MinGW: 64-bit file offsets
[git/platforms.git] / builtin-pack-objects.c
blobcb51916fe33a904b8ab11b7c43dfe3a84ec11705
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
20 #ifdef THREADED_DELTA_SEARCH
21 #include "thread-utils.h"
22 #include <pthread.h>
23 #endif
25 static const char pack_usage[] = "\
26 git pack-objects [{ -q | --progress | --all-progress }] \n\
27 [--max-pack-size=N] [--local] [--incremental] \n\
28 [--window=N] [--window-memory=N] [--depth=N] \n\
29 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
30 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
31 [--stdout | base-name] [--include-tag] \n\
32 [--keep-unreachable | --unpack-unreachable] \n\
33 [<ref-list | <object-list]";
35 struct object_entry {
36 struct pack_idx_entry idx;
37 unsigned long size; /* uncompressed size */
38 struct packed_git *in_pack; /* already in pack */
39 off_t in_pack_offset;
40 struct object_entry *delta; /* delta base object */
41 struct object_entry *delta_child; /* deltified objects who bases me */
42 struct object_entry *delta_sibling; /* other deltified objects who
43 * uses the same base as me
45 void *delta_data; /* cached delta (uncompressed) */
46 unsigned long delta_size; /* delta data size (uncompressed) */
47 unsigned long z_delta_size; /* delta data size (compressed) */
48 unsigned int hash; /* name hint hash */
49 enum object_type type;
50 enum object_type in_pack_type; /* could be delta */
51 unsigned char in_pack_header_size;
52 unsigned char preferred_base; /* we do not pack this, but is available
53 * to be used as the base object to delta
54 * objects against.
56 unsigned char no_try_delta;
60 * Objects we are going to pack are collected in objects array (dynamically
61 * expanded). nr_objects & nr_alloc controls this array. They are stored
62 * in the order we see -- typically rev-list --objects order that gives us
63 * nice "minimum seek" order.
65 static struct object_entry *objects;
66 static struct pack_idx_entry **written_list;
67 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
69 static int non_empty;
70 static int reuse_delta = 1, reuse_object = 1;
71 static int keep_unreachable, unpack_unreachable, include_tag;
72 static int local;
73 static int incremental;
74 static int ignore_packed_keep;
75 static int allow_ofs_delta;
76 static const char *base_name;
77 static int progress = 1;
78 static int window = 10;
79 static uint32_t pack_size_limit, pack_size_limit_cfg;
80 static int depth = 50;
81 static int delta_search_threads;
82 static int pack_to_stdout;
83 static int num_preferred_base;
84 static struct progress *progress_state;
85 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
86 static int pack_compression_seen;
88 static unsigned long delta_cache_size = 0;
89 static unsigned long max_delta_cache_size = 0;
90 static unsigned long cache_max_small_delta_size = 1000;
92 static unsigned long window_memory_limit = 0;
95 * The object names in objects array are hashed with this hashtable,
96 * to help looking up the entry by object name.
97 * This hashtable is built after all the objects are seen.
99 static int *object_ix;
100 static int object_ix_hashsz;
103 * stats
105 static uint32_t written, written_delta;
106 static uint32_t reused, reused_delta;
109 static void *get_delta(struct object_entry *entry)
111 unsigned long size, base_size, delta_size;
112 void *buf, *base_buf, *delta_buf;
113 enum object_type type;
115 buf = read_sha1_file(entry->idx.sha1, &type, &size);
116 if (!buf)
117 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
118 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
119 if (!base_buf)
120 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
121 delta_buf = diff_delta(base_buf, base_size,
122 buf, size, &delta_size, 0);
123 if (!delta_buf || delta_size != entry->delta_size)
124 die("delta size changed");
125 free(buf);
126 free(base_buf);
127 return delta_buf;
130 static unsigned long do_compress(void **pptr, unsigned long size)
132 z_stream stream;
133 void *in, *out;
134 unsigned long maxsize;
136 memset(&stream, 0, sizeof(stream));
137 deflateInit(&stream, pack_compression_level);
138 maxsize = deflateBound(&stream, size);
140 in = *pptr;
141 out = xmalloc(maxsize);
142 *pptr = out;
144 stream.next_in = in;
145 stream.avail_in = size;
146 stream.next_out = out;
147 stream.avail_out = maxsize;
148 while (deflate(&stream, Z_FINISH) == Z_OK)
149 ; /* nothing */
150 deflateEnd(&stream);
152 free(in);
153 return stream.total_out;
157 * The per-object header is a pretty dense thing, which is
158 * - first byte: low four bits are "size", then three bits of "type",
159 * and the high bit is "size continues".
160 * - each byte afterwards: low seven bits are size continuation,
161 * with the high bit being "size continues"
163 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
165 int n = 1;
166 unsigned char c;
168 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
169 die("bad type %d", type);
171 c = (type << 4) | (size & 15);
172 size >>= 4;
173 while (size) {
174 *hdr++ = c | 0x80;
175 c = size & 0x7f;
176 size >>= 7;
177 n++;
179 *hdr = c;
180 return n;
184 * we are going to reuse the existing object data as is. make
185 * sure it is not corrupt.
187 static int check_pack_inflate(struct packed_git *p,
188 struct pack_window **w_curs,
189 off_t offset,
190 off_t len,
191 unsigned long expect)
193 z_stream stream;
194 unsigned char fakebuf[4096], *in;
195 int st;
197 memset(&stream, 0, sizeof(stream));
198 git_inflate_init(&stream);
199 do {
200 in = use_pack(p, w_curs, offset, &stream.avail_in);
201 stream.next_in = in;
202 stream.next_out = fakebuf;
203 stream.avail_out = sizeof(fakebuf);
204 st = git_inflate(&stream, Z_FINISH);
205 offset += stream.next_in - in;
206 } while (st == Z_OK || st == Z_BUF_ERROR);
207 git_inflate_end(&stream);
208 return (st == Z_STREAM_END &&
209 stream.total_out == expect &&
210 stream.total_in == len) ? 0 : -1;
213 static void copy_pack_data(struct sha1file *f,
214 struct packed_git *p,
215 struct pack_window **w_curs,
216 off_t offset,
217 off_t len)
219 unsigned char *in;
220 unsigned int avail;
222 while (len) {
223 in = use_pack(p, w_curs, offset, &avail);
224 if (avail > len)
225 avail = (unsigned int)len;
226 sha1write(f, in, avail);
227 offset += avail;
228 len -= avail;
232 static unsigned long write_object(struct sha1file *f,
233 struct object_entry *entry,
234 off_t write_offset)
236 unsigned long size, limit, datalen;
237 void *buf;
238 unsigned char header[10], dheader[10];
239 unsigned hdrlen;
240 enum object_type type;
241 int usable_delta, to_reuse;
243 if (!pack_to_stdout)
244 crc32_begin(f);
246 type = entry->type;
248 /* write limit if limited packsize and not first object */
249 if (!pack_size_limit || !nr_written)
250 limit = 0;
251 else if (pack_size_limit <= write_offset)
253 * the earlier object did not fit the limit; avoid
254 * mistaking this with unlimited (i.e. limit = 0).
256 limit = 1;
257 else
258 limit = pack_size_limit - write_offset;
260 if (!entry->delta)
261 usable_delta = 0; /* no delta */
262 else if (!pack_size_limit)
263 usable_delta = 1; /* unlimited packfile */
264 else if (entry->delta->idx.offset == (off_t)-1)
265 usable_delta = 0; /* base was written to another pack */
266 else if (entry->delta->idx.offset)
267 usable_delta = 1; /* base already exists in this pack */
268 else
269 usable_delta = 0; /* base could end up in another pack */
271 if (!reuse_object)
272 to_reuse = 0; /* explicit */
273 else if (!entry->in_pack)
274 to_reuse = 0; /* can't reuse what we don't have */
275 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
276 /* check_object() decided it for us ... */
277 to_reuse = usable_delta;
278 /* ... but pack split may override that */
279 else if (type != entry->in_pack_type)
280 to_reuse = 0; /* pack has delta which is unusable */
281 else if (entry->delta)
282 to_reuse = 0; /* we want to pack afresh */
283 else
284 to_reuse = 1; /* we have it in-pack undeltified,
285 * and we do not need to deltify it.
288 if (!to_reuse) {
289 no_reuse:
290 if (!usable_delta) {
291 buf = read_sha1_file(entry->idx.sha1, &type, &size);
292 if (!buf)
293 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
295 * make sure no cached delta data remains from a
296 * previous attempt before a pack split occured.
298 free(entry->delta_data);
299 entry->delta_data = NULL;
300 entry->z_delta_size = 0;
301 } else if (entry->delta_data) {
302 size = entry->delta_size;
303 buf = entry->delta_data;
304 entry->delta_data = NULL;
305 type = (allow_ofs_delta && entry->delta->idx.offset) ?
306 OBJ_OFS_DELTA : OBJ_REF_DELTA;
307 } else {
308 buf = get_delta(entry);
309 size = entry->delta_size;
310 type = (allow_ofs_delta && entry->delta->idx.offset) ?
311 OBJ_OFS_DELTA : OBJ_REF_DELTA;
314 if (entry->z_delta_size)
315 datalen = entry->z_delta_size;
316 else
317 datalen = do_compress(&buf, size);
320 * The object header is a byte of 'type' followed by zero or
321 * more bytes of length.
323 hdrlen = encode_header(type, size, header);
325 if (type == OBJ_OFS_DELTA) {
327 * Deltas with relative base contain an additional
328 * encoding of the relative offset for the delta
329 * base from this object's position in the pack.
331 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
332 unsigned pos = sizeof(dheader) - 1;
333 dheader[pos] = ofs & 127;
334 while (ofs >>= 7)
335 dheader[--pos] = 128 | (--ofs & 127);
336 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
337 free(buf);
338 return 0;
340 sha1write(f, header, hdrlen);
341 sha1write(f, dheader + pos, sizeof(dheader) - pos);
342 hdrlen += sizeof(dheader) - pos;
343 } else if (type == OBJ_REF_DELTA) {
345 * Deltas with a base reference contain
346 * an additional 20 bytes for the base sha1.
348 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
349 free(buf);
350 return 0;
352 sha1write(f, header, hdrlen);
353 sha1write(f, entry->delta->idx.sha1, 20);
354 hdrlen += 20;
355 } else {
356 if (limit && hdrlen + datalen + 20 >= limit) {
357 free(buf);
358 return 0;
360 sha1write(f, header, hdrlen);
362 sha1write(f, buf, datalen);
363 free(buf);
365 else {
366 struct packed_git *p = entry->in_pack;
367 struct pack_window *w_curs = NULL;
368 struct revindex_entry *revidx;
369 off_t offset;
371 if (entry->delta)
372 type = (allow_ofs_delta && entry->delta->idx.offset) ?
373 OBJ_OFS_DELTA : OBJ_REF_DELTA;
374 hdrlen = encode_header(type, entry->size, header);
376 offset = entry->in_pack_offset;
377 revidx = find_pack_revindex(p, offset);
378 datalen = revidx[1].offset - offset;
379 if (!pack_to_stdout && p->index_version > 1 &&
380 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
381 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
382 unuse_pack(&w_curs);
383 goto no_reuse;
386 offset += entry->in_pack_header_size;
387 datalen -= entry->in_pack_header_size;
388 if (!pack_to_stdout && p->index_version == 1 &&
389 check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
390 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
391 unuse_pack(&w_curs);
392 goto no_reuse;
395 if (type == OBJ_OFS_DELTA) {
396 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
397 unsigned pos = sizeof(dheader) - 1;
398 dheader[pos] = ofs & 127;
399 while (ofs >>= 7)
400 dheader[--pos] = 128 | (--ofs & 127);
401 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
402 unuse_pack(&w_curs);
403 return 0;
405 sha1write(f, header, hdrlen);
406 sha1write(f, dheader + pos, sizeof(dheader) - pos);
407 hdrlen += sizeof(dheader) - pos;
408 reused_delta++;
409 } else if (type == OBJ_REF_DELTA) {
410 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
411 unuse_pack(&w_curs);
412 return 0;
414 sha1write(f, header, hdrlen);
415 sha1write(f, entry->delta->idx.sha1, 20);
416 hdrlen += 20;
417 reused_delta++;
418 } else {
419 if (limit && hdrlen + datalen + 20 >= limit) {
420 unuse_pack(&w_curs);
421 return 0;
423 sha1write(f, header, hdrlen);
425 copy_pack_data(f, p, &w_curs, offset, datalen);
426 unuse_pack(&w_curs);
427 reused++;
429 if (usable_delta)
430 written_delta++;
431 written++;
432 if (!pack_to_stdout)
433 entry->idx.crc32 = crc32_end(f);
434 return hdrlen + datalen;
437 static int write_one(struct sha1file *f,
438 struct object_entry *e,
439 off_t *offset)
441 unsigned long size;
443 /* offset is non zero if object is written already. */
444 if (e->idx.offset || e->preferred_base)
445 return 1;
447 /* if we are deltified, write out base object first. */
448 if (e->delta && !write_one(f, e->delta, offset))
449 return 0;
451 e->idx.offset = *offset;
452 size = write_object(f, e, *offset);
453 if (!size) {
454 e->idx.offset = 0;
455 return 0;
457 written_list[nr_written++] = &e->idx;
459 /* make sure off_t is sufficiently large not to wrap */
460 if (*offset > *offset + size)
461 die("pack too large for current definition of off_t");
462 *offset += size;
463 return 1;
466 /* forward declaration for write_pack_file */
467 static int adjust_perm(const char *path, mode_t mode);
469 static void write_pack_file(void)
471 uint32_t i = 0, j;
472 struct sha1file *f;
473 off_t offset;
474 struct pack_header hdr;
475 uint32_t nr_remaining = nr_result;
476 time_t last_mtime = 0;
478 if (progress > pack_to_stdout)
479 progress_state = start_progress("Writing objects", nr_result);
480 written_list = xmalloc(nr_objects * sizeof(*written_list));
482 do {
483 unsigned char sha1[20];
484 char *pack_tmp_name = NULL;
486 if (pack_to_stdout) {
487 f = sha1fd_throughput(1, "<stdout>", progress_state);
488 } else {
489 char tmpname[PATH_MAX];
490 int fd;
491 snprintf(tmpname, sizeof(tmpname),
492 "%s/pack/tmp_pack_XXXXXX", get_object_directory());
493 fd = xmkstemp(tmpname);
494 pack_tmp_name = xstrdup(tmpname);
495 f = sha1fd(fd, pack_tmp_name);
498 hdr.hdr_signature = htonl(PACK_SIGNATURE);
499 hdr.hdr_version = htonl(PACK_VERSION);
500 hdr.hdr_entries = htonl(nr_remaining);
501 sha1write(f, &hdr, sizeof(hdr));
502 offset = sizeof(hdr);
503 nr_written = 0;
504 for (; i < nr_objects; i++) {
505 if (!write_one(f, objects + i, &offset))
506 break;
507 display_progress(progress_state, written);
511 * Did we write the wrong # entries in the header?
512 * If so, rewrite it like in fast-import
514 if (pack_to_stdout) {
515 sha1close(f, sha1, CSUM_CLOSE);
516 } else if (nr_written == nr_remaining) {
517 sha1close(f, sha1, CSUM_FSYNC);
518 } else {
519 int fd = sha1close(f, sha1, 0);
520 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
521 nr_written, sha1, offset);
522 close(fd);
525 if (!pack_to_stdout) {
526 mode_t mode = umask(0);
527 struct stat st;
528 char *idx_tmp_name, tmpname[PATH_MAX];
530 umask(mode);
531 mode = 0444 & ~mode;
533 idx_tmp_name = write_idx_file(NULL, written_list,
534 nr_written, sha1);
536 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
537 base_name, sha1_to_hex(sha1));
538 free_pack_by_name(tmpname);
539 if (adjust_perm(pack_tmp_name, mode))
540 die("unable to make temporary pack file readable: %s",
541 strerror(errno));
542 if (rename(pack_tmp_name, tmpname))
543 die("unable to rename temporary pack file: %s",
544 strerror(errno));
547 * Packs are runtime accessed in their mtime
548 * order since newer packs are more likely to contain
549 * younger objects. So if we are creating multiple
550 * packs then we should modify the mtime of later ones
551 * to preserve this property.
553 if (stat(tmpname, &st) < 0) {
554 warning("failed to stat %s: %s",
555 tmpname, strerror(errno));
556 } else if (!last_mtime) {
557 last_mtime = st.st_mtime;
558 } else {
559 struct utimbuf utb;
560 utb.actime = st.st_atime;
561 utb.modtime = --last_mtime;
562 if (utime(tmpname, &utb) < 0)
563 warning("failed utime() on %s: %s",
564 tmpname, strerror(errno));
567 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
568 base_name, sha1_to_hex(sha1));
569 if (adjust_perm(idx_tmp_name, mode))
570 die("unable to make temporary index file readable: %s",
571 strerror(errno));
572 if (rename(idx_tmp_name, tmpname))
573 die("unable to rename temporary index file: %s",
574 strerror(errno));
576 free(idx_tmp_name);
577 free(pack_tmp_name);
578 puts(sha1_to_hex(sha1));
581 /* mark written objects as written to previous pack */
582 for (j = 0; j < nr_written; j++) {
583 written_list[j]->offset = (off_t)-1;
585 nr_remaining -= nr_written;
586 } while (nr_remaining && i < nr_objects);
588 free(written_list);
589 stop_progress(&progress_state);
590 if (written != nr_result)
591 die("wrote %"PRIu32" objects while expecting %"PRIu32,
592 written, nr_result);
594 * We have scanned through [0 ... i). Since we have written
595 * the correct number of objects, the remaining [i ... nr_objects)
596 * items must be either already written (due to out-of-order delta base)
597 * or a preferred base. Count those which are neither and complain if any.
599 for (j = 0; i < nr_objects; i++) {
600 struct object_entry *e = objects + i;
601 j += !e->idx.offset && !e->preferred_base;
603 if (j)
604 die("wrote %"PRIu32" objects as expected but %"PRIu32
605 " unwritten", written, j);
608 static int locate_object_entry_hash(const unsigned char *sha1)
610 int i;
611 unsigned int ui;
612 memcpy(&ui, sha1, sizeof(unsigned int));
613 i = ui % object_ix_hashsz;
614 while (0 < object_ix[i]) {
615 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
616 return i;
617 if (++i == object_ix_hashsz)
618 i = 0;
620 return -1 - i;
623 static struct object_entry *locate_object_entry(const unsigned char *sha1)
625 int i;
627 if (!object_ix_hashsz)
628 return NULL;
630 i = locate_object_entry_hash(sha1);
631 if (0 <= i)
632 return &objects[object_ix[i]-1];
633 return NULL;
636 static void rehash_objects(void)
638 uint32_t i;
639 struct object_entry *oe;
641 object_ix_hashsz = nr_objects * 3;
642 if (object_ix_hashsz < 1024)
643 object_ix_hashsz = 1024;
644 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
645 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
646 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
647 int ix = locate_object_entry_hash(oe->idx.sha1);
648 if (0 <= ix)
649 continue;
650 ix = -1 - ix;
651 object_ix[ix] = i + 1;
655 static unsigned name_hash(const char *name)
657 unsigned char c;
658 unsigned hash = 0;
660 if (!name)
661 return 0;
664 * This effectively just creates a sortable number from the
665 * last sixteen non-whitespace characters. Last characters
666 * count "most", so things that end in ".c" sort together.
668 while ((c = *name++) != 0) {
669 if (isspace(c))
670 continue;
671 hash = (hash >> 2) + (c << 24);
673 return hash;
676 static void setup_delta_attr_check(struct git_attr_check *check)
678 static struct git_attr *attr_delta;
680 if (!attr_delta)
681 attr_delta = git_attr("delta", 5);
683 check[0].attr = attr_delta;
686 static int no_try_delta(const char *path)
688 struct git_attr_check check[1];
690 setup_delta_attr_check(check);
691 if (git_checkattr(path, ARRAY_SIZE(check), check))
692 return 0;
693 if (ATTR_FALSE(check->value))
694 return 1;
695 return 0;
698 static int add_object_entry(const unsigned char *sha1, enum object_type type,
699 const char *name, int exclude)
701 struct object_entry *entry;
702 struct packed_git *p, *found_pack = NULL;
703 off_t found_offset = 0;
704 int ix;
705 unsigned hash = name_hash(name);
707 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
708 if (ix >= 0) {
709 if (exclude) {
710 entry = objects + object_ix[ix] - 1;
711 if (!entry->preferred_base)
712 nr_result--;
713 entry->preferred_base = 1;
715 return 0;
718 if (!exclude && local && has_loose_object_nonlocal(sha1))
719 return 0;
721 for (p = packed_git; p; p = p->next) {
722 off_t offset = find_pack_entry_one(sha1, p);
723 if (offset) {
724 if (!found_pack) {
725 found_offset = offset;
726 found_pack = p;
728 if (exclude)
729 break;
730 if (incremental)
731 return 0;
732 if (local && !p->pack_local)
733 return 0;
734 if (ignore_packed_keep && p->pack_local && p->pack_keep)
735 return 0;
739 if (nr_objects >= nr_alloc) {
740 nr_alloc = (nr_alloc + 1024) * 3 / 2;
741 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
744 entry = objects + nr_objects++;
745 memset(entry, 0, sizeof(*entry));
746 hashcpy(entry->idx.sha1, sha1);
747 entry->hash = hash;
748 if (type)
749 entry->type = type;
750 if (exclude)
751 entry->preferred_base = 1;
752 else
753 nr_result++;
754 if (found_pack) {
755 entry->in_pack = found_pack;
756 entry->in_pack_offset = found_offset;
759 if (object_ix_hashsz * 3 <= nr_objects * 4)
760 rehash_objects();
761 else
762 object_ix[-1 - ix] = nr_objects;
764 display_progress(progress_state, nr_objects);
766 if (name && no_try_delta(name))
767 entry->no_try_delta = 1;
769 return 1;
772 struct pbase_tree_cache {
773 unsigned char sha1[20];
774 int ref;
775 int temporary;
776 void *tree_data;
777 unsigned long tree_size;
780 static struct pbase_tree_cache *(pbase_tree_cache[256]);
781 static int pbase_tree_cache_ix(const unsigned char *sha1)
783 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
785 static int pbase_tree_cache_ix_incr(int ix)
787 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
790 static struct pbase_tree {
791 struct pbase_tree *next;
792 /* This is a phony "cache" entry; we are not
793 * going to evict it nor find it through _get()
794 * mechanism -- this is for the toplevel node that
795 * would almost always change with any commit.
797 struct pbase_tree_cache pcache;
798 } *pbase_tree;
800 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
802 struct pbase_tree_cache *ent, *nent;
803 void *data;
804 unsigned long size;
805 enum object_type type;
806 int neigh;
807 int my_ix = pbase_tree_cache_ix(sha1);
808 int available_ix = -1;
810 /* pbase-tree-cache acts as a limited hashtable.
811 * your object will be found at your index or within a few
812 * slots after that slot if it is cached.
814 for (neigh = 0; neigh < 8; neigh++) {
815 ent = pbase_tree_cache[my_ix];
816 if (ent && !hashcmp(ent->sha1, sha1)) {
817 ent->ref++;
818 return ent;
820 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
821 ((0 <= available_ix) &&
822 (!ent && pbase_tree_cache[available_ix])))
823 available_ix = my_ix;
824 if (!ent)
825 break;
826 my_ix = pbase_tree_cache_ix_incr(my_ix);
829 /* Did not find one. Either we got a bogus request or
830 * we need to read and perhaps cache.
832 data = read_sha1_file(sha1, &type, &size);
833 if (!data)
834 return NULL;
835 if (type != OBJ_TREE) {
836 free(data);
837 return NULL;
840 /* We need to either cache or return a throwaway copy */
842 if (available_ix < 0)
843 ent = NULL;
844 else {
845 ent = pbase_tree_cache[available_ix];
846 my_ix = available_ix;
849 if (!ent) {
850 nent = xmalloc(sizeof(*nent));
851 nent->temporary = (available_ix < 0);
853 else {
854 /* evict and reuse */
855 free(ent->tree_data);
856 nent = ent;
858 hashcpy(nent->sha1, sha1);
859 nent->tree_data = data;
860 nent->tree_size = size;
861 nent->ref = 1;
862 if (!nent->temporary)
863 pbase_tree_cache[my_ix] = nent;
864 return nent;
867 static void pbase_tree_put(struct pbase_tree_cache *cache)
869 if (!cache->temporary) {
870 cache->ref--;
871 return;
873 free(cache->tree_data);
874 free(cache);
877 static int name_cmp_len(const char *name)
879 int i;
880 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
882 return i;
885 static void add_pbase_object(struct tree_desc *tree,
886 const char *name,
887 int cmplen,
888 const char *fullname)
890 struct name_entry entry;
891 int cmp;
893 while (tree_entry(tree,&entry)) {
894 if (S_ISGITLINK(entry.mode))
895 continue;
896 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
897 memcmp(name, entry.path, cmplen);
898 if (cmp > 0)
899 continue;
900 if (cmp < 0)
901 return;
902 if (name[cmplen] != '/') {
903 add_object_entry(entry.sha1,
904 object_type(entry.mode),
905 fullname, 1);
906 return;
908 if (S_ISDIR(entry.mode)) {
909 struct tree_desc sub;
910 struct pbase_tree_cache *tree;
911 const char *down = name+cmplen+1;
912 int downlen = name_cmp_len(down);
914 tree = pbase_tree_get(entry.sha1);
915 if (!tree)
916 return;
917 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
919 add_pbase_object(&sub, down, downlen, fullname);
920 pbase_tree_put(tree);
925 static unsigned *done_pbase_paths;
926 static int done_pbase_paths_num;
927 static int done_pbase_paths_alloc;
928 static int done_pbase_path_pos(unsigned hash)
930 int lo = 0;
931 int hi = done_pbase_paths_num;
932 while (lo < hi) {
933 int mi = (hi + lo) / 2;
934 if (done_pbase_paths[mi] == hash)
935 return mi;
936 if (done_pbase_paths[mi] < hash)
937 hi = mi;
938 else
939 lo = mi + 1;
941 return -lo-1;
944 static int check_pbase_path(unsigned hash)
946 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
947 if (0 <= pos)
948 return 1;
949 pos = -pos - 1;
950 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
951 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
952 done_pbase_paths = xrealloc(done_pbase_paths,
953 done_pbase_paths_alloc *
954 sizeof(unsigned));
956 done_pbase_paths_num++;
957 if (pos < done_pbase_paths_num)
958 memmove(done_pbase_paths + pos + 1,
959 done_pbase_paths + pos,
960 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
961 done_pbase_paths[pos] = hash;
962 return 0;
965 static void add_preferred_base_object(const char *name)
967 struct pbase_tree *it;
968 int cmplen;
969 unsigned hash = name_hash(name);
971 if (!num_preferred_base || check_pbase_path(hash))
972 return;
974 cmplen = name_cmp_len(name);
975 for (it = pbase_tree; it; it = it->next) {
976 if (cmplen == 0) {
977 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
979 else {
980 struct tree_desc tree;
981 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
982 add_pbase_object(&tree, name, cmplen, name);
987 static void add_preferred_base(unsigned char *sha1)
989 struct pbase_tree *it;
990 void *data;
991 unsigned long size;
992 unsigned char tree_sha1[20];
994 if (window <= num_preferred_base++)
995 return;
997 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
998 if (!data)
999 return;
1001 for (it = pbase_tree; it; it = it->next) {
1002 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1003 free(data);
1004 return;
1008 it = xcalloc(1, sizeof(*it));
1009 it->next = pbase_tree;
1010 pbase_tree = it;
1012 hashcpy(it->pcache.sha1, tree_sha1);
1013 it->pcache.tree_data = data;
1014 it->pcache.tree_size = size;
1017 static void check_object(struct object_entry *entry)
1019 if (entry->in_pack) {
1020 struct packed_git *p = entry->in_pack;
1021 struct pack_window *w_curs = NULL;
1022 const unsigned char *base_ref = NULL;
1023 struct object_entry *base_entry;
1024 unsigned long used, used_0;
1025 unsigned int avail;
1026 off_t ofs;
1027 unsigned char *buf, c;
1029 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1032 * We want in_pack_type even if we do not reuse delta
1033 * since non-delta representations could still be reused.
1035 used = unpack_object_header_buffer(buf, avail,
1036 &entry->in_pack_type,
1037 &entry->size);
1038 if (used == 0)
1039 goto give_up;
1042 * Determine if this is a delta and if so whether we can
1043 * reuse it or not. Otherwise let's find out as cheaply as
1044 * possible what the actual type and size for this object is.
1046 switch (entry->in_pack_type) {
1047 default:
1048 /* Not a delta hence we've already got all we need. */
1049 entry->type = entry->in_pack_type;
1050 entry->in_pack_header_size = used;
1051 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1052 goto give_up;
1053 unuse_pack(&w_curs);
1054 return;
1055 case OBJ_REF_DELTA:
1056 if (reuse_delta && !entry->preferred_base)
1057 base_ref = use_pack(p, &w_curs,
1058 entry->in_pack_offset + used, NULL);
1059 entry->in_pack_header_size = used + 20;
1060 break;
1061 case OBJ_OFS_DELTA:
1062 buf = use_pack(p, &w_curs,
1063 entry->in_pack_offset + used, NULL);
1064 used_0 = 0;
1065 c = buf[used_0++];
1066 ofs = c & 127;
1067 while (c & 128) {
1068 ofs += 1;
1069 if (!ofs || MSB(ofs, 7)) {
1070 error("delta base offset overflow in pack for %s",
1071 sha1_to_hex(entry->idx.sha1));
1072 goto give_up;
1074 c = buf[used_0++];
1075 ofs = (ofs << 7) + (c & 127);
1077 ofs = entry->in_pack_offset - ofs;
1078 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1079 error("delta base offset out of bound for %s",
1080 sha1_to_hex(entry->idx.sha1));
1081 goto give_up;
1083 if (reuse_delta && !entry->preferred_base) {
1084 struct revindex_entry *revidx;
1085 revidx = find_pack_revindex(p, ofs);
1086 if (!revidx)
1087 goto give_up;
1088 base_ref = nth_packed_object_sha1(p, revidx->nr);
1090 entry->in_pack_header_size = used + used_0;
1091 break;
1094 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1096 * If base_ref was set above that means we wish to
1097 * reuse delta data, and we even found that base
1098 * in the list of objects we want to pack. Goodie!
1100 * Depth value does not matter - find_deltas() will
1101 * never consider reused delta as the base object to
1102 * deltify other objects against, in order to avoid
1103 * circular deltas.
1105 entry->type = entry->in_pack_type;
1106 entry->delta = base_entry;
1107 entry->delta_size = entry->size;
1108 entry->delta_sibling = base_entry->delta_child;
1109 base_entry->delta_child = entry;
1110 unuse_pack(&w_curs);
1111 return;
1114 if (entry->type) {
1116 * This must be a delta and we already know what the
1117 * final object type is. Let's extract the actual
1118 * object size from the delta header.
1120 entry->size = get_size_from_delta(p, &w_curs,
1121 entry->in_pack_offset + entry->in_pack_header_size);
1122 if (entry->size == 0)
1123 goto give_up;
1124 unuse_pack(&w_curs);
1125 return;
1129 * No choice but to fall back to the recursive delta walk
1130 * with sha1_object_info() to find about the object type
1131 * at this point...
1133 give_up:
1134 unuse_pack(&w_curs);
1137 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1139 * The error condition is checked in prepare_pack(). This is
1140 * to permit a missing preferred base object to be ignored
1141 * as a preferred base. Doing so can result in a larger
1142 * pack file, but the transfer will still take place.
1146 static int pack_offset_sort(const void *_a, const void *_b)
1148 const struct object_entry *a = *(struct object_entry **)_a;
1149 const struct object_entry *b = *(struct object_entry **)_b;
1151 /* avoid filesystem trashing with loose objects */
1152 if (!a->in_pack && !b->in_pack)
1153 return hashcmp(a->idx.sha1, b->idx.sha1);
1155 if (a->in_pack < b->in_pack)
1156 return -1;
1157 if (a->in_pack > b->in_pack)
1158 return 1;
1159 return a->in_pack_offset < b->in_pack_offset ? -1 :
1160 (a->in_pack_offset > b->in_pack_offset);
1163 static void get_object_details(void)
1165 uint32_t i;
1166 struct object_entry **sorted_by_offset;
1168 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1169 for (i = 0; i < nr_objects; i++)
1170 sorted_by_offset[i] = objects + i;
1171 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1173 for (i = 0; i < nr_objects; i++)
1174 check_object(sorted_by_offset[i]);
1176 free(sorted_by_offset);
1180 * We search for deltas in a list sorted by type, by filename hash, and then
1181 * by size, so that we see progressively smaller and smaller files.
1182 * That's because we prefer deltas to be from the bigger file
1183 * to the smaller -- deletes are potentially cheaper, but perhaps
1184 * more importantly, the bigger file is likely the more recent
1185 * one. The deepest deltas are therefore the oldest objects which are
1186 * less susceptible to be accessed often.
1188 static int type_size_sort(const void *_a, const void *_b)
1190 const struct object_entry *a = *(struct object_entry **)_a;
1191 const struct object_entry *b = *(struct object_entry **)_b;
1193 if (a->type > b->type)
1194 return -1;
1195 if (a->type < b->type)
1196 return 1;
1197 if (a->hash > b->hash)
1198 return -1;
1199 if (a->hash < b->hash)
1200 return 1;
1201 if (a->preferred_base > b->preferred_base)
1202 return -1;
1203 if (a->preferred_base < b->preferred_base)
1204 return 1;
1205 if (a->size > b->size)
1206 return -1;
1207 if (a->size < b->size)
1208 return 1;
1209 return a < b ? -1 : (a > b); /* newest first */
1212 struct unpacked {
1213 struct object_entry *entry;
1214 void *data;
1215 struct delta_index *index;
1216 unsigned depth;
1219 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1220 unsigned long delta_size)
1222 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1223 return 0;
1225 if (delta_size < cache_max_small_delta_size)
1226 return 1;
1228 /* cache delta, if objects are large enough compared to delta size */
1229 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1230 return 1;
1232 return 0;
1235 #ifdef THREADED_DELTA_SEARCH
1237 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1238 #define read_lock() pthread_mutex_lock(&read_mutex)
1239 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1241 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1242 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1243 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1245 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1246 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1247 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1249 #else
1251 #define read_lock() (void)0
1252 #define read_unlock() (void)0
1253 #define cache_lock() (void)0
1254 #define cache_unlock() (void)0
1255 #define progress_lock() (void)0
1256 #define progress_unlock() (void)0
1258 #endif
1260 static int try_delta(struct unpacked *trg, struct unpacked *src,
1261 unsigned max_depth, unsigned long *mem_usage)
1263 struct object_entry *trg_entry = trg->entry;
1264 struct object_entry *src_entry = src->entry;
1265 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1266 unsigned ref_depth;
1267 enum object_type type;
1268 void *delta_buf;
1270 /* Don't bother doing diffs between different types */
1271 if (trg_entry->type != src_entry->type)
1272 return -1;
1275 * We do not bother to try a delta that we discarded
1276 * on an earlier try, but only when reusing delta data.
1278 if (reuse_delta && trg_entry->in_pack &&
1279 trg_entry->in_pack == src_entry->in_pack &&
1280 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1281 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1282 return 0;
1284 /* Let's not bust the allowed depth. */
1285 if (src->depth >= max_depth)
1286 return 0;
1288 /* Now some size filtering heuristics. */
1289 trg_size = trg_entry->size;
1290 if (!trg_entry->delta) {
1291 max_size = trg_size/2 - 20;
1292 ref_depth = 1;
1293 } else {
1294 max_size = trg_entry->delta_size;
1295 ref_depth = trg->depth;
1297 max_size = max_size * (max_depth - src->depth) /
1298 (max_depth - ref_depth + 1);
1299 if (max_size == 0)
1300 return 0;
1301 src_size = src_entry->size;
1302 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1303 if (sizediff >= max_size)
1304 return 0;
1305 if (trg_size < src_size / 32)
1306 return 0;
1308 /* Load data if not already done */
1309 if (!trg->data) {
1310 read_lock();
1311 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1312 read_unlock();
1313 if (!trg->data)
1314 die("object %s cannot be read",
1315 sha1_to_hex(trg_entry->idx.sha1));
1316 if (sz != trg_size)
1317 die("object %s inconsistent object length (%lu vs %lu)",
1318 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1319 *mem_usage += sz;
1321 if (!src->data) {
1322 read_lock();
1323 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1324 read_unlock();
1325 if (!src->data)
1326 die("object %s cannot be read",
1327 sha1_to_hex(src_entry->idx.sha1));
1328 if (sz != src_size)
1329 die("object %s inconsistent object length (%lu vs %lu)",
1330 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1331 *mem_usage += sz;
1333 if (!src->index) {
1334 src->index = create_delta_index(src->data, src_size);
1335 if (!src->index) {
1336 static int warned = 0;
1337 if (!warned++)
1338 warning("suboptimal pack - out of memory");
1339 return 0;
1341 *mem_usage += sizeof_delta_index(src->index);
1344 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1345 if (!delta_buf)
1346 return 0;
1348 if (trg_entry->delta) {
1349 /* Prefer only shallower same-sized deltas. */
1350 if (delta_size == trg_entry->delta_size &&
1351 src->depth + 1 >= trg->depth) {
1352 free(delta_buf);
1353 return 0;
1358 * Handle memory allocation outside of the cache
1359 * accounting lock. Compiler will optimize the strangeness
1360 * away when THREADED_DELTA_SEARCH is not defined.
1362 free(trg_entry->delta_data);
1363 cache_lock();
1364 if (trg_entry->delta_data) {
1365 delta_cache_size -= trg_entry->delta_size;
1366 trg_entry->delta_data = NULL;
1368 if (delta_cacheable(src_size, trg_size, delta_size)) {
1369 delta_cache_size += delta_size;
1370 cache_unlock();
1371 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1372 } else {
1373 cache_unlock();
1374 free(delta_buf);
1377 trg_entry->delta = src_entry;
1378 trg_entry->delta_size = delta_size;
1379 trg->depth = src->depth + 1;
1381 return 1;
1384 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1386 struct object_entry *child = me->delta_child;
1387 unsigned int m = n;
1388 while (child) {
1389 unsigned int c = check_delta_limit(child, n + 1);
1390 if (m < c)
1391 m = c;
1392 child = child->delta_sibling;
1394 return m;
1397 static unsigned long free_unpacked(struct unpacked *n)
1399 unsigned long freed_mem = sizeof_delta_index(n->index);
1400 free_delta_index(n->index);
1401 n->index = NULL;
1402 if (n->data) {
1403 freed_mem += n->entry->size;
1404 free(n->data);
1405 n->data = NULL;
1407 n->entry = NULL;
1408 n->depth = 0;
1409 return freed_mem;
1412 static void find_deltas(struct object_entry **list, unsigned *list_size,
1413 int window, int depth, unsigned *processed)
1415 uint32_t i, idx = 0, count = 0;
1416 struct unpacked *array;
1417 unsigned long mem_usage = 0;
1419 array = xcalloc(window, sizeof(struct unpacked));
1421 for (;;) {
1422 struct object_entry *entry;
1423 struct unpacked *n = array + idx;
1424 int j, max_depth, best_base = -1;
1426 progress_lock();
1427 if (!*list_size) {
1428 progress_unlock();
1429 break;
1431 entry = *list++;
1432 (*list_size)--;
1433 if (!entry->preferred_base) {
1434 (*processed)++;
1435 display_progress(progress_state, *processed);
1437 progress_unlock();
1439 mem_usage -= free_unpacked(n);
1440 n->entry = entry;
1442 while (window_memory_limit &&
1443 mem_usage > window_memory_limit &&
1444 count > 1) {
1445 uint32_t tail = (idx + window - count) % window;
1446 mem_usage -= free_unpacked(array + tail);
1447 count--;
1450 /* We do not compute delta to *create* objects we are not
1451 * going to pack.
1453 if (entry->preferred_base)
1454 goto next;
1457 * If the current object is at pack edge, take the depth the
1458 * objects that depend on the current object into account
1459 * otherwise they would become too deep.
1461 max_depth = depth;
1462 if (entry->delta_child) {
1463 max_depth -= check_delta_limit(entry, 0);
1464 if (max_depth <= 0)
1465 goto next;
1468 j = window;
1469 while (--j > 0) {
1470 int ret;
1471 uint32_t other_idx = idx + j;
1472 struct unpacked *m;
1473 if (other_idx >= window)
1474 other_idx -= window;
1475 m = array + other_idx;
1476 if (!m->entry)
1477 break;
1478 ret = try_delta(n, m, max_depth, &mem_usage);
1479 if (ret < 0)
1480 break;
1481 else if (ret > 0)
1482 best_base = other_idx;
1486 * If we decided to cache the delta data, then it is best
1487 * to compress it right away. First because we have to do
1488 * it anyway, and doing it here while we're threaded will
1489 * save a lot of time in the non threaded write phase,
1490 * as well as allow for caching more deltas within
1491 * the same cache size limit.
1492 * ...
1493 * But only if not writing to stdout, since in that case
1494 * the network is most likely throttling writes anyway,
1495 * and therefore it is best to go to the write phase ASAP
1496 * instead, as we can afford spending more time compressing
1497 * between writes at that moment.
1499 if (entry->delta_data && !pack_to_stdout) {
1500 entry->z_delta_size = do_compress(&entry->delta_data,
1501 entry->delta_size);
1502 cache_lock();
1503 delta_cache_size -= entry->delta_size;
1504 delta_cache_size += entry->z_delta_size;
1505 cache_unlock();
1508 /* if we made n a delta, and if n is already at max
1509 * depth, leaving it in the window is pointless. we
1510 * should evict it first.
1512 if (entry->delta && max_depth <= n->depth)
1513 continue;
1516 * Move the best delta base up in the window, after the
1517 * currently deltified object, to keep it longer. It will
1518 * be the first base object to be attempted next.
1520 if (entry->delta) {
1521 struct unpacked swap = array[best_base];
1522 int dist = (window + idx - best_base) % window;
1523 int dst = best_base;
1524 while (dist--) {
1525 int src = (dst + 1) % window;
1526 array[dst] = array[src];
1527 dst = src;
1529 array[dst] = swap;
1532 next:
1533 idx++;
1534 if (count + 1 < window)
1535 count++;
1536 if (idx >= window)
1537 idx = 0;
1540 for (i = 0; i < window; ++i) {
1541 free_delta_index(array[i].index);
1542 free(array[i].data);
1544 free(array);
1547 #ifdef THREADED_DELTA_SEARCH
1550 * The main thread waits on the condition that (at least) one of the workers
1551 * has stopped working (which is indicated in the .working member of
1552 * struct thread_params).
1553 * When a work thread has completed its work, it sets .working to 0 and
1554 * signals the main thread and waits on the condition that .data_ready
1555 * becomes 1.
1558 struct thread_params {
1559 pthread_t thread;
1560 struct object_entry **list;
1561 unsigned list_size;
1562 unsigned remaining;
1563 int window;
1564 int depth;
1565 int working;
1566 int data_ready;
1567 pthread_mutex_t mutex;
1568 pthread_cond_t cond;
1569 unsigned *processed;
1572 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1574 static void *threaded_find_deltas(void *arg)
1576 struct thread_params *me = arg;
1578 while (me->remaining) {
1579 find_deltas(me->list, &me->remaining,
1580 me->window, me->depth, me->processed);
1582 progress_lock();
1583 me->working = 0;
1584 pthread_cond_signal(&progress_cond);
1585 progress_unlock();
1588 * We must not set ->data_ready before we wait on the
1589 * condition because the main thread may have set it to 1
1590 * before we get here. In order to be sure that new
1591 * work is available if we see 1 in ->data_ready, it
1592 * was initialized to 0 before this thread was spawned
1593 * and we reset it to 0 right away.
1595 pthread_mutex_lock(&me->mutex);
1596 while (!me->data_ready)
1597 pthread_cond_wait(&me->cond, &me->mutex);
1598 me->data_ready = 0;
1599 pthread_mutex_unlock(&me->mutex);
1601 /* leave ->working 1 so that this doesn't get more work assigned */
1602 return NULL;
1605 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1606 int window, int depth, unsigned *processed)
1608 struct thread_params p[delta_search_threads];
1609 int i, ret, active_threads = 0;
1611 if (delta_search_threads <= 1) {
1612 find_deltas(list, &list_size, window, depth, processed);
1613 return;
1615 if (progress > pack_to_stdout)
1616 fprintf(stderr, "Delta compression using %d threads.\n",
1617 delta_search_threads);
1619 /* Partition the work amongst work threads. */
1620 for (i = 0; i < delta_search_threads; i++) {
1621 unsigned sub_size = list_size / (delta_search_threads - i);
1623 /* don't use too small segments or no deltas will be found */
1624 if (sub_size < 2*window && i+1 < delta_search_threads)
1625 sub_size = 0;
1627 p[i].window = window;
1628 p[i].depth = depth;
1629 p[i].processed = processed;
1630 p[i].working = 1;
1631 p[i].data_ready = 0;
1633 /* try to split chunks on "path" boundaries */
1634 while (sub_size && sub_size < list_size &&
1635 list[sub_size]->hash &&
1636 list[sub_size]->hash == list[sub_size-1]->hash)
1637 sub_size++;
1639 p[i].list = list;
1640 p[i].list_size = sub_size;
1641 p[i].remaining = sub_size;
1643 list += sub_size;
1644 list_size -= sub_size;
1647 /* Start work threads. */
1648 for (i = 0; i < delta_search_threads; i++) {
1649 if (!p[i].list_size)
1650 continue;
1651 pthread_mutex_init(&p[i].mutex, NULL);
1652 pthread_cond_init(&p[i].cond, NULL);
1653 ret = pthread_create(&p[i].thread, NULL,
1654 threaded_find_deltas, &p[i]);
1655 if (ret)
1656 die("unable to create thread: %s", strerror(ret));
1657 active_threads++;
1661 * Now let's wait for work completion. Each time a thread is done
1662 * with its work, we steal half of the remaining work from the
1663 * thread with the largest number of unprocessed objects and give
1664 * it to that newly idle thread. This ensure good load balancing
1665 * until the remaining object list segments are simply too short
1666 * to be worth splitting anymore.
1668 while (active_threads) {
1669 struct thread_params *target = NULL;
1670 struct thread_params *victim = NULL;
1671 unsigned sub_size = 0;
1673 progress_lock();
1674 for (;;) {
1675 for (i = 0; !target && i < delta_search_threads; i++)
1676 if (!p[i].working)
1677 target = &p[i];
1678 if (target)
1679 break;
1680 pthread_cond_wait(&progress_cond, &progress_mutex);
1683 for (i = 0; i < delta_search_threads; i++)
1684 if (p[i].remaining > 2*window &&
1685 (!victim || victim->remaining < p[i].remaining))
1686 victim = &p[i];
1687 if (victim) {
1688 sub_size = victim->remaining / 2;
1689 list = victim->list + victim->list_size - sub_size;
1690 while (sub_size && list[0]->hash &&
1691 list[0]->hash == list[-1]->hash) {
1692 list++;
1693 sub_size--;
1695 if (!sub_size) {
1697 * It is possible for some "paths" to have
1698 * so many objects that no hash boundary
1699 * might be found. Let's just steal the
1700 * exact half in that case.
1702 sub_size = victim->remaining / 2;
1703 list -= sub_size;
1705 target->list = list;
1706 victim->list_size -= sub_size;
1707 victim->remaining -= sub_size;
1709 target->list_size = sub_size;
1710 target->remaining = sub_size;
1711 target->working = 1;
1712 progress_unlock();
1714 pthread_mutex_lock(&target->mutex);
1715 target->data_ready = 1;
1716 pthread_cond_signal(&target->cond);
1717 pthread_mutex_unlock(&target->mutex);
1719 if (!sub_size) {
1720 pthread_join(target->thread, NULL);
1721 pthread_cond_destroy(&target->cond);
1722 pthread_mutex_destroy(&target->mutex);
1723 active_threads--;
1728 #else
1729 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1730 #endif
1732 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1734 unsigned char peeled[20];
1736 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1737 !peel_ref(path, peeled) && /* peelable? */
1738 !is_null_sha1(peeled) && /* annotated tag? */
1739 locate_object_entry(peeled)) /* object packed? */
1740 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1741 return 0;
1744 static void prepare_pack(int window, int depth)
1746 struct object_entry **delta_list;
1747 uint32_t i, nr_deltas;
1748 unsigned n;
1750 get_object_details();
1753 * If we're locally repacking then we need to be doubly careful
1754 * from now on in order to make sure no stealth corruption gets
1755 * propagated to the new pack. Clients receiving streamed packs
1756 * should validate everything they get anyway so no need to incur
1757 * the additional cost here in that case.
1759 if (!pack_to_stdout)
1760 do_check_packed_object_crc = 1;
1762 if (!nr_objects || !window || !depth)
1763 return;
1765 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1766 nr_deltas = n = 0;
1768 for (i = 0; i < nr_objects; i++) {
1769 struct object_entry *entry = objects + i;
1771 if (entry->delta)
1772 /* This happens if we decided to reuse existing
1773 * delta from a pack. "reuse_delta &&" is implied.
1775 continue;
1777 if (entry->size < 50)
1778 continue;
1780 if (entry->no_try_delta)
1781 continue;
1783 if (!entry->preferred_base) {
1784 nr_deltas++;
1785 if (entry->type < 0)
1786 die("unable to get type of object %s",
1787 sha1_to_hex(entry->idx.sha1));
1788 } else {
1789 if (entry->type < 0) {
1791 * This object is not found, but we
1792 * don't have to include it anyway.
1794 continue;
1798 delta_list[n++] = entry;
1801 if (nr_deltas && n > 1) {
1802 unsigned nr_done = 0;
1803 if (progress)
1804 progress_state = start_progress("Compressing objects",
1805 nr_deltas);
1806 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1807 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1808 stop_progress(&progress_state);
1809 if (nr_done != nr_deltas)
1810 die("inconsistency with delta count");
1812 free(delta_list);
1815 static int git_pack_config(const char *k, const char *v, void *cb)
1817 if(!strcmp(k, "pack.window")) {
1818 window = git_config_int(k, v);
1819 return 0;
1821 if (!strcmp(k, "pack.windowmemory")) {
1822 window_memory_limit = git_config_ulong(k, v);
1823 return 0;
1825 if (!strcmp(k, "pack.depth")) {
1826 depth = git_config_int(k, v);
1827 return 0;
1829 if (!strcmp(k, "pack.compression")) {
1830 int level = git_config_int(k, v);
1831 if (level == -1)
1832 level = Z_DEFAULT_COMPRESSION;
1833 else if (level < 0 || level > Z_BEST_COMPRESSION)
1834 die("bad pack compression level %d", level);
1835 pack_compression_level = level;
1836 pack_compression_seen = 1;
1837 return 0;
1839 if (!strcmp(k, "pack.deltacachesize")) {
1840 max_delta_cache_size = git_config_int(k, v);
1841 return 0;
1843 if (!strcmp(k, "pack.deltacachelimit")) {
1844 cache_max_small_delta_size = git_config_int(k, v);
1845 return 0;
1847 if (!strcmp(k, "pack.threads")) {
1848 delta_search_threads = git_config_int(k, v);
1849 if (delta_search_threads < 0)
1850 die("invalid number of threads specified (%d)",
1851 delta_search_threads);
1852 #ifndef THREADED_DELTA_SEARCH
1853 if (delta_search_threads != 1)
1854 warning("no threads support, ignoring %s", k);
1855 #endif
1856 return 0;
1858 if (!strcmp(k, "pack.indexversion")) {
1859 pack_idx_default_version = git_config_int(k, v);
1860 if (pack_idx_default_version > 2)
1861 die("bad pack.indexversion=%"PRIu32,
1862 pack_idx_default_version);
1863 return 0;
1865 if (!strcmp(k, "pack.packsizelimit")) {
1866 pack_size_limit_cfg = git_config_ulong(k, v);
1867 return 0;
1869 return git_default_config(k, v, cb);
1872 static void read_object_list_from_stdin(void)
1874 char line[40 + 1 + PATH_MAX + 2];
1875 unsigned char sha1[20];
1877 for (;;) {
1878 if (!fgets(line, sizeof(line), stdin)) {
1879 if (feof(stdin))
1880 break;
1881 if (!ferror(stdin))
1882 die("fgets returned NULL, not EOF, not error!");
1883 if (errno != EINTR)
1884 die("fgets: %s", strerror(errno));
1885 clearerr(stdin);
1886 continue;
1888 if (line[0] == '-') {
1889 if (get_sha1_hex(line+1, sha1))
1890 die("expected edge sha1, got garbage:\n %s",
1891 line);
1892 add_preferred_base(sha1);
1893 continue;
1895 if (get_sha1_hex(line, sha1))
1896 die("expected sha1, got garbage:\n %s", line);
1898 add_preferred_base_object(line+41);
1899 add_object_entry(sha1, 0, line+41, 0);
1903 #define OBJECT_ADDED (1u<<20)
1905 static void show_commit(struct commit *commit)
1907 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1908 commit->object.flags |= OBJECT_ADDED;
1911 static void show_object(struct object_array_entry *p)
1913 add_preferred_base_object(p->name);
1914 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1915 p->item->flags |= OBJECT_ADDED;
1918 static void show_edge(struct commit *commit)
1920 add_preferred_base(commit->object.sha1);
1923 struct in_pack_object {
1924 off_t offset;
1925 struct object *object;
1928 struct in_pack {
1929 int alloc;
1930 int nr;
1931 struct in_pack_object *array;
1934 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1936 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1937 in_pack->array[in_pack->nr].object = object;
1938 in_pack->nr++;
1942 * Compare the objects in the offset order, in order to emulate the
1943 * "git rev-list --objects" output that produced the pack originally.
1945 static int ofscmp(const void *a_, const void *b_)
1947 struct in_pack_object *a = (struct in_pack_object *)a_;
1948 struct in_pack_object *b = (struct in_pack_object *)b_;
1950 if (a->offset < b->offset)
1951 return -1;
1952 else if (a->offset > b->offset)
1953 return 1;
1954 else
1955 return hashcmp(a->object->sha1, b->object->sha1);
1958 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1960 struct packed_git *p;
1961 struct in_pack in_pack;
1962 uint32_t i;
1964 memset(&in_pack, 0, sizeof(in_pack));
1966 for (p = packed_git; p; p = p->next) {
1967 const unsigned char *sha1;
1968 struct object *o;
1970 for (i = 0; i < revs->num_ignore_packed; i++) {
1971 if (matches_pack_name(p, revs->ignore_packed[i]))
1972 break;
1974 if (revs->num_ignore_packed <= i)
1975 continue;
1976 if (open_pack_index(p))
1977 die("cannot open pack index");
1979 ALLOC_GROW(in_pack.array,
1980 in_pack.nr + p->num_objects,
1981 in_pack.alloc);
1983 for (i = 0; i < p->num_objects; i++) {
1984 sha1 = nth_packed_object_sha1(p, i);
1985 o = lookup_unknown_object(sha1);
1986 if (!(o->flags & OBJECT_ADDED))
1987 mark_in_pack_object(o, p, &in_pack);
1988 o->flags |= OBJECT_ADDED;
1992 if (in_pack.nr) {
1993 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1994 ofscmp);
1995 for (i = 0; i < in_pack.nr; i++) {
1996 struct object *o = in_pack.array[i].object;
1997 add_object_entry(o->sha1, o->type, "", 0);
2000 free(in_pack.array);
2003 static void loosen_unused_packed_objects(struct rev_info *revs)
2005 struct packed_git *p;
2006 uint32_t i;
2007 const unsigned char *sha1;
2009 for (p = packed_git; p; p = p->next) {
2010 for (i = 0; i < revs->num_ignore_packed; i++) {
2011 if (matches_pack_name(p, revs->ignore_packed[i]))
2012 break;
2014 if (revs->num_ignore_packed <= i)
2015 continue;
2017 if (open_pack_index(p))
2018 die("cannot open pack index");
2020 for (i = 0; i < p->num_objects; i++) {
2021 sha1 = nth_packed_object_sha1(p, i);
2022 if (!locate_object_entry(sha1))
2023 if (force_object_loose(sha1, p->mtime))
2024 die("unable to force loose object");
2029 static void get_object_list(int ac, const char **av)
2031 struct rev_info revs;
2032 char line[1000];
2033 int flags = 0;
2035 init_revisions(&revs, NULL);
2036 save_commit_buffer = 0;
2037 setup_revisions(ac, av, &revs, NULL);
2039 while (fgets(line, sizeof(line), stdin) != NULL) {
2040 int len = strlen(line);
2041 if (len && line[len - 1] == '\n')
2042 line[--len] = 0;
2043 if (!len)
2044 break;
2045 if (*line == '-') {
2046 if (!strcmp(line, "--not")) {
2047 flags ^= UNINTERESTING;
2048 continue;
2050 die("not a rev '%s'", line);
2052 if (handle_revision_arg(line, &revs, flags, 1))
2053 die("bad revision '%s'", line);
2056 if (prepare_revision_walk(&revs))
2057 die("revision walk setup failed");
2058 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2059 traverse_commit_list(&revs, show_commit, show_object);
2061 if (keep_unreachable)
2062 add_objects_in_unpacked_packs(&revs);
2063 if (unpack_unreachable)
2064 loosen_unused_packed_objects(&revs);
2067 static int adjust_perm(const char *path, mode_t mode)
2069 if (chmod(path, mode))
2070 return -1;
2071 return adjust_shared_perm(path);
2074 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2076 int use_internal_rev_list = 0;
2077 int thin = 0;
2078 uint32_t i;
2079 const char **rp_av;
2080 int rp_ac_alloc = 64;
2081 int rp_ac;
2083 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2085 rp_av[0] = "pack-objects";
2086 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2087 rp_ac = 2;
2089 git_config(git_pack_config, NULL);
2090 if (!pack_compression_seen && core_compression_seen)
2091 pack_compression_level = core_compression_level;
2093 progress = isatty(2);
2094 for (i = 1; i < argc; i++) {
2095 const char *arg = argv[i];
2097 if (*arg != '-')
2098 break;
2100 if (!strcmp("--non-empty", arg)) {
2101 non_empty = 1;
2102 continue;
2104 if (!strcmp("--local", arg)) {
2105 local = 1;
2106 continue;
2108 if (!strcmp("--incremental", arg)) {
2109 incremental = 1;
2110 continue;
2112 if (!strcmp("--honor-pack-keep", arg)) {
2113 ignore_packed_keep = 1;
2114 continue;
2116 if (!prefixcmp(arg, "--compression=")) {
2117 char *end;
2118 int level = strtoul(arg+14, &end, 0);
2119 if (!arg[14] || *end)
2120 usage(pack_usage);
2121 if (level == -1)
2122 level = Z_DEFAULT_COMPRESSION;
2123 else if (level < 0 || level > Z_BEST_COMPRESSION)
2124 die("bad pack compression level %d", level);
2125 pack_compression_level = level;
2126 continue;
2128 if (!prefixcmp(arg, "--max-pack-size=")) {
2129 char *end;
2130 pack_size_limit_cfg = 0;
2131 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2132 if (!arg[16] || *end)
2133 usage(pack_usage);
2134 continue;
2136 if (!prefixcmp(arg, "--window=")) {
2137 char *end;
2138 window = strtoul(arg+9, &end, 0);
2139 if (!arg[9] || *end)
2140 usage(pack_usage);
2141 continue;
2143 if (!prefixcmp(arg, "--window-memory=")) {
2144 if (!git_parse_ulong(arg+16, &window_memory_limit))
2145 usage(pack_usage);
2146 continue;
2148 if (!prefixcmp(arg, "--threads=")) {
2149 char *end;
2150 delta_search_threads = strtoul(arg+10, &end, 0);
2151 if (!arg[10] || *end || delta_search_threads < 0)
2152 usage(pack_usage);
2153 #ifndef THREADED_DELTA_SEARCH
2154 if (delta_search_threads != 1)
2155 warning("no threads support, "
2156 "ignoring %s", arg);
2157 #endif
2158 continue;
2160 if (!prefixcmp(arg, "--depth=")) {
2161 char *end;
2162 depth = strtoul(arg+8, &end, 0);
2163 if (!arg[8] || *end)
2164 usage(pack_usage);
2165 continue;
2167 if (!strcmp("--progress", arg)) {
2168 progress = 1;
2169 continue;
2171 if (!strcmp("--all-progress", arg)) {
2172 progress = 2;
2173 continue;
2175 if (!strcmp("-q", arg)) {
2176 progress = 0;
2177 continue;
2179 if (!strcmp("--no-reuse-delta", arg)) {
2180 reuse_delta = 0;
2181 continue;
2183 if (!strcmp("--no-reuse-object", arg)) {
2184 reuse_object = reuse_delta = 0;
2185 continue;
2187 if (!strcmp("--delta-base-offset", arg)) {
2188 allow_ofs_delta = 1;
2189 continue;
2191 if (!strcmp("--stdout", arg)) {
2192 pack_to_stdout = 1;
2193 continue;
2195 if (!strcmp("--revs", arg)) {
2196 use_internal_rev_list = 1;
2197 continue;
2199 if (!strcmp("--keep-unreachable", arg)) {
2200 keep_unreachable = 1;
2201 continue;
2203 if (!strcmp("--unpack-unreachable", arg)) {
2204 unpack_unreachable = 1;
2205 continue;
2207 if (!strcmp("--include-tag", arg)) {
2208 include_tag = 1;
2209 continue;
2211 if (!strcmp("--unpacked", arg) ||
2212 !prefixcmp(arg, "--unpacked=") ||
2213 !strcmp("--reflog", arg) ||
2214 !strcmp("--all", arg)) {
2215 use_internal_rev_list = 1;
2216 if (rp_ac >= rp_ac_alloc - 1) {
2217 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2218 rp_av = xrealloc(rp_av,
2219 rp_ac_alloc * sizeof(*rp_av));
2221 rp_av[rp_ac++] = arg;
2222 continue;
2224 if (!strcmp("--thin", arg)) {
2225 use_internal_rev_list = 1;
2226 thin = 1;
2227 rp_av[1] = "--objects-edge";
2228 continue;
2230 if (!prefixcmp(arg, "--index-version=")) {
2231 char *c;
2232 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2233 if (pack_idx_default_version > 2)
2234 die("bad %s", arg);
2235 if (*c == ',')
2236 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2237 if (*c || pack_idx_off32_limit & 0x80000000)
2238 die("bad %s", arg);
2239 continue;
2241 usage(pack_usage);
2244 /* Traditionally "pack-objects [options] base extra" failed;
2245 * we would however want to take refs parameter that would
2246 * have been given to upstream rev-list ourselves, which means
2247 * we somehow want to say what the base name is. So the
2248 * syntax would be:
2250 * pack-objects [options] base <refs...>
2252 * in other words, we would treat the first non-option as the
2253 * base_name and send everything else to the internal revision
2254 * walker.
2257 if (!pack_to_stdout)
2258 base_name = argv[i++];
2260 if (pack_to_stdout != !base_name)
2261 usage(pack_usage);
2263 if (!pack_to_stdout && !pack_size_limit)
2264 pack_size_limit = pack_size_limit_cfg;
2266 if (pack_to_stdout && pack_size_limit)
2267 die("--max-pack-size cannot be used to build a pack for transfer.");
2269 if (!pack_to_stdout && thin)
2270 die("--thin cannot be used to build an indexable pack.");
2272 if (keep_unreachable && unpack_unreachable)
2273 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2275 #ifdef THREADED_DELTA_SEARCH
2276 if (!delta_search_threads) /* --threads=0 means autodetect */
2277 delta_search_threads = online_cpus();
2278 #endif
2280 prepare_packed_git();
2282 if (progress)
2283 progress_state = start_progress("Counting objects", 0);
2284 if (!use_internal_rev_list)
2285 read_object_list_from_stdin();
2286 else {
2287 rp_av[rp_ac] = NULL;
2288 get_object_list(rp_ac, rp_av);
2290 if (include_tag && nr_result)
2291 for_each_ref(add_ref_tag, NULL);
2292 stop_progress(&progress_state);
2294 if (non_empty && !nr_result)
2295 return 0;
2296 if (nr_result)
2297 prepare_pack(window, depth);
2298 write_pack_file();
2299 if (progress)
2300 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2301 " reused %"PRIu32" (delta %"PRIu32")\n",
2302 written, written_delta, reused, reused_delta);
2303 return 0;