![[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] [--keep-unreachable] \n\
32 [<ref-list | <object-list]";
34 struct object_entry {
35 struct pack_idx_entry idx;
36 unsigned long size; /* uncompressed size */
37 struct packed_git *in_pack; /* already in pack */
38 off_t in_pack_offset;
39 struct object_entry *delta; /* delta base object */
40 struct object_entry *delta_child; /* deltified objects who bases me */
41 struct object_entry *delta_sibling; /* other deltified objects who
42 * uses the same base as me
43 */
44 void *delta_data; /* cached delta (uncompressed) */
45 unsigned long delta_size; /* delta data size (uncompressed) */
46 unsigned int hash; /* name hint hash */
47 enum object_type type;
48 enum object_type in_pack_type; /* could be delta */
49 unsigned char in_pack_header_size;
50 unsigned char preferred_base; /* we do not pack this, but is available
51 * to be used as the base object to delta
52 * objects against.
53 */
54 unsigned char no_try_delta;
55 };
57 /*
58 * Objects we are going to pack are collected in objects array (dynamically
59 * expanded). nr_objects & nr_alloc controls this array. They are stored
60 * in the order we see -- typically rev-list --objects order that gives us
61 * nice "minimum seek" order.
62 */
63 static struct object_entry *objects;
64 static struct pack_idx_entry **written_list;
65 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
67 static int non_empty;
68 static int reuse_delta = 1, reuse_object = 1;
69 static int keep_unreachable, include_tag;
70 static int local;
71 static int incremental;
72 static int allow_ofs_delta;
73 static const char *base_name;
74 static int progress = 1;
75 static int window = 10;
76 static uint32_t pack_size_limit, pack_size_limit_cfg;
77 static int depth = 50;
78 static int delta_search_threads = 1;
79 static int pack_to_stdout;
80 static int num_preferred_base;
81 static struct progress *progress_state;
82 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
83 static int pack_compression_seen;
85 static unsigned long delta_cache_size = 0;
86 static unsigned long max_delta_cache_size = 0;
87 static unsigned long cache_max_small_delta_size = 1000;
89 static unsigned long window_memory_limit = 0;
91 /*
92 * The object names in objects array are hashed with this hashtable,
93 * to help looking up the entry by object name.
94 * This hashtable is built after all the objects are seen.
95 */
96 static int *object_ix;
97 static int object_ix_hashsz;
99 /*
100 * stats
101 */
102 static uint32_t written, written_delta;
103 static uint32_t reused, reused_delta;
106 static void *get_delta(struct object_entry *entry)
107 {
108 unsigned long size, base_size, delta_size;
109 void *buf, *base_buf, *delta_buf;
110 enum object_type type;
112 buf = read_sha1_file(entry->idx.sha1, &type, &size);
113 if (!buf)
114 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
115 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
116 if (!base_buf)
117 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
118 delta_buf = diff_delta(base_buf, base_size,
119 buf, size, &delta_size, 0);
120 if (!delta_buf || delta_size != entry->delta_size)
121 die("delta size changed");
122 free(buf);
123 free(base_buf);
124 return delta_buf;
125 }
127 static unsigned long do_compress(void **pptr, unsigned long size)
128 {
129 z_stream stream;
130 void *in, *out;
131 unsigned long maxsize;
133 memset(&stream, 0, sizeof(stream));
134 deflateInit(&stream, pack_compression_level);
135 maxsize = deflateBound(&stream, size);
137 in = *pptr;
138 out = xmalloc(maxsize);
139 *pptr = out;
141 stream.next_in = in;
142 stream.avail_in = size;
143 stream.next_out = out;
144 stream.avail_out = maxsize;
145 while (deflate(&stream, Z_FINISH) == Z_OK)
146 ; /* nothing */
147 deflateEnd(&stream);
149 free(in);
150 return stream.total_out;
151 }
153 /*
154 * The per-object header is a pretty dense thing, which is
155 * - first byte: low four bits are "size", then three bits of "type",
156 * and the high bit is "size continues".
157 * - each byte afterwards: low seven bits are size continuation,
158 * with the high bit being "size continues"
159 */
160 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
161 {
162 int n = 1;
163 unsigned char c;
165 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
166 die("bad type %d", type);
168 c = (type << 4) | (size & 15);
169 size >>= 4;
170 while (size) {
171 *hdr++ = c | 0x80;
172 c = size & 0x7f;
173 size >>= 7;
174 n++;
175 }
176 *hdr = c;
177 return n;
178 }
180 /*
181 * we are going to reuse the existing object data as is. make
182 * sure it is not corrupt.
183 */
184 static int check_pack_inflate(struct packed_git *p,
185 struct pack_window **w_curs,
186 off_t offset,
187 off_t len,
188 unsigned long expect)
189 {
190 z_stream stream;
191 unsigned char fakebuf[4096], *in;
192 int st;
194 memset(&stream, 0, sizeof(stream));
195 inflateInit(&stream);
196 do {
197 in = use_pack(p, w_curs, offset, &stream.avail_in);
198 stream.next_in = in;
199 stream.next_out = fakebuf;
200 stream.avail_out = sizeof(fakebuf);
201 st = inflate(&stream, Z_FINISH);
202 offset += stream.next_in - in;
203 } while (st == Z_OK || st == Z_BUF_ERROR);
204 inflateEnd(&stream);
205 return (st == Z_STREAM_END &&
206 stream.total_out == expect &&
207 stream.total_in == len) ? 0 : -1;
208 }
210 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
211 off_t offset, off_t len, unsigned int nr)
212 {
213 const uint32_t *index_crc;
214 uint32_t data_crc = crc32(0, Z_NULL, 0);
216 do {
217 unsigned int avail;
218 void *data = use_pack(p, w_curs, offset, &avail);
219 if (avail > len)
220 avail = len;
221 data_crc = crc32(data_crc, data, avail);
222 offset += avail;
223 len -= avail;
224 } while (len);
226 index_crc = p->index_data;
227 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
229 return data_crc != ntohl(*index_crc);
230 }
232 static void copy_pack_data(struct sha1file *f,
233 struct packed_git *p,
234 struct pack_window **w_curs,
235 off_t offset,
236 off_t len)
237 {
238 unsigned char *in;
239 unsigned int avail;
241 while (len) {
242 in = use_pack(p, w_curs, offset, &avail);
243 if (avail > len)
244 avail = (unsigned int)len;
245 sha1write(f, in, avail);
246 offset += avail;
247 len -= avail;
248 }
249 }
251 static unsigned long write_object(struct sha1file *f,
252 struct object_entry *entry,
253 off_t write_offset)
254 {
255 unsigned long size, limit, datalen;
256 void *buf;
257 unsigned char header[10], dheader[10];
258 unsigned hdrlen;
259 enum object_type type;
260 int usable_delta, to_reuse;
262 if (!pack_to_stdout)
263 crc32_begin(f);
265 type = entry->type;
267 /* write limit if limited packsize and not first object */
268 limit = pack_size_limit && nr_written ?
269 pack_size_limit - write_offset : 0;
271 if (!entry->delta)
272 usable_delta = 0; /* no delta */
273 else if (!pack_size_limit)
274 usable_delta = 1; /* unlimited packfile */
275 else if (entry->delta->idx.offset == (off_t)-1)
276 usable_delta = 0; /* base was written to another pack */
277 else if (entry->delta->idx.offset)
278 usable_delta = 1; /* base already exists in this pack */
279 else
280 usable_delta = 0; /* base could end up in another pack */
282 if (!reuse_object)
283 to_reuse = 0; /* explicit */
284 else if (!entry->in_pack)
285 to_reuse = 0; /* can't reuse what we don't have */
286 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
287 /* check_object() decided it for us ... */
288 to_reuse = usable_delta;
289 /* ... but pack split may override that */
290 else if (type != entry->in_pack_type)
291 to_reuse = 0; /* pack has delta which is unusable */
292 else if (entry->delta)
293 to_reuse = 0; /* we want to pack afresh */
294 else
295 to_reuse = 1; /* we have it in-pack undeltified,
296 * and we do not need to deltify it.
297 */
299 if (!to_reuse) {
300 if (!usable_delta) {
301 buf = read_sha1_file(entry->idx.sha1, &type, &size);
302 if (!buf)
303 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
304 } else if (entry->delta_data) {
305 size = entry->delta_size;
306 buf = entry->delta_data;
307 entry->delta_data = NULL;
308 type = (allow_ofs_delta && entry->delta->idx.offset) ?
309 OBJ_OFS_DELTA : OBJ_REF_DELTA;
310 } else {
311 buf = get_delta(entry);
312 size = entry->delta_size;
313 type = (allow_ofs_delta && entry->delta->idx.offset) ?
314 OBJ_OFS_DELTA : OBJ_REF_DELTA;
315 }
316 datalen = do_compress(&buf, size);
318 /*
319 * The object header is a byte of 'type' followed by zero or
320 * more bytes of length.
321 */
322 hdrlen = encode_header(type, size, header);
324 if (type == OBJ_OFS_DELTA) {
325 /*
326 * Deltas with relative base contain an additional
327 * encoding of the relative offset for the delta
328 * base from this object's position in the pack.
329 */
330 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
331 unsigned pos = sizeof(dheader) - 1;
332 dheader[pos] = ofs & 127;
333 while (ofs >>= 7)
334 dheader[--pos] = 128 | (--ofs & 127);
335 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
336 free(buf);
337 return 0;
338 }
339 sha1write(f, header, hdrlen);
340 sha1write(f, dheader + pos, sizeof(dheader) - pos);
341 hdrlen += sizeof(dheader) - pos;
342 } else if (type == OBJ_REF_DELTA) {
343 /*
344 * Deltas with a base reference contain
345 * an additional 20 bytes for the base sha1.
346 */
347 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
348 free(buf);
349 return 0;
350 }
351 sha1write(f, header, hdrlen);
352 sha1write(f, entry->delta->idx.sha1, 20);
353 hdrlen += 20;
354 } else {
355 if (limit && hdrlen + datalen + 20 >= limit) {
356 free(buf);
357 return 0;
358 }
359 sha1write(f, header, hdrlen);
360 }
361 sha1write(f, buf, datalen);
362 free(buf);
363 }
364 else {
365 struct packed_git *p = entry->in_pack;
366 struct pack_window *w_curs = NULL;
367 struct revindex_entry *revidx;
368 off_t offset;
370 if (entry->delta) {
371 type = (allow_ofs_delta && entry->delta->idx.offset) ?
372 OBJ_OFS_DELTA : OBJ_REF_DELTA;
373 reused_delta++;
374 }
375 hdrlen = encode_header(type, entry->size, header);
376 offset = entry->in_pack_offset;
377 revidx = find_pack_revindex(p, offset);
378 datalen = revidx[1].offset - offset;
379 if (!pack_to_stdout && p->index_version > 1 &&
380 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
381 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
382 offset += entry->in_pack_header_size;
383 datalen -= entry->in_pack_header_size;
384 if (type == OBJ_OFS_DELTA) {
385 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
386 unsigned pos = sizeof(dheader) - 1;
387 dheader[pos] = ofs & 127;
388 while (ofs >>= 7)
389 dheader[--pos] = 128 | (--ofs & 127);
390 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
391 return 0;
392 sha1write(f, header, hdrlen);
393 sha1write(f, dheader + pos, sizeof(dheader) - pos);
394 hdrlen += sizeof(dheader) - pos;
395 } else if (type == OBJ_REF_DELTA) {
396 if (limit && hdrlen + 20 + datalen + 20 >= limit)
397 return 0;
398 sha1write(f, header, hdrlen);
399 sha1write(f, entry->delta->idx.sha1, 20);
400 hdrlen += 20;
401 } else {
402 if (limit && hdrlen + datalen + 20 >= limit)
403 return 0;
404 sha1write(f, header, hdrlen);
405 }
407 if (!pack_to_stdout && p->index_version == 1 &&
408 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
409 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
410 copy_pack_data(f, p, &w_curs, offset, datalen);
411 unuse_pack(&w_curs);
412 reused++;
413 }
414 if (usable_delta)
415 written_delta++;
416 written++;
417 if (!pack_to_stdout)
418 entry->idx.crc32 = crc32_end(f);
419 return hdrlen + datalen;
420 }
422 static off_t write_one(struct sha1file *f,
423 struct object_entry *e,
424 off_t offset)
425 {
426 unsigned long size;
428 /* offset is non zero if object is written already. */
429 if (e->idx.offset || e->preferred_base)
430 return offset;
432 /* if we are deltified, write out base object first. */
433 if (e->delta) {
434 offset = write_one(f, e->delta, offset);
435 if (!offset)
436 return 0;
437 }
439 e->idx.offset = offset;
440 size = write_object(f, e, offset);
441 if (!size) {
442 e->idx.offset = 0;
443 return 0;
444 }
445 written_list[nr_written++] = &e->idx;
447 /* make sure off_t is sufficiently large not to wrap */
448 if (offset > offset + size)
449 die("pack too large for current definition of off_t");
450 return offset + size;
451 }
453 /* forward declaration for write_pack_file */
454 static int adjust_perm(const char *path, mode_t mode);
456 static void write_pack_file(void)
457 {
458 uint32_t i = 0, j;
459 struct sha1file *f;
460 off_t offset, offset_one, last_obj_offset = 0;
461 struct pack_header hdr;
462 uint32_t nr_remaining = nr_result;
463 time_t last_mtime = 0;
465 if (progress > pack_to_stdout)
466 progress_state = start_progress("Writing objects", nr_result);
467 written_list = xmalloc(nr_objects * sizeof(*written_list));
469 do {
470 unsigned char sha1[20];
471 char *pack_tmp_name = NULL;
473 if (pack_to_stdout) {
474 f = sha1fd_throughput(1, "<stdout>", progress_state);
475 } else {
476 char tmpname[PATH_MAX];
477 int fd;
478 snprintf(tmpname, sizeof(tmpname),
479 "%s/tmp_pack_XXXXXX", get_object_directory());
480 fd = xmkstemp(tmpname);
481 pack_tmp_name = xstrdup(tmpname);
482 f = sha1fd(fd, pack_tmp_name);
483 }
485 hdr.hdr_signature = htonl(PACK_SIGNATURE);
486 hdr.hdr_version = htonl(PACK_VERSION);
487 hdr.hdr_entries = htonl(nr_remaining);
488 sha1write(f, &hdr, sizeof(hdr));
489 offset = sizeof(hdr);
490 nr_written = 0;
491 for (; i < nr_objects; i++) {
492 last_obj_offset = offset;
493 offset_one = write_one(f, objects + i, offset);
494 if (!offset_one)
495 break;
496 offset = offset_one;
497 display_progress(progress_state, written);
498 }
500 /*
501 * Did we write the wrong # entries in the header?
502 * If so, rewrite it like in fast-import
503 */
504 if (pack_to_stdout || nr_written == nr_remaining) {
505 sha1close(f, sha1, 1);
506 } else {
507 int fd = sha1close(f, NULL, 0);
508 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
509 close(fd);
510 }
512 if (!pack_to_stdout) {
513 mode_t mode = umask(0);
514 struct stat st;
515 char *idx_tmp_name, tmpname[PATH_MAX];
517 umask(mode);
518 mode = 0444 & ~mode;
520 idx_tmp_name = write_idx_file(NULL, written_list,
521 nr_written, sha1);
523 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
524 base_name, sha1_to_hex(sha1));
525 if (adjust_perm(pack_tmp_name, mode))
526 die("unable to make temporary pack file readable: %s",
527 strerror(errno));
528 if (rename(pack_tmp_name, tmpname))
529 die("unable to rename temporary pack file: %s",
530 strerror(errno));
532 /*
533 * Packs are runtime accessed in their mtime
534 * order since newer packs are more likely to contain
535 * younger objects. So if we are creating multiple
536 * packs then we should modify the mtime of later ones
537 * to preserve this property.
538 */
539 if (stat(tmpname, &st) < 0) {
540 warning("failed to stat %s: %s",
541 tmpname, strerror(errno));
542 } else if (!last_mtime) {
543 last_mtime = st.st_mtime;
544 } else {
545 struct utimbuf utb;
546 utb.actime = st.st_atime;
547 utb.modtime = --last_mtime;
548 if (utime(tmpname, &utb) < 0)
549 warning("failed utime() on %s: %s",
550 tmpname, strerror(errno));
551 }
553 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
554 base_name, sha1_to_hex(sha1));
555 if (adjust_perm(idx_tmp_name, mode))
556 die("unable to make temporary index file readable: %s",
557 strerror(errno));
558 if (rename(idx_tmp_name, tmpname))
559 die("unable to rename temporary index file: %s",
560 strerror(errno));
562 free(idx_tmp_name);
563 free(pack_tmp_name);
564 puts(sha1_to_hex(sha1));
565 }
567 /* mark written objects as written to previous pack */
568 for (j = 0; j < nr_written; j++) {
569 written_list[j]->offset = (off_t)-1;
570 }
571 nr_remaining -= nr_written;
572 } while (nr_remaining && i < nr_objects);
574 free(written_list);
575 stop_progress(&progress_state);
576 if (written != nr_result)
577 die("wrote %u objects while expecting %u", written, nr_result);
578 /*
579 * We have scanned through [0 ... i). Since we have written
580 * the correct number of objects, the remaining [i ... nr_objects)
581 * items must be either already written (due to out-of-order delta base)
582 * or a preferred base. Count those which are neither and complain if any.
583 */
584 for (j = 0; i < nr_objects; i++) {
585 struct object_entry *e = objects + i;
586 j += !e->idx.offset && !e->preferred_base;
587 }
588 if (j)
589 die("wrote %u objects as expected but %u unwritten", written, j);
590 }
592 static int locate_object_entry_hash(const unsigned char *sha1)
593 {
594 int i;
595 unsigned int ui;
596 memcpy(&ui, sha1, sizeof(unsigned int));
597 i = ui % object_ix_hashsz;
598 while (0 < object_ix[i]) {
599 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
600 return i;
601 if (++i == object_ix_hashsz)
602 i = 0;
603 }
604 return -1 - i;
605 }
607 static struct object_entry *locate_object_entry(const unsigned char *sha1)
608 {
609 int i;
611 if (!object_ix_hashsz)
612 return NULL;
614 i = locate_object_entry_hash(sha1);
615 if (0 <= i)
616 return &objects[object_ix[i]-1];
617 return NULL;
618 }
620 static void rehash_objects(void)
621 {
622 uint32_t i;
623 struct object_entry *oe;
625 object_ix_hashsz = nr_objects * 3;
626 if (object_ix_hashsz < 1024)
627 object_ix_hashsz = 1024;
628 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
629 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
630 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
631 int ix = locate_object_entry_hash(oe->idx.sha1);
632 if (0 <= ix)
633 continue;
634 ix = -1 - ix;
635 object_ix[ix] = i + 1;
636 }
637 }
639 static unsigned name_hash(const char *name)
640 {
641 unsigned char c;
642 unsigned hash = 0;
644 if (!name)
645 return 0;
647 /*
648 * This effectively just creates a sortable number from the
649 * last sixteen non-whitespace characters. Last characters
650 * count "most", so things that end in ".c" sort together.
651 */
652 while ((c = *name++) != 0) {
653 if (isspace(c))
654 continue;
655 hash = (hash >> 2) + (c << 24);
656 }
657 return hash;
658 }
660 static void setup_delta_attr_check(struct git_attr_check *check)
661 {
662 static struct git_attr *attr_delta;
664 if (!attr_delta)
665 attr_delta = git_attr("delta", 5);
667 check[0].attr = attr_delta;
668 }
670 static int no_try_delta(const char *path)
671 {
672 struct git_attr_check check[1];
674 setup_delta_attr_check(check);
675 if (git_checkattr(path, ARRAY_SIZE(check), check))
676 return 0;
677 if (ATTR_FALSE(check->value))
678 return 1;
679 return 0;
680 }
682 static int add_object_entry(const unsigned char *sha1, enum object_type type,
683 const char *name, int exclude)
684 {
685 struct object_entry *entry;
686 struct packed_git *p, *found_pack = NULL;
687 off_t found_offset = 0;
688 int ix;
689 unsigned hash = name_hash(name);
691 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
692 if (ix >= 0) {
693 if (exclude) {
694 entry = objects + object_ix[ix] - 1;
695 if (!entry->preferred_base)
696 nr_result--;
697 entry->preferred_base = 1;
698 }
699 return 0;
700 }
702 for (p = packed_git; p; p = p->next) {
703 off_t offset = find_pack_entry_one(sha1, p);
704 if (offset) {
705 if (!found_pack) {
706 found_offset = offset;
707 found_pack = p;
708 }
709 if (exclude)
710 break;
711 if (incremental)
712 return 0;
713 if (local && !p->pack_local)
714 return 0;
715 }
716 }
718 if (nr_objects >= nr_alloc) {
719 nr_alloc = (nr_alloc + 1024) * 3 / 2;
720 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
721 }
723 entry = objects + nr_objects++;
724 memset(entry, 0, sizeof(*entry));
725 hashcpy(entry->idx.sha1, sha1);
726 entry->hash = hash;
727 if (type)
728 entry->type = type;
729 if (exclude)
730 entry->preferred_base = 1;
731 else
732 nr_result++;
733 if (found_pack) {
734 entry->in_pack = found_pack;
735 entry->in_pack_offset = found_offset;
736 }
738 if (object_ix_hashsz * 3 <= nr_objects * 4)
739 rehash_objects();
740 else
741 object_ix[-1 - ix] = nr_objects;
743 display_progress(progress_state, nr_objects);
745 if (name && no_try_delta(name))
746 entry->no_try_delta = 1;
748 return 1;
749 }
751 struct pbase_tree_cache {
752 unsigned char sha1[20];
753 int ref;
754 int temporary;
755 void *tree_data;
756 unsigned long tree_size;
757 };
759 static struct pbase_tree_cache *(pbase_tree_cache[256]);
760 static int pbase_tree_cache_ix(const unsigned char *sha1)
761 {
762 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
763 }
764 static int pbase_tree_cache_ix_incr(int ix)
765 {
766 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
767 }
769 static struct pbase_tree {
770 struct pbase_tree *next;
771 /* This is a phony "cache" entry; we are not
772 * going to evict it nor find it through _get()
773 * mechanism -- this is for the toplevel node that
774 * would almost always change with any commit.
775 */
776 struct pbase_tree_cache pcache;
777 } *pbase_tree;
779 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
780 {
781 struct pbase_tree_cache *ent, *nent;
782 void *data;
783 unsigned long size;
784 enum object_type type;
785 int neigh;
786 int my_ix = pbase_tree_cache_ix(sha1);
787 int available_ix = -1;
789 /* pbase-tree-cache acts as a limited hashtable.
790 * your object will be found at your index or within a few
791 * slots after that slot if it is cached.
792 */
793 for (neigh = 0; neigh < 8; neigh++) {
794 ent = pbase_tree_cache[my_ix];
795 if (ent && !hashcmp(ent->sha1, sha1)) {
796 ent->ref++;
797 return ent;
798 }
799 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
800 ((0 <= available_ix) &&
801 (!ent && pbase_tree_cache[available_ix])))
802 available_ix = my_ix;
803 if (!ent)
804 break;
805 my_ix = pbase_tree_cache_ix_incr(my_ix);
806 }
808 /* Did not find one. Either we got a bogus request or
809 * we need to read and perhaps cache.
810 */
811 data = read_sha1_file(sha1, &type, &size);
812 if (!data)
813 return NULL;
814 if (type != OBJ_TREE) {
815 free(data);
816 return NULL;
817 }
819 /* We need to either cache or return a throwaway copy */
821 if (available_ix < 0)
822 ent = NULL;
823 else {
824 ent = pbase_tree_cache[available_ix];
825 my_ix = available_ix;
826 }
828 if (!ent) {
829 nent = xmalloc(sizeof(*nent));
830 nent->temporary = (available_ix < 0);
831 }
832 else {
833 /* evict and reuse */
834 free(ent->tree_data);
835 nent = ent;
836 }
837 hashcpy(nent->sha1, sha1);
838 nent->tree_data = data;
839 nent->tree_size = size;
840 nent->ref = 1;
841 if (!nent->temporary)
842 pbase_tree_cache[my_ix] = nent;
843 return nent;
844 }
846 static void pbase_tree_put(struct pbase_tree_cache *cache)
847 {
848 if (!cache->temporary) {
849 cache->ref--;
850 return;
851 }
852 free(cache->tree_data);
853 free(cache);
854 }
856 static int name_cmp_len(const char *name)
857 {
858 int i;
859 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
860 ;
861 return i;
862 }
864 static void add_pbase_object(struct tree_desc *tree,
865 const char *name,
866 int cmplen,
867 const char *fullname)
868 {
869 struct name_entry entry;
870 int cmp;
872 while (tree_entry(tree,&entry)) {
873 if (S_ISGITLINK(entry.mode))
874 continue;
875 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
876 memcmp(name, entry.path, cmplen);
877 if (cmp > 0)
878 continue;
879 if (cmp < 0)
880 return;
881 if (name[cmplen] != '/') {
882 add_object_entry(entry.sha1,
883 object_type(entry.mode),
884 fullname, 1);
885 return;
886 }
887 if (S_ISDIR(entry.mode)) {
888 struct tree_desc sub;
889 struct pbase_tree_cache *tree;
890 const char *down = name+cmplen+1;
891 int downlen = name_cmp_len(down);
893 tree = pbase_tree_get(entry.sha1);
894 if (!tree)
895 return;
896 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
898 add_pbase_object(&sub, down, downlen, fullname);
899 pbase_tree_put(tree);
900 }
901 }
902 }
904 static unsigned *done_pbase_paths;
905 static int done_pbase_paths_num;
906 static int done_pbase_paths_alloc;
907 static int done_pbase_path_pos(unsigned hash)
908 {
909 int lo = 0;
910 int hi = done_pbase_paths_num;
911 while (lo < hi) {
912 int mi = (hi + lo) / 2;
913 if (done_pbase_paths[mi] == hash)
914 return mi;
915 if (done_pbase_paths[mi] < hash)
916 hi = mi;
917 else
918 lo = mi + 1;
919 }
920 return -lo-1;
921 }
923 static int check_pbase_path(unsigned hash)
924 {
925 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
926 if (0 <= pos)
927 return 1;
928 pos = -pos - 1;
929 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
930 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
931 done_pbase_paths = xrealloc(done_pbase_paths,
932 done_pbase_paths_alloc *
933 sizeof(unsigned));
934 }
935 done_pbase_paths_num++;
936 if (pos < done_pbase_paths_num)
937 memmove(done_pbase_paths + pos + 1,
938 done_pbase_paths + pos,
939 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
940 done_pbase_paths[pos] = hash;
941 return 0;
942 }
944 static void add_preferred_base_object(const char *name)
945 {
946 struct pbase_tree *it;
947 int cmplen;
948 unsigned hash = name_hash(name);
950 if (!num_preferred_base || check_pbase_path(hash))
951 return;
953 cmplen = name_cmp_len(name);
954 for (it = pbase_tree; it; it = it->next) {
955 if (cmplen == 0) {
956 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
957 }
958 else {
959 struct tree_desc tree;
960 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
961 add_pbase_object(&tree, name, cmplen, name);
962 }
963 }
964 }
966 static void add_preferred_base(unsigned char *sha1)
967 {
968 struct pbase_tree *it;
969 void *data;
970 unsigned long size;
971 unsigned char tree_sha1[20];
973 if (window <= num_preferred_base++)
974 return;
976 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
977 if (!data)
978 return;
980 for (it = pbase_tree; it; it = it->next) {
981 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
982 free(data);
983 return;
984 }
985 }
987 it = xcalloc(1, sizeof(*it));
988 it->next = pbase_tree;
989 pbase_tree = it;
991 hashcpy(it->pcache.sha1, tree_sha1);
992 it->pcache.tree_data = data;
993 it->pcache.tree_size = size;
994 }
996 static void check_object(struct object_entry *entry)
997 {
998 if (entry->in_pack) {
999 struct packed_git *p = entry->in_pack;
1000 struct pack_window *w_curs = NULL;
1001 const unsigned char *base_ref = NULL;
1002 struct object_entry *base_entry;
1003 unsigned long used, used_0;
1004 unsigned int avail;
1005 off_t ofs;
1006 unsigned char *buf, c;
1008 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1010 /*
1011 * We want in_pack_type even if we do not reuse delta
1012 * since non-delta representations could still be reused.
1013 */
1014 used = unpack_object_header_gently(buf, avail,
1015 &entry->in_pack_type,
1016 &entry->size);
1018 /*
1019 * Determine if this is a delta and if so whether we can
1020 * reuse it or not. Otherwise let's find out as cheaply as
1021 * possible what the actual type and size for this object is.
1022 */
1023 switch (entry->in_pack_type) {
1024 default:
1025 /* Not a delta hence we've already got all we need. */
1026 entry->type = entry->in_pack_type;
1027 entry->in_pack_header_size = used;
1028 unuse_pack(&w_curs);
1029 return;
1030 case OBJ_REF_DELTA:
1031 if (reuse_delta && !entry->preferred_base)
1032 base_ref = use_pack(p, &w_curs,
1033 entry->in_pack_offset + used, NULL);
1034 entry->in_pack_header_size = used + 20;
1035 break;
1036 case OBJ_OFS_DELTA:
1037 buf = use_pack(p, &w_curs,
1038 entry->in_pack_offset + used, NULL);
1039 used_0 = 0;
1040 c = buf[used_0++];
1041 ofs = c & 127;
1042 while (c & 128) {
1043 ofs += 1;
1044 if (!ofs || MSB(ofs, 7))
1045 die("delta base offset overflow in pack for %s",
1046 sha1_to_hex(entry->idx.sha1));
1047 c = buf[used_0++];
1048 ofs = (ofs << 7) + (c & 127);
1049 }
1050 if (ofs >= entry->in_pack_offset)
1051 die("delta base offset out of bound for %s",
1052 sha1_to_hex(entry->idx.sha1));
1053 ofs = entry->in_pack_offset - ofs;
1054 if (reuse_delta && !entry->preferred_base) {
1055 struct revindex_entry *revidx;
1056 revidx = find_pack_revindex(p, ofs);
1057 base_ref = nth_packed_object_sha1(p, revidx->nr);
1058 }
1059 entry->in_pack_header_size = used + used_0;
1060 break;
1061 }
1063 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1064 /*
1065 * If base_ref was set above that means we wish to
1066 * reuse delta data, and we even found that base
1067 * in the list of objects we want to pack. Goodie!
1068 *
1069 * Depth value does not matter - find_deltas() will
1070 * never consider reused delta as the base object to
1071 * deltify other objects against, in order to avoid
1072 * circular deltas.
1073 */
1074 entry->type = entry->in_pack_type;
1075 entry->delta = base_entry;
1076 entry->delta_sibling = base_entry->delta_child;
1077 base_entry->delta_child = entry;
1078 unuse_pack(&w_curs);
1079 return;
1080 }
1082 if (entry->type) {
1083 /*
1084 * This must be a delta and we already know what the
1085 * final object type is. Let's extract the actual
1086 * object size from the delta header.
1087 */
1088 entry->size = get_size_from_delta(p, &w_curs,
1089 entry->in_pack_offset + entry->in_pack_header_size);
1090 unuse_pack(&w_curs);
1091 return;
1092 }
1094 /*
1095 * No choice but to fall back to the recursive delta walk
1096 * with sha1_object_info() to find about the object type
1097 * at this point...
1098 */
1099 unuse_pack(&w_curs);
1100 }
1102 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1103 if (entry->type < 0)
1104 die("unable to get type of object %s",
1105 sha1_to_hex(entry->idx.sha1));
1106 }
1108 static int pack_offset_sort(const void *_a, const void *_b)
1109 {
1110 const struct object_entry *a = *(struct object_entry **)_a;
1111 const struct object_entry *b = *(struct object_entry **)_b;
1113 /* avoid filesystem trashing with loose objects */
1114 if (!a->in_pack && !b->in_pack)
1115 return hashcmp(a->idx.sha1, b->idx.sha1);
1117 if (a->in_pack < b->in_pack)
1118 return -1;
1119 if (a->in_pack > b->in_pack)
1120 return 1;
1121 return a->in_pack_offset < b->in_pack_offset ? -1 :
1122 (a->in_pack_offset > b->in_pack_offset);
1123 }
1125 static void get_object_details(void)
1126 {
1127 uint32_t i;
1128 struct object_entry **sorted_by_offset;
1130 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1131 for (i = 0; i < nr_objects; i++)
1132 sorted_by_offset[i] = objects + i;
1133 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1135 init_pack_revindex();
1137 for (i = 0; i < nr_objects; i++)
1138 check_object(sorted_by_offset[i]);
1140 free(sorted_by_offset);
1141 }
1143 /*
1144 * We search for deltas in a list sorted by type, by filename hash, and then
1145 * by size, so that we see progressively smaller and smaller files.
1146 * That's because we prefer deltas to be from the bigger file
1147 * to the smaller -- deletes are potentially cheaper, but perhaps
1148 * more importantly, the bigger file is likely the more recent
1149 * one. The deepest deltas are therefore the oldest objects which are
1150 * less susceptible to be accessed often.
1151 */
1152 static int type_size_sort(const void *_a, const void *_b)
1153 {
1154 const struct object_entry *a = *(struct object_entry **)_a;
1155 const struct object_entry *b = *(struct object_entry **)_b;
1157 if (a->type > b->type)
1158 return -1;
1159 if (a->type < b->type)
1160 return 1;
1161 if (a->hash > b->hash)
1162 return -1;
1163 if (a->hash < b->hash)
1164 return 1;
1165 if (a->preferred_base > b->preferred_base)
1166 return -1;
1167 if (a->preferred_base < b->preferred_base)
1168 return 1;
1169 if (a->size > b->size)
1170 return -1;
1171 if (a->size < b->size)
1172 return 1;
1173 return a < b ? -1 : (a > b); /* newest first */
1174 }
1176 struct unpacked {
1177 struct object_entry *entry;
1178 void *data;
1179 struct delta_index *index;
1180 unsigned depth;
1181 };
1183 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1184 unsigned long delta_size)
1185 {
1186 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1187 return 0;
1189 if (delta_size < cache_max_small_delta_size)
1190 return 1;
1192 /* cache delta, if objects are large enough compared to delta size */
1193 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1194 return 1;
1196 return 0;
1197 }
1199 #ifdef THREADED_DELTA_SEARCH
1201 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1202 #define read_lock() pthread_mutex_lock(&read_mutex)
1203 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1205 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1206 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1207 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1209 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1210 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1211 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1213 #else
1215 #define read_lock() (void)0
1216 #define read_unlock() (void)0
1217 #define cache_lock() (void)0
1218 #define cache_unlock() (void)0
1219 #define progress_lock() (void)0
1220 #define progress_unlock() (void)0
1222 #endif
1224 static int try_delta(struct unpacked *trg, struct unpacked *src,
1225 unsigned max_depth, unsigned long *mem_usage)
1226 {
1227 struct object_entry *trg_entry = trg->entry;
1228 struct object_entry *src_entry = src->entry;
1229 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1230 unsigned ref_depth;
1231 enum object_type type;
1232 void *delta_buf;
1234 /* Don't bother doing diffs between different types */
1235 if (trg_entry->type != src_entry->type)
1236 return -1;
1238 /*
1239 * We do not bother to try a delta that we discarded
1240 * on an earlier try, but only when reusing delta data.
1241 */
1242 if (reuse_delta && trg_entry->in_pack &&
1243 trg_entry->in_pack == src_entry->in_pack &&
1244 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1245 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1246 return 0;
1248 /* Let's not bust the allowed depth. */
1249 if (src->depth >= max_depth)
1250 return 0;
1252 /* Now some size filtering heuristics. */
1253 trg_size = trg_entry->size;
1254 if (!trg_entry->delta) {
1255 max_size = trg_size/2 - 20;
1256 ref_depth = 1;
1257 } else {
1258 max_size = trg_entry->delta_size;
1259 ref_depth = trg->depth;
1260 }
1261 max_size = max_size * (max_depth - src->depth) /
1262 (max_depth - ref_depth + 1);
1263 if (max_size == 0)
1264 return 0;
1265 src_size = src_entry->size;
1266 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1267 if (sizediff >= max_size)
1268 return 0;
1269 if (trg_size < src_size / 32)
1270 return 0;
1272 /* Load data if not already done */
1273 if (!trg->data) {
1274 read_lock();
1275 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1276 read_unlock();
1277 if (!trg->data)
1278 die("object %s cannot be read",
1279 sha1_to_hex(trg_entry->idx.sha1));
1280 if (sz != trg_size)
1281 die("object %s inconsistent object length (%lu vs %lu)",
1282 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1283 *mem_usage += sz;
1284 }
1285 if (!src->data) {
1286 read_lock();
1287 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1288 read_unlock();
1289 if (!src->data)
1290 die("object %s cannot be read",
1291 sha1_to_hex(src_entry->idx.sha1));
1292 if (sz != src_size)
1293 die("object %s inconsistent object length (%lu vs %lu)",
1294 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1295 *mem_usage += sz;
1296 }
1297 if (!src->index) {
1298 src->index = create_delta_index(src->data, src_size);
1299 if (!src->index) {
1300 static int warned = 0;
1301 if (!warned++)
1302 warning("suboptimal pack - out of memory");
1303 return 0;
1304 }
1305 *mem_usage += sizeof_delta_index(src->index);
1306 }
1308 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1309 if (!delta_buf)
1310 return 0;
1312 if (trg_entry->delta) {
1313 /* Prefer only shallower same-sized deltas. */
1314 if (delta_size == trg_entry->delta_size &&
1315 src->depth + 1 >= trg->depth) {
1316 free(delta_buf);
1317 return 0;
1318 }
1319 }
1321 /*
1322 * Handle memory allocation outside of the cache
1323 * accounting lock. Compiler will optimize the strangeness
1324 * away when THREADED_DELTA_SEARCH is not defined.
1325 */
1326 free(trg_entry->delta_data);
1327 cache_lock();
1328 if (trg_entry->delta_data) {
1329 delta_cache_size -= trg_entry->delta_size;
1330 trg_entry->delta_data = NULL;
1331 }
1332 if (delta_cacheable(src_size, trg_size, delta_size)) {
1333 delta_cache_size += delta_size;
1334 cache_unlock();
1335 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1336 } else {
1337 cache_unlock();
1338 free(delta_buf);
1339 }
1341 trg_entry->delta = src_entry;
1342 trg_entry->delta_size = delta_size;
1343 trg->depth = src->depth + 1;
1345 return 1;
1346 }
1348 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1349 {
1350 struct object_entry *child = me->delta_child;
1351 unsigned int m = n;
1352 while (child) {
1353 unsigned int c = check_delta_limit(child, n + 1);
1354 if (m < c)
1355 m = c;
1356 child = child->delta_sibling;
1357 }
1358 return m;
1359 }
1361 static unsigned long free_unpacked(struct unpacked *n)
1362 {
1363 unsigned long freed_mem = sizeof_delta_index(n->index);
1364 free_delta_index(n->index);
1365 n->index = NULL;
1366 if (n->data) {
1367 freed_mem += n->entry->size;
1368 free(n->data);
1369 n->data = NULL;
1370 }
1371 n->entry = NULL;
1372 n->depth = 0;
1373 return freed_mem;
1374 }
1376 static void find_deltas(struct object_entry **list, unsigned *list_size,
1377 int window, int depth, unsigned *processed)
1378 {
1379 uint32_t i, idx = 0, count = 0;
1380 unsigned int array_size = window * sizeof(struct unpacked);
1381 struct unpacked *array;
1382 unsigned long mem_usage = 0;
1384 array = xmalloc(array_size);
1385 memset(array, 0, array_size);
1387 for (;;) {
1388 struct object_entry *entry = *list++;
1389 struct unpacked *n = array + idx;
1390 int j, max_depth, best_base = -1;
1392 progress_lock();
1393 if (!*list_size) {
1394 progress_unlock();
1395 break;
1396 }
1397 (*list_size)--;
1398 if (!entry->preferred_base) {
1399 (*processed)++;
1400 display_progress(progress_state, *processed);
1401 }
1402 progress_unlock();
1404 mem_usage -= free_unpacked(n);
1405 n->entry = entry;
1407 while (window_memory_limit &&
1408 mem_usage > window_memory_limit &&
1409 count > 1) {
1410 uint32_t tail = (idx + window - count) % window;
1411 mem_usage -= free_unpacked(array + tail);
1412 count--;
1413 }
1415 /* We do not compute delta to *create* objects we are not
1416 * going to pack.
1417 */
1418 if (entry->preferred_base)
1419 goto next;
1421 /*
1422 * If the current object is at pack edge, take the depth the
1423 * objects that depend on the current object into account
1424 * otherwise they would become too deep.
1425 */
1426 max_depth = depth;
1427 if (entry->delta_child) {
1428 max_depth -= check_delta_limit(entry, 0);
1429 if (max_depth <= 0)
1430 goto next;
1431 }
1433 j = window;
1434 while (--j > 0) {
1435 int ret;
1436 uint32_t other_idx = idx + j;
1437 struct unpacked *m;
1438 if (other_idx >= window)
1439 other_idx -= window;
1440 m = array + other_idx;
1441 if (!m->entry)
1442 break;
1443 ret = try_delta(n, m, max_depth, &mem_usage);
1444 if (ret < 0)
1445 break;
1446 else if (ret > 0)
1447 best_base = other_idx;
1448 }
1450 /* if we made n a delta, and if n is already at max
1451 * depth, leaving it in the window is pointless. we
1452 * should evict it first.
1453 */
1454 if (entry->delta && depth <= n->depth)
1455 continue;
1457 /*
1458 * Move the best delta base up in the window, after the
1459 * currently deltified object, to keep it longer. It will
1460 * be the first base object to be attempted next.
1461 */
1462 if (entry->delta) {
1463 struct unpacked swap = array[best_base];
1464 int dist = (window + idx - best_base) % window;
1465 int dst = best_base;
1466 while (dist--) {
1467 int src = (dst + 1) % window;
1468 array[dst] = array[src];
1469 dst = src;
1470 }
1471 array[dst] = swap;
1472 }
1474 next:
1475 idx++;
1476 if (count + 1 < window)
1477 count++;
1478 if (idx >= window)
1479 idx = 0;
1480 }
1482 for (i = 0; i < window; ++i) {
1483 free_delta_index(array[i].index);
1484 free(array[i].data);
1485 }
1486 free(array);
1487 }
1489 #ifdef THREADED_DELTA_SEARCH
1491 /*
1492 * The main thread waits on the condition that (at least) one of the workers
1493 * has stopped working (which is indicated in the .working member of
1494 * struct thread_params).
1495 * When a work thread has completed its work, it sets .working to 0 and
1496 * signals the main thread and waits on the condition that .data_ready
1497 * becomes 1.
1498 */
1500 struct thread_params {
1501 pthread_t thread;
1502 struct object_entry **list;
1503 unsigned list_size;
1504 unsigned remaining;
1505 int window;
1506 int depth;
1507 int working;
1508 int data_ready;
1509 pthread_mutex_t mutex;
1510 pthread_cond_t cond;
1511 unsigned *processed;
1512 };
1514 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1516 static void *threaded_find_deltas(void *arg)
1517 {
1518 struct thread_params *me = arg;
1520 while (me->remaining) {
1521 find_deltas(me->list, &me->remaining,
1522 me->window, me->depth, me->processed);
1524 progress_lock();
1525 me->working = 0;
1526 pthread_cond_signal(&progress_cond);
1527 progress_unlock();
1529 /*
1530 * We must not set ->data_ready before we wait on the
1531 * condition because the main thread may have set it to 1
1532 * before we get here. In order to be sure that new
1533 * work is available if we see 1 in ->data_ready, it
1534 * was initialized to 0 before this thread was spawned
1535 * and we reset it to 0 right away.
1536 */
1537 pthread_mutex_lock(&me->mutex);
1538 while (!me->data_ready)
1539 pthread_cond_wait(&me->cond, &me->mutex);
1540 me->data_ready = 0;
1541 pthread_mutex_unlock(&me->mutex);
1542 }
1543 /* leave ->working 1 so that this doesn't get more work assigned */
1544 return NULL;
1545 }
1547 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1548 int window, int depth, unsigned *processed)
1549 {
1550 struct thread_params p[delta_search_threads];
1551 int i, ret, active_threads = 0;
1553 if (delta_search_threads <= 1) {
1554 find_deltas(list, &list_size, window, depth, processed);
1555 return;
1556 }
1558 /* Partition the work amongst work threads. */
1559 for (i = 0; i < delta_search_threads; i++) {
1560 unsigned sub_size = list_size / (delta_search_threads - i);
1562 p[i].window = window;
1563 p[i].depth = depth;
1564 p[i].processed = processed;
1565 p[i].working = 1;
1566 p[i].data_ready = 0;
1568 /* try to split chunks on "path" boundaries */
1569 while (sub_size && sub_size < list_size &&
1570 list[sub_size]->hash &&
1571 list[sub_size]->hash == list[sub_size-1]->hash)
1572 sub_size++;
1574 p[i].list = list;
1575 p[i].list_size = sub_size;
1576 p[i].remaining = sub_size;
1578 list += sub_size;
1579 list_size -= sub_size;
1580 }
1582 /* Start work threads. */
1583 for (i = 0; i < delta_search_threads; i++) {
1584 if (!p[i].list_size)
1585 continue;
1586 pthread_mutex_init(&p[i].mutex, NULL);
1587 pthread_cond_init(&p[i].cond, NULL);
1588 ret = pthread_create(&p[i].thread, NULL,
1589 threaded_find_deltas, &p[i]);
1590 if (ret)
1591 die("unable to create thread: %s", strerror(ret));
1592 active_threads++;
1593 }
1595 /*
1596 * Now let's wait for work completion. Each time a thread is done
1597 * with its work, we steal half of the remaining work from the
1598 * thread with the largest number of unprocessed objects and give
1599 * it to that newly idle thread. This ensure good load balancing
1600 * until the remaining object list segments are simply too short
1601 * to be worth splitting anymore.
1602 */
1603 while (active_threads) {
1604 struct thread_params *target = NULL;
1605 struct thread_params *victim = NULL;
1606 unsigned sub_size = 0;
1608 progress_lock();
1609 for (;;) {
1610 for (i = 0; !target && i < delta_search_threads; i++)
1611 if (!p[i].working)
1612 target = &p[i];
1613 if (target)
1614 break;
1615 pthread_cond_wait(&progress_cond, &progress_mutex);
1616 }
1618 for (i = 0; i < delta_search_threads; i++)
1619 if (p[i].remaining > 2*window &&
1620 (!victim || victim->remaining < p[i].remaining))
1621 victim = &p[i];
1622 if (victim) {
1623 sub_size = victim->remaining / 2;
1624 list = victim->list + victim->list_size - sub_size;
1625 while (sub_size && list[0]->hash &&
1626 list[0]->hash == list[-1]->hash) {
1627 list++;
1628 sub_size--;
1629 }
1630 if (!sub_size) {
1631 /*
1632 * It is possible for some "paths" to have
1633 * so many objects that no hash boundary
1634 * might be found. Let's just steal the
1635 * exact half in that case.
1636 */
1637 sub_size = victim->remaining / 2;
1638 list -= sub_size;
1639 }
1640 target->list = list;
1641 victim->list_size -= sub_size;
1642 victim->remaining -= sub_size;
1643 }
1644 target->list_size = sub_size;
1645 target->remaining = sub_size;
1646 target->working = 1;
1647 progress_unlock();
1649 pthread_mutex_lock(&target->mutex);
1650 target->data_ready = 1;
1651 pthread_cond_signal(&target->cond);
1652 pthread_mutex_unlock(&target->mutex);
1654 if (!sub_size) {
1655 pthread_join(target->thread, NULL);
1656 pthread_cond_destroy(&target->cond);
1657 pthread_mutex_destroy(&target->mutex);
1658 active_threads--;
1659 }
1660 }
1661 }
1663 #else
1664 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1665 #endif
1667 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1668 {
1669 unsigned char peeled[20];
1671 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1672 !peel_ref(path, peeled) && /* peelable? */
1673 !is_null_sha1(peeled) && /* annotated tag? */
1674 locate_object_entry(peeled)) /* object packed? */
1675 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1676 return 0;
1677 }
1679 static void prepare_pack(int window, int depth)
1680 {
1681 struct object_entry **delta_list;
1682 uint32_t i, n, nr_deltas;
1684 get_object_details();
1686 if (!nr_objects || !window || !depth)
1687 return;
1689 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1690 nr_deltas = n = 0;
1692 for (i = 0; i < nr_objects; i++) {
1693 struct object_entry *entry = objects + i;
1695 if (entry->delta)
1696 /* This happens if we decided to reuse existing
1697 * delta from a pack. "reuse_delta &&" is implied.
1698 */
1699 continue;
1701 if (entry->size < 50)
1702 continue;
1704 if (entry->no_try_delta)
1705 continue;
1707 if (!entry->preferred_base)
1708 nr_deltas++;
1710 delta_list[n++] = entry;
1711 }
1713 if (nr_deltas && n > 1) {
1714 unsigned nr_done = 0;
1715 if (progress)
1716 progress_state = start_progress("Compressing objects",
1717 nr_deltas);
1718 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1719 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1720 stop_progress(&progress_state);
1721 if (nr_done != nr_deltas)
1722 die("inconsistency with delta count");
1723 }
1724 free(delta_list);
1725 }
1727 static int git_pack_config(const char *k, const char *v)
1728 {
1729 if(!strcmp(k, "pack.window")) {
1730 window = git_config_int(k, v);
1731 return 0;
1732 }
1733 if (!strcmp(k, "pack.windowmemory")) {
1734 window_memory_limit = git_config_ulong(k, v);
1735 return 0;
1736 }
1737 if (!strcmp(k, "pack.depth")) {
1738 depth = git_config_int(k, v);
1739 return 0;
1740 }
1741 if (!strcmp(k, "pack.compression")) {
1742 int level = git_config_int(k, v);
1743 if (level == -1)
1744 level = Z_DEFAULT_COMPRESSION;
1745 else if (level < 0 || level > Z_BEST_COMPRESSION)
1746 die("bad pack compression level %d", level);
1747 pack_compression_level = level;
1748 pack_compression_seen = 1;
1749 return 0;
1750 }
1751 if (!strcmp(k, "pack.deltacachesize")) {
1752 max_delta_cache_size = git_config_int(k, v);
1753 return 0;
1754 }
1755 if (!strcmp(k, "pack.deltacachelimit")) {
1756 cache_max_small_delta_size = git_config_int(k, v);
1757 return 0;
1758 }
1759 if (!strcmp(k, "pack.threads")) {
1760 delta_search_threads = git_config_int(k, v);
1761 if (delta_search_threads < 0)
1762 die("invalid number of threads specified (%d)",
1763 delta_search_threads);
1764 #ifndef THREADED_DELTA_SEARCH
1765 if (delta_search_threads != 1)
1766 warning("no threads support, ignoring %s", k);
1767 #endif
1768 return 0;
1769 }
1770 if (!strcmp(k, "pack.indexversion")) {
1771 pack_idx_default_version = git_config_int(k, v);
1772 if (pack_idx_default_version > 2)
1773 die("bad pack.indexversion=%d", pack_idx_default_version);
1774 return 0;
1775 }
1776 if (!strcmp(k, "pack.packsizelimit")) {
1777 pack_size_limit_cfg = git_config_ulong(k, v);
1778 return 0;
1779 }
1780 return git_default_config(k, v);
1781 }
1783 static void read_object_list_from_stdin(void)
1784 {
1785 char line[40 + 1 + PATH_MAX + 2];
1786 unsigned char sha1[20];
1788 for (;;) {
1789 if (!fgets(line, sizeof(line), stdin)) {
1790 if (feof(stdin))
1791 break;
1792 if (!ferror(stdin))
1793 die("fgets returned NULL, not EOF, not error!");
1794 if (errno != EINTR)
1795 die("fgets: %s", strerror(errno));
1796 clearerr(stdin);
1797 continue;
1798 }
1799 if (line[0] == '-') {
1800 if (get_sha1_hex(line+1, sha1))
1801 die("expected edge sha1, got garbage:\n %s",
1802 line);
1803 add_preferred_base(sha1);
1804 continue;
1805 }
1806 if (get_sha1_hex(line, sha1))
1807 die("expected sha1, got garbage:\n %s", line);
1809 add_preferred_base_object(line+41);
1810 add_object_entry(sha1, 0, line+41, 0);
1811 }
1812 }
1814 #define OBJECT_ADDED (1u<<20)
1816 static void show_commit(struct commit *commit)
1817 {
1818 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1819 commit->object.flags |= OBJECT_ADDED;
1820 }
1822 static void show_object(struct object_array_entry *p)
1823 {
1824 add_preferred_base_object(p->name);
1825 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1826 p->item->flags |= OBJECT_ADDED;
1827 }
1829 static void show_edge(struct commit *commit)
1830 {
1831 add_preferred_base(commit->object.sha1);
1832 }
1834 struct in_pack_object {
1835 off_t offset;
1836 struct object *object;
1837 };
1839 struct in_pack {
1840 int alloc;
1841 int nr;
1842 struct in_pack_object *array;
1843 };
1845 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1846 {
1847 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1848 in_pack->array[in_pack->nr].object = object;
1849 in_pack->nr++;
1850 }
1852 /*
1853 * Compare the objects in the offset order, in order to emulate the
1854 * "git-rev-list --objects" output that produced the pack originally.
1855 */
1856 static int ofscmp(const void *a_, const void *b_)
1857 {
1858 struct in_pack_object *a = (struct in_pack_object *)a_;
1859 struct in_pack_object *b = (struct in_pack_object *)b_;
1861 if (a->offset < b->offset)
1862 return -1;
1863 else if (a->offset > b->offset)
1864 return 1;
1865 else
1866 return hashcmp(a->object->sha1, b->object->sha1);
1867 }
1869 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1870 {
1871 struct packed_git *p;
1872 struct in_pack in_pack;
1873 uint32_t i;
1875 memset(&in_pack, 0, sizeof(in_pack));
1877 for (p = packed_git; p; p = p->next) {
1878 const unsigned char *sha1;
1879 struct object *o;
1881 for (i = 0; i < revs->num_ignore_packed; i++) {
1882 if (matches_pack_name(p, revs->ignore_packed[i]))
1883 break;
1884 }
1885 if (revs->num_ignore_packed <= i)
1886 continue;
1887 if (open_pack_index(p))
1888 die("cannot open pack index");
1890 ALLOC_GROW(in_pack.array,
1891 in_pack.nr + p->num_objects,
1892 in_pack.alloc);
1894 for (i = 0; i < p->num_objects; i++) {
1895 sha1 = nth_packed_object_sha1(p, i);
1896 o = lookup_unknown_object(sha1);
1897 if (!(o->flags & OBJECT_ADDED))
1898 mark_in_pack_object(o, p, &in_pack);
1899 o->flags |= OBJECT_ADDED;
1900 }
1901 }
1903 if (in_pack.nr) {
1904 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1905 ofscmp);
1906 for (i = 0; i < in_pack.nr; i++) {
1907 struct object *o = in_pack.array[i].object;
1908 add_object_entry(o->sha1, o->type, "", 0);
1909 }
1910 }
1911 free(in_pack.array);
1912 }
1914 static void get_object_list(int ac, const char **av)
1915 {
1916 struct rev_info revs;
1917 char line[1000];
1918 int flags = 0;
1920 init_revisions(&revs, NULL);
1921 save_commit_buffer = 0;
1922 setup_revisions(ac, av, &revs, NULL);
1924 while (fgets(line, sizeof(line), stdin) != NULL) {
1925 int len = strlen(line);
1926 if (len && line[len - 1] == '\n')
1927 line[--len] = 0;
1928 if (!len)
1929 break;
1930 if (*line == '-') {
1931 if (!strcmp(line, "--not")) {
1932 flags ^= UNINTERESTING;
1933 continue;
1934 }
1935 die("not a rev '%s'", line);
1936 }
1937 if (handle_revision_arg(line, &revs, flags, 1))
1938 die("bad revision '%s'", line);
1939 }
1941 if (prepare_revision_walk(&revs))
1942 die("revision walk setup failed");
1943 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1944 traverse_commit_list(&revs, show_commit, show_object);
1946 if (keep_unreachable)
1947 add_objects_in_unpacked_packs(&revs);
1948 }
1950 static int adjust_perm(const char *path, mode_t mode)
1951 {
1952 if (chmod(path, mode))
1953 return -1;
1954 return adjust_shared_perm(path);
1955 }
1957 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1958 {
1959 int use_internal_rev_list = 0;
1960 int thin = 0;
1961 uint32_t i;
1962 const char **rp_av;
1963 int rp_ac_alloc = 64;
1964 int rp_ac;
1966 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
1968 rp_av[0] = "pack-objects";
1969 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
1970 rp_ac = 2;
1972 git_config(git_pack_config);
1973 if (!pack_compression_seen && core_compression_seen)
1974 pack_compression_level = core_compression_level;
1976 progress = isatty(2);
1977 for (i = 1; i < argc; i++) {
1978 const char *arg = argv[i];
1980 if (*arg != '-')
1981 break;
1983 if (!strcmp("--non-empty", arg)) {
1984 non_empty = 1;
1985 continue;
1986 }
1987 if (!strcmp("--local", arg)) {
1988 local = 1;
1989 continue;
1990 }
1991 if (!strcmp("--incremental", arg)) {
1992 incremental = 1;
1993 continue;
1994 }
1995 if (!prefixcmp(arg, "--compression=")) {
1996 char *end;
1997 int level = strtoul(arg+14, &end, 0);
1998 if (!arg[14] || *end)
1999 usage(pack_usage);
2000 if (level == -1)
2001 level = Z_DEFAULT_COMPRESSION;
2002 else if (level < 0 || level > Z_BEST_COMPRESSION)
2003 die("bad pack compression level %d", level);
2004 pack_compression_level = level;
2005 continue;
2006 }
2007 if (!prefixcmp(arg, "--max-pack-size=")) {
2008 char *end;
2009 pack_size_limit_cfg = 0;
2010 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2011 if (!arg[16] || *end)
2012 usage(pack_usage);
2013 continue;
2014 }
2015 if (!prefixcmp(arg, "--window=")) {
2016 char *end;
2017 window = strtoul(arg+9, &end, 0);
2018 if (!arg[9] || *end)
2019 usage(pack_usage);
2020 continue;
2021 }
2022 if (!prefixcmp(arg, "--window-memory=")) {
2023 if (!git_parse_ulong(arg+16, &window_memory_limit))
2024 usage(pack_usage);
2025 continue;
2026 }
2027 if (!prefixcmp(arg, "--threads=")) {
2028 char *end;
2029 delta_search_threads = strtoul(arg+10, &end, 0);
2030 if (!arg[10] || *end || delta_search_threads < 0)
2031 usage(pack_usage);
2032 #ifndef THREADED_DELTA_SEARCH
2033 if (delta_search_threads != 1)
2034 warning("no threads support, "
2035 "ignoring %s", arg);
2036 #endif
2037 continue;
2038 }
2039 if (!prefixcmp(arg, "--depth=")) {
2040 char *end;
2041 depth = strtoul(arg+8, &end, 0);
2042 if (!arg[8] || *end)
2043 usage(pack_usage);
2044 continue;
2045 }
2046 if (!strcmp("--progress", arg)) {
2047 progress = 1;
2048 continue;
2049 }
2050 if (!strcmp("--all-progress", arg)) {
2051 progress = 2;
2052 continue;
2053 }
2054 if (!strcmp("-q", arg)) {
2055 progress = 0;
2056 continue;
2057 }
2058 if (!strcmp("--no-reuse-delta", arg)) {
2059 reuse_delta = 0;
2060 continue;
2061 }
2062 if (!strcmp("--no-reuse-object", arg)) {
2063 reuse_object = reuse_delta = 0;
2064 continue;
2065 }
2066 if (!strcmp("--delta-base-offset", arg)) {
2067 allow_ofs_delta = 1;
2068 continue;
2069 }
2070 if (!strcmp("--stdout", arg)) {
2071 pack_to_stdout = 1;
2072 continue;
2073 }
2074 if (!strcmp("--revs", arg)) {
2075 use_internal_rev_list = 1;
2076 continue;
2077 }
2078 if (!strcmp("--keep-unreachable", arg)) {
2079 keep_unreachable = 1;
2080 continue;
2081 }
2082 if (!strcmp("--include-tag", arg)) {
2083 include_tag = 1;
2084 continue;
2085 }
2086 if (!strcmp("--unpacked", arg) ||
2087 !prefixcmp(arg, "--unpacked=") ||
2088 !strcmp("--reflog", arg) ||
2089 !strcmp("--all", arg)) {
2090 use_internal_rev_list = 1;
2091 if (rp_ac >= rp_ac_alloc - 1) {
2092 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2093 rp_av = xrealloc(rp_av,
2094 rp_ac_alloc * sizeof(*rp_av));
2095 }
2096 rp_av[rp_ac++] = arg;
2097 continue;
2098 }
2099 if (!strcmp("--thin", arg)) {
2100 use_internal_rev_list = 1;
2101 thin = 1;
2102 rp_av[1] = "--objects-edge";
2103 continue;
2104 }
2105 if (!prefixcmp(arg, "--index-version=")) {
2106 char *c;
2107 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2108 if (pack_idx_default_version > 2)
2109 die("bad %s", arg);
2110 if (*c == ',')
2111 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2112 if (*c || pack_idx_off32_limit & 0x80000000)
2113 die("bad %s", arg);
2114 continue;
2115 }
2116 usage(pack_usage);
2117 }
2119 /* Traditionally "pack-objects [options] base extra" failed;
2120 * we would however want to take refs parameter that would
2121 * have been given to upstream rev-list ourselves, which means
2122 * we somehow want to say what the base name is. So the
2123 * syntax would be:
2124 *
2125 * pack-objects [options] base <refs...>
2126 *
2127 * in other words, we would treat the first non-option as the
2128 * base_name and send everything else to the internal revision
2129 * walker.
2130 */
2132 if (!pack_to_stdout)
2133 base_name = argv[i++];
2135 if (pack_to_stdout != !base_name)
2136 usage(pack_usage);
2138 if (!pack_to_stdout && !pack_size_limit)
2139 pack_size_limit = pack_size_limit_cfg;
2141 if (pack_to_stdout && pack_size_limit)
2142 die("--max-pack-size cannot be used to build a pack for transfer.");
2144 if (!pack_to_stdout && thin)
2145 die("--thin cannot be used to build an indexable pack.");
2147 #ifdef THREADED_DELTA_SEARCH
2148 if (!delta_search_threads) /* --threads=0 means autodetect */
2149 delta_search_threads = online_cpus();
2150 #endif
2152 prepare_packed_git();
2154 if (progress)
2155 progress_state = start_progress("Counting objects", 0);
2156 if (!use_internal_rev_list)
2157 read_object_list_from_stdin();
2158 else {
2159 rp_av[rp_ac] = NULL;
2160 get_object_list(rp_ac, rp_av);
2161 }
2162 if (include_tag && nr_result)
2163 for_each_ref(add_ref_tag, NULL);
2164 stop_progress(&progress_state);
2166 if (non_empty && !nr_result)
2167 return 0;
2168 if (nr_result)
2169 prepare_pack(window, depth);
2170 write_pack_file();
2171 if (progress)
2172 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2173 written, written_delta, reused, reused_delta);
2174 return 0;
2175 }