Merge branch 'maint-1.5.6' into maint-1.6.0
[git/spearce.git] / builtin-pack-objects.c
bloba6adc8c271e20374b86462391016897f1d17a0f3
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
20 #ifdef THREADED_DELTA_SEARCH
21 #include "thread-utils.h"
22 #include <pthread.h>
23 #endif
25 static const char pack_usage[] = "\
26 git pack-objects [{ -q | --progress | --all-progress }] \n\
27 [--max-pack-size=N] [--local] [--incremental] \n\
28 [--window=N] [--window-memory=N] [--depth=N] \n\
29 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
30 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
31 [--stdout | base-name] [--include-tag] \n\
32 [--keep-unreachable | --unpack-unreachable] \n\
33 [<ref-list | <object-list]";
35 struct object_entry {
36 struct pack_idx_entry idx;
37 unsigned long size; /* uncompressed size */
38 struct packed_git *in_pack; /* already in pack */
39 off_t in_pack_offset;
40 struct object_entry *delta; /* delta base object */
41 struct object_entry *delta_child; /* deltified objects who bases me */
42 struct object_entry *delta_sibling; /* other deltified objects who
43 * uses the same base as me
45 void *delta_data; /* cached delta (uncompressed) */
46 unsigned long delta_size; /* delta data size (uncompressed) */
47 unsigned long z_delta_size; /* delta data size (compressed) */
48 unsigned int hash; /* name hint hash */
49 enum object_type type;
50 enum object_type in_pack_type; /* could be delta */
51 unsigned char in_pack_header_size;
52 unsigned char preferred_base; /* we do not pack this, but is available
53 * to be used as the base object to delta
54 * objects against.
56 unsigned char no_try_delta;
60 * Objects we are going to pack are collected in objects array (dynamically
61 * expanded). nr_objects & nr_alloc controls this array. They are stored
62 * in the order we see -- typically rev-list --objects order that gives us
63 * nice "minimum seek" order.
65 static struct object_entry *objects;
66 static struct pack_idx_entry **written_list;
67 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
69 static int non_empty;
70 static int reuse_delta = 1, reuse_object = 1;
71 static int keep_unreachable, unpack_unreachable, include_tag;
72 static int local;
73 static int incremental;
74 static int ignore_packed_keep;
75 static int allow_ofs_delta;
76 static const char *base_name;
77 static int progress = 1;
78 static int window = 10;
79 static uint32_t pack_size_limit, pack_size_limit_cfg;
80 static int depth = 50;
81 static int delta_search_threads = 1;
82 static int pack_to_stdout;
83 static int num_preferred_base;
84 static struct progress *progress_state;
85 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
86 static int pack_compression_seen;
88 static unsigned long delta_cache_size = 0;
89 static unsigned long max_delta_cache_size = 0;
90 static unsigned long cache_max_small_delta_size = 1000;
92 static unsigned long window_memory_limit = 0;
95 * The object names in objects array are hashed with this hashtable,
96 * to help looking up the entry by object name.
97 * This hashtable is built after all the objects are seen.
99 static int *object_ix;
100 static int object_ix_hashsz;
103 * stats
105 static uint32_t written, written_delta;
106 static uint32_t reused, reused_delta;
109 static void *get_delta(struct object_entry *entry)
111 unsigned long size, base_size, delta_size;
112 void *buf, *base_buf, *delta_buf;
113 enum object_type type;
115 buf = read_sha1_file(entry->idx.sha1, &type, &size);
116 if (!buf)
117 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
118 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
119 if (!base_buf)
120 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
121 delta_buf = diff_delta(base_buf, base_size,
122 buf, size, &delta_size, 0);
123 if (!delta_buf || delta_size != entry->delta_size)
124 die("delta size changed");
125 free(buf);
126 free(base_buf);
127 return delta_buf;
130 static unsigned long do_compress(void **pptr, unsigned long size)
132 z_stream stream;
133 void *in, *out;
134 unsigned long maxsize;
136 memset(&stream, 0, sizeof(stream));
137 deflateInit(&stream, pack_compression_level);
138 maxsize = deflateBound(&stream, size);
140 in = *pptr;
141 out = xmalloc(maxsize);
142 *pptr = out;
144 stream.next_in = in;
145 stream.avail_in = size;
146 stream.next_out = out;
147 stream.avail_out = maxsize;
148 while (deflate(&stream, Z_FINISH) == Z_OK)
149 ; /* nothing */
150 deflateEnd(&stream);
152 free(in);
153 return stream.total_out;
157 * The per-object header is a pretty dense thing, which is
158 * - first byte: low four bits are "size", then three bits of "type",
159 * and the high bit is "size continues".
160 * - each byte afterwards: low seven bits are size continuation,
161 * with the high bit being "size continues"
163 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
165 int n = 1;
166 unsigned char c;
168 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
169 die("bad type %d", type);
171 c = (type << 4) | (size & 15);
172 size >>= 4;
173 while (size) {
174 *hdr++ = c | 0x80;
175 c = size & 0x7f;
176 size >>= 7;
177 n++;
179 *hdr = c;
180 return n;
184 * we are going to reuse the existing object data as is. make
185 * sure it is not corrupt.
187 static int check_pack_inflate(struct packed_git *p,
188 struct pack_window **w_curs,
189 off_t offset,
190 off_t len,
191 unsigned long expect)
193 z_stream stream;
194 unsigned char fakebuf[4096], *in;
195 int st;
197 memset(&stream, 0, sizeof(stream));
198 inflateInit(&stream);
199 do {
200 in = use_pack(p, w_curs, offset, &stream.avail_in);
201 stream.next_in = in;
202 stream.next_out = fakebuf;
203 stream.avail_out = sizeof(fakebuf);
204 st = inflate(&stream, Z_FINISH);
205 offset += stream.next_in - in;
206 } while (st == Z_OK || st == Z_BUF_ERROR);
207 inflateEnd(&stream);
208 return (st == Z_STREAM_END &&
209 stream.total_out == expect &&
210 stream.total_in == len) ? 0 : -1;
213 static void copy_pack_data(struct sha1file *f,
214 struct packed_git *p,
215 struct pack_window **w_curs,
216 off_t offset,
217 off_t len)
219 unsigned char *in;
220 unsigned int avail;
222 while (len) {
223 in = use_pack(p, w_curs, offset, &avail);
224 if (avail > len)
225 avail = (unsigned int)len;
226 sha1write(f, in, avail);
227 offset += avail;
228 len -= avail;
232 static unsigned long write_object(struct sha1file *f,
233 struct object_entry *entry,
234 off_t write_offset)
236 unsigned long size, limit, datalen;
237 void *buf;
238 unsigned char header[10], dheader[10];
239 unsigned hdrlen;
240 enum object_type type;
241 int usable_delta, to_reuse;
243 if (!pack_to_stdout)
244 crc32_begin(f);
246 type = entry->type;
248 /* write limit if limited packsize and not first object */
249 if (!pack_size_limit || !nr_written)
250 limit = 0;
251 else if (pack_size_limit <= write_offset)
253 * the earlier object did not fit the limit; avoid
254 * mistaking this with unlimited (i.e. limit = 0).
256 limit = 1;
257 else
258 limit = pack_size_limit - write_offset;
260 if (!entry->delta)
261 usable_delta = 0; /* no delta */
262 else if (!pack_size_limit)
263 usable_delta = 1; /* unlimited packfile */
264 else if (entry->delta->idx.offset == (off_t)-1)
265 usable_delta = 0; /* base was written to another pack */
266 else if (entry->delta->idx.offset)
267 usable_delta = 1; /* base already exists in this pack */
268 else
269 usable_delta = 0; /* base could end up in another pack */
271 if (!reuse_object)
272 to_reuse = 0; /* explicit */
273 else if (!entry->in_pack)
274 to_reuse = 0; /* can't reuse what we don't have */
275 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
276 /* check_object() decided it for us ... */
277 to_reuse = usable_delta;
278 /* ... but pack split may override that */
279 else if (type != entry->in_pack_type)
280 to_reuse = 0; /* pack has delta which is unusable */
281 else if (entry->delta)
282 to_reuse = 0; /* we want to pack afresh */
283 else
284 to_reuse = 1; /* we have it in-pack undeltified,
285 * and we do not need to deltify it.
288 if (!to_reuse) {
289 if (!usable_delta) {
290 buf = read_sha1_file(entry->idx.sha1, &type, &size);
291 if (!buf)
292 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
294 * make sure no cached delta data remains from a
295 * previous attempt before a pack split occured.
297 free(entry->delta_data);
298 entry->delta_data = NULL;
299 entry->z_delta_size = 0;
300 } else if (entry->delta_data) {
301 size = entry->delta_size;
302 buf = entry->delta_data;
303 entry->delta_data = NULL;
304 type = (allow_ofs_delta && entry->delta->idx.offset) ?
305 OBJ_OFS_DELTA : OBJ_REF_DELTA;
306 } else {
307 buf = get_delta(entry);
308 size = entry->delta_size;
309 type = (allow_ofs_delta && entry->delta->idx.offset) ?
310 OBJ_OFS_DELTA : OBJ_REF_DELTA;
313 if (entry->z_delta_size)
314 datalen = entry->z_delta_size;
315 else
316 datalen = do_compress(&buf, size);
319 * The object header is a byte of 'type' followed by zero or
320 * more bytes of length.
322 hdrlen = encode_header(type, size, header);
324 if (type == OBJ_OFS_DELTA) {
326 * Deltas with relative base contain an additional
327 * encoding of the relative offset for the delta
328 * base from this object's position in the pack.
330 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
331 unsigned pos = sizeof(dheader) - 1;
332 dheader[pos] = ofs & 127;
333 while (ofs >>= 7)
334 dheader[--pos] = 128 | (--ofs & 127);
335 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
336 free(buf);
337 return 0;
339 sha1write(f, header, hdrlen);
340 sha1write(f, dheader + pos, sizeof(dheader) - pos);
341 hdrlen += sizeof(dheader) - pos;
342 } else if (type == OBJ_REF_DELTA) {
344 * Deltas with a base reference contain
345 * an additional 20 bytes for the base sha1.
347 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
348 free(buf);
349 return 0;
351 sha1write(f, header, hdrlen);
352 sha1write(f, entry->delta->idx.sha1, 20);
353 hdrlen += 20;
354 } else {
355 if (limit && hdrlen + datalen + 20 >= limit) {
356 free(buf);
357 return 0;
359 sha1write(f, header, hdrlen);
361 sha1write(f, buf, datalen);
362 free(buf);
364 else {
365 struct packed_git *p = entry->in_pack;
366 struct pack_window *w_curs = NULL;
367 struct revindex_entry *revidx;
368 off_t offset;
370 if (entry->delta) {
371 type = (allow_ofs_delta && entry->delta->idx.offset) ?
372 OBJ_OFS_DELTA : OBJ_REF_DELTA;
373 reused_delta++;
375 hdrlen = encode_header(type, entry->size, header);
376 offset = entry->in_pack_offset;
377 revidx = find_pack_revindex(p, offset);
378 datalen = revidx[1].offset - offset;
379 if (!pack_to_stdout && p->index_version > 1 &&
380 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
381 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
382 offset += entry->in_pack_header_size;
383 datalen -= entry->in_pack_header_size;
384 if (type == OBJ_OFS_DELTA) {
385 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
386 unsigned pos = sizeof(dheader) - 1;
387 dheader[pos] = ofs & 127;
388 while (ofs >>= 7)
389 dheader[--pos] = 128 | (--ofs & 127);
390 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
391 return 0;
392 sha1write(f, header, hdrlen);
393 sha1write(f, dheader + pos, sizeof(dheader) - pos);
394 hdrlen += sizeof(dheader) - pos;
395 } else if (type == OBJ_REF_DELTA) {
396 if (limit && hdrlen + 20 + datalen + 20 >= limit)
397 return 0;
398 sha1write(f, header, hdrlen);
399 sha1write(f, entry->delta->idx.sha1, 20);
400 hdrlen += 20;
401 } else {
402 if (limit && hdrlen + datalen + 20 >= limit)
403 return 0;
404 sha1write(f, header, hdrlen);
407 if (!pack_to_stdout && p->index_version == 1 &&
408 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
409 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
410 copy_pack_data(f, p, &w_curs, offset, datalen);
411 unuse_pack(&w_curs);
412 reused++;
414 if (usable_delta)
415 written_delta++;
416 written++;
417 if (!pack_to_stdout)
418 entry->idx.crc32 = crc32_end(f);
419 return hdrlen + datalen;
422 static int write_one(struct sha1file *f,
423 struct object_entry *e,
424 off_t *offset)
426 unsigned long size;
428 /* offset is non zero if object is written already. */
429 if (e->idx.offset || e->preferred_base)
430 return 1;
432 /* if we are deltified, write out base object first. */
433 if (e->delta && !write_one(f, e->delta, offset))
434 return 0;
436 e->idx.offset = *offset;
437 size = write_object(f, e, *offset);
438 if (!size) {
439 e->idx.offset = 0;
440 return 0;
442 written_list[nr_written++] = &e->idx;
444 /* make sure off_t is sufficiently large not to wrap */
445 if (*offset > *offset + size)
446 die("pack too large for current definition of off_t");
447 *offset += size;
448 return 1;
451 /* forward declaration for write_pack_file */
452 static int adjust_perm(const char *path, mode_t mode);
454 static void write_pack_file(void)
456 uint32_t i = 0, j;
457 struct sha1file *f;
458 off_t offset;
459 struct pack_header hdr;
460 uint32_t nr_remaining = nr_result;
461 time_t last_mtime = 0;
463 if (progress > pack_to_stdout)
464 progress_state = start_progress("Writing objects", nr_result);
465 written_list = xmalloc(nr_objects * sizeof(*written_list));
467 do {
468 unsigned char sha1[20];
469 char *pack_tmp_name = NULL;
471 if (pack_to_stdout) {
472 f = sha1fd_throughput(1, "<stdout>", progress_state);
473 } else {
474 char tmpname[PATH_MAX];
475 int fd;
476 snprintf(tmpname, sizeof(tmpname),
477 "%s/pack/tmp_pack_XXXXXX", get_object_directory());
478 fd = xmkstemp(tmpname);
479 pack_tmp_name = xstrdup(tmpname);
480 f = sha1fd(fd, pack_tmp_name);
483 hdr.hdr_signature = htonl(PACK_SIGNATURE);
484 hdr.hdr_version = htonl(PACK_VERSION);
485 hdr.hdr_entries = htonl(nr_remaining);
486 sha1write(f, &hdr, sizeof(hdr));
487 offset = sizeof(hdr);
488 nr_written = 0;
489 for (; i < nr_objects; i++) {
490 if (!write_one(f, objects + i, &offset))
491 break;
492 display_progress(progress_state, written);
496 * Did we write the wrong # entries in the header?
497 * If so, rewrite it like in fast-import
499 if (pack_to_stdout) {
500 sha1close(f, sha1, CSUM_CLOSE);
501 } else if (nr_written == nr_remaining) {
502 sha1close(f, sha1, CSUM_FSYNC);
503 } else {
504 int fd = sha1close(f, sha1, 0);
505 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
506 nr_written, sha1, offset);
507 close(fd);
510 if (!pack_to_stdout) {
511 mode_t mode = umask(0);
512 struct stat st;
513 char *idx_tmp_name, tmpname[PATH_MAX];
515 umask(mode);
516 mode = 0444 & ~mode;
518 idx_tmp_name = write_idx_file(NULL, written_list,
519 nr_written, sha1);
521 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
522 base_name, sha1_to_hex(sha1));
523 free_pack_by_name(tmpname);
524 if (adjust_perm(pack_tmp_name, mode))
525 die("unable to make temporary pack file readable: %s",
526 strerror(errno));
527 if (rename(pack_tmp_name, tmpname))
528 die("unable to rename temporary pack file: %s",
529 strerror(errno));
532 * Packs are runtime accessed in their mtime
533 * order since newer packs are more likely to contain
534 * younger objects. So if we are creating multiple
535 * packs then we should modify the mtime of later ones
536 * to preserve this property.
538 if (stat(tmpname, &st) < 0) {
539 warning("failed to stat %s: %s",
540 tmpname, strerror(errno));
541 } else if (!last_mtime) {
542 last_mtime = st.st_mtime;
543 } else {
544 struct utimbuf utb;
545 utb.actime = st.st_atime;
546 utb.modtime = --last_mtime;
547 if (utime(tmpname, &utb) < 0)
548 warning("failed utime() on %s: %s",
549 tmpname, strerror(errno));
552 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
553 base_name, sha1_to_hex(sha1));
554 if (adjust_perm(idx_tmp_name, mode))
555 die("unable to make temporary index file readable: %s",
556 strerror(errno));
557 if (rename(idx_tmp_name, tmpname))
558 die("unable to rename temporary index file: %s",
559 strerror(errno));
561 free(idx_tmp_name);
562 free(pack_tmp_name);
563 puts(sha1_to_hex(sha1));
566 /* mark written objects as written to previous pack */
567 for (j = 0; j < nr_written; j++) {
568 written_list[j]->offset = (off_t)-1;
570 nr_remaining -= nr_written;
571 } while (nr_remaining && i < nr_objects);
573 free(written_list);
574 stop_progress(&progress_state);
575 if (written != nr_result)
576 die("wrote %"PRIu32" objects while expecting %"PRIu32,
577 written, nr_result);
579 * We have scanned through [0 ... i). Since we have written
580 * the correct number of objects, the remaining [i ... nr_objects)
581 * items must be either already written (due to out-of-order delta base)
582 * or a preferred base. Count those which are neither and complain if any.
584 for (j = 0; i < nr_objects; i++) {
585 struct object_entry *e = objects + i;
586 j += !e->idx.offset && !e->preferred_base;
588 if (j)
589 die("wrote %"PRIu32" objects as expected but %"PRIu32
590 " unwritten", written, j);
593 static int locate_object_entry_hash(const unsigned char *sha1)
595 int i;
596 unsigned int ui;
597 memcpy(&ui, sha1, sizeof(unsigned int));
598 i = ui % object_ix_hashsz;
599 while (0 < object_ix[i]) {
600 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
601 return i;
602 if (++i == object_ix_hashsz)
603 i = 0;
605 return -1 - i;
608 static struct object_entry *locate_object_entry(const unsigned char *sha1)
610 int i;
612 if (!object_ix_hashsz)
613 return NULL;
615 i = locate_object_entry_hash(sha1);
616 if (0 <= i)
617 return &objects[object_ix[i]-1];
618 return NULL;
621 static void rehash_objects(void)
623 uint32_t i;
624 struct object_entry *oe;
626 object_ix_hashsz = nr_objects * 3;
627 if (object_ix_hashsz < 1024)
628 object_ix_hashsz = 1024;
629 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
630 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
631 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
632 int ix = locate_object_entry_hash(oe->idx.sha1);
633 if (0 <= ix)
634 continue;
635 ix = -1 - ix;
636 object_ix[ix] = i + 1;
640 static unsigned name_hash(const char *name)
642 unsigned char c;
643 unsigned hash = 0;
645 if (!name)
646 return 0;
649 * This effectively just creates a sortable number from the
650 * last sixteen non-whitespace characters. Last characters
651 * count "most", so things that end in ".c" sort together.
653 while ((c = *name++) != 0) {
654 if (isspace(c))
655 continue;
656 hash = (hash >> 2) + (c << 24);
658 return hash;
661 static void setup_delta_attr_check(struct git_attr_check *check)
663 static struct git_attr *attr_delta;
665 if (!attr_delta)
666 attr_delta = git_attr("delta", 5);
668 check[0].attr = attr_delta;
671 static int no_try_delta(const char *path)
673 struct git_attr_check check[1];
675 setup_delta_attr_check(check);
676 if (git_checkattr(path, ARRAY_SIZE(check), check))
677 return 0;
678 if (ATTR_FALSE(check->value))
679 return 1;
680 return 0;
683 static int add_object_entry(const unsigned char *sha1, enum object_type type,
684 const char *name, int exclude)
686 struct object_entry *entry;
687 struct packed_git *p, *found_pack = NULL;
688 off_t found_offset = 0;
689 int ix;
690 unsigned hash = name_hash(name);
692 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
693 if (ix >= 0) {
694 if (exclude) {
695 entry = objects + object_ix[ix] - 1;
696 if (!entry->preferred_base)
697 nr_result--;
698 entry->preferred_base = 1;
700 return 0;
703 if (!exclude && local && has_loose_object_nonlocal(sha1))
704 return 0;
706 for (p = packed_git; p; p = p->next) {
707 off_t offset = find_pack_entry_one(sha1, p);
708 if (offset) {
709 if (!found_pack) {
710 found_offset = offset;
711 found_pack = p;
713 if (exclude)
714 break;
715 if (incremental)
716 return 0;
717 if (local && !p->pack_local)
718 return 0;
719 if (ignore_packed_keep && p->pack_local && p->pack_keep)
720 return 0;
724 if (nr_objects >= nr_alloc) {
725 nr_alloc = (nr_alloc + 1024) * 3 / 2;
726 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
729 entry = objects + nr_objects++;
730 memset(entry, 0, sizeof(*entry));
731 hashcpy(entry->idx.sha1, sha1);
732 entry->hash = hash;
733 if (type)
734 entry->type = type;
735 if (exclude)
736 entry->preferred_base = 1;
737 else
738 nr_result++;
739 if (found_pack) {
740 entry->in_pack = found_pack;
741 entry->in_pack_offset = found_offset;
744 if (object_ix_hashsz * 3 <= nr_objects * 4)
745 rehash_objects();
746 else
747 object_ix[-1 - ix] = nr_objects;
749 display_progress(progress_state, nr_objects);
751 if (name && no_try_delta(name))
752 entry->no_try_delta = 1;
754 return 1;
757 struct pbase_tree_cache {
758 unsigned char sha1[20];
759 int ref;
760 int temporary;
761 void *tree_data;
762 unsigned long tree_size;
765 static struct pbase_tree_cache *(pbase_tree_cache[256]);
766 static int pbase_tree_cache_ix(const unsigned char *sha1)
768 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
770 static int pbase_tree_cache_ix_incr(int ix)
772 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
775 static struct pbase_tree {
776 struct pbase_tree *next;
777 /* This is a phony "cache" entry; we are not
778 * going to evict it nor find it through _get()
779 * mechanism -- this is for the toplevel node that
780 * would almost always change with any commit.
782 struct pbase_tree_cache pcache;
783 } *pbase_tree;
785 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
787 struct pbase_tree_cache *ent, *nent;
788 void *data;
789 unsigned long size;
790 enum object_type type;
791 int neigh;
792 int my_ix = pbase_tree_cache_ix(sha1);
793 int available_ix = -1;
795 /* pbase-tree-cache acts as a limited hashtable.
796 * your object will be found at your index or within a few
797 * slots after that slot if it is cached.
799 for (neigh = 0; neigh < 8; neigh++) {
800 ent = pbase_tree_cache[my_ix];
801 if (ent && !hashcmp(ent->sha1, sha1)) {
802 ent->ref++;
803 return ent;
805 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
806 ((0 <= available_ix) &&
807 (!ent && pbase_tree_cache[available_ix])))
808 available_ix = my_ix;
809 if (!ent)
810 break;
811 my_ix = pbase_tree_cache_ix_incr(my_ix);
814 /* Did not find one. Either we got a bogus request or
815 * we need to read and perhaps cache.
817 data = read_sha1_file(sha1, &type, &size);
818 if (!data)
819 return NULL;
820 if (type != OBJ_TREE) {
821 free(data);
822 return NULL;
825 /* We need to either cache or return a throwaway copy */
827 if (available_ix < 0)
828 ent = NULL;
829 else {
830 ent = pbase_tree_cache[available_ix];
831 my_ix = available_ix;
834 if (!ent) {
835 nent = xmalloc(sizeof(*nent));
836 nent->temporary = (available_ix < 0);
838 else {
839 /* evict and reuse */
840 free(ent->tree_data);
841 nent = ent;
843 hashcpy(nent->sha1, sha1);
844 nent->tree_data = data;
845 nent->tree_size = size;
846 nent->ref = 1;
847 if (!nent->temporary)
848 pbase_tree_cache[my_ix] = nent;
849 return nent;
852 static void pbase_tree_put(struct pbase_tree_cache *cache)
854 if (!cache->temporary) {
855 cache->ref--;
856 return;
858 free(cache->tree_data);
859 free(cache);
862 static int name_cmp_len(const char *name)
864 int i;
865 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
867 return i;
870 static void add_pbase_object(struct tree_desc *tree,
871 const char *name,
872 int cmplen,
873 const char *fullname)
875 struct name_entry entry;
876 int cmp;
878 while (tree_entry(tree,&entry)) {
879 if (S_ISGITLINK(entry.mode))
880 continue;
881 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
882 memcmp(name, entry.path, cmplen);
883 if (cmp > 0)
884 continue;
885 if (cmp < 0)
886 return;
887 if (name[cmplen] != '/') {
888 add_object_entry(entry.sha1,
889 object_type(entry.mode),
890 fullname, 1);
891 return;
893 if (S_ISDIR(entry.mode)) {
894 struct tree_desc sub;
895 struct pbase_tree_cache *tree;
896 const char *down = name+cmplen+1;
897 int downlen = name_cmp_len(down);
899 tree = pbase_tree_get(entry.sha1);
900 if (!tree)
901 return;
902 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
904 add_pbase_object(&sub, down, downlen, fullname);
905 pbase_tree_put(tree);
910 static unsigned *done_pbase_paths;
911 static int done_pbase_paths_num;
912 static int done_pbase_paths_alloc;
913 static int done_pbase_path_pos(unsigned hash)
915 int lo = 0;
916 int hi = done_pbase_paths_num;
917 while (lo < hi) {
918 int mi = (hi + lo) / 2;
919 if (done_pbase_paths[mi] == hash)
920 return mi;
921 if (done_pbase_paths[mi] < hash)
922 hi = mi;
923 else
924 lo = mi + 1;
926 return -lo-1;
929 static int check_pbase_path(unsigned hash)
931 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
932 if (0 <= pos)
933 return 1;
934 pos = -pos - 1;
935 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
936 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
937 done_pbase_paths = xrealloc(done_pbase_paths,
938 done_pbase_paths_alloc *
939 sizeof(unsigned));
941 done_pbase_paths_num++;
942 if (pos < done_pbase_paths_num)
943 memmove(done_pbase_paths + pos + 1,
944 done_pbase_paths + pos,
945 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
946 done_pbase_paths[pos] = hash;
947 return 0;
950 static void add_preferred_base_object(const char *name)
952 struct pbase_tree *it;
953 int cmplen;
954 unsigned hash = name_hash(name);
956 if (!num_preferred_base || check_pbase_path(hash))
957 return;
959 cmplen = name_cmp_len(name);
960 for (it = pbase_tree; it; it = it->next) {
961 if (cmplen == 0) {
962 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
964 else {
965 struct tree_desc tree;
966 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
967 add_pbase_object(&tree, name, cmplen, name);
972 static void add_preferred_base(unsigned char *sha1)
974 struct pbase_tree *it;
975 void *data;
976 unsigned long size;
977 unsigned char tree_sha1[20];
979 if (window <= num_preferred_base++)
980 return;
982 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
983 if (!data)
984 return;
986 for (it = pbase_tree; it; it = it->next) {
987 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
988 free(data);
989 return;
993 it = xcalloc(1, sizeof(*it));
994 it->next = pbase_tree;
995 pbase_tree = it;
997 hashcpy(it->pcache.sha1, tree_sha1);
998 it->pcache.tree_data = data;
999 it->pcache.tree_size = size;
1002 static void check_object(struct object_entry *entry)
1004 if (entry->in_pack) {
1005 struct packed_git *p = entry->in_pack;
1006 struct pack_window *w_curs = NULL;
1007 const unsigned char *base_ref = NULL;
1008 struct object_entry *base_entry;
1009 unsigned long used, used_0;
1010 unsigned int avail;
1011 off_t ofs;
1012 unsigned char *buf, c;
1014 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1017 * We want in_pack_type even if we do not reuse delta
1018 * since non-delta representations could still be reused.
1020 used = unpack_object_header_gently(buf, avail,
1021 &entry->in_pack_type,
1022 &entry->size);
1025 * Determine if this is a delta and if so whether we can
1026 * reuse it or not. Otherwise let's find out as cheaply as
1027 * possible what the actual type and size for this object is.
1029 switch (entry->in_pack_type) {
1030 default:
1031 /* Not a delta hence we've already got all we need. */
1032 entry->type = entry->in_pack_type;
1033 entry->in_pack_header_size = used;
1034 unuse_pack(&w_curs);
1035 return;
1036 case OBJ_REF_DELTA:
1037 if (reuse_delta && !entry->preferred_base)
1038 base_ref = use_pack(p, &w_curs,
1039 entry->in_pack_offset + used, NULL);
1040 entry->in_pack_header_size = used + 20;
1041 break;
1042 case OBJ_OFS_DELTA:
1043 buf = use_pack(p, &w_curs,
1044 entry->in_pack_offset + used, NULL);
1045 used_0 = 0;
1046 c = buf[used_0++];
1047 ofs = c & 127;
1048 while (c & 128) {
1049 ofs += 1;
1050 if (!ofs || MSB(ofs, 7))
1051 die("delta base offset overflow in pack for %s",
1052 sha1_to_hex(entry->idx.sha1));
1053 c = buf[used_0++];
1054 ofs = (ofs << 7) + (c & 127);
1056 if (ofs >= entry->in_pack_offset)
1057 die("delta base offset out of bound for %s",
1058 sha1_to_hex(entry->idx.sha1));
1059 ofs = entry->in_pack_offset - ofs;
1060 if (reuse_delta && !entry->preferred_base) {
1061 struct revindex_entry *revidx;
1062 revidx = find_pack_revindex(p, ofs);
1063 base_ref = nth_packed_object_sha1(p, revidx->nr);
1065 entry->in_pack_header_size = used + used_0;
1066 break;
1069 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1071 * If base_ref was set above that means we wish to
1072 * reuse delta data, and we even found that base
1073 * in the list of objects we want to pack. Goodie!
1075 * Depth value does not matter - find_deltas() will
1076 * never consider reused delta as the base object to
1077 * deltify other objects against, in order to avoid
1078 * circular deltas.
1080 entry->type = entry->in_pack_type;
1081 entry->delta = base_entry;
1082 entry->delta_sibling = base_entry->delta_child;
1083 base_entry->delta_child = entry;
1084 unuse_pack(&w_curs);
1085 return;
1088 if (entry->type) {
1090 * This must be a delta and we already know what the
1091 * final object type is. Let's extract the actual
1092 * object size from the delta header.
1094 entry->size = get_size_from_delta(p, &w_curs,
1095 entry->in_pack_offset + entry->in_pack_header_size);
1096 unuse_pack(&w_curs);
1097 return;
1101 * No choice but to fall back to the recursive delta walk
1102 * with sha1_object_info() to find about the object type
1103 * at this point...
1105 unuse_pack(&w_curs);
1108 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1110 * The error condition is checked in prepare_pack(). This is
1111 * to permit a missing preferred base object to be ignored
1112 * as a preferred base. Doing so can result in a larger
1113 * pack file, but the transfer will still take place.
1117 static int pack_offset_sort(const void *_a, const void *_b)
1119 const struct object_entry *a = *(struct object_entry **)_a;
1120 const struct object_entry *b = *(struct object_entry **)_b;
1122 /* avoid filesystem trashing with loose objects */
1123 if (!a->in_pack && !b->in_pack)
1124 return hashcmp(a->idx.sha1, b->idx.sha1);
1126 if (a->in_pack < b->in_pack)
1127 return -1;
1128 if (a->in_pack > b->in_pack)
1129 return 1;
1130 return a->in_pack_offset < b->in_pack_offset ? -1 :
1131 (a->in_pack_offset > b->in_pack_offset);
1134 static void get_object_details(void)
1136 uint32_t i;
1137 struct object_entry **sorted_by_offset;
1139 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1140 for (i = 0; i < nr_objects; i++)
1141 sorted_by_offset[i] = objects + i;
1142 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1144 for (i = 0; i < nr_objects; i++)
1145 check_object(sorted_by_offset[i]);
1147 free(sorted_by_offset);
1151 * We search for deltas in a list sorted by type, by filename hash, and then
1152 * by size, so that we see progressively smaller and smaller files.
1153 * That's because we prefer deltas to be from the bigger file
1154 * to the smaller -- deletes are potentially cheaper, but perhaps
1155 * more importantly, the bigger file is likely the more recent
1156 * one. The deepest deltas are therefore the oldest objects which are
1157 * less susceptible to be accessed often.
1159 static int type_size_sort(const void *_a, const void *_b)
1161 const struct object_entry *a = *(struct object_entry **)_a;
1162 const struct object_entry *b = *(struct object_entry **)_b;
1164 if (a->type > b->type)
1165 return -1;
1166 if (a->type < b->type)
1167 return 1;
1168 if (a->hash > b->hash)
1169 return -1;
1170 if (a->hash < b->hash)
1171 return 1;
1172 if (a->preferred_base > b->preferred_base)
1173 return -1;
1174 if (a->preferred_base < b->preferred_base)
1175 return 1;
1176 if (a->size > b->size)
1177 return -1;
1178 if (a->size < b->size)
1179 return 1;
1180 return a < b ? -1 : (a > b); /* newest first */
1183 struct unpacked {
1184 struct object_entry *entry;
1185 void *data;
1186 struct delta_index *index;
1187 unsigned depth;
1190 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1191 unsigned long delta_size)
1193 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1194 return 0;
1196 if (delta_size < cache_max_small_delta_size)
1197 return 1;
1199 /* cache delta, if objects are large enough compared to delta size */
1200 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1201 return 1;
1203 return 0;
1206 #ifdef THREADED_DELTA_SEARCH
1208 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1209 #define read_lock() pthread_mutex_lock(&read_mutex)
1210 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1212 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1213 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1214 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1216 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1217 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1218 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1220 #else
1222 #define read_lock() (void)0
1223 #define read_unlock() (void)0
1224 #define cache_lock() (void)0
1225 #define cache_unlock() (void)0
1226 #define progress_lock() (void)0
1227 #define progress_unlock() (void)0
1229 #endif
1231 static int try_delta(struct unpacked *trg, struct unpacked *src,
1232 unsigned max_depth, unsigned long *mem_usage)
1234 struct object_entry *trg_entry = trg->entry;
1235 struct object_entry *src_entry = src->entry;
1236 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1237 unsigned ref_depth;
1238 enum object_type type;
1239 void *delta_buf;
1241 /* Don't bother doing diffs between different types */
1242 if (trg_entry->type != src_entry->type)
1243 return -1;
1246 * We do not bother to try a delta that we discarded
1247 * on an earlier try, but only when reusing delta data.
1249 if (reuse_delta && trg_entry->in_pack &&
1250 trg_entry->in_pack == src_entry->in_pack &&
1251 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1252 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1253 return 0;
1255 /* Let's not bust the allowed depth. */
1256 if (src->depth >= max_depth)
1257 return 0;
1259 /* Now some size filtering heuristics. */
1260 trg_size = trg_entry->size;
1261 if (!trg_entry->delta) {
1262 max_size = trg_size/2 - 20;
1263 ref_depth = 1;
1264 } else {
1265 max_size = trg_entry->delta_size;
1266 ref_depth = trg->depth;
1268 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1269 (max_depth - ref_depth + 1);
1270 if (max_size == 0)
1271 return 0;
1272 src_size = src_entry->size;
1273 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1274 if (sizediff >= max_size)
1275 return 0;
1276 if (trg_size < src_size / 32)
1277 return 0;
1279 /* Load data if not already done */
1280 if (!trg->data) {
1281 read_lock();
1282 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1283 read_unlock();
1284 if (!trg->data)
1285 die("object %s cannot be read",
1286 sha1_to_hex(trg_entry->idx.sha1));
1287 if (sz != trg_size)
1288 die("object %s inconsistent object length (%lu vs %lu)",
1289 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1290 *mem_usage += sz;
1292 if (!src->data) {
1293 read_lock();
1294 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1295 read_unlock();
1296 if (!src->data)
1297 die("object %s cannot be read",
1298 sha1_to_hex(src_entry->idx.sha1));
1299 if (sz != src_size)
1300 die("object %s inconsistent object length (%lu vs %lu)",
1301 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1302 *mem_usage += sz;
1304 if (!src->index) {
1305 src->index = create_delta_index(src->data, src_size);
1306 if (!src->index) {
1307 static int warned = 0;
1308 if (!warned++)
1309 warning("suboptimal pack - out of memory");
1310 return 0;
1312 *mem_usage += sizeof_delta_index(src->index);
1315 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1316 if (!delta_buf)
1317 return 0;
1319 if (trg_entry->delta) {
1320 /* Prefer only shallower same-sized deltas. */
1321 if (delta_size == trg_entry->delta_size &&
1322 src->depth + 1 >= trg->depth) {
1323 free(delta_buf);
1324 return 0;
1329 * Handle memory allocation outside of the cache
1330 * accounting lock. Compiler will optimize the strangeness
1331 * away when THREADED_DELTA_SEARCH is not defined.
1333 free(trg_entry->delta_data);
1334 cache_lock();
1335 if (trg_entry->delta_data) {
1336 delta_cache_size -= trg_entry->delta_size;
1337 trg_entry->delta_data = NULL;
1339 if (delta_cacheable(src_size, trg_size, delta_size)) {
1340 delta_cache_size += delta_size;
1341 cache_unlock();
1342 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1343 } else {
1344 cache_unlock();
1345 free(delta_buf);
1348 trg_entry->delta = src_entry;
1349 trg_entry->delta_size = delta_size;
1350 trg->depth = src->depth + 1;
1352 return 1;
1355 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1357 struct object_entry *child = me->delta_child;
1358 unsigned int m = n;
1359 while (child) {
1360 unsigned int c = check_delta_limit(child, n + 1);
1361 if (m < c)
1362 m = c;
1363 child = child->delta_sibling;
1365 return m;
1368 static unsigned long free_unpacked(struct unpacked *n)
1370 unsigned long freed_mem = sizeof_delta_index(n->index);
1371 free_delta_index(n->index);
1372 n->index = NULL;
1373 if (n->data) {
1374 freed_mem += n->entry->size;
1375 free(n->data);
1376 n->data = NULL;
1378 n->entry = NULL;
1379 n->depth = 0;
1380 return freed_mem;
1383 static void find_deltas(struct object_entry **list, unsigned *list_size,
1384 int window, int depth, unsigned *processed)
1386 uint32_t i, idx = 0, count = 0;
1387 unsigned int array_size = window * sizeof(struct unpacked);
1388 struct unpacked *array;
1389 unsigned long mem_usage = 0;
1391 array = xmalloc(array_size);
1392 memset(array, 0, array_size);
1394 for (;;) {
1395 struct object_entry *entry;
1396 struct unpacked *n = array + idx;
1397 int j, max_depth, best_base = -1;
1399 progress_lock();
1400 if (!*list_size) {
1401 progress_unlock();
1402 break;
1404 entry = *list++;
1405 (*list_size)--;
1406 if (!entry->preferred_base) {
1407 (*processed)++;
1408 display_progress(progress_state, *processed);
1410 progress_unlock();
1412 mem_usage -= free_unpacked(n);
1413 n->entry = entry;
1415 while (window_memory_limit &&
1416 mem_usage > window_memory_limit &&
1417 count > 1) {
1418 uint32_t tail = (idx + window - count) % window;
1419 mem_usage -= free_unpacked(array + tail);
1420 count--;
1423 /* We do not compute delta to *create* objects we are not
1424 * going to pack.
1426 if (entry->preferred_base)
1427 goto next;
1430 * If the current object is at pack edge, take the depth the
1431 * objects that depend on the current object into account
1432 * otherwise they would become too deep.
1434 max_depth = depth;
1435 if (entry->delta_child) {
1436 max_depth -= check_delta_limit(entry, 0);
1437 if (max_depth <= 0)
1438 goto next;
1441 j = window;
1442 while (--j > 0) {
1443 int ret;
1444 uint32_t other_idx = idx + j;
1445 struct unpacked *m;
1446 if (other_idx >= window)
1447 other_idx -= window;
1448 m = array + other_idx;
1449 if (!m->entry)
1450 break;
1451 ret = try_delta(n, m, max_depth, &mem_usage);
1452 if (ret < 0)
1453 break;
1454 else if (ret > 0)
1455 best_base = other_idx;
1459 * If we decided to cache the delta data, then it is best
1460 * to compress it right away. First because we have to do
1461 * it anyway, and doing it here while we're threaded will
1462 * save a lot of time in the non threaded write phase,
1463 * as well as allow for caching more deltas within
1464 * the same cache size limit.
1465 * ...
1466 * But only if not writing to stdout, since in that case
1467 * the network is most likely throttling writes anyway,
1468 * and therefore it is best to go to the write phase ASAP
1469 * instead, as we can afford spending more time compressing
1470 * between writes at that moment.
1472 if (entry->delta_data && !pack_to_stdout) {
1473 entry->z_delta_size = do_compress(&entry->delta_data,
1474 entry->delta_size);
1475 cache_lock();
1476 delta_cache_size -= entry->delta_size;
1477 delta_cache_size += entry->z_delta_size;
1478 cache_unlock();
1481 /* if we made n a delta, and if n is already at max
1482 * depth, leaving it in the window is pointless. we
1483 * should evict it first.
1485 if (entry->delta && max_depth <= n->depth)
1486 continue;
1489 * Move the best delta base up in the window, after the
1490 * currently deltified object, to keep it longer. It will
1491 * be the first base object to be attempted next.
1493 if (entry->delta) {
1494 struct unpacked swap = array[best_base];
1495 int dist = (window + idx - best_base) % window;
1496 int dst = best_base;
1497 while (dist--) {
1498 int src = (dst + 1) % window;
1499 array[dst] = array[src];
1500 dst = src;
1502 array[dst] = swap;
1505 next:
1506 idx++;
1507 if (count + 1 < window)
1508 count++;
1509 if (idx >= window)
1510 idx = 0;
1513 for (i = 0; i < window; ++i) {
1514 free_delta_index(array[i].index);
1515 free(array[i].data);
1517 free(array);
1520 #ifdef THREADED_DELTA_SEARCH
1523 * The main thread waits on the condition that (at least) one of the workers
1524 * has stopped working (which is indicated in the .working member of
1525 * struct thread_params).
1526 * When a work thread has completed its work, it sets .working to 0 and
1527 * signals the main thread and waits on the condition that .data_ready
1528 * becomes 1.
1531 struct thread_params {
1532 pthread_t thread;
1533 struct object_entry **list;
1534 unsigned list_size;
1535 unsigned remaining;
1536 int window;
1537 int depth;
1538 int working;
1539 int data_ready;
1540 pthread_mutex_t mutex;
1541 pthread_cond_t cond;
1542 unsigned *processed;
1545 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1547 static void *threaded_find_deltas(void *arg)
1549 struct thread_params *me = arg;
1551 while (me->remaining) {
1552 find_deltas(me->list, &me->remaining,
1553 me->window, me->depth, me->processed);
1555 progress_lock();
1556 me->working = 0;
1557 pthread_cond_signal(&progress_cond);
1558 progress_unlock();
1561 * We must not set ->data_ready before we wait on the
1562 * condition because the main thread may have set it to 1
1563 * before we get here. In order to be sure that new
1564 * work is available if we see 1 in ->data_ready, it
1565 * was initialized to 0 before this thread was spawned
1566 * and we reset it to 0 right away.
1568 pthread_mutex_lock(&me->mutex);
1569 while (!me->data_ready)
1570 pthread_cond_wait(&me->cond, &me->mutex);
1571 me->data_ready = 0;
1572 pthread_mutex_unlock(&me->mutex);
1574 /* leave ->working 1 so that this doesn't get more work assigned */
1575 return NULL;
1578 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1579 int window, int depth, unsigned *processed)
1581 struct thread_params p[delta_search_threads];
1582 int i, ret, active_threads = 0;
1584 if (delta_search_threads <= 1) {
1585 find_deltas(list, &list_size, window, depth, processed);
1586 return;
1589 /* Partition the work amongst work threads. */
1590 for (i = 0; i < delta_search_threads; i++) {
1591 unsigned sub_size = list_size / (delta_search_threads - i);
1593 p[i].window = window;
1594 p[i].depth = depth;
1595 p[i].processed = processed;
1596 p[i].working = 1;
1597 p[i].data_ready = 0;
1599 /* try to split chunks on "path" boundaries */
1600 while (sub_size && sub_size < list_size &&
1601 list[sub_size]->hash &&
1602 list[sub_size]->hash == list[sub_size-1]->hash)
1603 sub_size++;
1605 p[i].list = list;
1606 p[i].list_size = sub_size;
1607 p[i].remaining = sub_size;
1609 list += sub_size;
1610 list_size -= sub_size;
1613 /* Start work threads. */
1614 for (i = 0; i < delta_search_threads; i++) {
1615 if (!p[i].list_size)
1616 continue;
1617 pthread_mutex_init(&p[i].mutex, NULL);
1618 pthread_cond_init(&p[i].cond, NULL);
1619 ret = pthread_create(&p[i].thread, NULL,
1620 threaded_find_deltas, &p[i]);
1621 if (ret)
1622 die("unable to create thread: %s", strerror(ret));
1623 active_threads++;
1627 * Now let's wait for work completion. Each time a thread is done
1628 * with its work, we steal half of the remaining work from the
1629 * thread with the largest number of unprocessed objects and give
1630 * it to that newly idle thread. This ensure good load balancing
1631 * until the remaining object list segments are simply too short
1632 * to be worth splitting anymore.
1634 while (active_threads) {
1635 struct thread_params *target = NULL;
1636 struct thread_params *victim = NULL;
1637 unsigned sub_size = 0;
1639 progress_lock();
1640 for (;;) {
1641 for (i = 0; !target && i < delta_search_threads; i++)
1642 if (!p[i].working)
1643 target = &p[i];
1644 if (target)
1645 break;
1646 pthread_cond_wait(&progress_cond, &progress_mutex);
1649 for (i = 0; i < delta_search_threads; i++)
1650 if (p[i].remaining > 2*window &&
1651 (!victim || victim->remaining < p[i].remaining))
1652 victim = &p[i];
1653 if (victim) {
1654 sub_size = victim->remaining / 2;
1655 list = victim->list + victim->list_size - sub_size;
1656 while (sub_size && list[0]->hash &&
1657 list[0]->hash == list[-1]->hash) {
1658 list++;
1659 sub_size--;
1661 if (!sub_size) {
1663 * It is possible for some "paths" to have
1664 * so many objects that no hash boundary
1665 * might be found. Let's just steal the
1666 * exact half in that case.
1668 sub_size = victim->remaining / 2;
1669 list -= sub_size;
1671 target->list = list;
1672 victim->list_size -= sub_size;
1673 victim->remaining -= sub_size;
1675 target->list_size = sub_size;
1676 target->remaining = sub_size;
1677 target->working = 1;
1678 progress_unlock();
1680 pthread_mutex_lock(&target->mutex);
1681 target->data_ready = 1;
1682 pthread_cond_signal(&target->cond);
1683 pthread_mutex_unlock(&target->mutex);
1685 if (!sub_size) {
1686 pthread_join(target->thread, NULL);
1687 pthread_cond_destroy(&target->cond);
1688 pthread_mutex_destroy(&target->mutex);
1689 active_threads--;
1694 #else
1695 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1696 #endif
1698 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1700 unsigned char peeled[20];
1702 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1703 !peel_ref(path, peeled) && /* peelable? */
1704 !is_null_sha1(peeled) && /* annotated tag? */
1705 locate_object_entry(peeled)) /* object packed? */
1706 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1707 return 0;
1710 static void prepare_pack(int window, int depth)
1712 struct object_entry **delta_list;
1713 uint32_t i, nr_deltas;
1714 unsigned n;
1716 get_object_details();
1718 if (!nr_objects || !window || !depth)
1719 return;
1721 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1722 nr_deltas = n = 0;
1724 for (i = 0; i < nr_objects; i++) {
1725 struct object_entry *entry = objects + i;
1727 if (entry->delta)
1728 /* This happens if we decided to reuse existing
1729 * delta from a pack. "reuse_delta &&" is implied.
1731 continue;
1733 if (entry->size < 50)
1734 continue;
1736 if (entry->no_try_delta)
1737 continue;
1739 if (!entry->preferred_base) {
1740 nr_deltas++;
1741 if (entry->type < 0)
1742 die("unable to get type of object %s",
1743 sha1_to_hex(entry->idx.sha1));
1746 delta_list[n++] = entry;
1749 if (nr_deltas && n > 1) {
1750 unsigned nr_done = 0;
1751 if (progress)
1752 progress_state = start_progress("Compressing objects",
1753 nr_deltas);
1754 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1755 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1756 stop_progress(&progress_state);
1757 if (nr_done != nr_deltas)
1758 die("inconsistency with delta count");
1760 free(delta_list);
1763 static int git_pack_config(const char *k, const char *v, void *cb)
1765 if(!strcmp(k, "pack.window")) {
1766 window = git_config_int(k, v);
1767 return 0;
1769 if (!strcmp(k, "pack.windowmemory")) {
1770 window_memory_limit = git_config_ulong(k, v);
1771 return 0;
1773 if (!strcmp(k, "pack.depth")) {
1774 depth = git_config_int(k, v);
1775 return 0;
1777 if (!strcmp(k, "pack.compression")) {
1778 int level = git_config_int(k, v);
1779 if (level == -1)
1780 level = Z_DEFAULT_COMPRESSION;
1781 else if (level < 0 || level > Z_BEST_COMPRESSION)
1782 die("bad pack compression level %d", level);
1783 pack_compression_level = level;
1784 pack_compression_seen = 1;
1785 return 0;
1787 if (!strcmp(k, "pack.deltacachesize")) {
1788 max_delta_cache_size = git_config_int(k, v);
1789 return 0;
1791 if (!strcmp(k, "pack.deltacachelimit")) {
1792 cache_max_small_delta_size = git_config_int(k, v);
1793 return 0;
1795 if (!strcmp(k, "pack.threads")) {
1796 delta_search_threads = git_config_int(k, v);
1797 if (delta_search_threads < 0)
1798 die("invalid number of threads specified (%d)",
1799 delta_search_threads);
1800 #ifndef THREADED_DELTA_SEARCH
1801 if (delta_search_threads != 1)
1802 warning("no threads support, ignoring %s", k);
1803 #endif
1804 return 0;
1806 if (!strcmp(k, "pack.indexversion")) {
1807 pack_idx_default_version = git_config_int(k, v);
1808 if (pack_idx_default_version > 2)
1809 die("bad pack.indexversion=%"PRIu32,
1810 pack_idx_default_version);
1811 return 0;
1813 if (!strcmp(k, "pack.packsizelimit")) {
1814 pack_size_limit_cfg = git_config_ulong(k, v);
1815 return 0;
1817 return git_default_config(k, v, cb);
1820 static void read_object_list_from_stdin(void)
1822 char line[40 + 1 + PATH_MAX + 2];
1823 unsigned char sha1[20];
1825 for (;;) {
1826 if (!fgets(line, sizeof(line), stdin)) {
1827 if (feof(stdin))
1828 break;
1829 if (!ferror(stdin))
1830 die("fgets returned NULL, not EOF, not error!");
1831 if (errno != EINTR)
1832 die("fgets: %s", strerror(errno));
1833 clearerr(stdin);
1834 continue;
1836 if (line[0] == '-') {
1837 if (get_sha1_hex(line+1, sha1))
1838 die("expected edge sha1, got garbage:\n %s",
1839 line);
1840 add_preferred_base(sha1);
1841 continue;
1843 if (get_sha1_hex(line, sha1))
1844 die("expected sha1, got garbage:\n %s", line);
1846 add_preferred_base_object(line+41);
1847 add_object_entry(sha1, 0, line+41, 0);
1851 #define OBJECT_ADDED (1u<<20)
1853 static void show_commit(struct commit *commit)
1855 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1856 commit->object.flags |= OBJECT_ADDED;
1859 static void show_object(struct object_array_entry *p)
1861 add_preferred_base_object(p->name);
1862 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1863 p->item->flags |= OBJECT_ADDED;
1864 free((char *)p->name);
1865 p->name = NULL;
1868 static void show_edge(struct commit *commit)
1870 add_preferred_base(commit->object.sha1);
1873 struct in_pack_object {
1874 off_t offset;
1875 struct object *object;
1878 struct in_pack {
1879 int alloc;
1880 int nr;
1881 struct in_pack_object *array;
1884 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1886 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1887 in_pack->array[in_pack->nr].object = object;
1888 in_pack->nr++;
1892 * Compare the objects in the offset order, in order to emulate the
1893 * "git rev-list --objects" output that produced the pack originally.
1895 static int ofscmp(const void *a_, const void *b_)
1897 struct in_pack_object *a = (struct in_pack_object *)a_;
1898 struct in_pack_object *b = (struct in_pack_object *)b_;
1900 if (a->offset < b->offset)
1901 return -1;
1902 else if (a->offset > b->offset)
1903 return 1;
1904 else
1905 return hashcmp(a->object->sha1, b->object->sha1);
1908 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1910 struct packed_git *p;
1911 struct in_pack in_pack;
1912 uint32_t i;
1914 memset(&in_pack, 0, sizeof(in_pack));
1916 for (p = packed_git; p; p = p->next) {
1917 const unsigned char *sha1;
1918 struct object *o;
1920 for (i = 0; i < revs->num_ignore_packed; i++) {
1921 if (matches_pack_name(p, revs->ignore_packed[i]))
1922 break;
1924 if (revs->num_ignore_packed <= i)
1925 continue;
1926 if (open_pack_index(p))
1927 die("cannot open pack index");
1929 ALLOC_GROW(in_pack.array,
1930 in_pack.nr + p->num_objects,
1931 in_pack.alloc);
1933 for (i = 0; i < p->num_objects; i++) {
1934 sha1 = nth_packed_object_sha1(p, i);
1935 o = lookup_unknown_object(sha1);
1936 if (!(o->flags & OBJECT_ADDED))
1937 mark_in_pack_object(o, p, &in_pack);
1938 o->flags |= OBJECT_ADDED;
1942 if (in_pack.nr) {
1943 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1944 ofscmp);
1945 for (i = 0; i < in_pack.nr; i++) {
1946 struct object *o = in_pack.array[i].object;
1947 add_object_entry(o->sha1, o->type, "", 0);
1950 free(in_pack.array);
1953 static void loosen_unused_packed_objects(struct rev_info *revs)
1955 struct packed_git *p;
1956 uint32_t i;
1957 const unsigned char *sha1;
1959 for (p = packed_git; p; p = p->next) {
1960 for (i = 0; i < revs->num_ignore_packed; i++) {
1961 if (matches_pack_name(p, revs->ignore_packed[i]))
1962 break;
1964 if (revs->num_ignore_packed <= i)
1965 continue;
1967 if (open_pack_index(p))
1968 die("cannot open pack index");
1970 for (i = 0; i < p->num_objects; i++) {
1971 sha1 = nth_packed_object_sha1(p, i);
1972 if (!locate_object_entry(sha1))
1973 if (force_object_loose(sha1, p->mtime))
1974 die("unable to force loose object");
1979 static void get_object_list(int ac, const char **av)
1981 struct rev_info revs;
1982 char line[1000];
1983 int flags = 0;
1985 init_revisions(&revs, NULL);
1986 save_commit_buffer = 0;
1987 setup_revisions(ac, av, &revs, NULL);
1989 while (fgets(line, sizeof(line), stdin) != NULL) {
1990 int len = strlen(line);
1991 if (len && line[len - 1] == '\n')
1992 line[--len] = 0;
1993 if (!len)
1994 break;
1995 if (*line == '-') {
1996 if (!strcmp(line, "--not")) {
1997 flags ^= UNINTERESTING;
1998 continue;
2000 die("not a rev '%s'", line);
2002 if (handle_revision_arg(line, &revs, flags, 1))
2003 die("bad revision '%s'", line);
2006 if (prepare_revision_walk(&revs))
2007 die("revision walk setup failed");
2008 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2009 traverse_commit_list(&revs, show_commit, show_object);
2011 if (keep_unreachable)
2012 add_objects_in_unpacked_packs(&revs);
2013 if (unpack_unreachable)
2014 loosen_unused_packed_objects(&revs);
2017 static int adjust_perm(const char *path, mode_t mode)
2019 if (chmod(path, mode))
2020 return -1;
2021 return adjust_shared_perm(path);
2024 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2026 int use_internal_rev_list = 0;
2027 int thin = 0;
2028 uint32_t i;
2029 const char **rp_av;
2030 int rp_ac_alloc = 64;
2031 int rp_ac;
2033 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2035 rp_av[0] = "pack-objects";
2036 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2037 rp_ac = 2;
2039 git_config(git_pack_config, NULL);
2040 if (!pack_compression_seen && core_compression_seen)
2041 pack_compression_level = core_compression_level;
2043 progress = isatty(2);
2044 for (i = 1; i < argc; i++) {
2045 const char *arg = argv[i];
2047 if (*arg != '-')
2048 break;
2050 if (!strcmp("--non-empty", arg)) {
2051 non_empty = 1;
2052 continue;
2054 if (!strcmp("--local", arg)) {
2055 local = 1;
2056 continue;
2058 if (!strcmp("--incremental", arg)) {
2059 incremental = 1;
2060 continue;
2062 if (!strcmp("--honor-pack-keep", arg)) {
2063 ignore_packed_keep = 1;
2064 continue;
2066 if (!prefixcmp(arg, "--compression=")) {
2067 char *end;
2068 int level = strtoul(arg+14, &end, 0);
2069 if (!arg[14] || *end)
2070 usage(pack_usage);
2071 if (level == -1)
2072 level = Z_DEFAULT_COMPRESSION;
2073 else if (level < 0 || level > Z_BEST_COMPRESSION)
2074 die("bad pack compression level %d", level);
2075 pack_compression_level = level;
2076 continue;
2078 if (!prefixcmp(arg, "--max-pack-size=")) {
2079 char *end;
2080 pack_size_limit_cfg = 0;
2081 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2082 if (!arg[16] || *end)
2083 usage(pack_usage);
2084 continue;
2086 if (!prefixcmp(arg, "--window=")) {
2087 char *end;
2088 window = strtoul(arg+9, &end, 0);
2089 if (!arg[9] || *end)
2090 usage(pack_usage);
2091 continue;
2093 if (!prefixcmp(arg, "--window-memory=")) {
2094 if (!git_parse_ulong(arg+16, &window_memory_limit))
2095 usage(pack_usage);
2096 continue;
2098 if (!prefixcmp(arg, "--threads=")) {
2099 char *end;
2100 delta_search_threads = strtoul(arg+10, &end, 0);
2101 if (!arg[10] || *end || delta_search_threads < 0)
2102 usage(pack_usage);
2103 #ifndef THREADED_DELTA_SEARCH
2104 if (delta_search_threads != 1)
2105 warning("no threads support, "
2106 "ignoring %s", arg);
2107 #endif
2108 continue;
2110 if (!prefixcmp(arg, "--depth=")) {
2111 char *end;
2112 depth = strtoul(arg+8, &end, 0);
2113 if (!arg[8] || *end)
2114 usage(pack_usage);
2115 continue;
2117 if (!strcmp("--progress", arg)) {
2118 progress = 1;
2119 continue;
2121 if (!strcmp("--all-progress", arg)) {
2122 progress = 2;
2123 continue;
2125 if (!strcmp("-q", arg)) {
2126 progress = 0;
2127 continue;
2129 if (!strcmp("--no-reuse-delta", arg)) {
2130 reuse_delta = 0;
2131 continue;
2133 if (!strcmp("--no-reuse-object", arg)) {
2134 reuse_object = reuse_delta = 0;
2135 continue;
2137 if (!strcmp("--delta-base-offset", arg)) {
2138 allow_ofs_delta = 1;
2139 continue;
2141 if (!strcmp("--stdout", arg)) {
2142 pack_to_stdout = 1;
2143 continue;
2145 if (!strcmp("--revs", arg)) {
2146 use_internal_rev_list = 1;
2147 continue;
2149 if (!strcmp("--keep-unreachable", arg)) {
2150 keep_unreachable = 1;
2151 continue;
2153 if (!strcmp("--unpack-unreachable", arg)) {
2154 unpack_unreachable = 1;
2155 continue;
2157 if (!strcmp("--include-tag", arg)) {
2158 include_tag = 1;
2159 continue;
2161 if (!strcmp("--unpacked", arg) ||
2162 !prefixcmp(arg, "--unpacked=") ||
2163 !strcmp("--reflog", arg) ||
2164 !strcmp("--all", arg)) {
2165 use_internal_rev_list = 1;
2166 if (rp_ac >= rp_ac_alloc - 1) {
2167 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2168 rp_av = xrealloc(rp_av,
2169 rp_ac_alloc * sizeof(*rp_av));
2171 rp_av[rp_ac++] = arg;
2172 continue;
2174 if (!strcmp("--thin", arg)) {
2175 use_internal_rev_list = 1;
2176 thin = 1;
2177 rp_av[1] = "--objects-edge";
2178 continue;
2180 if (!prefixcmp(arg, "--index-version=")) {
2181 char *c;
2182 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2183 if (pack_idx_default_version > 2)
2184 die("bad %s", arg);
2185 if (*c == ',')
2186 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2187 if (*c || pack_idx_off32_limit & 0x80000000)
2188 die("bad %s", arg);
2189 continue;
2191 usage(pack_usage);
2194 /* Traditionally "pack-objects [options] base extra" failed;
2195 * we would however want to take refs parameter that would
2196 * have been given to upstream rev-list ourselves, which means
2197 * we somehow want to say what the base name is. So the
2198 * syntax would be:
2200 * pack-objects [options] base <refs...>
2202 * in other words, we would treat the first non-option as the
2203 * base_name and send everything else to the internal revision
2204 * walker.
2207 if (!pack_to_stdout)
2208 base_name = argv[i++];
2210 if (pack_to_stdout != !base_name)
2211 usage(pack_usage);
2213 if (!pack_to_stdout && !pack_size_limit)
2214 pack_size_limit = pack_size_limit_cfg;
2216 if (pack_to_stdout && pack_size_limit)
2217 die("--max-pack-size cannot be used to build a pack for transfer.");
2219 if (!pack_to_stdout && thin)
2220 die("--thin cannot be used to build an indexable pack.");
2222 if (keep_unreachable && unpack_unreachable)
2223 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2225 #ifdef THREADED_DELTA_SEARCH
2226 if (!delta_search_threads) /* --threads=0 means autodetect */
2227 delta_search_threads = online_cpus();
2228 #endif
2230 prepare_packed_git();
2232 if (progress)
2233 progress_state = start_progress("Counting objects", 0);
2234 if (!use_internal_rev_list)
2235 read_object_list_from_stdin();
2236 else {
2237 rp_av[rp_ac] = NULL;
2238 get_object_list(rp_ac, rp_av);
2240 if (include_tag && nr_result)
2241 for_each_ref(add_ref_tag, NULL);
2242 stop_progress(&progress_state);
2244 if (non_empty && !nr_result)
2245 return 0;
2246 if (nr_result)
2247 prepare_pack(window, depth);
2248 write_pack_file();
2249 if (progress)
2250 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2251 " reused %"PRIu32" (delta %"PRIu32")\n",
2252 written, written_delta, reused, reused_delta);
2253 return 0;