Windows ANSI - hook write()
[git/mingw/4msysgit/peterh.git] / builtin-pack-objects.c
blob447d492dbbfb578b8903293fa1d2a988eecfefac
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) {
518 sha1close(f, sha1, CSUM_CLOSE);
519 } else if (nr_written == nr_remaining) {
520 sha1close(f, sha1, CSUM_FSYNC);
521 } else {
522 int fd = sha1close(f, NULL, 0);
523 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
524 fsync_or_die(fd, pack_tmp_name);
525 close(fd);
528 if (!pack_to_stdout) {
529 mode_t mode = umask(0);
530 struct stat st;
531 char *idx_tmp_name, tmpname[PATH_MAX];
533 umask(mode);
534 mode = 0444 & ~mode;
536 idx_tmp_name = write_idx_file(NULL, written_list,
537 nr_written, sha1);
539 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
540 base_name, sha1_to_hex(sha1));
541 if (adjust_perm(pack_tmp_name, mode))
542 die("unable to make temporary pack file readable: %s",
543 strerror(errno));
544 if (rename(pack_tmp_name, tmpname))
545 die("unable to rename temporary pack file: %s",
546 strerror(errno));
549 * Packs are runtime accessed in their mtime
550 * order since newer packs are more likely to contain
551 * younger objects. So if we are creating multiple
552 * packs then we should modify the mtime of later ones
553 * to preserve this property.
555 if (stat(tmpname, &st) < 0) {
556 warning("failed to stat %s: %s",
557 tmpname, strerror(errno));
558 } else if (!last_mtime) {
559 last_mtime = st.st_mtime;
560 } else {
561 struct utimbuf utb;
562 utb.actime = st.st_atime;
563 utb.modtime = --last_mtime;
564 if (utime(tmpname, &utb) < 0)
565 warning("failed utime() on %s: %s",
566 tmpname, strerror(errno));
569 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
570 base_name, sha1_to_hex(sha1));
571 if (adjust_perm(idx_tmp_name, mode))
572 die("unable to make temporary index file readable: %s",
573 strerror(errno));
574 if (rename(idx_tmp_name, tmpname))
575 die("unable to rename temporary index file: %s",
576 strerror(errno));
578 free(idx_tmp_name);
579 free(pack_tmp_name);
580 puts(sha1_to_hex(sha1));
583 /* mark written objects as written to previous pack */
584 for (j = 0; j < nr_written; j++) {
585 written_list[j]->offset = (off_t)-1;
587 nr_remaining -= nr_written;
588 } while (nr_remaining && i < nr_objects);
590 free(written_list);
591 stop_progress(&progress_state);
592 if (written != nr_result)
593 die("wrote %u objects while expecting %u", written, nr_result);
595 * We have scanned through [0 ... i). Since we have written
596 * the correct number of objects, the remaining [i ... nr_objects)
597 * items must be either already written (due to out-of-order delta base)
598 * or a preferred base. Count those which are neither and complain if any.
600 for (j = 0; i < nr_objects; i++) {
601 struct object_entry *e = objects + i;
602 j += !e->idx.offset && !e->preferred_base;
604 if (j)
605 die("wrote %u objects as expected but %u unwritten", written, j);
608 static int locate_object_entry_hash(const unsigned char *sha1)
610 int i;
611 unsigned int ui;
612 memcpy(&ui, sha1, sizeof(unsigned int));
613 i = ui % object_ix_hashsz;
614 while (0 < object_ix[i]) {
615 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
616 return i;
617 if (++i == object_ix_hashsz)
618 i = 0;
620 return -1 - i;
623 static struct object_entry *locate_object_entry(const unsigned char *sha1)
625 int i;
627 if (!object_ix_hashsz)
628 return NULL;
630 i = locate_object_entry_hash(sha1);
631 if (0 <= i)
632 return &objects[object_ix[i]-1];
633 return NULL;
636 static void rehash_objects(void)
638 uint32_t i;
639 struct object_entry *oe;
641 object_ix_hashsz = nr_objects * 3;
642 if (object_ix_hashsz < 1024)
643 object_ix_hashsz = 1024;
644 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
645 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
646 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
647 int ix = locate_object_entry_hash(oe->idx.sha1);
648 if (0 <= ix)
649 continue;
650 ix = -1 - ix;
651 object_ix[ix] = i + 1;
655 static unsigned name_hash(const char *name)
657 unsigned char c;
658 unsigned hash = 0;
660 if (!name)
661 return 0;
664 * This effectively just creates a sortable number from the
665 * last sixteen non-whitespace characters. Last characters
666 * count "most", so things that end in ".c" sort together.
668 while ((c = *name++) != 0) {
669 if (isspace(c))
670 continue;
671 hash = (hash >> 2) + (c << 24);
673 return hash;
676 static void setup_delta_attr_check(struct git_attr_check *check)
678 static struct git_attr *attr_delta;
680 if (!attr_delta)
681 attr_delta = git_attr("delta", 5);
683 check[0].attr = attr_delta;
686 static int no_try_delta(const char *path)
688 struct git_attr_check check[1];
690 setup_delta_attr_check(check);
691 if (git_checkattr(path, ARRAY_SIZE(check), check))
692 return 0;
693 if (ATTR_FALSE(check->value))
694 return 1;
695 return 0;
698 static int add_object_entry(const unsigned char *sha1, enum object_type type,
699 const char *name, int exclude)
701 struct object_entry *entry;
702 struct packed_git *p, *found_pack = NULL;
703 off_t found_offset = 0;
704 int ix;
705 unsigned hash = name_hash(name);
707 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
708 if (ix >= 0) {
709 if (exclude) {
710 entry = objects + object_ix[ix] - 1;
711 if (!entry->preferred_base)
712 nr_result--;
713 entry->preferred_base = 1;
715 return 0;
718 for (p = packed_git; p; p = p->next) {
719 off_t offset = find_pack_entry_one(sha1, p);
720 if (offset) {
721 if (!found_pack) {
722 found_offset = offset;
723 found_pack = p;
725 if (exclude)
726 break;
727 if (incremental)
728 return 0;
729 if (local && !p->pack_local)
730 return 0;
734 if (nr_objects >= nr_alloc) {
735 nr_alloc = (nr_alloc + 1024) * 3 / 2;
736 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
739 entry = objects + nr_objects++;
740 memset(entry, 0, sizeof(*entry));
741 hashcpy(entry->idx.sha1, sha1);
742 entry->hash = hash;
743 if (type)
744 entry->type = type;
745 if (exclude)
746 entry->preferred_base = 1;
747 else
748 nr_result++;
749 if (found_pack) {
750 entry->in_pack = found_pack;
751 entry->in_pack_offset = found_offset;
754 if (object_ix_hashsz * 3 <= nr_objects * 4)
755 rehash_objects();
756 else
757 object_ix[-1 - ix] = nr_objects;
759 display_progress(progress_state, nr_objects);
761 if (name && no_try_delta(name))
762 entry->no_try_delta = 1;
764 return 1;
767 struct pbase_tree_cache {
768 unsigned char sha1[20];
769 int ref;
770 int temporary;
771 void *tree_data;
772 unsigned long tree_size;
775 static struct pbase_tree_cache *(pbase_tree_cache[256]);
776 static int pbase_tree_cache_ix(const unsigned char *sha1)
778 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
780 static int pbase_tree_cache_ix_incr(int ix)
782 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
785 static struct pbase_tree {
786 struct pbase_tree *next;
787 /* This is a phony "cache" entry; we are not
788 * going to evict it nor find it through _get()
789 * mechanism -- this is for the toplevel node that
790 * would almost always change with any commit.
792 struct pbase_tree_cache pcache;
793 } *pbase_tree;
795 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
797 struct pbase_tree_cache *ent, *nent;
798 void *data;
799 unsigned long size;
800 enum object_type type;
801 int neigh;
802 int my_ix = pbase_tree_cache_ix(sha1);
803 int available_ix = -1;
805 /* pbase-tree-cache acts as a limited hashtable.
806 * your object will be found at your index or within a few
807 * slots after that slot if it is cached.
809 for (neigh = 0; neigh < 8; neigh++) {
810 ent = pbase_tree_cache[my_ix];
811 if (ent && !hashcmp(ent->sha1, sha1)) {
812 ent->ref++;
813 return ent;
815 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
816 ((0 <= available_ix) &&
817 (!ent && pbase_tree_cache[available_ix])))
818 available_ix = my_ix;
819 if (!ent)
820 break;
821 my_ix = pbase_tree_cache_ix_incr(my_ix);
824 /* Did not find one. Either we got a bogus request or
825 * we need to read and perhaps cache.
827 data = read_sha1_file(sha1, &type, &size);
828 if (!data)
829 return NULL;
830 if (type != OBJ_TREE) {
831 free(data);
832 return NULL;
835 /* We need to either cache or return a throwaway copy */
837 if (available_ix < 0)
838 ent = NULL;
839 else {
840 ent = pbase_tree_cache[available_ix];
841 my_ix = available_ix;
844 if (!ent) {
845 nent = xmalloc(sizeof(*nent));
846 nent->temporary = (available_ix < 0);
848 else {
849 /* evict and reuse */
850 free(ent->tree_data);
851 nent = ent;
853 hashcpy(nent->sha1, sha1);
854 nent->tree_data = data;
855 nent->tree_size = size;
856 nent->ref = 1;
857 if (!nent->temporary)
858 pbase_tree_cache[my_ix] = nent;
859 return nent;
862 static void pbase_tree_put(struct pbase_tree_cache *cache)
864 if (!cache->temporary) {
865 cache->ref--;
866 return;
868 free(cache->tree_data);
869 free(cache);
872 static int name_cmp_len(const char *name)
874 int i;
875 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
877 return i;
880 static void add_pbase_object(struct tree_desc *tree,
881 const char *name,
882 int cmplen,
883 const char *fullname)
885 struct name_entry entry;
886 int cmp;
888 while (tree_entry(tree,&entry)) {
889 if (S_ISGITLINK(entry.mode))
890 continue;
891 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
892 memcmp(name, entry.path, cmplen);
893 if (cmp > 0)
894 continue;
895 if (cmp < 0)
896 return;
897 if (name[cmplen] != '/') {
898 add_object_entry(entry.sha1,
899 object_type(entry.mode),
900 fullname, 1);
901 return;
903 if (S_ISDIR(entry.mode)) {
904 struct tree_desc sub;
905 struct pbase_tree_cache *tree;
906 const char *down = name+cmplen+1;
907 int downlen = name_cmp_len(down);
909 tree = pbase_tree_get(entry.sha1);
910 if (!tree)
911 return;
912 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
914 add_pbase_object(&sub, down, downlen, fullname);
915 pbase_tree_put(tree);
920 static unsigned *done_pbase_paths;
921 static int done_pbase_paths_num;
922 static int done_pbase_paths_alloc;
923 static int done_pbase_path_pos(unsigned hash)
925 int lo = 0;
926 int hi = done_pbase_paths_num;
927 while (lo < hi) {
928 int mi = (hi + lo) / 2;
929 if (done_pbase_paths[mi] == hash)
930 return mi;
931 if (done_pbase_paths[mi] < hash)
932 hi = mi;
933 else
934 lo = mi + 1;
936 return -lo-1;
939 static int check_pbase_path(unsigned hash)
941 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
942 if (0 <= pos)
943 return 1;
944 pos = -pos - 1;
945 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
946 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
947 done_pbase_paths = xrealloc(done_pbase_paths,
948 done_pbase_paths_alloc *
949 sizeof(unsigned));
951 done_pbase_paths_num++;
952 if (pos < done_pbase_paths_num)
953 memmove(done_pbase_paths + pos + 1,
954 done_pbase_paths + pos,
955 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
956 done_pbase_paths[pos] = hash;
957 return 0;
960 static void add_preferred_base_object(const char *name)
962 struct pbase_tree *it;
963 int cmplen;
964 unsigned hash = name_hash(name);
966 if (!num_preferred_base || check_pbase_path(hash))
967 return;
969 cmplen = name_cmp_len(name);
970 for (it = pbase_tree; it; it = it->next) {
971 if (cmplen == 0) {
972 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
974 else {
975 struct tree_desc tree;
976 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
977 add_pbase_object(&tree, name, cmplen, name);
982 static void add_preferred_base(unsigned char *sha1)
984 struct pbase_tree *it;
985 void *data;
986 unsigned long size;
987 unsigned char tree_sha1[20];
989 if (window <= num_preferred_base++)
990 return;
992 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
993 if (!data)
994 return;
996 for (it = pbase_tree; it; it = it->next) {
997 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
998 free(data);
999 return;
1003 it = xcalloc(1, sizeof(*it));
1004 it->next = pbase_tree;
1005 pbase_tree = it;
1007 hashcpy(it->pcache.sha1, tree_sha1);
1008 it->pcache.tree_data = data;
1009 it->pcache.tree_size = size;
1012 static void check_object(struct object_entry *entry)
1014 if (entry->in_pack) {
1015 struct packed_git *p = entry->in_pack;
1016 struct pack_window *w_curs = NULL;
1017 const unsigned char *base_ref = NULL;
1018 struct object_entry *base_entry;
1019 unsigned long used, used_0;
1020 unsigned int avail;
1021 off_t ofs;
1022 unsigned char *buf, c;
1024 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1027 * We want in_pack_type even if we do not reuse delta
1028 * since non-delta representations could still be reused.
1030 used = unpack_object_header_gently(buf, avail,
1031 &entry->in_pack_type,
1032 &entry->size);
1035 * Determine if this is a delta and if so whether we can
1036 * reuse it or not. Otherwise let's find out as cheaply as
1037 * possible what the actual type and size for this object is.
1039 switch (entry->in_pack_type) {
1040 default:
1041 /* Not a delta hence we've already got all we need. */
1042 entry->type = entry->in_pack_type;
1043 entry->in_pack_header_size = used;
1044 unuse_pack(&w_curs);
1045 return;
1046 case OBJ_REF_DELTA:
1047 if (reuse_delta && !entry->preferred_base)
1048 base_ref = use_pack(p, &w_curs,
1049 entry->in_pack_offset + used, NULL);
1050 entry->in_pack_header_size = used + 20;
1051 break;
1052 case OBJ_OFS_DELTA:
1053 buf = use_pack(p, &w_curs,
1054 entry->in_pack_offset + used, NULL);
1055 used_0 = 0;
1056 c = buf[used_0++];
1057 ofs = c & 127;
1058 while (c & 128) {
1059 ofs += 1;
1060 if (!ofs || MSB(ofs, 7))
1061 die("delta base offset overflow in pack for %s",
1062 sha1_to_hex(entry->idx.sha1));
1063 c = buf[used_0++];
1064 ofs = (ofs << 7) + (c & 127);
1066 if (ofs >= entry->in_pack_offset)
1067 die("delta base offset out of bound for %s",
1068 sha1_to_hex(entry->idx.sha1));
1069 ofs = entry->in_pack_offset - ofs;
1070 if (reuse_delta && !entry->preferred_base) {
1071 struct revindex_entry *revidx;
1072 revidx = find_pack_revindex(p, ofs);
1073 base_ref = nth_packed_object_sha1(p, revidx->nr);
1075 entry->in_pack_header_size = used + used_0;
1076 break;
1079 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1081 * If base_ref was set above that means we wish to
1082 * reuse delta data, and we even found that base
1083 * in the list of objects we want to pack. Goodie!
1085 * Depth value does not matter - find_deltas() will
1086 * never consider reused delta as the base object to
1087 * deltify other objects against, in order to avoid
1088 * circular deltas.
1090 entry->type = entry->in_pack_type;
1091 entry->delta = base_entry;
1092 entry->delta_sibling = base_entry->delta_child;
1093 base_entry->delta_child = entry;
1094 unuse_pack(&w_curs);
1095 return;
1098 if (entry->type) {
1100 * This must be a delta and we already know what the
1101 * final object type is. Let's extract the actual
1102 * object size from the delta header.
1104 entry->size = get_size_from_delta(p, &w_curs,
1105 entry->in_pack_offset + entry->in_pack_header_size);
1106 unuse_pack(&w_curs);
1107 return;
1111 * No choice but to fall back to the recursive delta walk
1112 * with sha1_object_info() to find about the object type
1113 * at this point...
1115 unuse_pack(&w_curs);
1118 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1119 if (entry->type < 0)
1120 die("unable to get type of object %s",
1121 sha1_to_hex(entry->idx.sha1));
1124 static int pack_offset_sort(const void *_a, const void *_b)
1126 const struct object_entry *a = *(struct object_entry **)_a;
1127 const struct object_entry *b = *(struct object_entry **)_b;
1129 /* avoid filesystem trashing with loose objects */
1130 if (!a->in_pack && !b->in_pack)
1131 return hashcmp(a->idx.sha1, b->idx.sha1);
1133 if (a->in_pack < b->in_pack)
1134 return -1;
1135 if (a->in_pack > b->in_pack)
1136 return 1;
1137 return a->in_pack_offset < b->in_pack_offset ? -1 :
1138 (a->in_pack_offset > b->in_pack_offset);
1141 static void get_object_details(void)
1143 uint32_t i;
1144 struct object_entry **sorted_by_offset;
1146 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1147 for (i = 0; i < nr_objects; i++)
1148 sorted_by_offset[i] = objects + i;
1149 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1151 init_pack_revindex();
1153 for (i = 0; i < nr_objects; i++)
1154 check_object(sorted_by_offset[i]);
1156 free(sorted_by_offset);
1160 * We search for deltas in a list sorted by type, by filename hash, and then
1161 * by size, so that we see progressively smaller and smaller files.
1162 * That's because we prefer deltas to be from the bigger file
1163 * to the smaller -- deletes are potentially cheaper, but perhaps
1164 * more importantly, the bigger file is likely the more recent
1165 * one. The deepest deltas are therefore the oldest objects which are
1166 * less susceptible to be accessed often.
1168 static int type_size_sort(const void *_a, const void *_b)
1170 const struct object_entry *a = *(struct object_entry **)_a;
1171 const struct object_entry *b = *(struct object_entry **)_b;
1173 if (a->type > b->type)
1174 return -1;
1175 if (a->type < b->type)
1176 return 1;
1177 if (a->hash > b->hash)
1178 return -1;
1179 if (a->hash < b->hash)
1180 return 1;
1181 if (a->preferred_base > b->preferred_base)
1182 return -1;
1183 if (a->preferred_base < b->preferred_base)
1184 return 1;
1185 if (a->size > b->size)
1186 return -1;
1187 if (a->size < b->size)
1188 return 1;
1189 return a < b ? -1 : (a > b); /* newest first */
1192 struct unpacked {
1193 struct object_entry *entry;
1194 void *data;
1195 struct delta_index *index;
1196 unsigned depth;
1199 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1200 unsigned long delta_size)
1202 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1203 return 0;
1205 if (delta_size < cache_max_small_delta_size)
1206 return 1;
1208 /* cache delta, if objects are large enough compared to delta size */
1209 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1210 return 1;
1212 return 0;
1215 #ifdef THREADED_DELTA_SEARCH
1217 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1218 #define read_lock() pthread_mutex_lock(&read_mutex)
1219 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1221 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1222 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1223 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1225 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1226 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1227 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1229 #else
1231 #define read_lock() (void)0
1232 #define read_unlock() (void)0
1233 #define cache_lock() (void)0
1234 #define cache_unlock() (void)0
1235 #define progress_lock() (void)0
1236 #define progress_unlock() (void)0
1238 #endif
1240 static int try_delta(struct unpacked *trg, struct unpacked *src,
1241 unsigned max_depth, unsigned long *mem_usage)
1243 struct object_entry *trg_entry = trg->entry;
1244 struct object_entry *src_entry = src->entry;
1245 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1246 unsigned ref_depth;
1247 enum object_type type;
1248 void *delta_buf;
1250 /* Don't bother doing diffs between different types */
1251 if (trg_entry->type != src_entry->type)
1252 return -1;
1255 * We do not bother to try a delta that we discarded
1256 * on an earlier try, but only when reusing delta data.
1258 if (reuse_delta && trg_entry->in_pack &&
1259 trg_entry->in_pack == src_entry->in_pack &&
1260 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1261 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1262 return 0;
1264 /* Let's not bust the allowed depth. */
1265 if (src->depth >= max_depth)
1266 return 0;
1268 /* Now some size filtering heuristics. */
1269 trg_size = trg_entry->size;
1270 if (!trg_entry->delta) {
1271 max_size = trg_size/2 - 20;
1272 ref_depth = 1;
1273 } else {
1274 max_size = trg_entry->delta_size;
1275 ref_depth = trg->depth;
1277 max_size = max_size * (max_depth - src->depth) /
1278 (max_depth - ref_depth + 1);
1279 if (max_size == 0)
1280 return 0;
1281 src_size = src_entry->size;
1282 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1283 if (sizediff >= max_size)
1284 return 0;
1285 if (trg_size < src_size / 32)
1286 return 0;
1288 /* Load data if not already done */
1289 if (!trg->data) {
1290 read_lock();
1291 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1292 read_unlock();
1293 if (!trg->data)
1294 die("object %s cannot be read",
1295 sha1_to_hex(trg_entry->idx.sha1));
1296 if (sz != trg_size)
1297 die("object %s inconsistent object length (%lu vs %lu)",
1298 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1299 *mem_usage += sz;
1301 if (!src->data) {
1302 read_lock();
1303 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1304 read_unlock();
1305 if (!src->data)
1306 die("object %s cannot be read",
1307 sha1_to_hex(src_entry->idx.sha1));
1308 if (sz != src_size)
1309 die("object %s inconsistent object length (%lu vs %lu)",
1310 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1311 *mem_usage += sz;
1313 if (!src->index) {
1314 src->index = create_delta_index(src->data, src_size);
1315 if (!src->index) {
1316 static int warned = 0;
1317 if (!warned++)
1318 warning("suboptimal pack - out of memory");
1319 return 0;
1321 *mem_usage += sizeof_delta_index(src->index);
1324 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1325 if (!delta_buf)
1326 return 0;
1328 if (trg_entry->delta) {
1329 /* Prefer only shallower same-sized deltas. */
1330 if (delta_size == trg_entry->delta_size &&
1331 src->depth + 1 >= trg->depth) {
1332 free(delta_buf);
1333 return 0;
1338 * Handle memory allocation outside of the cache
1339 * accounting lock. Compiler will optimize the strangeness
1340 * away when THREADED_DELTA_SEARCH is not defined.
1342 free(trg_entry->delta_data);
1343 cache_lock();
1344 if (trg_entry->delta_data) {
1345 delta_cache_size -= trg_entry->delta_size;
1346 trg_entry->delta_data = NULL;
1348 if (delta_cacheable(src_size, trg_size, delta_size)) {
1349 delta_cache_size += delta_size;
1350 cache_unlock();
1351 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1352 } else {
1353 cache_unlock();
1354 free(delta_buf);
1357 trg_entry->delta = src_entry;
1358 trg_entry->delta_size = delta_size;
1359 trg->depth = src->depth + 1;
1361 return 1;
1364 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1366 struct object_entry *child = me->delta_child;
1367 unsigned int m = n;
1368 while (child) {
1369 unsigned int c = check_delta_limit(child, n + 1);
1370 if (m < c)
1371 m = c;
1372 child = child->delta_sibling;
1374 return m;
1377 static unsigned long free_unpacked(struct unpacked *n)
1379 unsigned long freed_mem = sizeof_delta_index(n->index);
1380 free_delta_index(n->index);
1381 n->index = NULL;
1382 if (n->data) {
1383 freed_mem += n->entry->size;
1384 free(n->data);
1385 n->data = NULL;
1387 n->entry = NULL;
1388 n->depth = 0;
1389 return freed_mem;
1392 static void find_deltas(struct object_entry **list, unsigned *list_size,
1393 int window, int depth, unsigned *processed)
1395 uint32_t i, idx = 0, count = 0;
1396 unsigned int array_size = window * sizeof(struct unpacked);
1397 struct unpacked *array;
1398 unsigned long mem_usage = 0;
1400 array = xmalloc(array_size);
1401 memset(array, 0, array_size);
1403 for (;;) {
1404 struct object_entry *entry = *list++;
1405 struct unpacked *n = array + idx;
1406 int j, max_depth, best_base = -1;
1408 progress_lock();
1409 if (!*list_size) {
1410 progress_unlock();
1411 break;
1413 (*list_size)--;
1414 if (!entry->preferred_base) {
1415 (*processed)++;
1416 display_progress(progress_state, *processed);
1418 progress_unlock();
1420 mem_usage -= free_unpacked(n);
1421 n->entry = entry;
1423 while (window_memory_limit &&
1424 mem_usage > window_memory_limit &&
1425 count > 1) {
1426 uint32_t tail = (idx + window - count) % window;
1427 mem_usage -= free_unpacked(array + tail);
1428 count--;
1431 /* We do not compute delta to *create* objects we are not
1432 * going to pack.
1434 if (entry->preferred_base)
1435 goto next;
1438 * If the current object is at pack edge, take the depth the
1439 * objects that depend on the current object into account
1440 * otherwise they would become too deep.
1442 max_depth = depth;
1443 if (entry->delta_child) {
1444 max_depth -= check_delta_limit(entry, 0);
1445 if (max_depth <= 0)
1446 goto next;
1449 j = window;
1450 while (--j > 0) {
1451 int ret;
1452 uint32_t other_idx = idx + j;
1453 struct unpacked *m;
1454 if (other_idx >= window)
1455 other_idx -= window;
1456 m = array + other_idx;
1457 if (!m->entry)
1458 break;
1459 ret = try_delta(n, m, max_depth, &mem_usage);
1460 if (ret < 0)
1461 break;
1462 else if (ret > 0)
1463 best_base = other_idx;
1467 * If we decided to cache the delta data, then it is best
1468 * to compress it right away. First because we have to do
1469 * it anyway, and doing it here while we're threaded will
1470 * save a lot of time in the non threaded write phase,
1471 * as well as allow for caching more deltas within
1472 * the same cache size limit.
1473 * ...
1474 * But only if not writing to stdout, since in that case
1475 * the network is most likely throttling writes anyway,
1476 * and therefore it is best to go to the write phase ASAP
1477 * instead, as we can afford spending more time compressing
1478 * between writes at that moment.
1480 if (entry->delta_data && !pack_to_stdout) {
1481 entry->z_delta_size = do_compress(&entry->delta_data,
1482 entry->delta_size);
1483 cache_lock();
1484 delta_cache_size -= entry->delta_size;
1485 delta_cache_size += entry->z_delta_size;
1486 cache_unlock();
1489 /* if we made n a delta, and if n is already at max
1490 * depth, leaving it in the window is pointless. we
1491 * should evict it first.
1493 if (entry->delta && max_depth <= n->depth)
1494 continue;
1497 * Move the best delta base up in the window, after the
1498 * currently deltified object, to keep it longer. It will
1499 * be the first base object to be attempted next.
1501 if (entry->delta) {
1502 struct unpacked swap = array[best_base];
1503 int dist = (window + idx - best_base) % window;
1504 int dst = best_base;
1505 while (dist--) {
1506 int src = (dst + 1) % window;
1507 array[dst] = array[src];
1508 dst = src;
1510 array[dst] = swap;
1513 next:
1514 idx++;
1515 if (count + 1 < window)
1516 count++;
1517 if (idx >= window)
1518 idx = 0;
1521 for (i = 0; i < window; ++i) {
1522 free_delta_index(array[i].index);
1523 free(array[i].data);
1525 free(array);
1528 #ifdef THREADED_DELTA_SEARCH
1531 * The main thread waits on the condition that (at least) one of the workers
1532 * has stopped working (which is indicated in the .working member of
1533 * struct thread_params).
1534 * When a work thread has completed its work, it sets .working to 0 and
1535 * signals the main thread and waits on the condition that .data_ready
1536 * becomes 1.
1539 struct thread_params {
1540 pthread_t thread;
1541 struct object_entry **list;
1542 unsigned list_size;
1543 unsigned remaining;
1544 int window;
1545 int depth;
1546 int working;
1547 int data_ready;
1548 pthread_mutex_t mutex;
1549 pthread_cond_t cond;
1550 unsigned *processed;
1553 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1555 static void *threaded_find_deltas(void *arg)
1557 struct thread_params *me = arg;
1559 while (me->remaining) {
1560 find_deltas(me->list, &me->remaining,
1561 me->window, me->depth, me->processed);
1563 progress_lock();
1564 me->working = 0;
1565 pthread_cond_signal(&progress_cond);
1566 progress_unlock();
1569 * We must not set ->data_ready before we wait on the
1570 * condition because the main thread may have set it to 1
1571 * before we get here. In order to be sure that new
1572 * work is available if we see 1 in ->data_ready, it
1573 * was initialized to 0 before this thread was spawned
1574 * and we reset it to 0 right away.
1576 pthread_mutex_lock(&me->mutex);
1577 while (!me->data_ready)
1578 pthread_cond_wait(&me->cond, &me->mutex);
1579 me->data_ready = 0;
1580 pthread_mutex_unlock(&me->mutex);
1582 /* leave ->working 1 so that this doesn't get more work assigned */
1583 return NULL;
1586 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1587 int window, int depth, unsigned *processed)
1589 struct thread_params p[delta_search_threads];
1590 int i, ret, active_threads = 0;
1592 if (delta_search_threads <= 1) {
1593 find_deltas(list, &list_size, window, depth, processed);
1594 return;
1597 /* Partition the work amongst work threads. */
1598 for (i = 0; i < delta_search_threads; i++) {
1599 unsigned sub_size = list_size / (delta_search_threads - i);
1601 p[i].window = window;
1602 p[i].depth = depth;
1603 p[i].processed = processed;
1604 p[i].working = 1;
1605 p[i].data_ready = 0;
1607 /* try to split chunks on "path" boundaries */
1608 while (sub_size && sub_size < list_size &&
1609 list[sub_size]->hash &&
1610 list[sub_size]->hash == list[sub_size-1]->hash)
1611 sub_size++;
1613 p[i].list = list;
1614 p[i].list_size = sub_size;
1615 p[i].remaining = sub_size;
1617 list += sub_size;
1618 list_size -= sub_size;
1621 /* Start work threads. */
1622 for (i = 0; i < delta_search_threads; i++) {
1623 if (!p[i].list_size)
1624 continue;
1625 pthread_mutex_init(&p[i].mutex, NULL);
1626 pthread_cond_init(&p[i].cond, NULL);
1627 ret = pthread_create(&p[i].thread, NULL,
1628 threaded_find_deltas, &p[i]);
1629 if (ret)
1630 die("unable to create thread: %s", strerror(ret));
1631 active_threads++;
1635 * Now let's wait for work completion. Each time a thread is done
1636 * with its work, we steal half of the remaining work from the
1637 * thread with the largest number of unprocessed objects and give
1638 * it to that newly idle thread. This ensure good load balancing
1639 * until the remaining object list segments are simply too short
1640 * to be worth splitting anymore.
1642 while (active_threads) {
1643 struct thread_params *target = NULL;
1644 struct thread_params *victim = NULL;
1645 unsigned sub_size = 0;
1647 progress_lock();
1648 for (;;) {
1649 for (i = 0; !target && i < delta_search_threads; i++)
1650 if (!p[i].working)
1651 target = &p[i];
1652 if (target)
1653 break;
1654 pthread_cond_wait(&progress_cond, &progress_mutex);
1657 for (i = 0; i < delta_search_threads; i++)
1658 if (p[i].remaining > 2*window &&
1659 (!victim || victim->remaining < p[i].remaining))
1660 victim = &p[i];
1661 if (victim) {
1662 sub_size = victim->remaining / 2;
1663 list = victim->list + victim->list_size - sub_size;
1664 while (sub_size && list[0]->hash &&
1665 list[0]->hash == list[-1]->hash) {
1666 list++;
1667 sub_size--;
1669 if (!sub_size) {
1671 * It is possible for some "paths" to have
1672 * so many objects that no hash boundary
1673 * might be found. Let's just steal the
1674 * exact half in that case.
1676 sub_size = victim->remaining / 2;
1677 list -= sub_size;
1679 target->list = list;
1680 victim->list_size -= sub_size;
1681 victim->remaining -= sub_size;
1683 target->list_size = sub_size;
1684 target->remaining = sub_size;
1685 target->working = 1;
1686 progress_unlock();
1688 pthread_mutex_lock(&target->mutex);
1689 target->data_ready = 1;
1690 pthread_cond_signal(&target->cond);
1691 pthread_mutex_unlock(&target->mutex);
1693 if (!sub_size) {
1694 pthread_join(target->thread, NULL);
1695 pthread_cond_destroy(&target->cond);
1696 pthread_mutex_destroy(&target->mutex);
1697 active_threads--;
1702 #else
1703 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1704 #endif
1706 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1708 unsigned char peeled[20];
1710 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1711 !peel_ref(path, peeled) && /* peelable? */
1712 !is_null_sha1(peeled) && /* annotated tag? */
1713 locate_object_entry(peeled)) /* object packed? */
1714 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1715 return 0;
1718 static void prepare_pack(int window, int depth)
1720 struct object_entry **delta_list;
1721 uint32_t i, n, nr_deltas;
1723 get_object_details();
1725 if (!nr_objects || !window || !depth)
1726 return;
1728 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1729 nr_deltas = n = 0;
1731 for (i = 0; i < nr_objects; i++) {
1732 struct object_entry *entry = objects + i;
1734 if (entry->delta)
1735 /* This happens if we decided to reuse existing
1736 * delta from a pack. "reuse_delta &&" is implied.
1738 continue;
1740 if (entry->size < 50)
1741 continue;
1743 if (entry->no_try_delta)
1744 continue;
1746 if (!entry->preferred_base)
1747 nr_deltas++;
1749 delta_list[n++] = entry;
1752 if (nr_deltas && n > 1) {
1753 unsigned nr_done = 0;
1754 if (progress)
1755 progress_state = start_progress("Compressing objects",
1756 nr_deltas);
1757 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1758 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1759 stop_progress(&progress_state);
1760 if (nr_done != nr_deltas)
1761 die("inconsistency with delta count");
1763 free(delta_list);
1766 static int git_pack_config(const char *k, const char *v, void *cb)
1768 if(!strcmp(k, "pack.window")) {
1769 window = git_config_int(k, v);
1770 return 0;
1772 if (!strcmp(k, "pack.windowmemory")) {
1773 window_memory_limit = git_config_ulong(k, v);
1774 return 0;
1776 if (!strcmp(k, "pack.depth")) {
1777 depth = git_config_int(k, v);
1778 return 0;
1780 if (!strcmp(k, "pack.compression")) {
1781 int level = git_config_int(k, v);
1782 if (level == -1)
1783 level = Z_DEFAULT_COMPRESSION;
1784 else if (level < 0 || level > Z_BEST_COMPRESSION)
1785 die("bad pack compression level %d", level);
1786 pack_compression_level = level;
1787 pack_compression_seen = 1;
1788 return 0;
1790 if (!strcmp(k, "pack.deltacachesize")) {
1791 max_delta_cache_size = git_config_int(k, v);
1792 return 0;
1794 if (!strcmp(k, "pack.deltacachelimit")) {
1795 cache_max_small_delta_size = git_config_int(k, v);
1796 return 0;
1798 if (!strcmp(k, "pack.threads")) {
1799 delta_search_threads = git_config_int(k, v);
1800 if (delta_search_threads < 0)
1801 die("invalid number of threads specified (%d)",
1802 delta_search_threads);
1803 #ifndef THREADED_DELTA_SEARCH
1804 if (delta_search_threads != 1)
1805 warning("no threads support, ignoring %s", k);
1806 #endif
1807 return 0;
1809 if (!strcmp(k, "pack.indexversion")) {
1810 pack_idx_default_version = git_config_int(k, v);
1811 if (pack_idx_default_version > 2)
1812 die("bad pack.indexversion=%d", pack_idx_default_version);
1813 return 0;
1815 if (!strcmp(k, "pack.packsizelimit")) {
1816 pack_size_limit_cfg = git_config_ulong(k, v);
1817 return 0;
1819 return git_default_config(k, v, cb);
1822 static void read_object_list_from_stdin(void)
1824 char line[40 + 1 + PATH_MAX + 2];
1825 unsigned char sha1[20];
1827 for (;;) {
1828 if (!fgets(line, sizeof(line), stdin)) {
1829 if (feof(stdin))
1830 break;
1831 if (!ferror(stdin))
1832 die("fgets returned NULL, not EOF, not error!");
1833 if (errno != EINTR)
1834 die("fgets: %s", strerror(errno));
1835 clearerr(stdin);
1836 continue;
1838 if (line[0] == '-') {
1839 if (get_sha1_hex(line+1, sha1))
1840 die("expected edge sha1, got garbage:\n %s",
1841 line);
1842 add_preferred_base(sha1);
1843 continue;
1845 if (get_sha1_hex(line, sha1))
1846 die("expected sha1, got garbage:\n %s", line);
1848 add_preferred_base_object(line+41);
1849 add_object_entry(sha1, 0, line+41, 0);
1853 #define OBJECT_ADDED (1u<<20)
1855 static void show_commit(struct commit *commit)
1857 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1858 commit->object.flags |= OBJECT_ADDED;
1861 static void show_object(struct object_array_entry *p)
1863 add_preferred_base_object(p->name);
1864 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1865 p->item->flags |= OBJECT_ADDED;
1868 static void show_edge(struct commit *commit)
1870 add_preferred_base(commit->object.sha1);
1873 struct in_pack_object {
1874 off_t offset;
1875 struct object *object;
1878 struct in_pack {
1879 int alloc;
1880 int nr;
1881 struct in_pack_object *array;
1884 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1886 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1887 in_pack->array[in_pack->nr].object = object;
1888 in_pack->nr++;
1892 * Compare the objects in the offset order, in order to emulate the
1893 * "git-rev-list --objects" output that produced the pack originally.
1895 static int ofscmp(const void *a_, const void *b_)
1897 struct in_pack_object *a = (struct in_pack_object *)a_;
1898 struct in_pack_object *b = (struct in_pack_object *)b_;
1900 if (a->offset < b->offset)
1901 return -1;
1902 else if (a->offset > b->offset)
1903 return 1;
1904 else
1905 return hashcmp(a->object->sha1, b->object->sha1);
1908 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1910 struct packed_git *p;
1911 struct in_pack in_pack;
1912 uint32_t i;
1914 memset(&in_pack, 0, sizeof(in_pack));
1916 for (p = packed_git; p; p = p->next) {
1917 const unsigned char *sha1;
1918 struct object *o;
1920 for (i = 0; i < revs->num_ignore_packed; i++) {
1921 if (matches_pack_name(p, revs->ignore_packed[i]))
1922 break;
1924 if (revs->num_ignore_packed <= i)
1925 continue;
1926 if (open_pack_index(p))
1927 die("cannot open pack index");
1929 ALLOC_GROW(in_pack.array,
1930 in_pack.nr + p->num_objects,
1931 in_pack.alloc);
1933 for (i = 0; i < p->num_objects; i++) {
1934 sha1 = nth_packed_object_sha1(p, i);
1935 o = lookup_unknown_object(sha1);
1936 if (!(o->flags & OBJECT_ADDED))
1937 mark_in_pack_object(o, p, &in_pack);
1938 o->flags |= OBJECT_ADDED;
1942 if (in_pack.nr) {
1943 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1944 ofscmp);
1945 for (i = 0; i < in_pack.nr; i++) {
1946 struct object *o = in_pack.array[i].object;
1947 add_object_entry(o->sha1, o->type, "", 0);
1950 free(in_pack.array);
1953 static void loosen_unused_packed_objects(struct rev_info *revs)
1955 struct packed_git *p;
1956 uint32_t i;
1957 const unsigned char *sha1;
1959 for (p = packed_git; p; p = p->next) {
1960 for (i = 0; i < revs->num_ignore_packed; i++) {
1961 if (matches_pack_name(p, revs->ignore_packed[i]))
1962 break;
1964 if (revs->num_ignore_packed <= i)
1965 continue;
1967 if (open_pack_index(p))
1968 die("cannot open pack index");
1970 for (i = 0; i < p->num_objects; i++) {
1971 sha1 = nth_packed_object_sha1(p, i);
1972 if (!locate_object_entry(sha1))
1973 if (force_object_loose(sha1, p->mtime))
1974 die("unable to force loose object");
1979 static void get_object_list(int ac, const char **av)
1981 struct rev_info revs;
1982 char line[1000];
1983 int flags = 0;
1985 init_revisions(&revs, NULL);
1986 save_commit_buffer = 0;
1987 setup_revisions(ac, av, &revs, NULL);
1989 while (fgets(line, sizeof(line), stdin) != NULL) {
1990 int len = strlen(line);
1991 if (len && line[len - 1] == '\n')
1992 line[--len] = 0;
1993 if (!len)
1994 break;
1995 if (*line == '-') {
1996 if (!strcmp(line, "--not")) {
1997 flags ^= UNINTERESTING;
1998 continue;
2000 die("not a rev '%s'", line);
2002 if (handle_revision_arg(line, &revs, flags, 1))
2003 die("bad revision '%s'", line);
2006 if (prepare_revision_walk(&revs))
2007 die("revision walk setup failed");
2008 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2009 traverse_commit_list(&revs, show_commit, show_object);
2011 if (keep_unreachable)
2012 add_objects_in_unpacked_packs(&revs);
2013 if (unpack_unreachable)
2014 loosen_unused_packed_objects(&revs);
2017 static int adjust_perm(const char *path, mode_t mode)
2019 if (chmod(path, mode))
2020 return -1;
2021 return adjust_shared_perm(path);
2024 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2026 int use_internal_rev_list = 0;
2027 int thin = 0;
2028 uint32_t i;
2029 const char **rp_av;
2030 int rp_ac_alloc = 64;
2031 int rp_ac;
2033 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2035 rp_av[0] = "pack-objects";
2036 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2037 rp_ac = 2;
2039 git_config(git_pack_config, NULL);
2040 if (!pack_compression_seen && core_compression_seen)
2041 pack_compression_level = core_compression_level;
2043 progress = isatty(2);
2044 for (i = 1; i < argc; i++) {
2045 const char *arg = argv[i];
2047 if (*arg != '-')
2048 break;
2050 if (!strcmp("--non-empty", arg)) {
2051 non_empty = 1;
2052 continue;
2054 if (!strcmp("--local", arg)) {
2055 local = 1;
2056 continue;
2058 if (!strcmp("--incremental", arg)) {
2059 incremental = 1;
2060 continue;
2062 if (!prefixcmp(arg, "--compression=")) {
2063 char *end;
2064 int level = strtoul(arg+14, &end, 0);
2065 if (!arg[14] || *end)
2066 usage(pack_usage);
2067 if (level == -1)
2068 level = Z_DEFAULT_COMPRESSION;
2069 else if (level < 0 || level > Z_BEST_COMPRESSION)
2070 die("bad pack compression level %d", level);
2071 pack_compression_level = level;
2072 continue;
2074 if (!prefixcmp(arg, "--max-pack-size=")) {
2075 char *end;
2076 pack_size_limit_cfg = 0;
2077 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2078 if (!arg[16] || *end)
2079 usage(pack_usage);
2080 continue;
2082 if (!prefixcmp(arg, "--window=")) {
2083 char *end;
2084 window = strtoul(arg+9, &end, 0);
2085 if (!arg[9] || *end)
2086 usage(pack_usage);
2087 continue;
2089 if (!prefixcmp(arg, "--window-memory=")) {
2090 if (!git_parse_ulong(arg+16, &window_memory_limit))
2091 usage(pack_usage);
2092 continue;
2094 if (!prefixcmp(arg, "--threads=")) {
2095 char *end;
2096 delta_search_threads = strtoul(arg+10, &end, 0);
2097 if (!arg[10] || *end || delta_search_threads < 0)
2098 usage(pack_usage);
2099 #ifndef THREADED_DELTA_SEARCH
2100 if (delta_search_threads != 1)
2101 warning("no threads support, "
2102 "ignoring %s", arg);
2103 #endif
2104 continue;
2106 if (!prefixcmp(arg, "--depth=")) {
2107 char *end;
2108 depth = strtoul(arg+8, &end, 0);
2109 if (!arg[8] || *end)
2110 usage(pack_usage);
2111 continue;
2113 if (!strcmp("--progress", arg)) {
2114 progress = 1;
2115 continue;
2117 if (!strcmp("--all-progress", arg)) {
2118 progress = 2;
2119 continue;
2121 if (!strcmp("-q", arg)) {
2122 progress = 0;
2123 continue;
2125 if (!strcmp("--no-reuse-delta", arg)) {
2126 reuse_delta = 0;
2127 continue;
2129 if (!strcmp("--no-reuse-object", arg)) {
2130 reuse_object = reuse_delta = 0;
2131 continue;
2133 if (!strcmp("--delta-base-offset", arg)) {
2134 allow_ofs_delta = 1;
2135 continue;
2137 if (!strcmp("--stdout", arg)) {
2138 pack_to_stdout = 1;
2139 continue;
2141 if (!strcmp("--revs", arg)) {
2142 use_internal_rev_list = 1;
2143 continue;
2145 if (!strcmp("--keep-unreachable", arg)) {
2146 keep_unreachable = 1;
2147 continue;
2149 if (!strcmp("--unpack-unreachable", arg)) {
2150 unpack_unreachable = 1;
2151 continue;
2153 if (!strcmp("--include-tag", arg)) {
2154 include_tag = 1;
2155 continue;
2157 if (!strcmp("--unpacked", arg) ||
2158 !prefixcmp(arg, "--unpacked=") ||
2159 !strcmp("--reflog", arg) ||
2160 !strcmp("--all", arg)) {
2161 use_internal_rev_list = 1;
2162 if (rp_ac >= rp_ac_alloc - 1) {
2163 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2164 rp_av = xrealloc(rp_av,
2165 rp_ac_alloc * sizeof(*rp_av));
2167 rp_av[rp_ac++] = arg;
2168 continue;
2170 if (!strcmp("--thin", arg)) {
2171 use_internal_rev_list = 1;
2172 thin = 1;
2173 rp_av[1] = "--objects-edge";
2174 continue;
2176 if (!prefixcmp(arg, "--index-version=")) {
2177 char *c;
2178 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2179 if (pack_idx_default_version > 2)
2180 die("bad %s", arg);
2181 if (*c == ',')
2182 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2183 if (*c || pack_idx_off32_limit & 0x80000000)
2184 die("bad %s", arg);
2185 continue;
2187 usage(pack_usage);
2190 /* Traditionally "pack-objects [options] base extra" failed;
2191 * we would however want to take refs parameter that would
2192 * have been given to upstream rev-list ourselves, which means
2193 * we somehow want to say what the base name is. So the
2194 * syntax would be:
2196 * pack-objects [options] base <refs...>
2198 * in other words, we would treat the first non-option as the
2199 * base_name and send everything else to the internal revision
2200 * walker.
2203 if (!pack_to_stdout)
2204 base_name = argv[i++];
2206 if (pack_to_stdout != !base_name)
2207 usage(pack_usage);
2209 if (!pack_to_stdout && !pack_size_limit)
2210 pack_size_limit = pack_size_limit_cfg;
2212 if (pack_to_stdout && pack_size_limit)
2213 die("--max-pack-size cannot be used to build a pack for transfer.");
2215 if (!pack_to_stdout && thin)
2216 die("--thin cannot be used to build an indexable pack.");
2218 if (keep_unreachable && unpack_unreachable)
2219 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2221 #ifdef THREADED_DELTA_SEARCH
2222 if (!delta_search_threads) /* --threads=0 means autodetect */
2223 delta_search_threads = online_cpus();
2224 #endif
2226 prepare_packed_git();
2228 if (progress)
2229 progress_state = start_progress("Counting objects", 0);
2230 if (!use_internal_rev_list)
2231 read_object_list_from_stdin();
2232 else {
2233 rp_av[rp_ac] = NULL;
2234 get_object_list(rp_ac, rp_av);
2236 if (include_tag && nr_result)
2237 for_each_ref(add_ref_tag, NULL);
2238 stop_progress(&progress_state);
2240 if (non_empty && !nr_result)
2241 return 0;
2242 if (nr_result)
2243 prepare_pack(window, depth);
2244 write_pack_file();
2245 if (progress)
2246 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2247 written, written_delta, reused, reused_delta);
2248 return 0;