![[tokkee]](http://tokkee.org/images/avatar.png){kind=avatar}
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
44 */
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.
55 */
56 unsigned char no_try_delta;
57 };
59 /*
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.
64 */
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;
93 /*
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.
97 */
98 static int *object_ix;
99 static int object_ix_hashsz;
101 /*
102 * stats
103 */
104 static uint32_t written, written_delta;
105 static uint32_t reused, reused_delta;
108 static void *get_delta(struct object_entry *entry)
109 {
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;
127 }
129 static unsigned long do_compress(void **pptr, unsigned long size)
130 {
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;
153 }
155 /*
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"
161 */
162 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
163 {
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++;
177 }
178 *hdr = c;
179 return n;
180 }
182 /*
183 * we are going to reuse the existing object data as is. make
184 * sure it is not corrupt.
185 */
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)
191 {
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;
210 }
212 static void copy_pack_data(struct sha1file *f,
213 struct packed_git *p,
214 struct pack_window **w_curs,
215 off_t offset,
216 off_t len)
217 {
218 unsigned char *in;
219 unsigned int avail;
221 while (len) {
222 in = use_pack(p, w_curs, offset, &avail);
223 if (avail > len)
224 avail = (unsigned int)len;
225 sha1write(f, in, avail);
226 offset += avail;
227 len -= avail;
228 }
229 }
231 static unsigned long write_object(struct sha1file *f,
232 struct object_entry *entry,
233 off_t write_offset)
234 {
235 unsigned long size, limit, datalen;
236 void *buf;
237 unsigned char header[10], dheader[10];
238 unsigned hdrlen;
239 enum object_type type;
240 int usable_delta, to_reuse;
242 if (!pack_to_stdout)
243 crc32_begin(f);
245 type = entry->type;
247 /* write limit if limited packsize and not first object */
248 limit = pack_size_limit && nr_written ?
249 pack_size_limit - write_offset : 0;
251 if (!entry->delta)
252 usable_delta = 0; /* no delta */
253 else if (!pack_size_limit)
254 usable_delta = 1; /* unlimited packfile */
255 else if (entry->delta->idx.offset == (off_t)-1)
256 usable_delta = 0; /* base was written to another pack */
257 else if (entry->delta->idx.offset)
258 usable_delta = 1; /* base already exists in this pack */
259 else
260 usable_delta = 0; /* base could end up in another pack */
262 if (!reuse_object)
263 to_reuse = 0; /* explicit */
264 else if (!entry->in_pack)
265 to_reuse = 0; /* can't reuse what we don't have */
266 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
267 /* check_object() decided it for us ... */
268 to_reuse = usable_delta;
269 /* ... but pack split may override that */
270 else if (type != entry->in_pack_type)
271 to_reuse = 0; /* pack has delta which is unusable */
272 else if (entry->delta)
273 to_reuse = 0; /* we want to pack afresh */
274 else
275 to_reuse = 1; /* we have it in-pack undeltified,
276 * and we do not need to deltify it.
277 */
279 if (!to_reuse) {
280 if (!usable_delta) {
281 buf = read_sha1_file(entry->idx.sha1, &type, &size);
282 if (!buf)
283 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
284 /*
285 * make sure no cached delta data remains from a
286 * previous attempt before a pack split occured.
287 */
288 free(entry->delta_data);
289 entry->delta_data = NULL;
290 entry->z_delta_size = 0;
291 } else if (entry->delta_data) {
292 size = entry->delta_size;
293 buf = entry->delta_data;
294 entry->delta_data = NULL;
295 type = (allow_ofs_delta && entry->delta->idx.offset) ?
296 OBJ_OFS_DELTA : OBJ_REF_DELTA;
297 } else {
298 buf = get_delta(entry);
299 size = entry->delta_size;
300 type = (allow_ofs_delta && entry->delta->idx.offset) ?
301 OBJ_OFS_DELTA : OBJ_REF_DELTA;
302 }
304 if (entry->z_delta_size)
305 datalen = entry->z_delta_size;
306 else
307 datalen = do_compress(&buf, size);
309 /*
310 * The object header is a byte of 'type' followed by zero or
311 * more bytes of length.
312 */
313 hdrlen = encode_header(type, size, header);
315 if (type == OBJ_OFS_DELTA) {
316 /*
317 * Deltas with relative base contain an additional
318 * encoding of the relative offset for the delta
319 * base from this object's position in the pack.
320 */
321 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
322 unsigned pos = sizeof(dheader) - 1;
323 dheader[pos] = ofs & 127;
324 while (ofs >>= 7)
325 dheader[--pos] = 128 | (--ofs & 127);
326 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
327 free(buf);
328 return 0;
329 }
330 sha1write(f, header, hdrlen);
331 sha1write(f, dheader + pos, sizeof(dheader) - pos);
332 hdrlen += sizeof(dheader) - pos;
333 } else if (type == OBJ_REF_DELTA) {
334 /*
335 * Deltas with a base reference contain
336 * an additional 20 bytes for the base sha1.
337 */
338 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
339 free(buf);
340 return 0;
341 }
342 sha1write(f, header, hdrlen);
343 sha1write(f, entry->delta->idx.sha1, 20);
344 hdrlen += 20;
345 } else {
346 if (limit && hdrlen + datalen + 20 >= limit) {
347 free(buf);
348 return 0;
349 }
350 sha1write(f, header, hdrlen);
351 }
352 sha1write(f, buf, datalen);
353 free(buf);
354 }
355 else {
356 struct packed_git *p = entry->in_pack;
357 struct pack_window *w_curs = NULL;
358 struct revindex_entry *revidx;
359 off_t offset;
361 if (entry->delta) {
362 type = (allow_ofs_delta && entry->delta->idx.offset) ?
363 OBJ_OFS_DELTA : OBJ_REF_DELTA;
364 reused_delta++;
365 }
366 hdrlen = encode_header(type, entry->size, header);
367 offset = entry->in_pack_offset;
368 revidx = find_pack_revindex(p, offset);
369 datalen = revidx[1].offset - offset;
370 if (!pack_to_stdout && p->index_version > 1 &&
371 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
372 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
373 offset += entry->in_pack_header_size;
374 datalen -= entry->in_pack_header_size;
375 if (type == OBJ_OFS_DELTA) {
376 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
377 unsigned pos = sizeof(dheader) - 1;
378 dheader[pos] = ofs & 127;
379 while (ofs >>= 7)
380 dheader[--pos] = 128 | (--ofs & 127);
381 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
382 return 0;
383 sha1write(f, header, hdrlen);
384 sha1write(f, dheader + pos, sizeof(dheader) - pos);
385 hdrlen += sizeof(dheader) - pos;
386 } else if (type == OBJ_REF_DELTA) {
387 if (limit && hdrlen + 20 + datalen + 20 >= limit)
388 return 0;
389 sha1write(f, header, hdrlen);
390 sha1write(f, entry->delta->idx.sha1, 20);
391 hdrlen += 20;
392 } else {
393 if (limit && hdrlen + datalen + 20 >= limit)
394 return 0;
395 sha1write(f, header, hdrlen);
396 }
398 if (!pack_to_stdout && p->index_version == 1 &&
399 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
400 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
401 copy_pack_data(f, p, &w_curs, offset, datalen);
402 unuse_pack(&w_curs);
403 reused++;
404 }
405 if (usable_delta)
406 written_delta++;
407 written++;
408 if (!pack_to_stdout)
409 entry->idx.crc32 = crc32_end(f);
410 return hdrlen + datalen;
411 }
413 static int write_one(struct sha1file *f,
414 struct object_entry *e,
415 off_t *offset)
416 {
417 unsigned long size;
419 /* offset is non zero if object is written already. */
420 if (e->idx.offset || e->preferred_base)
421 return 1;
423 /* if we are deltified, write out base object first. */
424 if (e->delta && !write_one(f, e->delta, offset))
425 return 0;
427 e->idx.offset = *offset;
428 size = write_object(f, e, *offset);
429 if (!size) {
430 e->idx.offset = 0;
431 return 0;
432 }
433 written_list[nr_written++] = &e->idx;
435 /* make sure off_t is sufficiently large not to wrap */
436 if (*offset > *offset + size)
437 die("pack too large for current definition of off_t");
438 *offset += size;
439 return 1;
440 }
442 /* forward declaration for write_pack_file */
443 static int adjust_perm(const char *path, mode_t mode);
445 static void write_pack_file(void)
446 {
447 uint32_t i = 0, j;
448 struct sha1file *f;
449 off_t offset;
450 struct pack_header hdr;
451 uint32_t nr_remaining = nr_result;
452 time_t last_mtime = 0;
454 if (progress > pack_to_stdout)
455 progress_state = start_progress("Writing objects", nr_result);
456 written_list = xmalloc(nr_objects * sizeof(*written_list));
458 do {
459 unsigned char sha1[20];
460 char *pack_tmp_name = NULL;
462 if (pack_to_stdout) {
463 f = sha1fd_throughput(1, "<stdout>", progress_state);
464 } else {
465 char tmpname[PATH_MAX];
466 int fd;
467 snprintf(tmpname, sizeof(tmpname),
468 "%s/tmp_pack_XXXXXX", get_object_directory());
469 fd = xmkstemp(tmpname);
470 pack_tmp_name = xstrdup(tmpname);
471 f = sha1fd(fd, pack_tmp_name);
472 }
474 hdr.hdr_signature = htonl(PACK_SIGNATURE);
475 hdr.hdr_version = htonl(PACK_VERSION);
476 hdr.hdr_entries = htonl(nr_remaining);
477 sha1write(f, &hdr, sizeof(hdr));
478 offset = sizeof(hdr);
479 nr_written = 0;
480 for (; i < nr_objects; i++) {
481 if (!write_one(f, objects + i, &offset))
482 break;
483 display_progress(progress_state, written);
484 }
486 /*
487 * Did we write the wrong # entries in the header?
488 * If so, rewrite it like in fast-import
489 */
490 if (pack_to_stdout) {
491 sha1close(f, sha1, CSUM_CLOSE);
492 } else if (nr_written == nr_remaining) {
493 sha1close(f, sha1, CSUM_FSYNC);
494 } else {
495 int fd = sha1close(f, NULL, 0);
496 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
497 nr_written, NULL, 0);
498 close(fd);
499 }
501 if (!pack_to_stdout) {
502 mode_t mode = umask(0);
503 struct stat st;
504 char *idx_tmp_name, tmpname[PATH_MAX];
506 umask(mode);
507 mode = 0444 & ~mode;
509 idx_tmp_name = write_idx_file(NULL, written_list,
510 nr_written, sha1);
512 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
513 base_name, sha1_to_hex(sha1));
514 if (adjust_perm(pack_tmp_name, mode))
515 die("unable to make temporary pack file readable: %s",
516 strerror(errno));
517 if (rename(pack_tmp_name, tmpname))
518 die("unable to rename temporary pack file: %s",
519 strerror(errno));
521 /*
522 * Packs are runtime accessed in their mtime
523 * order since newer packs are more likely to contain
524 * younger objects. So if we are creating multiple
525 * packs then we should modify the mtime of later ones
526 * to preserve this property.
527 */
528 if (stat(tmpname, &st) < 0) {
529 warning("failed to stat %s: %s",
530 tmpname, strerror(errno));
531 } else if (!last_mtime) {
532 last_mtime = st.st_mtime;
533 } else {
534 struct utimbuf utb;
535 utb.actime = st.st_atime;
536 utb.modtime = --last_mtime;
537 if (utime(tmpname, &utb) < 0)
538 warning("failed utime() on %s: %s",
539 tmpname, strerror(errno));
540 }
542 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
543 base_name, sha1_to_hex(sha1));
544 if (adjust_perm(idx_tmp_name, mode))
545 die("unable to make temporary index file readable: %s",
546 strerror(errno));
547 if (rename(idx_tmp_name, tmpname))
548 die("unable to rename temporary index file: %s",
549 strerror(errno));
551 free(idx_tmp_name);
552 free(pack_tmp_name);
553 puts(sha1_to_hex(sha1));
554 }
556 /* mark written objects as written to previous pack */
557 for (j = 0; j < nr_written; j++) {
558 written_list[j]->offset = (off_t)-1;
559 }
560 nr_remaining -= nr_written;
561 } while (nr_remaining && i < nr_objects);
563 free(written_list);
564 stop_progress(&progress_state);
565 if (written != nr_result)
566 die("wrote %"PRIu32" objects while expecting %"PRIu32,
567 written, nr_result);
568 /*
569 * We have scanned through [0 ... i). Since we have written
570 * the correct number of objects, the remaining [i ... nr_objects)
571 * items must be either already written (due to out-of-order delta base)
572 * or a preferred base. Count those which are neither and complain if any.
573 */
574 for (j = 0; i < nr_objects; i++) {
575 struct object_entry *e = objects + i;
576 j += !e->idx.offset && !e->preferred_base;
577 }
578 if (j)
579 die("wrote %"PRIu32" objects as expected but %"PRIu32
580 " unwritten", written, j);
581 }
583 static int locate_object_entry_hash(const unsigned char *sha1)
584 {
585 int i;
586 unsigned int ui;
587 memcpy(&ui, sha1, sizeof(unsigned int));
588 i = ui % object_ix_hashsz;
589 while (0 < object_ix[i]) {
590 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
591 return i;
592 if (++i == object_ix_hashsz)
593 i = 0;
594 }
595 return -1 - i;
596 }
598 static struct object_entry *locate_object_entry(const unsigned char *sha1)
599 {
600 int i;
602 if (!object_ix_hashsz)
603 return NULL;
605 i = locate_object_entry_hash(sha1);
606 if (0 <= i)
607 return &objects[object_ix[i]-1];
608 return NULL;
609 }
611 static void rehash_objects(void)
612 {
613 uint32_t i;
614 struct object_entry *oe;
616 object_ix_hashsz = nr_objects * 3;
617 if (object_ix_hashsz < 1024)
618 object_ix_hashsz = 1024;
619 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
620 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
621 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
622 int ix = locate_object_entry_hash(oe->idx.sha1);
623 if (0 <= ix)
624 continue;
625 ix = -1 - ix;
626 object_ix[ix] = i + 1;
627 }
628 }
630 static unsigned name_hash(const char *name)
631 {
632 unsigned char c;
633 unsigned hash = 0;
635 if (!name)
636 return 0;
638 /*
639 * This effectively just creates a sortable number from the
640 * last sixteen non-whitespace characters. Last characters
641 * count "most", so things that end in ".c" sort together.
642 */
643 while ((c = *name++) != 0) {
644 if (isspace(c))
645 continue;
646 hash = (hash >> 2) + (c << 24);
647 }
648 return hash;
649 }
651 static void setup_delta_attr_check(struct git_attr_check *check)
652 {
653 static struct git_attr *attr_delta;
655 if (!attr_delta)
656 attr_delta = git_attr("delta", 5);
658 check[0].attr = attr_delta;
659 }
661 static int no_try_delta(const char *path)
662 {
663 struct git_attr_check check[1];
665 setup_delta_attr_check(check);
666 if (git_checkattr(path, ARRAY_SIZE(check), check))
667 return 0;
668 if (ATTR_FALSE(check->value))
669 return 1;
670 return 0;
671 }
673 static int add_object_entry(const unsigned char *sha1, enum object_type type,
674 const char *name, int exclude)
675 {
676 struct object_entry *entry;
677 struct packed_git *p, *found_pack = NULL;
678 off_t found_offset = 0;
679 int ix;
680 unsigned hash = name_hash(name);
682 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
683 if (ix >= 0) {
684 if (exclude) {
685 entry = objects + object_ix[ix] - 1;
686 if (!entry->preferred_base)
687 nr_result--;
688 entry->preferred_base = 1;
689 }
690 return 0;
691 }
693 for (p = packed_git; p; p = p->next) {
694 off_t offset = find_pack_entry_one(sha1, p);
695 if (offset) {
696 if (!found_pack) {
697 found_offset = offset;
698 found_pack = p;
699 }
700 if (exclude)
701 break;
702 if (incremental)
703 return 0;
704 if (local && !p->pack_local)
705 return 0;
706 }
707 }
709 if (nr_objects >= nr_alloc) {
710 nr_alloc = (nr_alloc + 1024) * 3 / 2;
711 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
712 }
714 entry = objects + nr_objects++;
715 memset(entry, 0, sizeof(*entry));
716 hashcpy(entry->idx.sha1, sha1);
717 entry->hash = hash;
718 if (type)
719 entry->type = type;
720 if (exclude)
721 entry->preferred_base = 1;
722 else
723 nr_result++;
724 if (found_pack) {
725 entry->in_pack = found_pack;
726 entry->in_pack_offset = found_offset;
727 }
729 if (object_ix_hashsz * 3 <= nr_objects * 4)
730 rehash_objects();
731 else
732 object_ix[-1 - ix] = nr_objects;
734 display_progress(progress_state, nr_objects);
736 if (name && no_try_delta(name))
737 entry->no_try_delta = 1;
739 return 1;
740 }
742 struct pbase_tree_cache {
743 unsigned char sha1[20];
744 int ref;
745 int temporary;
746 void *tree_data;
747 unsigned long tree_size;
748 };
750 static struct pbase_tree_cache *(pbase_tree_cache[256]);
751 static int pbase_tree_cache_ix(const unsigned char *sha1)
752 {
753 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
754 }
755 static int pbase_tree_cache_ix_incr(int ix)
756 {
757 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
758 }
760 static struct pbase_tree {
761 struct pbase_tree *next;
762 /* This is a phony "cache" entry; we are not
763 * going to evict it nor find it through _get()
764 * mechanism -- this is for the toplevel node that
765 * would almost always change with any commit.
766 */
767 struct pbase_tree_cache pcache;
768 } *pbase_tree;
770 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
771 {
772 struct pbase_tree_cache *ent, *nent;
773 void *data;
774 unsigned long size;
775 enum object_type type;
776 int neigh;
777 int my_ix = pbase_tree_cache_ix(sha1);
778 int available_ix = -1;
780 /* pbase-tree-cache acts as a limited hashtable.
781 * your object will be found at your index or within a few
782 * slots after that slot if it is cached.
783 */
784 for (neigh = 0; neigh < 8; neigh++) {
785 ent = pbase_tree_cache[my_ix];
786 if (ent && !hashcmp(ent->sha1, sha1)) {
787 ent->ref++;
788 return ent;
789 }
790 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
791 ((0 <= available_ix) &&
792 (!ent && pbase_tree_cache[available_ix])))
793 available_ix = my_ix;
794 if (!ent)
795 break;
796 my_ix = pbase_tree_cache_ix_incr(my_ix);
797 }
799 /* Did not find one. Either we got a bogus request or
800 * we need to read and perhaps cache.
801 */
802 data = read_sha1_file(sha1, &type, &size);
803 if (!data)
804 return NULL;
805 if (type != OBJ_TREE) {
806 free(data);
807 return NULL;
808 }
810 /* We need to either cache or return a throwaway copy */
812 if (available_ix < 0)
813 ent = NULL;
814 else {
815 ent = pbase_tree_cache[available_ix];
816 my_ix = available_ix;
817 }
819 if (!ent) {
820 nent = xmalloc(sizeof(*nent));
821 nent->temporary = (available_ix < 0);
822 }
823 else {
824 /* evict and reuse */
825 free(ent->tree_data);
826 nent = ent;
827 }
828 hashcpy(nent->sha1, sha1);
829 nent->tree_data = data;
830 nent->tree_size = size;
831 nent->ref = 1;
832 if (!nent->temporary)
833 pbase_tree_cache[my_ix] = nent;
834 return nent;
835 }
837 static void pbase_tree_put(struct pbase_tree_cache *cache)
838 {
839 if (!cache->temporary) {
840 cache->ref--;
841 return;
842 }
843 free(cache->tree_data);
844 free(cache);
845 }
847 static int name_cmp_len(const char *name)
848 {
849 int i;
850 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
851 ;
852 return i;
853 }
855 static void add_pbase_object(struct tree_desc *tree,
856 const char *name,
857 int cmplen,
858 const char *fullname)
859 {
860 struct name_entry entry;
861 int cmp;
863 while (tree_entry(tree,&entry)) {
864 if (S_ISGITLINK(entry.mode))
865 continue;
866 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
867 memcmp(name, entry.path, cmplen);
868 if (cmp > 0)
869 continue;
870 if (cmp < 0)
871 return;
872 if (name[cmplen] != '/') {
873 add_object_entry(entry.sha1,
874 object_type(entry.mode),
875 fullname, 1);
876 return;
877 }
878 if (S_ISDIR(entry.mode)) {
879 struct tree_desc sub;
880 struct pbase_tree_cache *tree;
881 const char *down = name+cmplen+1;
882 int downlen = name_cmp_len(down);
884 tree = pbase_tree_get(entry.sha1);
885 if (!tree)
886 return;
887 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
889 add_pbase_object(&sub, down, downlen, fullname);
890 pbase_tree_put(tree);
891 }
892 }
893 }
895 static unsigned *done_pbase_paths;
896 static int done_pbase_paths_num;
897 static int done_pbase_paths_alloc;
898 static int done_pbase_path_pos(unsigned hash)
899 {
900 int lo = 0;
901 int hi = done_pbase_paths_num;
902 while (lo < hi) {
903 int mi = (hi + lo) / 2;
904 if (done_pbase_paths[mi] == hash)
905 return mi;
906 if (done_pbase_paths[mi] < hash)
907 hi = mi;
908 else
909 lo = mi + 1;
910 }
911 return -lo-1;
912 }
914 static int check_pbase_path(unsigned hash)
915 {
916 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
917 if (0 <= pos)
918 return 1;
919 pos = -pos - 1;
920 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
921 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
922 done_pbase_paths = xrealloc(done_pbase_paths,
923 done_pbase_paths_alloc *
924 sizeof(unsigned));
925 }
926 done_pbase_paths_num++;
927 if (pos < done_pbase_paths_num)
928 memmove(done_pbase_paths + pos + 1,
929 done_pbase_paths + pos,
930 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
931 done_pbase_paths[pos] = hash;
932 return 0;
933 }
935 static void add_preferred_base_object(const char *name)
936 {
937 struct pbase_tree *it;
938 int cmplen;
939 unsigned hash = name_hash(name);
941 if (!num_preferred_base || check_pbase_path(hash))
942 return;
944 cmplen = name_cmp_len(name);
945 for (it = pbase_tree; it; it = it->next) {
946 if (cmplen == 0) {
947 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
948 }
949 else {
950 struct tree_desc tree;
951 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
952 add_pbase_object(&tree, name, cmplen, name);
953 }
954 }
955 }
957 static void add_preferred_base(unsigned char *sha1)
958 {
959 struct pbase_tree *it;
960 void *data;
961 unsigned long size;
962 unsigned char tree_sha1[20];
964 if (window <= num_preferred_base++)
965 return;
967 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
968 if (!data)
969 return;
971 for (it = pbase_tree; it; it = it->next) {
972 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
973 free(data);
974 return;
975 }
976 }
978 it = xcalloc(1, sizeof(*it));
979 it->next = pbase_tree;
980 pbase_tree = it;
982 hashcpy(it->pcache.sha1, tree_sha1);
983 it->pcache.tree_data = data;
984 it->pcache.tree_size = size;
985 }
987 static void check_object(struct object_entry *entry)
988 {
989 if (entry->in_pack) {
990 struct packed_git *p = entry->in_pack;
991 struct pack_window *w_curs = NULL;
992 const unsigned char *base_ref = NULL;
993 struct object_entry *base_entry;
994 unsigned long used, used_0;
995 unsigned int avail;
996 off_t ofs;
997 unsigned char *buf, c;
999 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1001 /*
1002 * We want in_pack_type even if we do not reuse delta
1003 * since non-delta representations could still be reused.
1004 */
1005 used = unpack_object_header_gently(buf, avail,
1006 &entry->in_pack_type,
1007 &entry->size);
1009 /*
1010 * Determine if this is a delta and if so whether we can
1011 * reuse it or not. Otherwise let's find out as cheaply as
1012 * possible what the actual type and size for this object is.
1013 */
1014 switch (entry->in_pack_type) {
1015 default:
1016 /* Not a delta hence we've already got all we need. */
1017 entry->type = entry->in_pack_type;
1018 entry->in_pack_header_size = used;
1019 unuse_pack(&w_curs);
1020 return;
1021 case OBJ_REF_DELTA:
1022 if (reuse_delta && !entry->preferred_base)
1023 base_ref = use_pack(p, &w_curs,
1024 entry->in_pack_offset + used, NULL);
1025 entry->in_pack_header_size = used + 20;
1026 break;
1027 case OBJ_OFS_DELTA:
1028 buf = use_pack(p, &w_curs,
1029 entry->in_pack_offset + used, NULL);
1030 used_0 = 0;
1031 c = buf[used_0++];
1032 ofs = c & 127;
1033 while (c & 128) {
1034 ofs += 1;
1035 if (!ofs || MSB(ofs, 7))
1036 die("delta base offset overflow in pack for %s",
1037 sha1_to_hex(entry->idx.sha1));
1038 c = buf[used_0++];
1039 ofs = (ofs << 7) + (c & 127);
1040 }
1041 if (ofs >= entry->in_pack_offset)
1042 die("delta base offset out of bound for %s",
1043 sha1_to_hex(entry->idx.sha1));
1044 ofs = entry->in_pack_offset - ofs;
1045 if (reuse_delta && !entry->preferred_base) {
1046 struct revindex_entry *revidx;
1047 revidx = find_pack_revindex(p, ofs);
1048 base_ref = nth_packed_object_sha1(p, revidx->nr);
1049 }
1050 entry->in_pack_header_size = used + used_0;
1051 break;
1052 }
1054 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1055 /*
1056 * If base_ref was set above that means we wish to
1057 * reuse delta data, and we even found that base
1058 * in the list of objects we want to pack. Goodie!
1059 *
1060 * Depth value does not matter - find_deltas() will
1061 * never consider reused delta as the base object to
1062 * deltify other objects against, in order to avoid
1063 * circular deltas.
1064 */
1065 entry->type = entry->in_pack_type;
1066 entry->delta = base_entry;
1067 entry->delta_sibling = base_entry->delta_child;
1068 base_entry->delta_child = entry;
1069 unuse_pack(&w_curs);
1070 return;
1071 }
1073 if (entry->type) {
1074 /*
1075 * This must be a delta and we already know what the
1076 * final object type is. Let's extract the actual
1077 * object size from the delta header.
1078 */
1079 entry->size = get_size_from_delta(p, &w_curs,
1080 entry->in_pack_offset + entry->in_pack_header_size);
1081 unuse_pack(&w_curs);
1082 return;
1083 }
1085 /*
1086 * No choice but to fall back to the recursive delta walk
1087 * with sha1_object_info() to find about the object type
1088 * at this point...
1089 */
1090 unuse_pack(&w_curs);
1091 }
1093 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1094 if (entry->type < 0)
1095 die("unable to get type of object %s",
1096 sha1_to_hex(entry->idx.sha1));
1097 }
1099 static int pack_offset_sort(const void *_a, const void *_b)
1100 {
1101 const struct object_entry *a = *(struct object_entry **)_a;
1102 const struct object_entry *b = *(struct object_entry **)_b;
1104 /* avoid filesystem trashing with loose objects */
1105 if (!a->in_pack && !b->in_pack)
1106 return hashcmp(a->idx.sha1, b->idx.sha1);
1108 if (a->in_pack < b->in_pack)
1109 return -1;
1110 if (a->in_pack > b->in_pack)
1111 return 1;
1112 return a->in_pack_offset < b->in_pack_offset ? -1 :
1113 (a->in_pack_offset > b->in_pack_offset);
1114 }
1116 static void get_object_details(void)
1117 {
1118 uint32_t i;
1119 struct object_entry **sorted_by_offset;
1121 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1122 for (i = 0; i < nr_objects; i++)
1123 sorted_by_offset[i] = objects + i;
1124 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1126 for (i = 0; i < nr_objects; i++)
1127 check_object(sorted_by_offset[i]);
1129 free(sorted_by_offset);
1130 }
1132 /*
1133 * We search for deltas in a list sorted by type, by filename hash, and then
1134 * by size, so that we see progressively smaller and smaller files.
1135 * That's because we prefer deltas to be from the bigger file
1136 * to the smaller -- deletes are potentially cheaper, but perhaps
1137 * more importantly, the bigger file is likely the more recent
1138 * one. The deepest deltas are therefore the oldest objects which are
1139 * less susceptible to be accessed often.
1140 */
1141 static int type_size_sort(const void *_a, const void *_b)
1142 {
1143 const struct object_entry *a = *(struct object_entry **)_a;
1144 const struct object_entry *b = *(struct object_entry **)_b;
1146 if (a->type > b->type)
1147 return -1;
1148 if (a->type < b->type)
1149 return 1;
1150 if (a->hash > b->hash)
1151 return -1;
1152 if (a->hash < b->hash)
1153 return 1;
1154 if (a->preferred_base > b->preferred_base)
1155 return -1;
1156 if (a->preferred_base < b->preferred_base)
1157 return 1;
1158 if (a->size > b->size)
1159 return -1;
1160 if (a->size < b->size)
1161 return 1;
1162 return a < b ? -1 : (a > b); /* newest first */
1163 }
1165 struct unpacked {
1166 struct object_entry *entry;
1167 void *data;
1168 struct delta_index *index;
1169 unsigned depth;
1170 };
1172 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1173 unsigned long delta_size)
1174 {
1175 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1176 return 0;
1178 if (delta_size < cache_max_small_delta_size)
1179 return 1;
1181 /* cache delta, if objects are large enough compared to delta size */
1182 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1183 return 1;
1185 return 0;
1186 }
1188 #ifdef THREADED_DELTA_SEARCH
1190 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1191 #define read_lock() pthread_mutex_lock(&read_mutex)
1192 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1194 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1195 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1196 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1198 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1199 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1200 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1202 #else
1204 #define read_lock() (void)0
1205 #define read_unlock() (void)0
1206 #define cache_lock() (void)0
1207 #define cache_unlock() (void)0
1208 #define progress_lock() (void)0
1209 #define progress_unlock() (void)0
1211 #endif
1213 static int try_delta(struct unpacked *trg, struct unpacked *src,
1214 unsigned max_depth, unsigned long *mem_usage)
1215 {
1216 struct object_entry *trg_entry = trg->entry;
1217 struct object_entry *src_entry = src->entry;
1218 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1219 unsigned ref_depth;
1220 enum object_type type;
1221 void *delta_buf;
1223 /* Don't bother doing diffs between different types */
1224 if (trg_entry->type != src_entry->type)
1225 return -1;
1227 /*
1228 * We do not bother to try a delta that we discarded
1229 * on an earlier try, but only when reusing delta data.
1230 */
1231 if (reuse_delta && trg_entry->in_pack &&
1232 trg_entry->in_pack == src_entry->in_pack &&
1233 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1234 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1235 return 0;
1237 /* Let's not bust the allowed depth. */
1238 if (src->depth >= max_depth)
1239 return 0;
1241 /* Now some size filtering heuristics. */
1242 trg_size = trg_entry->size;
1243 if (!trg_entry->delta) {
1244 max_size = trg_size/2 - 20;
1245 ref_depth = 1;
1246 } else {
1247 max_size = trg_entry->delta_size;
1248 ref_depth = trg->depth;
1249 }
1250 max_size = max_size * (max_depth - src->depth) /
1251 (max_depth - ref_depth + 1);
1252 if (max_size == 0)
1253 return 0;
1254 src_size = src_entry->size;
1255 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1256 if (sizediff >= max_size)
1257 return 0;
1258 if (trg_size < src_size / 32)
1259 return 0;
1261 /* Load data if not already done */
1262 if (!trg->data) {
1263 read_lock();
1264 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1265 read_unlock();
1266 if (!trg->data)
1267 die("object %s cannot be read",
1268 sha1_to_hex(trg_entry->idx.sha1));
1269 if (sz != trg_size)
1270 die("object %s inconsistent object length (%lu vs %lu)",
1271 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1272 *mem_usage += sz;
1273 }
1274 if (!src->data) {
1275 read_lock();
1276 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1277 read_unlock();
1278 if (!src->data)
1279 die("object %s cannot be read",
1280 sha1_to_hex(src_entry->idx.sha1));
1281 if (sz != src_size)
1282 die("object %s inconsistent object length (%lu vs %lu)",
1283 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1284 *mem_usage += sz;
1285 }
1286 if (!src->index) {
1287 src->index = create_delta_index(src->data, src_size);
1288 if (!src->index) {
1289 static int warned = 0;
1290 if (!warned++)
1291 warning("suboptimal pack - out of memory");
1292 return 0;
1293 }
1294 *mem_usage += sizeof_delta_index(src->index);
1295 }
1297 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1298 if (!delta_buf)
1299 return 0;
1301 if (trg_entry->delta) {
1302 /* Prefer only shallower same-sized deltas. */
1303 if (delta_size == trg_entry->delta_size &&
1304 src->depth + 1 >= trg->depth) {
1305 free(delta_buf);
1306 return 0;
1307 }
1308 }
1310 /*
1311 * Handle memory allocation outside of the cache
1312 * accounting lock. Compiler will optimize the strangeness
1313 * away when THREADED_DELTA_SEARCH is not defined.
1314 */
1315 free(trg_entry->delta_data);
1316 cache_lock();
1317 if (trg_entry->delta_data) {
1318 delta_cache_size -= trg_entry->delta_size;
1319 trg_entry->delta_data = NULL;
1320 }
1321 if (delta_cacheable(src_size, trg_size, delta_size)) {
1322 delta_cache_size += delta_size;
1323 cache_unlock();
1324 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1325 } else {
1326 cache_unlock();
1327 free(delta_buf);
1328 }
1330 trg_entry->delta = src_entry;
1331 trg_entry->delta_size = delta_size;
1332 trg->depth = src->depth + 1;
1334 return 1;
1335 }
1337 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1338 {
1339 struct object_entry *child = me->delta_child;
1340 unsigned int m = n;
1341 while (child) {
1342 unsigned int c = check_delta_limit(child, n + 1);
1343 if (m < c)
1344 m = c;
1345 child = child->delta_sibling;
1346 }
1347 return m;
1348 }
1350 static unsigned long free_unpacked(struct unpacked *n)
1351 {
1352 unsigned long freed_mem = sizeof_delta_index(n->index);
1353 free_delta_index(n->index);
1354 n->index = NULL;
1355 if (n->data) {
1356 freed_mem += n->entry->size;
1357 free(n->data);
1358 n->data = NULL;
1359 }
1360 n->entry = NULL;
1361 n->depth = 0;
1362 return freed_mem;
1363 }
1365 static void find_deltas(struct object_entry **list, unsigned *list_size,
1366 int window, int depth, unsigned *processed)
1367 {
1368 uint32_t i, idx = 0, count = 0;
1369 unsigned int array_size = window * sizeof(struct unpacked);
1370 struct unpacked *array;
1371 unsigned long mem_usage = 0;
1373 array = xmalloc(array_size);
1374 memset(array, 0, array_size);
1376 for (;;) {
1377 struct object_entry *entry = *list++;
1378 struct unpacked *n = array + idx;
1379 int j, max_depth, best_base = -1;
1381 progress_lock();
1382 if (!*list_size) {
1383 progress_unlock();
1384 break;
1385 }
1386 (*list_size)--;
1387 if (!entry->preferred_base) {
1388 (*processed)++;
1389 display_progress(progress_state, *processed);
1390 }
1391 progress_unlock();
1393 mem_usage -= free_unpacked(n);
1394 n->entry = entry;
1396 while (window_memory_limit &&
1397 mem_usage > window_memory_limit &&
1398 count > 1) {
1399 uint32_t tail = (idx + window - count) % window;
1400 mem_usage -= free_unpacked(array + tail);
1401 count--;
1402 }
1404 /* We do not compute delta to *create* objects we are not
1405 * going to pack.
1406 */
1407 if (entry->preferred_base)
1408 goto next;
1410 /*
1411 * If the current object is at pack edge, take the depth the
1412 * objects that depend on the current object into account
1413 * otherwise they would become too deep.
1414 */
1415 max_depth = depth;
1416 if (entry->delta_child) {
1417 max_depth -= check_delta_limit(entry, 0);
1418 if (max_depth <= 0)
1419 goto next;
1420 }
1422 j = window;
1423 while (--j > 0) {
1424 int ret;
1425 uint32_t other_idx = idx + j;
1426 struct unpacked *m;
1427 if (other_idx >= window)
1428 other_idx -= window;
1429 m = array + other_idx;
1430 if (!m->entry)
1431 break;
1432 ret = try_delta(n, m, max_depth, &mem_usage);
1433 if (ret < 0)
1434 break;
1435 else if (ret > 0)
1436 best_base = other_idx;
1437 }
1439 /*
1440 * If we decided to cache the delta data, then it is best
1441 * to compress it right away. First because we have to do
1442 * it anyway, and doing it here while we're threaded will
1443 * save a lot of time in the non threaded write phase,
1444 * as well as allow for caching more deltas within
1445 * the same cache size limit.
1446 * ...
1447 * But only if not writing to stdout, since in that case
1448 * the network is most likely throttling writes anyway,
1449 * and therefore it is best to go to the write phase ASAP
1450 * instead, as we can afford spending more time compressing
1451 * between writes at that moment.
1452 */
1453 if (entry->delta_data && !pack_to_stdout) {
1454 entry->z_delta_size = do_compress(&entry->delta_data,
1455 entry->delta_size);
1456 cache_lock();
1457 delta_cache_size -= entry->delta_size;
1458 delta_cache_size += entry->z_delta_size;
1459 cache_unlock();
1460 }
1462 /* if we made n a delta, and if n is already at max
1463 * depth, leaving it in the window is pointless. we
1464 * should evict it first.
1465 */
1466 if (entry->delta && max_depth <= n->depth)
1467 continue;
1469 /*
1470 * Move the best delta base up in the window, after the
1471 * currently deltified object, to keep it longer. It will
1472 * be the first base object to be attempted next.
1473 */
1474 if (entry->delta) {
1475 struct unpacked swap = array[best_base];
1476 int dist = (window + idx - best_base) % window;
1477 int dst = best_base;
1478 while (dist--) {
1479 int src = (dst + 1) % window;
1480 array[dst] = array[src];
1481 dst = src;
1482 }
1483 array[dst] = swap;
1484 }
1486 next:
1487 idx++;
1488 if (count + 1 < window)
1489 count++;
1490 if (idx >= window)
1491 idx = 0;
1492 }
1494 for (i = 0; i < window; ++i) {
1495 free_delta_index(array[i].index);
1496 free(array[i].data);
1497 }
1498 free(array);
1499 }
1501 #ifdef THREADED_DELTA_SEARCH
1503 /*
1504 * The main thread waits on the condition that (at least) one of the workers
1505 * has stopped working (which is indicated in the .working member of
1506 * struct thread_params).
1507 * When a work thread has completed its work, it sets .working to 0 and
1508 * signals the main thread and waits on the condition that .data_ready
1509 * becomes 1.
1510 */
1512 struct thread_params {
1513 pthread_t thread;
1514 struct object_entry **list;
1515 unsigned list_size;
1516 unsigned remaining;
1517 int window;
1518 int depth;
1519 int working;
1520 int data_ready;
1521 pthread_mutex_t mutex;
1522 pthread_cond_t cond;
1523 unsigned *processed;
1524 };
1526 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1528 static void *threaded_find_deltas(void *arg)
1529 {
1530 struct thread_params *me = arg;
1532 while (me->remaining) {
1533 find_deltas(me->list, &me->remaining,
1534 me->window, me->depth, me->processed);
1536 progress_lock();
1537 me->working = 0;
1538 pthread_cond_signal(&progress_cond);
1539 progress_unlock();
1541 /*
1542 * We must not set ->data_ready before we wait on the
1543 * condition because the main thread may have set it to 1
1544 * before we get here. In order to be sure that new
1545 * work is available if we see 1 in ->data_ready, it
1546 * was initialized to 0 before this thread was spawned
1547 * and we reset it to 0 right away.
1548 */
1549 pthread_mutex_lock(&me->mutex);
1550 while (!me->data_ready)
1551 pthread_cond_wait(&me->cond, &me->mutex);
1552 me->data_ready = 0;
1553 pthread_mutex_unlock(&me->mutex);
1554 }
1555 /* leave ->working 1 so that this doesn't get more work assigned */
1556 return NULL;
1557 }
1559 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1560 int window, int depth, unsigned *processed)
1561 {
1562 struct thread_params p[delta_search_threads];
1563 int i, ret, active_threads = 0;
1565 if (delta_search_threads <= 1) {
1566 find_deltas(list, &list_size, window, depth, processed);
1567 return;
1568 }
1570 /* Partition the work amongst work threads. */
1571 for (i = 0; i < delta_search_threads; i++) {
1572 unsigned sub_size = list_size / (delta_search_threads - i);
1574 p[i].window = window;
1575 p[i].depth = depth;
1576 p[i].processed = processed;
1577 p[i].working = 1;
1578 p[i].data_ready = 0;
1580 /* try to split chunks on "path" boundaries */
1581 while (sub_size && sub_size < list_size &&
1582 list[sub_size]->hash &&
1583 list[sub_size]->hash == list[sub_size-1]->hash)
1584 sub_size++;
1586 p[i].list = list;
1587 p[i].list_size = sub_size;
1588 p[i].remaining = sub_size;
1590 list += sub_size;
1591 list_size -= sub_size;
1592 }
1594 /* Start work threads. */
1595 for (i = 0; i < delta_search_threads; i++) {
1596 if (!p[i].list_size)
1597 continue;
1598 pthread_mutex_init(&p[i].mutex, NULL);
1599 pthread_cond_init(&p[i].cond, NULL);
1600 ret = pthread_create(&p[i].thread, NULL,
1601 threaded_find_deltas, &p[i]);
1602 if (ret)
1603 die("unable to create thread: %s", strerror(ret));
1604 active_threads++;
1605 }
1607 /*
1608 * Now let's wait for work completion. Each time a thread is done
1609 * with its work, we steal half of the remaining work from the
1610 * thread with the largest number of unprocessed objects and give
1611 * it to that newly idle thread. This ensure good load balancing
1612 * until the remaining object list segments are simply too short
1613 * to be worth splitting anymore.
1614 */
1615 while (active_threads) {
1616 struct thread_params *target = NULL;
1617 struct thread_params *victim = NULL;
1618 unsigned sub_size = 0;
1620 progress_lock();
1621 for (;;) {
1622 for (i = 0; !target && i < delta_search_threads; i++)
1623 if (!p[i].working)
1624 target = &p[i];
1625 if (target)
1626 break;
1627 pthread_cond_wait(&progress_cond, &progress_mutex);
1628 }
1630 for (i = 0; i < delta_search_threads; i++)
1631 if (p[i].remaining > 2*window &&
1632 (!victim || victim->remaining < p[i].remaining))
1633 victim = &p[i];
1634 if (victim) {
1635 sub_size = victim->remaining / 2;
1636 list = victim->list + victim->list_size - sub_size;
1637 while (sub_size && list[0]->hash &&
1638 list[0]->hash == list[-1]->hash) {
1639 list++;
1640 sub_size--;
1641 }
1642 if (!sub_size) {
1643 /*
1644 * It is possible for some "paths" to have
1645 * so many objects that no hash boundary
1646 * might be found. Let's just steal the
1647 * exact half in that case.
1648 */
1649 sub_size = victim->remaining / 2;
1650 list -= sub_size;
1651 }
1652 target->list = list;
1653 victim->list_size -= sub_size;
1654 victim->remaining -= sub_size;
1655 }
1656 target->list_size = sub_size;
1657 target->remaining = sub_size;
1658 target->working = 1;
1659 progress_unlock();
1661 pthread_mutex_lock(&target->mutex);
1662 target->data_ready = 1;
1663 pthread_cond_signal(&target->cond);
1664 pthread_mutex_unlock(&target->mutex);
1666 if (!sub_size) {
1667 pthread_join(target->thread, NULL);
1668 pthread_cond_destroy(&target->cond);
1669 pthread_mutex_destroy(&target->mutex);
1670 active_threads--;
1671 }
1672 }
1673 }
1675 #else
1676 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1677 #endif
1679 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1680 {
1681 unsigned char peeled[20];
1683 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1684 !peel_ref(path, peeled) && /* peelable? */
1685 !is_null_sha1(peeled) && /* annotated tag? */
1686 locate_object_entry(peeled)) /* object packed? */
1687 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1688 return 0;
1689 }
1691 static void prepare_pack(int window, int depth)
1692 {
1693 struct object_entry **delta_list;
1694 uint32_t i, nr_deltas;
1695 unsigned n;
1697 get_object_details();
1699 if (!nr_objects || !window || !depth)
1700 return;
1702 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1703 nr_deltas = n = 0;
1705 for (i = 0; i < nr_objects; i++) {
1706 struct object_entry *entry = objects + i;
1708 if (entry->delta)
1709 /* This happens if we decided to reuse existing
1710 * delta from a pack. "reuse_delta &&" is implied.
1711 */
1712 continue;
1714 if (entry->size < 50)
1715 continue;
1717 if (entry->no_try_delta)
1718 continue;
1720 if (!entry->preferred_base)
1721 nr_deltas++;
1723 delta_list[n++] = entry;
1724 }
1726 if (nr_deltas && n > 1) {
1727 unsigned nr_done = 0;
1728 if (progress)
1729 progress_state = start_progress("Compressing objects",
1730 nr_deltas);
1731 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1732 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1733 stop_progress(&progress_state);
1734 if (nr_done != nr_deltas)
1735 die("inconsistency with delta count");
1736 }
1737 free(delta_list);
1738 }
1740 static int git_pack_config(const char *k, const char *v, void *cb)
1741 {
1742 if(!strcmp(k, "pack.window")) {
1743 window = git_config_int(k, v);
1744 return 0;
1745 }
1746 if (!strcmp(k, "pack.windowmemory")) {
1747 window_memory_limit = git_config_ulong(k, v);
1748 return 0;
1749 }
1750 if (!strcmp(k, "pack.depth")) {
1751 depth = git_config_int(k, v);
1752 return 0;
1753 }
1754 if (!strcmp(k, "pack.compression")) {
1755 int level = git_config_int(k, v);
1756 if (level == -1)
1757 level = Z_DEFAULT_COMPRESSION;
1758 else if (level < 0 || level > Z_BEST_COMPRESSION)
1759 die("bad pack compression level %d", level);
1760 pack_compression_level = level;
1761 pack_compression_seen = 1;
1762 return 0;
1763 }
1764 if (!strcmp(k, "pack.deltacachesize")) {
1765 max_delta_cache_size = git_config_int(k, v);
1766 return 0;
1767 }
1768 if (!strcmp(k, "pack.deltacachelimit")) {
1769 cache_max_small_delta_size = git_config_int(k, v);
1770 return 0;
1771 }
1772 if (!strcmp(k, "pack.threads")) {
1773 delta_search_threads = git_config_int(k, v);
1774 if (delta_search_threads < 0)
1775 die("invalid number of threads specified (%d)",
1776 delta_search_threads);
1777 #ifndef THREADED_DELTA_SEARCH
1778 if (delta_search_threads != 1)
1779 warning("no threads support, ignoring %s", k);
1780 #endif
1781 return 0;
1782 }
1783 if (!strcmp(k, "pack.indexversion")) {
1784 pack_idx_default_version = git_config_int(k, v);
1785 if (pack_idx_default_version > 2)
1786 die("bad pack.indexversion=%"PRIu32,
1787 pack_idx_default_version);
1788 return 0;
1789 }
1790 if (!strcmp(k, "pack.packsizelimit")) {
1791 pack_size_limit_cfg = git_config_ulong(k, v);
1792 return 0;
1793 }
1794 return git_default_config(k, v, cb);
1795 }
1797 static void read_object_list_from_stdin(void)
1798 {
1799 char line[40 + 1 + PATH_MAX + 2];
1800 unsigned char sha1[20];
1802 for (;;) {
1803 if (!fgets(line, sizeof(line), stdin)) {
1804 if (feof(stdin))
1805 break;
1806 if (!ferror(stdin))
1807 die("fgets returned NULL, not EOF, not error!");
1808 if (errno != EINTR)
1809 die("fgets: %s", strerror(errno));
1810 clearerr(stdin);
1811 continue;
1812 }
1813 if (line[0] == '-') {
1814 if (get_sha1_hex(line+1, sha1))
1815 die("expected edge sha1, got garbage:\n %s",
1816 line);
1817 add_preferred_base(sha1);
1818 continue;
1819 }
1820 if (get_sha1_hex(line, sha1))
1821 die("expected sha1, got garbage:\n %s", line);
1823 add_preferred_base_object(line+41);
1824 add_object_entry(sha1, 0, line+41, 0);
1825 }
1826 }
1828 #define OBJECT_ADDED (1u<<20)
1830 static void show_commit(struct commit *commit)
1831 {
1832 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1833 commit->object.flags |= OBJECT_ADDED;
1834 }
1836 static void show_object(struct object_array_entry *p)
1837 {
1838 add_preferred_base_object(p->name);
1839 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1840 p->item->flags |= OBJECT_ADDED;
1841 }
1843 static void show_edge(struct commit *commit)
1844 {
1845 add_preferred_base(commit->object.sha1);
1846 }
1848 struct in_pack_object {
1849 off_t offset;
1850 struct object *object;
1851 };
1853 struct in_pack {
1854 int alloc;
1855 int nr;
1856 struct in_pack_object *array;
1857 };
1859 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1860 {
1861 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1862 in_pack->array[in_pack->nr].object = object;
1863 in_pack->nr++;
1864 }
1866 /*
1867 * Compare the objects in the offset order, in order to emulate the
1868 * "git-rev-list --objects" output that produced the pack originally.
1869 */
1870 static int ofscmp(const void *a_, const void *b_)
1871 {
1872 struct in_pack_object *a = (struct in_pack_object *)a_;
1873 struct in_pack_object *b = (struct in_pack_object *)b_;
1875 if (a->offset < b->offset)
1876 return -1;
1877 else if (a->offset > b->offset)
1878 return 1;
1879 else
1880 return hashcmp(a->object->sha1, b->object->sha1);
1881 }
1883 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1884 {
1885 struct packed_git *p;
1886 struct in_pack in_pack;
1887 uint32_t i;
1889 memset(&in_pack, 0, sizeof(in_pack));
1891 for (p = packed_git; p; p = p->next) {
1892 const unsigned char *sha1;
1893 struct object *o;
1895 for (i = 0; i < revs->num_ignore_packed; i++) {
1896 if (matches_pack_name(p, revs->ignore_packed[i]))
1897 break;
1898 }
1899 if (revs->num_ignore_packed <= i)
1900 continue;
1901 if (open_pack_index(p))
1902 die("cannot open pack index");
1904 ALLOC_GROW(in_pack.array,
1905 in_pack.nr + p->num_objects,
1906 in_pack.alloc);
1908 for (i = 0; i < p->num_objects; i++) {
1909 sha1 = nth_packed_object_sha1(p, i);
1910 o = lookup_unknown_object(sha1);
1911 if (!(o->flags & OBJECT_ADDED))
1912 mark_in_pack_object(o, p, &in_pack);
1913 o->flags |= OBJECT_ADDED;
1914 }
1915 }
1917 if (in_pack.nr) {
1918 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1919 ofscmp);
1920 for (i = 0; i < in_pack.nr; i++) {
1921 struct object *o = in_pack.array[i].object;
1922 add_object_entry(o->sha1, o->type, "", 0);
1923 }
1924 }
1925 free(in_pack.array);
1926 }
1928 static void loosen_unused_packed_objects(struct rev_info *revs)
1929 {
1930 struct packed_git *p;
1931 uint32_t i;
1932 const unsigned char *sha1;
1934 for (p = packed_git; p; p = p->next) {
1935 for (i = 0; i < revs->num_ignore_packed; i++) {
1936 if (matches_pack_name(p, revs->ignore_packed[i]))
1937 break;
1938 }
1939 if (revs->num_ignore_packed <= i)
1940 continue;
1942 if (open_pack_index(p))
1943 die("cannot open pack index");
1945 for (i = 0; i < p->num_objects; i++) {
1946 sha1 = nth_packed_object_sha1(p, i);
1947 if (!locate_object_entry(sha1))
1948 if (force_object_loose(sha1, p->mtime))
1949 die("unable to force loose object");
1950 }
1951 }
1952 }
1954 static void get_object_list(int ac, const char **av)
1955 {
1956 struct rev_info revs;
1957 char line[1000];
1958 int flags = 0;
1960 init_revisions(&revs, NULL);
1961 save_commit_buffer = 0;
1962 setup_revisions(ac, av, &revs, NULL);
1964 while (fgets(line, sizeof(line), stdin) != NULL) {
1965 int len = strlen(line);
1966 if (len && line[len - 1] == '\n')
1967 line[--len] = 0;
1968 if (!len)
1969 break;
1970 if (*line == '-') {
1971 if (!strcmp(line, "--not")) {
1972 flags ^= UNINTERESTING;
1973 continue;
1974 }
1975 die("not a rev '%s'", line);
1976 }
1977 if (handle_revision_arg(line, &revs, flags, 1))
1978 die("bad revision '%s'", line);
1979 }
1981 if (prepare_revision_walk(&revs))
1982 die("revision walk setup failed");
1983 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1984 traverse_commit_list(&revs, show_commit, show_object);
1986 if (keep_unreachable)
1987 add_objects_in_unpacked_packs(&revs);
1988 if (unpack_unreachable)
1989 loosen_unused_packed_objects(&revs);
1990 }
1992 static int adjust_perm(const char *path, mode_t mode)
1993 {
1994 if (chmod(path, mode))
1995 return -1;
1996 return adjust_shared_perm(path);
1997 }
1999 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2000 {
2001 int use_internal_rev_list = 0;
2002 int thin = 0;
2003 uint32_t i;
2004 const char **rp_av;
2005 int rp_ac_alloc = 64;
2006 int rp_ac;
2008 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2010 rp_av[0] = "pack-objects";
2011 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2012 rp_ac = 2;
2014 git_config(git_pack_config, NULL);
2015 if (!pack_compression_seen && core_compression_seen)
2016 pack_compression_level = core_compression_level;
2018 progress = isatty(2);
2019 for (i = 1; i < argc; i++) {
2020 const char *arg = argv[i];
2022 if (*arg != '-')
2023 break;
2025 if (!strcmp("--non-empty", arg)) {
2026 non_empty = 1;
2027 continue;
2028 }
2029 if (!strcmp("--local", arg)) {
2030 local = 1;
2031 continue;
2032 }
2033 if (!strcmp("--incremental", arg)) {
2034 incremental = 1;
2035 continue;
2036 }
2037 if (!prefixcmp(arg, "--compression=")) {
2038 char *end;
2039 int level = strtoul(arg+14, &end, 0);
2040 if (!arg[14] || *end)
2041 usage(pack_usage);
2042 if (level == -1)
2043 level = Z_DEFAULT_COMPRESSION;
2044 else if (level < 0 || level > Z_BEST_COMPRESSION)
2045 die("bad pack compression level %d", level);
2046 pack_compression_level = level;
2047 continue;
2048 }
2049 if (!prefixcmp(arg, "--max-pack-size=")) {
2050 char *end;
2051 pack_size_limit_cfg = 0;
2052 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2053 if (!arg[16] || *end)
2054 usage(pack_usage);
2055 continue;
2056 }
2057 if (!prefixcmp(arg, "--window=")) {
2058 char *end;
2059 window = strtoul(arg+9, &end, 0);
2060 if (!arg[9] || *end)
2061 usage(pack_usage);
2062 continue;
2063 }
2064 if (!prefixcmp(arg, "--window-memory=")) {
2065 if (!git_parse_ulong(arg+16, &window_memory_limit))
2066 usage(pack_usage);
2067 continue;
2068 }
2069 if (!prefixcmp(arg, "--threads=")) {
2070 char *end;
2071 delta_search_threads = strtoul(arg+10, &end, 0);
2072 if (!arg[10] || *end || delta_search_threads < 0)
2073 usage(pack_usage);
2074 #ifndef THREADED_DELTA_SEARCH
2075 if (delta_search_threads != 1)
2076 warning("no threads support, "
2077 "ignoring %s", arg);
2078 #endif
2079 continue;
2080 }
2081 if (!prefixcmp(arg, "--depth=")) {
2082 char *end;
2083 depth = strtoul(arg+8, &end, 0);
2084 if (!arg[8] || *end)
2085 usage(pack_usage);
2086 continue;
2087 }
2088 if (!strcmp("--progress", arg)) {
2089 progress = 1;
2090 continue;
2091 }
2092 if (!strcmp("--all-progress", arg)) {
2093 progress = 2;
2094 continue;
2095 }
2096 if (!strcmp("-q", arg)) {
2097 progress = 0;
2098 continue;
2099 }
2100 if (!strcmp("--no-reuse-delta", arg)) {
2101 reuse_delta = 0;
2102 continue;
2103 }
2104 if (!strcmp("--no-reuse-object", arg)) {
2105 reuse_object = reuse_delta = 0;
2106 continue;
2107 }
2108 if (!strcmp("--delta-base-offset", arg)) {
2109 allow_ofs_delta = 1;
2110 continue;
2111 }
2112 if (!strcmp("--stdout", arg)) {
2113 pack_to_stdout = 1;
2114 continue;
2115 }
2116 if (!strcmp("--revs", arg)) {
2117 use_internal_rev_list = 1;
2118 continue;
2119 }
2120 if (!strcmp("--keep-unreachable", arg)) {
2121 keep_unreachable = 1;
2122 continue;
2123 }
2124 if (!strcmp("--unpack-unreachable", arg)) {
2125 unpack_unreachable = 1;
2126 continue;
2127 }
2128 if (!strcmp("--include-tag", arg)) {
2129 include_tag = 1;
2130 continue;
2131 }
2132 if (!strcmp("--unpacked", arg) ||
2133 !prefixcmp(arg, "--unpacked=") ||
2134 !strcmp("--reflog", arg) ||
2135 !strcmp("--all", arg)) {
2136 use_internal_rev_list = 1;
2137 if (rp_ac >= rp_ac_alloc - 1) {
2138 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2139 rp_av = xrealloc(rp_av,
2140 rp_ac_alloc * sizeof(*rp_av));
2141 }
2142 rp_av[rp_ac++] = arg;
2143 continue;
2144 }
2145 if (!strcmp("--thin", arg)) {
2146 use_internal_rev_list = 1;
2147 thin = 1;
2148 rp_av[1] = "--objects-edge";
2149 continue;
2150 }
2151 if (!prefixcmp(arg, "--index-version=")) {
2152 char *c;
2153 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2154 if (pack_idx_default_version > 2)
2155 die("bad %s", arg);
2156 if (*c == ',')
2157 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2158 if (*c || pack_idx_off32_limit & 0x80000000)
2159 die("bad %s", arg);
2160 continue;
2161 }
2162 usage(pack_usage);
2163 }
2165 /* Traditionally "pack-objects [options] base extra" failed;
2166 * we would however want to take refs parameter that would
2167 * have been given to upstream rev-list ourselves, which means
2168 * we somehow want to say what the base name is. So the
2169 * syntax would be:
2170 *
2171 * pack-objects [options] base <refs...>
2172 *
2173 * in other words, we would treat the first non-option as the
2174 * base_name and send everything else to the internal revision
2175 * walker.
2176 */
2178 if (!pack_to_stdout)
2179 base_name = argv[i++];
2181 if (pack_to_stdout != !base_name)
2182 usage(pack_usage);
2184 if (!pack_to_stdout && !pack_size_limit)
2185 pack_size_limit = pack_size_limit_cfg;
2187 if (pack_to_stdout && pack_size_limit)
2188 die("--max-pack-size cannot be used to build a pack for transfer.");
2190 if (!pack_to_stdout && thin)
2191 die("--thin cannot be used to build an indexable pack.");
2193 if (keep_unreachable && unpack_unreachable)
2194 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2196 #ifdef THREADED_DELTA_SEARCH
2197 if (!delta_search_threads) /* --threads=0 means autodetect */
2198 delta_search_threads = online_cpus();
2199 #endif
2201 prepare_packed_git();
2203 if (progress)
2204 progress_state = start_progress("Counting objects", 0);
2205 if (!use_internal_rev_list)
2206 read_object_list_from_stdin();
2207 else {
2208 rp_av[rp_ac] = NULL;
2209 get_object_list(rp_ac, rp_av);
2210 }
2211 if (include_tag && nr_result)
2212 for_each_ref(add_ref_tag, NULL);
2213 stop_progress(&progress_state);
2215 if (non_empty && !nr_result)
2216 return 0;
2217 if (nr_result)
2218 prepare_pack(window, depth);
2219 write_pack_file();
2220 if (progress)
2221 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2222 " reused %"PRIu32" (delta %"PRIu32")\n",
2223 written, written_delta, reused, reused_delta);
2224 return 0;
2225 }