pretty=format: respect date format options
[git/dscho.git] / builtin-pack-objects.c
blobd394c494a55d45a7af6506371f3432374dffe424
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 allow_ofs_delta;
75 static const char *base_name;
76 static int progress = 1;
77 static int window = 10;
78 static uint32_t pack_size_limit, pack_size_limit_cfg;
79 static int depth = 50;
80 static int delta_search_threads = 1;
81 static int pack_to_stdout;
82 static int num_preferred_base;
83 static struct progress *progress_state;
84 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
85 static int pack_compression_seen;
87 static unsigned long delta_cache_size = 0;
88 static unsigned long max_delta_cache_size = 0;
89 static unsigned long cache_max_small_delta_size = 1000;
91 static unsigned long window_memory_limit = 0;
94 * The object names in objects array are hashed with this hashtable,
95 * to help looking up the entry by object name.
96 * This hashtable is built after all the objects are seen.
98 static int *object_ix;
99 static int object_ix_hashsz;
102 * stats
104 static uint32_t written, written_delta;
105 static uint32_t reused, reused_delta;
108 static void *get_delta(struct object_entry *entry)
110 unsigned long size, base_size, delta_size;
111 void *buf, *base_buf, *delta_buf;
112 enum object_type type;
114 buf = read_sha1_file(entry->idx.sha1, &type, &size);
115 if (!buf)
116 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
117 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
118 if (!base_buf)
119 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
120 delta_buf = diff_delta(base_buf, base_size,
121 buf, size, &delta_size, 0);
122 if (!delta_buf || delta_size != entry->delta_size)
123 die("delta size changed");
124 free(buf);
125 free(base_buf);
126 return delta_buf;
129 static unsigned long do_compress(void **pptr, unsigned long size)
131 z_stream stream;
132 void *in, *out;
133 unsigned long maxsize;
135 memset(&stream, 0, sizeof(stream));
136 deflateInit(&stream, pack_compression_level);
137 maxsize = deflateBound(&stream, size);
139 in = *pptr;
140 out = xmalloc(maxsize);
141 *pptr = out;
143 stream.next_in = in;
144 stream.avail_in = size;
145 stream.next_out = out;
146 stream.avail_out = maxsize;
147 while (deflate(&stream, Z_FINISH) == Z_OK)
148 ; /* nothing */
149 deflateEnd(&stream);
151 free(in);
152 return stream.total_out;
156 * The per-object header is a pretty dense thing, which is
157 * - first byte: low four bits are "size", then three bits of "type",
158 * and the high bit is "size continues".
159 * - each byte afterwards: low seven bits are size continuation,
160 * with the high bit being "size continues"
162 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
164 int n = 1;
165 unsigned char c;
167 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
168 die("bad type %d", type);
170 c = (type << 4) | (size & 15);
171 size >>= 4;
172 while (size) {
173 *hdr++ = c | 0x80;
174 c = size & 0x7f;
175 size >>= 7;
176 n++;
178 *hdr = c;
179 return n;
183 * we are going to reuse the existing object data as is. make
184 * sure it is not corrupt.
186 static int check_pack_inflate(struct packed_git *p,
187 struct pack_window **w_curs,
188 off_t offset,
189 off_t len,
190 unsigned long expect)
192 z_stream stream;
193 unsigned char fakebuf[4096], *in;
194 int st;
196 memset(&stream, 0, sizeof(stream));
197 inflateInit(&stream);
198 do {
199 in = use_pack(p, w_curs, offset, &stream.avail_in);
200 stream.next_in = in;
201 stream.next_out = fakebuf;
202 stream.avail_out = sizeof(fakebuf);
203 st = inflate(&stream, Z_FINISH);
204 offset += stream.next_in - in;
205 } while (st == Z_OK || st == Z_BUF_ERROR);
206 inflateEnd(&stream);
207 return (st == Z_STREAM_END &&
208 stream.total_out == expect &&
209 stream.total_in == len) ? 0 : -1;
212 static void copy_pack_data(struct sha1file *f,
213 struct packed_git *p,
214 struct pack_window **w_curs,
215 off_t offset,
216 off_t len)
218 unsigned char *in;
219 unsigned int avail;
221 while (len) {
222 in = use_pack(p, w_curs, offset, &avail);
223 if (avail > len)
224 avail = (unsigned int)len;
225 sha1write(f, in, avail);
226 offset += avail;
227 len -= avail;
231 static unsigned long write_object(struct sha1file *f,
232 struct object_entry *entry,
233 off_t write_offset)
235 unsigned long size, limit, datalen;
236 void *buf;
237 unsigned char header[10], dheader[10];
238 unsigned hdrlen;
239 enum object_type type;
240 int usable_delta, to_reuse;
242 if (!pack_to_stdout)
243 crc32_begin(f);
245 type = entry->type;
247 /* write limit if limited packsize and not first object */
248 limit = pack_size_limit && nr_written ?
249 pack_size_limit - write_offset : 0;
251 if (!entry->delta)
252 usable_delta = 0; /* no delta */
253 else if (!pack_size_limit)
254 usable_delta = 1; /* unlimited packfile */
255 else if (entry->delta->idx.offset == (off_t)-1)
256 usable_delta = 0; /* base was written to another pack */
257 else if (entry->delta->idx.offset)
258 usable_delta = 1; /* base already exists in this pack */
259 else
260 usable_delta = 0; /* base could end up in another pack */
262 if (!reuse_object)
263 to_reuse = 0; /* explicit */
264 else if (!entry->in_pack)
265 to_reuse = 0; /* can't reuse what we don't have */
266 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
267 /* check_object() decided it for us ... */
268 to_reuse = usable_delta;
269 /* ... but pack split may override that */
270 else if (type != entry->in_pack_type)
271 to_reuse = 0; /* pack has delta which is unusable */
272 else if (entry->delta)
273 to_reuse = 0; /* we want to pack afresh */
274 else
275 to_reuse = 1; /* we have it in-pack undeltified,
276 * and we do not need to deltify it.
279 if (!to_reuse) {
280 if (!usable_delta) {
281 buf = read_sha1_file(entry->idx.sha1, &type, &size);
282 if (!buf)
283 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
285 * make sure no cached delta data remains from a
286 * previous attempt before a pack split occured.
288 free(entry->delta_data);
289 entry->delta_data = NULL;
290 entry->z_delta_size = 0;
291 } else if (entry->delta_data) {
292 size = entry->delta_size;
293 buf = entry->delta_data;
294 entry->delta_data = NULL;
295 type = (allow_ofs_delta && entry->delta->idx.offset) ?
296 OBJ_OFS_DELTA : OBJ_REF_DELTA;
297 } else {
298 buf = get_delta(entry);
299 size = entry->delta_size;
300 type = (allow_ofs_delta && entry->delta->idx.offset) ?
301 OBJ_OFS_DELTA : OBJ_REF_DELTA;
304 if (entry->z_delta_size)
305 datalen = entry->z_delta_size;
306 else
307 datalen = do_compress(&buf, size);
310 * The object header is a byte of 'type' followed by zero or
311 * more bytes of length.
313 hdrlen = encode_header(type, size, header);
315 if (type == OBJ_OFS_DELTA) {
317 * Deltas with relative base contain an additional
318 * encoding of the relative offset for the delta
319 * base from this object's position in the pack.
321 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
322 unsigned pos = sizeof(dheader) - 1;
323 dheader[pos] = ofs & 127;
324 while (ofs >>= 7)
325 dheader[--pos] = 128 | (--ofs & 127);
326 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
327 free(buf);
328 return 0;
330 sha1write(f, header, hdrlen);
331 sha1write(f, dheader + pos, sizeof(dheader) - pos);
332 hdrlen += sizeof(dheader) - pos;
333 } else if (type == OBJ_REF_DELTA) {
335 * Deltas with a base reference contain
336 * an additional 20 bytes for the base sha1.
338 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
339 free(buf);
340 return 0;
342 sha1write(f, header, hdrlen);
343 sha1write(f, entry->delta->idx.sha1, 20);
344 hdrlen += 20;
345 } else {
346 if (limit && hdrlen + datalen + 20 >= limit) {
347 free(buf);
348 return 0;
350 sha1write(f, header, hdrlen);
352 sha1write(f, buf, datalen);
353 free(buf);
355 else {
356 struct packed_git *p = entry->in_pack;
357 struct pack_window *w_curs = NULL;
358 struct revindex_entry *revidx;
359 off_t offset;
361 if (entry->delta) {
362 type = (allow_ofs_delta && entry->delta->idx.offset) ?
363 OBJ_OFS_DELTA : OBJ_REF_DELTA;
364 reused_delta++;
366 hdrlen = encode_header(type, entry->size, header);
367 offset = entry->in_pack_offset;
368 revidx = find_pack_revindex(p, offset);
369 datalen = revidx[1].offset - offset;
370 if (!pack_to_stdout && p->index_version > 1 &&
371 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
372 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
373 offset += entry->in_pack_header_size;
374 datalen -= entry->in_pack_header_size;
375 if (type == OBJ_OFS_DELTA) {
376 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
377 unsigned pos = sizeof(dheader) - 1;
378 dheader[pos] = ofs & 127;
379 while (ofs >>= 7)
380 dheader[--pos] = 128 | (--ofs & 127);
381 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
382 return 0;
383 sha1write(f, header, hdrlen);
384 sha1write(f, dheader + pos, sizeof(dheader) - pos);
385 hdrlen += sizeof(dheader) - pos;
386 } else if (type == OBJ_REF_DELTA) {
387 if (limit && hdrlen + 20 + datalen + 20 >= limit)
388 return 0;
389 sha1write(f, header, hdrlen);
390 sha1write(f, entry->delta->idx.sha1, 20);
391 hdrlen += 20;
392 } else {
393 if (limit && hdrlen + datalen + 20 >= limit)
394 return 0;
395 sha1write(f, header, hdrlen);
398 if (!pack_to_stdout && p->index_version == 1 &&
399 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
400 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
401 copy_pack_data(f, p, &w_curs, offset, datalen);
402 unuse_pack(&w_curs);
403 reused++;
405 if (usable_delta)
406 written_delta++;
407 written++;
408 if (!pack_to_stdout)
409 entry->idx.crc32 = crc32_end(f);
410 return hdrlen + datalen;
413 static off_t write_one(struct sha1file *f,
414 struct object_entry *e,
415 off_t offset)
417 unsigned long size;
419 /* offset is non zero if object is written already. */
420 if (e->idx.offset || e->preferred_base)
421 return offset;
423 /* if we are deltified, write out base object first. */
424 if (e->delta) {
425 offset = write_one(f, e->delta, offset);
426 if (!offset)
427 return 0;
430 e->idx.offset = offset;
431 size = write_object(f, e, offset);
432 if (!size) {
433 e->idx.offset = 0;
434 return 0;
436 written_list[nr_written++] = &e->idx;
438 /* make sure off_t is sufficiently large not to wrap */
439 if (offset > offset + size)
440 die("pack too large for current definition of off_t");
441 return offset + size;
444 /* forward declaration for write_pack_file */
445 static int adjust_perm(const char *path, mode_t mode);
447 static void write_pack_file(void)
449 uint32_t i = 0, j;
450 struct sha1file *f;
451 off_t offset, offset_one, last_obj_offset = 0;
452 struct pack_header hdr;
453 uint32_t nr_remaining = nr_result;
454 time_t last_mtime = 0;
456 if (progress > pack_to_stdout)
457 progress_state = start_progress("Writing objects", nr_result);
458 written_list = xmalloc(nr_objects * sizeof(*written_list));
460 do {
461 unsigned char sha1[20];
462 char *pack_tmp_name = NULL;
464 if (pack_to_stdout) {
465 f = sha1fd_throughput(1, "<stdout>", progress_state);
466 } else {
467 char tmpname[PATH_MAX];
468 int fd;
469 snprintf(tmpname, sizeof(tmpname),
470 "%s/tmp_pack_XXXXXX", get_object_directory());
471 fd = xmkstemp(tmpname);
472 pack_tmp_name = xstrdup(tmpname);
473 f = sha1fd(fd, pack_tmp_name);
476 hdr.hdr_signature = htonl(PACK_SIGNATURE);
477 hdr.hdr_version = htonl(PACK_VERSION);
478 hdr.hdr_entries = htonl(nr_remaining);
479 sha1write(f, &hdr, sizeof(hdr));
480 offset = sizeof(hdr);
481 nr_written = 0;
482 for (; i < nr_objects; i++) {
483 last_obj_offset = offset;
484 offset_one = write_one(f, objects + i, offset);
485 if (!offset_one)
486 break;
487 offset = offset_one;
488 display_progress(progress_state, written);
492 * Did we write the wrong # entries in the header?
493 * If so, rewrite it like in fast-import
495 if (pack_to_stdout) {
496 sha1close(f, sha1, CSUM_CLOSE);
497 } else if (nr_written == nr_remaining) {
498 sha1close(f, sha1, CSUM_FSYNC);
499 } else {
500 int fd = sha1close(f, NULL, 0);
501 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
502 close(fd);
505 if (!pack_to_stdout) {
506 mode_t mode = umask(0);
507 struct stat st;
508 char *idx_tmp_name, tmpname[PATH_MAX];
510 umask(mode);
511 mode = 0444 & ~mode;
513 idx_tmp_name = write_idx_file(NULL, written_list,
514 nr_written, sha1);
516 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
517 base_name, sha1_to_hex(sha1));
518 if (adjust_perm(pack_tmp_name, mode))
519 die("unable to make temporary pack file readable: %s",
520 strerror(errno));
521 if (rename(pack_tmp_name, tmpname))
522 die("unable to rename temporary pack file: %s",
523 strerror(errno));
526 * Packs are runtime accessed in their mtime
527 * order since newer packs are more likely to contain
528 * younger objects. So if we are creating multiple
529 * packs then we should modify the mtime of later ones
530 * to preserve this property.
532 if (stat(tmpname, &st) < 0) {
533 warning("failed to stat %s: %s",
534 tmpname, strerror(errno));
535 } else if (!last_mtime) {
536 last_mtime = st.st_mtime;
537 } else {
538 struct utimbuf utb;
539 utb.actime = st.st_atime;
540 utb.modtime = --last_mtime;
541 if (utime(tmpname, &utb) < 0)
542 warning("failed utime() on %s: %s",
543 tmpname, strerror(errno));
546 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
547 base_name, sha1_to_hex(sha1));
548 if (adjust_perm(idx_tmp_name, mode))
549 die("unable to make temporary index file readable: %s",
550 strerror(errno));
551 if (rename(idx_tmp_name, tmpname))
552 die("unable to rename temporary index file: %s",
553 strerror(errno));
555 free(idx_tmp_name);
556 free(pack_tmp_name);
557 puts(sha1_to_hex(sha1));
560 /* mark written objects as written to previous pack */
561 for (j = 0; j < nr_written; j++) {
562 written_list[j]->offset = (off_t)-1;
564 nr_remaining -= nr_written;
565 } while (nr_remaining && i < nr_objects);
567 free(written_list);
568 stop_progress(&progress_state);
569 if (written != nr_result)
570 die("wrote %"PRIu32" objects while expecting %"PRIu32,
571 written, nr_result);
573 * We have scanned through [0 ... i). Since we have written
574 * the correct number of objects, the remaining [i ... nr_objects)
575 * items must be either already written (due to out-of-order delta base)
576 * or a preferred base. Count those which are neither and complain if any.
578 for (j = 0; i < nr_objects; i++) {
579 struct object_entry *e = objects + i;
580 j += !e->idx.offset && !e->preferred_base;
582 if (j)
583 die("wrote %"PRIu32" objects as expected but %"PRIu32
584 " unwritten", written, j);
587 static int locate_object_entry_hash(const unsigned char *sha1)
589 int i;
590 unsigned int ui;
591 memcpy(&ui, sha1, sizeof(unsigned int));
592 i = ui % object_ix_hashsz;
593 while (0 < object_ix[i]) {
594 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
595 return i;
596 if (++i == object_ix_hashsz)
597 i = 0;
599 return -1 - i;
602 static struct object_entry *locate_object_entry(const unsigned char *sha1)
604 int i;
606 if (!object_ix_hashsz)
607 return NULL;
609 i = locate_object_entry_hash(sha1);
610 if (0 <= i)
611 return &objects[object_ix[i]-1];
612 return NULL;
615 static void rehash_objects(void)
617 uint32_t i;
618 struct object_entry *oe;
620 object_ix_hashsz = nr_objects * 3;
621 if (object_ix_hashsz < 1024)
622 object_ix_hashsz = 1024;
623 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
624 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
625 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
626 int ix = locate_object_entry_hash(oe->idx.sha1);
627 if (0 <= ix)
628 continue;
629 ix = -1 - ix;
630 object_ix[ix] = i + 1;
634 static unsigned name_hash(const char *name)
636 unsigned char c;
637 unsigned hash = 0;
639 if (!name)
640 return 0;
643 * This effectively just creates a sortable number from the
644 * last sixteen non-whitespace characters. Last characters
645 * count "most", so things that end in ".c" sort together.
647 while ((c = *name++) != 0) {
648 if (isspace(c))
649 continue;
650 hash = (hash >> 2) + (c << 24);
652 return hash;
655 static void setup_delta_attr_check(struct git_attr_check *check)
657 static struct git_attr *attr_delta;
659 if (!attr_delta)
660 attr_delta = git_attr("delta", 5);
662 check[0].attr = attr_delta;
665 static int no_try_delta(const char *path)
667 struct git_attr_check check[1];
669 setup_delta_attr_check(check);
670 if (git_checkattr(path, ARRAY_SIZE(check), check))
671 return 0;
672 if (ATTR_FALSE(check->value))
673 return 1;
674 return 0;
677 static int add_object_entry(const unsigned char *sha1, enum object_type type,
678 const char *name, int exclude)
680 struct object_entry *entry;
681 struct packed_git *p, *found_pack = NULL;
682 off_t found_offset = 0;
683 int ix;
684 unsigned hash = name_hash(name);
686 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
687 if (ix >= 0) {
688 if (exclude) {
689 entry = objects + object_ix[ix] - 1;
690 if (!entry->preferred_base)
691 nr_result--;
692 entry->preferred_base = 1;
694 return 0;
697 for (p = packed_git; p; p = p->next) {
698 off_t offset = find_pack_entry_one(sha1, p);
699 if (offset) {
700 if (!found_pack) {
701 found_offset = offset;
702 found_pack = p;
704 if (exclude)
705 break;
706 if (incremental)
707 return 0;
708 if (local && !p->pack_local)
709 return 0;
713 if (nr_objects >= nr_alloc) {
714 nr_alloc = (nr_alloc + 1024) * 3 / 2;
715 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
718 entry = objects + nr_objects++;
719 memset(entry, 0, sizeof(*entry));
720 hashcpy(entry->idx.sha1, sha1);
721 entry->hash = hash;
722 if (type)
723 entry->type = type;
724 if (exclude)
725 entry->preferred_base = 1;
726 else
727 nr_result++;
728 if (found_pack) {
729 entry->in_pack = found_pack;
730 entry->in_pack_offset = found_offset;
733 if (object_ix_hashsz * 3 <= nr_objects * 4)
734 rehash_objects();
735 else
736 object_ix[-1 - ix] = nr_objects;
738 display_progress(progress_state, nr_objects);
740 if (name && no_try_delta(name))
741 entry->no_try_delta = 1;
743 return 1;
746 struct pbase_tree_cache {
747 unsigned char sha1[20];
748 int ref;
749 int temporary;
750 void *tree_data;
751 unsigned long tree_size;
754 static struct pbase_tree_cache *(pbase_tree_cache[256]);
755 static int pbase_tree_cache_ix(const unsigned char *sha1)
757 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
759 static int pbase_tree_cache_ix_incr(int ix)
761 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
764 static struct pbase_tree {
765 struct pbase_tree *next;
766 /* This is a phony "cache" entry; we are not
767 * going to evict it nor find it through _get()
768 * mechanism -- this is for the toplevel node that
769 * would almost always change with any commit.
771 struct pbase_tree_cache pcache;
772 } *pbase_tree;
774 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
776 struct pbase_tree_cache *ent, *nent;
777 void *data;
778 unsigned long size;
779 enum object_type type;
780 int neigh;
781 int my_ix = pbase_tree_cache_ix(sha1);
782 int available_ix = -1;
784 /* pbase-tree-cache acts as a limited hashtable.
785 * your object will be found at your index or within a few
786 * slots after that slot if it is cached.
788 for (neigh = 0; neigh < 8; neigh++) {
789 ent = pbase_tree_cache[my_ix];
790 if (ent && !hashcmp(ent->sha1, sha1)) {
791 ent->ref++;
792 return ent;
794 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
795 ((0 <= available_ix) &&
796 (!ent && pbase_tree_cache[available_ix])))
797 available_ix = my_ix;
798 if (!ent)
799 break;
800 my_ix = pbase_tree_cache_ix_incr(my_ix);
803 /* Did not find one. Either we got a bogus request or
804 * we need to read and perhaps cache.
806 data = read_sha1_file(sha1, &type, &size);
807 if (!data)
808 return NULL;
809 if (type != OBJ_TREE) {
810 free(data);
811 return NULL;
814 /* We need to either cache or return a throwaway copy */
816 if (available_ix < 0)
817 ent = NULL;
818 else {
819 ent = pbase_tree_cache[available_ix];
820 my_ix = available_ix;
823 if (!ent) {
824 nent = xmalloc(sizeof(*nent));
825 nent->temporary = (available_ix < 0);
827 else {
828 /* evict and reuse */
829 free(ent->tree_data);
830 nent = ent;
832 hashcpy(nent->sha1, sha1);
833 nent->tree_data = data;
834 nent->tree_size = size;
835 nent->ref = 1;
836 if (!nent->temporary)
837 pbase_tree_cache[my_ix] = nent;
838 return nent;
841 static void pbase_tree_put(struct pbase_tree_cache *cache)
843 if (!cache->temporary) {
844 cache->ref--;
845 return;
847 free(cache->tree_data);
848 free(cache);
851 static int name_cmp_len(const char *name)
853 int i;
854 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
856 return i;
859 static void add_pbase_object(struct tree_desc *tree,
860 const char *name,
861 int cmplen,
862 const char *fullname)
864 struct name_entry entry;
865 int cmp;
867 while (tree_entry(tree,&entry)) {
868 if (S_ISGITLINK(entry.mode))
869 continue;
870 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
871 memcmp(name, entry.path, cmplen);
872 if (cmp > 0)
873 continue;
874 if (cmp < 0)
875 return;
876 if (name[cmplen] != '/') {
877 add_object_entry(entry.sha1,
878 object_type(entry.mode),
879 fullname, 1);
880 return;
882 if (S_ISDIR(entry.mode)) {
883 struct tree_desc sub;
884 struct pbase_tree_cache *tree;
885 const char *down = name+cmplen+1;
886 int downlen = name_cmp_len(down);
888 tree = pbase_tree_get(entry.sha1);
889 if (!tree)
890 return;
891 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
893 add_pbase_object(&sub, down, downlen, fullname);
894 pbase_tree_put(tree);
899 static unsigned *done_pbase_paths;
900 static int done_pbase_paths_num;
901 static int done_pbase_paths_alloc;
902 static int done_pbase_path_pos(unsigned hash)
904 int lo = 0;
905 int hi = done_pbase_paths_num;
906 while (lo < hi) {
907 int mi = (hi + lo) / 2;
908 if (done_pbase_paths[mi] == hash)
909 return mi;
910 if (done_pbase_paths[mi] < hash)
911 hi = mi;
912 else
913 lo = mi + 1;
915 return -lo-1;
918 static int check_pbase_path(unsigned hash)
920 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
921 if (0 <= pos)
922 return 1;
923 pos = -pos - 1;
924 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
925 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
926 done_pbase_paths = xrealloc(done_pbase_paths,
927 done_pbase_paths_alloc *
928 sizeof(unsigned));
930 done_pbase_paths_num++;
931 if (pos < done_pbase_paths_num)
932 memmove(done_pbase_paths + pos + 1,
933 done_pbase_paths + pos,
934 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
935 done_pbase_paths[pos] = hash;
936 return 0;
939 static void add_preferred_base_object(const char *name)
941 struct pbase_tree *it;
942 int cmplen;
943 unsigned hash = name_hash(name);
945 if (!num_preferred_base || check_pbase_path(hash))
946 return;
948 cmplen = name_cmp_len(name);
949 for (it = pbase_tree; it; it = it->next) {
950 if (cmplen == 0) {
951 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
953 else {
954 struct tree_desc tree;
955 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
956 add_pbase_object(&tree, name, cmplen, name);
961 static void add_preferred_base(unsigned char *sha1)
963 struct pbase_tree *it;
964 void *data;
965 unsigned long size;
966 unsigned char tree_sha1[20];
968 if (window <= num_preferred_base++)
969 return;
971 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
972 if (!data)
973 return;
975 for (it = pbase_tree; it; it = it->next) {
976 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
977 free(data);
978 return;
982 it = xcalloc(1, sizeof(*it));
983 it->next = pbase_tree;
984 pbase_tree = it;
986 hashcpy(it->pcache.sha1, tree_sha1);
987 it->pcache.tree_data = data;
988 it->pcache.tree_size = size;
991 static void check_object(struct object_entry *entry)
993 if (entry->in_pack) {
994 struct packed_git *p = entry->in_pack;
995 struct pack_window *w_curs = NULL;
996 const unsigned char *base_ref = NULL;
997 struct object_entry *base_entry;
998 unsigned long used, used_0;
999 unsigned int avail;
1000 off_t ofs;
1001 unsigned char *buf, c;
1003 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1006 * We want in_pack_type even if we do not reuse delta
1007 * since non-delta representations could still be reused.
1009 used = unpack_object_header_gently(buf, avail,
1010 &entry->in_pack_type,
1011 &entry->size);
1014 * Determine if this is a delta and if so whether we can
1015 * reuse it or not. Otherwise let's find out as cheaply as
1016 * possible what the actual type and size for this object is.
1018 switch (entry->in_pack_type) {
1019 default:
1020 /* Not a delta hence we've already got all we need. */
1021 entry->type = entry->in_pack_type;
1022 entry->in_pack_header_size = used;
1023 unuse_pack(&w_curs);
1024 return;
1025 case OBJ_REF_DELTA:
1026 if (reuse_delta && !entry->preferred_base)
1027 base_ref = use_pack(p, &w_curs,
1028 entry->in_pack_offset + used, NULL);
1029 entry->in_pack_header_size = used + 20;
1030 break;
1031 case OBJ_OFS_DELTA:
1032 buf = use_pack(p, &w_curs,
1033 entry->in_pack_offset + used, NULL);
1034 used_0 = 0;
1035 c = buf[used_0++];
1036 ofs = c & 127;
1037 while (c & 128) {
1038 ofs += 1;
1039 if (!ofs || MSB(ofs, 7))
1040 die("delta base offset overflow in pack for %s",
1041 sha1_to_hex(entry->idx.sha1));
1042 c = buf[used_0++];
1043 ofs = (ofs << 7) + (c & 127);
1045 if (ofs >= entry->in_pack_offset)
1046 die("delta base offset out of bound for %s",
1047 sha1_to_hex(entry->idx.sha1));
1048 ofs = entry->in_pack_offset - ofs;
1049 if (reuse_delta && !entry->preferred_base) {
1050 struct revindex_entry *revidx;
1051 revidx = find_pack_revindex(p, ofs);
1052 base_ref = nth_packed_object_sha1(p, revidx->nr);
1054 entry->in_pack_header_size = used + used_0;
1055 break;
1058 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1060 * If base_ref was set above that means we wish to
1061 * reuse delta data, and we even found that base
1062 * in the list of objects we want to pack. Goodie!
1064 * Depth value does not matter - find_deltas() will
1065 * never consider reused delta as the base object to
1066 * deltify other objects against, in order to avoid
1067 * circular deltas.
1069 entry->type = entry->in_pack_type;
1070 entry->delta = base_entry;
1071 entry->delta_sibling = base_entry->delta_child;
1072 base_entry->delta_child = entry;
1073 unuse_pack(&w_curs);
1074 return;
1077 if (entry->type) {
1079 * This must be a delta and we already know what the
1080 * final object type is. Let's extract the actual
1081 * object size from the delta header.
1083 entry->size = get_size_from_delta(p, &w_curs,
1084 entry->in_pack_offset + entry->in_pack_header_size);
1085 unuse_pack(&w_curs);
1086 return;
1090 * No choice but to fall back to the recursive delta walk
1091 * with sha1_object_info() to find about the object type
1092 * at this point...
1094 unuse_pack(&w_curs);
1097 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1098 if (entry->type < 0)
1099 die("unable to get type of object %s",
1100 sha1_to_hex(entry->idx.sha1));
1103 static int pack_offset_sort(const void *_a, const void *_b)
1105 const struct object_entry *a = *(struct object_entry **)_a;
1106 const struct object_entry *b = *(struct object_entry **)_b;
1108 /* avoid filesystem trashing with loose objects */
1109 if (!a->in_pack && !b->in_pack)
1110 return hashcmp(a->idx.sha1, b->idx.sha1);
1112 if (a->in_pack < b->in_pack)
1113 return -1;
1114 if (a->in_pack > b->in_pack)
1115 return 1;
1116 return a->in_pack_offset < b->in_pack_offset ? -1 :
1117 (a->in_pack_offset > b->in_pack_offset);
1120 static void get_object_details(void)
1122 uint32_t i;
1123 struct object_entry **sorted_by_offset;
1125 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1126 for (i = 0; i < nr_objects; i++)
1127 sorted_by_offset[i] = objects + i;
1128 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1130 for (i = 0; i < nr_objects; i++)
1131 check_object(sorted_by_offset[i]);
1133 free(sorted_by_offset);
1137 * We search for deltas in a list sorted by type, by filename hash, and then
1138 * by size, so that we see progressively smaller and smaller files.
1139 * That's because we prefer deltas to be from the bigger file
1140 * to the smaller -- deletes are potentially cheaper, but perhaps
1141 * more importantly, the bigger file is likely the more recent
1142 * one. The deepest deltas are therefore the oldest objects which are
1143 * less susceptible to be accessed often.
1145 static int type_size_sort(const void *_a, const void *_b)
1147 const struct object_entry *a = *(struct object_entry **)_a;
1148 const struct object_entry *b = *(struct object_entry **)_b;
1150 if (a->type > b->type)
1151 return -1;
1152 if (a->type < b->type)
1153 return 1;
1154 if (a->hash > b->hash)
1155 return -1;
1156 if (a->hash < b->hash)
1157 return 1;
1158 if (a->preferred_base > b->preferred_base)
1159 return -1;
1160 if (a->preferred_base < b->preferred_base)
1161 return 1;
1162 if (a->size > b->size)
1163 return -1;
1164 if (a->size < b->size)
1165 return 1;
1166 return a < b ? -1 : (a > b); /* newest first */
1169 struct unpacked {
1170 struct object_entry *entry;
1171 void *data;
1172 struct delta_index *index;
1173 unsigned depth;
1176 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1177 unsigned long delta_size)
1179 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1180 return 0;
1182 if (delta_size < cache_max_small_delta_size)
1183 return 1;
1185 /* cache delta, if objects are large enough compared to delta size */
1186 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1187 return 1;
1189 return 0;
1192 #ifdef THREADED_DELTA_SEARCH
1194 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1195 #define read_lock() pthread_mutex_lock(&read_mutex)
1196 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1198 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1199 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1200 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1202 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1203 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1204 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1206 #else
1208 #define read_lock() (void)0
1209 #define read_unlock() (void)0
1210 #define cache_lock() (void)0
1211 #define cache_unlock() (void)0
1212 #define progress_lock() (void)0
1213 #define progress_unlock() (void)0
1215 #endif
1217 static int try_delta(struct unpacked *trg, struct unpacked *src,
1218 unsigned max_depth, unsigned long *mem_usage)
1220 struct object_entry *trg_entry = trg->entry;
1221 struct object_entry *src_entry = src->entry;
1222 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1223 unsigned ref_depth;
1224 enum object_type type;
1225 void *delta_buf;
1227 /* Don't bother doing diffs between different types */
1228 if (trg_entry->type != src_entry->type)
1229 return -1;
1232 * We do not bother to try a delta that we discarded
1233 * on an earlier try, but only when reusing delta data.
1235 if (reuse_delta && trg_entry->in_pack &&
1236 trg_entry->in_pack == src_entry->in_pack &&
1237 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1238 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1239 return 0;
1241 /* Let's not bust the allowed depth. */
1242 if (src->depth >= max_depth)
1243 return 0;
1245 /* Now some size filtering heuristics. */
1246 trg_size = trg_entry->size;
1247 if (!trg_entry->delta) {
1248 max_size = trg_size/2 - 20;
1249 ref_depth = 1;
1250 } else {
1251 max_size = trg_entry->delta_size;
1252 ref_depth = trg->depth;
1254 max_size = max_size * (max_depth - src->depth) /
1255 (max_depth - ref_depth + 1);
1256 if (max_size == 0)
1257 return 0;
1258 src_size = src_entry->size;
1259 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1260 if (sizediff >= max_size)
1261 return 0;
1262 if (trg_size < src_size / 32)
1263 return 0;
1265 /* Load data if not already done */
1266 if (!trg->data) {
1267 read_lock();
1268 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1269 read_unlock();
1270 if (!trg->data)
1271 die("object %s cannot be read",
1272 sha1_to_hex(trg_entry->idx.sha1));
1273 if (sz != trg_size)
1274 die("object %s inconsistent object length (%lu vs %lu)",
1275 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1276 *mem_usage += sz;
1278 if (!src->data) {
1279 read_lock();
1280 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1281 read_unlock();
1282 if (!src->data)
1283 die("object %s cannot be read",
1284 sha1_to_hex(src_entry->idx.sha1));
1285 if (sz != src_size)
1286 die("object %s inconsistent object length (%lu vs %lu)",
1287 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1288 *mem_usage += sz;
1290 if (!src->index) {
1291 src->index = create_delta_index(src->data, src_size);
1292 if (!src->index) {
1293 static int warned = 0;
1294 if (!warned++)
1295 warning("suboptimal pack - out of memory");
1296 return 0;
1298 *mem_usage += sizeof_delta_index(src->index);
1301 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1302 if (!delta_buf)
1303 return 0;
1305 if (trg_entry->delta) {
1306 /* Prefer only shallower same-sized deltas. */
1307 if (delta_size == trg_entry->delta_size &&
1308 src->depth + 1 >= trg->depth) {
1309 free(delta_buf);
1310 return 0;
1315 * Handle memory allocation outside of the cache
1316 * accounting lock. Compiler will optimize the strangeness
1317 * away when THREADED_DELTA_SEARCH is not defined.
1319 free(trg_entry->delta_data);
1320 cache_lock();
1321 if (trg_entry->delta_data) {
1322 delta_cache_size -= trg_entry->delta_size;
1323 trg_entry->delta_data = NULL;
1325 if (delta_cacheable(src_size, trg_size, delta_size)) {
1326 delta_cache_size += delta_size;
1327 cache_unlock();
1328 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1329 } else {
1330 cache_unlock();
1331 free(delta_buf);
1334 trg_entry->delta = src_entry;
1335 trg_entry->delta_size = delta_size;
1336 trg->depth = src->depth + 1;
1338 return 1;
1341 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1343 struct object_entry *child = me->delta_child;
1344 unsigned int m = n;
1345 while (child) {
1346 unsigned int c = check_delta_limit(child, n + 1);
1347 if (m < c)
1348 m = c;
1349 child = child->delta_sibling;
1351 return m;
1354 static unsigned long free_unpacked(struct unpacked *n)
1356 unsigned long freed_mem = sizeof_delta_index(n->index);
1357 free_delta_index(n->index);
1358 n->index = NULL;
1359 if (n->data) {
1360 freed_mem += n->entry->size;
1361 free(n->data);
1362 n->data = NULL;
1364 n->entry = NULL;
1365 n->depth = 0;
1366 return freed_mem;
1369 static void find_deltas(struct object_entry **list, unsigned *list_size,
1370 int window, int depth, unsigned *processed)
1372 uint32_t i, idx = 0, count = 0;
1373 unsigned int array_size = window * sizeof(struct unpacked);
1374 struct unpacked *array;
1375 unsigned long mem_usage = 0;
1377 array = xmalloc(array_size);
1378 memset(array, 0, array_size);
1380 for (;;) {
1381 struct object_entry *entry = *list++;
1382 struct unpacked *n = array + idx;
1383 int j, max_depth, best_base = -1;
1385 progress_lock();
1386 if (!*list_size) {
1387 progress_unlock();
1388 break;
1390 (*list_size)--;
1391 if (!entry->preferred_base) {
1392 (*processed)++;
1393 display_progress(progress_state, *processed);
1395 progress_unlock();
1397 mem_usage -= free_unpacked(n);
1398 n->entry = entry;
1400 while (window_memory_limit &&
1401 mem_usage > window_memory_limit &&
1402 count > 1) {
1403 uint32_t tail = (idx + window - count) % window;
1404 mem_usage -= free_unpacked(array + tail);
1405 count--;
1408 /* We do not compute delta to *create* objects we are not
1409 * going to pack.
1411 if (entry->preferred_base)
1412 goto next;
1415 * If the current object is at pack edge, take the depth the
1416 * objects that depend on the current object into account
1417 * otherwise they would become too deep.
1419 max_depth = depth;
1420 if (entry->delta_child) {
1421 max_depth -= check_delta_limit(entry, 0);
1422 if (max_depth <= 0)
1423 goto next;
1426 j = window;
1427 while (--j > 0) {
1428 int ret;
1429 uint32_t other_idx = idx + j;
1430 struct unpacked *m;
1431 if (other_idx >= window)
1432 other_idx -= window;
1433 m = array + other_idx;
1434 if (!m->entry)
1435 break;
1436 ret = try_delta(n, m, max_depth, &mem_usage);
1437 if (ret < 0)
1438 break;
1439 else if (ret > 0)
1440 best_base = other_idx;
1444 * If we decided to cache the delta data, then it is best
1445 * to compress it right away. First because we have to do
1446 * it anyway, and doing it here while we're threaded will
1447 * save a lot of time in the non threaded write phase,
1448 * as well as allow for caching more deltas within
1449 * the same cache size limit.
1450 * ...
1451 * But only if not writing to stdout, since in that case
1452 * the network is most likely throttling writes anyway,
1453 * and therefore it is best to go to the write phase ASAP
1454 * instead, as we can afford spending more time compressing
1455 * between writes at that moment.
1457 if (entry->delta_data && !pack_to_stdout) {
1458 entry->z_delta_size = do_compress(&entry->delta_data,
1459 entry->delta_size);
1460 cache_lock();
1461 delta_cache_size -= entry->delta_size;
1462 delta_cache_size += entry->z_delta_size;
1463 cache_unlock();
1466 /* if we made n a delta, and if n is already at max
1467 * depth, leaving it in the window is pointless. we
1468 * should evict it first.
1470 if (entry->delta && max_depth <= n->depth)
1471 continue;
1474 * Move the best delta base up in the window, after the
1475 * currently deltified object, to keep it longer. It will
1476 * be the first base object to be attempted next.
1478 if (entry->delta) {
1479 struct unpacked swap = array[best_base];
1480 int dist = (window + idx - best_base) % window;
1481 int dst = best_base;
1482 while (dist--) {
1483 int src = (dst + 1) % window;
1484 array[dst] = array[src];
1485 dst = src;
1487 array[dst] = swap;
1490 next:
1491 idx++;
1492 if (count + 1 < window)
1493 count++;
1494 if (idx >= window)
1495 idx = 0;
1498 for (i = 0; i < window; ++i) {
1499 free_delta_index(array[i].index);
1500 free(array[i].data);
1502 free(array);
1505 #ifdef THREADED_DELTA_SEARCH
1508 * The main thread waits on the condition that (at least) one of the workers
1509 * has stopped working (which is indicated in the .working member of
1510 * struct thread_params).
1511 * When a work thread has completed its work, it sets .working to 0 and
1512 * signals the main thread and waits on the condition that .data_ready
1513 * becomes 1.
1516 struct thread_params {
1517 pthread_t thread;
1518 struct object_entry **list;
1519 unsigned list_size;
1520 unsigned remaining;
1521 int window;
1522 int depth;
1523 int working;
1524 int data_ready;
1525 pthread_mutex_t mutex;
1526 pthread_cond_t cond;
1527 unsigned *processed;
1530 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1532 static void *threaded_find_deltas(void *arg)
1534 struct thread_params *me = arg;
1536 while (me->remaining) {
1537 find_deltas(me->list, &me->remaining,
1538 me->window, me->depth, me->processed);
1540 progress_lock();
1541 me->working = 0;
1542 pthread_cond_signal(&progress_cond);
1543 progress_unlock();
1546 * We must not set ->data_ready before we wait on the
1547 * condition because the main thread may have set it to 1
1548 * before we get here. In order to be sure that new
1549 * work is available if we see 1 in ->data_ready, it
1550 * was initialized to 0 before this thread was spawned
1551 * and we reset it to 0 right away.
1553 pthread_mutex_lock(&me->mutex);
1554 while (!me->data_ready)
1555 pthread_cond_wait(&me->cond, &me->mutex);
1556 me->data_ready = 0;
1557 pthread_mutex_unlock(&me->mutex);
1559 /* leave ->working 1 so that this doesn't get more work assigned */
1560 return NULL;
1563 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1564 int window, int depth, unsigned *processed)
1566 struct thread_params p[delta_search_threads];
1567 int i, ret, active_threads = 0;
1569 if (delta_search_threads <= 1) {
1570 find_deltas(list, &list_size, window, depth, processed);
1571 return;
1574 /* Partition the work amongst work threads. */
1575 for (i = 0; i < delta_search_threads; i++) {
1576 unsigned sub_size = list_size / (delta_search_threads - i);
1578 p[i].window = window;
1579 p[i].depth = depth;
1580 p[i].processed = processed;
1581 p[i].working = 1;
1582 p[i].data_ready = 0;
1584 /* try to split chunks on "path" boundaries */
1585 while (sub_size && sub_size < list_size &&
1586 list[sub_size]->hash &&
1587 list[sub_size]->hash == list[sub_size-1]->hash)
1588 sub_size++;
1590 p[i].list = list;
1591 p[i].list_size = sub_size;
1592 p[i].remaining = sub_size;
1594 list += sub_size;
1595 list_size -= sub_size;
1598 /* Start work threads. */
1599 for (i = 0; i < delta_search_threads; i++) {
1600 if (!p[i].list_size)
1601 continue;
1602 pthread_mutex_init(&p[i].mutex, NULL);
1603 pthread_cond_init(&p[i].cond, NULL);
1604 ret = pthread_create(&p[i].thread, NULL,
1605 threaded_find_deltas, &p[i]);
1606 if (ret)
1607 die("unable to create thread: %s", strerror(ret));
1608 active_threads++;
1612 * Now let's wait for work completion. Each time a thread is done
1613 * with its work, we steal half of the remaining work from the
1614 * thread with the largest number of unprocessed objects and give
1615 * it to that newly idle thread. This ensure good load balancing
1616 * until the remaining object list segments are simply too short
1617 * to be worth splitting anymore.
1619 while (active_threads) {
1620 struct thread_params *target = NULL;
1621 struct thread_params *victim = NULL;
1622 unsigned sub_size = 0;
1624 progress_lock();
1625 for (;;) {
1626 for (i = 0; !target && i < delta_search_threads; i++)
1627 if (!p[i].working)
1628 target = &p[i];
1629 if (target)
1630 break;
1631 pthread_cond_wait(&progress_cond, &progress_mutex);
1634 for (i = 0; i < delta_search_threads; i++)
1635 if (p[i].remaining > 2*window &&
1636 (!victim || victim->remaining < p[i].remaining))
1637 victim = &p[i];
1638 if (victim) {
1639 sub_size = victim->remaining / 2;
1640 list = victim->list + victim->list_size - sub_size;
1641 while (sub_size && list[0]->hash &&
1642 list[0]->hash == list[-1]->hash) {
1643 list++;
1644 sub_size--;
1646 if (!sub_size) {
1648 * It is possible for some "paths" to have
1649 * so many objects that no hash boundary
1650 * might be found. Let's just steal the
1651 * exact half in that case.
1653 sub_size = victim->remaining / 2;
1654 list -= sub_size;
1656 target->list = list;
1657 victim->list_size -= sub_size;
1658 victim->remaining -= sub_size;
1660 target->list_size = sub_size;
1661 target->remaining = sub_size;
1662 target->working = 1;
1663 progress_unlock();
1665 pthread_mutex_lock(&target->mutex);
1666 target->data_ready = 1;
1667 pthread_cond_signal(&target->cond);
1668 pthread_mutex_unlock(&target->mutex);
1670 if (!sub_size) {
1671 pthread_join(target->thread, NULL);
1672 pthread_cond_destroy(&target->cond);
1673 pthread_mutex_destroy(&target->mutex);
1674 active_threads--;
1679 #else
1680 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1681 #endif
1683 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1685 unsigned char peeled[20];
1687 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1688 !peel_ref(path, peeled) && /* peelable? */
1689 !is_null_sha1(peeled) && /* annotated tag? */
1690 locate_object_entry(peeled)) /* object packed? */
1691 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1692 return 0;
1695 static void prepare_pack(int window, int depth)
1697 struct object_entry **delta_list;
1698 uint32_t i, nr_deltas;
1699 unsigned n;
1701 get_object_details();
1703 if (!nr_objects || !window || !depth)
1704 return;
1706 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1707 nr_deltas = n = 0;
1709 for (i = 0; i < nr_objects; i++) {
1710 struct object_entry *entry = objects + i;
1712 if (entry->delta)
1713 /* This happens if we decided to reuse existing
1714 * delta from a pack. "reuse_delta &&" is implied.
1716 continue;
1718 if (entry->size < 50)
1719 continue;
1721 if (entry->no_try_delta)
1722 continue;
1724 if (!entry->preferred_base)
1725 nr_deltas++;
1727 delta_list[n++] = entry;
1730 if (nr_deltas && n > 1) {
1731 unsigned nr_done = 0;
1732 if (progress)
1733 progress_state = start_progress("Compressing objects",
1734 nr_deltas);
1735 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1736 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1737 stop_progress(&progress_state);
1738 if (nr_done != nr_deltas)
1739 die("inconsistency with delta count");
1741 free(delta_list);
1744 static int git_pack_config(const char *k, const char *v, void *cb)
1746 if(!strcmp(k, "pack.window")) {
1747 window = git_config_int(k, v);
1748 return 0;
1750 if (!strcmp(k, "pack.windowmemory")) {
1751 window_memory_limit = git_config_ulong(k, v);
1752 return 0;
1754 if (!strcmp(k, "pack.depth")) {
1755 depth = git_config_int(k, v);
1756 return 0;
1758 if (!strcmp(k, "pack.compression")) {
1759 int level = git_config_int(k, v);
1760 if (level == -1)
1761 level = Z_DEFAULT_COMPRESSION;
1762 else if (level < 0 || level > Z_BEST_COMPRESSION)
1763 die("bad pack compression level %d", level);
1764 pack_compression_level = level;
1765 pack_compression_seen = 1;
1766 return 0;
1768 if (!strcmp(k, "pack.deltacachesize")) {
1769 max_delta_cache_size = git_config_int(k, v);
1770 return 0;
1772 if (!strcmp(k, "pack.deltacachelimit")) {
1773 cache_max_small_delta_size = git_config_int(k, v);
1774 return 0;
1776 if (!strcmp(k, "pack.threads")) {
1777 delta_search_threads = git_config_int(k, v);
1778 if (delta_search_threads < 0)
1779 die("invalid number of threads specified (%d)",
1780 delta_search_threads);
1781 #ifndef THREADED_DELTA_SEARCH
1782 if (delta_search_threads != 1)
1783 warning("no threads support, ignoring %s", k);
1784 #endif
1785 return 0;
1787 if (!strcmp(k, "pack.indexversion")) {
1788 pack_idx_default_version = git_config_int(k, v);
1789 if (pack_idx_default_version > 2)
1790 die("bad pack.indexversion=%"PRIu32,
1791 pack_idx_default_version);
1792 return 0;
1794 if (!strcmp(k, "pack.packsizelimit")) {
1795 pack_size_limit_cfg = git_config_ulong(k, v);
1796 return 0;
1798 return git_default_config(k, v, cb);
1801 static void read_object_list_from_stdin(void)
1803 char line[40 + 1 + PATH_MAX + 2];
1804 unsigned char sha1[20];
1806 for (;;) {
1807 if (!fgets(line, sizeof(line), stdin)) {
1808 if (feof(stdin))
1809 break;
1810 if (!ferror(stdin))
1811 die("fgets returned NULL, not EOF, not error!");
1812 if (errno != EINTR)
1813 die("fgets: %s", strerror(errno));
1814 clearerr(stdin);
1815 continue;
1817 if (line[0] == '-') {
1818 if (get_sha1_hex(line+1, sha1))
1819 die("expected edge sha1, got garbage:\n %s",
1820 line);
1821 add_preferred_base(sha1);
1822 continue;
1824 if (get_sha1_hex(line, sha1))
1825 die("expected sha1, got garbage:\n %s", line);
1827 add_preferred_base_object(line+41);
1828 add_object_entry(sha1, 0, line+41, 0);
1832 #define OBJECT_ADDED (1u<<20)
1834 static void show_commit(struct commit *commit)
1836 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1837 commit->object.flags |= OBJECT_ADDED;
1840 static void show_object(struct object_array_entry *p)
1842 add_preferred_base_object(p->name);
1843 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1844 p->item->flags |= OBJECT_ADDED;
1847 static void show_edge(struct commit *commit)
1849 add_preferred_base(commit->object.sha1);
1852 struct in_pack_object {
1853 off_t offset;
1854 struct object *object;
1857 struct in_pack {
1858 int alloc;
1859 int nr;
1860 struct in_pack_object *array;
1863 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1865 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1866 in_pack->array[in_pack->nr].object = object;
1867 in_pack->nr++;
1871 * Compare the objects in the offset order, in order to emulate the
1872 * "git-rev-list --objects" output that produced the pack originally.
1874 static int ofscmp(const void *a_, const void *b_)
1876 struct in_pack_object *a = (struct in_pack_object *)a_;
1877 struct in_pack_object *b = (struct in_pack_object *)b_;
1879 if (a->offset < b->offset)
1880 return -1;
1881 else if (a->offset > b->offset)
1882 return 1;
1883 else
1884 return hashcmp(a->object->sha1, b->object->sha1);
1887 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1889 struct packed_git *p;
1890 struct in_pack in_pack;
1891 uint32_t i;
1893 memset(&in_pack, 0, sizeof(in_pack));
1895 for (p = packed_git; p; p = p->next) {
1896 const unsigned char *sha1;
1897 struct object *o;
1899 for (i = 0; i < revs->num_ignore_packed; i++) {
1900 if (matches_pack_name(p, revs->ignore_packed[i]))
1901 break;
1903 if (revs->num_ignore_packed <= i)
1904 continue;
1905 if (open_pack_index(p))
1906 die("cannot open pack index");
1908 ALLOC_GROW(in_pack.array,
1909 in_pack.nr + p->num_objects,
1910 in_pack.alloc);
1912 for (i = 0; i < p->num_objects; i++) {
1913 sha1 = nth_packed_object_sha1(p, i);
1914 o = lookup_unknown_object(sha1);
1915 if (!(o->flags & OBJECT_ADDED))
1916 mark_in_pack_object(o, p, &in_pack);
1917 o->flags |= OBJECT_ADDED;
1921 if (in_pack.nr) {
1922 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1923 ofscmp);
1924 for (i = 0; i < in_pack.nr; i++) {
1925 struct object *o = in_pack.array[i].object;
1926 add_object_entry(o->sha1, o->type, "", 0);
1929 free(in_pack.array);
1932 static void loosen_unused_packed_objects(struct rev_info *revs)
1934 struct packed_git *p;
1935 uint32_t i;
1936 const unsigned char *sha1;
1938 for (p = packed_git; p; p = p->next) {
1939 for (i = 0; i < revs->num_ignore_packed; i++) {
1940 if (matches_pack_name(p, revs->ignore_packed[i]))
1941 break;
1943 if (revs->num_ignore_packed <= i)
1944 continue;
1946 if (open_pack_index(p))
1947 die("cannot open pack index");
1949 for (i = 0; i < p->num_objects; i++) {
1950 sha1 = nth_packed_object_sha1(p, i);
1951 if (!locate_object_entry(sha1))
1952 if (force_object_loose(sha1, p->mtime))
1953 die("unable to force loose object");
1958 static void get_object_list(int ac, const char **av)
1960 struct rev_info revs;
1961 char line[1000];
1962 int flags = 0;
1964 init_revisions(&revs, NULL);
1965 save_commit_buffer = 0;
1966 setup_revisions(ac, av, &revs, NULL);
1968 while (fgets(line, sizeof(line), stdin) != NULL) {
1969 int len = strlen(line);
1970 if (len && line[len - 1] == '\n')
1971 line[--len] = 0;
1972 if (!len)
1973 break;
1974 if (*line == '-') {
1975 if (!strcmp(line, "--not")) {
1976 flags ^= UNINTERESTING;
1977 continue;
1979 die("not a rev '%s'", line);
1981 if (handle_revision_arg(line, &revs, flags, 1))
1982 die("bad revision '%s'", line);
1985 if (prepare_revision_walk(&revs))
1986 die("revision walk setup failed");
1987 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1988 traverse_commit_list(&revs, show_commit, show_object);
1990 if (keep_unreachable)
1991 add_objects_in_unpacked_packs(&revs);
1992 if (unpack_unreachable)
1993 loosen_unused_packed_objects(&revs);
1996 static int adjust_perm(const char *path, mode_t mode)
1998 if (chmod(path, mode))
1999 return -1;
2000 return adjust_shared_perm(path);
2003 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2005 int use_internal_rev_list = 0;
2006 int thin = 0;
2007 uint32_t i;
2008 const char **rp_av;
2009 int rp_ac_alloc = 64;
2010 int rp_ac;
2012 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2014 rp_av[0] = "pack-objects";
2015 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2016 rp_ac = 2;
2018 git_config(git_pack_config, NULL);
2019 if (!pack_compression_seen && core_compression_seen)
2020 pack_compression_level = core_compression_level;
2022 progress = isatty(2);
2023 for (i = 1; i < argc; i++) {
2024 const char *arg = argv[i];
2026 if (*arg != '-')
2027 break;
2029 if (!strcmp("--non-empty", arg)) {
2030 non_empty = 1;
2031 continue;
2033 if (!strcmp("--local", arg)) {
2034 local = 1;
2035 continue;
2037 if (!strcmp("--incremental", arg)) {
2038 incremental = 1;
2039 continue;
2041 if (!prefixcmp(arg, "--compression=")) {
2042 char *end;
2043 int level = strtoul(arg+14, &end, 0);
2044 if (!arg[14] || *end)
2045 usage(pack_usage);
2046 if (level == -1)
2047 level = Z_DEFAULT_COMPRESSION;
2048 else if (level < 0 || level > Z_BEST_COMPRESSION)
2049 die("bad pack compression level %d", level);
2050 pack_compression_level = level;
2051 continue;
2053 if (!prefixcmp(arg, "--max-pack-size=")) {
2054 char *end;
2055 pack_size_limit_cfg = 0;
2056 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2057 if (!arg[16] || *end)
2058 usage(pack_usage);
2059 continue;
2061 if (!prefixcmp(arg, "--window=")) {
2062 char *end;
2063 window = strtoul(arg+9, &end, 0);
2064 if (!arg[9] || *end)
2065 usage(pack_usage);
2066 continue;
2068 if (!prefixcmp(arg, "--window-memory=")) {
2069 if (!git_parse_ulong(arg+16, &window_memory_limit))
2070 usage(pack_usage);
2071 continue;
2073 if (!prefixcmp(arg, "--threads=")) {
2074 char *end;
2075 delta_search_threads = strtoul(arg+10, &end, 0);
2076 if (!arg[10] || *end || delta_search_threads < 0)
2077 usage(pack_usage);
2078 #ifndef THREADED_DELTA_SEARCH
2079 if (delta_search_threads != 1)
2080 warning("no threads support, "
2081 "ignoring %s", arg);
2082 #endif
2083 continue;
2085 if (!prefixcmp(arg, "--depth=")) {
2086 char *end;
2087 depth = strtoul(arg+8, &end, 0);
2088 if (!arg[8] || *end)
2089 usage(pack_usage);
2090 continue;
2092 if (!strcmp("--progress", arg)) {
2093 progress = 1;
2094 continue;
2096 if (!strcmp("--all-progress", arg)) {
2097 progress = 2;
2098 continue;
2100 if (!strcmp("-q", arg)) {
2101 progress = 0;
2102 continue;
2104 if (!strcmp("--no-reuse-delta", arg)) {
2105 reuse_delta = 0;
2106 continue;
2108 if (!strcmp("--no-reuse-object", arg)) {
2109 reuse_object = reuse_delta = 0;
2110 continue;
2112 if (!strcmp("--delta-base-offset", arg)) {
2113 allow_ofs_delta = 1;
2114 continue;
2116 if (!strcmp("--stdout", arg)) {
2117 pack_to_stdout = 1;
2118 continue;
2120 if (!strcmp("--revs", arg)) {
2121 use_internal_rev_list = 1;
2122 continue;
2124 if (!strcmp("--keep-unreachable", arg)) {
2125 keep_unreachable = 1;
2126 continue;
2128 if (!strcmp("--unpack-unreachable", arg)) {
2129 unpack_unreachable = 1;
2130 continue;
2132 if (!strcmp("--include-tag", arg)) {
2133 include_tag = 1;
2134 continue;
2136 if (!strcmp("--unpacked", arg) ||
2137 !prefixcmp(arg, "--unpacked=") ||
2138 !strcmp("--reflog", arg) ||
2139 !strcmp("--all", arg)) {
2140 use_internal_rev_list = 1;
2141 if (rp_ac >= rp_ac_alloc - 1) {
2142 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2143 rp_av = xrealloc(rp_av,
2144 rp_ac_alloc * sizeof(*rp_av));
2146 rp_av[rp_ac++] = arg;
2147 continue;
2149 if (!strcmp("--thin", arg)) {
2150 use_internal_rev_list = 1;
2151 thin = 1;
2152 rp_av[1] = "--objects-edge";
2153 continue;
2155 if (!prefixcmp(arg, "--index-version=")) {
2156 char *c;
2157 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2158 if (pack_idx_default_version > 2)
2159 die("bad %s", arg);
2160 if (*c == ',')
2161 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2162 if (*c || pack_idx_off32_limit & 0x80000000)
2163 die("bad %s", arg);
2164 continue;
2166 usage(pack_usage);
2169 /* Traditionally "pack-objects [options] base extra" failed;
2170 * we would however want to take refs parameter that would
2171 * have been given to upstream rev-list ourselves, which means
2172 * we somehow want to say what the base name is. So the
2173 * syntax would be:
2175 * pack-objects [options] base <refs...>
2177 * in other words, we would treat the first non-option as the
2178 * base_name and send everything else to the internal revision
2179 * walker.
2182 if (!pack_to_stdout)
2183 base_name = argv[i++];
2185 if (pack_to_stdout != !base_name)
2186 usage(pack_usage);
2188 if (!pack_to_stdout && !pack_size_limit)
2189 pack_size_limit = pack_size_limit_cfg;
2191 if (pack_to_stdout && pack_size_limit)
2192 die("--max-pack-size cannot be used to build a pack for transfer.");
2194 if (!pack_to_stdout && thin)
2195 die("--thin cannot be used to build an indexable pack.");
2197 if (keep_unreachable && unpack_unreachable)
2198 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2200 #ifdef THREADED_DELTA_SEARCH
2201 if (!delta_search_threads) /* --threads=0 means autodetect */
2202 delta_search_threads = online_cpus();
2203 #endif
2205 prepare_packed_git();
2207 if (progress)
2208 progress_state = start_progress("Counting objects", 0);
2209 if (!use_internal_rev_list)
2210 read_object_list_from_stdin();
2211 else {
2212 rp_av[rp_ac] = NULL;
2213 get_object_list(rp_ac, rp_av);
2215 if (include_tag && nr_result)
2216 for_each_ref(add_ref_tag, NULL);
2217 stop_progress(&progress_state);
2219 if (non_empty && !nr_result)
2220 return 0;
2221 if (nr_result)
2222 prepare_pack(window, depth);
2223 write_pack_file();
2224 if (progress)
2225 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2226 " reused %"PRIu32" (delta %"PRIu32")\n",
2227 written, written_delta, reused, reused_delta);
2228 return 0;