gitweb: use new Git::Repo API, and add optional caching
[git/gitweb-caching.git] / builtin-pack-objects.c
blobb43e0b8f27b1ead12c04d99a393e8b36f4d2cb46
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
20 #ifdef THREADED_DELTA_SEARCH
21 #include "thread-utils.h"
22 #include <pthread.h>
23 #endif
25 static const char pack_usage[] = "\
26 git-pack-objects [{ -q | --progress | --all-progress }] \n\
27 [--max-pack-size=N] [--local] [--incremental] \n\
28 [--window=N] [--window-memory=N] [--depth=N] \n\
29 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
30 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
31 [--stdout | base-name] [--include-tag] \n\
32 [--keep-unreachable | --unpack-unreachable] \n\
33 [<ref-list | <object-list]";
35 struct object_entry {
36 struct pack_idx_entry idx;
37 unsigned long size; /* uncompressed size */
38 struct packed_git *in_pack; /* already in pack */
39 off_t in_pack_offset;
40 struct object_entry *delta; /* delta base object */
41 struct object_entry *delta_child; /* deltified objects who bases me */
42 struct object_entry *delta_sibling; /* other deltified objects who
43 * uses the same base as me
45 void *delta_data; /* cached delta (uncompressed) */
46 unsigned long delta_size; /* delta data size (uncompressed) */
47 unsigned long z_delta_size; /* delta data size (compressed) */
48 unsigned int hash; /* name hint hash */
49 enum object_type type;
50 enum object_type in_pack_type; /* could be delta */
51 unsigned char in_pack_header_size;
52 unsigned char preferred_base; /* we do not pack this, but is available
53 * to be used as the base object to delta
54 * objects against.
56 unsigned char no_try_delta;
60 * Objects we are going to pack are collected in objects array (dynamically
61 * expanded). nr_objects & nr_alloc controls this array. They are stored
62 * in the order we see -- typically rev-list --objects order that gives us
63 * nice "minimum seek" order.
65 static struct object_entry *objects;
66 static struct pack_idx_entry **written_list;
67 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
69 static int non_empty;
70 static int reuse_delta = 1, reuse_object = 1;
71 static int keep_unreachable, unpack_unreachable, include_tag;
72 static int local;
73 static int incremental;
74 static int allow_ofs_delta;
75 static const char *base_name;
76 static int progress = 1;
77 static int window = 10;
78 static uint32_t pack_size_limit, pack_size_limit_cfg;
79 static int depth = 50;
80 static int delta_search_threads = 1;
81 static int pack_to_stdout;
82 static int num_preferred_base;
83 static struct progress *progress_state;
84 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
85 static int pack_compression_seen;
87 static unsigned long delta_cache_size = 0;
88 static unsigned long max_delta_cache_size = 0;
89 static unsigned long cache_max_small_delta_size = 1000;
91 static unsigned long window_memory_limit = 0;
94 * The object names in objects array are hashed with this hashtable,
95 * to help looking up the entry by object name.
96 * This hashtable is built after all the objects are seen.
98 static int *object_ix;
99 static int object_ix_hashsz;
102 * stats
104 static uint32_t written, written_delta;
105 static uint32_t reused, reused_delta;
108 static void *get_delta(struct object_entry *entry)
110 unsigned long size, base_size, delta_size;
111 void *buf, *base_buf, *delta_buf;
112 enum object_type type;
114 buf = read_sha1_file(entry->idx.sha1, &type, &size);
115 if (!buf)
116 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
117 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
118 if (!base_buf)
119 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
120 delta_buf = diff_delta(base_buf, base_size,
121 buf, size, &delta_size, 0);
122 if (!delta_buf || delta_size != entry->delta_size)
123 die("delta size changed");
124 free(buf);
125 free(base_buf);
126 return delta_buf;
129 static unsigned long do_compress(void **pptr, unsigned long size)
131 z_stream stream;
132 void *in, *out;
133 unsigned long maxsize;
135 memset(&stream, 0, sizeof(stream));
136 deflateInit(&stream, pack_compression_level);
137 maxsize = deflateBound(&stream, size);
139 in = *pptr;
140 out = xmalloc(maxsize);
141 *pptr = out;
143 stream.next_in = in;
144 stream.avail_in = size;
145 stream.next_out = out;
146 stream.avail_out = maxsize;
147 while (deflate(&stream, Z_FINISH) == Z_OK)
148 ; /* nothing */
149 deflateEnd(&stream);
151 free(in);
152 return stream.total_out;
156 * The per-object header is a pretty dense thing, which is
157 * - first byte: low four bits are "size", then three bits of "type",
158 * and the high bit is "size continues".
159 * - each byte afterwards: low seven bits are size continuation,
160 * with the high bit being "size continues"
162 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
164 int n = 1;
165 unsigned char c;
167 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
168 die("bad type %d", type);
170 c = (type << 4) | (size & 15);
171 size >>= 4;
172 while (size) {
173 *hdr++ = c | 0x80;
174 c = size & 0x7f;
175 size >>= 7;
176 n++;
178 *hdr = c;
179 return n;
183 * we are going to reuse the existing object data as is. make
184 * sure it is not corrupt.
186 static int check_pack_inflate(struct packed_git *p,
187 struct pack_window **w_curs,
188 off_t offset,
189 off_t len,
190 unsigned long expect)
192 z_stream stream;
193 unsigned char fakebuf[4096], *in;
194 int st;
196 memset(&stream, 0, sizeof(stream));
197 inflateInit(&stream);
198 do {
199 in = use_pack(p, w_curs, offset, &stream.avail_in);
200 stream.next_in = in;
201 stream.next_out = fakebuf;
202 stream.avail_out = sizeof(fakebuf);
203 st = inflate(&stream, Z_FINISH);
204 offset += stream.next_in - in;
205 } while (st == Z_OK || st == Z_BUF_ERROR);
206 inflateEnd(&stream);
207 return (st == Z_STREAM_END &&
208 stream.total_out == expect &&
209 stream.total_in == len) ? 0 : -1;
212 static void copy_pack_data(struct sha1file *f,
213 struct packed_git *p,
214 struct pack_window **w_curs,
215 off_t offset,
216 off_t len)
218 unsigned char *in;
219 unsigned int avail;
221 while (len) {
222 in = use_pack(p, w_curs, offset, &avail);
223 if (avail > len)
224 avail = (unsigned int)len;
225 sha1write(f, in, avail);
226 offset += avail;
227 len -= avail;
231 static unsigned long write_object(struct sha1file *f,
232 struct object_entry *entry,
233 off_t write_offset)
235 unsigned long size, limit, datalen;
236 void *buf;
237 unsigned char header[10], dheader[10];
238 unsigned hdrlen;
239 enum object_type type;
240 int usable_delta, to_reuse;
242 if (!pack_to_stdout)
243 crc32_begin(f);
245 type = entry->type;
247 /* write limit if limited packsize and not first object */
248 limit = pack_size_limit && nr_written ?
249 pack_size_limit - write_offset : 0;
251 if (!entry->delta)
252 usable_delta = 0; /* no delta */
253 else if (!pack_size_limit)
254 usable_delta = 1; /* unlimited packfile */
255 else if (entry->delta->idx.offset == (off_t)-1)
256 usable_delta = 0; /* base was written to another pack */
257 else if (entry->delta->idx.offset)
258 usable_delta = 1; /* base already exists in this pack */
259 else
260 usable_delta = 0; /* base could end up in another pack */
262 if (!reuse_object)
263 to_reuse = 0; /* explicit */
264 else if (!entry->in_pack)
265 to_reuse = 0; /* can't reuse what we don't have */
266 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
267 /* check_object() decided it for us ... */
268 to_reuse = usable_delta;
269 /* ... but pack split may override that */
270 else if (type != entry->in_pack_type)
271 to_reuse = 0; /* pack has delta which is unusable */
272 else if (entry->delta)
273 to_reuse = 0; /* we want to pack afresh */
274 else
275 to_reuse = 1; /* we have it in-pack undeltified,
276 * and we do not need to deltify it.
279 if (!to_reuse) {
280 if (!usable_delta) {
281 buf = read_sha1_file(entry->idx.sha1, &type, &size);
282 if (!buf)
283 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
285 * make sure no cached delta data remains from a
286 * previous attempt before a pack split occured.
288 free(entry->delta_data);
289 entry->delta_data = NULL;
290 entry->z_delta_size = 0;
291 } else if (entry->delta_data) {
292 size = entry->delta_size;
293 buf = entry->delta_data;
294 entry->delta_data = NULL;
295 type = (allow_ofs_delta && entry->delta->idx.offset) ?
296 OBJ_OFS_DELTA : OBJ_REF_DELTA;
297 } else {
298 buf = get_delta(entry);
299 size = entry->delta_size;
300 type = (allow_ofs_delta && entry->delta->idx.offset) ?
301 OBJ_OFS_DELTA : OBJ_REF_DELTA;
304 if (entry->z_delta_size)
305 datalen = entry->z_delta_size;
306 else
307 datalen = do_compress(&buf, size);
310 * The object header is a byte of 'type' followed by zero or
311 * more bytes of length.
313 hdrlen = encode_header(type, size, header);
315 if (type == OBJ_OFS_DELTA) {
317 * Deltas with relative base contain an additional
318 * encoding of the relative offset for the delta
319 * base from this object's position in the pack.
321 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
322 unsigned pos = sizeof(dheader) - 1;
323 dheader[pos] = ofs & 127;
324 while (ofs >>= 7)
325 dheader[--pos] = 128 | (--ofs & 127);
326 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
327 free(buf);
328 return 0;
330 sha1write(f, header, hdrlen);
331 sha1write(f, dheader + pos, sizeof(dheader) - pos);
332 hdrlen += sizeof(dheader) - pos;
333 } else if (type == OBJ_REF_DELTA) {
335 * Deltas with a base reference contain
336 * an additional 20 bytes for the base sha1.
338 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
339 free(buf);
340 return 0;
342 sha1write(f, header, hdrlen);
343 sha1write(f, entry->delta->idx.sha1, 20);
344 hdrlen += 20;
345 } else {
346 if (limit && hdrlen + datalen + 20 >= limit) {
347 free(buf);
348 return 0;
350 sha1write(f, header, hdrlen);
352 sha1write(f, buf, datalen);
353 free(buf);
355 else {
356 struct packed_git *p = entry->in_pack;
357 struct pack_window *w_curs = NULL;
358 struct revindex_entry *revidx;
359 off_t offset;
361 if (entry->delta) {
362 type = (allow_ofs_delta && entry->delta->idx.offset) ?
363 OBJ_OFS_DELTA : OBJ_REF_DELTA;
364 reused_delta++;
366 hdrlen = encode_header(type, entry->size, header);
367 offset = entry->in_pack_offset;
368 revidx = find_pack_revindex(p, offset);
369 datalen = revidx[1].offset - offset;
370 if (!pack_to_stdout && p->index_version > 1 &&
371 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
372 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
373 offset += entry->in_pack_header_size;
374 datalen -= entry->in_pack_header_size;
375 if (type == OBJ_OFS_DELTA) {
376 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
377 unsigned pos = sizeof(dheader) - 1;
378 dheader[pos] = ofs & 127;
379 while (ofs >>= 7)
380 dheader[--pos] = 128 | (--ofs & 127);
381 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
382 return 0;
383 sha1write(f, header, hdrlen);
384 sha1write(f, dheader + pos, sizeof(dheader) - pos);
385 hdrlen += sizeof(dheader) - pos;
386 } else if (type == OBJ_REF_DELTA) {
387 if (limit && hdrlen + 20 + datalen + 20 >= limit)
388 return 0;
389 sha1write(f, header, hdrlen);
390 sha1write(f, entry->delta->idx.sha1, 20);
391 hdrlen += 20;
392 } else {
393 if (limit && hdrlen + datalen + 20 >= limit)
394 return 0;
395 sha1write(f, header, hdrlen);
398 if (!pack_to_stdout && p->index_version == 1 &&
399 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
400 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
401 copy_pack_data(f, p, &w_curs, offset, datalen);
402 unuse_pack(&w_curs);
403 reused++;
405 if (usable_delta)
406 written_delta++;
407 written++;
408 if (!pack_to_stdout)
409 entry->idx.crc32 = crc32_end(f);
410 return hdrlen + datalen;
413 static off_t write_one(struct sha1file *f,
414 struct object_entry *e,
415 off_t offset)
417 unsigned long size;
419 /* offset is non zero if object is written already. */
420 if (e->idx.offset || e->preferred_base)
421 return offset;
423 /* if we are deltified, write out base object first. */
424 if (e->delta) {
425 offset = write_one(f, e->delta, offset);
426 if (!offset)
427 return 0;
430 e->idx.offset = offset;
431 size = write_object(f, e, offset);
432 if (!size) {
433 e->idx.offset = 0;
434 return 0;
436 written_list[nr_written++] = &e->idx;
438 /* make sure off_t is sufficiently large not to wrap */
439 if (offset > offset + size)
440 die("pack too large for current definition of off_t");
441 return offset + size;
444 /* forward declaration for write_pack_file */
445 static int adjust_perm(const char *path, mode_t mode);
447 static void write_pack_file(void)
449 uint32_t i = 0, j;
450 struct sha1file *f;
451 off_t offset, offset_one, last_obj_offset = 0;
452 struct pack_header hdr;
453 uint32_t nr_remaining = nr_result;
454 time_t last_mtime = 0;
456 if (progress > pack_to_stdout)
457 progress_state = start_progress("Writing objects", nr_result);
458 written_list = xmalloc(nr_objects * sizeof(*written_list));
460 do {
461 unsigned char sha1[20];
462 char *pack_tmp_name = NULL;
464 if (pack_to_stdout) {
465 f = sha1fd_throughput(1, "<stdout>", progress_state);
466 } else {
467 char tmpname[PATH_MAX];
468 int fd;
469 snprintf(tmpname, sizeof(tmpname),
470 "%s/tmp_pack_XXXXXX", get_object_directory());
471 fd = xmkstemp(tmpname);
472 pack_tmp_name = xstrdup(tmpname);
473 f = sha1fd(fd, pack_tmp_name);
476 hdr.hdr_signature = htonl(PACK_SIGNATURE);
477 hdr.hdr_version = htonl(PACK_VERSION);
478 hdr.hdr_entries = htonl(nr_remaining);
479 sha1write(f, &hdr, sizeof(hdr));
480 offset = sizeof(hdr);
481 nr_written = 0;
482 for (; i < nr_objects; i++) {
483 last_obj_offset = offset;
484 offset_one = write_one(f, objects + i, offset);
485 if (!offset_one)
486 break;
487 offset = offset_one;
488 display_progress(progress_state, written);
492 * Did we write the wrong # entries in the header?
493 * If so, rewrite it like in fast-import
495 if (pack_to_stdout) {
496 sha1close(f, sha1, CSUM_CLOSE);
497 } else if (nr_written == nr_remaining) {
498 sha1close(f, sha1, CSUM_FSYNC);
499 } else {
500 int fd = sha1close(f, NULL, 0);
501 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
502 fsync_or_die(fd, pack_tmp_name);
503 close(fd);
506 if (!pack_to_stdout) {
507 mode_t mode = umask(0);
508 struct stat st;
509 char *idx_tmp_name, tmpname[PATH_MAX];
511 umask(mode);
512 mode = 0444 & ~mode;
514 idx_tmp_name = write_idx_file(NULL, written_list,
515 nr_written, sha1);
517 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
518 base_name, sha1_to_hex(sha1));
519 if (adjust_perm(pack_tmp_name, mode))
520 die("unable to make temporary pack file readable: %s",
521 strerror(errno));
522 if (rename(pack_tmp_name, tmpname))
523 die("unable to rename temporary pack file: %s",
524 strerror(errno));
527 * Packs are runtime accessed in their mtime
528 * order since newer packs are more likely to contain
529 * younger objects. So if we are creating multiple
530 * packs then we should modify the mtime of later ones
531 * to preserve this property.
533 if (stat(tmpname, &st) < 0) {
534 warning("failed to stat %s: %s",
535 tmpname, strerror(errno));
536 } else if (!last_mtime) {
537 last_mtime = st.st_mtime;
538 } else {
539 struct utimbuf utb;
540 utb.actime = st.st_atime;
541 utb.modtime = --last_mtime;
542 if (utime(tmpname, &utb) < 0)
543 warning("failed utime() on %s: %s",
544 tmpname, strerror(errno));
547 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
548 base_name, sha1_to_hex(sha1));
549 if (adjust_perm(idx_tmp_name, mode))
550 die("unable to make temporary index file readable: %s",
551 strerror(errno));
552 if (rename(idx_tmp_name, tmpname))
553 die("unable to rename temporary index file: %s",
554 strerror(errno));
556 free(idx_tmp_name);
557 free(pack_tmp_name);
558 puts(sha1_to_hex(sha1));
561 /* mark written objects as written to previous pack */
562 for (j = 0; j < nr_written; j++) {
563 written_list[j]->offset = (off_t)-1;
565 nr_remaining -= nr_written;
566 } while (nr_remaining && i < nr_objects);
568 free(written_list);
569 stop_progress(&progress_state);
570 if (written != nr_result)
571 die("wrote %"PRIu32" objects while expecting %"PRIu32,
572 written, nr_result);
574 * We have scanned through [0 ... i). Since we have written
575 * the correct number of objects, the remaining [i ... nr_objects)
576 * items must be either already written (due to out-of-order delta base)
577 * or a preferred base. Count those which are neither and complain if any.
579 for (j = 0; i < nr_objects; i++) {
580 struct object_entry *e = objects + i;
581 j += !e->idx.offset && !e->preferred_base;
583 if (j)
584 die("wrote %"PRIu32" objects as expected but %"PRIu32
585 " unwritten", written, j);
588 static int locate_object_entry_hash(const unsigned char *sha1)
590 int i;
591 unsigned int ui;
592 memcpy(&ui, sha1, sizeof(unsigned int));
593 i = ui % object_ix_hashsz;
594 while (0 < object_ix[i]) {
595 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
596 return i;
597 if (++i == object_ix_hashsz)
598 i = 0;
600 return -1 - i;
603 static struct object_entry *locate_object_entry(const unsigned char *sha1)
605 int i;
607 if (!object_ix_hashsz)
608 return NULL;
610 i = locate_object_entry_hash(sha1);
611 if (0 <= i)
612 return &objects[object_ix[i]-1];
613 return NULL;
616 static void rehash_objects(void)
618 uint32_t i;
619 struct object_entry *oe;
621 object_ix_hashsz = nr_objects * 3;
622 if (object_ix_hashsz < 1024)
623 object_ix_hashsz = 1024;
624 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
625 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
626 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
627 int ix = locate_object_entry_hash(oe->idx.sha1);
628 if (0 <= ix)
629 continue;
630 ix = -1 - ix;
631 object_ix[ix] = i + 1;
635 static unsigned name_hash(const char *name)
637 unsigned char c;
638 unsigned hash = 0;
640 if (!name)
641 return 0;
644 * This effectively just creates a sortable number from the
645 * last sixteen non-whitespace characters. Last characters
646 * count "most", so things that end in ".c" sort together.
648 while ((c = *name++) != 0) {
649 if (isspace(c))
650 continue;
651 hash = (hash >> 2) + (c << 24);
653 return hash;
656 static void setup_delta_attr_check(struct git_attr_check *check)
658 static struct git_attr *attr_delta;
660 if (!attr_delta)
661 attr_delta = git_attr("delta", 5);
663 check[0].attr = attr_delta;
666 static int no_try_delta(const char *path)
668 struct git_attr_check check[1];
670 setup_delta_attr_check(check);
671 if (git_checkattr(path, ARRAY_SIZE(check), check))
672 return 0;
673 if (ATTR_FALSE(check->value))
674 return 1;
675 return 0;
678 static int add_object_entry(const unsigned char *sha1, enum object_type type,
679 const char *name, int exclude)
681 struct object_entry *entry;
682 struct packed_git *p, *found_pack = NULL;
683 off_t found_offset = 0;
684 int ix;
685 unsigned hash = name_hash(name);
687 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
688 if (ix >= 0) {
689 if (exclude) {
690 entry = objects + object_ix[ix] - 1;
691 if (!entry->preferred_base)
692 nr_result--;
693 entry->preferred_base = 1;
695 return 0;
698 for (p = packed_git; p; p = p->next) {
699 off_t offset = find_pack_entry_one(sha1, p);
700 if (offset) {
701 if (!found_pack) {
702 found_offset = offset;
703 found_pack = p;
705 if (exclude)
706 break;
707 if (incremental)
708 return 0;
709 if (local && !p->pack_local)
710 return 0;
714 if (nr_objects >= nr_alloc) {
715 nr_alloc = (nr_alloc + 1024) * 3 / 2;
716 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
719 entry = objects + nr_objects++;
720 memset(entry, 0, sizeof(*entry));
721 hashcpy(entry->idx.sha1, sha1);
722 entry->hash = hash;
723 if (type)
724 entry->type = type;
725 if (exclude)
726 entry->preferred_base = 1;
727 else
728 nr_result++;
729 if (found_pack) {
730 entry->in_pack = found_pack;
731 entry->in_pack_offset = found_offset;
734 if (object_ix_hashsz * 3 <= nr_objects * 4)
735 rehash_objects();
736 else
737 object_ix[-1 - ix] = nr_objects;
739 display_progress(progress_state, nr_objects);
741 if (name && no_try_delta(name))
742 entry->no_try_delta = 1;
744 return 1;
747 struct pbase_tree_cache {
748 unsigned char sha1[20];
749 int ref;
750 int temporary;
751 void *tree_data;
752 unsigned long tree_size;
755 static struct pbase_tree_cache *(pbase_tree_cache[256]);
756 static int pbase_tree_cache_ix(const unsigned char *sha1)
758 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
760 static int pbase_tree_cache_ix_incr(int ix)
762 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
765 static struct pbase_tree {
766 struct pbase_tree *next;
767 /* This is a phony "cache" entry; we are not
768 * going to evict it nor find it through _get()
769 * mechanism -- this is for the toplevel node that
770 * would almost always change with any commit.
772 struct pbase_tree_cache pcache;
773 } *pbase_tree;
775 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
777 struct pbase_tree_cache *ent, *nent;
778 void *data;
779 unsigned long size;
780 enum object_type type;
781 int neigh;
782 int my_ix = pbase_tree_cache_ix(sha1);
783 int available_ix = -1;
785 /* pbase-tree-cache acts as a limited hashtable.
786 * your object will be found at your index or within a few
787 * slots after that slot if it is cached.
789 for (neigh = 0; neigh < 8; neigh++) {
790 ent = pbase_tree_cache[my_ix];
791 if (ent && !hashcmp(ent->sha1, sha1)) {
792 ent->ref++;
793 return ent;
795 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
796 ((0 <= available_ix) &&
797 (!ent && pbase_tree_cache[available_ix])))
798 available_ix = my_ix;
799 if (!ent)
800 break;
801 my_ix = pbase_tree_cache_ix_incr(my_ix);
804 /* Did not find one. Either we got a bogus request or
805 * we need to read and perhaps cache.
807 data = read_sha1_file(sha1, &type, &size);
808 if (!data)
809 return NULL;
810 if (type != OBJ_TREE) {
811 free(data);
812 return NULL;
815 /* We need to either cache or return a throwaway copy */
817 if (available_ix < 0)
818 ent = NULL;
819 else {
820 ent = pbase_tree_cache[available_ix];
821 my_ix = available_ix;
824 if (!ent) {
825 nent = xmalloc(sizeof(*nent));
826 nent->temporary = (available_ix < 0);
828 else {
829 /* evict and reuse */
830 free(ent->tree_data);
831 nent = ent;
833 hashcpy(nent->sha1, sha1);
834 nent->tree_data = data;
835 nent->tree_size = size;
836 nent->ref = 1;
837 if (!nent->temporary)
838 pbase_tree_cache[my_ix] = nent;
839 return nent;
842 static void pbase_tree_put(struct pbase_tree_cache *cache)
844 if (!cache->temporary) {
845 cache->ref--;
846 return;
848 free(cache->tree_data);
849 free(cache);
852 static int name_cmp_len(const char *name)
854 int i;
855 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
857 return i;
860 static void add_pbase_object(struct tree_desc *tree,
861 const char *name,
862 int cmplen,
863 const char *fullname)
865 struct name_entry entry;
866 int cmp;
868 while (tree_entry(tree,&entry)) {
869 if (S_ISGITLINK(entry.mode))
870 continue;
871 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
872 memcmp(name, entry.path, cmplen);
873 if (cmp > 0)
874 continue;
875 if (cmp < 0)
876 return;
877 if (name[cmplen] != '/') {
878 add_object_entry(entry.sha1,
879 object_type(entry.mode),
880 fullname, 1);
881 return;
883 if (S_ISDIR(entry.mode)) {
884 struct tree_desc sub;
885 struct pbase_tree_cache *tree;
886 const char *down = name+cmplen+1;
887 int downlen = name_cmp_len(down);
889 tree = pbase_tree_get(entry.sha1);
890 if (!tree)
891 return;
892 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
894 add_pbase_object(&sub, down, downlen, fullname);
895 pbase_tree_put(tree);
900 static unsigned *done_pbase_paths;
901 static int done_pbase_paths_num;
902 static int done_pbase_paths_alloc;
903 static int done_pbase_path_pos(unsigned hash)
905 int lo = 0;
906 int hi = done_pbase_paths_num;
907 while (lo < hi) {
908 int mi = (hi + lo) / 2;
909 if (done_pbase_paths[mi] == hash)
910 return mi;
911 if (done_pbase_paths[mi] < hash)
912 hi = mi;
913 else
914 lo = mi + 1;
916 return -lo-1;
919 static int check_pbase_path(unsigned hash)
921 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
922 if (0 <= pos)
923 return 1;
924 pos = -pos - 1;
925 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
926 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
927 done_pbase_paths = xrealloc(done_pbase_paths,
928 done_pbase_paths_alloc *
929 sizeof(unsigned));
931 done_pbase_paths_num++;
932 if (pos < done_pbase_paths_num)
933 memmove(done_pbase_paths + pos + 1,
934 done_pbase_paths + pos,
935 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
936 done_pbase_paths[pos] = hash;
937 return 0;
940 static void add_preferred_base_object(const char *name)
942 struct pbase_tree *it;
943 int cmplen;
944 unsigned hash = name_hash(name);
946 if (!num_preferred_base || check_pbase_path(hash))
947 return;
949 cmplen = name_cmp_len(name);
950 for (it = pbase_tree; it; it = it->next) {
951 if (cmplen == 0) {
952 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
954 else {
955 struct tree_desc tree;
956 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
957 add_pbase_object(&tree, name, cmplen, name);
962 static void add_preferred_base(unsigned char *sha1)
964 struct pbase_tree *it;
965 void *data;
966 unsigned long size;
967 unsigned char tree_sha1[20];
969 if (window <= num_preferred_base++)
970 return;
972 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
973 if (!data)
974 return;
976 for (it = pbase_tree; it; it = it->next) {
977 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
978 free(data);
979 return;
983 it = xcalloc(1, sizeof(*it));
984 it->next = pbase_tree;
985 pbase_tree = it;
987 hashcpy(it->pcache.sha1, tree_sha1);
988 it->pcache.tree_data = data;
989 it->pcache.tree_size = size;
992 static void check_object(struct object_entry *entry)
994 if (entry->in_pack) {
995 struct packed_git *p = entry->in_pack;
996 struct pack_window *w_curs = NULL;
997 const unsigned char *base_ref = NULL;
998 struct object_entry *base_entry;
999 unsigned long used, used_0;
1000 unsigned int avail;
1001 off_t ofs;
1002 unsigned char *buf, c;
1004 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1007 * We want in_pack_type even if we do not reuse delta
1008 * since non-delta representations could still be reused.
1010 used = unpack_object_header_gently(buf, avail,
1011 &entry->in_pack_type,
1012 &entry->size);
1015 * Determine if this is a delta and if so whether we can
1016 * reuse it or not. Otherwise let's find out as cheaply as
1017 * possible what the actual type and size for this object is.
1019 switch (entry->in_pack_type) {
1020 default:
1021 /* Not a delta hence we've already got all we need. */
1022 entry->type = entry->in_pack_type;
1023 entry->in_pack_header_size = used;
1024 unuse_pack(&w_curs);
1025 return;
1026 case OBJ_REF_DELTA:
1027 if (reuse_delta && !entry->preferred_base)
1028 base_ref = use_pack(p, &w_curs,
1029 entry->in_pack_offset + used, NULL);
1030 entry->in_pack_header_size = used + 20;
1031 break;
1032 case OBJ_OFS_DELTA:
1033 buf = use_pack(p, &w_curs,
1034 entry->in_pack_offset + used, NULL);
1035 used_0 = 0;
1036 c = buf[used_0++];
1037 ofs = c & 127;
1038 while (c & 128) {
1039 ofs += 1;
1040 if (!ofs || MSB(ofs, 7))
1041 die("delta base offset overflow in pack for %s",
1042 sha1_to_hex(entry->idx.sha1));
1043 c = buf[used_0++];
1044 ofs = (ofs << 7) + (c & 127);
1046 if (ofs >= entry->in_pack_offset)
1047 die("delta base offset out of bound for %s",
1048 sha1_to_hex(entry->idx.sha1));
1049 ofs = entry->in_pack_offset - ofs;
1050 if (reuse_delta && !entry->preferred_base) {
1051 struct revindex_entry *revidx;
1052 revidx = find_pack_revindex(p, ofs);
1053 base_ref = nth_packed_object_sha1(p, revidx->nr);
1055 entry->in_pack_header_size = used + used_0;
1056 break;
1059 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1061 * If base_ref was set above that means we wish to
1062 * reuse delta data, and we even found that base
1063 * in the list of objects we want to pack. Goodie!
1065 * Depth value does not matter - find_deltas() will
1066 * never consider reused delta as the base object to
1067 * deltify other objects against, in order to avoid
1068 * circular deltas.
1070 entry->type = entry->in_pack_type;
1071 entry->delta = base_entry;
1072 entry->delta_sibling = base_entry->delta_child;
1073 base_entry->delta_child = entry;
1074 unuse_pack(&w_curs);
1075 return;
1078 if (entry->type) {
1080 * This must be a delta and we already know what the
1081 * final object type is. Let's extract the actual
1082 * object size from the delta header.
1084 entry->size = get_size_from_delta(p, &w_curs,
1085 entry->in_pack_offset + entry->in_pack_header_size);
1086 unuse_pack(&w_curs);
1087 return;
1091 * No choice but to fall back to the recursive delta walk
1092 * with sha1_object_info() to find about the object type
1093 * at this point...
1095 unuse_pack(&w_curs);
1098 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1100 * The error condition is checked in prepare_pack(). This is
1101 * to permit a missing preferred base object to be ignored
1102 * as a preferred base. Doing so can result in a larger
1103 * pack file, but the transfer will still take place.
1107 static int pack_offset_sort(const void *_a, const void *_b)
1109 const struct object_entry *a = *(struct object_entry **)_a;
1110 const struct object_entry *b = *(struct object_entry **)_b;
1112 /* avoid filesystem trashing with loose objects */
1113 if (!a->in_pack && !b->in_pack)
1114 return hashcmp(a->idx.sha1, b->idx.sha1);
1116 if (a->in_pack < b->in_pack)
1117 return -1;
1118 if (a->in_pack > b->in_pack)
1119 return 1;
1120 return a->in_pack_offset < b->in_pack_offset ? -1 :
1121 (a->in_pack_offset > b->in_pack_offset);
1124 static void get_object_details(void)
1126 uint32_t i;
1127 struct object_entry **sorted_by_offset;
1129 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1130 for (i = 0; i < nr_objects; i++)
1131 sorted_by_offset[i] = objects + i;
1132 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1134 for (i = 0; i < nr_objects; i++)
1135 check_object(sorted_by_offset[i]);
1137 free(sorted_by_offset);
1141 * We search for deltas in a list sorted by type, by filename hash, and then
1142 * by size, so that we see progressively smaller and smaller files.
1143 * That's because we prefer deltas to be from the bigger file
1144 * to the smaller -- deletes are potentially cheaper, but perhaps
1145 * more importantly, the bigger file is likely the more recent
1146 * one. The deepest deltas are therefore the oldest objects which are
1147 * less susceptible to be accessed often.
1149 static int type_size_sort(const void *_a, const void *_b)
1151 const struct object_entry *a = *(struct object_entry **)_a;
1152 const struct object_entry *b = *(struct object_entry **)_b;
1154 if (a->type > b->type)
1155 return -1;
1156 if (a->type < b->type)
1157 return 1;
1158 if (a->hash > b->hash)
1159 return -1;
1160 if (a->hash < b->hash)
1161 return 1;
1162 if (a->preferred_base > b->preferred_base)
1163 return -1;
1164 if (a->preferred_base < b->preferred_base)
1165 return 1;
1166 if (a->size > b->size)
1167 return -1;
1168 if (a->size < b->size)
1169 return 1;
1170 return a < b ? -1 : (a > b); /* newest first */
1173 struct unpacked {
1174 struct object_entry *entry;
1175 void *data;
1176 struct delta_index *index;
1177 unsigned depth;
1180 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1181 unsigned long delta_size)
1183 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1184 return 0;
1186 if (delta_size < cache_max_small_delta_size)
1187 return 1;
1189 /* cache delta, if objects are large enough compared to delta size */
1190 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1191 return 1;
1193 return 0;
1196 #ifdef THREADED_DELTA_SEARCH
1198 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1199 #define read_lock() pthread_mutex_lock(&read_mutex)
1200 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1202 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1203 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1204 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1206 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1207 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1208 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1210 #else
1212 #define read_lock() (void)0
1213 #define read_unlock() (void)0
1214 #define cache_lock() (void)0
1215 #define cache_unlock() (void)0
1216 #define progress_lock() (void)0
1217 #define progress_unlock() (void)0
1219 #endif
1221 static int try_delta(struct unpacked *trg, struct unpacked *src,
1222 unsigned max_depth, unsigned long *mem_usage)
1224 struct object_entry *trg_entry = trg->entry;
1225 struct object_entry *src_entry = src->entry;
1226 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1227 unsigned ref_depth;
1228 enum object_type type;
1229 void *delta_buf;
1231 /* Don't bother doing diffs between different types */
1232 if (trg_entry->type != src_entry->type)
1233 return -1;
1236 * We do not bother to try a delta that we discarded
1237 * on an earlier try, but only when reusing delta data.
1239 if (reuse_delta && trg_entry->in_pack &&
1240 trg_entry->in_pack == src_entry->in_pack &&
1241 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1242 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1243 return 0;
1245 /* Let's not bust the allowed depth. */
1246 if (src->depth >= max_depth)
1247 return 0;
1249 /* Now some size filtering heuristics. */
1250 trg_size = trg_entry->size;
1251 if (!trg_entry->delta) {
1252 max_size = trg_size/2 - 20;
1253 ref_depth = 1;
1254 } else {
1255 max_size = trg_entry->delta_size;
1256 ref_depth = trg->depth;
1258 max_size = max_size * (max_depth - src->depth) /
1259 (max_depth - ref_depth + 1);
1260 if (max_size == 0)
1261 return 0;
1262 src_size = src_entry->size;
1263 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1264 if (sizediff >= max_size)
1265 return 0;
1266 if (trg_size < src_size / 32)
1267 return 0;
1269 /* Load data if not already done */
1270 if (!trg->data) {
1271 read_lock();
1272 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1273 read_unlock();
1274 if (!trg->data)
1275 die("object %s cannot be read",
1276 sha1_to_hex(trg_entry->idx.sha1));
1277 if (sz != trg_size)
1278 die("object %s inconsistent object length (%lu vs %lu)",
1279 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1280 *mem_usage += sz;
1282 if (!src->data) {
1283 read_lock();
1284 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1285 read_unlock();
1286 if (!src->data)
1287 die("object %s cannot be read",
1288 sha1_to_hex(src_entry->idx.sha1));
1289 if (sz != src_size)
1290 die("object %s inconsistent object length (%lu vs %lu)",
1291 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1292 *mem_usage += sz;
1294 if (!src->index) {
1295 src->index = create_delta_index(src->data, src_size);
1296 if (!src->index) {
1297 static int warned = 0;
1298 if (!warned++)
1299 warning("suboptimal pack - out of memory");
1300 return 0;
1302 *mem_usage += sizeof_delta_index(src->index);
1305 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1306 if (!delta_buf)
1307 return 0;
1309 if (trg_entry->delta) {
1310 /* Prefer only shallower same-sized deltas. */
1311 if (delta_size == trg_entry->delta_size &&
1312 src->depth + 1 >= trg->depth) {
1313 free(delta_buf);
1314 return 0;
1319 * Handle memory allocation outside of the cache
1320 * accounting lock. Compiler will optimize the strangeness
1321 * away when THREADED_DELTA_SEARCH is not defined.
1323 free(trg_entry->delta_data);
1324 cache_lock();
1325 if (trg_entry->delta_data) {
1326 delta_cache_size -= trg_entry->delta_size;
1327 trg_entry->delta_data = NULL;
1329 if (delta_cacheable(src_size, trg_size, delta_size)) {
1330 delta_cache_size += delta_size;
1331 cache_unlock();
1332 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1333 } else {
1334 cache_unlock();
1335 free(delta_buf);
1338 trg_entry->delta = src_entry;
1339 trg_entry->delta_size = delta_size;
1340 trg->depth = src->depth + 1;
1342 return 1;
1345 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1347 struct object_entry *child = me->delta_child;
1348 unsigned int m = n;
1349 while (child) {
1350 unsigned int c = check_delta_limit(child, n + 1);
1351 if (m < c)
1352 m = c;
1353 child = child->delta_sibling;
1355 return m;
1358 static unsigned long free_unpacked(struct unpacked *n)
1360 unsigned long freed_mem = sizeof_delta_index(n->index);
1361 free_delta_index(n->index);
1362 n->index = NULL;
1363 if (n->data) {
1364 freed_mem += n->entry->size;
1365 free(n->data);
1366 n->data = NULL;
1368 n->entry = NULL;
1369 n->depth = 0;
1370 return freed_mem;
1373 static void find_deltas(struct object_entry **list, unsigned *list_size,
1374 int window, int depth, unsigned *processed)
1376 uint32_t i, idx = 0, count = 0;
1377 unsigned int array_size = window * sizeof(struct unpacked);
1378 struct unpacked *array;
1379 unsigned long mem_usage = 0;
1381 array = xmalloc(array_size);
1382 memset(array, 0, array_size);
1384 for (;;) {
1385 struct object_entry *entry = *list++;
1386 struct unpacked *n = array + idx;
1387 int j, max_depth, best_base = -1;
1389 progress_lock();
1390 if (!*list_size) {
1391 progress_unlock();
1392 break;
1394 (*list_size)--;
1395 if (!entry->preferred_base) {
1396 (*processed)++;
1397 display_progress(progress_state, *processed);
1399 progress_unlock();
1401 mem_usage -= free_unpacked(n);
1402 n->entry = entry;
1404 while (window_memory_limit &&
1405 mem_usage > window_memory_limit &&
1406 count > 1) {
1407 uint32_t tail = (idx + window - count) % window;
1408 mem_usage -= free_unpacked(array + tail);
1409 count--;
1412 /* We do not compute delta to *create* objects we are not
1413 * going to pack.
1415 if (entry->preferred_base)
1416 goto next;
1419 * If the current object is at pack edge, take the depth the
1420 * objects that depend on the current object into account
1421 * otherwise they would become too deep.
1423 max_depth = depth;
1424 if (entry->delta_child) {
1425 max_depth -= check_delta_limit(entry, 0);
1426 if (max_depth <= 0)
1427 goto next;
1430 j = window;
1431 while (--j > 0) {
1432 int ret;
1433 uint32_t other_idx = idx + j;
1434 struct unpacked *m;
1435 if (other_idx >= window)
1436 other_idx -= window;
1437 m = array + other_idx;
1438 if (!m->entry)
1439 break;
1440 ret = try_delta(n, m, max_depth, &mem_usage);
1441 if (ret < 0)
1442 break;
1443 else if (ret > 0)
1444 best_base = other_idx;
1448 * If we decided to cache the delta data, then it is best
1449 * to compress it right away. First because we have to do
1450 * it anyway, and doing it here while we're threaded will
1451 * save a lot of time in the non threaded write phase,
1452 * as well as allow for caching more deltas within
1453 * the same cache size limit.
1454 * ...
1455 * But only if not writing to stdout, since in that case
1456 * the network is most likely throttling writes anyway,
1457 * and therefore it is best to go to the write phase ASAP
1458 * instead, as we can afford spending more time compressing
1459 * between writes at that moment.
1461 if (entry->delta_data && !pack_to_stdout) {
1462 entry->z_delta_size = do_compress(&entry->delta_data,
1463 entry->delta_size);
1464 cache_lock();
1465 delta_cache_size -= entry->delta_size;
1466 delta_cache_size += entry->z_delta_size;
1467 cache_unlock();
1470 /* if we made n a delta, and if n is already at max
1471 * depth, leaving it in the window is pointless. we
1472 * should evict it first.
1474 if (entry->delta && max_depth <= n->depth)
1475 continue;
1478 * Move the best delta base up in the window, after the
1479 * currently deltified object, to keep it longer. It will
1480 * be the first base object to be attempted next.
1482 if (entry->delta) {
1483 struct unpacked swap = array[best_base];
1484 int dist = (window + idx - best_base) % window;
1485 int dst = best_base;
1486 while (dist--) {
1487 int src = (dst + 1) % window;
1488 array[dst] = array[src];
1489 dst = src;
1491 array[dst] = swap;
1494 next:
1495 idx++;
1496 if (count + 1 < window)
1497 count++;
1498 if (idx >= window)
1499 idx = 0;
1502 for (i = 0; i < window; ++i) {
1503 free_delta_index(array[i].index);
1504 free(array[i].data);
1506 free(array);
1509 #ifdef THREADED_DELTA_SEARCH
1512 * The main thread waits on the condition that (at least) one of the workers
1513 * has stopped working (which is indicated in the .working member of
1514 * struct thread_params).
1515 * When a work thread has completed its work, it sets .working to 0 and
1516 * signals the main thread and waits on the condition that .data_ready
1517 * becomes 1.
1520 struct thread_params {
1521 pthread_t thread;
1522 struct object_entry **list;
1523 unsigned list_size;
1524 unsigned remaining;
1525 int window;
1526 int depth;
1527 int working;
1528 int data_ready;
1529 pthread_mutex_t mutex;
1530 pthread_cond_t cond;
1531 unsigned *processed;
1534 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1536 static void *threaded_find_deltas(void *arg)
1538 struct thread_params *me = arg;
1540 while (me->remaining) {
1541 find_deltas(me->list, &me->remaining,
1542 me->window, me->depth, me->processed);
1544 progress_lock();
1545 me->working = 0;
1546 pthread_cond_signal(&progress_cond);
1547 progress_unlock();
1550 * We must not set ->data_ready before we wait on the
1551 * condition because the main thread may have set it to 1
1552 * before we get here. In order to be sure that new
1553 * work is available if we see 1 in ->data_ready, it
1554 * was initialized to 0 before this thread was spawned
1555 * and we reset it to 0 right away.
1557 pthread_mutex_lock(&me->mutex);
1558 while (!me->data_ready)
1559 pthread_cond_wait(&me->cond, &me->mutex);
1560 me->data_ready = 0;
1561 pthread_mutex_unlock(&me->mutex);
1563 /* leave ->working 1 so that this doesn't get more work assigned */
1564 return NULL;
1567 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1568 int window, int depth, unsigned *processed)
1570 struct thread_params p[delta_search_threads];
1571 int i, ret, active_threads = 0;
1573 if (delta_search_threads <= 1) {
1574 find_deltas(list, &list_size, window, depth, processed);
1575 return;
1578 /* Partition the work amongst work threads. */
1579 for (i = 0; i < delta_search_threads; i++) {
1580 unsigned sub_size = list_size / (delta_search_threads - i);
1582 p[i].window = window;
1583 p[i].depth = depth;
1584 p[i].processed = processed;
1585 p[i].working = 1;
1586 p[i].data_ready = 0;
1588 /* try to split chunks on "path" boundaries */
1589 while (sub_size && sub_size < list_size &&
1590 list[sub_size]->hash &&
1591 list[sub_size]->hash == list[sub_size-1]->hash)
1592 sub_size++;
1594 p[i].list = list;
1595 p[i].list_size = sub_size;
1596 p[i].remaining = sub_size;
1598 list += sub_size;
1599 list_size -= sub_size;
1602 /* Start work threads. */
1603 for (i = 0; i < delta_search_threads; i++) {
1604 if (!p[i].list_size)
1605 continue;
1606 pthread_mutex_init(&p[i].mutex, NULL);
1607 pthread_cond_init(&p[i].cond, NULL);
1608 ret = pthread_create(&p[i].thread, NULL,
1609 threaded_find_deltas, &p[i]);
1610 if (ret)
1611 die("unable to create thread: %s", strerror(ret));
1612 active_threads++;
1616 * Now let's wait for work completion. Each time a thread is done
1617 * with its work, we steal half of the remaining work from the
1618 * thread with the largest number of unprocessed objects and give
1619 * it to that newly idle thread. This ensure good load balancing
1620 * until the remaining object list segments are simply too short
1621 * to be worth splitting anymore.
1623 while (active_threads) {
1624 struct thread_params *target = NULL;
1625 struct thread_params *victim = NULL;
1626 unsigned sub_size = 0;
1628 progress_lock();
1629 for (;;) {
1630 for (i = 0; !target && i < delta_search_threads; i++)
1631 if (!p[i].working)
1632 target = &p[i];
1633 if (target)
1634 break;
1635 pthread_cond_wait(&progress_cond, &progress_mutex);
1638 for (i = 0; i < delta_search_threads; i++)
1639 if (p[i].remaining > 2*window &&
1640 (!victim || victim->remaining < p[i].remaining))
1641 victim = &p[i];
1642 if (victim) {
1643 sub_size = victim->remaining / 2;
1644 list = victim->list + victim->list_size - sub_size;
1645 while (sub_size && list[0]->hash &&
1646 list[0]->hash == list[-1]->hash) {
1647 list++;
1648 sub_size--;
1650 if (!sub_size) {
1652 * It is possible for some "paths" to have
1653 * so many objects that no hash boundary
1654 * might be found. Let's just steal the
1655 * exact half in that case.
1657 sub_size = victim->remaining / 2;
1658 list -= sub_size;
1660 target->list = list;
1661 victim->list_size -= sub_size;
1662 victim->remaining -= sub_size;
1664 target->list_size = sub_size;
1665 target->remaining = sub_size;
1666 target->working = 1;
1667 progress_unlock();
1669 pthread_mutex_lock(&target->mutex);
1670 target->data_ready = 1;
1671 pthread_cond_signal(&target->cond);
1672 pthread_mutex_unlock(&target->mutex);
1674 if (!sub_size) {
1675 pthread_join(target->thread, NULL);
1676 pthread_cond_destroy(&target->cond);
1677 pthread_mutex_destroy(&target->mutex);
1678 active_threads--;
1683 #else
1684 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1685 #endif
1687 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1689 unsigned char peeled[20];
1691 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1692 !peel_ref(path, peeled) && /* peelable? */
1693 !is_null_sha1(peeled) && /* annotated tag? */
1694 locate_object_entry(peeled)) /* object packed? */
1695 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1696 return 0;
1699 static void prepare_pack(int window, int depth)
1701 struct object_entry **delta_list;
1702 uint32_t i, nr_deltas;
1703 unsigned n;
1705 get_object_details();
1707 if (!nr_objects || !window || !depth)
1708 return;
1710 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1711 nr_deltas = n = 0;
1713 for (i = 0; i < nr_objects; i++) {
1714 struct object_entry *entry = objects + i;
1716 if (entry->delta)
1717 /* This happens if we decided to reuse existing
1718 * delta from a pack. "reuse_delta &&" is implied.
1720 continue;
1722 if (entry->size < 50)
1723 continue;
1725 if (entry->no_try_delta)
1726 continue;
1728 if (!entry->preferred_base) {
1729 nr_deltas++;
1730 if (entry->type < 0)
1731 die("unable to get type of object %s",
1732 sha1_to_hex(entry->idx.sha1));
1735 delta_list[n++] = entry;
1738 if (nr_deltas && n > 1) {
1739 unsigned nr_done = 0;
1740 if (progress)
1741 progress_state = start_progress("Compressing objects",
1742 nr_deltas);
1743 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1744 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1745 stop_progress(&progress_state);
1746 if (nr_done != nr_deltas)
1747 die("inconsistency with delta count");
1749 free(delta_list);
1752 static int git_pack_config(const char *k, const char *v, void *cb)
1754 if(!strcmp(k, "pack.window")) {
1755 window = git_config_int(k, v);
1756 return 0;
1758 if (!strcmp(k, "pack.windowmemory")) {
1759 window_memory_limit = git_config_ulong(k, v);
1760 return 0;
1762 if (!strcmp(k, "pack.depth")) {
1763 depth = git_config_int(k, v);
1764 return 0;
1766 if (!strcmp(k, "pack.compression")) {
1767 int level = git_config_int(k, v);
1768 if (level == -1)
1769 level = Z_DEFAULT_COMPRESSION;
1770 else if (level < 0 || level > Z_BEST_COMPRESSION)
1771 die("bad pack compression level %d", level);
1772 pack_compression_level = level;
1773 pack_compression_seen = 1;
1774 return 0;
1776 if (!strcmp(k, "pack.deltacachesize")) {
1777 max_delta_cache_size = git_config_int(k, v);
1778 return 0;
1780 if (!strcmp(k, "pack.deltacachelimit")) {
1781 cache_max_small_delta_size = git_config_int(k, v);
1782 return 0;
1784 if (!strcmp(k, "pack.threads")) {
1785 delta_search_threads = git_config_int(k, v);
1786 if (delta_search_threads < 0)
1787 die("invalid number of threads specified (%d)",
1788 delta_search_threads);
1789 #ifndef THREADED_DELTA_SEARCH
1790 if (delta_search_threads != 1)
1791 warning("no threads support, ignoring %s", k);
1792 #endif
1793 return 0;
1795 if (!strcmp(k, "pack.indexversion")) {
1796 pack_idx_default_version = git_config_int(k, v);
1797 if (pack_idx_default_version > 2)
1798 die("bad pack.indexversion=%"PRIu32,
1799 pack_idx_default_version);
1800 return 0;
1802 if (!strcmp(k, "pack.packsizelimit")) {
1803 pack_size_limit_cfg = git_config_ulong(k, v);
1804 return 0;
1806 return git_default_config(k, v, cb);
1809 static void read_object_list_from_stdin(void)
1811 char line[40 + 1 + PATH_MAX + 2];
1812 unsigned char sha1[20];
1814 for (;;) {
1815 if (!fgets(line, sizeof(line), stdin)) {
1816 if (feof(stdin))
1817 break;
1818 if (!ferror(stdin))
1819 die("fgets returned NULL, not EOF, not error!");
1820 if (errno != EINTR)
1821 die("fgets: %s", strerror(errno));
1822 clearerr(stdin);
1823 continue;
1825 if (line[0] == '-') {
1826 if (get_sha1_hex(line+1, sha1))
1827 die("expected edge sha1, got garbage:\n %s",
1828 line);
1829 add_preferred_base(sha1);
1830 continue;
1832 if (get_sha1_hex(line, sha1))
1833 die("expected sha1, got garbage:\n %s", line);
1835 add_preferred_base_object(line+41);
1836 add_object_entry(sha1, 0, line+41, 0);
1840 #define OBJECT_ADDED (1u<<20)
1842 static void show_commit(struct commit *commit)
1844 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1845 commit->object.flags |= OBJECT_ADDED;
1848 static void show_object(struct object_array_entry *p)
1850 add_preferred_base_object(p->name);
1851 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1852 p->item->flags |= OBJECT_ADDED;
1855 static void show_edge(struct commit *commit)
1857 add_preferred_base(commit->object.sha1);
1860 struct in_pack_object {
1861 off_t offset;
1862 struct object *object;
1865 struct in_pack {
1866 int alloc;
1867 int nr;
1868 struct in_pack_object *array;
1871 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1873 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1874 in_pack->array[in_pack->nr].object = object;
1875 in_pack->nr++;
1879 * Compare the objects in the offset order, in order to emulate the
1880 * "git-rev-list --objects" output that produced the pack originally.
1882 static int ofscmp(const void *a_, const void *b_)
1884 struct in_pack_object *a = (struct in_pack_object *)a_;
1885 struct in_pack_object *b = (struct in_pack_object *)b_;
1887 if (a->offset < b->offset)
1888 return -1;
1889 else if (a->offset > b->offset)
1890 return 1;
1891 else
1892 return hashcmp(a->object->sha1, b->object->sha1);
1895 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1897 struct packed_git *p;
1898 struct in_pack in_pack;
1899 uint32_t i;
1901 memset(&in_pack, 0, sizeof(in_pack));
1903 for (p = packed_git; p; p = p->next) {
1904 const unsigned char *sha1;
1905 struct object *o;
1907 for (i = 0; i < revs->num_ignore_packed; i++) {
1908 if (matches_pack_name(p, revs->ignore_packed[i]))
1909 break;
1911 if (revs->num_ignore_packed <= i)
1912 continue;
1913 if (open_pack_index(p))
1914 die("cannot open pack index");
1916 ALLOC_GROW(in_pack.array,
1917 in_pack.nr + p->num_objects,
1918 in_pack.alloc);
1920 for (i = 0; i < p->num_objects; i++) {
1921 sha1 = nth_packed_object_sha1(p, i);
1922 o = lookup_unknown_object(sha1);
1923 if (!(o->flags & OBJECT_ADDED))
1924 mark_in_pack_object(o, p, &in_pack);
1925 o->flags |= OBJECT_ADDED;
1929 if (in_pack.nr) {
1930 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1931 ofscmp);
1932 for (i = 0; i < in_pack.nr; i++) {
1933 struct object *o = in_pack.array[i].object;
1934 add_object_entry(o->sha1, o->type, "", 0);
1937 free(in_pack.array);
1940 static void loosen_unused_packed_objects(struct rev_info *revs)
1942 struct packed_git *p;
1943 uint32_t i;
1944 const unsigned char *sha1;
1946 for (p = packed_git; p; p = p->next) {
1947 for (i = 0; i < revs->num_ignore_packed; i++) {
1948 if (matches_pack_name(p, revs->ignore_packed[i]))
1949 break;
1951 if (revs->num_ignore_packed <= i)
1952 continue;
1954 if (open_pack_index(p))
1955 die("cannot open pack index");
1957 for (i = 0; i < p->num_objects; i++) {
1958 sha1 = nth_packed_object_sha1(p, i);
1959 if (!locate_object_entry(sha1))
1960 if (force_object_loose(sha1, p->mtime))
1961 die("unable to force loose object");
1966 static void get_object_list(int ac, const char **av)
1968 struct rev_info revs;
1969 char line[1000];
1970 int flags = 0;
1972 init_revisions(&revs, NULL);
1973 save_commit_buffer = 0;
1974 setup_revisions(ac, av, &revs, NULL);
1976 while (fgets(line, sizeof(line), stdin) != NULL) {
1977 int len = strlen(line);
1978 if (len && line[len - 1] == '\n')
1979 line[--len] = 0;
1980 if (!len)
1981 break;
1982 if (*line == '-') {
1983 if (!strcmp(line, "--not")) {
1984 flags ^= UNINTERESTING;
1985 continue;
1987 die("not a rev '%s'", line);
1989 if (handle_revision_arg(line, &revs, flags, 1))
1990 die("bad revision '%s'", line);
1993 if (prepare_revision_walk(&revs))
1994 die("revision walk setup failed");
1995 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1996 traverse_commit_list(&revs, show_commit, show_object);
1998 if (keep_unreachable)
1999 add_objects_in_unpacked_packs(&revs);
2000 if (unpack_unreachable)
2001 loosen_unused_packed_objects(&revs);
2004 static int adjust_perm(const char *path, mode_t mode)
2006 if (chmod(path, mode))
2007 return -1;
2008 return adjust_shared_perm(path);
2011 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2013 int use_internal_rev_list = 0;
2014 int thin = 0;
2015 uint32_t i;
2016 const char **rp_av;
2017 int rp_ac_alloc = 64;
2018 int rp_ac;
2020 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2022 rp_av[0] = "pack-objects";
2023 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2024 rp_ac = 2;
2026 git_config(git_pack_config, NULL);
2027 if (!pack_compression_seen && core_compression_seen)
2028 pack_compression_level = core_compression_level;
2030 progress = isatty(2);
2031 for (i = 1; i < argc; i++) {
2032 const char *arg = argv[i];
2034 if (*arg != '-')
2035 break;
2037 if (!strcmp("--non-empty", arg)) {
2038 non_empty = 1;
2039 continue;
2041 if (!strcmp("--local", arg)) {
2042 local = 1;
2043 continue;
2045 if (!strcmp("--incremental", arg)) {
2046 incremental = 1;
2047 continue;
2049 if (!prefixcmp(arg, "--compression=")) {
2050 char *end;
2051 int level = strtoul(arg+14, &end, 0);
2052 if (!arg[14] || *end)
2053 usage(pack_usage);
2054 if (level == -1)
2055 level = Z_DEFAULT_COMPRESSION;
2056 else if (level < 0 || level > Z_BEST_COMPRESSION)
2057 die("bad pack compression level %d", level);
2058 pack_compression_level = level;
2059 continue;
2061 if (!prefixcmp(arg, "--max-pack-size=")) {
2062 char *end;
2063 pack_size_limit_cfg = 0;
2064 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2065 if (!arg[16] || *end)
2066 usage(pack_usage);
2067 continue;
2069 if (!prefixcmp(arg, "--window=")) {
2070 char *end;
2071 window = strtoul(arg+9, &end, 0);
2072 if (!arg[9] || *end)
2073 usage(pack_usage);
2074 continue;
2076 if (!prefixcmp(arg, "--window-memory=")) {
2077 if (!git_parse_ulong(arg+16, &window_memory_limit))
2078 usage(pack_usage);
2079 continue;
2081 if (!prefixcmp(arg, "--threads=")) {
2082 char *end;
2083 delta_search_threads = strtoul(arg+10, &end, 0);
2084 if (!arg[10] || *end || delta_search_threads < 0)
2085 usage(pack_usage);
2086 #ifndef THREADED_DELTA_SEARCH
2087 if (delta_search_threads != 1)
2088 warning("no threads support, "
2089 "ignoring %s", arg);
2090 #endif
2091 continue;
2093 if (!prefixcmp(arg, "--depth=")) {
2094 char *end;
2095 depth = strtoul(arg+8, &end, 0);
2096 if (!arg[8] || *end)
2097 usage(pack_usage);
2098 continue;
2100 if (!strcmp("--progress", arg)) {
2101 progress = 1;
2102 continue;
2104 if (!strcmp("--all-progress", arg)) {
2105 progress = 2;
2106 continue;
2108 if (!strcmp("-q", arg)) {
2109 progress = 0;
2110 continue;
2112 if (!strcmp("--no-reuse-delta", arg)) {
2113 reuse_delta = 0;
2114 continue;
2116 if (!strcmp("--no-reuse-object", arg)) {
2117 reuse_object = reuse_delta = 0;
2118 continue;
2120 if (!strcmp("--delta-base-offset", arg)) {
2121 allow_ofs_delta = 1;
2122 continue;
2124 if (!strcmp("--stdout", arg)) {
2125 pack_to_stdout = 1;
2126 continue;
2128 if (!strcmp("--revs", arg)) {
2129 use_internal_rev_list = 1;
2130 continue;
2132 if (!strcmp("--keep-unreachable", arg)) {
2133 keep_unreachable = 1;
2134 continue;
2136 if (!strcmp("--unpack-unreachable", arg)) {
2137 unpack_unreachable = 1;
2138 continue;
2140 if (!strcmp("--include-tag", arg)) {
2141 include_tag = 1;
2142 continue;
2144 if (!strcmp("--unpacked", arg) ||
2145 !prefixcmp(arg, "--unpacked=") ||
2146 !strcmp("--reflog", arg) ||
2147 !strcmp("--all", arg)) {
2148 use_internal_rev_list = 1;
2149 if (rp_ac >= rp_ac_alloc - 1) {
2150 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2151 rp_av = xrealloc(rp_av,
2152 rp_ac_alloc * sizeof(*rp_av));
2154 rp_av[rp_ac++] = arg;
2155 continue;
2157 if (!strcmp("--thin", arg)) {
2158 use_internal_rev_list = 1;
2159 thin = 1;
2160 rp_av[1] = "--objects-edge";
2161 continue;
2163 if (!prefixcmp(arg, "--index-version=")) {
2164 char *c;
2165 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2166 if (pack_idx_default_version > 2)
2167 die("bad %s", arg);
2168 if (*c == ',')
2169 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2170 if (*c || pack_idx_off32_limit & 0x80000000)
2171 die("bad %s", arg);
2172 continue;
2174 usage(pack_usage);
2177 /* Traditionally "pack-objects [options] base extra" failed;
2178 * we would however want to take refs parameter that would
2179 * have been given to upstream rev-list ourselves, which means
2180 * we somehow want to say what the base name is. So the
2181 * syntax would be:
2183 * pack-objects [options] base <refs...>
2185 * in other words, we would treat the first non-option as the
2186 * base_name and send everything else to the internal revision
2187 * walker.
2190 if (!pack_to_stdout)
2191 base_name = argv[i++];
2193 if (pack_to_stdout != !base_name)
2194 usage(pack_usage);
2196 if (!pack_to_stdout && !pack_size_limit)
2197 pack_size_limit = pack_size_limit_cfg;
2199 if (pack_to_stdout && pack_size_limit)
2200 die("--max-pack-size cannot be used to build a pack for transfer.");
2202 if (!pack_to_stdout && thin)
2203 die("--thin cannot be used to build an indexable pack.");
2205 if (keep_unreachable && unpack_unreachable)
2206 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2208 #ifdef THREADED_DELTA_SEARCH
2209 if (!delta_search_threads) /* --threads=0 means autodetect */
2210 delta_search_threads = online_cpus();
2211 #endif
2213 prepare_packed_git();
2215 if (progress)
2216 progress_state = start_progress("Counting objects", 0);
2217 if (!use_internal_rev_list)
2218 read_object_list_from_stdin();
2219 else {
2220 rp_av[rp_ac] = NULL;
2221 get_object_list(rp_ac, rp_av);
2223 if (include_tag && nr_result)
2224 for_each_ref(add_ref_tag, NULL);
2225 stop_progress(&progress_state);
2227 if (non_empty && !nr_result)
2228 return 0;
2229 if (nr_result)
2230 prepare_pack(window, depth);
2231 write_pack_file();
2232 if (progress)
2233 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2234 " reused %"PRIu32" (delta %"PRIu32")\n",
2235 written, written_delta, reused, reused_delta);
2236 return 0;