Rename the test trash directory to contain spaces.
[git/bdonlan.git] / builtin-pack-objects.c
blob777f272668c39931b330be555f41399adfbce397
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] [--keep-unreachable] \n\
32 [<ref-list | <object-list]";
34 struct object_entry {
35 struct pack_idx_entry idx;
36 unsigned long size; /* uncompressed size */
37 struct packed_git *in_pack; /* already in pack */
38 off_t in_pack_offset;
39 struct object_entry *delta; /* delta base object */
40 struct object_entry *delta_child; /* deltified objects who bases me */
41 struct object_entry *delta_sibling; /* other deltified objects who
42 * uses the same base as me
44 void *delta_data; /* cached delta (uncompressed) */
45 unsigned long delta_size; /* delta data size (uncompressed) */
46 unsigned int hash; /* name hint hash */
47 enum object_type type;
48 enum object_type in_pack_type; /* could be delta */
49 unsigned char in_pack_header_size;
50 unsigned char preferred_base; /* we do not pack this, but is available
51 * to be used as the base object to delta
52 * objects against.
54 unsigned char no_try_delta;
58 * Objects we are going to pack are collected in objects array (dynamically
59 * expanded). nr_objects & nr_alloc controls this array. They are stored
60 * in the order we see -- typically rev-list --objects order that gives us
61 * nice "minimum seek" order.
63 static struct object_entry *objects;
64 static struct pack_idx_entry **written_list;
65 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
67 static int non_empty;
68 static int no_reuse_delta, no_reuse_object, keep_unreachable, include_tag;
69 static int local;
70 static int incremental;
71 static int allow_ofs_delta;
72 static const char *base_name;
73 static int progress = 1;
74 static int window = 10;
75 static uint32_t pack_size_limit, pack_size_limit_cfg;
76 static int depth = 50;
77 static int delta_search_threads = 1;
78 static int pack_to_stdout;
79 static int num_preferred_base;
80 static struct progress *progress_state;
81 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
82 static int pack_compression_seen;
84 static unsigned long delta_cache_size = 0;
85 static unsigned long max_delta_cache_size = 0;
86 static unsigned long cache_max_small_delta_size = 1000;
88 static unsigned long window_memory_limit = 0;
91 * The object names in objects array are hashed with this hashtable,
92 * to help looking up the entry by object name.
93 * This hashtable is built after all the objects are seen.
95 static int *object_ix;
96 static int object_ix_hashsz;
99 * stats
101 static uint32_t written, written_delta;
102 static uint32_t reused, reused_delta;
105 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
107 unsigned long othersize, delta_size;
108 enum object_type type;
109 void *otherbuf = read_sha1_file(entry->delta->idx.sha1, &type, &othersize);
110 void *delta_buf;
112 if (!otherbuf)
113 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
114 delta_buf = diff_delta(otherbuf, othersize,
115 buf, size, &delta_size, 0);
116 if (!delta_buf || delta_size != entry->delta_size)
117 die("delta size changed");
118 free(buf);
119 free(otherbuf);
120 return delta_buf;
124 * The per-object header is a pretty dense thing, which is
125 * - first byte: low four bits are "size", then three bits of "type",
126 * and the high bit is "size continues".
127 * - each byte afterwards: low seven bits are size continuation,
128 * with the high bit being "size continues"
130 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
132 int n = 1;
133 unsigned char c;
135 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
136 die("bad type %d", type);
138 c = (type << 4) | (size & 15);
139 size >>= 4;
140 while (size) {
141 *hdr++ = c | 0x80;
142 c = size & 0x7f;
143 size >>= 7;
144 n++;
146 *hdr = c;
147 return n;
151 * we are going to reuse the existing object data as is. make
152 * sure it is not corrupt.
154 static int check_pack_inflate(struct packed_git *p,
155 struct pack_window **w_curs,
156 off_t offset,
157 off_t len,
158 unsigned long expect)
160 z_stream stream;
161 unsigned char fakebuf[4096], *in;
162 int st;
164 memset(&stream, 0, sizeof(stream));
165 inflateInit(&stream);
166 do {
167 in = use_pack(p, w_curs, offset, &stream.avail_in);
168 stream.next_in = in;
169 stream.next_out = fakebuf;
170 stream.avail_out = sizeof(fakebuf);
171 st = inflate(&stream, Z_FINISH);
172 offset += stream.next_in - in;
173 } while (st == Z_OK || st == Z_BUF_ERROR);
174 inflateEnd(&stream);
175 return (st == Z_STREAM_END &&
176 stream.total_out == expect &&
177 stream.total_in == len) ? 0 : -1;
180 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
181 off_t offset, off_t len, unsigned int nr)
183 const uint32_t *index_crc;
184 uint32_t data_crc = crc32(0, Z_NULL, 0);
186 do {
187 unsigned int avail;
188 void *data = use_pack(p, w_curs, offset, &avail);
189 if (avail > len)
190 avail = len;
191 data_crc = crc32(data_crc, data, avail);
192 offset += avail;
193 len -= avail;
194 } while (len);
196 index_crc = p->index_data;
197 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
199 return data_crc != ntohl(*index_crc);
202 static void copy_pack_data(struct sha1file *f,
203 struct packed_git *p,
204 struct pack_window **w_curs,
205 off_t offset,
206 off_t len)
208 unsigned char *in;
209 unsigned int avail;
211 while (len) {
212 in = use_pack(p, w_curs, offset, &avail);
213 if (avail > len)
214 avail = (unsigned int)len;
215 sha1write(f, in, avail);
216 offset += avail;
217 len -= avail;
221 static unsigned long write_object(struct sha1file *f,
222 struct object_entry *entry,
223 off_t write_offset)
225 unsigned long size;
226 enum object_type type;
227 void *buf;
228 unsigned char header[10];
229 unsigned char dheader[10];
230 unsigned hdrlen;
231 off_t datalen;
232 enum object_type obj_type;
233 int to_reuse = 0;
234 /* write limit if limited packsize and not first object */
235 unsigned long limit = pack_size_limit && nr_written ?
236 pack_size_limit - write_offset : 0;
237 /* no if no delta */
238 int usable_delta = !entry->delta ? 0 :
239 /* yes if unlimited packfile */
240 !pack_size_limit ? 1 :
241 /* no if base written to previous pack */
242 entry->delta->idx.offset == (off_t)-1 ? 0 :
243 /* otherwise double-check written to this
244 * pack, like we do below
246 entry->delta->idx.offset ? 1 : 0;
248 if (!pack_to_stdout)
249 crc32_begin(f);
251 obj_type = entry->type;
252 if (no_reuse_object)
253 to_reuse = 0; /* explicit */
254 else if (!entry->in_pack)
255 to_reuse = 0; /* can't reuse what we don't have */
256 else if (obj_type == OBJ_REF_DELTA || obj_type == OBJ_OFS_DELTA)
257 /* check_object() decided it for us ... */
258 to_reuse = usable_delta;
259 /* ... but pack split may override that */
260 else if (obj_type != entry->in_pack_type)
261 to_reuse = 0; /* pack has delta which is unusable */
262 else if (entry->delta)
263 to_reuse = 0; /* we want to pack afresh */
264 else
265 to_reuse = 1; /* we have it in-pack undeltified,
266 * and we do not need to deltify it.
269 if (!to_reuse) {
270 z_stream stream;
271 unsigned long maxsize;
272 void *out;
273 if (!usable_delta) {
274 buf = read_sha1_file(entry->idx.sha1, &obj_type, &size);
275 if (!buf)
276 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
277 } else if (entry->delta_data) {
278 size = entry->delta_size;
279 buf = entry->delta_data;
280 entry->delta_data = NULL;
281 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
282 OBJ_OFS_DELTA : OBJ_REF_DELTA;
283 } else {
284 buf = read_sha1_file(entry->idx.sha1, &type, &size);
285 if (!buf)
286 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
287 buf = delta_against(buf, size, entry);
288 size = entry->delta_size;
289 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
290 OBJ_OFS_DELTA : OBJ_REF_DELTA;
292 /* compress the data to store and put compressed length in datalen */
293 memset(&stream, 0, sizeof(stream));
294 deflateInit(&stream, pack_compression_level);
295 maxsize = deflateBound(&stream, size);
296 out = xmalloc(maxsize);
297 /* Compress it */
298 stream.next_in = buf;
299 stream.avail_in = size;
300 stream.next_out = out;
301 stream.avail_out = maxsize;
302 while (deflate(&stream, Z_FINISH) == Z_OK)
303 /* nothing */;
304 deflateEnd(&stream);
305 datalen = stream.total_out;
308 * The object header is a byte of 'type' followed by zero or
309 * more bytes of length.
311 hdrlen = encode_header(obj_type, size, header);
313 if (obj_type == OBJ_OFS_DELTA) {
315 * Deltas with relative base contain an additional
316 * encoding of the relative offset for the delta
317 * base from this object's position in the pack.
319 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
320 unsigned pos = sizeof(dheader) - 1;
321 dheader[pos] = ofs & 127;
322 while (ofs >>= 7)
323 dheader[--pos] = 128 | (--ofs & 127);
324 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
325 free(out);
326 free(buf);
327 return 0;
329 sha1write(f, header, hdrlen);
330 sha1write(f, dheader + pos, sizeof(dheader) - pos);
331 hdrlen += sizeof(dheader) - pos;
332 } else if (obj_type == OBJ_REF_DELTA) {
334 * Deltas with a base reference contain
335 * an additional 20 bytes for the base sha1.
337 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
338 free(out);
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(out);
348 free(buf);
349 return 0;
351 sha1write(f, header, hdrlen);
353 sha1write(f, out, datalen);
354 free(out);
355 free(buf);
357 else {
358 struct packed_git *p = entry->in_pack;
359 struct pack_window *w_curs = NULL;
360 struct revindex_entry *revidx;
361 off_t offset;
363 if (entry->delta) {
364 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
365 OBJ_OFS_DELTA : OBJ_REF_DELTA;
366 reused_delta++;
368 hdrlen = encode_header(obj_type, entry->size, header);
369 offset = entry->in_pack_offset;
370 revidx = find_pack_revindex(p, offset);
371 datalen = revidx[1].offset - offset;
372 if (!pack_to_stdout && p->index_version > 1 &&
373 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
374 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
375 offset += entry->in_pack_header_size;
376 datalen -= entry->in_pack_header_size;
377 if (obj_type == OBJ_OFS_DELTA) {
378 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
379 unsigned pos = sizeof(dheader) - 1;
380 dheader[pos] = ofs & 127;
381 while (ofs >>= 7)
382 dheader[--pos] = 128 | (--ofs & 127);
383 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
384 return 0;
385 sha1write(f, header, hdrlen);
386 sha1write(f, dheader + pos, sizeof(dheader) - pos);
387 hdrlen += sizeof(dheader) - pos;
388 } else if (obj_type == OBJ_REF_DELTA) {
389 if (limit && hdrlen + 20 + datalen + 20 >= limit)
390 return 0;
391 sha1write(f, header, hdrlen);
392 sha1write(f, entry->delta->idx.sha1, 20);
393 hdrlen += 20;
394 } else {
395 if (limit && hdrlen + datalen + 20 >= limit)
396 return 0;
397 sha1write(f, header, hdrlen);
400 if (!pack_to_stdout && p->index_version == 1 &&
401 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
402 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
403 copy_pack_data(f, p, &w_curs, offset, datalen);
404 unuse_pack(&w_curs);
405 reused++;
407 if (usable_delta)
408 written_delta++;
409 written++;
410 if (!pack_to_stdout)
411 entry->idx.crc32 = crc32_end(f);
412 return hdrlen + datalen;
415 static off_t write_one(struct sha1file *f,
416 struct object_entry *e,
417 off_t offset)
419 unsigned long size;
421 /* offset is non zero if object is written already. */
422 if (e->idx.offset || e->preferred_base)
423 return offset;
425 /* if we are deltified, write out base object first. */
426 if (e->delta) {
427 offset = write_one(f, e->delta, offset);
428 if (!offset)
429 return 0;
432 e->idx.offset = offset;
433 size = write_object(f, e, offset);
434 if (!size) {
435 e->idx.offset = 0;
436 return 0;
438 written_list[nr_written++] = &e->idx;
440 /* make sure off_t is sufficiently large not to wrap */
441 if (offset > offset + size)
442 die("pack too large for current definition of off_t");
443 return offset + size;
446 /* forward declaration for write_pack_file */
447 static int adjust_perm(const char *path, mode_t mode);
449 static void write_pack_file(void)
451 uint32_t i = 0, j;
452 struct sha1file *f;
453 off_t offset, offset_one, last_obj_offset = 0;
454 struct pack_header hdr;
455 int do_progress = progress >> pack_to_stdout;
456 uint32_t nr_remaining = nr_result;
457 time_t last_mtime = 0;
459 if (do_progress)
460 progress_state = start_progress("Writing objects", nr_result);
461 written_list = xmalloc(nr_objects * sizeof(*written_list));
463 do {
464 unsigned char sha1[20];
465 char *pack_tmp_name = NULL;
467 if (pack_to_stdout) {
468 f = sha1fd_throughput(1, "<stdout>", progress_state);
469 } else {
470 char tmpname[PATH_MAX];
471 int fd;
472 snprintf(tmpname, sizeof(tmpname),
473 "%s/tmp_pack_XXXXXX", get_object_directory());
474 fd = xmkstemp(tmpname);
475 pack_tmp_name = xstrdup(tmpname);
476 f = sha1fd(fd, pack_tmp_name);
479 hdr.hdr_signature = htonl(PACK_SIGNATURE);
480 hdr.hdr_version = htonl(PACK_VERSION);
481 hdr.hdr_entries = htonl(nr_remaining);
482 sha1write(f, &hdr, sizeof(hdr));
483 offset = sizeof(hdr);
484 nr_written = 0;
485 for (; i < nr_objects; i++) {
486 last_obj_offset = offset;
487 offset_one = write_one(f, objects + i, offset);
488 if (!offset_one)
489 break;
490 offset = offset_one;
491 display_progress(progress_state, written);
495 * Did we write the wrong # entries in the header?
496 * If so, rewrite it like in fast-import
498 if (pack_to_stdout || nr_written == nr_remaining) {
499 sha1close(f, sha1, 1);
500 } else {
501 int fd = sha1close(f, NULL, 0);
502 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
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 %u objects while expecting %u", written, nr_result);
573 * We have scanned through [0 ... i). Since we have written
574 * the correct number of objects, the remaining [i ... nr_objects)
575 * items must be either already written (due to out-of-order delta base)
576 * or a preferred base. Count those which are neither and complain if any.
578 for (j = 0; i < nr_objects; i++) {
579 struct object_entry *e = objects + i;
580 j += !e->idx.offset && !e->preferred_base;
582 if (j)
583 die("wrote %u objects as expected but %u unwritten", written, j);
586 static int locate_object_entry_hash(const unsigned char *sha1)
588 int i;
589 unsigned int ui;
590 memcpy(&ui, sha1, sizeof(unsigned int));
591 i = ui % object_ix_hashsz;
592 while (0 < object_ix[i]) {
593 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
594 return i;
595 if (++i == object_ix_hashsz)
596 i = 0;
598 return -1 - i;
601 static struct object_entry *locate_object_entry(const unsigned char *sha1)
603 int i;
605 if (!object_ix_hashsz)
606 return NULL;
608 i = locate_object_entry_hash(sha1);
609 if (0 <= i)
610 return &objects[object_ix[i]-1];
611 return NULL;
614 static void rehash_objects(void)
616 uint32_t i;
617 struct object_entry *oe;
619 object_ix_hashsz = nr_objects * 3;
620 if (object_ix_hashsz < 1024)
621 object_ix_hashsz = 1024;
622 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
623 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
624 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
625 int ix = locate_object_entry_hash(oe->idx.sha1);
626 if (0 <= ix)
627 continue;
628 ix = -1 - ix;
629 object_ix[ix] = i + 1;
633 static unsigned name_hash(const char *name)
635 unsigned char c;
636 unsigned hash = 0;
638 if (!name)
639 return 0;
642 * This effectively just creates a sortable number from the
643 * last sixteen non-whitespace characters. Last characters
644 * count "most", so things that end in ".c" sort together.
646 while ((c = *name++) != 0) {
647 if (isspace(c))
648 continue;
649 hash = (hash >> 2) + (c << 24);
651 return hash;
654 static void setup_delta_attr_check(struct git_attr_check *check)
656 static struct git_attr *attr_delta;
658 if (!attr_delta)
659 attr_delta = git_attr("delta", 5);
661 check[0].attr = attr_delta;
664 static int no_try_delta(const char *path)
666 struct git_attr_check check[1];
668 setup_delta_attr_check(check);
669 if (git_checkattr(path, ARRAY_SIZE(check), check))
670 return 0;
671 if (ATTR_FALSE(check->value))
672 return 1;
673 return 0;
676 static int add_object_entry(const unsigned char *sha1, enum object_type type,
677 const char *name, int exclude)
679 struct object_entry *entry;
680 struct packed_git *p, *found_pack = NULL;
681 off_t found_offset = 0;
682 int ix;
683 unsigned hash = name_hash(name);
685 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
686 if (ix >= 0) {
687 if (exclude) {
688 entry = objects + object_ix[ix] - 1;
689 if (!entry->preferred_base)
690 nr_result--;
691 entry->preferred_base = 1;
693 return 0;
696 for (p = packed_git; p; p = p->next) {
697 off_t offset = find_pack_entry_one(sha1, p);
698 if (offset) {
699 if (!found_pack) {
700 found_offset = offset;
701 found_pack = p;
703 if (exclude)
704 break;
705 if (incremental)
706 return 0;
707 if (local && !p->pack_local)
708 return 0;
712 if (nr_objects >= nr_alloc) {
713 nr_alloc = (nr_alloc + 1024) * 3 / 2;
714 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
717 entry = objects + nr_objects++;
718 memset(entry, 0, sizeof(*entry));
719 hashcpy(entry->idx.sha1, sha1);
720 entry->hash = hash;
721 if (type)
722 entry->type = type;
723 if (exclude)
724 entry->preferred_base = 1;
725 else
726 nr_result++;
727 if (found_pack) {
728 entry->in_pack = found_pack;
729 entry->in_pack_offset = found_offset;
732 if (object_ix_hashsz * 3 <= nr_objects * 4)
733 rehash_objects();
734 else
735 object_ix[-1 - ix] = nr_objects;
737 display_progress(progress_state, nr_objects);
739 if (name && no_try_delta(name))
740 entry->no_try_delta = 1;
742 return 1;
745 struct pbase_tree_cache {
746 unsigned char sha1[20];
747 int ref;
748 int temporary;
749 void *tree_data;
750 unsigned long tree_size;
753 static struct pbase_tree_cache *(pbase_tree_cache[256]);
754 static int pbase_tree_cache_ix(const unsigned char *sha1)
756 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
758 static int pbase_tree_cache_ix_incr(int ix)
760 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
763 static struct pbase_tree {
764 struct pbase_tree *next;
765 /* This is a phony "cache" entry; we are not
766 * going to evict it nor find it through _get()
767 * mechanism -- this is for the toplevel node that
768 * would almost always change with any commit.
770 struct pbase_tree_cache pcache;
771 } *pbase_tree;
773 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
775 struct pbase_tree_cache *ent, *nent;
776 void *data;
777 unsigned long size;
778 enum object_type type;
779 int neigh;
780 int my_ix = pbase_tree_cache_ix(sha1);
781 int available_ix = -1;
783 /* pbase-tree-cache acts as a limited hashtable.
784 * your object will be found at your index or within a few
785 * slots after that slot if it is cached.
787 for (neigh = 0; neigh < 8; neigh++) {
788 ent = pbase_tree_cache[my_ix];
789 if (ent && !hashcmp(ent->sha1, sha1)) {
790 ent->ref++;
791 return ent;
793 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
794 ((0 <= available_ix) &&
795 (!ent && pbase_tree_cache[available_ix])))
796 available_ix = my_ix;
797 if (!ent)
798 break;
799 my_ix = pbase_tree_cache_ix_incr(my_ix);
802 /* Did not find one. Either we got a bogus request or
803 * we need to read and perhaps cache.
805 data = read_sha1_file(sha1, &type, &size);
806 if (!data)
807 return NULL;
808 if (type != OBJ_TREE) {
809 free(data);
810 return NULL;
813 /* We need to either cache or return a throwaway copy */
815 if (available_ix < 0)
816 ent = NULL;
817 else {
818 ent = pbase_tree_cache[available_ix];
819 my_ix = available_ix;
822 if (!ent) {
823 nent = xmalloc(sizeof(*nent));
824 nent->temporary = (available_ix < 0);
826 else {
827 /* evict and reuse */
828 free(ent->tree_data);
829 nent = ent;
831 hashcpy(nent->sha1, sha1);
832 nent->tree_data = data;
833 nent->tree_size = size;
834 nent->ref = 1;
835 if (!nent->temporary)
836 pbase_tree_cache[my_ix] = nent;
837 return nent;
840 static void pbase_tree_put(struct pbase_tree_cache *cache)
842 if (!cache->temporary) {
843 cache->ref--;
844 return;
846 free(cache->tree_data);
847 free(cache);
850 static int name_cmp_len(const char *name)
852 int i;
853 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
855 return i;
858 static void add_pbase_object(struct tree_desc *tree,
859 const char *name,
860 int cmplen,
861 const char *fullname)
863 struct name_entry entry;
864 int cmp;
866 while (tree_entry(tree,&entry)) {
867 if (S_ISGITLINK(entry.mode))
868 continue;
869 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
870 memcmp(name, entry.path, cmplen);
871 if (cmp > 0)
872 continue;
873 if (cmp < 0)
874 return;
875 if (name[cmplen] != '/') {
876 add_object_entry(entry.sha1,
877 object_type(entry.mode),
878 fullname, 1);
879 return;
881 if (S_ISDIR(entry.mode)) {
882 struct tree_desc sub;
883 struct pbase_tree_cache *tree;
884 const char *down = name+cmplen+1;
885 int downlen = name_cmp_len(down);
887 tree = pbase_tree_get(entry.sha1);
888 if (!tree)
889 return;
890 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
892 add_pbase_object(&sub, down, downlen, fullname);
893 pbase_tree_put(tree);
898 static unsigned *done_pbase_paths;
899 static int done_pbase_paths_num;
900 static int done_pbase_paths_alloc;
901 static int done_pbase_path_pos(unsigned hash)
903 int lo = 0;
904 int hi = done_pbase_paths_num;
905 while (lo < hi) {
906 int mi = (hi + lo) / 2;
907 if (done_pbase_paths[mi] == hash)
908 return mi;
909 if (done_pbase_paths[mi] < hash)
910 hi = mi;
911 else
912 lo = mi + 1;
914 return -lo-1;
917 static int check_pbase_path(unsigned hash)
919 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
920 if (0 <= pos)
921 return 1;
922 pos = -pos - 1;
923 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
924 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
925 done_pbase_paths = xrealloc(done_pbase_paths,
926 done_pbase_paths_alloc *
927 sizeof(unsigned));
929 done_pbase_paths_num++;
930 if (pos < done_pbase_paths_num)
931 memmove(done_pbase_paths + pos + 1,
932 done_pbase_paths + pos,
933 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
934 done_pbase_paths[pos] = hash;
935 return 0;
938 static void add_preferred_base_object(const char *name)
940 struct pbase_tree *it;
941 int cmplen;
942 unsigned hash = name_hash(name);
944 if (!num_preferred_base || check_pbase_path(hash))
945 return;
947 cmplen = name_cmp_len(name);
948 for (it = pbase_tree; it; it = it->next) {
949 if (cmplen == 0) {
950 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
952 else {
953 struct tree_desc tree;
954 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
955 add_pbase_object(&tree, name, cmplen, name);
960 static void add_preferred_base(unsigned char *sha1)
962 struct pbase_tree *it;
963 void *data;
964 unsigned long size;
965 unsigned char tree_sha1[20];
967 if (window <= num_preferred_base++)
968 return;
970 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
971 if (!data)
972 return;
974 for (it = pbase_tree; it; it = it->next) {
975 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
976 free(data);
977 return;
981 it = xcalloc(1, sizeof(*it));
982 it->next = pbase_tree;
983 pbase_tree = it;
985 hashcpy(it->pcache.sha1, tree_sha1);
986 it->pcache.tree_data = data;
987 it->pcache.tree_size = size;
990 static void check_object(struct object_entry *entry)
992 if (entry->in_pack) {
993 struct packed_git *p = entry->in_pack;
994 struct pack_window *w_curs = NULL;
995 const unsigned char *base_ref = NULL;
996 struct object_entry *base_entry;
997 unsigned long used, used_0;
998 unsigned int avail;
999 off_t ofs;
1000 unsigned char *buf, c;
1002 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1005 * We want in_pack_type even if we do not reuse delta
1006 * since non-delta representations could still be reused.
1008 used = unpack_object_header_gently(buf, avail,
1009 &entry->in_pack_type,
1010 &entry->size);
1013 * Determine if this is a delta and if so whether we can
1014 * reuse it or not. Otherwise let's find out as cheaply as
1015 * possible what the actual type and size for this object is.
1017 switch (entry->in_pack_type) {
1018 default:
1019 /* Not a delta hence we've already got all we need. */
1020 entry->type = entry->in_pack_type;
1021 entry->in_pack_header_size = used;
1022 unuse_pack(&w_curs);
1023 return;
1024 case OBJ_REF_DELTA:
1025 if (!no_reuse_delta && !entry->preferred_base)
1026 base_ref = use_pack(p, &w_curs,
1027 entry->in_pack_offset + used, NULL);
1028 entry->in_pack_header_size = used + 20;
1029 break;
1030 case OBJ_OFS_DELTA:
1031 buf = use_pack(p, &w_curs,
1032 entry->in_pack_offset + used, NULL);
1033 used_0 = 0;
1034 c = buf[used_0++];
1035 ofs = c & 127;
1036 while (c & 128) {
1037 ofs += 1;
1038 if (!ofs || MSB(ofs, 7))
1039 die("delta base offset overflow in pack for %s",
1040 sha1_to_hex(entry->idx.sha1));
1041 c = buf[used_0++];
1042 ofs = (ofs << 7) + (c & 127);
1044 if (ofs >= entry->in_pack_offset)
1045 die("delta base offset out of bound for %s",
1046 sha1_to_hex(entry->idx.sha1));
1047 ofs = entry->in_pack_offset - ofs;
1048 if (!no_reuse_delta && !entry->preferred_base) {
1049 struct revindex_entry *revidx;
1050 revidx = find_pack_revindex(p, ofs);
1051 base_ref = nth_packed_object_sha1(p, revidx->nr);
1053 entry->in_pack_header_size = used + used_0;
1054 break;
1057 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1059 * If base_ref was set above that means we wish to
1060 * reuse delta data, and we even found that base
1061 * in the list of objects we want to pack. Goodie!
1063 * Depth value does not matter - find_deltas() will
1064 * never consider reused delta as the base object to
1065 * deltify other objects against, in order to avoid
1066 * circular deltas.
1068 entry->type = entry->in_pack_type;
1069 entry->delta = base_entry;
1070 entry->delta_sibling = base_entry->delta_child;
1071 base_entry->delta_child = entry;
1072 unuse_pack(&w_curs);
1073 return;
1076 if (entry->type) {
1078 * This must be a delta and we already know what the
1079 * final object type is. Let's extract the actual
1080 * object size from the delta header.
1082 entry->size = get_size_from_delta(p, &w_curs,
1083 entry->in_pack_offset + entry->in_pack_header_size);
1084 unuse_pack(&w_curs);
1085 return;
1089 * No choice but to fall back to the recursive delta walk
1090 * with sha1_object_info() to find about the object type
1091 * at this point...
1093 unuse_pack(&w_curs);
1096 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1097 if (entry->type < 0)
1098 die("unable to get type of object %s",
1099 sha1_to_hex(entry->idx.sha1));
1102 static int pack_offset_sort(const void *_a, const void *_b)
1104 const struct object_entry *a = *(struct object_entry **)_a;
1105 const struct object_entry *b = *(struct object_entry **)_b;
1107 /* avoid filesystem trashing with loose objects */
1108 if (!a->in_pack && !b->in_pack)
1109 return hashcmp(a->idx.sha1, b->idx.sha1);
1111 if (a->in_pack < b->in_pack)
1112 return -1;
1113 if (a->in_pack > b->in_pack)
1114 return 1;
1115 return a->in_pack_offset < b->in_pack_offset ? -1 :
1116 (a->in_pack_offset > b->in_pack_offset);
1119 static void get_object_details(void)
1121 uint32_t i;
1122 struct object_entry **sorted_by_offset;
1124 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1125 for (i = 0; i < nr_objects; i++)
1126 sorted_by_offset[i] = objects + i;
1127 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1129 init_pack_revindex();
1131 for (i = 0; i < nr_objects; i++)
1132 check_object(sorted_by_offset[i]);
1134 free(sorted_by_offset);
1138 * We search for deltas in a list sorted by type, by filename hash, and then
1139 * by size, so that we see progressively smaller and smaller files.
1140 * That's because we prefer deltas to be from the bigger file
1141 * to the smaller -- deletes are potentially cheaper, but perhaps
1142 * more importantly, the bigger file is likely the more recent
1143 * one. The deepest deltas are therefore the oldest objects which are
1144 * less susceptible to be accessed often.
1146 static int type_size_sort(const void *_a, const void *_b)
1148 const struct object_entry *a = *(struct object_entry **)_a;
1149 const struct object_entry *b = *(struct object_entry **)_b;
1151 if (a->type > b->type)
1152 return -1;
1153 if (a->type < b->type)
1154 return 1;
1155 if (a->hash > b->hash)
1156 return -1;
1157 if (a->hash < b->hash)
1158 return 1;
1159 if (a->preferred_base > b->preferred_base)
1160 return -1;
1161 if (a->preferred_base < b->preferred_base)
1162 return 1;
1163 if (a->size > b->size)
1164 return -1;
1165 if (a->size < b->size)
1166 return 1;
1167 return a < b ? -1 : (a > b); /* newest first */
1170 struct unpacked {
1171 struct object_entry *entry;
1172 void *data;
1173 struct delta_index *index;
1174 unsigned depth;
1177 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1178 unsigned long delta_size)
1180 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1181 return 0;
1183 if (delta_size < cache_max_small_delta_size)
1184 return 1;
1186 /* cache delta, if objects are large enough compared to delta size */
1187 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1188 return 1;
1190 return 0;
1193 #ifdef THREADED_DELTA_SEARCH
1195 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1196 #define read_lock() pthread_mutex_lock(&read_mutex)
1197 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1199 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1200 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1201 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1203 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1204 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1205 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1207 #else
1209 #define read_lock() (void)0
1210 #define read_unlock() (void)0
1211 #define cache_lock() (void)0
1212 #define cache_unlock() (void)0
1213 #define progress_lock() (void)0
1214 #define progress_unlock() (void)0
1216 #endif
1218 static int try_delta(struct unpacked *trg, struct unpacked *src,
1219 unsigned max_depth, unsigned long *mem_usage)
1221 struct object_entry *trg_entry = trg->entry;
1222 struct object_entry *src_entry = src->entry;
1223 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1224 unsigned ref_depth;
1225 enum object_type type;
1226 void *delta_buf;
1228 /* Don't bother doing diffs between different types */
1229 if (trg_entry->type != src_entry->type)
1230 return -1;
1233 * We do not bother to try a delta that we discarded
1234 * on an earlier try, but only when reusing delta data.
1236 if (!no_reuse_delta && trg_entry->in_pack &&
1237 trg_entry->in_pack == src_entry->in_pack &&
1238 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1239 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1240 return 0;
1242 /* Let's not bust the allowed depth. */
1243 if (src->depth >= max_depth)
1244 return 0;
1246 /* Now some size filtering heuristics. */
1247 trg_size = trg_entry->size;
1248 if (!trg_entry->delta) {
1249 max_size = trg_size/2 - 20;
1250 ref_depth = 1;
1251 } else {
1252 max_size = trg_entry->delta_size;
1253 ref_depth = trg->depth;
1255 max_size = max_size * (max_depth - src->depth) /
1256 (max_depth - ref_depth + 1);
1257 if (max_size == 0)
1258 return 0;
1259 src_size = src_entry->size;
1260 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1261 if (sizediff >= max_size)
1262 return 0;
1263 if (trg_size < src_size / 32)
1264 return 0;
1266 /* Load data if not already done */
1267 if (!trg->data) {
1268 read_lock();
1269 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1270 read_unlock();
1271 if (!trg->data)
1272 die("object %s cannot be read",
1273 sha1_to_hex(trg_entry->idx.sha1));
1274 if (sz != trg_size)
1275 die("object %s inconsistent object length (%lu vs %lu)",
1276 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1277 *mem_usage += sz;
1279 if (!src->data) {
1280 read_lock();
1281 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1282 read_unlock();
1283 if (!src->data)
1284 die("object %s cannot be read",
1285 sha1_to_hex(src_entry->idx.sha1));
1286 if (sz != src_size)
1287 die("object %s inconsistent object length (%lu vs %lu)",
1288 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1289 *mem_usage += sz;
1291 if (!src->index) {
1292 src->index = create_delta_index(src->data, src_size);
1293 if (!src->index) {
1294 static int warned = 0;
1295 if (!warned++)
1296 warning("suboptimal pack - out of memory");
1297 return 0;
1299 *mem_usage += sizeof_delta_index(src->index);
1302 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1303 if (!delta_buf)
1304 return 0;
1306 if (trg_entry->delta) {
1307 /* Prefer only shallower same-sized deltas. */
1308 if (delta_size == trg_entry->delta_size &&
1309 src->depth + 1 >= trg->depth) {
1310 free(delta_buf);
1311 return 0;
1316 * Handle memory allocation outside of the cache
1317 * accounting lock. Compiler will optimize the strangeness
1318 * away when THREADED_DELTA_SEARCH is not defined.
1320 free(trg_entry->delta_data);
1321 cache_lock();
1322 if (trg_entry->delta_data) {
1323 delta_cache_size -= trg_entry->delta_size;
1324 trg_entry->delta_data = NULL;
1326 if (delta_cacheable(src_size, trg_size, delta_size)) {
1327 delta_cache_size += delta_size;
1328 cache_unlock();
1329 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1330 } else {
1331 cache_unlock();
1332 free(delta_buf);
1335 trg_entry->delta = src_entry;
1336 trg_entry->delta_size = delta_size;
1337 trg->depth = src->depth + 1;
1339 return 1;
1342 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1344 struct object_entry *child = me->delta_child;
1345 unsigned int m = n;
1346 while (child) {
1347 unsigned int c = check_delta_limit(child, n + 1);
1348 if (m < c)
1349 m = c;
1350 child = child->delta_sibling;
1352 return m;
1355 static unsigned long free_unpacked(struct unpacked *n)
1357 unsigned long freed_mem = sizeof_delta_index(n->index);
1358 free_delta_index(n->index);
1359 n->index = NULL;
1360 if (n->data) {
1361 freed_mem += n->entry->size;
1362 free(n->data);
1363 n->data = NULL;
1365 n->entry = NULL;
1366 n->depth = 0;
1367 return freed_mem;
1370 static void find_deltas(struct object_entry **list, unsigned *list_size,
1371 int window, int depth, unsigned *processed)
1373 uint32_t i, idx = 0, count = 0;
1374 unsigned int array_size = window * sizeof(struct unpacked);
1375 struct unpacked *array;
1376 unsigned long mem_usage = 0;
1378 array = xmalloc(array_size);
1379 memset(array, 0, array_size);
1381 for (;;) {
1382 struct object_entry *entry = *list++;
1383 struct unpacked *n = array + idx;
1384 int j, max_depth, best_base = -1;
1386 progress_lock();
1387 if (!*list_size) {
1388 progress_unlock();
1389 break;
1391 (*list_size)--;
1392 if (!entry->preferred_base) {
1393 (*processed)++;
1394 display_progress(progress_state, *processed);
1396 progress_unlock();
1398 mem_usage -= free_unpacked(n);
1399 n->entry = entry;
1401 while (window_memory_limit &&
1402 mem_usage > window_memory_limit &&
1403 count > 1) {
1404 uint32_t tail = (idx + window - count) % window;
1405 mem_usage -= free_unpacked(array + tail);
1406 count--;
1409 /* We do not compute delta to *create* objects we are not
1410 * going to pack.
1412 if (entry->preferred_base)
1413 goto next;
1416 * If the current object is at pack edge, take the depth the
1417 * objects that depend on the current object into account
1418 * otherwise they would become too deep.
1420 max_depth = depth;
1421 if (entry->delta_child) {
1422 max_depth -= check_delta_limit(entry, 0);
1423 if (max_depth <= 0)
1424 goto next;
1427 j = window;
1428 while (--j > 0) {
1429 int ret;
1430 uint32_t other_idx = idx + j;
1431 struct unpacked *m;
1432 if (other_idx >= window)
1433 other_idx -= window;
1434 m = array + other_idx;
1435 if (!m->entry)
1436 break;
1437 ret = try_delta(n, m, max_depth, &mem_usage);
1438 if (ret < 0)
1439 break;
1440 else if (ret > 0)
1441 best_base = other_idx;
1444 /* if we made n a delta, and if n is already at max
1445 * depth, leaving it in the window is pointless. we
1446 * should evict it first.
1448 if (entry->delta && depth <= n->depth)
1449 continue;
1452 * Move the best delta base up in the window, after the
1453 * currently deltified object, to keep it longer. It will
1454 * be the first base object to be attempted next.
1456 if (entry->delta) {
1457 struct unpacked swap = array[best_base];
1458 int dist = (window + idx - best_base) % window;
1459 int dst = best_base;
1460 while (dist--) {
1461 int src = (dst + 1) % window;
1462 array[dst] = array[src];
1463 dst = src;
1465 array[dst] = swap;
1468 next:
1469 idx++;
1470 if (count + 1 < window)
1471 count++;
1472 if (idx >= window)
1473 idx = 0;
1476 for (i = 0; i < window; ++i) {
1477 free_delta_index(array[i].index);
1478 free(array[i].data);
1480 free(array);
1483 #ifdef THREADED_DELTA_SEARCH
1486 * The main thread waits on the condition that (at least) one of the workers
1487 * has stopped working (which is indicated in the .working member of
1488 * struct thread_params).
1489 * When a work thread has completed its work, it sets .working to 0 and
1490 * signals the main thread and waits on the condition that .data_ready
1491 * becomes 1.
1494 struct thread_params {
1495 pthread_t thread;
1496 struct object_entry **list;
1497 unsigned list_size;
1498 unsigned remaining;
1499 int window;
1500 int depth;
1501 int working;
1502 int data_ready;
1503 pthread_mutex_t mutex;
1504 pthread_cond_t cond;
1505 unsigned *processed;
1508 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1510 static void *threaded_find_deltas(void *arg)
1512 struct thread_params *me = arg;
1514 while (me->remaining) {
1515 find_deltas(me->list, &me->remaining,
1516 me->window, me->depth, me->processed);
1518 progress_lock();
1519 me->working = 0;
1520 pthread_cond_signal(&progress_cond);
1521 progress_unlock();
1524 * We must not set ->data_ready before we wait on the
1525 * condition because the main thread may have set it to 1
1526 * before we get here. In order to be sure that new
1527 * work is available if we see 1 in ->data_ready, it
1528 * was initialized to 0 before this thread was spawned
1529 * and we reset it to 0 right away.
1531 pthread_mutex_lock(&me->mutex);
1532 while (!me->data_ready)
1533 pthread_cond_wait(&me->cond, &me->mutex);
1534 me->data_ready = 0;
1535 pthread_mutex_unlock(&me->mutex);
1537 /* leave ->working 1 so that this doesn't get more work assigned */
1538 return NULL;
1541 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1542 int window, int depth, unsigned *processed)
1544 struct thread_params p[delta_search_threads];
1545 int i, ret, active_threads = 0;
1547 if (delta_search_threads <= 1) {
1548 find_deltas(list, &list_size, window, depth, processed);
1549 return;
1552 /* Partition the work amongst work threads. */
1553 for (i = 0; i < delta_search_threads; i++) {
1554 unsigned sub_size = list_size / (delta_search_threads - i);
1556 p[i].window = window;
1557 p[i].depth = depth;
1558 p[i].processed = processed;
1559 p[i].working = 1;
1560 p[i].data_ready = 0;
1562 /* try to split chunks on "path" boundaries */
1563 while (sub_size && sub_size < list_size &&
1564 list[sub_size]->hash &&
1565 list[sub_size]->hash == list[sub_size-1]->hash)
1566 sub_size++;
1568 p[i].list = list;
1569 p[i].list_size = sub_size;
1570 p[i].remaining = sub_size;
1572 list += sub_size;
1573 list_size -= sub_size;
1576 /* Start work threads. */
1577 for (i = 0; i < delta_search_threads; i++) {
1578 if (!p[i].list_size)
1579 continue;
1580 pthread_mutex_init(&p[i].mutex, NULL);
1581 pthread_cond_init(&p[i].cond, NULL);
1582 ret = pthread_create(&p[i].thread, NULL,
1583 threaded_find_deltas, &p[i]);
1584 if (ret)
1585 die("unable to create thread: %s", strerror(ret));
1586 active_threads++;
1590 * Now let's wait for work completion. Each time a thread is done
1591 * with its work, we steal half of the remaining work from the
1592 * thread with the largest number of unprocessed objects and give
1593 * it to that newly idle thread. This ensure good load balancing
1594 * until the remaining object list segments are simply too short
1595 * to be worth splitting anymore.
1597 while (active_threads) {
1598 struct thread_params *target = NULL;
1599 struct thread_params *victim = NULL;
1600 unsigned sub_size = 0;
1602 progress_lock();
1603 for (;;) {
1604 for (i = 0; !target && i < delta_search_threads; i++)
1605 if (!p[i].working)
1606 target = &p[i];
1607 if (target)
1608 break;
1609 pthread_cond_wait(&progress_cond, &progress_mutex);
1612 for (i = 0; i < delta_search_threads; i++)
1613 if (p[i].remaining > 2*window &&
1614 (!victim || victim->remaining < p[i].remaining))
1615 victim = &p[i];
1616 if (victim) {
1617 sub_size = victim->remaining / 2;
1618 list = victim->list + victim->list_size - sub_size;
1619 while (sub_size && list[0]->hash &&
1620 list[0]->hash == list[-1]->hash) {
1621 list++;
1622 sub_size--;
1624 if (!sub_size) {
1626 * It is possible for some "paths" to have
1627 * so many objects that no hash boundary
1628 * might be found. Let's just steal the
1629 * exact half in that case.
1631 sub_size = victim->remaining / 2;
1632 list -= sub_size;
1634 target->list = list;
1635 victim->list_size -= sub_size;
1636 victim->remaining -= sub_size;
1638 target->list_size = sub_size;
1639 target->remaining = sub_size;
1640 target->working = 1;
1641 progress_unlock();
1643 pthread_mutex_lock(&target->mutex);
1644 target->data_ready = 1;
1645 pthread_cond_signal(&target->cond);
1646 pthread_mutex_unlock(&target->mutex);
1648 if (!sub_size) {
1649 pthread_join(target->thread, NULL);
1650 pthread_cond_destroy(&target->cond);
1651 pthread_mutex_destroy(&target->mutex);
1652 active_threads--;
1657 #else
1658 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1659 #endif
1661 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1663 unsigned char peeled[20];
1665 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1666 !peel_ref(path, peeled) && /* peelable? */
1667 !is_null_sha1(peeled) && /* annotated tag? */
1668 locate_object_entry(peeled)) /* object packed? */
1669 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1670 return 0;
1673 static void prepare_pack(int window, int depth)
1675 struct object_entry **delta_list;
1676 uint32_t i, n, nr_deltas;
1678 get_object_details();
1680 if (!nr_objects || !window || !depth)
1681 return;
1683 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1684 nr_deltas = n = 0;
1686 for (i = 0; i < nr_objects; i++) {
1687 struct object_entry *entry = objects + i;
1689 if (entry->delta)
1690 /* This happens if we decided to reuse existing
1691 * delta from a pack. "!no_reuse_delta &&" is implied.
1693 continue;
1695 if (entry->size < 50)
1696 continue;
1698 if (entry->no_try_delta)
1699 continue;
1701 if (!entry->preferred_base)
1702 nr_deltas++;
1704 delta_list[n++] = entry;
1707 if (nr_deltas && n > 1) {
1708 unsigned nr_done = 0;
1709 if (progress)
1710 progress_state = start_progress("Compressing objects",
1711 nr_deltas);
1712 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1713 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1714 stop_progress(&progress_state);
1715 if (nr_done != nr_deltas)
1716 die("inconsistency with delta count");
1718 free(delta_list);
1721 static int git_pack_config(const char *k, const char *v)
1723 if(!strcmp(k, "pack.window")) {
1724 window = git_config_int(k, v);
1725 return 0;
1727 if (!strcmp(k, "pack.windowmemory")) {
1728 window_memory_limit = git_config_ulong(k, v);
1729 return 0;
1731 if (!strcmp(k, "pack.depth")) {
1732 depth = git_config_int(k, v);
1733 return 0;
1735 if (!strcmp(k, "pack.compression")) {
1736 int level = git_config_int(k, v);
1737 if (level == -1)
1738 level = Z_DEFAULT_COMPRESSION;
1739 else if (level < 0 || level > Z_BEST_COMPRESSION)
1740 die("bad pack compression level %d", level);
1741 pack_compression_level = level;
1742 pack_compression_seen = 1;
1743 return 0;
1745 if (!strcmp(k, "pack.deltacachesize")) {
1746 max_delta_cache_size = git_config_int(k, v);
1747 return 0;
1749 if (!strcmp(k, "pack.deltacachelimit")) {
1750 cache_max_small_delta_size = git_config_int(k, v);
1751 return 0;
1753 if (!strcmp(k, "pack.threads")) {
1754 delta_search_threads = git_config_int(k, v);
1755 if (delta_search_threads < 0)
1756 die("invalid number of threads specified (%d)",
1757 delta_search_threads);
1758 #ifndef THREADED_DELTA_SEARCH
1759 if (delta_search_threads != 1)
1760 warning("no threads support, ignoring %s", k);
1761 #endif
1762 return 0;
1764 if (!strcmp(k, "pack.indexversion")) {
1765 pack_idx_default_version = git_config_int(k, v);
1766 if (pack_idx_default_version > 2)
1767 die("bad pack.indexversion=%d", pack_idx_default_version);
1768 return 0;
1770 if (!strcmp(k, "pack.packsizelimit")) {
1771 pack_size_limit_cfg = git_config_ulong(k, v);
1772 return 0;
1774 return git_default_config(k, v);
1777 static void read_object_list_from_stdin(void)
1779 char line[40 + 1 + PATH_MAX + 2];
1780 unsigned char sha1[20];
1782 for (;;) {
1783 if (!fgets(line, sizeof(line), stdin)) {
1784 if (feof(stdin))
1785 break;
1786 if (!ferror(stdin))
1787 die("fgets returned NULL, not EOF, not error!");
1788 if (errno != EINTR)
1789 die("fgets: %s", strerror(errno));
1790 clearerr(stdin);
1791 continue;
1793 if (line[0] == '-') {
1794 if (get_sha1_hex(line+1, sha1))
1795 die("expected edge sha1, got garbage:\n %s",
1796 line);
1797 add_preferred_base(sha1);
1798 continue;
1800 if (get_sha1_hex(line, sha1))
1801 die("expected sha1, got garbage:\n %s", line);
1803 add_preferred_base_object(line+41);
1804 add_object_entry(sha1, 0, line+41, 0);
1808 #define OBJECT_ADDED (1u<<20)
1810 static void show_commit(struct commit *commit)
1812 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1813 commit->object.flags |= OBJECT_ADDED;
1816 static void show_object(struct object_array_entry *p)
1818 add_preferred_base_object(p->name);
1819 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1820 p->item->flags |= OBJECT_ADDED;
1823 static void show_edge(struct commit *commit)
1825 add_preferred_base(commit->object.sha1);
1828 struct in_pack_object {
1829 off_t offset;
1830 struct object *object;
1833 struct in_pack {
1834 int alloc;
1835 int nr;
1836 struct in_pack_object *array;
1839 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1841 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1842 in_pack->array[in_pack->nr].object = object;
1843 in_pack->nr++;
1847 * Compare the objects in the offset order, in order to emulate the
1848 * "git-rev-list --objects" output that produced the pack originally.
1850 static int ofscmp(const void *a_, const void *b_)
1852 struct in_pack_object *a = (struct in_pack_object *)a_;
1853 struct in_pack_object *b = (struct in_pack_object *)b_;
1855 if (a->offset < b->offset)
1856 return -1;
1857 else if (a->offset > b->offset)
1858 return 1;
1859 else
1860 return hashcmp(a->object->sha1, b->object->sha1);
1863 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1865 struct packed_git *p;
1866 struct in_pack in_pack;
1867 uint32_t i;
1869 memset(&in_pack, 0, sizeof(in_pack));
1871 for (p = packed_git; p; p = p->next) {
1872 const unsigned char *sha1;
1873 struct object *o;
1875 for (i = 0; i < revs->num_ignore_packed; i++) {
1876 if (matches_pack_name(p, revs->ignore_packed[i]))
1877 break;
1879 if (revs->num_ignore_packed <= i)
1880 continue;
1881 if (open_pack_index(p))
1882 die("cannot open pack index");
1884 ALLOC_GROW(in_pack.array,
1885 in_pack.nr + p->num_objects,
1886 in_pack.alloc);
1888 for (i = 0; i < p->num_objects; i++) {
1889 sha1 = nth_packed_object_sha1(p, i);
1890 o = lookup_unknown_object(sha1);
1891 if (!(o->flags & OBJECT_ADDED))
1892 mark_in_pack_object(o, p, &in_pack);
1893 o->flags |= OBJECT_ADDED;
1897 if (in_pack.nr) {
1898 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1899 ofscmp);
1900 for (i = 0; i < in_pack.nr; i++) {
1901 struct object *o = in_pack.array[i].object;
1902 add_object_entry(o->sha1, o->type, "", 0);
1905 free(in_pack.array);
1908 static void get_object_list(int ac, const char **av)
1910 struct rev_info revs;
1911 char line[1000];
1912 int flags = 0;
1914 init_revisions(&revs, NULL);
1915 save_commit_buffer = 0;
1916 setup_revisions(ac, av, &revs, NULL);
1918 while (fgets(line, sizeof(line), stdin) != NULL) {
1919 int len = strlen(line);
1920 if (len && line[len - 1] == '\n')
1921 line[--len] = 0;
1922 if (!len)
1923 break;
1924 if (*line == '-') {
1925 if (!strcmp(line, "--not")) {
1926 flags ^= UNINTERESTING;
1927 continue;
1929 die("not a rev '%s'", line);
1931 if (handle_revision_arg(line, &revs, flags, 1))
1932 die("bad revision '%s'", line);
1935 if (prepare_revision_walk(&revs))
1936 die("revision walk setup failed");
1937 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1938 traverse_commit_list(&revs, show_commit, show_object);
1940 if (keep_unreachable)
1941 add_objects_in_unpacked_packs(&revs);
1944 static int adjust_perm(const char *path, mode_t mode)
1946 if (chmod(path, mode))
1947 return -1;
1948 return adjust_shared_perm(path);
1951 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1953 int use_internal_rev_list = 0;
1954 int thin = 0;
1955 uint32_t i;
1956 const char **rp_av;
1957 int rp_ac_alloc = 64;
1958 int rp_ac;
1960 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
1962 rp_av[0] = "pack-objects";
1963 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
1964 rp_ac = 2;
1966 git_config(git_pack_config);
1967 if (!pack_compression_seen && core_compression_seen)
1968 pack_compression_level = core_compression_level;
1970 progress = isatty(2);
1971 for (i = 1; i < argc; i++) {
1972 const char *arg = argv[i];
1974 if (*arg != '-')
1975 break;
1977 if (!strcmp("--non-empty", arg)) {
1978 non_empty = 1;
1979 continue;
1981 if (!strcmp("--local", arg)) {
1982 local = 1;
1983 continue;
1985 if (!strcmp("--incremental", arg)) {
1986 incremental = 1;
1987 continue;
1989 if (!prefixcmp(arg, "--compression=")) {
1990 char *end;
1991 int level = strtoul(arg+14, &end, 0);
1992 if (!arg[14] || *end)
1993 usage(pack_usage);
1994 if (level == -1)
1995 level = Z_DEFAULT_COMPRESSION;
1996 else if (level < 0 || level > Z_BEST_COMPRESSION)
1997 die("bad pack compression level %d", level);
1998 pack_compression_level = level;
1999 continue;
2001 if (!prefixcmp(arg, "--max-pack-size=")) {
2002 char *end;
2003 pack_size_limit_cfg = 0;
2004 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2005 if (!arg[16] || *end)
2006 usage(pack_usage);
2007 continue;
2009 if (!prefixcmp(arg, "--window=")) {
2010 char *end;
2011 window = strtoul(arg+9, &end, 0);
2012 if (!arg[9] || *end)
2013 usage(pack_usage);
2014 continue;
2016 if (!prefixcmp(arg, "--window-memory=")) {
2017 if (!git_parse_ulong(arg+16, &window_memory_limit))
2018 usage(pack_usage);
2019 continue;
2021 if (!prefixcmp(arg, "--threads=")) {
2022 char *end;
2023 delta_search_threads = strtoul(arg+10, &end, 0);
2024 if (!arg[10] || *end || delta_search_threads < 0)
2025 usage(pack_usage);
2026 #ifndef THREADED_DELTA_SEARCH
2027 if (delta_search_threads != 1)
2028 warning("no threads support, "
2029 "ignoring %s", arg);
2030 #endif
2031 continue;
2033 if (!prefixcmp(arg, "--depth=")) {
2034 char *end;
2035 depth = strtoul(arg+8, &end, 0);
2036 if (!arg[8] || *end)
2037 usage(pack_usage);
2038 continue;
2040 if (!strcmp("--progress", arg)) {
2041 progress = 1;
2042 continue;
2044 if (!strcmp("--all-progress", arg)) {
2045 progress = 2;
2046 continue;
2048 if (!strcmp("-q", arg)) {
2049 progress = 0;
2050 continue;
2052 if (!strcmp("--no-reuse-delta", arg)) {
2053 no_reuse_delta = 1;
2054 continue;
2056 if (!strcmp("--no-reuse-object", arg)) {
2057 no_reuse_object = no_reuse_delta = 1;
2058 continue;
2060 if (!strcmp("--delta-base-offset", arg)) {
2061 allow_ofs_delta = 1;
2062 continue;
2064 if (!strcmp("--stdout", arg)) {
2065 pack_to_stdout = 1;
2066 continue;
2068 if (!strcmp("--revs", arg)) {
2069 use_internal_rev_list = 1;
2070 continue;
2072 if (!strcmp("--keep-unreachable", arg)) {
2073 keep_unreachable = 1;
2074 continue;
2076 if (!strcmp("--include-tag", arg)) {
2077 include_tag = 1;
2078 continue;
2080 if (!strcmp("--unpacked", arg) ||
2081 !prefixcmp(arg, "--unpacked=") ||
2082 !strcmp("--reflog", arg) ||
2083 !strcmp("--all", arg)) {
2084 use_internal_rev_list = 1;
2085 if (rp_ac >= rp_ac_alloc - 1) {
2086 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2087 rp_av = xrealloc(rp_av,
2088 rp_ac_alloc * sizeof(*rp_av));
2090 rp_av[rp_ac++] = arg;
2091 continue;
2093 if (!strcmp("--thin", arg)) {
2094 use_internal_rev_list = 1;
2095 thin = 1;
2096 rp_av[1] = "--objects-edge";
2097 continue;
2099 if (!prefixcmp(arg, "--index-version=")) {
2100 char *c;
2101 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2102 if (pack_idx_default_version > 2)
2103 die("bad %s", arg);
2104 if (*c == ',')
2105 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2106 if (*c || pack_idx_off32_limit & 0x80000000)
2107 die("bad %s", arg);
2108 continue;
2110 usage(pack_usage);
2113 /* Traditionally "pack-objects [options] base extra" failed;
2114 * we would however want to take refs parameter that would
2115 * have been given to upstream rev-list ourselves, which means
2116 * we somehow want to say what the base name is. So the
2117 * syntax would be:
2119 * pack-objects [options] base <refs...>
2121 * in other words, we would treat the first non-option as the
2122 * base_name and send everything else to the internal revision
2123 * walker.
2126 if (!pack_to_stdout)
2127 base_name = argv[i++];
2129 if (pack_to_stdout != !base_name)
2130 usage(pack_usage);
2132 if (!pack_to_stdout && !pack_size_limit)
2133 pack_size_limit = pack_size_limit_cfg;
2135 if (pack_to_stdout && pack_size_limit)
2136 die("--max-pack-size cannot be used to build a pack for transfer.");
2138 if (!pack_to_stdout && thin)
2139 die("--thin cannot be used to build an indexable pack.");
2141 #ifdef THREADED_DELTA_SEARCH
2142 if (!delta_search_threads) /* --threads=0 means autodetect */
2143 delta_search_threads = online_cpus();
2144 #endif
2146 prepare_packed_git();
2148 if (progress)
2149 progress_state = start_progress("Counting objects", 0);
2150 if (!use_internal_rev_list)
2151 read_object_list_from_stdin();
2152 else {
2153 rp_av[rp_ac] = NULL;
2154 get_object_list(rp_ac, rp_av);
2156 if (include_tag && nr_result)
2157 for_each_ref(add_ref_tag, NULL);
2158 stop_progress(&progress_state);
2160 if (non_empty && !nr_result)
2161 return 0;
2162 if (nr_result)
2163 prepare_pack(window, depth);
2164 write_pack_file();
2165 if (progress)
2166 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2167 written, written_delta, reused, reused_delta);
2168 return 0;