Merge branch 'bc/repack'
[git/git-svn.git] / builtin-pack-objects.c
blobf43eb67016ee719d34d2e2f4ae8d77e1ec74b396
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
20 #ifdef THREADED_DELTA_SEARCH
21 #include "thread-utils.h"
22 #include <pthread.h>
23 #endif
25 static const char pack_usage[] = "\
26 git-pack-objects [{ -q | --progress | --all-progress }] \n\
27 [--max-pack-size=N] [--local] [--incremental] \n\
28 [--window=N] [--window-memory=N] [--depth=N] \n\
29 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
30 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
31 [--stdout | base-name] [--include-tag] \n\
32 [--keep-unreachable | --unpack-unreachable] \n\
33 [<ref-list | <object-list]";
35 struct object_entry {
36 struct pack_idx_entry idx;
37 unsigned long size; /* uncompressed size */
38 struct packed_git *in_pack; /* already in pack */
39 off_t in_pack_offset;
40 struct object_entry *delta; /* delta base object */
41 struct object_entry *delta_child; /* deltified objects who bases me */
42 struct object_entry *delta_sibling; /* other deltified objects who
43 * uses the same base as me
45 void *delta_data; /* cached delta (uncompressed) */
46 unsigned long delta_size; /* delta data size (uncompressed) */
47 unsigned long z_delta_size; /* delta data size (compressed) */
48 unsigned int hash; /* name hint hash */
49 enum object_type type;
50 enum object_type in_pack_type; /* could be delta */
51 unsigned char in_pack_header_size;
52 unsigned char preferred_base; /* we do not pack this, but is available
53 * to be used as the base object to delta
54 * objects against.
56 unsigned char no_try_delta;
60 * Objects we are going to pack are collected in objects array (dynamically
61 * expanded). nr_objects & nr_alloc controls this array. They are stored
62 * in the order we see -- typically rev-list --objects order that gives us
63 * nice "minimum seek" order.
65 static struct object_entry *objects;
66 static struct pack_idx_entry **written_list;
67 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
69 static int non_empty;
70 static int reuse_delta = 1, reuse_object = 1;
71 static int keep_unreachable, unpack_unreachable, include_tag;
72 static int local;
73 static int incremental;
74 static int allow_ofs_delta;
75 static const char *base_name;
76 static int progress = 1;
77 static int window = 10;
78 static uint32_t pack_size_limit, pack_size_limit_cfg;
79 static int depth = 50;
80 static int delta_search_threads = 1;
81 static int pack_to_stdout;
82 static int num_preferred_base;
83 static struct progress *progress_state;
84 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
85 static int pack_compression_seen;
87 static unsigned long delta_cache_size = 0;
88 static unsigned long max_delta_cache_size = 0;
89 static unsigned long cache_max_small_delta_size = 1000;
91 static unsigned long window_memory_limit = 0;
94 * The object names in objects array are hashed with this hashtable,
95 * to help looking up the entry by object name.
96 * This hashtable is built after all the objects are seen.
98 static int *object_ix;
99 static int object_ix_hashsz;
102 * stats
104 static uint32_t written, written_delta;
105 static uint32_t reused, reused_delta;
108 static void *get_delta(struct object_entry *entry)
110 unsigned long size, base_size, delta_size;
111 void *buf, *base_buf, *delta_buf;
112 enum object_type type;
114 buf = read_sha1_file(entry->idx.sha1, &type, &size);
115 if (!buf)
116 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
117 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
118 if (!base_buf)
119 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
120 delta_buf = diff_delta(base_buf, base_size,
121 buf, size, &delta_size, 0);
122 if (!delta_buf || delta_size != entry->delta_size)
123 die("delta size changed");
124 free(buf);
125 free(base_buf);
126 return delta_buf;
129 static unsigned long do_compress(void **pptr, unsigned long size)
131 z_stream stream;
132 void *in, *out;
133 unsigned long maxsize;
135 memset(&stream, 0, sizeof(stream));
136 deflateInit(&stream, pack_compression_level);
137 maxsize = deflateBound(&stream, size);
139 in = *pptr;
140 out = xmalloc(maxsize);
141 *pptr = out;
143 stream.next_in = in;
144 stream.avail_in = size;
145 stream.next_out = out;
146 stream.avail_out = maxsize;
147 while (deflate(&stream, Z_FINISH) == Z_OK)
148 ; /* nothing */
149 deflateEnd(&stream);
151 free(in);
152 return stream.total_out;
156 * The per-object header is a pretty dense thing, which is
157 * - first byte: low four bits are "size", then three bits of "type",
158 * and the high bit is "size continues".
159 * - each byte afterwards: low seven bits are size continuation,
160 * with the high bit being "size continues"
162 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
164 int n = 1;
165 unsigned char c;
167 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
168 die("bad type %d", type);
170 c = (type << 4) | (size & 15);
171 size >>= 4;
172 while (size) {
173 *hdr++ = c | 0x80;
174 c = size & 0x7f;
175 size >>= 7;
176 n++;
178 *hdr = c;
179 return n;
183 * we are going to reuse the existing object data as is. make
184 * sure it is not corrupt.
186 static int check_pack_inflate(struct packed_git *p,
187 struct pack_window **w_curs,
188 off_t offset,
189 off_t len,
190 unsigned long expect)
192 z_stream stream;
193 unsigned char fakebuf[4096], *in;
194 int st;
196 memset(&stream, 0, sizeof(stream));
197 inflateInit(&stream);
198 do {
199 in = use_pack(p, w_curs, offset, &stream.avail_in);
200 stream.next_in = in;
201 stream.next_out = fakebuf;
202 stream.avail_out = sizeof(fakebuf);
203 st = inflate(&stream, Z_FINISH);
204 offset += stream.next_in - in;
205 } while (st == Z_OK || st == Z_BUF_ERROR);
206 inflateEnd(&stream);
207 return (st == Z_STREAM_END &&
208 stream.total_out == expect &&
209 stream.total_in == len) ? 0 : -1;
212 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
213 off_t offset, off_t len, unsigned int nr)
215 const uint32_t *index_crc;
216 uint32_t data_crc = crc32(0, Z_NULL, 0);
218 do {
219 unsigned int avail;
220 void *data = use_pack(p, w_curs, offset, &avail);
221 if (avail > len)
222 avail = len;
223 data_crc = crc32(data_crc, data, avail);
224 offset += avail;
225 len -= avail;
226 } while (len);
228 index_crc = p->index_data;
229 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
231 return data_crc != ntohl(*index_crc);
234 static void copy_pack_data(struct sha1file *f,
235 struct packed_git *p,
236 struct pack_window **w_curs,
237 off_t offset,
238 off_t len)
240 unsigned char *in;
241 unsigned int avail;
243 while (len) {
244 in = use_pack(p, w_curs, offset, &avail);
245 if (avail > len)
246 avail = (unsigned int)len;
247 sha1write(f, in, avail);
248 offset += avail;
249 len -= avail;
253 static unsigned long write_object(struct sha1file *f,
254 struct object_entry *entry,
255 off_t write_offset)
257 unsigned long size, limit, datalen;
258 void *buf;
259 unsigned char header[10], dheader[10];
260 unsigned hdrlen;
261 enum object_type type;
262 int usable_delta, to_reuse;
264 if (!pack_to_stdout)
265 crc32_begin(f);
267 type = entry->type;
269 /* write limit if limited packsize and not first object */
270 limit = pack_size_limit && nr_written ?
271 pack_size_limit - write_offset : 0;
273 if (!entry->delta)
274 usable_delta = 0; /* no delta */
275 else if (!pack_size_limit)
276 usable_delta = 1; /* unlimited packfile */
277 else if (entry->delta->idx.offset == (off_t)-1)
278 usable_delta = 0; /* base was written to another pack */
279 else if (entry->delta->idx.offset)
280 usable_delta = 1; /* base already exists in this pack */
281 else
282 usable_delta = 0; /* base could end up in another pack */
284 if (!reuse_object)
285 to_reuse = 0; /* explicit */
286 else if (!entry->in_pack)
287 to_reuse = 0; /* can't reuse what we don't have */
288 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
289 /* check_object() decided it for us ... */
290 to_reuse = usable_delta;
291 /* ... but pack split may override that */
292 else if (type != entry->in_pack_type)
293 to_reuse = 0; /* pack has delta which is unusable */
294 else if (entry->delta)
295 to_reuse = 0; /* we want to pack afresh */
296 else
297 to_reuse = 1; /* we have it in-pack undeltified,
298 * and we do not need to deltify it.
301 if (!to_reuse) {
302 if (!usable_delta) {
303 buf = read_sha1_file(entry->idx.sha1, &type, &size);
304 if (!buf)
305 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
307 * make sure no cached delta data remains from a
308 * previous attempt before a pack split occured.
310 free(entry->delta_data);
311 entry->delta_data = NULL;
312 entry->z_delta_size = 0;
313 } else if (entry->delta_data) {
314 size = entry->delta_size;
315 buf = entry->delta_data;
316 entry->delta_data = NULL;
317 type = (allow_ofs_delta && entry->delta->idx.offset) ?
318 OBJ_OFS_DELTA : OBJ_REF_DELTA;
319 } else {
320 buf = get_delta(entry);
321 size = entry->delta_size;
322 type = (allow_ofs_delta && entry->delta->idx.offset) ?
323 OBJ_OFS_DELTA : OBJ_REF_DELTA;
326 if (entry->z_delta_size)
327 datalen = entry->z_delta_size;
328 else
329 datalen = do_compress(&buf, size);
332 * The object header is a byte of 'type' followed by zero or
333 * more bytes of length.
335 hdrlen = encode_header(type, size, header);
337 if (type == OBJ_OFS_DELTA) {
339 * Deltas with relative base contain an additional
340 * encoding of the relative offset for the delta
341 * base from this object's position in the pack.
343 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
344 unsigned pos = sizeof(dheader) - 1;
345 dheader[pos] = ofs & 127;
346 while (ofs >>= 7)
347 dheader[--pos] = 128 | (--ofs & 127);
348 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
349 free(buf);
350 return 0;
352 sha1write(f, header, hdrlen);
353 sha1write(f, dheader + pos, sizeof(dheader) - pos);
354 hdrlen += sizeof(dheader) - pos;
355 } else if (type == OBJ_REF_DELTA) {
357 * Deltas with a base reference contain
358 * an additional 20 bytes for the base sha1.
360 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
361 free(buf);
362 return 0;
364 sha1write(f, header, hdrlen);
365 sha1write(f, entry->delta->idx.sha1, 20);
366 hdrlen += 20;
367 } else {
368 if (limit && hdrlen + datalen + 20 >= limit) {
369 free(buf);
370 return 0;
372 sha1write(f, header, hdrlen);
374 sha1write(f, buf, datalen);
375 free(buf);
377 else {
378 struct packed_git *p = entry->in_pack;
379 struct pack_window *w_curs = NULL;
380 struct revindex_entry *revidx;
381 off_t offset;
383 if (entry->delta) {
384 type = (allow_ofs_delta && entry->delta->idx.offset) ?
385 OBJ_OFS_DELTA : OBJ_REF_DELTA;
386 reused_delta++;
388 hdrlen = encode_header(type, entry->size, header);
389 offset = entry->in_pack_offset;
390 revidx = find_pack_revindex(p, offset);
391 datalen = revidx[1].offset - offset;
392 if (!pack_to_stdout && p->index_version > 1 &&
393 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
394 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
395 offset += entry->in_pack_header_size;
396 datalen -= entry->in_pack_header_size;
397 if (type == OBJ_OFS_DELTA) {
398 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
399 unsigned pos = sizeof(dheader) - 1;
400 dheader[pos] = ofs & 127;
401 while (ofs >>= 7)
402 dheader[--pos] = 128 | (--ofs & 127);
403 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
404 return 0;
405 sha1write(f, header, hdrlen);
406 sha1write(f, dheader + pos, sizeof(dheader) - pos);
407 hdrlen += sizeof(dheader) - pos;
408 } else if (type == OBJ_REF_DELTA) {
409 if (limit && hdrlen + 20 + datalen + 20 >= limit)
410 return 0;
411 sha1write(f, header, hdrlen);
412 sha1write(f, entry->delta->idx.sha1, 20);
413 hdrlen += 20;
414 } else {
415 if (limit && hdrlen + datalen + 20 >= limit)
416 return 0;
417 sha1write(f, header, hdrlen);
420 if (!pack_to_stdout && p->index_version == 1 &&
421 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
422 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
423 copy_pack_data(f, p, &w_curs, offset, datalen);
424 unuse_pack(&w_curs);
425 reused++;
427 if (usable_delta)
428 written_delta++;
429 written++;
430 if (!pack_to_stdout)
431 entry->idx.crc32 = crc32_end(f);
432 return hdrlen + datalen;
435 static off_t write_one(struct sha1file *f,
436 struct object_entry *e,
437 off_t offset)
439 unsigned long size;
441 /* offset is non zero if object is written already. */
442 if (e->idx.offset || e->preferred_base)
443 return offset;
445 /* if we are deltified, write out base object first. */
446 if (e->delta) {
447 offset = write_one(f, e->delta, offset);
448 if (!offset)
449 return 0;
452 e->idx.offset = offset;
453 size = write_object(f, e, offset);
454 if (!size) {
455 e->idx.offset = 0;
456 return 0;
458 written_list[nr_written++] = &e->idx;
460 /* make sure off_t is sufficiently large not to wrap */
461 if (offset > offset + size)
462 die("pack too large for current definition of off_t");
463 return offset + size;
466 /* forward declaration for write_pack_file */
467 static int adjust_perm(const char *path, mode_t mode);
469 static void write_pack_file(void)
471 uint32_t i = 0, j;
472 struct sha1file *f;
473 off_t offset, offset_one, last_obj_offset = 0;
474 struct pack_header hdr;
475 uint32_t nr_remaining = nr_result;
476 time_t last_mtime = 0;
478 if (progress > pack_to_stdout)
479 progress_state = start_progress("Writing objects", nr_result);
480 written_list = xmalloc(nr_objects * sizeof(*written_list));
482 do {
483 unsigned char sha1[20];
484 char *pack_tmp_name = NULL;
486 if (pack_to_stdout) {
487 f = sha1fd_throughput(1, "<stdout>", progress_state);
488 } else {
489 char tmpname[PATH_MAX];
490 int fd;
491 snprintf(tmpname, sizeof(tmpname),
492 "%s/tmp_pack_XXXXXX", get_object_directory());
493 fd = xmkstemp(tmpname);
494 pack_tmp_name = xstrdup(tmpname);
495 f = sha1fd(fd, pack_tmp_name);
498 hdr.hdr_signature = htonl(PACK_SIGNATURE);
499 hdr.hdr_version = htonl(PACK_VERSION);
500 hdr.hdr_entries = htonl(nr_remaining);
501 sha1write(f, &hdr, sizeof(hdr));
502 offset = sizeof(hdr);
503 nr_written = 0;
504 for (; i < nr_objects; i++) {
505 last_obj_offset = offset;
506 offset_one = write_one(f, objects + i, offset);
507 if (!offset_one)
508 break;
509 offset = offset_one;
510 display_progress(progress_state, written);
514 * Did we write the wrong # entries in the header?
515 * If so, rewrite it like in fast-import
517 if (pack_to_stdout || nr_written == nr_remaining) {
518 sha1close(f, sha1, 1);
519 } else {
520 int fd = sha1close(f, NULL, 0);
521 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
522 close(fd);
525 if (!pack_to_stdout) {
526 mode_t mode = umask(0);
527 struct stat st;
528 char *idx_tmp_name, tmpname[PATH_MAX];
530 umask(mode);
531 mode = 0444 & ~mode;
533 idx_tmp_name = write_idx_file(NULL, written_list,
534 nr_written, sha1);
536 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
537 base_name, sha1_to_hex(sha1));
538 if (adjust_perm(pack_tmp_name, mode))
539 die("unable to make temporary pack file readable: %s",
540 strerror(errno));
541 if (rename(pack_tmp_name, tmpname))
542 die("unable to rename temporary pack file: %s",
543 strerror(errno));
546 * Packs are runtime accessed in their mtime
547 * order since newer packs are more likely to contain
548 * younger objects. So if we are creating multiple
549 * packs then we should modify the mtime of later ones
550 * to preserve this property.
552 if (stat(tmpname, &st) < 0) {
553 warning("failed to stat %s: %s",
554 tmpname, strerror(errno));
555 } else if (!last_mtime) {
556 last_mtime = st.st_mtime;
557 } else {
558 struct utimbuf utb;
559 utb.actime = st.st_atime;
560 utb.modtime = --last_mtime;
561 if (utime(tmpname, &utb) < 0)
562 warning("failed utime() on %s: %s",
563 tmpname, strerror(errno));
566 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
567 base_name, sha1_to_hex(sha1));
568 if (adjust_perm(idx_tmp_name, mode))
569 die("unable to make temporary index file readable: %s",
570 strerror(errno));
571 if (rename(idx_tmp_name, tmpname))
572 die("unable to rename temporary index file: %s",
573 strerror(errno));
575 free(idx_tmp_name);
576 free(pack_tmp_name);
577 puts(sha1_to_hex(sha1));
580 /* mark written objects as written to previous pack */
581 for (j = 0; j < nr_written; j++) {
582 written_list[j]->offset = (off_t)-1;
584 nr_remaining -= nr_written;
585 } while (nr_remaining && i < nr_objects);
587 free(written_list);
588 stop_progress(&progress_state);
589 if (written != nr_result)
590 die("wrote %u objects while expecting %u", written, nr_result);
592 * We have scanned through [0 ... i). Since we have written
593 * the correct number of objects, the remaining [i ... nr_objects)
594 * items must be either already written (due to out-of-order delta base)
595 * or a preferred base. Count those which are neither and complain if any.
597 for (j = 0; i < nr_objects; i++) {
598 struct object_entry *e = objects + i;
599 j += !e->idx.offset && !e->preferred_base;
601 if (j)
602 die("wrote %u objects as expected but %u unwritten", written, j);
605 static int locate_object_entry_hash(const unsigned char *sha1)
607 int i;
608 unsigned int ui;
609 memcpy(&ui, sha1, sizeof(unsigned int));
610 i = ui % object_ix_hashsz;
611 while (0 < object_ix[i]) {
612 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
613 return i;
614 if (++i == object_ix_hashsz)
615 i = 0;
617 return -1 - i;
620 static struct object_entry *locate_object_entry(const unsigned char *sha1)
622 int i;
624 if (!object_ix_hashsz)
625 return NULL;
627 i = locate_object_entry_hash(sha1);
628 if (0 <= i)
629 return &objects[object_ix[i]-1];
630 return NULL;
633 static void rehash_objects(void)
635 uint32_t i;
636 struct object_entry *oe;
638 object_ix_hashsz = nr_objects * 3;
639 if (object_ix_hashsz < 1024)
640 object_ix_hashsz = 1024;
641 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
642 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
643 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
644 int ix = locate_object_entry_hash(oe->idx.sha1);
645 if (0 <= ix)
646 continue;
647 ix = -1 - ix;
648 object_ix[ix] = i + 1;
652 static unsigned name_hash(const char *name)
654 unsigned char c;
655 unsigned hash = 0;
657 if (!name)
658 return 0;
661 * This effectively just creates a sortable number from the
662 * last sixteen non-whitespace characters. Last characters
663 * count "most", so things that end in ".c" sort together.
665 while ((c = *name++) != 0) {
666 if (isspace(c))
667 continue;
668 hash = (hash >> 2) + (c << 24);
670 return hash;
673 static void setup_delta_attr_check(struct git_attr_check *check)
675 static struct git_attr *attr_delta;
677 if (!attr_delta)
678 attr_delta = git_attr("delta", 5);
680 check[0].attr = attr_delta;
683 static int no_try_delta(const char *path)
685 struct git_attr_check check[1];
687 setup_delta_attr_check(check);
688 if (git_checkattr(path, ARRAY_SIZE(check), check))
689 return 0;
690 if (ATTR_FALSE(check->value))
691 return 1;
692 return 0;
695 static int add_object_entry(const unsigned char *sha1, enum object_type type,
696 const char *name, int exclude)
698 struct object_entry *entry;
699 struct packed_git *p, *found_pack = NULL;
700 off_t found_offset = 0;
701 int ix;
702 unsigned hash = name_hash(name);
704 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
705 if (ix >= 0) {
706 if (exclude) {
707 entry = objects + object_ix[ix] - 1;
708 if (!entry->preferred_base)
709 nr_result--;
710 entry->preferred_base = 1;
712 return 0;
715 for (p = packed_git; p; p = p->next) {
716 off_t offset = find_pack_entry_one(sha1, p);
717 if (offset) {
718 if (!found_pack) {
719 found_offset = offset;
720 found_pack = p;
722 if (exclude)
723 break;
724 if (incremental)
725 return 0;
726 if (local && !p->pack_local)
727 return 0;
731 if (nr_objects >= nr_alloc) {
732 nr_alloc = (nr_alloc + 1024) * 3 / 2;
733 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
736 entry = objects + nr_objects++;
737 memset(entry, 0, sizeof(*entry));
738 hashcpy(entry->idx.sha1, sha1);
739 entry->hash = hash;
740 if (type)
741 entry->type = type;
742 if (exclude)
743 entry->preferred_base = 1;
744 else
745 nr_result++;
746 if (found_pack) {
747 entry->in_pack = found_pack;
748 entry->in_pack_offset = found_offset;
751 if (object_ix_hashsz * 3 <= nr_objects * 4)
752 rehash_objects();
753 else
754 object_ix[-1 - ix] = nr_objects;
756 display_progress(progress_state, nr_objects);
758 if (name && no_try_delta(name))
759 entry->no_try_delta = 1;
761 return 1;
764 struct pbase_tree_cache {
765 unsigned char sha1[20];
766 int ref;
767 int temporary;
768 void *tree_data;
769 unsigned long tree_size;
772 static struct pbase_tree_cache *(pbase_tree_cache[256]);
773 static int pbase_tree_cache_ix(const unsigned char *sha1)
775 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
777 static int pbase_tree_cache_ix_incr(int ix)
779 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
782 static struct pbase_tree {
783 struct pbase_tree *next;
784 /* This is a phony "cache" entry; we are not
785 * going to evict it nor find it through _get()
786 * mechanism -- this is for the toplevel node that
787 * would almost always change with any commit.
789 struct pbase_tree_cache pcache;
790 } *pbase_tree;
792 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
794 struct pbase_tree_cache *ent, *nent;
795 void *data;
796 unsigned long size;
797 enum object_type type;
798 int neigh;
799 int my_ix = pbase_tree_cache_ix(sha1);
800 int available_ix = -1;
802 /* pbase-tree-cache acts as a limited hashtable.
803 * your object will be found at your index or within a few
804 * slots after that slot if it is cached.
806 for (neigh = 0; neigh < 8; neigh++) {
807 ent = pbase_tree_cache[my_ix];
808 if (ent && !hashcmp(ent->sha1, sha1)) {
809 ent->ref++;
810 return ent;
812 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
813 ((0 <= available_ix) &&
814 (!ent && pbase_tree_cache[available_ix])))
815 available_ix = my_ix;
816 if (!ent)
817 break;
818 my_ix = pbase_tree_cache_ix_incr(my_ix);
821 /* Did not find one. Either we got a bogus request or
822 * we need to read and perhaps cache.
824 data = read_sha1_file(sha1, &type, &size);
825 if (!data)
826 return NULL;
827 if (type != OBJ_TREE) {
828 free(data);
829 return NULL;
832 /* We need to either cache or return a throwaway copy */
834 if (available_ix < 0)
835 ent = NULL;
836 else {
837 ent = pbase_tree_cache[available_ix];
838 my_ix = available_ix;
841 if (!ent) {
842 nent = xmalloc(sizeof(*nent));
843 nent->temporary = (available_ix < 0);
845 else {
846 /* evict and reuse */
847 free(ent->tree_data);
848 nent = ent;
850 hashcpy(nent->sha1, sha1);
851 nent->tree_data = data;
852 nent->tree_size = size;
853 nent->ref = 1;
854 if (!nent->temporary)
855 pbase_tree_cache[my_ix] = nent;
856 return nent;
859 static void pbase_tree_put(struct pbase_tree_cache *cache)
861 if (!cache->temporary) {
862 cache->ref--;
863 return;
865 free(cache->tree_data);
866 free(cache);
869 static int name_cmp_len(const char *name)
871 int i;
872 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
874 return i;
877 static void add_pbase_object(struct tree_desc *tree,
878 const char *name,
879 int cmplen,
880 const char *fullname)
882 struct name_entry entry;
883 int cmp;
885 while (tree_entry(tree,&entry)) {
886 if (S_ISGITLINK(entry.mode))
887 continue;
888 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
889 memcmp(name, entry.path, cmplen);
890 if (cmp > 0)
891 continue;
892 if (cmp < 0)
893 return;
894 if (name[cmplen] != '/') {
895 add_object_entry(entry.sha1,
896 object_type(entry.mode),
897 fullname, 1);
898 return;
900 if (S_ISDIR(entry.mode)) {
901 struct tree_desc sub;
902 struct pbase_tree_cache *tree;
903 const char *down = name+cmplen+1;
904 int downlen = name_cmp_len(down);
906 tree = pbase_tree_get(entry.sha1);
907 if (!tree)
908 return;
909 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
911 add_pbase_object(&sub, down, downlen, fullname);
912 pbase_tree_put(tree);
917 static unsigned *done_pbase_paths;
918 static int done_pbase_paths_num;
919 static int done_pbase_paths_alloc;
920 static int done_pbase_path_pos(unsigned hash)
922 int lo = 0;
923 int hi = done_pbase_paths_num;
924 while (lo < hi) {
925 int mi = (hi + lo) / 2;
926 if (done_pbase_paths[mi] == hash)
927 return mi;
928 if (done_pbase_paths[mi] < hash)
929 hi = mi;
930 else
931 lo = mi + 1;
933 return -lo-1;
936 static int check_pbase_path(unsigned hash)
938 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
939 if (0 <= pos)
940 return 1;
941 pos = -pos - 1;
942 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
943 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
944 done_pbase_paths = xrealloc(done_pbase_paths,
945 done_pbase_paths_alloc *
946 sizeof(unsigned));
948 done_pbase_paths_num++;
949 if (pos < done_pbase_paths_num)
950 memmove(done_pbase_paths + pos + 1,
951 done_pbase_paths + pos,
952 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
953 done_pbase_paths[pos] = hash;
954 return 0;
957 static void add_preferred_base_object(const char *name)
959 struct pbase_tree *it;
960 int cmplen;
961 unsigned hash = name_hash(name);
963 if (!num_preferred_base || check_pbase_path(hash))
964 return;
966 cmplen = name_cmp_len(name);
967 for (it = pbase_tree; it; it = it->next) {
968 if (cmplen == 0) {
969 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
971 else {
972 struct tree_desc tree;
973 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
974 add_pbase_object(&tree, name, cmplen, name);
979 static void add_preferred_base(unsigned char *sha1)
981 struct pbase_tree *it;
982 void *data;
983 unsigned long size;
984 unsigned char tree_sha1[20];
986 if (window <= num_preferred_base++)
987 return;
989 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
990 if (!data)
991 return;
993 for (it = pbase_tree; it; it = it->next) {
994 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
995 free(data);
996 return;
1000 it = xcalloc(1, sizeof(*it));
1001 it->next = pbase_tree;
1002 pbase_tree = it;
1004 hashcpy(it->pcache.sha1, tree_sha1);
1005 it->pcache.tree_data = data;
1006 it->pcache.tree_size = size;
1009 static void check_object(struct object_entry *entry)
1011 if (entry->in_pack) {
1012 struct packed_git *p = entry->in_pack;
1013 struct pack_window *w_curs = NULL;
1014 const unsigned char *base_ref = NULL;
1015 struct object_entry *base_entry;
1016 unsigned long used, used_0;
1017 unsigned int avail;
1018 off_t ofs;
1019 unsigned char *buf, c;
1021 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1024 * We want in_pack_type even if we do not reuse delta
1025 * since non-delta representations could still be reused.
1027 used = unpack_object_header_gently(buf, avail,
1028 &entry->in_pack_type,
1029 &entry->size);
1032 * Determine if this is a delta and if so whether we can
1033 * reuse it or not. Otherwise let's find out as cheaply as
1034 * possible what the actual type and size for this object is.
1036 switch (entry->in_pack_type) {
1037 default:
1038 /* Not a delta hence we've already got all we need. */
1039 entry->type = entry->in_pack_type;
1040 entry->in_pack_header_size = used;
1041 unuse_pack(&w_curs);
1042 return;
1043 case OBJ_REF_DELTA:
1044 if (reuse_delta && !entry->preferred_base)
1045 base_ref = use_pack(p, &w_curs,
1046 entry->in_pack_offset + used, NULL);
1047 entry->in_pack_header_size = used + 20;
1048 break;
1049 case OBJ_OFS_DELTA:
1050 buf = use_pack(p, &w_curs,
1051 entry->in_pack_offset + used, NULL);
1052 used_0 = 0;
1053 c = buf[used_0++];
1054 ofs = c & 127;
1055 while (c & 128) {
1056 ofs += 1;
1057 if (!ofs || MSB(ofs, 7))
1058 die("delta base offset overflow in pack for %s",
1059 sha1_to_hex(entry->idx.sha1));
1060 c = buf[used_0++];
1061 ofs = (ofs << 7) + (c & 127);
1063 if (ofs >= entry->in_pack_offset)
1064 die("delta base offset out of bound for %s",
1065 sha1_to_hex(entry->idx.sha1));
1066 ofs = entry->in_pack_offset - ofs;
1067 if (reuse_delta && !entry->preferred_base) {
1068 struct revindex_entry *revidx;
1069 revidx = find_pack_revindex(p, ofs);
1070 base_ref = nth_packed_object_sha1(p, revidx->nr);
1072 entry->in_pack_header_size = used + used_0;
1073 break;
1076 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1078 * If base_ref was set above that means we wish to
1079 * reuse delta data, and we even found that base
1080 * in the list of objects we want to pack. Goodie!
1082 * Depth value does not matter - find_deltas() will
1083 * never consider reused delta as the base object to
1084 * deltify other objects against, in order to avoid
1085 * circular deltas.
1087 entry->type = entry->in_pack_type;
1088 entry->delta = base_entry;
1089 entry->delta_sibling = base_entry->delta_child;
1090 base_entry->delta_child = entry;
1091 unuse_pack(&w_curs);
1092 return;
1095 if (entry->type) {
1097 * This must be a delta and we already know what the
1098 * final object type is. Let's extract the actual
1099 * object size from the delta header.
1101 entry->size = get_size_from_delta(p, &w_curs,
1102 entry->in_pack_offset + entry->in_pack_header_size);
1103 unuse_pack(&w_curs);
1104 return;
1108 * No choice but to fall back to the recursive delta walk
1109 * with sha1_object_info() to find about the object type
1110 * at this point...
1112 unuse_pack(&w_curs);
1115 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1116 if (entry->type < 0)
1117 die("unable to get type of object %s",
1118 sha1_to_hex(entry->idx.sha1));
1121 static int pack_offset_sort(const void *_a, const void *_b)
1123 const struct object_entry *a = *(struct object_entry **)_a;
1124 const struct object_entry *b = *(struct object_entry **)_b;
1126 /* avoid filesystem trashing with loose objects */
1127 if (!a->in_pack && !b->in_pack)
1128 return hashcmp(a->idx.sha1, b->idx.sha1);
1130 if (a->in_pack < b->in_pack)
1131 return -1;
1132 if (a->in_pack > b->in_pack)
1133 return 1;
1134 return a->in_pack_offset < b->in_pack_offset ? -1 :
1135 (a->in_pack_offset > b->in_pack_offset);
1138 static void get_object_details(void)
1140 uint32_t i;
1141 struct object_entry **sorted_by_offset;
1143 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1144 for (i = 0; i < nr_objects; i++)
1145 sorted_by_offset[i] = objects + i;
1146 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1148 init_pack_revindex();
1150 for (i = 0; i < nr_objects; i++)
1151 check_object(sorted_by_offset[i]);
1153 free(sorted_by_offset);
1157 * We search for deltas in a list sorted by type, by filename hash, and then
1158 * by size, so that we see progressively smaller and smaller files.
1159 * That's because we prefer deltas to be from the bigger file
1160 * to the smaller -- deletes are potentially cheaper, but perhaps
1161 * more importantly, the bigger file is likely the more recent
1162 * one. The deepest deltas are therefore the oldest objects which are
1163 * less susceptible to be accessed often.
1165 static int type_size_sort(const void *_a, const void *_b)
1167 const struct object_entry *a = *(struct object_entry **)_a;
1168 const struct object_entry *b = *(struct object_entry **)_b;
1170 if (a->type > b->type)
1171 return -1;
1172 if (a->type < b->type)
1173 return 1;
1174 if (a->hash > b->hash)
1175 return -1;
1176 if (a->hash < b->hash)
1177 return 1;
1178 if (a->preferred_base > b->preferred_base)
1179 return -1;
1180 if (a->preferred_base < b->preferred_base)
1181 return 1;
1182 if (a->size > b->size)
1183 return -1;
1184 if (a->size < b->size)
1185 return 1;
1186 return a < b ? -1 : (a > b); /* newest first */
1189 struct unpacked {
1190 struct object_entry *entry;
1191 void *data;
1192 struct delta_index *index;
1193 unsigned depth;
1196 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1197 unsigned long delta_size)
1199 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1200 return 0;
1202 if (delta_size < cache_max_small_delta_size)
1203 return 1;
1205 /* cache delta, if objects are large enough compared to delta size */
1206 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1207 return 1;
1209 return 0;
1212 #ifdef THREADED_DELTA_SEARCH
1214 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1215 #define read_lock() pthread_mutex_lock(&read_mutex)
1216 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1218 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1219 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1220 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1222 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1223 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1224 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1226 #else
1228 #define read_lock() (void)0
1229 #define read_unlock() (void)0
1230 #define cache_lock() (void)0
1231 #define cache_unlock() (void)0
1232 #define progress_lock() (void)0
1233 #define progress_unlock() (void)0
1235 #endif
1237 static int try_delta(struct unpacked *trg, struct unpacked *src,
1238 unsigned max_depth, unsigned long *mem_usage)
1240 struct object_entry *trg_entry = trg->entry;
1241 struct object_entry *src_entry = src->entry;
1242 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1243 unsigned ref_depth;
1244 enum object_type type;
1245 void *delta_buf;
1247 /* Don't bother doing diffs between different types */
1248 if (trg_entry->type != src_entry->type)
1249 return -1;
1252 * We do not bother to try a delta that we discarded
1253 * on an earlier try, but only when reusing delta data.
1255 if (reuse_delta && trg_entry->in_pack &&
1256 trg_entry->in_pack == src_entry->in_pack &&
1257 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1258 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1259 return 0;
1261 /* Let's not bust the allowed depth. */
1262 if (src->depth >= max_depth)
1263 return 0;
1265 /* Now some size filtering heuristics. */
1266 trg_size = trg_entry->size;
1267 if (!trg_entry->delta) {
1268 max_size = trg_size/2 - 20;
1269 ref_depth = 1;
1270 } else {
1271 max_size = trg_entry->delta_size;
1272 ref_depth = trg->depth;
1274 max_size = max_size * (max_depth - src->depth) /
1275 (max_depth - ref_depth + 1);
1276 if (max_size == 0)
1277 return 0;
1278 src_size = src_entry->size;
1279 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1280 if (sizediff >= max_size)
1281 return 0;
1282 if (trg_size < src_size / 32)
1283 return 0;
1285 /* Load data if not already done */
1286 if (!trg->data) {
1287 read_lock();
1288 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1289 read_unlock();
1290 if (!trg->data)
1291 die("object %s cannot be read",
1292 sha1_to_hex(trg_entry->idx.sha1));
1293 if (sz != trg_size)
1294 die("object %s inconsistent object length (%lu vs %lu)",
1295 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1296 *mem_usage += sz;
1298 if (!src->data) {
1299 read_lock();
1300 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1301 read_unlock();
1302 if (!src->data)
1303 die("object %s cannot be read",
1304 sha1_to_hex(src_entry->idx.sha1));
1305 if (sz != src_size)
1306 die("object %s inconsistent object length (%lu vs %lu)",
1307 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1308 *mem_usage += sz;
1310 if (!src->index) {
1311 src->index = create_delta_index(src->data, src_size);
1312 if (!src->index) {
1313 static int warned = 0;
1314 if (!warned++)
1315 warning("suboptimal pack - out of memory");
1316 return 0;
1318 *mem_usage += sizeof_delta_index(src->index);
1321 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1322 if (!delta_buf)
1323 return 0;
1325 if (trg_entry->delta) {
1326 /* Prefer only shallower same-sized deltas. */
1327 if (delta_size == trg_entry->delta_size &&
1328 src->depth + 1 >= trg->depth) {
1329 free(delta_buf);
1330 return 0;
1335 * Handle memory allocation outside of the cache
1336 * accounting lock. Compiler will optimize the strangeness
1337 * away when THREADED_DELTA_SEARCH is not defined.
1339 free(trg_entry->delta_data);
1340 cache_lock();
1341 if (trg_entry->delta_data) {
1342 delta_cache_size -= trg_entry->delta_size;
1343 trg_entry->delta_data = NULL;
1345 if (delta_cacheable(src_size, trg_size, delta_size)) {
1346 delta_cache_size += delta_size;
1347 cache_unlock();
1348 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1349 } else {
1350 cache_unlock();
1351 free(delta_buf);
1354 trg_entry->delta = src_entry;
1355 trg_entry->delta_size = delta_size;
1356 trg->depth = src->depth + 1;
1358 return 1;
1361 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1363 struct object_entry *child = me->delta_child;
1364 unsigned int m = n;
1365 while (child) {
1366 unsigned int c = check_delta_limit(child, n + 1);
1367 if (m < c)
1368 m = c;
1369 child = child->delta_sibling;
1371 return m;
1374 static unsigned long free_unpacked(struct unpacked *n)
1376 unsigned long freed_mem = sizeof_delta_index(n->index);
1377 free_delta_index(n->index);
1378 n->index = NULL;
1379 if (n->data) {
1380 freed_mem += n->entry->size;
1381 free(n->data);
1382 n->data = NULL;
1384 n->entry = NULL;
1385 n->depth = 0;
1386 return freed_mem;
1389 static void find_deltas(struct object_entry **list, unsigned *list_size,
1390 int window, int depth, unsigned *processed)
1392 uint32_t i, idx = 0, count = 0;
1393 unsigned int array_size = window * sizeof(struct unpacked);
1394 struct unpacked *array;
1395 unsigned long mem_usage = 0;
1397 array = xmalloc(array_size);
1398 memset(array, 0, array_size);
1400 for (;;) {
1401 struct object_entry *entry = *list++;
1402 struct unpacked *n = array + idx;
1403 int j, max_depth, best_base = -1;
1405 progress_lock();
1406 if (!*list_size) {
1407 progress_unlock();
1408 break;
1410 (*list_size)--;
1411 if (!entry->preferred_base) {
1412 (*processed)++;
1413 display_progress(progress_state, *processed);
1415 progress_unlock();
1417 mem_usage -= free_unpacked(n);
1418 n->entry = entry;
1420 while (window_memory_limit &&
1421 mem_usage > window_memory_limit &&
1422 count > 1) {
1423 uint32_t tail = (idx + window - count) % window;
1424 mem_usage -= free_unpacked(array + tail);
1425 count--;
1428 /* We do not compute delta to *create* objects we are not
1429 * going to pack.
1431 if (entry->preferred_base)
1432 goto next;
1435 * If the current object is at pack edge, take the depth the
1436 * objects that depend on the current object into account
1437 * otherwise they would become too deep.
1439 max_depth = depth;
1440 if (entry->delta_child) {
1441 max_depth -= check_delta_limit(entry, 0);
1442 if (max_depth <= 0)
1443 goto next;
1446 j = window;
1447 while (--j > 0) {
1448 int ret;
1449 uint32_t other_idx = idx + j;
1450 struct unpacked *m;
1451 if (other_idx >= window)
1452 other_idx -= window;
1453 m = array + other_idx;
1454 if (!m->entry)
1455 break;
1456 ret = try_delta(n, m, max_depth, &mem_usage);
1457 if (ret < 0)
1458 break;
1459 else if (ret > 0)
1460 best_base = other_idx;
1464 * If we decided to cache the delta data, then it is best
1465 * to compress it right away. First because we have to do
1466 * it anyway, and doing it here while we're threaded will
1467 * save a lot of time in the non threaded write phase,
1468 * as well as allow for caching more deltas within
1469 * the same cache size limit.
1470 * ...
1471 * But only if not writing to stdout, since in that case
1472 * the network is most likely throttling writes anyway,
1473 * and therefore it is best to go to the write phase ASAP
1474 * instead, as we can afford spending more time compressing
1475 * between writes at that moment.
1477 if (entry->delta_data && !pack_to_stdout) {
1478 entry->z_delta_size = do_compress(&entry->delta_data,
1479 entry->delta_size);
1480 cache_lock();
1481 delta_cache_size -= entry->delta_size;
1482 delta_cache_size += entry->z_delta_size;
1483 cache_unlock();
1486 /* if we made n a delta, and if n is already at max
1487 * depth, leaving it in the window is pointless. we
1488 * should evict it first.
1490 if (entry->delta && max_depth <= n->depth)
1491 continue;
1494 * Move the best delta base up in the window, after the
1495 * currently deltified object, to keep it longer. It will
1496 * be the first base object to be attempted next.
1498 if (entry->delta) {
1499 struct unpacked swap = array[best_base];
1500 int dist = (window + idx - best_base) % window;
1501 int dst = best_base;
1502 while (dist--) {
1503 int src = (dst + 1) % window;
1504 array[dst] = array[src];
1505 dst = src;
1507 array[dst] = swap;
1510 next:
1511 idx++;
1512 if (count + 1 < window)
1513 count++;
1514 if (idx >= window)
1515 idx = 0;
1518 for (i = 0; i < window; ++i) {
1519 free_delta_index(array[i].index);
1520 free(array[i].data);
1522 free(array);
1525 #ifdef THREADED_DELTA_SEARCH
1528 * The main thread waits on the condition that (at least) one of the workers
1529 * has stopped working (which is indicated in the .working member of
1530 * struct thread_params).
1531 * When a work thread has completed its work, it sets .working to 0 and
1532 * signals the main thread and waits on the condition that .data_ready
1533 * becomes 1.
1536 struct thread_params {
1537 pthread_t thread;
1538 struct object_entry **list;
1539 unsigned list_size;
1540 unsigned remaining;
1541 int window;
1542 int depth;
1543 int working;
1544 int data_ready;
1545 pthread_mutex_t mutex;
1546 pthread_cond_t cond;
1547 unsigned *processed;
1550 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1552 static void *threaded_find_deltas(void *arg)
1554 struct thread_params *me = arg;
1556 while (me->remaining) {
1557 find_deltas(me->list, &me->remaining,
1558 me->window, me->depth, me->processed);
1560 progress_lock();
1561 me->working = 0;
1562 pthread_cond_signal(&progress_cond);
1563 progress_unlock();
1566 * We must not set ->data_ready before we wait on the
1567 * condition because the main thread may have set it to 1
1568 * before we get here. In order to be sure that new
1569 * work is available if we see 1 in ->data_ready, it
1570 * was initialized to 0 before this thread was spawned
1571 * and we reset it to 0 right away.
1573 pthread_mutex_lock(&me->mutex);
1574 while (!me->data_ready)
1575 pthread_cond_wait(&me->cond, &me->mutex);
1576 me->data_ready = 0;
1577 pthread_mutex_unlock(&me->mutex);
1579 /* leave ->working 1 so that this doesn't get more work assigned */
1580 return NULL;
1583 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1584 int window, int depth, unsigned *processed)
1586 struct thread_params p[delta_search_threads];
1587 int i, ret, active_threads = 0;
1589 if (delta_search_threads <= 1) {
1590 find_deltas(list, &list_size, window, depth, processed);
1591 return;
1594 /* Partition the work amongst work threads. */
1595 for (i = 0; i < delta_search_threads; i++) {
1596 unsigned sub_size = list_size / (delta_search_threads - i);
1598 p[i].window = window;
1599 p[i].depth = depth;
1600 p[i].processed = processed;
1601 p[i].working = 1;
1602 p[i].data_ready = 0;
1604 /* try to split chunks on "path" boundaries */
1605 while (sub_size && sub_size < list_size &&
1606 list[sub_size]->hash &&
1607 list[sub_size]->hash == list[sub_size-1]->hash)
1608 sub_size++;
1610 p[i].list = list;
1611 p[i].list_size = sub_size;
1612 p[i].remaining = sub_size;
1614 list += sub_size;
1615 list_size -= sub_size;
1618 /* Start work threads. */
1619 for (i = 0; i < delta_search_threads; i++) {
1620 if (!p[i].list_size)
1621 continue;
1622 pthread_mutex_init(&p[i].mutex, NULL);
1623 pthread_cond_init(&p[i].cond, NULL);
1624 ret = pthread_create(&p[i].thread, NULL,
1625 threaded_find_deltas, &p[i]);
1626 if (ret)
1627 die("unable to create thread: %s", strerror(ret));
1628 active_threads++;
1632 * Now let's wait for work completion. Each time a thread is done
1633 * with its work, we steal half of the remaining work from the
1634 * thread with the largest number of unprocessed objects and give
1635 * it to that newly idle thread. This ensure good load balancing
1636 * until the remaining object list segments are simply too short
1637 * to be worth splitting anymore.
1639 while (active_threads) {
1640 struct thread_params *target = NULL;
1641 struct thread_params *victim = NULL;
1642 unsigned sub_size = 0;
1644 progress_lock();
1645 for (;;) {
1646 for (i = 0; !target && i < delta_search_threads; i++)
1647 if (!p[i].working)
1648 target = &p[i];
1649 if (target)
1650 break;
1651 pthread_cond_wait(&progress_cond, &progress_mutex);
1654 for (i = 0; i < delta_search_threads; i++)
1655 if (p[i].remaining > 2*window &&
1656 (!victim || victim->remaining < p[i].remaining))
1657 victim = &p[i];
1658 if (victim) {
1659 sub_size = victim->remaining / 2;
1660 list = victim->list + victim->list_size - sub_size;
1661 while (sub_size && list[0]->hash &&
1662 list[0]->hash == list[-1]->hash) {
1663 list++;
1664 sub_size--;
1666 if (!sub_size) {
1668 * It is possible for some "paths" to have
1669 * so many objects that no hash boundary
1670 * might be found. Let's just steal the
1671 * exact half in that case.
1673 sub_size = victim->remaining / 2;
1674 list -= sub_size;
1676 target->list = list;
1677 victim->list_size -= sub_size;
1678 victim->remaining -= sub_size;
1680 target->list_size = sub_size;
1681 target->remaining = sub_size;
1682 target->working = 1;
1683 progress_unlock();
1685 pthread_mutex_lock(&target->mutex);
1686 target->data_ready = 1;
1687 pthread_cond_signal(&target->cond);
1688 pthread_mutex_unlock(&target->mutex);
1690 if (!sub_size) {
1691 pthread_join(target->thread, NULL);
1692 pthread_cond_destroy(&target->cond);
1693 pthread_mutex_destroy(&target->mutex);
1694 active_threads--;
1699 #else
1700 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1701 #endif
1703 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1705 unsigned char peeled[20];
1707 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1708 !peel_ref(path, peeled) && /* peelable? */
1709 !is_null_sha1(peeled) && /* annotated tag? */
1710 locate_object_entry(peeled)) /* object packed? */
1711 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1712 return 0;
1715 static void prepare_pack(int window, int depth)
1717 struct object_entry **delta_list;
1718 uint32_t i, n, nr_deltas;
1720 get_object_details();
1722 if (!nr_objects || !window || !depth)
1723 return;
1725 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1726 nr_deltas = n = 0;
1728 for (i = 0; i < nr_objects; i++) {
1729 struct object_entry *entry = objects + i;
1731 if (entry->delta)
1732 /* This happens if we decided to reuse existing
1733 * delta from a pack. "reuse_delta &&" is implied.
1735 continue;
1737 if (entry->size < 50)
1738 continue;
1740 if (entry->no_try_delta)
1741 continue;
1743 if (!entry->preferred_base)
1744 nr_deltas++;
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)
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=%d", pack_idx_default_version);
1810 return 0;
1812 if (!strcmp(k, "pack.packsizelimit")) {
1813 pack_size_limit_cfg = git_config_ulong(k, v);
1814 return 0;
1816 return git_default_config(k, v);
1819 static void read_object_list_from_stdin(void)
1821 char line[40 + 1 + PATH_MAX + 2];
1822 unsigned char sha1[20];
1824 for (;;) {
1825 if (!fgets(line, sizeof(line), stdin)) {
1826 if (feof(stdin))
1827 break;
1828 if (!ferror(stdin))
1829 die("fgets returned NULL, not EOF, not error!");
1830 if (errno != EINTR)
1831 die("fgets: %s", strerror(errno));
1832 clearerr(stdin);
1833 continue;
1835 if (line[0] == '-') {
1836 if (get_sha1_hex(line+1, sha1))
1837 die("expected edge sha1, got garbage:\n %s",
1838 line);
1839 add_preferred_base(sha1);
1840 continue;
1842 if (get_sha1_hex(line, sha1))
1843 die("expected sha1, got garbage:\n %s", line);
1845 add_preferred_base_object(line+41);
1846 add_object_entry(sha1, 0, line+41, 0);
1850 #define OBJECT_ADDED (1u<<20)
1852 static void show_commit(struct commit *commit)
1854 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1855 commit->object.flags |= OBJECT_ADDED;
1858 static void show_object(struct object_array_entry *p)
1860 add_preferred_base_object(p->name);
1861 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1862 p->item->flags |= OBJECT_ADDED;
1865 static void show_edge(struct commit *commit)
1867 add_preferred_base(commit->object.sha1);
1870 struct in_pack_object {
1871 off_t offset;
1872 struct object *object;
1875 struct in_pack {
1876 int alloc;
1877 int nr;
1878 struct in_pack_object *array;
1881 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1883 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1884 in_pack->array[in_pack->nr].object = object;
1885 in_pack->nr++;
1889 * Compare the objects in the offset order, in order to emulate the
1890 * "git-rev-list --objects" output that produced the pack originally.
1892 static int ofscmp(const void *a_, const void *b_)
1894 struct in_pack_object *a = (struct in_pack_object *)a_;
1895 struct in_pack_object *b = (struct in_pack_object *)b_;
1897 if (a->offset < b->offset)
1898 return -1;
1899 else if (a->offset > b->offset)
1900 return 1;
1901 else
1902 return hashcmp(a->object->sha1, b->object->sha1);
1905 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1907 struct packed_git *p;
1908 struct in_pack in_pack;
1909 uint32_t i;
1911 memset(&in_pack, 0, sizeof(in_pack));
1913 for (p = packed_git; p; p = p->next) {
1914 const unsigned char *sha1;
1915 struct object *o;
1917 for (i = 0; i < revs->num_ignore_packed; i++) {
1918 if (matches_pack_name(p, revs->ignore_packed[i]))
1919 break;
1921 if (revs->num_ignore_packed <= i)
1922 continue;
1923 if (open_pack_index(p))
1924 die("cannot open pack index");
1926 ALLOC_GROW(in_pack.array,
1927 in_pack.nr + p->num_objects,
1928 in_pack.alloc);
1930 for (i = 0; i < p->num_objects; i++) {
1931 sha1 = nth_packed_object_sha1(p, i);
1932 o = lookup_unknown_object(sha1);
1933 if (!(o->flags & OBJECT_ADDED))
1934 mark_in_pack_object(o, p, &in_pack);
1935 o->flags |= OBJECT_ADDED;
1939 if (in_pack.nr) {
1940 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1941 ofscmp);
1942 for (i = 0; i < in_pack.nr; i++) {
1943 struct object *o = in_pack.array[i].object;
1944 add_object_entry(o->sha1, o->type, "", 0);
1947 free(in_pack.array);
1950 static void loosen_unused_packed_objects(struct rev_info *revs)
1952 struct packed_git *p;
1953 uint32_t i;
1954 const unsigned char *sha1;
1956 for (p = packed_git; p; p = p->next) {
1957 for (i = 0; i < revs->num_ignore_packed; i++) {
1958 if (matches_pack_name(p, revs->ignore_packed[i]))
1959 break;
1961 if (revs->num_ignore_packed <= i)
1962 continue;
1964 if (open_pack_index(p))
1965 die("cannot open pack index");
1967 for (i = 0; i < p->num_objects; i++) {
1968 sha1 = nth_packed_object_sha1(p, i);
1969 if (!locate_object_entry(sha1))
1970 if (force_object_loose(sha1, p->mtime))
1971 die("unable to force loose object");
1976 static void get_object_list(int ac, const char **av)
1978 struct rev_info revs;
1979 char line[1000];
1980 int flags = 0;
1982 init_revisions(&revs, NULL);
1983 save_commit_buffer = 0;
1984 setup_revisions(ac, av, &revs, NULL);
1986 while (fgets(line, sizeof(line), stdin) != NULL) {
1987 int len = strlen(line);
1988 if (len && line[len - 1] == '\n')
1989 line[--len] = 0;
1990 if (!len)
1991 break;
1992 if (*line == '-') {
1993 if (!strcmp(line, "--not")) {
1994 flags ^= UNINTERESTING;
1995 continue;
1997 die("not a rev '%s'", line);
1999 if (handle_revision_arg(line, &revs, flags, 1))
2000 die("bad revision '%s'", line);
2003 if (prepare_revision_walk(&revs))
2004 die("revision walk setup failed");
2005 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2006 traverse_commit_list(&revs, show_commit, show_object);
2008 if (keep_unreachable)
2009 add_objects_in_unpacked_packs(&revs);
2010 if (unpack_unreachable)
2011 loosen_unused_packed_objects(&revs);
2014 static int adjust_perm(const char *path, mode_t mode)
2016 if (chmod(path, mode))
2017 return -1;
2018 return adjust_shared_perm(path);
2021 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2023 int use_internal_rev_list = 0;
2024 int thin = 0;
2025 uint32_t i;
2026 const char **rp_av;
2027 int rp_ac_alloc = 64;
2028 int rp_ac;
2030 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2032 rp_av[0] = "pack-objects";
2033 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2034 rp_ac = 2;
2036 git_config(git_pack_config);
2037 if (!pack_compression_seen && core_compression_seen)
2038 pack_compression_level = core_compression_level;
2040 progress = isatty(2);
2041 for (i = 1; i < argc; i++) {
2042 const char *arg = argv[i];
2044 if (*arg != '-')
2045 break;
2047 if (!strcmp("--non-empty", arg)) {
2048 non_empty = 1;
2049 continue;
2051 if (!strcmp("--local", arg)) {
2052 local = 1;
2053 continue;
2055 if (!strcmp("--incremental", arg)) {
2056 incremental = 1;
2057 continue;
2059 if (!prefixcmp(arg, "--compression=")) {
2060 char *end;
2061 int level = strtoul(arg+14, &end, 0);
2062 if (!arg[14] || *end)
2063 usage(pack_usage);
2064 if (level == -1)
2065 level = Z_DEFAULT_COMPRESSION;
2066 else if (level < 0 || level > Z_BEST_COMPRESSION)
2067 die("bad pack compression level %d", level);
2068 pack_compression_level = level;
2069 continue;
2071 if (!prefixcmp(arg, "--max-pack-size=")) {
2072 char *end;
2073 pack_size_limit_cfg = 0;
2074 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2075 if (!arg[16] || *end)
2076 usage(pack_usage);
2077 continue;
2079 if (!prefixcmp(arg, "--window=")) {
2080 char *end;
2081 window = strtoul(arg+9, &end, 0);
2082 if (!arg[9] || *end)
2083 usage(pack_usage);
2084 continue;
2086 if (!prefixcmp(arg, "--window-memory=")) {
2087 if (!git_parse_ulong(arg+16, &window_memory_limit))
2088 usage(pack_usage);
2089 continue;
2091 if (!prefixcmp(arg, "--threads=")) {
2092 char *end;
2093 delta_search_threads = strtoul(arg+10, &end, 0);
2094 if (!arg[10] || *end || delta_search_threads < 0)
2095 usage(pack_usage);
2096 #ifndef THREADED_DELTA_SEARCH
2097 if (delta_search_threads != 1)
2098 warning("no threads support, "
2099 "ignoring %s", arg);
2100 #endif
2101 continue;
2103 if (!prefixcmp(arg, "--depth=")) {
2104 char *end;
2105 depth = strtoul(arg+8, &end, 0);
2106 if (!arg[8] || *end)
2107 usage(pack_usage);
2108 continue;
2110 if (!strcmp("--progress", arg)) {
2111 progress = 1;
2112 continue;
2114 if (!strcmp("--all-progress", arg)) {
2115 progress = 2;
2116 continue;
2118 if (!strcmp("-q", arg)) {
2119 progress = 0;
2120 continue;
2122 if (!strcmp("--no-reuse-delta", arg)) {
2123 reuse_delta = 0;
2124 continue;
2126 if (!strcmp("--no-reuse-object", arg)) {
2127 reuse_object = reuse_delta = 0;
2128 continue;
2130 if (!strcmp("--delta-base-offset", arg)) {
2131 allow_ofs_delta = 1;
2132 continue;
2134 if (!strcmp("--stdout", arg)) {
2135 pack_to_stdout = 1;
2136 continue;
2138 if (!strcmp("--revs", arg)) {
2139 use_internal_rev_list = 1;
2140 continue;
2142 if (!strcmp("--keep-unreachable", arg)) {
2143 keep_unreachable = 1;
2144 continue;
2146 if (!strcmp("--unpack-unreachable", arg)) {
2147 unpack_unreachable = 1;
2148 continue;
2150 if (!strcmp("--include-tag", arg)) {
2151 include_tag = 1;
2152 continue;
2154 if (!strcmp("--unpacked", arg) ||
2155 !prefixcmp(arg, "--unpacked=") ||
2156 !strcmp("--reflog", arg) ||
2157 !strcmp("--all", arg)) {
2158 use_internal_rev_list = 1;
2159 if (rp_ac >= rp_ac_alloc - 1) {
2160 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2161 rp_av = xrealloc(rp_av,
2162 rp_ac_alloc * sizeof(*rp_av));
2164 rp_av[rp_ac++] = arg;
2165 continue;
2167 if (!strcmp("--thin", arg)) {
2168 use_internal_rev_list = 1;
2169 thin = 1;
2170 rp_av[1] = "--objects-edge";
2171 continue;
2173 if (!prefixcmp(arg, "--index-version=")) {
2174 char *c;
2175 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2176 if (pack_idx_default_version > 2)
2177 die("bad %s", arg);
2178 if (*c == ',')
2179 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2180 if (*c || pack_idx_off32_limit & 0x80000000)
2181 die("bad %s", arg);
2182 continue;
2184 usage(pack_usage);
2187 /* Traditionally "pack-objects [options] base extra" failed;
2188 * we would however want to take refs parameter that would
2189 * have been given to upstream rev-list ourselves, which means
2190 * we somehow want to say what the base name is. So the
2191 * syntax would be:
2193 * pack-objects [options] base <refs...>
2195 * in other words, we would treat the first non-option as the
2196 * base_name and send everything else to the internal revision
2197 * walker.
2200 if (!pack_to_stdout)
2201 base_name = argv[i++];
2203 if (pack_to_stdout != !base_name)
2204 usage(pack_usage);
2206 if (!pack_to_stdout && !pack_size_limit)
2207 pack_size_limit = pack_size_limit_cfg;
2209 if (pack_to_stdout && pack_size_limit)
2210 die("--max-pack-size cannot be used to build a pack for transfer.");
2212 if (!pack_to_stdout && thin)
2213 die("--thin cannot be used to build an indexable pack.");
2215 if (keep_unreachable && unpack_unreachable)
2216 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2218 #ifdef THREADED_DELTA_SEARCH
2219 if (!delta_search_threads) /* --threads=0 means autodetect */
2220 delta_search_threads = online_cpus();
2221 #endif
2223 prepare_packed_git();
2225 if (progress)
2226 progress_state = start_progress("Counting objects", 0);
2227 if (!use_internal_rev_list)
2228 read_object_list_from_stdin();
2229 else {
2230 rp_av[rp_ac] = NULL;
2231 get_object_list(rp_ac, rp_av);
2233 if (include_tag && nr_result)
2234 for_each_ref(add_ref_tag, NULL);
2235 stop_progress(&progress_state);
2237 if (non_empty && !nr_result)
2238 return 0;
2239 if (nr_result)
2240 prepare_pack(window, depth);
2241 write_pack_file();
2242 if (progress)
2243 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2244 written, written_delta, reused, reused_delta);
2245 return 0;