pack_size_limit && pack_size_limit < 1024*1024) {
2323 warning("minimum pack size limit is 1 MiB");
2324 pack_size_limit = 1024*1024;
2327 if (!pack_to_stdout && thin)
2328 die("--thin cannot be used to build an indexable pack.");
2330 if (keep_unreachable && unpack_unreachable)
2331 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2333 if (progress && all_progress_implied)
2334 progress = 2;
2336 prepare_packed_git();
2338 if (progress)
2339 progress_state = start_progress("Counting objects", 0);
2340 if (!use_internal_rev_list)
2341 read_object_list_from_stdin();
2342 else {
2343 rp_av[rp_ac] = NULL;
2344 get_object_list(rp_ac, rp_av);
2346 cleanup_preferred_base();
2347 if (include_tag && nr_result)
2348 for_each_ref(add_ref_tag, NULL);
2349 stop_progress(&progress_state);
2351 if (non_empty && !nr_result)
2352 return 0;
2353 if (nr_result)
2354 prepare_pack(window, depth);
2355 write_pack_file();
2356 if (progress)
2357 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2358 " reused %"PRIu32" (delta %"PRIu32")\n",
2359 written, written_delta, reused, reused_delta);
2360 return 0;