add storage size output to 'git verify-pack -v'
[git/spearce.git] / builtin-pack-objects.c
blobc7834bb3684131b508257e0e11c81a8934d5b700
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"
19 #ifdef THREADED_DELTA_SEARCH
20 #include "thread-utils.h"
21 #include <pthread.h>
22 #endif
24 static const char pack_usage[] = "\
25 git-pack-objects [{ -q | --progress | --all-progress }] \n\
26 [--max-pack-size=N] [--local] [--incremental] \n\
27 [--window=N] [--window-memory=N] [--depth=N] \n\
28 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
29 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
30 [--stdout | base-name] [--keep-unreachable] [<ref-list | <object-list]";
32 struct object_entry {
33 struct pack_idx_entry idx;
34 unsigned long size; /* uncompressed size */
35 struct packed_git *in_pack; /* already in pack */
36 off_t in_pack_offset;
37 struct object_entry *delta; /* delta base object */
38 struct object_entry *delta_child; /* deltified objects who bases me */
39 struct object_entry *delta_sibling; /* other deltified objects who
40 * uses the same base as me
42 void *delta_data; /* cached delta (uncompressed) */
43 unsigned long delta_size; /* delta data size (uncompressed) */
44 unsigned int hash; /* name hint hash */
45 enum object_type type;
46 enum object_type in_pack_type; /* could be delta */
47 unsigned char in_pack_header_size;
48 unsigned char preferred_base; /* we do not pack this, but is available
49 * to be used as the base object to delta
50 * objects against.
52 unsigned char no_try_delta;
56 * Objects we are going to pack are collected in objects array (dynamically
57 * expanded). nr_objects & nr_alloc controls this array. They are stored
58 * in the order we see -- typically rev-list --objects order that gives us
59 * nice "minimum seek" order.
61 static struct object_entry *objects;
62 static struct pack_idx_entry **written_list;
63 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
65 static int non_empty;
66 static int no_reuse_delta, no_reuse_object, keep_unreachable;
67 static int local;
68 static int incremental;
69 static int allow_ofs_delta;
70 static const char *base_name;
71 static int progress = 1;
72 static int window = 10;
73 static uint32_t pack_size_limit, pack_size_limit_cfg;
74 static int depth = 50;
75 static int delta_search_threads = 1;
76 static int pack_to_stdout;
77 static int num_preferred_base;
78 static struct progress *progress_state;
79 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
80 static int pack_compression_seen;
82 static unsigned long delta_cache_size = 0;
83 static unsigned long max_delta_cache_size = 0;
84 static unsigned long cache_max_small_delta_size = 1000;
86 static unsigned long window_memory_limit = 0;
89 * The object names in objects array are hashed with this hashtable,
90 * to help looking up the entry by object name.
91 * This hashtable is built after all the objects are seen.
93 static int *object_ix;
94 static int object_ix_hashsz;
97 * stats
99 static uint32_t written, written_delta;
100 static uint32_t reused, reused_delta;
103 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
105 unsigned long othersize, delta_size;
106 enum object_type type;
107 void *otherbuf = read_sha1_file(entry->delta->idx.sha1, &type, &othersize);
108 void *delta_buf;
110 if (!otherbuf)
111 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
112 delta_buf = diff_delta(otherbuf, othersize,
113 buf, size, &delta_size, 0);
114 if (!delta_buf || delta_size != entry->delta_size)
115 die("delta size changed");
116 free(buf);
117 free(otherbuf);
118 return delta_buf;
122 * The per-object header is a pretty dense thing, which is
123 * - first byte: low four bits are "size", then three bits of "type",
124 * and the high bit is "size continues".
125 * - each byte afterwards: low seven bits are size continuation,
126 * with the high bit being "size continues"
128 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
130 int n = 1;
131 unsigned char c;
133 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
134 die("bad type %d", type);
136 c = (type << 4) | (size & 15);
137 size >>= 4;
138 while (size) {
139 *hdr++ = c | 0x80;
140 c = size & 0x7f;
141 size >>= 7;
142 n++;
144 *hdr = c;
145 return n;
149 * we are going to reuse the existing object data as is. make
150 * sure it is not corrupt.
152 static int check_pack_inflate(struct packed_git *p,
153 struct pack_window **w_curs,
154 off_t offset,
155 off_t len,
156 unsigned long expect)
158 z_stream stream;
159 unsigned char fakebuf[4096], *in;
160 int st;
162 memset(&stream, 0, sizeof(stream));
163 inflateInit(&stream);
164 do {
165 in = use_pack(p, w_curs, offset, &stream.avail_in);
166 stream.next_in = in;
167 stream.next_out = fakebuf;
168 stream.avail_out = sizeof(fakebuf);
169 st = inflate(&stream, Z_FINISH);
170 offset += stream.next_in - in;
171 } while (st == Z_OK || st == Z_BUF_ERROR);
172 inflateEnd(&stream);
173 return (st == Z_STREAM_END &&
174 stream.total_out == expect &&
175 stream.total_in == len) ? 0 : -1;
178 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
179 off_t offset, off_t len, unsigned int nr)
181 const uint32_t *index_crc;
182 uint32_t data_crc = crc32(0, Z_NULL, 0);
184 do {
185 unsigned int avail;
186 void *data = use_pack(p, w_curs, offset, &avail);
187 if (avail > len)
188 avail = len;
189 data_crc = crc32(data_crc, data, avail);
190 offset += avail;
191 len -= avail;
192 } while (len);
194 index_crc = p->index_data;
195 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
197 return data_crc != ntohl(*index_crc);
200 static void copy_pack_data(struct sha1file *f,
201 struct packed_git *p,
202 struct pack_window **w_curs,
203 off_t offset,
204 off_t len)
206 unsigned char *in;
207 unsigned int avail;
209 while (len) {
210 in = use_pack(p, w_curs, offset, &avail);
211 if (avail > len)
212 avail = (unsigned int)len;
213 sha1write(f, in, avail);
214 offset += avail;
215 len -= avail;
219 static unsigned long write_object(struct sha1file *f,
220 struct object_entry *entry,
221 off_t write_offset)
223 unsigned long size;
224 enum object_type type;
225 void *buf;
226 unsigned char header[10];
227 unsigned char dheader[10];
228 unsigned hdrlen;
229 off_t datalen;
230 enum object_type obj_type;
231 int to_reuse = 0;
232 /* write limit if limited packsize and not first object */
233 unsigned long limit = pack_size_limit && nr_written ?
234 pack_size_limit - write_offset : 0;
235 /* no if no delta */
236 int usable_delta = !entry->delta ? 0 :
237 /* yes if unlimited packfile */
238 !pack_size_limit ? 1 :
239 /* no if base written to previous pack */
240 entry->delta->idx.offset == (off_t)-1 ? 0 :
241 /* otherwise double-check written to this
242 * pack, like we do below
244 entry->delta->idx.offset ? 1 : 0;
246 if (!pack_to_stdout)
247 crc32_begin(f);
249 obj_type = entry->type;
250 if (no_reuse_object)
251 to_reuse = 0; /* explicit */
252 else if (!entry->in_pack)
253 to_reuse = 0; /* can't reuse what we don't have */
254 else if (obj_type == OBJ_REF_DELTA || obj_type == OBJ_OFS_DELTA)
255 /* check_object() decided it for us ... */
256 to_reuse = usable_delta;
257 /* ... but pack split may override that */
258 else if (obj_type != entry->in_pack_type)
259 to_reuse = 0; /* pack has delta which is unusable */
260 else if (entry->delta)
261 to_reuse = 0; /* we want to pack afresh */
262 else
263 to_reuse = 1; /* we have it in-pack undeltified,
264 * and we do not need to deltify it.
267 if (!to_reuse) {
268 z_stream stream;
269 unsigned long maxsize;
270 void *out;
271 if (!usable_delta) {
272 buf = read_sha1_file(entry->idx.sha1, &obj_type, &size);
273 if (!buf)
274 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
275 } else if (entry->delta_data) {
276 size = entry->delta_size;
277 buf = entry->delta_data;
278 entry->delta_data = NULL;
279 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
280 OBJ_OFS_DELTA : OBJ_REF_DELTA;
281 } else {
282 buf = read_sha1_file(entry->idx.sha1, &type, &size);
283 if (!buf)
284 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
285 buf = delta_against(buf, size, entry);
286 size = entry->delta_size;
287 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
288 OBJ_OFS_DELTA : OBJ_REF_DELTA;
290 /* compress the data to store and put compressed length in datalen */
291 memset(&stream, 0, sizeof(stream));
292 deflateInit(&stream, pack_compression_level);
293 maxsize = deflateBound(&stream, size);
294 out = xmalloc(maxsize);
295 /* Compress it */
296 stream.next_in = buf;
297 stream.avail_in = size;
298 stream.next_out = out;
299 stream.avail_out = maxsize;
300 while (deflate(&stream, Z_FINISH) == Z_OK)
301 /* nothing */;
302 deflateEnd(&stream);
303 datalen = stream.total_out;
306 * The object header is a byte of 'type' followed by zero or
307 * more bytes of length.
309 hdrlen = encode_header(obj_type, size, header);
311 if (obj_type == OBJ_OFS_DELTA) {
313 * Deltas with relative base contain an additional
314 * encoding of the relative offset for the delta
315 * base from this object's position in the pack.
317 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
318 unsigned pos = sizeof(dheader) - 1;
319 dheader[pos] = ofs & 127;
320 while (ofs >>= 7)
321 dheader[--pos] = 128 | (--ofs & 127);
322 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
323 free(out);
324 free(buf);
325 return 0;
327 sha1write(f, header, hdrlen);
328 sha1write(f, dheader + pos, sizeof(dheader) - pos);
329 hdrlen += sizeof(dheader) - pos;
330 } else if (obj_type == OBJ_REF_DELTA) {
332 * Deltas with a base reference contain
333 * an additional 20 bytes for the base sha1.
335 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
336 free(out);
337 free(buf);
338 return 0;
340 sha1write(f, header, hdrlen);
341 sha1write(f, entry->delta->idx.sha1, 20);
342 hdrlen += 20;
343 } else {
344 if (limit && hdrlen + datalen + 20 >= limit) {
345 free(out);
346 free(buf);
347 return 0;
349 sha1write(f, header, hdrlen);
351 sha1write(f, out, datalen);
352 free(out);
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 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
363 OBJ_OFS_DELTA : OBJ_REF_DELTA;
364 reused_delta++;
366 hdrlen = encode_header(obj_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 (obj_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 (obj_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 int do_progress = progress >> pack_to_stdout;
454 uint32_t nr_remaining = nr_result;
456 if (do_progress)
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 || nr_written == nr_remaining) {
496 sha1close(f, sha1, 1);
497 } else {
498 int fd = sha1close(f, NULL, 0);
499 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
500 close(fd);
503 if (!pack_to_stdout) {
504 mode_t mode = umask(0);
505 char *idx_tmp_name, tmpname[PATH_MAX];
507 umask(mode);
508 mode = 0444 & ~mode;
510 idx_tmp_name = write_idx_file(NULL, written_list,
511 nr_written, sha1);
512 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
513 base_name, sha1_to_hex(sha1));
514 if (adjust_perm(pack_tmp_name, mode))
515 die("unable to make temporary pack file readable: %s",
516 strerror(errno));
517 if (rename(pack_tmp_name, tmpname))
518 die("unable to rename temporary pack file: %s",
519 strerror(errno));
520 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
521 base_name, sha1_to_hex(sha1));
522 if (adjust_perm(idx_tmp_name, mode))
523 die("unable to make temporary index file readable: %s",
524 strerror(errno));
525 if (rename(idx_tmp_name, tmpname))
526 die("unable to rename temporary index file: %s",
527 strerror(errno));
528 free(idx_tmp_name);
529 free(pack_tmp_name);
530 puts(sha1_to_hex(sha1));
533 /* mark written objects as written to previous pack */
534 for (j = 0; j < nr_written; j++) {
535 written_list[j]->offset = (off_t)-1;
537 nr_remaining -= nr_written;
538 } while (nr_remaining && i < nr_objects);
540 free(written_list);
541 stop_progress(&progress_state);
542 if (written != nr_result)
543 die("wrote %u objects while expecting %u", written, nr_result);
545 * We have scanned through [0 ... i). Since we have written
546 * the correct number of objects, the remaining [i ... nr_objects)
547 * items must be either already written (due to out-of-order delta base)
548 * or a preferred base. Count those which are neither and complain if any.
550 for (j = 0; i < nr_objects; i++) {
551 struct object_entry *e = objects + i;
552 j += !e->idx.offset && !e->preferred_base;
554 if (j)
555 die("wrote %u objects as expected but %u unwritten", written, j);
558 static int locate_object_entry_hash(const unsigned char *sha1)
560 int i;
561 unsigned int ui;
562 memcpy(&ui, sha1, sizeof(unsigned int));
563 i = ui % object_ix_hashsz;
564 while (0 < object_ix[i]) {
565 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
566 return i;
567 if (++i == object_ix_hashsz)
568 i = 0;
570 return -1 - i;
573 static struct object_entry *locate_object_entry(const unsigned char *sha1)
575 int i;
577 if (!object_ix_hashsz)
578 return NULL;
580 i = locate_object_entry_hash(sha1);
581 if (0 <= i)
582 return &objects[object_ix[i]-1];
583 return NULL;
586 static void rehash_objects(void)
588 uint32_t i;
589 struct object_entry *oe;
591 object_ix_hashsz = nr_objects * 3;
592 if (object_ix_hashsz < 1024)
593 object_ix_hashsz = 1024;
594 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
595 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
596 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
597 int ix = locate_object_entry_hash(oe->idx.sha1);
598 if (0 <= ix)
599 continue;
600 ix = -1 - ix;
601 object_ix[ix] = i + 1;
605 static unsigned name_hash(const char *name)
607 unsigned char c;
608 unsigned hash = 0;
610 if (!name)
611 return 0;
614 * This effectively just creates a sortable number from the
615 * last sixteen non-whitespace characters. Last characters
616 * count "most", so things that end in ".c" sort together.
618 while ((c = *name++) != 0) {
619 if (isspace(c))
620 continue;
621 hash = (hash >> 2) + (c << 24);
623 return hash;
626 static void setup_delta_attr_check(struct git_attr_check *check)
628 static struct git_attr *attr_delta;
630 if (!attr_delta)
631 attr_delta = git_attr("delta", 5);
633 check[0].attr = attr_delta;
636 static int no_try_delta(const char *path)
638 struct git_attr_check check[1];
640 setup_delta_attr_check(check);
641 if (git_checkattr(path, ARRAY_SIZE(check), check))
642 return 0;
643 if (ATTR_FALSE(check->value))
644 return 1;
645 return 0;
648 static int add_object_entry(const unsigned char *sha1, enum object_type type,
649 const char *name, int exclude)
651 struct object_entry *entry;
652 struct packed_git *p, *found_pack = NULL;
653 off_t found_offset = 0;
654 int ix;
655 unsigned hash = name_hash(name);
657 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
658 if (ix >= 0) {
659 if (exclude) {
660 entry = objects + object_ix[ix] - 1;
661 if (!entry->preferred_base)
662 nr_result--;
663 entry->preferred_base = 1;
665 return 0;
668 for (p = packed_git; p; p = p->next) {
669 off_t offset = find_pack_entry_one(sha1, p);
670 if (offset) {
671 if (!found_pack) {
672 found_offset = offset;
673 found_pack = p;
675 if (exclude)
676 break;
677 if (incremental)
678 return 0;
679 if (local && !p->pack_local)
680 return 0;
684 if (nr_objects >= nr_alloc) {
685 nr_alloc = (nr_alloc + 1024) * 3 / 2;
686 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
689 entry = objects + nr_objects++;
690 memset(entry, 0, sizeof(*entry));
691 hashcpy(entry->idx.sha1, sha1);
692 entry->hash = hash;
693 if (type)
694 entry->type = type;
695 if (exclude)
696 entry->preferred_base = 1;
697 else
698 nr_result++;
699 if (found_pack) {
700 entry->in_pack = found_pack;
701 entry->in_pack_offset = found_offset;
704 if (object_ix_hashsz * 3 <= nr_objects * 4)
705 rehash_objects();
706 else
707 object_ix[-1 - ix] = nr_objects;
709 display_progress(progress_state, nr_objects);
711 if (name && no_try_delta(name))
712 entry->no_try_delta = 1;
714 return 1;
717 struct pbase_tree_cache {
718 unsigned char sha1[20];
719 int ref;
720 int temporary;
721 void *tree_data;
722 unsigned long tree_size;
725 static struct pbase_tree_cache *(pbase_tree_cache[256]);
726 static int pbase_tree_cache_ix(const unsigned char *sha1)
728 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
730 static int pbase_tree_cache_ix_incr(int ix)
732 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
735 static struct pbase_tree {
736 struct pbase_tree *next;
737 /* This is a phony "cache" entry; we are not
738 * going to evict it nor find it through _get()
739 * mechanism -- this is for the toplevel node that
740 * would almost always change with any commit.
742 struct pbase_tree_cache pcache;
743 } *pbase_tree;
745 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
747 struct pbase_tree_cache *ent, *nent;
748 void *data;
749 unsigned long size;
750 enum object_type type;
751 int neigh;
752 int my_ix = pbase_tree_cache_ix(sha1);
753 int available_ix = -1;
755 /* pbase-tree-cache acts as a limited hashtable.
756 * your object will be found at your index or within a few
757 * slots after that slot if it is cached.
759 for (neigh = 0; neigh < 8; neigh++) {
760 ent = pbase_tree_cache[my_ix];
761 if (ent && !hashcmp(ent->sha1, sha1)) {
762 ent->ref++;
763 return ent;
765 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
766 ((0 <= available_ix) &&
767 (!ent && pbase_tree_cache[available_ix])))
768 available_ix = my_ix;
769 if (!ent)
770 break;
771 my_ix = pbase_tree_cache_ix_incr(my_ix);
774 /* Did not find one. Either we got a bogus request or
775 * we need to read and perhaps cache.
777 data = read_sha1_file(sha1, &type, &size);
778 if (!data)
779 return NULL;
780 if (type != OBJ_TREE) {
781 free(data);
782 return NULL;
785 /* We need to either cache or return a throwaway copy */
787 if (available_ix < 0)
788 ent = NULL;
789 else {
790 ent = pbase_tree_cache[available_ix];
791 my_ix = available_ix;
794 if (!ent) {
795 nent = xmalloc(sizeof(*nent));
796 nent->temporary = (available_ix < 0);
798 else {
799 /* evict and reuse */
800 free(ent->tree_data);
801 nent = ent;
803 hashcpy(nent->sha1, sha1);
804 nent->tree_data = data;
805 nent->tree_size = size;
806 nent->ref = 1;
807 if (!nent->temporary)
808 pbase_tree_cache[my_ix] = nent;
809 return nent;
812 static void pbase_tree_put(struct pbase_tree_cache *cache)
814 if (!cache->temporary) {
815 cache->ref--;
816 return;
818 free(cache->tree_data);
819 free(cache);
822 static int name_cmp_len(const char *name)
824 int i;
825 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
827 return i;
830 static void add_pbase_object(struct tree_desc *tree,
831 const char *name,
832 int cmplen,
833 const char *fullname)
835 struct name_entry entry;
836 int cmp;
838 while (tree_entry(tree,&entry)) {
839 if (S_ISGITLINK(entry.mode))
840 continue;
841 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
842 memcmp(name, entry.path, cmplen);
843 if (cmp > 0)
844 continue;
845 if (cmp < 0)
846 return;
847 if (name[cmplen] != '/') {
848 add_object_entry(entry.sha1,
849 object_type(entry.mode),
850 fullname, 1);
851 return;
853 if (S_ISDIR(entry.mode)) {
854 struct tree_desc sub;
855 struct pbase_tree_cache *tree;
856 const char *down = name+cmplen+1;
857 int downlen = name_cmp_len(down);
859 tree = pbase_tree_get(entry.sha1);
860 if (!tree)
861 return;
862 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
864 add_pbase_object(&sub, down, downlen, fullname);
865 pbase_tree_put(tree);
870 static unsigned *done_pbase_paths;
871 static int done_pbase_paths_num;
872 static int done_pbase_paths_alloc;
873 static int done_pbase_path_pos(unsigned hash)
875 int lo = 0;
876 int hi = done_pbase_paths_num;
877 while (lo < hi) {
878 int mi = (hi + lo) / 2;
879 if (done_pbase_paths[mi] == hash)
880 return mi;
881 if (done_pbase_paths[mi] < hash)
882 hi = mi;
883 else
884 lo = mi + 1;
886 return -lo-1;
889 static int check_pbase_path(unsigned hash)
891 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
892 if (0 <= pos)
893 return 1;
894 pos = -pos - 1;
895 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
896 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
897 done_pbase_paths = xrealloc(done_pbase_paths,
898 done_pbase_paths_alloc *
899 sizeof(unsigned));
901 done_pbase_paths_num++;
902 if (pos < done_pbase_paths_num)
903 memmove(done_pbase_paths + pos + 1,
904 done_pbase_paths + pos,
905 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
906 done_pbase_paths[pos] = hash;
907 return 0;
910 static void add_preferred_base_object(const char *name)
912 struct pbase_tree *it;
913 int cmplen;
914 unsigned hash = name_hash(name);
916 if (!num_preferred_base || check_pbase_path(hash))
917 return;
919 cmplen = name_cmp_len(name);
920 for (it = pbase_tree; it; it = it->next) {
921 if (cmplen == 0) {
922 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
924 else {
925 struct tree_desc tree;
926 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
927 add_pbase_object(&tree, name, cmplen, name);
932 static void add_preferred_base(unsigned char *sha1)
934 struct pbase_tree *it;
935 void *data;
936 unsigned long size;
937 unsigned char tree_sha1[20];
939 if (window <= num_preferred_base++)
940 return;
942 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
943 if (!data)
944 return;
946 for (it = pbase_tree; it; it = it->next) {
947 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
948 free(data);
949 return;
953 it = xcalloc(1, sizeof(*it));
954 it->next = pbase_tree;
955 pbase_tree = it;
957 hashcpy(it->pcache.sha1, tree_sha1);
958 it->pcache.tree_data = data;
959 it->pcache.tree_size = size;
962 static void check_object(struct object_entry *entry)
964 if (entry->in_pack) {
965 struct packed_git *p = entry->in_pack;
966 struct pack_window *w_curs = NULL;
967 const unsigned char *base_ref = NULL;
968 struct object_entry *base_entry;
969 unsigned long used, used_0;
970 unsigned int avail;
971 off_t ofs;
972 unsigned char *buf, c;
974 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
977 * We want in_pack_type even if we do not reuse delta
978 * since non-delta representations could still be reused.
980 used = unpack_object_header_gently(buf, avail,
981 &entry->in_pack_type,
982 &entry->size);
985 * Determine if this is a delta and if so whether we can
986 * reuse it or not. Otherwise let's find out as cheaply as
987 * possible what the actual type and size for this object is.
989 switch (entry->in_pack_type) {
990 default:
991 /* Not a delta hence we've already got all we need. */
992 entry->type = entry->in_pack_type;
993 entry->in_pack_header_size = used;
994 unuse_pack(&w_curs);
995 return;
996 case OBJ_REF_DELTA:
997 if (!no_reuse_delta && !entry->preferred_base)
998 base_ref = use_pack(p, &w_curs,
999 entry->in_pack_offset + used, NULL);
1000 entry->in_pack_header_size = used + 20;
1001 break;
1002 case OBJ_OFS_DELTA:
1003 buf = use_pack(p, &w_curs,
1004 entry->in_pack_offset + used, NULL);
1005 used_0 = 0;
1006 c = buf[used_0++];
1007 ofs = c & 127;
1008 while (c & 128) {
1009 ofs += 1;
1010 if (!ofs || MSB(ofs, 7))
1011 die("delta base offset overflow in pack for %s",
1012 sha1_to_hex(entry->idx.sha1));
1013 c = buf[used_0++];
1014 ofs = (ofs << 7) + (c & 127);
1016 if (ofs >= entry->in_pack_offset)
1017 die("delta base offset out of bound for %s",
1018 sha1_to_hex(entry->idx.sha1));
1019 ofs = entry->in_pack_offset - ofs;
1020 if (!no_reuse_delta && !entry->preferred_base) {
1021 struct revindex_entry *revidx;
1022 revidx = find_pack_revindex(p, ofs);
1023 base_ref = nth_packed_object_sha1(p, revidx->nr);
1025 entry->in_pack_header_size = used + used_0;
1026 break;
1029 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1031 * If base_ref was set above that means we wish to
1032 * reuse delta data, and we even found that base
1033 * in the list of objects we want to pack. Goodie!
1035 * Depth value does not matter - find_deltas() will
1036 * never consider reused delta as the base object to
1037 * deltify other objects against, in order to avoid
1038 * circular deltas.
1040 entry->type = entry->in_pack_type;
1041 entry->delta = base_entry;
1042 entry->delta_sibling = base_entry->delta_child;
1043 base_entry->delta_child = entry;
1044 unuse_pack(&w_curs);
1045 return;
1048 if (entry->type) {
1050 * This must be a delta and we already know what the
1051 * final object type is. Let's extract the actual
1052 * object size from the delta header.
1054 entry->size = get_size_from_delta(p, &w_curs,
1055 entry->in_pack_offset + entry->in_pack_header_size);
1056 unuse_pack(&w_curs);
1057 return;
1061 * No choice but to fall back to the recursive delta walk
1062 * with sha1_object_info() to find about the object type
1063 * at this point...
1065 unuse_pack(&w_curs);
1068 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1069 if (entry->type < 0)
1070 die("unable to get type of object %s",
1071 sha1_to_hex(entry->idx.sha1));
1074 static int pack_offset_sort(const void *_a, const void *_b)
1076 const struct object_entry *a = *(struct object_entry **)_a;
1077 const struct object_entry *b = *(struct object_entry **)_b;
1079 /* avoid filesystem trashing with loose objects */
1080 if (!a->in_pack && !b->in_pack)
1081 return hashcmp(a->idx.sha1, b->idx.sha1);
1083 if (a->in_pack < b->in_pack)
1084 return -1;
1085 if (a->in_pack > b->in_pack)
1086 return 1;
1087 return a->in_pack_offset < b->in_pack_offset ? -1 :
1088 (a->in_pack_offset > b->in_pack_offset);
1091 static void get_object_details(void)
1093 uint32_t i;
1094 struct object_entry **sorted_by_offset;
1096 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1097 for (i = 0; i < nr_objects; i++)
1098 sorted_by_offset[i] = objects + i;
1099 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1101 init_pack_revindex();
1103 for (i = 0; i < nr_objects; i++)
1104 check_object(sorted_by_offset[i]);
1106 free(sorted_by_offset);
1110 * We search for deltas in a list sorted by type, by filename hash, and then
1111 * by size, so that we see progressively smaller and smaller files.
1112 * That's because we prefer deltas to be from the bigger file
1113 * to the smaller -- deletes are potentially cheaper, but perhaps
1114 * more importantly, the bigger file is likely the more recent
1115 * one. The deepest deltas are therefore the oldest objects which are
1116 * less susceptible to be accessed often.
1118 static int type_size_sort(const void *_a, const void *_b)
1120 const struct object_entry *a = *(struct object_entry **)_a;
1121 const struct object_entry *b = *(struct object_entry **)_b;
1123 if (a->type > b->type)
1124 return -1;
1125 if (a->type < b->type)
1126 return 1;
1127 if (a->hash > b->hash)
1128 return -1;
1129 if (a->hash < b->hash)
1130 return 1;
1131 if (a->preferred_base > b->preferred_base)
1132 return -1;
1133 if (a->preferred_base < b->preferred_base)
1134 return 1;
1135 if (a->size > b->size)
1136 return -1;
1137 if (a->size < b->size)
1138 return 1;
1139 return a < b ? -1 : (a > b); /* newest first */
1142 struct unpacked {
1143 struct object_entry *entry;
1144 void *data;
1145 struct delta_index *index;
1146 unsigned depth;
1149 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1150 unsigned long delta_size)
1152 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1153 return 0;
1155 if (delta_size < cache_max_small_delta_size)
1156 return 1;
1158 /* cache delta, if objects are large enough compared to delta size */
1159 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1160 return 1;
1162 return 0;
1165 #ifdef THREADED_DELTA_SEARCH
1167 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1168 #define read_lock() pthread_mutex_lock(&read_mutex)
1169 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1171 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1172 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1173 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1175 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1176 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1177 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1179 #else
1181 #define read_lock() (void)0
1182 #define read_unlock() (void)0
1183 #define cache_lock() (void)0
1184 #define cache_unlock() (void)0
1185 #define progress_lock() (void)0
1186 #define progress_unlock() (void)0
1188 #endif
1190 static int try_delta(struct unpacked *trg, struct unpacked *src,
1191 unsigned max_depth, unsigned long *mem_usage)
1193 struct object_entry *trg_entry = trg->entry;
1194 struct object_entry *src_entry = src->entry;
1195 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1196 unsigned ref_depth;
1197 enum object_type type;
1198 void *delta_buf;
1200 /* Don't bother doing diffs between different types */
1201 if (trg_entry->type != src_entry->type)
1202 return -1;
1205 * We do not bother to try a delta that we discarded
1206 * on an earlier try, but only when reusing delta data.
1208 if (!no_reuse_delta && trg_entry->in_pack &&
1209 trg_entry->in_pack == src_entry->in_pack &&
1210 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1211 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1212 return 0;
1214 /* Let's not bust the allowed depth. */
1215 if (src->depth >= max_depth)
1216 return 0;
1218 /* Now some size filtering heuristics. */
1219 trg_size = trg_entry->size;
1220 if (!trg_entry->delta) {
1221 max_size = trg_size/2 - 20;
1222 ref_depth = 1;
1223 } else {
1224 max_size = trg_entry->delta_size;
1225 ref_depth = trg->depth;
1227 max_size = max_size * (max_depth - src->depth) /
1228 (max_depth - ref_depth + 1);
1229 if (max_size == 0)
1230 return 0;
1231 src_size = src_entry->size;
1232 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1233 if (sizediff >= max_size)
1234 return 0;
1235 if (trg_size < src_size / 32)
1236 return 0;
1238 /* Load data if not already done */
1239 if (!trg->data) {
1240 read_lock();
1241 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1242 read_unlock();
1243 if (!trg->data)
1244 die("object %s cannot be read",
1245 sha1_to_hex(trg_entry->idx.sha1));
1246 if (sz != trg_size)
1247 die("object %s inconsistent object length (%lu vs %lu)",
1248 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1249 *mem_usage += sz;
1251 if (!src->data) {
1252 read_lock();
1253 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1254 read_unlock();
1255 if (!src->data)
1256 die("object %s cannot be read",
1257 sha1_to_hex(src_entry->idx.sha1));
1258 if (sz != src_size)
1259 die("object %s inconsistent object length (%lu vs %lu)",
1260 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1261 *mem_usage += sz;
1263 if (!src->index) {
1264 src->index = create_delta_index(src->data, src_size);
1265 if (!src->index) {
1266 static int warned = 0;
1267 if (!warned++)
1268 warning("suboptimal pack - out of memory");
1269 return 0;
1271 *mem_usage += sizeof_delta_index(src->index);
1274 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1275 if (!delta_buf)
1276 return 0;
1278 if (trg_entry->delta) {
1279 /* Prefer only shallower same-sized deltas. */
1280 if (delta_size == trg_entry->delta_size &&
1281 src->depth + 1 >= trg->depth) {
1282 free(delta_buf);
1283 return 0;
1288 * Handle memory allocation outside of the cache
1289 * accounting lock. Compiler will optimize the strangeness
1290 * away when THREADED_DELTA_SEARCH is not defined.
1292 free(trg_entry->delta_data);
1293 cache_lock();
1294 if (trg_entry->delta_data) {
1295 delta_cache_size -= trg_entry->delta_size;
1296 trg_entry->delta_data = NULL;
1298 if (delta_cacheable(src_size, trg_size, delta_size)) {
1299 delta_cache_size += delta_size;
1300 cache_unlock();
1301 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1302 } else {
1303 cache_unlock();
1304 free(delta_buf);
1307 trg_entry->delta = src_entry;
1308 trg_entry->delta_size = delta_size;
1309 trg->depth = src->depth + 1;
1311 return 1;
1314 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1316 struct object_entry *child = me->delta_child;
1317 unsigned int m = n;
1318 while (child) {
1319 unsigned int c = check_delta_limit(child, n + 1);
1320 if (m < c)
1321 m = c;
1322 child = child->delta_sibling;
1324 return m;
1327 static unsigned long free_unpacked(struct unpacked *n)
1329 unsigned long freed_mem = sizeof_delta_index(n->index);
1330 free_delta_index(n->index);
1331 n->index = NULL;
1332 if (n->data) {
1333 freed_mem += n->entry->size;
1334 free(n->data);
1335 n->data = NULL;
1337 n->entry = NULL;
1338 n->depth = 0;
1339 return freed_mem;
1342 static void find_deltas(struct object_entry **list, unsigned *list_size,
1343 int window, int depth, unsigned *processed)
1345 uint32_t i, idx = 0, count = 0;
1346 unsigned int array_size = window * sizeof(struct unpacked);
1347 struct unpacked *array;
1348 unsigned long mem_usage = 0;
1350 array = xmalloc(array_size);
1351 memset(array, 0, array_size);
1353 for (;;) {
1354 struct object_entry *entry = *list++;
1355 struct unpacked *n = array + idx;
1356 int j, max_depth, best_base = -1;
1358 progress_lock();
1359 if (!*list_size) {
1360 progress_unlock();
1361 break;
1363 (*list_size)--;
1364 if (!entry->preferred_base) {
1365 (*processed)++;
1366 display_progress(progress_state, *processed);
1368 progress_unlock();
1370 mem_usage -= free_unpacked(n);
1371 n->entry = entry;
1373 while (window_memory_limit &&
1374 mem_usage > window_memory_limit &&
1375 count > 1) {
1376 uint32_t tail = (idx + window - count) % window;
1377 mem_usage -= free_unpacked(array + tail);
1378 count--;
1381 /* We do not compute delta to *create* objects we are not
1382 * going to pack.
1384 if (entry->preferred_base)
1385 goto next;
1388 * If the current object is at pack edge, take the depth the
1389 * objects that depend on the current object into account
1390 * otherwise they would become too deep.
1392 max_depth = depth;
1393 if (entry->delta_child) {
1394 max_depth -= check_delta_limit(entry, 0);
1395 if (max_depth <= 0)
1396 goto next;
1399 j = window;
1400 while (--j > 0) {
1401 int ret;
1402 uint32_t other_idx = idx + j;
1403 struct unpacked *m;
1404 if (other_idx >= window)
1405 other_idx -= window;
1406 m = array + other_idx;
1407 if (!m->entry)
1408 break;
1409 ret = try_delta(n, m, max_depth, &mem_usage);
1410 if (ret < 0)
1411 break;
1412 else if (ret > 0)
1413 best_base = other_idx;
1416 /* if we made n a delta, and if n is already at max
1417 * depth, leaving it in the window is pointless. we
1418 * should evict it first.
1420 if (entry->delta && depth <= n->depth)
1421 continue;
1424 * Move the best delta base up in the window, after the
1425 * currently deltified object, to keep it longer. It will
1426 * be the first base object to be attempted next.
1428 if (entry->delta) {
1429 struct unpacked swap = array[best_base];
1430 int dist = (window + idx - best_base) % window;
1431 int dst = best_base;
1432 while (dist--) {
1433 int src = (dst + 1) % window;
1434 array[dst] = array[src];
1435 dst = src;
1437 array[dst] = swap;
1440 next:
1441 idx++;
1442 if (count + 1 < window)
1443 count++;
1444 if (idx >= window)
1445 idx = 0;
1448 for (i = 0; i < window; ++i) {
1449 free_delta_index(array[i].index);
1450 free(array[i].data);
1452 free(array);
1455 #ifdef THREADED_DELTA_SEARCH
1458 * The main thread waits on the condition that (at least) one of the workers
1459 * has stopped working (which is indicated in the .working member of
1460 * struct thread_params).
1461 * When a work thread has completed its work, it sets .working to 0 and
1462 * signals the main thread and waits on the condition that .data_ready
1463 * becomes 1.
1466 struct thread_params {
1467 pthread_t thread;
1468 struct object_entry **list;
1469 unsigned list_size;
1470 unsigned remaining;
1471 int window;
1472 int depth;
1473 int working;
1474 int data_ready;
1475 pthread_mutex_t mutex;
1476 pthread_cond_t cond;
1477 unsigned *processed;
1480 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1482 static void *threaded_find_deltas(void *arg)
1484 struct thread_params *me = arg;
1486 while (me->remaining) {
1487 find_deltas(me->list, &me->remaining,
1488 me->window, me->depth, me->processed);
1490 progress_lock();
1491 me->working = 0;
1492 pthread_cond_signal(&progress_cond);
1493 progress_unlock();
1496 * We must not set ->data_ready before we wait on the
1497 * condition because the main thread may have set it to 1
1498 * before we get here. In order to be sure that new
1499 * work is available if we see 1 in ->data_ready, it
1500 * was initialized to 0 before this thread was spawned
1501 * and we reset it to 0 right away.
1503 pthread_mutex_lock(&me->mutex);
1504 while (!me->data_ready)
1505 pthread_cond_wait(&me->cond, &me->mutex);
1506 me->data_ready = 0;
1507 pthread_mutex_unlock(&me->mutex);
1509 /* leave ->working 1 so that this doesn't get more work assigned */
1510 return NULL;
1513 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1514 int window, int depth, unsigned *processed)
1516 struct thread_params p[delta_search_threads];
1517 int i, ret, active_threads = 0;
1519 if (delta_search_threads <= 1) {
1520 find_deltas(list, &list_size, window, depth, processed);
1521 return;
1524 /* Partition the work amongst work threads. */
1525 for (i = 0; i < delta_search_threads; i++) {
1526 unsigned sub_size = list_size / (delta_search_threads - i);
1528 p[i].window = window;
1529 p[i].depth = depth;
1530 p[i].processed = processed;
1531 p[i].working = 1;
1532 p[i].data_ready = 0;
1534 /* try to split chunks on "path" boundaries */
1535 while (sub_size && sub_size < list_size &&
1536 list[sub_size]->hash &&
1537 list[sub_size]->hash == list[sub_size-1]->hash)
1538 sub_size++;
1540 p[i].list = list;
1541 p[i].list_size = sub_size;
1542 p[i].remaining = sub_size;
1544 list += sub_size;
1545 list_size -= sub_size;
1548 /* Start work threads. */
1549 for (i = 0; i < delta_search_threads; i++) {
1550 if (!p[i].list_size)
1551 continue;
1552 pthread_mutex_init(&p[i].mutex, NULL);
1553 pthread_cond_init(&p[i].cond, NULL);
1554 ret = pthread_create(&p[i].thread, NULL,
1555 threaded_find_deltas, &p[i]);
1556 if (ret)
1557 die("unable to create thread: %s", strerror(ret));
1558 active_threads++;
1562 * Now let's wait for work completion. Each time a thread is done
1563 * with its work, we steal half of the remaining work from the
1564 * thread with the largest number of unprocessed objects and give
1565 * it to that newly idle thread. This ensure good load balancing
1566 * until the remaining object list segments are simply too short
1567 * to be worth splitting anymore.
1569 while (active_threads) {
1570 struct thread_params *target = NULL;
1571 struct thread_params *victim = NULL;
1572 unsigned sub_size = 0;
1574 progress_lock();
1575 for (;;) {
1576 for (i = 0; !target && i < delta_search_threads; i++)
1577 if (!p[i].working)
1578 target = &p[i];
1579 if (target)
1580 break;
1581 pthread_cond_wait(&progress_cond, &progress_mutex);
1584 for (i = 0; i < delta_search_threads; i++)
1585 if (p[i].remaining > 2*window &&
1586 (!victim || victim->remaining < p[i].remaining))
1587 victim = &p[i];
1588 if (victim) {
1589 sub_size = victim->remaining / 2;
1590 list = victim->list + victim->list_size - sub_size;
1591 while (sub_size && list[0]->hash &&
1592 list[0]->hash == list[-1]->hash) {
1593 list++;
1594 sub_size--;
1596 if (!sub_size) {
1598 * It is possible for some "paths" to have
1599 * so many objects that no hash boundary
1600 * might be found. Let's just steal the
1601 * exact half in that case.
1603 sub_size = victim->remaining / 2;
1604 list -= sub_size;
1606 target->list = list;
1607 victim->list_size -= sub_size;
1608 victim->remaining -= sub_size;
1610 target->list_size = sub_size;
1611 target->remaining = sub_size;
1612 target->working = 1;
1613 progress_unlock();
1615 pthread_mutex_lock(&target->mutex);
1616 target->data_ready = 1;
1617 pthread_cond_signal(&target->cond);
1618 pthread_mutex_unlock(&target->mutex);
1620 if (!sub_size) {
1621 pthread_join(target->thread, NULL);
1622 pthread_cond_destroy(&target->cond);
1623 pthread_mutex_destroy(&target->mutex);
1624 active_threads--;
1629 #else
1630 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1631 #endif
1633 static void prepare_pack(int window, int depth)
1635 struct object_entry **delta_list;
1636 uint32_t i, n, nr_deltas;
1638 get_object_details();
1640 if (!nr_objects || !window || !depth)
1641 return;
1643 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1644 nr_deltas = n = 0;
1646 for (i = 0; i < nr_objects; i++) {
1647 struct object_entry *entry = objects + i;
1649 if (entry->delta)
1650 /* This happens if we decided to reuse existing
1651 * delta from a pack. "!no_reuse_delta &&" is implied.
1653 continue;
1655 if (entry->size < 50)
1656 continue;
1658 if (entry->no_try_delta)
1659 continue;
1661 if (!entry->preferred_base)
1662 nr_deltas++;
1664 delta_list[n++] = entry;
1667 if (nr_deltas && n > 1) {
1668 unsigned nr_done = 0;
1669 if (progress)
1670 progress_state = start_progress("Compressing objects",
1671 nr_deltas);
1672 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1673 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1674 stop_progress(&progress_state);
1675 if (nr_done != nr_deltas)
1676 die("inconsistency with delta count");
1678 free(delta_list);
1681 static int git_pack_config(const char *k, const char *v)
1683 if(!strcmp(k, "pack.window")) {
1684 window = git_config_int(k, v);
1685 return 0;
1687 if (!strcmp(k, "pack.windowmemory")) {
1688 window_memory_limit = git_config_ulong(k, v);
1689 return 0;
1691 if (!strcmp(k, "pack.depth")) {
1692 depth = git_config_int(k, v);
1693 return 0;
1695 if (!strcmp(k, "pack.compression")) {
1696 int level = git_config_int(k, v);
1697 if (level == -1)
1698 level = Z_DEFAULT_COMPRESSION;
1699 else if (level < 0 || level > Z_BEST_COMPRESSION)
1700 die("bad pack compression level %d", level);
1701 pack_compression_level = level;
1702 pack_compression_seen = 1;
1703 return 0;
1705 if (!strcmp(k, "pack.deltacachesize")) {
1706 max_delta_cache_size = git_config_int(k, v);
1707 return 0;
1709 if (!strcmp(k, "pack.deltacachelimit")) {
1710 cache_max_small_delta_size = git_config_int(k, v);
1711 return 0;
1713 if (!strcmp(k, "pack.threads")) {
1714 delta_search_threads = git_config_int(k, v);
1715 if (delta_search_threads < 0)
1716 die("invalid number of threads specified (%d)",
1717 delta_search_threads);
1718 #ifndef THREADED_DELTA_SEARCH
1719 if (delta_search_threads != 1)
1720 warning("no threads support, ignoring %s", k);
1721 #endif
1722 return 0;
1724 if (!strcmp(k, "pack.indexversion")) {
1725 pack_idx_default_version = git_config_int(k, v);
1726 if (pack_idx_default_version > 2)
1727 die("bad pack.indexversion=%d", pack_idx_default_version);
1728 return 0;
1730 if (!strcmp(k, "pack.packsizelimit")) {
1731 pack_size_limit_cfg = git_config_ulong(k, v);
1732 return 0;
1734 return git_default_config(k, v);
1737 static void read_object_list_from_stdin(void)
1739 char line[40 + 1 + PATH_MAX + 2];
1740 unsigned char sha1[20];
1742 for (;;) {
1743 if (!fgets(line, sizeof(line), stdin)) {
1744 if (feof(stdin))
1745 break;
1746 if (!ferror(stdin))
1747 die("fgets returned NULL, not EOF, not error!");
1748 if (errno != EINTR)
1749 die("fgets: %s", strerror(errno));
1750 clearerr(stdin);
1751 continue;
1753 if (line[0] == '-') {
1754 if (get_sha1_hex(line+1, sha1))
1755 die("expected edge sha1, got garbage:\n %s",
1756 line);
1757 add_preferred_base(sha1);
1758 continue;
1760 if (get_sha1_hex(line, sha1))
1761 die("expected sha1, got garbage:\n %s", line);
1763 add_preferred_base_object(line+41);
1764 add_object_entry(sha1, 0, line+41, 0);
1768 #define OBJECT_ADDED (1u<<20)
1770 static void show_commit(struct commit *commit)
1772 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1773 commit->object.flags |= OBJECT_ADDED;
1776 static void show_object(struct object_array_entry *p)
1778 add_preferred_base_object(p->name);
1779 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1780 p->item->flags |= OBJECT_ADDED;
1783 static void show_edge(struct commit *commit)
1785 add_preferred_base(commit->object.sha1);
1788 struct in_pack_object {
1789 off_t offset;
1790 struct object *object;
1793 struct in_pack {
1794 int alloc;
1795 int nr;
1796 struct in_pack_object *array;
1799 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1801 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1802 in_pack->array[in_pack->nr].object = object;
1803 in_pack->nr++;
1807 * Compare the objects in the offset order, in order to emulate the
1808 * "git-rev-list --objects" output that produced the pack originally.
1810 static int ofscmp(const void *a_, const void *b_)
1812 struct in_pack_object *a = (struct in_pack_object *)a_;
1813 struct in_pack_object *b = (struct in_pack_object *)b_;
1815 if (a->offset < b->offset)
1816 return -1;
1817 else if (a->offset > b->offset)
1818 return 1;
1819 else
1820 return hashcmp(a->object->sha1, b->object->sha1);
1823 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1825 struct packed_git *p;
1826 struct in_pack in_pack;
1827 uint32_t i;
1829 memset(&in_pack, 0, sizeof(in_pack));
1831 for (p = packed_git; p; p = p->next) {
1832 const unsigned char *sha1;
1833 struct object *o;
1835 for (i = 0; i < revs->num_ignore_packed; i++) {
1836 if (matches_pack_name(p, revs->ignore_packed[i]))
1837 break;
1839 if (revs->num_ignore_packed <= i)
1840 continue;
1841 if (open_pack_index(p))
1842 die("cannot open pack index");
1844 ALLOC_GROW(in_pack.array,
1845 in_pack.nr + p->num_objects,
1846 in_pack.alloc);
1848 for (i = 0; i < p->num_objects; i++) {
1849 sha1 = nth_packed_object_sha1(p, i);
1850 o = lookup_unknown_object(sha1);
1851 if (!(o->flags & OBJECT_ADDED))
1852 mark_in_pack_object(o, p, &in_pack);
1853 o->flags |= OBJECT_ADDED;
1857 if (in_pack.nr) {
1858 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1859 ofscmp);
1860 for (i = 0; i < in_pack.nr; i++) {
1861 struct object *o = in_pack.array[i].object;
1862 add_object_entry(o->sha1, o->type, "", 0);
1865 free(in_pack.array);
1868 static void get_object_list(int ac, const char **av)
1870 struct rev_info revs;
1871 char line[1000];
1872 int flags = 0;
1874 init_revisions(&revs, NULL);
1875 save_commit_buffer = 0;
1876 track_object_refs = 0;
1877 setup_revisions(ac, av, &revs, NULL);
1879 while (fgets(line, sizeof(line), stdin) != NULL) {
1880 int len = strlen(line);
1881 if (len && line[len - 1] == '\n')
1882 line[--len] = 0;
1883 if (!len)
1884 break;
1885 if (*line == '-') {
1886 if (!strcmp(line, "--not")) {
1887 flags ^= UNINTERESTING;
1888 continue;
1890 die("not a rev '%s'", line);
1892 if (handle_revision_arg(line, &revs, flags, 1))
1893 die("bad revision '%s'", line);
1896 if (prepare_revision_walk(&revs))
1897 die("revision walk setup failed");
1898 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1899 traverse_commit_list(&revs, show_commit, show_object);
1901 if (keep_unreachable)
1902 add_objects_in_unpacked_packs(&revs);
1905 static int adjust_perm(const char *path, mode_t mode)
1907 if (chmod(path, mode))
1908 return -1;
1909 return adjust_shared_perm(path);
1912 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1914 int use_internal_rev_list = 0;
1915 int thin = 0;
1916 uint32_t i;
1917 const char **rp_av;
1918 int rp_ac_alloc = 64;
1919 int rp_ac;
1921 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
1923 rp_av[0] = "pack-objects";
1924 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
1925 rp_ac = 2;
1927 git_config(git_pack_config);
1928 if (!pack_compression_seen && core_compression_seen)
1929 pack_compression_level = core_compression_level;
1931 progress = isatty(2);
1932 for (i = 1; i < argc; i++) {
1933 const char *arg = argv[i];
1935 if (*arg != '-')
1936 break;
1938 if (!strcmp("--non-empty", arg)) {
1939 non_empty = 1;
1940 continue;
1942 if (!strcmp("--local", arg)) {
1943 local = 1;
1944 continue;
1946 if (!strcmp("--incremental", arg)) {
1947 incremental = 1;
1948 continue;
1950 if (!prefixcmp(arg, "--compression=")) {
1951 char *end;
1952 int level = strtoul(arg+14, &end, 0);
1953 if (!arg[14] || *end)
1954 usage(pack_usage);
1955 if (level == -1)
1956 level = Z_DEFAULT_COMPRESSION;
1957 else if (level < 0 || level > Z_BEST_COMPRESSION)
1958 die("bad pack compression level %d", level);
1959 pack_compression_level = level;
1960 continue;
1962 if (!prefixcmp(arg, "--max-pack-size=")) {
1963 char *end;
1964 pack_size_limit_cfg = 0;
1965 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
1966 if (!arg[16] || *end)
1967 usage(pack_usage);
1968 continue;
1970 if (!prefixcmp(arg, "--window=")) {
1971 char *end;
1972 window = strtoul(arg+9, &end, 0);
1973 if (!arg[9] || *end)
1974 usage(pack_usage);
1975 continue;
1977 if (!prefixcmp(arg, "--window-memory=")) {
1978 if (!git_parse_ulong(arg+16, &window_memory_limit))
1979 usage(pack_usage);
1980 continue;
1982 if (!prefixcmp(arg, "--threads=")) {
1983 char *end;
1984 delta_search_threads = strtoul(arg+10, &end, 0);
1985 if (!arg[10] || *end || delta_search_threads < 0)
1986 usage(pack_usage);
1987 #ifndef THREADED_DELTA_SEARCH
1988 if (delta_search_threads != 1)
1989 warning("no threads support, "
1990 "ignoring %s", arg);
1991 #endif
1992 continue;
1994 if (!prefixcmp(arg, "--depth=")) {
1995 char *end;
1996 depth = strtoul(arg+8, &end, 0);
1997 if (!arg[8] || *end)
1998 usage(pack_usage);
1999 continue;
2001 if (!strcmp("--progress", arg)) {
2002 progress = 1;
2003 continue;
2005 if (!strcmp("--all-progress", arg)) {
2006 progress = 2;
2007 continue;
2009 if (!strcmp("-q", arg)) {
2010 progress = 0;
2011 continue;
2013 if (!strcmp("--no-reuse-delta", arg)) {
2014 no_reuse_delta = 1;
2015 continue;
2017 if (!strcmp("--no-reuse-object", arg)) {
2018 no_reuse_object = no_reuse_delta = 1;
2019 continue;
2021 if (!strcmp("--delta-base-offset", arg)) {
2022 allow_ofs_delta = 1;
2023 continue;
2025 if (!strcmp("--stdout", arg)) {
2026 pack_to_stdout = 1;
2027 continue;
2029 if (!strcmp("--revs", arg)) {
2030 use_internal_rev_list = 1;
2031 continue;
2033 if (!strcmp("--keep-unreachable", arg)) {
2034 keep_unreachable = 1;
2035 continue;
2037 if (!strcmp("--unpacked", arg) ||
2038 !prefixcmp(arg, "--unpacked=") ||
2039 !strcmp("--reflog", arg) ||
2040 !strcmp("--all", arg)) {
2041 use_internal_rev_list = 1;
2042 if (rp_ac >= rp_ac_alloc - 1) {
2043 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2044 rp_av = xrealloc(rp_av,
2045 rp_ac_alloc * sizeof(*rp_av));
2047 rp_av[rp_ac++] = arg;
2048 continue;
2050 if (!strcmp("--thin", arg)) {
2051 use_internal_rev_list = 1;
2052 thin = 1;
2053 rp_av[1] = "--objects-edge";
2054 continue;
2056 if (!prefixcmp(arg, "--index-version=")) {
2057 char *c;
2058 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2059 if (pack_idx_default_version > 2)
2060 die("bad %s", arg);
2061 if (*c == ',')
2062 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2063 if (*c || pack_idx_off32_limit & 0x80000000)
2064 die("bad %s", arg);
2065 continue;
2067 usage(pack_usage);
2070 /* Traditionally "pack-objects [options] base extra" failed;
2071 * we would however want to take refs parameter that would
2072 * have been given to upstream rev-list ourselves, which means
2073 * we somehow want to say what the base name is. So the
2074 * syntax would be:
2076 * pack-objects [options] base <refs...>
2078 * in other words, we would treat the first non-option as the
2079 * base_name and send everything else to the internal revision
2080 * walker.
2083 if (!pack_to_stdout)
2084 base_name = argv[i++];
2086 if (pack_to_stdout != !base_name)
2087 usage(pack_usage);
2089 if (!pack_to_stdout && !pack_size_limit)
2090 pack_size_limit = pack_size_limit_cfg;
2092 if (pack_to_stdout && pack_size_limit)
2093 die("--max-pack-size cannot be used to build a pack for transfer.");
2095 if (!pack_to_stdout && thin)
2096 die("--thin cannot be used to build an indexable pack.");
2098 #ifdef THREADED_DELTA_SEARCH
2099 if (!delta_search_threads) /* --threads=0 means autodetect */
2100 delta_search_threads = online_cpus();
2101 #endif
2103 prepare_packed_git();
2105 if (progress)
2106 progress_state = start_progress("Counting objects", 0);
2107 if (!use_internal_rev_list)
2108 read_object_list_from_stdin();
2109 else {
2110 rp_av[rp_ac] = NULL;
2111 get_object_list(rp_ac, rp_av);
2113 stop_progress(&progress_state);
2115 if (non_empty && !nr_result)
2116 return 0;
2117 if (nr_result)
2118 prepare_pack(window, depth);
2119 write_pack_file();
2120 if (progress)
2121 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2122 written, written_delta, reused, reused_delta);
2123 return 0;