Merge branch 'cb/maint-orphan-merge-noclobber'
[git/kirr.git] / builtin / pack-objects.c
blobf8eba53c82242d0c435524ab9ffab70ba2e2d5ad
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 <pthread.h>
22 #include "thread-utils.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 * we are going to reuse the existing object data as is. make
159 * sure it is not corrupt.
161 static int check_pack_inflate(struct packed_git *p,
162 struct pack_window **w_curs,
163 off_t offset,
164 off_t len,
165 unsigned long expect)
167 z_stream stream;
168 unsigned char fakebuf[4096], *in;
169 int st;
171 memset(&stream, 0, sizeof(stream));
172 git_inflate_init(&stream);
173 do {
174 in = use_pack(p, w_curs, offset, &stream.avail_in);
175 stream.next_in = in;
176 stream.next_out = fakebuf;
177 stream.avail_out = sizeof(fakebuf);
178 st = git_inflate(&stream, Z_FINISH);
179 offset += stream.next_in - in;
180 } while (st == Z_OK || st == Z_BUF_ERROR);
181 git_inflate_end(&stream);
182 return (st == Z_STREAM_END &&
183 stream.total_out == expect &&
184 stream.total_in == len) ? 0 : -1;
187 static void copy_pack_data(struct sha1file *f,
188 struct packed_git *p,
189 struct pack_window **w_curs,
190 off_t offset,
191 off_t len)
193 unsigned char *in;
194 unsigned int avail;
196 while (len) {
197 in = use_pack(p, w_curs, offset, &avail);
198 if (avail > len)
199 avail = (unsigned int)len;
200 sha1write(f, in, avail);
201 offset += avail;
202 len -= avail;
206 static unsigned long write_object(struct sha1file *f,
207 struct object_entry *entry,
208 off_t write_offset)
210 unsigned long size, limit, datalen;
211 void *buf;
212 unsigned char header[10], dheader[10];
213 unsigned hdrlen;
214 enum object_type type;
215 int usable_delta, to_reuse;
217 if (!pack_to_stdout)
218 crc32_begin(f);
220 type = entry->type;
222 /* apply size limit if limited packsize and not first object */
223 if (!pack_size_limit || !nr_written)
224 limit = 0;
225 else if (pack_size_limit <= write_offset)
227 * the earlier object did not fit the limit; avoid
228 * mistaking this with unlimited (i.e. limit = 0).
230 limit = 1;
231 else
232 limit = pack_size_limit - write_offset;
234 if (!entry->delta)
235 usable_delta = 0; /* no delta */
236 else if (!pack_size_limit)
237 usable_delta = 1; /* unlimited packfile */
238 else if (entry->delta->idx.offset == (off_t)-1)
239 usable_delta = 0; /* base was written to another pack */
240 else if (entry->delta->idx.offset)
241 usable_delta = 1; /* base already exists in this pack */
242 else
243 usable_delta = 0; /* base could end up in another pack */
245 if (!reuse_object)
246 to_reuse = 0; /* explicit */
247 else if (!entry->in_pack)
248 to_reuse = 0; /* can't reuse what we don't have */
249 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
250 /* check_object() decided it for us ... */
251 to_reuse = usable_delta;
252 /* ... but pack split may override that */
253 else if (type != entry->in_pack_type)
254 to_reuse = 0; /* pack has delta which is unusable */
255 else if (entry->delta)
256 to_reuse = 0; /* we want to pack afresh */
257 else
258 to_reuse = 1; /* we have it in-pack undeltified,
259 * and we do not need to deltify it.
262 if (!to_reuse) {
263 no_reuse:
264 if (!usable_delta) {
265 buf = read_sha1_file(entry->idx.sha1, &type, &size);
266 if (!buf)
267 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
269 * make sure no cached delta data remains from a
270 * previous attempt before a pack split occurred.
272 free(entry->delta_data);
273 entry->delta_data = NULL;
274 entry->z_delta_size = 0;
275 } else if (entry->delta_data) {
276 size = entry->delta_size;
277 buf = entry->delta_data;
278 entry->delta_data = NULL;
279 type = (allow_ofs_delta && entry->delta->idx.offset) ?
280 OBJ_OFS_DELTA : OBJ_REF_DELTA;
281 } else {
282 buf = get_delta(entry);
283 size = entry->delta_size;
284 type = (allow_ofs_delta && entry->delta->idx.offset) ?
285 OBJ_OFS_DELTA : OBJ_REF_DELTA;
288 if (entry->z_delta_size)
289 datalen = entry->z_delta_size;
290 else
291 datalen = do_compress(&buf, size);
294 * The object header is a byte of 'type' followed by zero or
295 * more bytes of length.
297 hdrlen = encode_in_pack_object_header(type, size, header);
299 if (type == OBJ_OFS_DELTA) {
301 * Deltas with relative base contain an additional
302 * encoding of the relative offset for the delta
303 * base from this object's position in the pack.
305 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
306 unsigned pos = sizeof(dheader) - 1;
307 dheader[pos] = ofs & 127;
308 while (ofs >>= 7)
309 dheader[--pos] = 128 | (--ofs & 127);
310 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
311 free(buf);
312 return 0;
314 sha1write(f, header, hdrlen);
315 sha1write(f, dheader + pos, sizeof(dheader) - pos);
316 hdrlen += sizeof(dheader) - pos;
317 } else if (type == OBJ_REF_DELTA) {
319 * Deltas with a base reference contain
320 * an additional 20 bytes for the base sha1.
322 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
323 free(buf);
324 return 0;
326 sha1write(f, header, hdrlen);
327 sha1write(f, entry->delta->idx.sha1, 20);
328 hdrlen += 20;
329 } else {
330 if (limit && hdrlen + datalen + 20 >= limit) {
331 free(buf);
332 return 0;
334 sha1write(f, header, hdrlen);
336 sha1write(f, buf, datalen);
337 free(buf);
339 else {
340 struct packed_git *p = entry->in_pack;
341 struct pack_window *w_curs = NULL;
342 struct revindex_entry *revidx;
343 off_t offset;
345 if (entry->delta)
346 type = (allow_ofs_delta && entry->delta->idx.offset) ?
347 OBJ_OFS_DELTA : OBJ_REF_DELTA;
348 hdrlen = encode_in_pack_object_header(type, entry->size, header);
350 offset = entry->in_pack_offset;
351 revidx = find_pack_revindex(p, offset);
352 datalen = revidx[1].offset - offset;
353 if (!pack_to_stdout && p->index_version > 1 &&
354 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
355 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
356 unuse_pack(&w_curs);
357 goto no_reuse;
360 offset += entry->in_pack_header_size;
361 datalen -= entry->in_pack_header_size;
362 if (!pack_to_stdout && p->index_version == 1 &&
363 check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
364 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
365 unuse_pack(&w_curs);
366 goto no_reuse;
369 if (type == OBJ_OFS_DELTA) {
370 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
371 unsigned pos = sizeof(dheader) - 1;
372 dheader[pos] = ofs & 127;
373 while (ofs >>= 7)
374 dheader[--pos] = 128 | (--ofs & 127);
375 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
376 unuse_pack(&w_curs);
377 return 0;
379 sha1write(f, header, hdrlen);
380 sha1write(f, dheader + pos, sizeof(dheader) - pos);
381 hdrlen += sizeof(dheader) - pos;
382 reused_delta++;
383 } else if (type == OBJ_REF_DELTA) {
384 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
385 unuse_pack(&w_curs);
386 return 0;
388 sha1write(f, header, hdrlen);
389 sha1write(f, entry->delta->idx.sha1, 20);
390 hdrlen += 20;
391 reused_delta++;
392 } else {
393 if (limit && hdrlen + datalen + 20 >= limit) {
394 unuse_pack(&w_curs);
395 return 0;
397 sha1write(f, header, hdrlen);
399 copy_pack_data(f, p, &w_curs, offset, datalen);
400 unuse_pack(&w_curs);
401 reused++;
403 if (usable_delta)
404 written_delta++;
405 written++;
406 if (!pack_to_stdout)
407 entry->idx.crc32 = crc32_end(f);
408 return hdrlen + datalen;
411 static int write_one(struct sha1file *f,
412 struct object_entry *e,
413 off_t *offset)
415 unsigned long size;
417 /* offset is non zero if object is written already. */
418 if (e->idx.offset || e->preferred_base)
419 return -1;
421 /* if we are deltified, write out base object first. */
422 if (e->delta && !write_one(f, e->delta, offset))
423 return 0;
425 e->idx.offset = *offset;
426 size = write_object(f, e, *offset);
427 if (!size) {
428 e->idx.offset = 0;
429 return 0;
431 written_list[nr_written++] = &e->idx;
433 /* make sure off_t is sufficiently large not to wrap */
434 if (signed_add_overflows(*offset, size))
435 die("pack too large for current definition of off_t");
436 *offset += size;
437 return 1;
440 static void write_pack_file(void)
442 uint32_t i = 0, j;
443 struct sha1file *f;
444 off_t offset;
445 struct pack_header hdr;
446 uint32_t nr_remaining = nr_result;
447 time_t last_mtime = 0;
449 if (progress > pack_to_stdout)
450 progress_state = start_progress("Writing objects", nr_result);
451 written_list = xmalloc(nr_objects * sizeof(*written_list));
453 do {
454 unsigned char sha1[20];
455 char *pack_tmp_name = NULL;
457 if (pack_to_stdout) {
458 f = sha1fd_throughput(1, "<stdout>", progress_state);
459 } else {
460 char tmpname[PATH_MAX];
461 int fd;
462 fd = odb_mkstemp(tmpname, sizeof(tmpname),
463 "pack/tmp_pack_XXXXXX");
464 pack_tmp_name = xstrdup(tmpname);
465 f = sha1fd(fd, pack_tmp_name);
468 hdr.hdr_signature = htonl(PACK_SIGNATURE);
469 hdr.hdr_version = htonl(PACK_VERSION);
470 hdr.hdr_entries = htonl(nr_remaining);
471 sha1write(f, &hdr, sizeof(hdr));
472 offset = sizeof(hdr);
473 nr_written = 0;
474 for (; i < nr_objects; i++) {
475 if (!write_one(f, objects + i, &offset))
476 break;
477 display_progress(progress_state, written);
481 * Did we write the wrong # entries in the header?
482 * If so, rewrite it like in fast-import
484 if (pack_to_stdout) {
485 sha1close(f, sha1, CSUM_CLOSE);
486 } else if (nr_written == nr_remaining) {
487 sha1close(f, sha1, CSUM_FSYNC);
488 } else {
489 int fd = sha1close(f, sha1, 0);
490 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
491 nr_written, sha1, offset);
492 close(fd);
495 if (!pack_to_stdout) {
496 struct stat st;
497 const char *idx_tmp_name;
498 char tmpname[PATH_MAX];
500 idx_tmp_name = write_idx_file(NULL, written_list,
501 nr_written, sha1);
503 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
504 base_name, sha1_to_hex(sha1));
505 free_pack_by_name(tmpname);
506 if (adjust_shared_perm(pack_tmp_name))
507 die_errno("unable to make temporary pack file readable");
508 if (rename(pack_tmp_name, tmpname))
509 die_errno("unable to rename temporary pack file");
512 * Packs are runtime accessed in their mtime
513 * order since newer packs are more likely to contain
514 * younger objects. So if we are creating multiple
515 * packs then we should modify the mtime of later ones
516 * to preserve this property.
518 if (stat(tmpname, &st) < 0) {
519 warning("failed to stat %s: %s",
520 tmpname, strerror(errno));
521 } else if (!last_mtime) {
522 last_mtime = st.st_mtime;
523 } else {
524 struct utimbuf utb;
525 utb.actime = st.st_atime;
526 utb.modtime = --last_mtime;
527 if (utime(tmpname, &utb) < 0)
528 warning("failed utime() on %s: %s",
529 tmpname, strerror(errno));
532 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
533 base_name, sha1_to_hex(sha1));
534 if (adjust_shared_perm(idx_tmp_name))
535 die_errno("unable to make temporary index file readable");
536 if (rename(idx_tmp_name, tmpname))
537 die_errno("unable to rename temporary index file");
539 free((void *) idx_tmp_name);
540 free(pack_tmp_name);
541 puts(sha1_to_hex(sha1));
544 /* mark written objects as written to previous pack */
545 for (j = 0; j < nr_written; j++) {
546 written_list[j]->offset = (off_t)-1;
548 nr_remaining -= nr_written;
549 } while (nr_remaining && i < nr_objects);
551 free(written_list);
552 stop_progress(&progress_state);
553 if (written != nr_result)
554 die("wrote %"PRIu32" objects while expecting %"PRIu32,
555 written, nr_result);
558 static int locate_object_entry_hash(const unsigned char *sha1)
560 int i;
561 unsigned int ui;
562 memcpy(&ui, sha1, sizeof(unsigned int));
563 i = ui % object_ix_hashsz;
564 while (0 < object_ix[i]) {
565 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
566 return i;
567 if (++i == object_ix_hashsz)
568 i = 0;
570 return -1 - i;
573 static struct object_entry *locate_object_entry(const unsigned char *sha1)
575 int i;
577 if (!object_ix_hashsz)
578 return NULL;
580 i = locate_object_entry_hash(sha1);
581 if (0 <= i)
582 return &objects[object_ix[i]-1];
583 return NULL;
586 static void rehash_objects(void)
588 uint32_t i;
589 struct object_entry *oe;
591 object_ix_hashsz = nr_objects * 3;
592 if (object_ix_hashsz < 1024)
593 object_ix_hashsz = 1024;
594 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
595 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
596 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
597 int ix = locate_object_entry_hash(oe->idx.sha1);
598 if (0 <= ix)
599 continue;
600 ix = -1 - ix;
601 object_ix[ix] = i + 1;
605 static unsigned name_hash(const char *name)
607 unsigned c, hash = 0;
609 if (!name)
610 return 0;
613 * This effectively just creates a sortable number from the
614 * last sixteen non-whitespace characters. Last characters
615 * count "most", so things that end in ".c" sort together.
617 while ((c = *name++) != 0) {
618 if (isspace(c))
619 continue;
620 hash = (hash >> 2) + (c << 24);
622 return hash;
625 static void setup_delta_attr_check(struct git_attr_check *check)
627 static struct git_attr *attr_delta;
629 if (!attr_delta)
630 attr_delta = git_attr("delta");
632 check[0].attr = attr_delta;
635 static int no_try_delta(const char *path)
637 struct git_attr_check check[1];
639 setup_delta_attr_check(check);
640 if (git_checkattr(path, ARRAY_SIZE(check), check))
641 return 0;
642 if (ATTR_FALSE(check->value))
643 return 1;
644 return 0;
647 static int add_object_entry(const unsigned char *sha1, enum object_type type,
648 const char *name, int exclude)
650 struct object_entry *entry;
651 struct packed_git *p, *found_pack = NULL;
652 off_t found_offset = 0;
653 int ix;
654 unsigned hash = name_hash(name);
656 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
657 if (ix >= 0) {
658 if (exclude) {
659 entry = objects + object_ix[ix] - 1;
660 if (!entry->preferred_base)
661 nr_result--;
662 entry->preferred_base = 1;
664 return 0;
667 if (!exclude && local && has_loose_object_nonlocal(sha1))
668 return 0;
670 for (p = packed_git; p; p = p->next) {
671 off_t offset = find_pack_entry_one(sha1, p);
672 if (offset) {
673 if (!found_pack) {
674 found_offset = offset;
675 found_pack = p;
677 if (exclude)
678 break;
679 if (incremental)
680 return 0;
681 if (local && !p->pack_local)
682 return 0;
683 if (ignore_packed_keep && p->pack_local && p->pack_keep)
684 return 0;
688 if (nr_objects >= nr_alloc) {
689 nr_alloc = (nr_alloc + 1024) * 3 / 2;
690 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
693 entry = objects + nr_objects++;
694 memset(entry, 0, sizeof(*entry));
695 hashcpy(entry->idx.sha1, sha1);
696 entry->hash = hash;
697 if (type)
698 entry->type = type;
699 if (exclude)
700 entry->preferred_base = 1;
701 else
702 nr_result++;
703 if (found_pack) {
704 entry->in_pack = found_pack;
705 entry->in_pack_offset = found_offset;
708 if (object_ix_hashsz * 3 <= nr_objects * 4)
709 rehash_objects();
710 else
711 object_ix[-1 - ix] = nr_objects;
713 display_progress(progress_state, nr_objects);
715 if (name && no_try_delta(name))
716 entry->no_try_delta = 1;
718 return 1;
721 struct pbase_tree_cache {
722 unsigned char sha1[20];
723 int ref;
724 int temporary;
725 void *tree_data;
726 unsigned long tree_size;
729 static struct pbase_tree_cache *(pbase_tree_cache[256]);
730 static int pbase_tree_cache_ix(const unsigned char *sha1)
732 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
734 static int pbase_tree_cache_ix_incr(int ix)
736 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
739 static struct pbase_tree {
740 struct pbase_tree *next;
741 /* This is a phony "cache" entry; we are not
742 * going to evict it nor find it through _get()
743 * mechanism -- this is for the toplevel node that
744 * would almost always change with any commit.
746 struct pbase_tree_cache pcache;
747 } *pbase_tree;
749 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
751 struct pbase_tree_cache *ent, *nent;
752 void *data;
753 unsigned long size;
754 enum object_type type;
755 int neigh;
756 int my_ix = pbase_tree_cache_ix(sha1);
757 int available_ix = -1;
759 /* pbase-tree-cache acts as a limited hashtable.
760 * your object will be found at your index or within a few
761 * slots after that slot if it is cached.
763 for (neigh = 0; neigh < 8; neigh++) {
764 ent = pbase_tree_cache[my_ix];
765 if (ent && !hashcmp(ent->sha1, sha1)) {
766 ent->ref++;
767 return ent;
769 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
770 ((0 <= available_ix) &&
771 (!ent && pbase_tree_cache[available_ix])))
772 available_ix = my_ix;
773 if (!ent)
774 break;
775 my_ix = pbase_tree_cache_ix_incr(my_ix);
778 /* Did not find one. Either we got a bogus request or
779 * we need to read and perhaps cache.
781 data = read_sha1_file(sha1, &type, &size);
782 if (!data)
783 return NULL;
784 if (type != OBJ_TREE) {
785 free(data);
786 return NULL;
789 /* We need to either cache or return a throwaway copy */
791 if (available_ix < 0)
792 ent = NULL;
793 else {
794 ent = pbase_tree_cache[available_ix];
795 my_ix = available_ix;
798 if (!ent) {
799 nent = xmalloc(sizeof(*nent));
800 nent->temporary = (available_ix < 0);
802 else {
803 /* evict and reuse */
804 free(ent->tree_data);
805 nent = ent;
807 hashcpy(nent->sha1, sha1);
808 nent->tree_data = data;
809 nent->tree_size = size;
810 nent->ref = 1;
811 if (!nent->temporary)
812 pbase_tree_cache[my_ix] = nent;
813 return nent;
816 static void pbase_tree_put(struct pbase_tree_cache *cache)
818 if (!cache->temporary) {
819 cache->ref--;
820 return;
822 free(cache->tree_data);
823 free(cache);
826 static int name_cmp_len(const char *name)
828 int i;
829 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
831 return i;
834 static void add_pbase_object(struct tree_desc *tree,
835 const char *name,
836 int cmplen,
837 const char *fullname)
839 struct name_entry entry;
840 int cmp;
842 while (tree_entry(tree,&entry)) {
843 if (S_ISGITLINK(entry.mode))
844 continue;
845 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
846 memcmp(name, entry.path, cmplen);
847 if (cmp > 0)
848 continue;
849 if (cmp < 0)
850 return;
851 if (name[cmplen] != '/') {
852 add_object_entry(entry.sha1,
853 object_type(entry.mode),
854 fullname, 1);
855 return;
857 if (S_ISDIR(entry.mode)) {
858 struct tree_desc sub;
859 struct pbase_tree_cache *tree;
860 const char *down = name+cmplen+1;
861 int downlen = name_cmp_len(down);
863 tree = pbase_tree_get(entry.sha1);
864 if (!tree)
865 return;
866 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
868 add_pbase_object(&sub, down, downlen, fullname);
869 pbase_tree_put(tree);
874 static unsigned *done_pbase_paths;
875 static int done_pbase_paths_num;
876 static int done_pbase_paths_alloc;
877 static int done_pbase_path_pos(unsigned hash)
879 int lo = 0;
880 int hi = done_pbase_paths_num;
881 while (lo < hi) {
882 int mi = (hi + lo) / 2;
883 if (done_pbase_paths[mi] == hash)
884 return mi;
885 if (done_pbase_paths[mi] < hash)
886 hi = mi;
887 else
888 lo = mi + 1;
890 return -lo-1;
893 static int check_pbase_path(unsigned hash)
895 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
896 if (0 <= pos)
897 return 1;
898 pos = -pos - 1;
899 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
900 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
901 done_pbase_paths = xrealloc(done_pbase_paths,
902 done_pbase_paths_alloc *
903 sizeof(unsigned));
905 done_pbase_paths_num++;
906 if (pos < done_pbase_paths_num)
907 memmove(done_pbase_paths + pos + 1,
908 done_pbase_paths + pos,
909 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
910 done_pbase_paths[pos] = hash;
911 return 0;
914 static void add_preferred_base_object(const char *name)
916 struct pbase_tree *it;
917 int cmplen;
918 unsigned hash = name_hash(name);
920 if (!num_preferred_base || check_pbase_path(hash))
921 return;
923 cmplen = name_cmp_len(name);
924 for (it = pbase_tree; it; it = it->next) {
925 if (cmplen == 0) {
926 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
928 else {
929 struct tree_desc tree;
930 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
931 add_pbase_object(&tree, name, cmplen, name);
936 static void add_preferred_base(unsigned char *sha1)
938 struct pbase_tree *it;
939 void *data;
940 unsigned long size;
941 unsigned char tree_sha1[20];
943 if (window <= num_preferred_base++)
944 return;
946 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
947 if (!data)
948 return;
950 for (it = pbase_tree; it; it = it->next) {
951 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
952 free(data);
953 return;
957 it = xcalloc(1, sizeof(*it));
958 it->next = pbase_tree;
959 pbase_tree = it;
961 hashcpy(it->pcache.sha1, tree_sha1);
962 it->pcache.tree_data = data;
963 it->pcache.tree_size = size;
966 static void cleanup_preferred_base(void)
968 struct pbase_tree *it;
969 unsigned i;
971 it = pbase_tree;
972 pbase_tree = NULL;
973 while (it) {
974 struct pbase_tree *this = it;
975 it = this->next;
976 free(this->pcache.tree_data);
977 free(this);
980 for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
981 if (!pbase_tree_cache[i])
982 continue;
983 free(pbase_tree_cache[i]->tree_data);
984 free(pbase_tree_cache[i]);
985 pbase_tree_cache[i] = NULL;
988 free(done_pbase_paths);
989 done_pbase_paths = NULL;
990 done_pbase_paths_num = done_pbase_paths_alloc = 0;
993 static void check_object(struct object_entry *entry)
995 if (entry->in_pack) {
996 struct packed_git *p = entry->in_pack;
997 struct pack_window *w_curs = NULL;
998 const unsigned char *base_ref = NULL;
999 struct object_entry *base_entry;
1000 unsigned long used, used_0;
1001 unsigned int avail;
1002 off_t ofs;
1003 unsigned char *buf, c;
1005 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1008 * We want in_pack_type even if we do not reuse delta
1009 * since non-delta representations could still be reused.
1011 used = unpack_object_header_buffer(buf, avail,
1012 &entry->in_pack_type,
1013 &entry->size);
1014 if (used == 0)
1015 goto give_up;
1018 * Determine if this is a delta and if so whether we can
1019 * reuse it or not. Otherwise let's find out as cheaply as
1020 * possible what the actual type and size for this object is.
1022 switch (entry->in_pack_type) {
1023 default:
1024 /* Not a delta hence we've already got all we need. */
1025 entry->type = entry->in_pack_type;
1026 entry->in_pack_header_size = used;
1027 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1028 goto give_up;
1029 unuse_pack(&w_curs);
1030 return;
1031 case OBJ_REF_DELTA:
1032 if (reuse_delta && !entry->preferred_base)
1033 base_ref = use_pack(p, &w_curs,
1034 entry->in_pack_offset + used, NULL);
1035 entry->in_pack_header_size = used + 20;
1036 break;
1037 case OBJ_OFS_DELTA:
1038 buf = use_pack(p, &w_curs,
1039 entry->in_pack_offset + used, NULL);
1040 used_0 = 0;
1041 c = buf[used_0++];
1042 ofs = c & 127;
1043 while (c & 128) {
1044 ofs += 1;
1045 if (!ofs || MSB(ofs, 7)) {
1046 error("delta base offset overflow in pack for %s",
1047 sha1_to_hex(entry->idx.sha1));
1048 goto give_up;
1050 c = buf[used_0++];
1051 ofs = (ofs << 7) + (c & 127);
1053 ofs = entry->in_pack_offset - ofs;
1054 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1055 error("delta base offset out of bound for %s",
1056 sha1_to_hex(entry->idx.sha1));
1057 goto give_up;
1059 if (reuse_delta && !entry->preferred_base) {
1060 struct revindex_entry *revidx;
1061 revidx = find_pack_revindex(p, ofs);
1062 if (!revidx)
1063 goto give_up;
1064 base_ref = nth_packed_object_sha1(p, revidx->nr);
1066 entry->in_pack_header_size = used + used_0;
1067 break;
1070 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1072 * If base_ref was set above that means we wish to
1073 * reuse delta data, and we even found that base
1074 * in the list of objects we want to pack. Goodie!
1076 * Depth value does not matter - find_deltas() will
1077 * never consider reused delta as the base object to
1078 * deltify other objects against, in order to avoid
1079 * circular deltas.
1081 entry->type = entry->in_pack_type;
1082 entry->delta = base_entry;
1083 entry->delta_size = entry->size;
1084 entry->delta_sibling = base_entry->delta_child;
1085 base_entry->delta_child = entry;
1086 unuse_pack(&w_curs);
1087 return;
1090 if (entry->type) {
1092 * This must be a delta and we already know what the
1093 * final object type is. Let's extract the actual
1094 * object size from the delta header.
1096 entry->size = get_size_from_delta(p, &w_curs,
1097 entry->in_pack_offset + entry->in_pack_header_size);
1098 if (entry->size == 0)
1099 goto give_up;
1100 unuse_pack(&w_curs);
1101 return;
1105 * No choice but to fall back to the recursive delta walk
1106 * with sha1_object_info() to find about the object type
1107 * at this point...
1109 give_up:
1110 unuse_pack(&w_curs);
1113 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1115 * The error condition is checked in prepare_pack(). This is
1116 * to permit a missing preferred base object to be ignored
1117 * as a preferred base. Doing so can result in a larger
1118 * pack file, but the transfer will still take place.
1122 static int pack_offset_sort(const void *_a, const void *_b)
1124 const struct object_entry *a = *(struct object_entry **)_a;
1125 const struct object_entry *b = *(struct object_entry **)_b;
1127 /* avoid filesystem trashing with loose objects */
1128 if (!a->in_pack && !b->in_pack)
1129 return hashcmp(a->idx.sha1, b->idx.sha1);
1131 if (a->in_pack < b->in_pack)
1132 return -1;
1133 if (a->in_pack > b->in_pack)
1134 return 1;
1135 return a->in_pack_offset < b->in_pack_offset ? -1 :
1136 (a->in_pack_offset > b->in_pack_offset);
1139 static void get_object_details(void)
1141 uint32_t i;
1142 struct object_entry **sorted_by_offset;
1144 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1145 for (i = 0; i < nr_objects; i++)
1146 sorted_by_offset[i] = objects + i;
1147 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1149 for (i = 0; i < nr_objects; i++)
1150 check_object(sorted_by_offset[i]);
1152 free(sorted_by_offset);
1156 * We search for deltas in a list sorted by type, by filename hash, and then
1157 * by size, so that we see progressively smaller and smaller files.
1158 * That's because we prefer deltas to be from the bigger file
1159 * to the smaller -- deletes are potentially cheaper, but perhaps
1160 * more importantly, the bigger file is likely the more recent
1161 * one. The deepest deltas are therefore the oldest objects which are
1162 * less susceptible to be accessed often.
1164 static int type_size_sort(const void *_a, const void *_b)
1166 const struct object_entry *a = *(struct object_entry **)_a;
1167 const struct object_entry *b = *(struct object_entry **)_b;
1169 if (a->type > b->type)
1170 return -1;
1171 if (a->type < b->type)
1172 return 1;
1173 if (a->hash > b->hash)
1174 return -1;
1175 if (a->hash < b->hash)
1176 return 1;
1177 if (a->preferred_base > b->preferred_base)
1178 return -1;
1179 if (a->preferred_base < b->preferred_base)
1180 return 1;
1181 if (a->size > b->size)
1182 return -1;
1183 if (a->size < b->size)
1184 return 1;
1185 return a < b ? -1 : (a > b); /* newest first */
1188 struct unpacked {
1189 struct object_entry *entry;
1190 void *data;
1191 struct delta_index *index;
1192 unsigned depth;
1195 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1196 unsigned long delta_size)
1198 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1199 return 0;
1201 if (delta_size < cache_max_small_delta_size)
1202 return 1;
1204 /* cache delta, if objects are large enough compared to delta size */
1205 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1206 return 1;
1208 return 0;
1211 #ifndef NO_PTHREADS
1213 static pthread_mutex_t read_mutex;
1214 #define read_lock() pthread_mutex_lock(&read_mutex)
1215 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1217 static pthread_mutex_t cache_mutex;
1218 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1219 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1221 static pthread_mutex_t progress_mutex;
1222 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1223 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1225 #else
1227 #define read_lock() (void)0
1228 #define read_unlock() (void)0
1229 #define cache_lock() (void)0
1230 #define cache_unlock() (void)0
1231 #define progress_lock() (void)0
1232 #define progress_unlock() (void)0
1234 #endif
1236 static int try_delta(struct unpacked *trg, struct unpacked *src,
1237 unsigned max_depth, unsigned long *mem_usage)
1239 struct object_entry *trg_entry = trg->entry;
1240 struct object_entry *src_entry = src->entry;
1241 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1242 unsigned ref_depth;
1243 enum object_type type;
1244 void *delta_buf;
1246 /* Don't bother doing diffs between different types */
1247 if (trg_entry->type != src_entry->type)
1248 return -1;
1251 * We do not bother to try a delta that we discarded
1252 * on an earlier try, but only when reusing delta data.
1254 if (reuse_delta && trg_entry->in_pack &&
1255 trg_entry->in_pack == src_entry->in_pack &&
1256 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1257 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1258 return 0;
1260 /* Let's not bust the allowed depth. */
1261 if (src->depth >= max_depth)
1262 return 0;
1264 /* Now some size filtering heuristics. */
1265 trg_size = trg_entry->size;
1266 if (!trg_entry->delta) {
1267 max_size = trg_size/2 - 20;
1268 ref_depth = 1;
1269 } else {
1270 max_size = trg_entry->delta_size;
1271 ref_depth = trg->depth;
1273 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1274 (max_depth - ref_depth + 1);
1275 if (max_size == 0)
1276 return 0;
1277 src_size = src_entry->size;
1278 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1279 if (sizediff >= max_size)
1280 return 0;
1281 if (trg_size < src_size / 32)
1282 return 0;
1284 /* Load data if not already done */
1285 if (!trg->data) {
1286 read_lock();
1287 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1288 read_unlock();
1289 if (!trg->data)
1290 die("object %s cannot be read",
1291 sha1_to_hex(trg_entry->idx.sha1));
1292 if (sz != trg_size)
1293 die("object %s inconsistent object length (%lu vs %lu)",
1294 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1295 *mem_usage += sz;
1297 if (!src->data) {
1298 read_lock();
1299 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1300 read_unlock();
1301 if (!src->data)
1302 die("object %s cannot be read",
1303 sha1_to_hex(src_entry->idx.sha1));
1304 if (sz != src_size)
1305 die("object %s inconsistent object length (%lu vs %lu)",
1306 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1307 *mem_usage += sz;
1309 if (!src->index) {
1310 src->index = create_delta_index(src->data, src_size);
1311 if (!src->index) {
1312 static int warned = 0;
1313 if (!warned++)
1314 warning("suboptimal pack - out of memory");
1315 return 0;
1317 *mem_usage += sizeof_delta_index(src->index);
1320 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1321 if (!delta_buf)
1322 return 0;
1324 if (trg_entry->delta) {
1325 /* Prefer only shallower same-sized deltas. */
1326 if (delta_size == trg_entry->delta_size &&
1327 src->depth + 1 >= trg->depth) {
1328 free(delta_buf);
1329 return 0;
1334 * Handle memory allocation outside of the cache
1335 * accounting lock. Compiler will optimize the strangeness
1336 * away when NO_PTHREADS is defined.
1338 free(trg_entry->delta_data);
1339 cache_lock();
1340 if (trg_entry->delta_data) {
1341 delta_cache_size -= trg_entry->delta_size;
1342 trg_entry->delta_data = NULL;
1344 if (delta_cacheable(src_size, trg_size, delta_size)) {
1345 delta_cache_size += delta_size;
1346 cache_unlock();
1347 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1348 } else {
1349 cache_unlock();
1350 free(delta_buf);
1353 trg_entry->delta = src_entry;
1354 trg_entry->delta_size = delta_size;
1355 trg->depth = src->depth + 1;
1357 return 1;
1360 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1362 struct object_entry *child = me->delta_child;
1363 unsigned int m = n;
1364 while (child) {
1365 unsigned int c = check_delta_limit(child, n + 1);
1366 if (m < c)
1367 m = c;
1368 child = child->delta_sibling;
1370 return m;
1373 static unsigned long free_unpacked(struct unpacked *n)
1375 unsigned long freed_mem = sizeof_delta_index(n->index);
1376 free_delta_index(n->index);
1377 n->index = NULL;
1378 if (n->data) {
1379 freed_mem += n->entry->size;
1380 free(n->data);
1381 n->data = NULL;
1383 n->entry = NULL;
1384 n->depth = 0;
1385 return freed_mem;
1388 static void find_deltas(struct object_entry **list, unsigned *list_size,
1389 int window, int depth, unsigned *processed)
1391 uint32_t i, idx = 0, count = 0;
1392 struct unpacked *array;
1393 unsigned long mem_usage = 0;
1395 array = xcalloc(window, sizeof(struct unpacked));
1397 for (;;) {
1398 struct object_entry *entry;
1399 struct unpacked *n = array + idx;
1400 int j, max_depth, best_base = -1;
1402 progress_lock();
1403 if (!*list_size) {
1404 progress_unlock();
1405 break;
1407 entry = *list++;
1408 (*list_size)--;
1409 if (!entry->preferred_base) {
1410 (*processed)++;
1411 display_progress(progress_state, *processed);
1413 progress_unlock();
1415 mem_usage -= free_unpacked(n);
1416 n->entry = entry;
1418 while (window_memory_limit &&
1419 mem_usage > window_memory_limit &&
1420 count > 1) {
1421 uint32_t tail = (idx + window - count) % window;
1422 mem_usage -= free_unpacked(array + tail);
1423 count--;
1426 /* We do not compute delta to *create* objects we are not
1427 * going to pack.
1429 if (entry->preferred_base)
1430 goto next;
1433 * If the current object is at pack edge, take the depth the
1434 * objects that depend on the current object into account
1435 * otherwise they would become too deep.
1437 max_depth = depth;
1438 if (entry->delta_child) {
1439 max_depth -= check_delta_limit(entry, 0);
1440 if (max_depth <= 0)
1441 goto next;
1444 j = window;
1445 while (--j > 0) {
1446 int ret;
1447 uint32_t other_idx = idx + j;
1448 struct unpacked *m;
1449 if (other_idx >= window)
1450 other_idx -= window;
1451 m = array + other_idx;
1452 if (!m->entry)
1453 break;
1454 ret = try_delta(n, m, max_depth, &mem_usage);
1455 if (ret < 0)
1456 break;
1457 else if (ret > 0)
1458 best_base = other_idx;
1462 * If we decided to cache the delta data, then it is best
1463 * to compress it right away. First because we have to do
1464 * it anyway, and doing it here while we're threaded will
1465 * save a lot of time in the non threaded write phase,
1466 * as well as allow for caching more deltas within
1467 * the same cache size limit.
1468 * ...
1469 * But only if not writing to stdout, since in that case
1470 * the network is most likely throttling writes anyway,
1471 * and therefore it is best to go to the write phase ASAP
1472 * instead, as we can afford spending more time compressing
1473 * between writes at that moment.
1475 if (entry->delta_data && !pack_to_stdout) {
1476 entry->z_delta_size = do_compress(&entry->delta_data,
1477 entry->delta_size);
1478 cache_lock();
1479 delta_cache_size -= entry->delta_size;
1480 delta_cache_size += entry->z_delta_size;
1481 cache_unlock();
1484 /* if we made n a delta, and if n is already at max
1485 * depth, leaving it in the window is pointless. we
1486 * should evict it first.
1488 if (entry->delta && max_depth <= n->depth)
1489 continue;
1492 * Move the best delta base up in the window, after the
1493 * currently deltified object, to keep it longer. It will
1494 * be the first base object to be attempted next.
1496 if (entry->delta) {
1497 struct unpacked swap = array[best_base];
1498 int dist = (window + idx - best_base) % window;
1499 int dst = best_base;
1500 while (dist--) {
1501 int src = (dst + 1) % window;
1502 array[dst] = array[src];
1503 dst = src;
1505 array[dst] = swap;
1508 next:
1509 idx++;
1510 if (count + 1 < window)
1511 count++;
1512 if (idx >= window)
1513 idx = 0;
1516 for (i = 0; i < window; ++i) {
1517 free_delta_index(array[i].index);
1518 free(array[i].data);
1520 free(array);
1523 #ifndef NO_PTHREADS
1525 static void try_to_free_from_threads(size_t size)
1527 read_lock();
1528 release_pack_memory(size, -1);
1529 read_unlock();
1532 try_to_free_t old_try_to_free_routine;
1535 * The main thread waits on the condition that (at least) one of the workers
1536 * has stopped working (which is indicated in the .working member of
1537 * struct thread_params).
1538 * When a work thread has completed its work, it sets .working to 0 and
1539 * signals the main thread and waits on the condition that .data_ready
1540 * becomes 1.
1543 struct thread_params {
1544 pthread_t thread;
1545 struct object_entry **list;
1546 unsigned list_size;
1547 unsigned remaining;
1548 int window;
1549 int depth;
1550 int working;
1551 int data_ready;
1552 pthread_mutex_t mutex;
1553 pthread_cond_t cond;
1554 unsigned *processed;
1557 static pthread_cond_t progress_cond;
1560 * Mutex and conditional variable can't be statically-initialized on Windows.
1562 static void init_threaded_search(void)
1564 init_recursive_mutex(&read_mutex);
1565 pthread_mutex_init(&cache_mutex, NULL);
1566 pthread_mutex_init(&progress_mutex, NULL);
1567 pthread_cond_init(&progress_cond, NULL);
1568 old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1571 static void cleanup_threaded_search(void)
1573 set_try_to_free_routine(old_try_to_free_routine);
1574 pthread_cond_destroy(&progress_cond);
1575 pthread_mutex_destroy(&read_mutex);
1576 pthread_mutex_destroy(&cache_mutex);
1577 pthread_mutex_destroy(&progress_mutex);
1580 static void *threaded_find_deltas(void *arg)
1582 struct thread_params *me = arg;
1584 while (me->remaining) {
1585 find_deltas(me->list, &me->remaining,
1586 me->window, me->depth, me->processed);
1588 progress_lock();
1589 me->working = 0;
1590 pthread_cond_signal(&progress_cond);
1591 progress_unlock();
1594 * We must not set ->data_ready before we wait on the
1595 * condition because the main thread may have set it to 1
1596 * before we get here. In order to be sure that new
1597 * work is available if we see 1 in ->data_ready, it
1598 * was initialized to 0 before this thread was spawned
1599 * and we reset it to 0 right away.
1601 pthread_mutex_lock(&me->mutex);
1602 while (!me->data_ready)
1603 pthread_cond_wait(&me->cond, &me->mutex);
1604 me->data_ready = 0;
1605 pthread_mutex_unlock(&me->mutex);
1607 /* leave ->working 1 so that this doesn't get more work assigned */
1608 return NULL;
1611 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1612 int window, int depth, unsigned *processed)
1614 struct thread_params *p;
1615 int i, ret, active_threads = 0;
1617 init_threaded_search();
1619 if (!delta_search_threads) /* --threads=0 means autodetect */
1620 delta_search_threads = online_cpus();
1621 if (delta_search_threads <= 1) {
1622 find_deltas(list, &list_size, window, depth, processed);
1623 cleanup_threaded_search();
1624 return;
1626 if (progress > pack_to_stdout)
1627 fprintf(stderr, "Delta compression using up to %d threads.\n",
1628 delta_search_threads);
1629 p = xcalloc(delta_search_threads, sizeof(*p));
1631 /* Partition the work amongst work threads. */
1632 for (i = 0; i < delta_search_threads; i++) {
1633 unsigned sub_size = list_size / (delta_search_threads - i);
1635 /* don't use too small segments or no deltas will be found */
1636 if (sub_size < 2*window && i+1 < delta_search_threads)
1637 sub_size = 0;
1639 p[i].window = window;
1640 p[i].depth = depth;
1641 p[i].processed = processed;
1642 p[i].working = 1;
1643 p[i].data_ready = 0;
1645 /* try to split chunks on "path" boundaries */
1646 while (sub_size && sub_size < list_size &&
1647 list[sub_size]->hash &&
1648 list[sub_size]->hash == list[sub_size-1]->hash)
1649 sub_size++;
1651 p[i].list = list;
1652 p[i].list_size = sub_size;
1653 p[i].remaining = sub_size;
1655 list += sub_size;
1656 list_size -= sub_size;
1659 /* Start work threads. */
1660 for (i = 0; i < delta_search_threads; i++) {
1661 if (!p[i].list_size)
1662 continue;
1663 pthread_mutex_init(&p[i].mutex, NULL);
1664 pthread_cond_init(&p[i].cond, NULL);
1665 ret = pthread_create(&p[i].thread, NULL,
1666 threaded_find_deltas, &p[i]);
1667 if (ret)
1668 die("unable to create thread: %s", strerror(ret));
1669 active_threads++;
1673 * Now let's wait for work completion. Each time a thread is done
1674 * with its work, we steal half of the remaining work from the
1675 * thread with the largest number of unprocessed objects and give
1676 * it to that newly idle thread. This ensure good load balancing
1677 * until the remaining object list segments are simply too short
1678 * to be worth splitting anymore.
1680 while (active_threads) {
1681 struct thread_params *target = NULL;
1682 struct thread_params *victim = NULL;
1683 unsigned sub_size = 0;
1685 progress_lock();
1686 for (;;) {
1687 for (i = 0; !target && i < delta_search_threads; i++)
1688 if (!p[i].working)
1689 target = &p[i];
1690 if (target)
1691 break;
1692 pthread_cond_wait(&progress_cond, &progress_mutex);
1695 for (i = 0; i < delta_search_threads; i++)
1696 if (p[i].remaining > 2*window &&
1697 (!victim || victim->remaining < p[i].remaining))
1698 victim = &p[i];
1699 if (victim) {
1700 sub_size = victim->remaining / 2;
1701 list = victim->list + victim->list_size - sub_size;
1702 while (sub_size && list[0]->hash &&
1703 list[0]->hash == list[-1]->hash) {
1704 list++;
1705 sub_size--;
1707 if (!sub_size) {
1709 * It is possible for some "paths" to have
1710 * so many objects that no hash boundary
1711 * might be found. Let's just steal the
1712 * exact half in that case.
1714 sub_size = victim->remaining / 2;
1715 list -= sub_size;
1717 target->list = list;
1718 victim->list_size -= sub_size;
1719 victim->remaining -= sub_size;
1721 target->list_size = sub_size;
1722 target->remaining = sub_size;
1723 target->working = 1;
1724 progress_unlock();
1726 pthread_mutex_lock(&target->mutex);
1727 target->data_ready = 1;
1728 pthread_cond_signal(&target->cond);
1729 pthread_mutex_unlock(&target->mutex);
1731 if (!sub_size) {
1732 pthread_join(target->thread, NULL);
1733 pthread_cond_destroy(&target->cond);
1734 pthread_mutex_destroy(&target->mutex);
1735 active_threads--;
1738 cleanup_threaded_search();
1739 free(p);
1742 #else
1743 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1744 #endif
1746 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1748 unsigned char peeled[20];
1750 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1751 !peel_ref(path, peeled) && /* peelable? */
1752 !is_null_sha1(peeled) && /* annotated tag? */
1753 locate_object_entry(peeled)) /* object packed? */
1754 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1755 return 0;
1758 static void prepare_pack(int window, int depth)
1760 struct object_entry **delta_list;
1761 uint32_t i, nr_deltas;
1762 unsigned n;
1764 get_object_details();
1767 * If we're locally repacking then we need to be doubly careful
1768 * from now on in order to make sure no stealth corruption gets
1769 * propagated to the new pack. Clients receiving streamed packs
1770 * should validate everything they get anyway so no need to incur
1771 * the additional cost here in that case.
1773 if (!pack_to_stdout)
1774 do_check_packed_object_crc = 1;
1776 if (!nr_objects || !window || !depth)
1777 return;
1779 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1780 nr_deltas = n = 0;
1782 for (i = 0; i < nr_objects; i++) {
1783 struct object_entry *entry = objects + i;
1785 if (entry->delta)
1786 /* This happens if we decided to reuse existing
1787 * delta from a pack. "reuse_delta &&" is implied.
1789 continue;
1791 if (entry->size < 50)
1792 continue;
1794 if (entry->no_try_delta)
1795 continue;
1797 if (!entry->preferred_base) {
1798 nr_deltas++;
1799 if (entry->type < 0)
1800 die("unable to get type of object %s",
1801 sha1_to_hex(entry->idx.sha1));
1802 } else {
1803 if (entry->type < 0) {
1805 * This object is not found, but we
1806 * don't have to include it anyway.
1808 continue;
1812 delta_list[n++] = entry;
1815 if (nr_deltas && n > 1) {
1816 unsigned nr_done = 0;
1817 if (progress)
1818 progress_state = start_progress("Compressing objects",
1819 nr_deltas);
1820 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1821 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1822 stop_progress(&progress_state);
1823 if (nr_done != nr_deltas)
1824 die("inconsistency with delta count");
1826 free(delta_list);
1829 static int git_pack_config(const char *k, const char *v, void *cb)
1831 if (!strcmp(k, "pack.window")) {
1832 window = git_config_int(k, v);
1833 return 0;
1835 if (!strcmp(k, "pack.windowmemory")) {
1836 window_memory_limit = git_config_ulong(k, v);
1837 return 0;
1839 if (!strcmp(k, "pack.depth")) {
1840 depth = git_config_int(k, v);
1841 return 0;
1843 if (!strcmp(k, "pack.compression")) {
1844 int level = git_config_int(k, v);
1845 if (level == -1)
1846 level = Z_DEFAULT_COMPRESSION;
1847 else if (level < 0 || level > Z_BEST_COMPRESSION)
1848 die("bad pack compression level %d", level);
1849 pack_compression_level = level;
1850 pack_compression_seen = 1;
1851 return 0;
1853 if (!strcmp(k, "pack.deltacachesize")) {
1854 max_delta_cache_size = git_config_int(k, v);
1855 return 0;
1857 if (!strcmp(k, "pack.deltacachelimit")) {
1858 cache_max_small_delta_size = git_config_int(k, v);
1859 return 0;
1861 if (!strcmp(k, "pack.threads")) {
1862 delta_search_threads = git_config_int(k, v);
1863 if (delta_search_threads < 0)
1864 die("invalid number of threads specified (%d)",
1865 delta_search_threads);
1866 #ifdef NO_PTHREADS
1867 if (delta_search_threads != 1)
1868 warning("no threads support, ignoring %s", k);
1869 #endif
1870 return 0;
1872 if (!strcmp(k, "pack.indexversion")) {
1873 pack_idx_default_version = git_config_int(k, v);
1874 if (pack_idx_default_version > 2)
1875 die("bad pack.indexversion=%"PRIu32,
1876 pack_idx_default_version);
1877 return 0;
1879 if (!strcmp(k, "pack.packsizelimit")) {
1880 pack_size_limit_cfg = git_config_ulong(k, v);
1881 return 0;
1883 return git_default_config(k, v, cb);
1886 static void read_object_list_from_stdin(void)
1888 char line[40 + 1 + PATH_MAX + 2];
1889 unsigned char sha1[20];
1891 for (;;) {
1892 if (!fgets(line, sizeof(line), stdin)) {
1893 if (feof(stdin))
1894 break;
1895 if (!ferror(stdin))
1896 die("fgets returned NULL, not EOF, not error!");
1897 if (errno != EINTR)
1898 die_errno("fgets");
1899 clearerr(stdin);
1900 continue;
1902 if (line[0] == '-') {
1903 if (get_sha1_hex(line+1, sha1))
1904 die("expected edge sha1, got garbage:\n %s",
1905 line);
1906 add_preferred_base(sha1);
1907 continue;
1909 if (get_sha1_hex(line, sha1))
1910 die("expected sha1, got garbage:\n %s", line);
1912 add_preferred_base_object(line+41);
1913 add_object_entry(sha1, 0, line+41, 0);
1917 #define OBJECT_ADDED (1u<<20)
1919 static void show_commit(struct commit *commit, void *data)
1921 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1922 commit->object.flags |= OBJECT_ADDED;
1925 static void show_object(struct object *obj, const struct name_path *path, const char *last)
1927 char *name = path_name(path, last);
1929 add_preferred_base_object(name);
1930 add_object_entry(obj->sha1, obj->type, name, 0);
1931 obj->flags |= OBJECT_ADDED;
1934 * We will have generated the hash from the name,
1935 * but not saved a pointer to it - we can free it
1937 free((char *)name);
1940 static void show_edge(struct commit *commit)
1942 add_preferred_base(commit->object.sha1);
1945 struct in_pack_object {
1946 off_t offset;
1947 struct object *object;
1950 struct in_pack {
1951 int alloc;
1952 int nr;
1953 struct in_pack_object *array;
1956 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1958 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1959 in_pack->array[in_pack->nr].object = object;
1960 in_pack->nr++;
1964 * Compare the objects in the offset order, in order to emulate the
1965 * "git rev-list --objects" output that produced the pack originally.
1967 static int ofscmp(const void *a_, const void *b_)
1969 struct in_pack_object *a = (struct in_pack_object *)a_;
1970 struct in_pack_object *b = (struct in_pack_object *)b_;
1972 if (a->offset < b->offset)
1973 return -1;
1974 else if (a->offset > b->offset)
1975 return 1;
1976 else
1977 return hashcmp(a->object->sha1, b->object->sha1);
1980 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1982 struct packed_git *p;
1983 struct in_pack in_pack;
1984 uint32_t i;
1986 memset(&in_pack, 0, sizeof(in_pack));
1988 for (p = packed_git; p; p = p->next) {
1989 const unsigned char *sha1;
1990 struct object *o;
1992 if (!p->pack_local || p->pack_keep)
1993 continue;
1994 if (open_pack_index(p))
1995 die("cannot open pack index");
1997 ALLOC_GROW(in_pack.array,
1998 in_pack.nr + p->num_objects,
1999 in_pack.alloc);
2001 for (i = 0; i < p->num_objects; i++) {
2002 sha1 = nth_packed_object_sha1(p, i);
2003 o = lookup_unknown_object(sha1);
2004 if (!(o->flags & OBJECT_ADDED))
2005 mark_in_pack_object(o, p, &in_pack);
2006 o->flags |= OBJECT_ADDED;
2010 if (in_pack.nr) {
2011 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2012 ofscmp);
2013 for (i = 0; i < in_pack.nr; i++) {
2014 struct object *o = in_pack.array[i].object;
2015 add_object_entry(o->sha1, o->type, "", 0);
2018 free(in_pack.array);
2021 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2023 static struct packed_git *last_found = (void *)1;
2024 struct packed_git *p;
2026 p = (last_found != (void *)1) ? last_found : packed_git;
2028 while (p) {
2029 if ((!p->pack_local || p->pack_keep) &&
2030 find_pack_entry_one(sha1, p)) {
2031 last_found = p;
2032 return 1;
2034 if (p == last_found)
2035 p = packed_git;
2036 else
2037 p = p->next;
2038 if (p == last_found)
2039 p = p->next;
2041 return 0;
2044 static void loosen_unused_packed_objects(struct rev_info *revs)
2046 struct packed_git *p;
2047 uint32_t i;
2048 const unsigned char *sha1;
2050 for (p = packed_git; p; p = p->next) {
2051 if (!p->pack_local || p->pack_keep)
2052 continue;
2054 if (open_pack_index(p))
2055 die("cannot open pack index");
2057 for (i = 0; i < p->num_objects; i++) {
2058 sha1 = nth_packed_object_sha1(p, i);
2059 if (!locate_object_entry(sha1) &&
2060 !has_sha1_pack_kept_or_nonlocal(sha1))
2061 if (force_object_loose(sha1, p->mtime))
2062 die("unable to force loose object");
2067 static void get_object_list(int ac, const char **av)
2069 struct rev_info revs;
2070 char line[1000];
2071 int flags = 0;
2073 init_revisions(&revs, NULL);
2074 save_commit_buffer = 0;
2075 setup_revisions(ac, av, &revs, NULL);
2077 while (fgets(line, sizeof(line), stdin) != NULL) {
2078 int len = strlen(line);
2079 if (len && line[len - 1] == '\n')
2080 line[--len] = 0;
2081 if (!len)
2082 break;
2083 if (*line == '-') {
2084 if (!strcmp(line, "--not")) {
2085 flags ^= UNINTERESTING;
2086 continue;
2088 die("not a rev '%s'", line);
2090 if (handle_revision_arg(line, &revs, flags, 1))
2091 die("bad revision '%s'", line);
2094 if (prepare_revision_walk(&revs))
2095 die("revision walk setup failed");
2096 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2097 traverse_commit_list(&revs, show_commit, show_object, NULL);
2099 if (keep_unreachable)
2100 add_objects_in_unpacked_packs(&revs);
2101 if (unpack_unreachable)
2102 loosen_unused_packed_objects(&revs);
2105 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2107 int use_internal_rev_list = 0;
2108 int thin = 0;
2109 int all_progress_implied = 0;
2110 uint32_t i;
2111 const char **rp_av;
2112 int rp_ac_alloc = 64;
2113 int rp_ac;
2115 read_replace_refs = 0;
2117 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2119 rp_av[0] = "pack-objects";
2120 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2121 rp_ac = 2;
2123 git_config(git_pack_config, NULL);
2124 if (!pack_compression_seen && core_compression_seen)
2125 pack_compression_level = core_compression_level;
2127 progress = isatty(2);
2128 for (i = 1; i < argc; i++) {
2129 const char *arg = argv[i];
2131 if (*arg != '-')
2132 break;
2134 if (!strcmp("--non-empty", arg)) {
2135 non_empty = 1;
2136 continue;
2138 if (!strcmp("--local", arg)) {
2139 local = 1;
2140 continue;
2142 if (!strcmp("--incremental", arg)) {
2143 incremental = 1;
2144 continue;
2146 if (!strcmp("--honor-pack-keep", arg)) {
2147 ignore_packed_keep = 1;
2148 continue;
2150 if (!prefixcmp(arg, "--compression=")) {
2151 char *end;
2152 int level = strtoul(arg+14, &end, 0);
2153 if (!arg[14] || *end)
2154 usage(pack_usage);
2155 if (level == -1)
2156 level = Z_DEFAULT_COMPRESSION;
2157 else if (level < 0 || level > Z_BEST_COMPRESSION)
2158 die("bad pack compression level %d", level);
2159 pack_compression_level = level;
2160 continue;
2162 if (!prefixcmp(arg, "--max-pack-size=")) {
2163 pack_size_limit_cfg = 0;
2164 if (!git_parse_ulong(arg+16, &pack_size_limit))
2165 usage(pack_usage);
2166 continue;
2168 if (!prefixcmp(arg, "--window=")) {
2169 char *end;
2170 window = strtoul(arg+9, &end, 0);
2171 if (!arg[9] || *end)
2172 usage(pack_usage);
2173 continue;
2175 if (!prefixcmp(arg, "--window-memory=")) {
2176 if (!git_parse_ulong(arg+16, &window_memory_limit))
2177 usage(pack_usage);
2178 continue;
2180 if (!prefixcmp(arg, "--threads=")) {
2181 char *end;
2182 delta_search_threads = strtoul(arg+10, &end, 0);
2183 if (!arg[10] || *end || delta_search_threads < 0)
2184 usage(pack_usage);
2185 #ifdef NO_PTHREADS
2186 if (delta_search_threads != 1)
2187 warning("no threads support, "
2188 "ignoring %s", arg);
2189 #endif
2190 continue;
2192 if (!prefixcmp(arg, "--depth=")) {
2193 char *end;
2194 depth = strtoul(arg+8, &end, 0);
2195 if (!arg[8] || *end)
2196 usage(pack_usage);
2197 continue;
2199 if (!strcmp("--progress", arg)) {
2200 progress = 1;
2201 continue;
2203 if (!strcmp("--all-progress", arg)) {
2204 progress = 2;
2205 continue;
2207 if (!strcmp("--all-progress-implied", arg)) {
2208 all_progress_implied = 1;
2209 continue;
2211 if (!strcmp("-q", arg)) {
2212 progress = 0;
2213 continue;
2215 if (!strcmp("--no-reuse-delta", arg)) {
2216 reuse_delta = 0;
2217 continue;
2219 if (!strcmp("--no-reuse-object", arg)) {
2220 reuse_object = reuse_delta = 0;
2221 continue;
2223 if (!strcmp("--delta-base-offset", arg)) {
2224 allow_ofs_delta = 1;
2225 continue;
2227 if (!strcmp("--stdout", arg)) {
2228 pack_to_stdout = 1;
2229 continue;
2231 if (!strcmp("--revs", arg)) {
2232 use_internal_rev_list = 1;
2233 continue;
2235 if (!strcmp("--keep-unreachable", arg)) {
2236 keep_unreachable = 1;
2237 continue;
2239 if (!strcmp("--unpack-unreachable", arg)) {
2240 unpack_unreachable = 1;
2241 continue;
2243 if (!strcmp("--include-tag", arg)) {
2244 include_tag = 1;
2245 continue;
2247 if (!strcmp("--unpacked", arg) ||
2248 !strcmp("--reflog", arg) ||
2249 !strcmp("--all", arg)) {
2250 use_internal_rev_list = 1;
2251 if (rp_ac >= rp_ac_alloc - 1) {
2252 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2253 rp_av = xrealloc(rp_av,
2254 rp_ac_alloc * sizeof(*rp_av));
2256 rp_av[rp_ac++] = arg;
2257 continue;
2259 if (!strcmp("--thin", arg)) {
2260 use_internal_rev_list = 1;
2261 thin = 1;
2262 rp_av[1] = "--objects-edge";
2263 continue;
2265 if (!prefixcmp(arg, "--index-version=")) {
2266 char *c;
2267 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2268 if (pack_idx_default_version > 2)
2269 die("bad %s", arg);
2270 if (*c == ',')
2271 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2272 if (*c || pack_idx_off32_limit & 0x80000000)
2273 die("bad %s", arg);
2274 continue;
2276 if (!strcmp(arg, "--keep-true-parents")) {
2277 grafts_replace_parents = 0;
2278 continue;
2280 usage(pack_usage);
2283 /* Traditionally "pack-objects [options] base extra" failed;
2284 * we would however want to take refs parameter that would
2285 * have been given to upstream rev-list ourselves, which means
2286 * we somehow want to say what the base name is. So the
2287 * syntax would be:
2289 * pack-objects [options] base <refs...>
2291 * in other words, we would treat the first non-option as the
2292 * base_name and send everything else to the internal revision
2293 * walker.
2296 if (!pack_to_stdout)
2297 base_name = argv[i++];
2299 if (pack_to_stdout != !base_name)
2300 usage(pack_usage);
2302 if (!pack_to_stdout && !pack_size_limit)
2303 pack_size_limit = pack_size_limit_cfg;
2304 if (pack_to_stdout && pack_size_limit)
2305 die("--max-pack-size cannot be used to build a pack for transfer.");
2306 if (pack_size_limit && pack_size_limit < 1024*1024) {
2307 warning("minimum pack size limit is 1 MiB");
2308 pack_size_limit = 1024*1024;
2311 if (!pack_to_stdout && thin)
2312 die("--thin cannot be used to build an indexable pack.");
2314 if (keep_unreachable && unpack_unreachable)
2315 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2317 if (progress && all_progress_implied)
2318 progress = 2;
2320 prepare_packed_git();
2322 if (progress)
2323 progress_state = start_progress("Counting objects", 0);
2324 if (!use_internal_rev_list)
2325 read_object_list_from_stdin();
2326 else {
2327 rp_av[rp_ac] = NULL;
2328 get_object_list(rp_ac, rp_av);
2330 cleanup_preferred_base();
2331 if (include_tag && nr_result)
2332 for_each_ref(add_ref_tag, NULL);
2333 stop_progress(&progress_state);
2335 if (non_empty && !nr_result)
2336 return 0;
2337 if (nr_result)
2338 prepare_pack(window, depth);
2339 write_pack_file();
2340 if (progress)
2341 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2342 " reused %"PRIu32" (delta %"PRIu32")\n",
2343 written, written_delta, reused, reused_delta);
2344 return 0;