Merge commit 'refs/top-bases/girocco' into girocco
[git/repo.git] / builtin-pack-objects.c
blob59c30d1caa37416041177ff4aaf01b67f4e8add4
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 int 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 1;
423 /* if we are deltified, write out base object first. */
424 if (e->delta && !write_one(f, e->delta, offset))
425 return 0;
427 e->idx.offset = *offset;
428 size = write_object(f, e, *offset);
429 if (!size) {
430 e->idx.offset = 0;
431 return 0;
433 written_list[nr_written++] = &e->idx;
435 /* make sure off_t is sufficiently large not to wrap */
436 if (*offset > *offset + size)
437 die("pack too large for current definition of off_t");
438 *offset += size;
439 return 1;
442 /* forward declaration for write_pack_file */
443 static int adjust_perm(const char *path, mode_t mode);
445 static void write_pack_file(void)
447 uint32_t i = 0, j;
448 struct sha1file *f;
449 off_t offset;
450 struct pack_header hdr;
451 uint32_t nr_remaining = nr_result;
452 time_t last_mtime = 0;
454 if (progress > pack_to_stdout)
455 progress_state = start_progress("Writing objects", nr_result);
456 written_list = xmalloc(nr_objects * sizeof(*written_list));
458 do {
459 unsigned char sha1[20];
460 char *pack_tmp_name = NULL;
462 if (pack_to_stdout) {
463 f = sha1fd_throughput(1, "<stdout>", progress_state);
464 } else {
465 char tmpname[PATH_MAX];
466 int fd;
467 snprintf(tmpname, sizeof(tmpname),
468 "%s/pack/tmp_pack_XXXXXX", get_object_directory());
469 fd = xmkstemp(tmpname);
470 pack_tmp_name = xstrdup(tmpname);
471 f = sha1fd(fd, pack_tmp_name);
474 hdr.hdr_signature = htonl(PACK_SIGNATURE);
475 hdr.hdr_version = htonl(PACK_VERSION);
476 hdr.hdr_entries = htonl(nr_remaining);
477 sha1write(f, &hdr, sizeof(hdr));
478 offset = sizeof(hdr);
479 nr_written = 0;
480 for (; i < nr_objects; i++) {
481 if (!write_one(f, objects + i, &offset))
482 break;
483 display_progress(progress_state, written);
487 * Did we write the wrong # entries in the header?
488 * If so, rewrite it like in fast-import
490 if (pack_to_stdout) {
491 sha1close(f, sha1, CSUM_CLOSE);
492 } else if (nr_written == nr_remaining) {
493 sha1close(f, sha1, CSUM_FSYNC);
494 } else {
495 int fd = sha1close(f, sha1, 0);
496 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
497 nr_written, sha1, offset);
498 close(fd);
501 if (!pack_to_stdout) {
502 mode_t mode = umask(0);
503 struct stat st;
504 char *idx_tmp_name, tmpname[PATH_MAX];
506 umask(mode);
507 mode = 0444 & ~mode;
509 idx_tmp_name = write_idx_file(NULL, written_list,
510 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));
522 * Packs are runtime accessed in their mtime
523 * order since newer packs are more likely to contain
524 * younger objects. So if we are creating multiple
525 * packs then we should modify the mtime of later ones
526 * to preserve this property.
528 if (stat(tmpname, &st) < 0) {
529 warning("failed to stat %s: %s",
530 tmpname, strerror(errno));
531 } else if (!last_mtime) {
532 last_mtime = st.st_mtime;
533 } else {
534 struct utimbuf utb;
535 utb.actime = st.st_atime;
536 utb.modtime = --last_mtime;
537 if (utime(tmpname, &utb) < 0)
538 warning("failed utime() on %s: %s",
539 tmpname, strerror(errno));
542 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
543 base_name, sha1_to_hex(sha1));
544 if (adjust_perm(idx_tmp_name, mode))
545 die("unable to make temporary index file readable: %s",
546 strerror(errno));
547 if (rename(idx_tmp_name, tmpname))
548 die("unable to rename temporary index file: %s",
549 strerror(errno));
551 free(idx_tmp_name);
552 free(pack_tmp_name);
553 puts(sha1_to_hex(sha1));
556 /* mark written objects as written to previous pack */
557 for (j = 0; j < nr_written; j++) {
558 written_list[j]->offset = (off_t)-1;
560 nr_remaining -= nr_written;
561 } while (nr_remaining && i < nr_objects);
563 free(written_list);
564 stop_progress(&progress_state);
565 if (written != nr_result)
566 die("wrote %"PRIu32" objects while expecting %"PRIu32,
567 written, nr_result);
569 * We have scanned through [0 ... i). Since we have written
570 * the correct number of objects, the remaining [i ... nr_objects)
571 * items must be either already written (due to out-of-order delta base)
572 * or a preferred base. Count those which are neither and complain if any.
574 for (j = 0; i < nr_objects; i++) {
575 struct object_entry *e = objects + i;
576 j += !e->idx.offset && !e->preferred_base;
578 if (j)
579 die("wrote %"PRIu32" objects as expected but %"PRIu32
580 " unwritten", written, j);
583 static int locate_object_entry_hash(const unsigned char *sha1)
585 int i;
586 unsigned int ui;
587 memcpy(&ui, sha1, sizeof(unsigned int));
588 i = ui % object_ix_hashsz;
589 while (0 < object_ix[i]) {
590 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
591 return i;
592 if (++i == object_ix_hashsz)
593 i = 0;
595 return -1 - i;
598 static struct object_entry *locate_object_entry(const unsigned char *sha1)
600 int i;
602 if (!object_ix_hashsz)
603 return NULL;
605 i = locate_object_entry_hash(sha1);
606 if (0 <= i)
607 return &objects[object_ix[i]-1];
608 return NULL;
611 static void rehash_objects(void)
613 uint32_t i;
614 struct object_entry *oe;
616 object_ix_hashsz = nr_objects * 3;
617 if (object_ix_hashsz < 1024)
618 object_ix_hashsz = 1024;
619 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
620 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
621 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
622 int ix = locate_object_entry_hash(oe->idx.sha1);
623 if (0 <= ix)
624 continue;
625 ix = -1 - ix;
626 object_ix[ix] = i + 1;
630 static unsigned name_hash(const char *name)
632 unsigned char c;
633 unsigned hash = 0;
635 if (!name)
636 return 0;
639 * This effectively just creates a sortable number from the
640 * last sixteen non-whitespace characters. Last characters
641 * count "most", so things that end in ".c" sort together.
643 while ((c = *name++) != 0) {
644 if (isspace(c))
645 continue;
646 hash = (hash >> 2) + (c << 24);
648 return hash;
651 static void setup_delta_attr_check(struct git_attr_check *check)
653 static struct git_attr *attr_delta;
655 if (!attr_delta)
656 attr_delta = git_attr("delta", 5);
658 check[0].attr = attr_delta;
661 static int no_try_delta(const char *path)
663 struct git_attr_check check[1];
665 setup_delta_attr_check(check);
666 if (git_checkattr(path, ARRAY_SIZE(check), check))
667 return 0;
668 if (ATTR_FALSE(check->value))
669 return 1;
670 return 0;
673 static int add_object_entry(const unsigned char *sha1, enum object_type type,
674 const char *name, int exclude)
676 struct object_entry *entry;
677 struct packed_git *p, *found_pack = NULL;
678 off_t found_offset = 0;
679 int ix;
680 unsigned hash = name_hash(name);
682 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
683 if (ix >= 0) {
684 if (exclude) {
685 entry = objects + object_ix[ix] - 1;
686 if (!entry->preferred_base)
687 nr_result--;
688 entry->preferred_base = 1;
690 return 0;
693 for (p = packed_git; p; p = p->next) {
694 off_t offset = find_pack_entry_one(sha1, p);
695 if (offset) {
696 if (!found_pack) {
697 found_offset = offset;
698 found_pack = p;
700 if (exclude)
701 break;
702 if (incremental)
703 return 0;
704 if (local && !p->pack_local)
705 return 0;
709 if (nr_objects >= nr_alloc) {
710 nr_alloc = (nr_alloc + 1024) * 3 / 2;
711 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
714 entry = objects + nr_objects++;
715 memset(entry, 0, sizeof(*entry));
716 hashcpy(entry->idx.sha1, sha1);
717 entry->hash = hash;
718 if (type)
719 entry->type = type;
720 if (exclude)
721 entry->preferred_base = 1;
722 else
723 nr_result++;
724 if (found_pack) {
725 entry->in_pack = found_pack;
726 entry->in_pack_offset = found_offset;
729 if (object_ix_hashsz * 3 <= nr_objects * 4)
730 rehash_objects();
731 else
732 object_ix[-1 - ix] = nr_objects;
734 display_progress(progress_state, nr_objects);
736 if (name && no_try_delta(name))
737 entry->no_try_delta = 1;
739 return 1;
742 struct pbase_tree_cache {
743 unsigned char sha1[20];
744 int ref;
745 int temporary;
746 void *tree_data;
747 unsigned long tree_size;
750 static struct pbase_tree_cache *(pbase_tree_cache[256]);
751 static int pbase_tree_cache_ix(const unsigned char *sha1)
753 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
755 static int pbase_tree_cache_ix_incr(int ix)
757 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
760 static struct pbase_tree {
761 struct pbase_tree *next;
762 /* This is a phony "cache" entry; we are not
763 * going to evict it nor find it through _get()
764 * mechanism -- this is for the toplevel node that
765 * would almost always change with any commit.
767 struct pbase_tree_cache pcache;
768 } *pbase_tree;
770 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
772 struct pbase_tree_cache *ent, *nent;
773 void *data;
774 unsigned long size;
775 enum object_type type;
776 int neigh;
777 int my_ix = pbase_tree_cache_ix(sha1);
778 int available_ix = -1;
780 /* pbase-tree-cache acts as a limited hashtable.
781 * your object will be found at your index or within a few
782 * slots after that slot if it is cached.
784 for (neigh = 0; neigh < 8; neigh++) {
785 ent = pbase_tree_cache[my_ix];
786 if (ent && !hashcmp(ent->sha1, sha1)) {
787 ent->ref++;
788 return ent;
790 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
791 ((0 <= available_ix) &&
792 (!ent && pbase_tree_cache[available_ix])))
793 available_ix = my_ix;
794 if (!ent)
795 break;
796 my_ix = pbase_tree_cache_ix_incr(my_ix);
799 /* Did not find one. Either we got a bogus request or
800 * we need to read and perhaps cache.
802 data = read_sha1_file(sha1, &type, &size);
803 if (!data)
804 return NULL;
805 if (type != OBJ_TREE) {
806 free(data);
807 return NULL;
810 /* We need to either cache or return a throwaway copy */
812 if (available_ix < 0)
813 ent = NULL;
814 else {
815 ent = pbase_tree_cache[available_ix];
816 my_ix = available_ix;
819 if (!ent) {
820 nent = xmalloc(sizeof(*nent));
821 nent->temporary = (available_ix < 0);
823 else {
824 /* evict and reuse */
825 free(ent->tree_data);
826 nent = ent;
828 hashcpy(nent->sha1, sha1);
829 nent->tree_data = data;
830 nent->tree_size = size;
831 nent->ref = 1;
832 if (!nent->temporary)
833 pbase_tree_cache[my_ix] = nent;
834 return nent;
837 static void pbase_tree_put(struct pbase_tree_cache *cache)
839 if (!cache->temporary) {
840 cache->ref--;
841 return;
843 free(cache->tree_data);
844 free(cache);
847 static int name_cmp_len(const char *name)
849 int i;
850 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
852 return i;
855 static void add_pbase_object(struct tree_desc *tree,
856 const char *name,
857 int cmplen,
858 const char *fullname)
860 struct name_entry entry;
861 int cmp;
863 while (tree_entry(tree,&entry)) {
864 if (S_ISGITLINK(entry.mode))
865 continue;
866 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
867 memcmp(name, entry.path, cmplen);
868 if (cmp > 0)
869 continue;
870 if (cmp < 0)
871 return;
872 if (name[cmplen] != '/') {
873 add_object_entry(entry.sha1,
874 object_type(entry.mode),
875 fullname, 1);
876 return;
878 if (S_ISDIR(entry.mode)) {
879 struct tree_desc sub;
880 struct pbase_tree_cache *tree;
881 const char *down = name+cmplen+1;
882 int downlen = name_cmp_len(down);
884 tree = pbase_tree_get(entry.sha1);
885 if (!tree)
886 return;
887 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
889 add_pbase_object(&sub, down, downlen, fullname);
890 pbase_tree_put(tree);
895 static unsigned *done_pbase_paths;
896 static int done_pbase_paths_num;
897 static int done_pbase_paths_alloc;
898 static int done_pbase_path_pos(unsigned hash)
900 int lo = 0;
901 int hi = done_pbase_paths_num;
902 while (lo < hi) {
903 int mi = (hi + lo) / 2;
904 if (done_pbase_paths[mi] == hash)
905 return mi;
906 if (done_pbase_paths[mi] < hash)
907 hi = mi;
908 else
909 lo = mi + 1;
911 return -lo-1;
914 static int check_pbase_path(unsigned hash)
916 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
917 if (0 <= pos)
918 return 1;
919 pos = -pos - 1;
920 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
921 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
922 done_pbase_paths = xrealloc(done_pbase_paths,
923 done_pbase_paths_alloc *
924 sizeof(unsigned));
926 done_pbase_paths_num++;
927 if (pos < done_pbase_paths_num)
928 memmove(done_pbase_paths + pos + 1,
929 done_pbase_paths + pos,
930 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
931 done_pbase_paths[pos] = hash;
932 return 0;
935 static void add_preferred_base_object(const char *name)
937 struct pbase_tree *it;
938 int cmplen;
939 unsigned hash = name_hash(name);
941 if (!num_preferred_base || check_pbase_path(hash))
942 return;
944 cmplen = name_cmp_len(name);
945 for (it = pbase_tree; it; it = it->next) {
946 if (cmplen == 0) {
947 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
949 else {
950 struct tree_desc tree;
951 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
952 add_pbase_object(&tree, name, cmplen, name);
957 static void add_preferred_base(unsigned char *sha1)
959 struct pbase_tree *it;
960 void *data;
961 unsigned long size;
962 unsigned char tree_sha1[20];
964 if (window <= num_preferred_base++)
965 return;
967 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
968 if (!data)
969 return;
971 for (it = pbase_tree; it; it = it->next) {
972 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
973 free(data);
974 return;
978 it = xcalloc(1, sizeof(*it));
979 it->next = pbase_tree;
980 pbase_tree = it;
982 hashcpy(it->pcache.sha1, tree_sha1);
983 it->pcache.tree_data = data;
984 it->pcache.tree_size = size;
987 static void check_object(struct object_entry *entry)
989 if (entry->in_pack) {
990 struct packed_git *p = entry->in_pack;
991 struct pack_window *w_curs = NULL;
992 const unsigned char *base_ref = NULL;
993 struct object_entry *base_entry;
994 unsigned long used, used_0;
995 unsigned int avail;
996 off_t ofs;
997 unsigned char *buf, c;
999 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1002 * We want in_pack_type even if we do not reuse delta
1003 * since non-delta representations could still be reused.
1005 used = unpack_object_header_gently(buf, avail,
1006 &entry->in_pack_type,
1007 &entry->size);
1010 * Determine if this is a delta and if so whether we can
1011 * reuse it or not. Otherwise let's find out as cheaply as
1012 * possible what the actual type and size for this object is.
1014 switch (entry->in_pack_type) {
1015 default:
1016 /* Not a delta hence we've already got all we need. */
1017 entry->type = entry->in_pack_type;
1018 entry->in_pack_header_size = used;
1019 unuse_pack(&w_curs);
1020 return;
1021 case OBJ_REF_DELTA:
1022 if (reuse_delta && !entry->preferred_base)
1023 base_ref = use_pack(p, &w_curs,
1024 entry->in_pack_offset + used, NULL);
1025 entry->in_pack_header_size = used + 20;
1026 break;
1027 case OBJ_OFS_DELTA:
1028 buf = use_pack(p, &w_curs,
1029 entry->in_pack_offset + used, NULL);
1030 used_0 = 0;
1031 c = buf[used_0++];
1032 ofs = c & 127;
1033 while (c & 128) {
1034 ofs += 1;
1035 if (!ofs || MSB(ofs, 7))
1036 die("delta base offset overflow in pack for %s",
1037 sha1_to_hex(entry->idx.sha1));
1038 c = buf[used_0++];
1039 ofs = (ofs << 7) + (c & 127);
1041 if (ofs >= entry->in_pack_offset)
1042 die("delta base offset out of bound for %s",
1043 sha1_to_hex(entry->idx.sha1));
1044 ofs = entry->in_pack_offset - ofs;
1045 if (reuse_delta && !entry->preferred_base) {
1046 struct revindex_entry *revidx;
1047 revidx = find_pack_revindex(p, ofs);
1048 base_ref = nth_packed_object_sha1(p, revidx->nr);
1050 entry->in_pack_header_size = used + used_0;
1051 break;
1054 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1056 * If base_ref was set above that means we wish to
1057 * reuse delta data, and we even found that base
1058 * in the list of objects we want to pack. Goodie!
1060 * Depth value does not matter - find_deltas() will
1061 * never consider reused delta as the base object to
1062 * deltify other objects against, in order to avoid
1063 * circular deltas.
1065 entry->type = entry->in_pack_type;
1066 entry->delta = base_entry;
1067 entry->delta_sibling = base_entry->delta_child;
1068 base_entry->delta_child = entry;
1069 unuse_pack(&w_curs);
1070 return;
1073 if (entry->type) {
1075 * This must be a delta and we already know what the
1076 * final object type is. Let's extract the actual
1077 * object size from the delta header.
1079 entry->size = get_size_from_delta(p, &w_curs,
1080 entry->in_pack_offset + entry->in_pack_header_size);
1081 unuse_pack(&w_curs);
1082 return;
1086 * No choice but to fall back to the recursive delta walk
1087 * with sha1_object_info() to find about the object type
1088 * at this point...
1090 unuse_pack(&w_curs);
1093 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1095 * The error condition is checked in prepare_pack(). This is
1096 * to permit a missing preferred base object to be ignored
1097 * as a preferred base. Doing so can result in a larger
1098 * pack file, but the transfer will still take place.
1102 static int pack_offset_sort(const void *_a, const void *_b)
1104 const struct object_entry *a = *(struct object_entry **)_a;
1105 const struct object_entry *b = *(struct object_entry **)_b;
1107 /* avoid filesystem trashing with loose objects */
1108 if (!a->in_pack && !b->in_pack)
1109 return hashcmp(a->idx.sha1, b->idx.sha1);
1111 if (a->in_pack < b->in_pack)
1112 return -1;
1113 if (a->in_pack > b->in_pack)
1114 return 1;
1115 return a->in_pack_offset < b->in_pack_offset ? -1 :
1116 (a->in_pack_offset > b->in_pack_offset);
1119 static void get_object_details(void)
1121 uint32_t i;
1122 struct object_entry **sorted_by_offset;
1124 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1125 for (i = 0; i < nr_objects; i++)
1126 sorted_by_offset[i] = objects + i;
1127 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1129 for (i = 0; i < nr_objects; i++)
1130 check_object(sorted_by_offset[i]);
1132 free(sorted_by_offset);
1136 * We search for deltas in a list sorted by type, by filename hash, and then
1137 * by size, so that we see progressively smaller and smaller files.
1138 * That's because we prefer deltas to be from the bigger file
1139 * to the smaller -- deletes are potentially cheaper, but perhaps
1140 * more importantly, the bigger file is likely the more recent
1141 * one. The deepest deltas are therefore the oldest objects which are
1142 * less susceptible to be accessed often.
1144 static int type_size_sort(const void *_a, const void *_b)
1146 const struct object_entry *a = *(struct object_entry **)_a;
1147 const struct object_entry *b = *(struct object_entry **)_b;
1149 if (a->type > b->type)
1150 return -1;
1151 if (a->type < b->type)
1152 return 1;
1153 if (a->hash > b->hash)
1154 return -1;
1155 if (a->hash < b->hash)
1156 return 1;
1157 if (a->preferred_base > b->preferred_base)
1158 return -1;
1159 if (a->preferred_base < b->preferred_base)
1160 return 1;
1161 if (a->size > b->size)
1162 return -1;
1163 if (a->size < b->size)
1164 return 1;
1165 return a < b ? -1 : (a > b); /* newest first */
1168 struct unpacked {
1169 struct object_entry *entry;
1170 void *data;
1171 struct delta_index *index;
1172 unsigned depth;
1175 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1176 unsigned long delta_size)
1178 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1179 return 0;
1181 if (delta_size < cache_max_small_delta_size)
1182 return 1;
1184 /* cache delta, if objects are large enough compared to delta size */
1185 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1186 return 1;
1188 return 0;
1191 #ifdef THREADED_DELTA_SEARCH
1193 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1194 #define read_lock() pthread_mutex_lock(&read_mutex)
1195 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1197 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1198 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1199 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1201 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1202 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1203 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1205 #else
1207 #define read_lock() (void)0
1208 #define read_unlock() (void)0
1209 #define cache_lock() (void)0
1210 #define cache_unlock() (void)0
1211 #define progress_lock() (void)0
1212 #define progress_unlock() (void)0
1214 #endif
1216 static int try_delta(struct unpacked *trg, struct unpacked *src,
1217 unsigned max_depth, unsigned long *mem_usage)
1219 struct object_entry *trg_entry = trg->entry;
1220 struct object_entry *src_entry = src->entry;
1221 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1222 unsigned ref_depth;
1223 enum object_type type;
1224 void *delta_buf;
1226 /* Don't bother doing diffs between different types */
1227 if (trg_entry->type != src_entry->type)
1228 return -1;
1231 * We do not bother to try a delta that we discarded
1232 * on an earlier try, but only when reusing delta data.
1234 if (reuse_delta && trg_entry->in_pack &&
1235 trg_entry->in_pack == src_entry->in_pack &&
1236 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1237 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1238 return 0;
1240 /* Let's not bust the allowed depth. */
1241 if (src->depth >= max_depth)
1242 return 0;
1244 /* Now some size filtering heuristics. */
1245 trg_size = trg_entry->size;
1246 if (!trg_entry->delta) {
1247 max_size = trg_size/2 - 20;
1248 ref_depth = 1;
1249 } else {
1250 max_size = trg_entry->delta_size;
1251 ref_depth = trg->depth;
1253 max_size = max_size * (max_depth - src->depth) /
1254 (max_depth - ref_depth + 1);
1255 if (max_size == 0)
1256 return 0;
1257 src_size = src_entry->size;
1258 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1259 if (sizediff >= max_size)
1260 return 0;
1261 if (trg_size < src_size / 32)
1262 return 0;
1264 /* Load data if not already done */
1265 if (!trg->data) {
1266 read_lock();
1267 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1268 read_unlock();
1269 if (!trg->data)
1270 die("object %s cannot be read",
1271 sha1_to_hex(trg_entry->idx.sha1));
1272 if (sz != trg_size)
1273 die("object %s inconsistent object length (%lu vs %lu)",
1274 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1275 *mem_usage += sz;
1277 if (!src->data) {
1278 read_lock();
1279 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1280 read_unlock();
1281 if (!src->data)
1282 die("object %s cannot be read",
1283 sha1_to_hex(src_entry->idx.sha1));
1284 if (sz != src_size)
1285 die("object %s inconsistent object length (%lu vs %lu)",
1286 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1287 *mem_usage += sz;
1289 if (!src->index) {
1290 src->index = create_delta_index(src->data, src_size);
1291 if (!src->index) {
1292 static int warned = 0;
1293 if (!warned++)
1294 warning("suboptimal pack - out of memory");
1295 return 0;
1297 *mem_usage += sizeof_delta_index(src->index);
1300 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1301 if (!delta_buf)
1302 return 0;
1304 if (trg_entry->delta) {
1305 /* Prefer only shallower same-sized deltas. */
1306 if (delta_size == trg_entry->delta_size &&
1307 src->depth + 1 >= trg->depth) {
1308 free(delta_buf);
1309 return 0;
1314 * Handle memory allocation outside of the cache
1315 * accounting lock. Compiler will optimize the strangeness
1316 * away when THREADED_DELTA_SEARCH is not defined.
1318 free(trg_entry->delta_data);
1319 cache_lock();
1320 if (trg_entry->delta_data) {
1321 delta_cache_size -= trg_entry->delta_size;
1322 trg_entry->delta_data = NULL;
1324 if (delta_cacheable(src_size, trg_size, delta_size)) {
1325 delta_cache_size += delta_size;
1326 cache_unlock();
1327 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1328 } else {
1329 cache_unlock();
1330 free(delta_buf);
1333 trg_entry->delta = src_entry;
1334 trg_entry->delta_size = delta_size;
1335 trg->depth = src->depth + 1;
1337 return 1;
1340 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1342 struct object_entry *child = me->delta_child;
1343 unsigned int m = n;
1344 while (child) {
1345 unsigned int c = check_delta_limit(child, n + 1);
1346 if (m < c)
1347 m = c;
1348 child = child->delta_sibling;
1350 return m;
1353 static unsigned long free_unpacked(struct unpacked *n)
1355 unsigned long freed_mem = sizeof_delta_index(n->index);
1356 free_delta_index(n->index);
1357 n->index = NULL;
1358 if (n->data) {
1359 freed_mem += n->entry->size;
1360 free(n->data);
1361 n->data = NULL;
1363 n->entry = NULL;
1364 n->depth = 0;
1365 return freed_mem;
1368 static void find_deltas(struct object_entry **list, unsigned *list_size,
1369 int window, int depth, unsigned *processed)
1371 uint32_t i, idx = 0, count = 0;
1372 struct unpacked *array;
1373 unsigned long mem_usage = 0;
1375 array = xcalloc(window, sizeof(struct unpacked));
1377 for (;;) {
1378 struct object_entry *entry = *list++;
1379 struct unpacked *n = array + idx;
1380 int j, max_depth, best_base = -1;
1382 progress_lock();
1383 if (!*list_size) {
1384 progress_unlock();
1385 break;
1387 (*list_size)--;
1388 if (!entry->preferred_base) {
1389 (*processed)++;
1390 display_progress(progress_state, *processed);
1392 progress_unlock();
1394 mem_usage -= free_unpacked(n);
1395 n->entry = entry;
1397 while (window_memory_limit &&
1398 mem_usage > window_memory_limit &&
1399 count > 1) {
1400 uint32_t tail = (idx + window - count) % window;
1401 mem_usage -= free_unpacked(array + tail);
1402 count--;
1405 /* We do not compute delta to *create* objects we are not
1406 * going to pack.
1408 if (entry->preferred_base)
1409 goto next;
1412 * If the current object is at pack edge, take the depth the
1413 * objects that depend on the current object into account
1414 * otherwise they would become too deep.
1416 max_depth = depth;
1417 if (entry->delta_child) {
1418 max_depth -= check_delta_limit(entry, 0);
1419 if (max_depth <= 0)
1420 goto next;
1423 j = window;
1424 while (--j > 0) {
1425 int ret;
1426 uint32_t other_idx = idx + j;
1427 struct unpacked *m;
1428 if (other_idx >= window)
1429 other_idx -= window;
1430 m = array + other_idx;
1431 if (!m->entry)
1432 break;
1433 ret = try_delta(n, m, max_depth, &mem_usage);
1434 if (ret < 0)
1435 break;
1436 else if (ret > 0)
1437 best_base = other_idx;
1441 * If we decided to cache the delta data, then it is best
1442 * to compress it right away. First because we have to do
1443 * it anyway, and doing it here while we're threaded will
1444 * save a lot of time in the non threaded write phase,
1445 * as well as allow for caching more deltas within
1446 * the same cache size limit.
1447 * ...
1448 * But only if not writing to stdout, since in that case
1449 * the network is most likely throttling writes anyway,
1450 * and therefore it is best to go to the write phase ASAP
1451 * instead, as we can afford spending more time compressing
1452 * between writes at that moment.
1454 if (entry->delta_data && !pack_to_stdout) {
1455 entry->z_delta_size = do_compress(&entry->delta_data,
1456 entry->delta_size);
1457 cache_lock();
1458 delta_cache_size -= entry->delta_size;
1459 delta_cache_size += entry->z_delta_size;
1460 cache_unlock();
1463 /* if we made n a delta, and if n is already at max
1464 * depth, leaving it in the window is pointless. we
1465 * should evict it first.
1467 if (entry->delta && max_depth <= n->depth)
1468 continue;
1471 * Move the best delta base up in the window, after the
1472 * currently deltified object, to keep it longer. It will
1473 * be the first base object to be attempted next.
1475 if (entry->delta) {
1476 struct unpacked swap = array[best_base];
1477 int dist = (window + idx - best_base) % window;
1478 int dst = best_base;
1479 while (dist--) {
1480 int src = (dst + 1) % window;
1481 array[dst] = array[src];
1482 dst = src;
1484 array[dst] = swap;
1487 next:
1488 idx++;
1489 if (count + 1 < window)
1490 count++;
1491 if (idx >= window)
1492 idx = 0;
1495 for (i = 0; i < window; ++i) {
1496 free_delta_index(array[i].index);
1497 free(array[i].data);
1499 free(array);
1502 #ifdef THREADED_DELTA_SEARCH
1505 * The main thread waits on the condition that (at least) one of the workers
1506 * has stopped working (which is indicated in the .working member of
1507 * struct thread_params).
1508 * When a work thread has completed its work, it sets .working to 0 and
1509 * signals the main thread and waits on the condition that .data_ready
1510 * becomes 1.
1513 struct thread_params {
1514 pthread_t thread;
1515 struct object_entry **list;
1516 unsigned list_size;
1517 unsigned remaining;
1518 int window;
1519 int depth;
1520 int working;
1521 int data_ready;
1522 pthread_mutex_t mutex;
1523 pthread_cond_t cond;
1524 unsigned *processed;
1527 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1529 static void *threaded_find_deltas(void *arg)
1531 struct thread_params *me = arg;
1533 while (me->remaining) {
1534 find_deltas(me->list, &me->remaining,
1535 me->window, me->depth, me->processed);
1537 progress_lock();
1538 me->working = 0;
1539 pthread_cond_signal(&progress_cond);
1540 progress_unlock();
1543 * We must not set ->data_ready before we wait on the
1544 * condition because the main thread may have set it to 1
1545 * before we get here. In order to be sure that new
1546 * work is available if we see 1 in ->data_ready, it
1547 * was initialized to 0 before this thread was spawned
1548 * and we reset it to 0 right away.
1550 pthread_mutex_lock(&me->mutex);
1551 while (!me->data_ready)
1552 pthread_cond_wait(&me->cond, &me->mutex);
1553 me->data_ready = 0;
1554 pthread_mutex_unlock(&me->mutex);
1556 /* leave ->working 1 so that this doesn't get more work assigned */
1557 return NULL;
1560 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1561 int window, int depth, unsigned *processed)
1563 struct thread_params p[delta_search_threads];
1564 int i, ret, active_threads = 0;
1566 if (delta_search_threads <= 1) {
1567 find_deltas(list, &list_size, window, depth, processed);
1568 return;
1571 /* Partition the work amongst work threads. */
1572 for (i = 0; i < delta_search_threads; i++) {
1573 unsigned sub_size = list_size / (delta_search_threads - i);
1575 p[i].window = window;
1576 p[i].depth = depth;
1577 p[i].processed = processed;
1578 p[i].working = 1;
1579 p[i].data_ready = 0;
1581 /* try to split chunks on "path" boundaries */
1582 while (sub_size && sub_size < list_size &&
1583 list[sub_size]->hash &&
1584 list[sub_size]->hash == list[sub_size-1]->hash)
1585 sub_size++;
1587 p[i].list = list;
1588 p[i].list_size = sub_size;
1589 p[i].remaining = sub_size;
1591 list += sub_size;
1592 list_size -= sub_size;
1595 /* Start work threads. */
1596 for (i = 0; i < delta_search_threads; i++) {
1597 if (!p[i].list_size)
1598 continue;
1599 pthread_mutex_init(&p[i].mutex, NULL);
1600 pthread_cond_init(&p[i].cond, NULL);
1601 ret = pthread_create(&p[i].thread, NULL,
1602 threaded_find_deltas, &p[i]);
1603 if (ret)
1604 die("unable to create thread: %s", strerror(ret));
1605 active_threads++;
1609 * Now let's wait for work completion. Each time a thread is done
1610 * with its work, we steal half of the remaining work from the
1611 * thread with the largest number of unprocessed objects and give
1612 * it to that newly idle thread. This ensure good load balancing
1613 * until the remaining object list segments are simply too short
1614 * to be worth splitting anymore.
1616 while (active_threads) {
1617 struct thread_params *target = NULL;
1618 struct thread_params *victim = NULL;
1619 unsigned sub_size = 0;
1621 progress_lock();
1622 for (;;) {
1623 for (i = 0; !target && i < delta_search_threads; i++)
1624 if (!p[i].working)
1625 target = &p[i];
1626 if (target)
1627 break;
1628 pthread_cond_wait(&progress_cond, &progress_mutex);
1631 for (i = 0; i < delta_search_threads; i++)
1632 if (p[i].remaining > 2*window &&
1633 (!victim || victim->remaining < p[i].remaining))
1634 victim = &p[i];
1635 if (victim) {
1636 sub_size = victim->remaining / 2;
1637 list = victim->list + victim->list_size - sub_size;
1638 while (sub_size && list[0]->hash &&
1639 list[0]->hash == list[-1]->hash) {
1640 list++;
1641 sub_size--;
1643 if (!sub_size) {
1645 * It is possible for some "paths" to have
1646 * so many objects that no hash boundary
1647 * might be found. Let's just steal the
1648 * exact half in that case.
1650 sub_size = victim->remaining / 2;
1651 list -= sub_size;
1653 target->list = list;
1654 victim->list_size -= sub_size;
1655 victim->remaining -= sub_size;
1657 target->list_size = sub_size;
1658 target->remaining = sub_size;
1659 target->working = 1;
1660 progress_unlock();
1662 pthread_mutex_lock(&target->mutex);
1663 target->data_ready = 1;
1664 pthread_cond_signal(&target->cond);
1665 pthread_mutex_unlock(&target->mutex);
1667 if (!sub_size) {
1668 pthread_join(target->thread, NULL);
1669 pthread_cond_destroy(&target->cond);
1670 pthread_mutex_destroy(&target->mutex);
1671 active_threads--;
1676 #else
1677 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1678 #endif
1680 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1682 unsigned char peeled[20];
1684 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1685 !peel_ref(path, peeled) && /* peelable? */
1686 !is_null_sha1(peeled) && /* annotated tag? */
1687 locate_object_entry(peeled)) /* object packed? */
1688 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1689 return 0;
1692 static void prepare_pack(int window, int depth)
1694 struct object_entry **delta_list;
1695 uint32_t i, nr_deltas;
1696 unsigned n;
1698 get_object_details();
1700 if (!nr_objects || !window || !depth)
1701 return;
1703 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1704 nr_deltas = n = 0;
1706 for (i = 0; i < nr_objects; i++) {
1707 struct object_entry *entry = objects + i;
1709 if (entry->delta)
1710 /* This happens if we decided to reuse existing
1711 * delta from a pack. "reuse_delta &&" is implied.
1713 continue;
1715 if (entry->size < 50)
1716 continue;
1718 if (entry->no_try_delta)
1719 continue;
1721 if (!entry->preferred_base) {
1722 nr_deltas++;
1723 if (entry->type < 0)
1724 die("unable to get type of object %s",
1725 sha1_to_hex(entry->idx.sha1));
1726 } else {
1727 if (entry->type < 0) {
1729 * This object is not found, but we
1730 * don't have to include it anyway.
1732 continue;
1736 delta_list[n++] = entry;
1739 if (nr_deltas && n > 1) {
1740 unsigned nr_done = 0;
1741 if (progress)
1742 progress_state = start_progress("Compressing objects",
1743 nr_deltas);
1744 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1745 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1746 stop_progress(&progress_state);
1747 if (nr_done != nr_deltas)
1748 die("inconsistency with delta count");
1750 free(delta_list);
1753 static int git_pack_config(const char *k, const char *v, void *cb)
1755 if(!strcmp(k, "pack.window")) {
1756 window = git_config_int(k, v);
1757 return 0;
1759 if (!strcmp(k, "pack.windowmemory")) {
1760 window_memory_limit = git_config_ulong(k, v);
1761 return 0;
1763 if (!strcmp(k, "pack.depth")) {
1764 depth = git_config_int(k, v);
1765 return 0;
1767 if (!strcmp(k, "pack.compression")) {
1768 int level = git_config_int(k, v);
1769 if (level == -1)
1770 level = Z_DEFAULT_COMPRESSION;
1771 else if (level < 0 || level > Z_BEST_COMPRESSION)
1772 die("bad pack compression level %d", level);
1773 pack_compression_level = level;
1774 pack_compression_seen = 1;
1775 return 0;
1777 if (!strcmp(k, "pack.deltacachesize")) {
1778 max_delta_cache_size = git_config_int(k, v);
1779 return 0;
1781 if (!strcmp(k, "pack.deltacachelimit")) {
1782 cache_max_small_delta_size = git_config_int(k, v);
1783 return 0;
1785 if (!strcmp(k, "pack.threads")) {
1786 delta_search_threads = git_config_int(k, v);
1787 if (delta_search_threads < 0)
1788 die("invalid number of threads specified (%d)",
1789 delta_search_threads);
1790 #ifndef THREADED_DELTA_SEARCH
1791 if (delta_search_threads != 1)
1792 warning("no threads support, ignoring %s", k);
1793 #endif
1794 return 0;
1796 if (!strcmp(k, "pack.indexversion")) {
1797 pack_idx_default_version = git_config_int(k, v);
1798 if (pack_idx_default_version > 2)
1799 die("bad pack.indexversion=%"PRIu32,
1800 pack_idx_default_version);
1801 return 0;
1803 if (!strcmp(k, "pack.packsizelimit")) {
1804 pack_size_limit_cfg = git_config_ulong(k, v);
1805 return 0;
1807 return git_default_config(k, v, cb);
1810 static void read_object_list_from_stdin(void)
1812 char line[40 + 1 + PATH_MAX + 2];
1813 unsigned char sha1[20];
1815 for (;;) {
1816 if (!fgets(line, sizeof(line), stdin)) {
1817 if (feof(stdin))
1818 break;
1819 if (!ferror(stdin))
1820 die("fgets returned NULL, not EOF, not error!");
1821 if (errno != EINTR)
1822 die("fgets: %s", strerror(errno));
1823 clearerr(stdin);
1824 continue;
1826 if (line[0] == '-') {
1827 if (get_sha1_hex(line+1, sha1))
1828 die("expected edge sha1, got garbage:\n %s",
1829 line);
1830 add_preferred_base(sha1);
1831 continue;
1833 if (get_sha1_hex(line, sha1))
1834 die("expected sha1, got garbage:\n %s", line);
1836 add_preferred_base_object(line+41);
1837 add_object_entry(sha1, 0, line+41, 0);
1841 #define OBJECT_ADDED (1u<<20)
1843 static void show_commit(struct commit *commit)
1845 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1846 commit->object.flags |= OBJECT_ADDED;
1849 static void show_object(struct object_array_entry *p)
1851 add_preferred_base_object(p->name);
1852 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1853 p->item->flags |= OBJECT_ADDED;
1856 static void show_edge(struct commit *commit)
1858 add_preferred_base(commit->object.sha1);
1861 struct in_pack_object {
1862 off_t offset;
1863 struct object *object;
1866 struct in_pack {
1867 int alloc;
1868 int nr;
1869 struct in_pack_object *array;
1872 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1874 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1875 in_pack->array[in_pack->nr].object = object;
1876 in_pack->nr++;
1880 * Compare the objects in the offset order, in order to emulate the
1881 * "git rev-list --objects" output that produced the pack originally.
1883 static int ofscmp(const void *a_, const void *b_)
1885 struct in_pack_object *a = (struct in_pack_object *)a_;
1886 struct in_pack_object *b = (struct in_pack_object *)b_;
1888 if (a->offset < b->offset)
1889 return -1;
1890 else if (a->offset > b->offset)
1891 return 1;
1892 else
1893 return hashcmp(a->object->sha1, b->object->sha1);
1896 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1898 struct packed_git *p;
1899 struct in_pack in_pack;
1900 uint32_t i;
1902 memset(&in_pack, 0, sizeof(in_pack));
1904 for (p = packed_git; p; p = p->next) {
1905 const unsigned char *sha1;
1906 struct object *o;
1908 for (i = 0; i < revs->num_ignore_packed; i++) {
1909 if (matches_pack_name(p, revs->ignore_packed[i]))
1910 break;
1912 if (revs->num_ignore_packed <= i)
1913 continue;
1914 if (open_pack_index(p))
1915 die("cannot open pack index");
1917 ALLOC_GROW(in_pack.array,
1918 in_pack.nr + p->num_objects,
1919 in_pack.alloc);
1921 for (i = 0; i < p->num_objects; i++) {
1922 sha1 = nth_packed_object_sha1(p, i);
1923 o = lookup_unknown_object(sha1);
1924 if (!(o->flags & OBJECT_ADDED))
1925 mark_in_pack_object(o, p, &in_pack);
1926 o->flags |= OBJECT_ADDED;
1930 if (in_pack.nr) {
1931 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1932 ofscmp);
1933 for (i = 0; i < in_pack.nr; i++) {
1934 struct object *o = in_pack.array[i].object;
1935 add_object_entry(o->sha1, o->type, "", 0);
1938 free(in_pack.array);
1941 static void loosen_unused_packed_objects(struct rev_info *revs)
1943 struct packed_git *p;
1944 uint32_t i;
1945 const unsigned char *sha1;
1947 for (p = packed_git; p; p = p->next) {
1948 for (i = 0; i < revs->num_ignore_packed; i++) {
1949 if (matches_pack_name(p, revs->ignore_packed[i]))
1950 break;
1952 if (revs->num_ignore_packed <= i)
1953 continue;
1955 if (open_pack_index(p))
1956 die("cannot open pack index");
1958 for (i = 0; i < p->num_objects; i++) {
1959 sha1 = nth_packed_object_sha1(p, i);
1960 if (!locate_object_entry(sha1))
1961 if (force_object_loose(sha1, p->mtime))
1962 die("unable to force loose object");
1967 static void get_object_list(int ac, const char **av)
1969 struct rev_info revs;
1970 char line[1000];
1971 int flags = 0;
1973 init_revisions(&revs, NULL);
1974 save_commit_buffer = 0;
1975 setup_revisions(ac, av, &revs, NULL);
1977 while (fgets(line, sizeof(line), stdin) != NULL) {
1978 int len = strlen(line);
1979 if (len && line[len - 1] == '\n')
1980 line[--len] = 0;
1981 if (!len)
1982 break;
1983 if (*line == '-') {
1984 if (!strcmp(line, "--not")) {
1985 flags ^= UNINTERESTING;
1986 continue;
1988 die("not a rev '%s'", line);
1990 if (handle_revision_arg(line, &revs, flags, 1))
1991 die("bad revision '%s'", line);
1994 if (prepare_revision_walk(&revs))
1995 die("revision walk setup failed");
1996 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1997 traverse_commit_list(&revs, show_commit, show_object);
1999 if (keep_unreachable)
2000 add_objects_in_unpacked_packs(&revs);
2001 if (unpack_unreachable)
2002 loosen_unused_packed_objects(&revs);
2005 static int adjust_perm(const char *path, mode_t mode)
2007 if (chmod(path, mode))
2008 return -1;
2009 return adjust_shared_perm(path);
2012 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2014 int use_internal_rev_list = 0;
2015 int thin = 0;
2016 uint32_t i;
2017 const char **rp_av;
2018 int rp_ac_alloc = 64;
2019 int rp_ac;
2021 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2023 rp_av[0] = "pack-objects";
2024 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2025 rp_ac = 2;
2027 git_config(git_pack_config, NULL);
2028 if (!pack_compression_seen && core_compression_seen)
2029 pack_compression_level = core_compression_level;
2031 progress = isatty(2);
2032 for (i = 1; i < argc; i++) {
2033 const char *arg = argv[i];
2035 if (*arg != '-')
2036 break;
2038 if (!strcmp("--non-empty", arg)) {
2039 non_empty = 1;
2040 continue;
2042 if (!strcmp("--local", arg)) {
2043 local = 1;
2044 continue;
2046 if (!strcmp("--incremental", arg)) {
2047 incremental = 1;
2048 continue;
2050 if (!prefixcmp(arg, "--compression=")) {
2051 char *end;
2052 int level = strtoul(arg+14, &end, 0);
2053 if (!arg[14] || *end)
2054 usage(pack_usage);
2055 if (level == -1)
2056 level = Z_DEFAULT_COMPRESSION;
2057 else if (level < 0 || level > Z_BEST_COMPRESSION)
2058 die("bad pack compression level %d", level);
2059 pack_compression_level = level;
2060 continue;
2062 if (!prefixcmp(arg, "--max-pack-size=")) {
2063 char *end;
2064 pack_size_limit_cfg = 0;
2065 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2066 if (!arg[16] || *end)
2067 usage(pack_usage);
2068 continue;
2070 if (!prefixcmp(arg, "--window=")) {
2071 char *end;
2072 window = strtoul(arg+9, &end, 0);
2073 if (!arg[9] || *end)
2074 usage(pack_usage);
2075 continue;
2077 if (!prefixcmp(arg, "--window-memory=")) {
2078 if (!git_parse_ulong(arg+16, &window_memory_limit))
2079 usage(pack_usage);
2080 continue;
2082 if (!prefixcmp(arg, "--threads=")) {
2083 char *end;
2084 delta_search_threads = strtoul(arg+10, &end, 0);
2085 if (!arg[10] || *end || delta_search_threads < 0)
2086 usage(pack_usage);
2087 #ifndef THREADED_DELTA_SEARCH
2088 if (delta_search_threads != 1)
2089 warning("no threads support, "
2090 "ignoring %s", arg);
2091 #endif
2092 continue;
2094 if (!prefixcmp(arg, "--depth=")) {
2095 char *end;
2096 depth = strtoul(arg+8, &end, 0);
2097 if (!arg[8] || *end)
2098 usage(pack_usage);
2099 continue;
2101 if (!strcmp("--progress", arg)) {
2102 progress = 1;
2103 continue;
2105 if (!strcmp("--all-progress", arg)) {
2106 progress = 2;
2107 continue;
2109 if (!strcmp("-q", arg)) {
2110 progress = 0;
2111 continue;
2113 if (!strcmp("--no-reuse-delta", arg)) {
2114 reuse_delta = 0;
2115 continue;
2117 if (!strcmp("--no-reuse-object", arg)) {
2118 reuse_object = reuse_delta = 0;
2119 continue;
2121 if (!strcmp("--delta-base-offset", arg)) {
2122 allow_ofs_delta = 1;
2123 continue;
2125 if (!strcmp("--stdout", arg)) {
2126 pack_to_stdout = 1;
2127 continue;
2129 if (!strcmp("--revs", arg)) {
2130 use_internal_rev_list = 1;
2131 continue;
2133 if (!strcmp("--keep-unreachable", arg)) {
2134 keep_unreachable = 1;
2135 continue;
2137 if (!strcmp("--unpack-unreachable", arg)) {
2138 unpack_unreachable = 1;
2139 continue;
2141 if (!strcmp("--include-tag", arg)) {
2142 include_tag = 1;
2143 continue;
2145 if (!strcmp("--unpacked", arg) ||
2146 !prefixcmp(arg, "--unpacked=") ||
2147 !strcmp("--reflog", arg) ||
2148 !strcmp("--all", arg)) {
2149 use_internal_rev_list = 1;
2150 if (rp_ac >= rp_ac_alloc - 1) {
2151 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2152 rp_av = xrealloc(rp_av,
2153 rp_ac_alloc * sizeof(*rp_av));
2155 rp_av[rp_ac++] = arg;
2156 continue;
2158 if (!strcmp("--thin", arg)) {
2159 use_internal_rev_list = 1;
2160 thin = 1;
2161 rp_av[1] = "--objects-edge";
2162 continue;
2164 if (!prefixcmp(arg, "--index-version=")) {
2165 char *c;
2166 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2167 if (pack_idx_default_version > 2)
2168 die("bad %s", arg);
2169 if (*c == ',')
2170 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2171 if (*c || pack_idx_off32_limit & 0x80000000)
2172 die("bad %s", arg);
2173 continue;
2175 usage(pack_usage);
2178 /* Traditionally "pack-objects [options] base extra" failed;
2179 * we would however want to take refs parameter that would
2180 * have been given to upstream rev-list ourselves, which means
2181 * we somehow want to say what the base name is. So the
2182 * syntax would be:
2184 * pack-objects [options] base <refs...>
2186 * in other words, we would treat the first non-option as the
2187 * base_name and send everything else to the internal revision
2188 * walker.
2191 if (!pack_to_stdout)
2192 base_name = argv[i++];
2194 if (pack_to_stdout != !base_name)
2195 usage(pack_usage);
2197 if (!pack_to_stdout && !pack_size_limit)
2198 pack_size_limit = pack_size_limit_cfg;
2200 if (pack_to_stdout && pack_size_limit)
2201 die("--max-pack-size cannot be used to build a pack for transfer.");
2203 if (!pack_to_stdout && thin)
2204 die("--thin cannot be used to build an indexable pack.");
2206 if (keep_unreachable && unpack_unreachable)
2207 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2209 #ifdef THREADED_DELTA_SEARCH
2210 if (!delta_search_threads) /* --threads=0 means autodetect */
2211 delta_search_threads = online_cpus();
2212 #endif
2214 prepare_packed_git();
2216 if (progress)
2217 progress_state = start_progress("Counting objects", 0);
2218 if (!use_internal_rev_list)
2219 read_object_list_from_stdin();
2220 else {
2221 rp_av[rp_ac] = NULL;
2222 get_object_list(rp_ac, rp_av);
2224 if (include_tag && nr_result)
2225 for_each_ref(add_ref_tag, NULL);
2226 stop_progress(&progress_state);
2228 if (non_empty && !nr_result)
2229 return 0;
2230 if (nr_result)
2231 prepare_pack(window, depth);
2232 write_pack_file();
2233 if (progress)
2234 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2235 " reused %"PRIu32" (delta %"PRIu32")\n",
2236 written, written_delta, reused, reused_delta);
2237 return 0;