![[tokkee]](http://tokkee.org/images/avatar.png){kind=avatar}
827673ce4e81e14e13ae4b41ac89c18f6201ddb6
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 int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
213 off_t offset, off_t len, unsigned int nr)
214 {
215 const uint32_t *index_crc;
216 uint32_t data_crc = crc32(0, Z_NULL, 0);
218 do {
219 unsigned int avail;
220 void *data = use_pack(p, w_curs, offset, &avail);
221 if (avail > len)
222 avail = len;
223 data_crc = crc32(data_crc, data, avail);
224 offset += avail;
225 len -= avail;
226 } while (len);
228 index_crc = p->index_data;
229 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
231 return data_crc != ntohl(*index_crc);
232 }
234 static void copy_pack_data(struct sha1file *f,
235 struct packed_git *p,
236 struct pack_window **w_curs,
237 off_t offset,
238 off_t len)
239 {
240 unsigned char *in;
241 unsigned int avail;
243 while (len) {
244 in = use_pack(p, w_curs, offset, &avail);
245 if (avail > len)
246 avail = (unsigned int)len;
247 sha1write(f, in, avail);
248 offset += avail;
249 len -= avail;
250 }
251 }
253 static unsigned long write_object(struct sha1file *f,
254 struct object_entry *entry,
255 off_t write_offset)
256 {
257 unsigned long size, limit, datalen;
258 void *buf;
259 unsigned char header[10], dheader[10];
260 unsigned hdrlen;
261 enum object_type type;
262 int usable_delta, to_reuse;
264 if (!pack_to_stdout)
265 crc32_begin(f);
267 type = entry->type;
269 /* write limit if limited packsize and not first object */
270 limit = pack_size_limit && nr_written ?
271 pack_size_limit - write_offset : 0;
273 if (!entry->delta)
274 usable_delta = 0; /* no delta */
275 else if (!pack_size_limit)
276 usable_delta = 1; /* unlimited packfile */
277 else if (entry->delta->idx.offset == (off_t)-1)
278 usable_delta = 0; /* base was written to another pack */
279 else if (entry->delta->idx.offset)
280 usable_delta = 1; /* base already exists in this pack */
281 else
282 usable_delta = 0; /* base could end up in another pack */
284 if (!reuse_object)
285 to_reuse = 0; /* explicit */
286 else if (!entry->in_pack)
287 to_reuse = 0; /* can't reuse what we don't have */
288 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
289 /* check_object() decided it for us ... */
290 to_reuse = usable_delta;
291 /* ... but pack split may override that */
292 else if (type != entry->in_pack_type)
293 to_reuse = 0; /* pack has delta which is unusable */
294 else if (entry->delta)
295 to_reuse = 0; /* we want to pack afresh */
296 else
297 to_reuse = 1; /* we have it in-pack undeltified,
298 * and we do not need to deltify it.
299 */
301 if (!to_reuse) {
302 if (!usable_delta) {
303 buf = read_sha1_file(entry->idx.sha1, &type, &size);
304 if (!buf)
305 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
306 /*
307 * make sure no cached delta data remains from a
308 * previous attempt before a pack split occured.
309 */
310 free(entry->delta_data);
311 entry->delta_data = NULL;
312 entry->z_delta_size = 0;
313 } else if (entry->delta_data) {
314 size = entry->delta_size;
315 buf = entry->delta_data;
316 entry->delta_data = NULL;
317 type = (allow_ofs_delta && entry->delta->idx.offset) ?
318 OBJ_OFS_DELTA : OBJ_REF_DELTA;
319 } else {
320 buf = get_delta(entry);
321 size = entry->delta_size;
322 type = (allow_ofs_delta && entry->delta->idx.offset) ?
323 OBJ_OFS_DELTA : OBJ_REF_DELTA;
324 }
326 if (entry->z_delta_size)
327 datalen = entry->z_delta_size;
328 else
329 datalen = do_compress(&buf, size);
331 /*
332 * The object header is a byte of 'type' followed by zero or
333 * more bytes of length.
334 */
335 hdrlen = encode_header(type, size, header);
337 if (type == OBJ_OFS_DELTA) {
338 /*
339 * Deltas with relative base contain an additional
340 * encoding of the relative offset for the delta
341 * base from this object's position in the pack.
342 */
343 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
344 unsigned pos = sizeof(dheader) - 1;
345 dheader[pos] = ofs & 127;
346 while (ofs >>= 7)
347 dheader[--pos] = 128 | (--ofs & 127);
348 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
349 free(buf);
350 return 0;
351 }
352 sha1write(f, header, hdrlen);
353 sha1write(f, dheader + pos, sizeof(dheader) - pos);
354 hdrlen += sizeof(dheader) - pos;
355 } else if (type == OBJ_REF_DELTA) {
356 /*
357 * Deltas with a base reference contain
358 * an additional 20 bytes for the base sha1.
359 */
360 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
361 free(buf);
362 return 0;
363 }
364 sha1write(f, header, hdrlen);
365 sha1write(f, entry->delta->idx.sha1, 20);
366 hdrlen += 20;
367 } else {
368 if (limit && hdrlen + datalen + 20 >= limit) {
369 free(buf);
370 return 0;
371 }
372 sha1write(f, header, hdrlen);
373 }
374 sha1write(f, buf, datalen);
375 free(buf);
376 }
377 else {
378 struct packed_git *p = entry->in_pack;
379 struct pack_window *w_curs = NULL;
380 struct revindex_entry *revidx;
381 off_t offset;
383 if (entry->delta) {
384 type = (allow_ofs_delta && entry->delta->idx.offset) ?
385 OBJ_OFS_DELTA : OBJ_REF_DELTA;
386 reused_delta++;
387 }
388 hdrlen = encode_header(type, entry->size, header);
389 offset = entry->in_pack_offset;
390 revidx = find_pack_revindex(p, offset);
391 datalen = revidx[1].offset - offset;
392 if (!pack_to_stdout && p->index_version > 1 &&
393 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
394 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
395 offset += entry->in_pack_header_size;
396 datalen -= entry->in_pack_header_size;
397 if (type == OBJ_OFS_DELTA) {
398 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
399 unsigned pos = sizeof(dheader) - 1;
400 dheader[pos] = ofs & 127;
401 while (ofs >>= 7)
402 dheader[--pos] = 128 | (--ofs & 127);
403 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
404 return 0;
405 sha1write(f, header, hdrlen);
406 sha1write(f, dheader + pos, sizeof(dheader) - pos);
407 hdrlen += sizeof(dheader) - pos;
408 } else if (type == OBJ_REF_DELTA) {
409 if (limit && hdrlen + 20 + datalen + 20 >= limit)
410 return 0;
411 sha1write(f, header, hdrlen);
412 sha1write(f, entry->delta->idx.sha1, 20);
413 hdrlen += 20;
414 } else {
415 if (limit && hdrlen + datalen + 20 >= limit)
416 return 0;
417 sha1write(f, header, hdrlen);
418 }
420 if (!pack_to_stdout && p->index_version == 1 &&
421 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
422 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
423 copy_pack_data(f, p, &w_curs, offset, datalen);
424 unuse_pack(&w_curs);
425 reused++;
426 }
427 if (usable_delta)
428 written_delta++;
429 written++;
430 if (!pack_to_stdout)
431 entry->idx.crc32 = crc32_end(f);
432 return hdrlen + datalen;
433 }
435 static off_t write_one(struct sha1file *f,
436 struct object_entry *e,
437 off_t offset)
438 {
439 unsigned long size;
441 /* offset is non zero if object is written already. */
442 if (e->idx.offset || e->preferred_base)
443 return offset;
445 /* if we are deltified, write out base object first. */
446 if (e->delta) {
447 offset = write_one(f, e->delta, offset);
448 if (!offset)
449 return 0;
450 }
452 e->idx.offset = offset;
453 size = write_object(f, e, offset);
454 if (!size) {
455 e->idx.offset = 0;
456 return 0;
457 }
458 written_list[nr_written++] = &e->idx;
460 /* make sure off_t is sufficiently large not to wrap */
461 if (offset > offset + size)
462 die("pack too large for current definition of off_t");
463 return offset + size;
464 }
466 /* forward declaration for write_pack_file */
467 static int adjust_perm(const char *path, mode_t mode);
469 static void write_pack_file(void)
470 {
471 uint32_t i = 0, j;
472 struct sha1file *f;
473 off_t offset, offset_one, last_obj_offset = 0;
474 struct pack_header hdr;
475 uint32_t nr_remaining = nr_result;
476 time_t last_mtime = 0;
478 if (progress > pack_to_stdout)
479 progress_state = start_progress("Writing objects", nr_result);
480 written_list = xmalloc(nr_objects * sizeof(*written_list));
482 do {
483 unsigned char sha1[20];
484 char *pack_tmp_name = NULL;
486 if (pack_to_stdout) {
487 f = sha1fd_throughput(1, "<stdout>", progress_state);
488 } else {
489 char tmpname[PATH_MAX];
490 int fd;
491 snprintf(tmpname, sizeof(tmpname),
492 "%s/tmp_pack_XXXXXX", get_object_directory());
493 fd = xmkstemp(tmpname);
494 pack_tmp_name = xstrdup(tmpname);
495 f = sha1fd(fd, pack_tmp_name);
496 }
498 hdr.hdr_signature = htonl(PACK_SIGNATURE);
499 hdr.hdr_version = htonl(PACK_VERSION);
500 hdr.hdr_entries = htonl(nr_remaining);
501 sha1write(f, &hdr, sizeof(hdr));
502 offset = sizeof(hdr);
503 nr_written = 0;
504 for (; i < nr_objects; i++) {
505 last_obj_offset = offset;
506 offset_one = write_one(f, objects + i, offset);
507 if (!offset_one)
508 break;
509 offset = offset_one;
510 display_progress(progress_state, written);
511 }
513 /*
514 * Did we write the wrong # entries in the header?
515 * If so, rewrite it like in fast-import
516 */
517 if (pack_to_stdout) {
518 sha1close(f, sha1, CSUM_CLOSE);
519 } else if (nr_written == nr_remaining) {
520 sha1close(f, sha1, CSUM_FSYNC);
521 } else {
522 int fd = sha1close(f, NULL, 0);
523 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
524 fsync_or_die(fd, pack_tmp_name);
525 close(fd);
526 }
528 if (!pack_to_stdout) {
529 mode_t mode = umask(0);
530 struct stat st;
531 char *idx_tmp_name, tmpname[PATH_MAX];
533 umask(mode);
534 mode = 0444 & ~mode;
536 idx_tmp_name = write_idx_file(NULL, written_list,
537 nr_written, sha1);
539 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
540 base_name, sha1_to_hex(sha1));
541 if (adjust_perm(pack_tmp_name, mode))
542 die("unable to make temporary pack file readable: %s",
543 strerror(errno));
544 if (rename(pack_tmp_name, tmpname))
545 die("unable to rename temporary pack file: %s",
546 strerror(errno));
548 /*
549 * Packs are runtime accessed in their mtime
550 * order since newer packs are more likely to contain
551 * younger objects. So if we are creating multiple
552 * packs then we should modify the mtime of later ones
553 * to preserve this property.
554 */
555 if (stat(tmpname, &st) < 0) {
556 warning("failed to stat %s: %s",
557 tmpname, strerror(errno));
558 } else if (!last_mtime) {
559 last_mtime = st.st_mtime;
560 } else {
561 struct utimbuf utb;
562 utb.actime = st.st_atime;
563 utb.modtime = --last_mtime;
564 if (utime(tmpname, &utb) < 0)
565 warning("failed utime() on %s: %s",
566 tmpname, strerror(errno));
567 }
569 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
570 base_name, sha1_to_hex(sha1));
571 if (adjust_perm(idx_tmp_name, mode))
572 die("unable to make temporary index file readable: %s",
573 strerror(errno));
574 if (rename(idx_tmp_name, tmpname))
575 die("unable to rename temporary index file: %s",
576 strerror(errno));
578 free(idx_tmp_name);
579 free(pack_tmp_name);
580 puts(sha1_to_hex(sha1));
581 }
583 /* mark written objects as written to previous pack */
584 for (j = 0; j < nr_written; j++) {
585 written_list[j]->offset = (off_t)-1;
586 }
587 nr_remaining -= nr_written;
588 } while (nr_remaining && i < nr_objects);
590 free(written_list);
591 stop_progress(&progress_state);
592 if (written != nr_result)
593 die("wrote %u objects while expecting %u", written, nr_result);
594 /*
595 * We have scanned through [0 ... i). Since we have written
596 * the correct number of objects, the remaining [i ... nr_objects)
597 * items must be either already written (due to out-of-order delta base)
598 * or a preferred base. Count those which are neither and complain if any.
599 */
600 for (j = 0; i < nr_objects; i++) {
601 struct object_entry *e = objects + i;
602 j += !e->idx.offset && !e->preferred_base;
603 }
604 if (j)
605 die("wrote %u objects as expected but %u unwritten", written, j);
606 }
608 static int locate_object_entry_hash(const unsigned char *sha1)
609 {
610 int i;
611 unsigned int ui;
612 memcpy(&ui, sha1, sizeof(unsigned int));
613 i = ui % object_ix_hashsz;
614 while (0 < object_ix[i]) {
615 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
616 return i;
617 if (++i == object_ix_hashsz)
618 i = 0;
619 }
620 return -1 - i;
621 }
623 static struct object_entry *locate_object_entry(const unsigned char *sha1)
624 {
625 int i;
627 if (!object_ix_hashsz)
628 return NULL;
630 i = locate_object_entry_hash(sha1);
631 if (0 <= i)
632 return &objects[object_ix[i]-1];
633 return NULL;
634 }
636 static void rehash_objects(void)
637 {
638 uint32_t i;
639 struct object_entry *oe;
641 object_ix_hashsz = nr_objects * 3;
642 if (object_ix_hashsz < 1024)
643 object_ix_hashsz = 1024;
644 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
645 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
646 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
647 int ix = locate_object_entry_hash(oe->idx.sha1);
648 if (0 <= ix)
649 continue;
650 ix = -1 - ix;
651 object_ix[ix] = i + 1;
652 }
653 }
655 static unsigned name_hash(const char *name)
656 {
657 unsigned char c;
658 unsigned hash = 0;
660 if (!name)
661 return 0;
663 /*
664 * This effectively just creates a sortable number from the
665 * last sixteen non-whitespace characters. Last characters
666 * count "most", so things that end in ".c" sort together.
667 */
668 while ((c = *name++) != 0) {
669 if (isspace(c))
670 continue;
671 hash = (hash >> 2) + (c << 24);
672 }
673 return hash;
674 }
676 static void setup_delta_attr_check(struct git_attr_check *check)
677 {
678 static struct git_attr *attr_delta;
680 if (!attr_delta)
681 attr_delta = git_attr("delta", 5);
683 check[0].attr = attr_delta;
684 }
686 static int no_try_delta(const char *path)
687 {
688 struct git_attr_check check[1];
690 setup_delta_attr_check(check);
691 if (git_checkattr(path, ARRAY_SIZE(check), check))
692 return 0;
693 if (ATTR_FALSE(check->value))
694 return 1;
695 return 0;
696 }
698 static int add_object_entry(const unsigned char *sha1, enum object_type type,
699 const char *name, int exclude)
700 {
701 struct object_entry *entry;
702 struct packed_git *p, *found_pack = NULL;
703 off_t found_offset = 0;
704 int ix;
705 unsigned hash = name_hash(name);
707 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
708 if (ix >= 0) {
709 if (exclude) {
710 entry = objects + object_ix[ix] - 1;
711 if (!entry->preferred_base)
712 nr_result--;
713 entry->preferred_base = 1;
714 }
715 return 0;
716 }
718 for (p = packed_git; p; p = p->next) {
719 off_t offset = find_pack_entry_one(sha1, p);
720 if (offset) {
721 if (!found_pack) {
722 found_offset = offset;
723 found_pack = p;
724 }
725 if (exclude)
726 break;
727 if (incremental)
728 return 0;
729 if (local && !p->pack_local)
730 return 0;
731 }
732 }
734 if (nr_objects >= nr_alloc) {
735 nr_alloc = (nr_alloc + 1024) * 3 / 2;
736 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
737 }
739 entry = objects + nr_objects++;
740 memset(entry, 0, sizeof(*entry));
741 hashcpy(entry->idx.sha1, sha1);
742 entry->hash = hash;
743 if (type)
744 entry->type = type;
745 if (exclude)
746 entry->preferred_base = 1;
747 else
748 nr_result++;
749 if (found_pack) {
750 entry->in_pack = found_pack;
751 entry->in_pack_offset = found_offset;
752 }
754 if (object_ix_hashsz * 3 <= nr_objects * 4)
755 rehash_objects();
756 else
757 object_ix[-1 - ix] = nr_objects;
759 display_progress(progress_state, nr_objects);
761 if (name && no_try_delta(name))
762 entry->no_try_delta = 1;
764 return 1;
765 }
767 struct pbase_tree_cache {
768 unsigned char sha1[20];
769 int ref;
770 int temporary;
771 void *tree_data;
772 unsigned long tree_size;
773 };
775 static struct pbase_tree_cache *(pbase_tree_cache[256]);
776 static int pbase_tree_cache_ix(const unsigned char *sha1)
777 {
778 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
779 }
780 static int pbase_tree_cache_ix_incr(int ix)
781 {
782 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
783 }
785 static struct pbase_tree {
786 struct pbase_tree *next;
787 /* This is a phony "cache" entry; we are not
788 * going to evict it nor find it through _get()
789 * mechanism -- this is for the toplevel node that
790 * would almost always change with any commit.
791 */
792 struct pbase_tree_cache pcache;
793 } *pbase_tree;
795 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
796 {
797 struct pbase_tree_cache *ent, *nent;
798 void *data;
799 unsigned long size;
800 enum object_type type;
801 int neigh;
802 int my_ix = pbase_tree_cache_ix(sha1);
803 int available_ix = -1;
805 /* pbase-tree-cache acts as a limited hashtable.
806 * your object will be found at your index or within a few
807 * slots after that slot if it is cached.
808 */
809 for (neigh = 0; neigh < 8; neigh++) {
810 ent = pbase_tree_cache[my_ix];
811 if (ent && !hashcmp(ent->sha1, sha1)) {
812 ent->ref++;
813 return ent;
814 }
815 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
816 ((0 <= available_ix) &&
817 (!ent && pbase_tree_cache[available_ix])))
818 available_ix = my_ix;
819 if (!ent)
820 break;
821 my_ix = pbase_tree_cache_ix_incr(my_ix);
822 }
824 /* Did not find one. Either we got a bogus request or
825 * we need to read and perhaps cache.
826 */
827 data = read_sha1_file(sha1, &type, &size);
828 if (!data)
829 return NULL;
830 if (type != OBJ_TREE) {
831 free(data);
832 return NULL;
833 }
835 /* We need to either cache or return a throwaway copy */
837 if (available_ix < 0)
838 ent = NULL;
839 else {
840 ent = pbase_tree_cache[available_ix];
841 my_ix = available_ix;
842 }
844 if (!ent) {
845 nent = xmalloc(sizeof(*nent));
846 nent->temporary = (available_ix < 0);
847 }
848 else {
849 /* evict and reuse */
850 free(ent->tree_data);
851 nent = ent;
852 }
853 hashcpy(nent->sha1, sha1);
854 nent->tree_data = data;
855 nent->tree_size = size;
856 nent->ref = 1;
857 if (!nent->temporary)
858 pbase_tree_cache[my_ix] = nent;
859 return nent;
860 }
862 static void pbase_tree_put(struct pbase_tree_cache *cache)
863 {
864 if (!cache->temporary) {
865 cache->ref--;
866 return;
867 }
868 free(cache->tree_data);
869 free(cache);
870 }
872 static int name_cmp_len(const char *name)
873 {
874 int i;
875 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
876 ;
877 return i;
878 }
880 static void add_pbase_object(struct tree_desc *tree,
881 const char *name,
882 int cmplen,
883 const char *fullname)
884 {
885 struct name_entry entry;
886 int cmp;
888 while (tree_entry(tree,&entry)) {
889 if (S_ISGITLINK(entry.mode))
890 continue;
891 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
892 memcmp(name, entry.path, cmplen);
893 if (cmp > 0)
894 continue;
895 if (cmp < 0)
896 return;
897 if (name[cmplen] != '/') {
898 add_object_entry(entry.sha1,
899 object_type(entry.mode),
900 fullname, 1);
901 return;
902 }
903 if (S_ISDIR(entry.mode)) {
904 struct tree_desc sub;
905 struct pbase_tree_cache *tree;
906 const char *down = name+cmplen+1;
907 int downlen = name_cmp_len(down);
909 tree = pbase_tree_get(entry.sha1);
910 if (!tree)
911 return;
912 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
914 add_pbase_object(&sub, down, downlen, fullname);
915 pbase_tree_put(tree);
916 }
917 }
918 }
920 static unsigned *done_pbase_paths;
921 static int done_pbase_paths_num;
922 static int done_pbase_paths_alloc;
923 static int done_pbase_path_pos(unsigned hash)
924 {
925 int lo = 0;
926 int hi = done_pbase_paths_num;
927 while (lo < hi) {
928 int mi = (hi + lo) / 2;
929 if (done_pbase_paths[mi] == hash)
930 return mi;
931 if (done_pbase_paths[mi] < hash)
932 hi = mi;
933 else
934 lo = mi + 1;
935 }
936 return -lo-1;
937 }
939 static int check_pbase_path(unsigned hash)
940 {
941 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
942 if (0 <= pos)
943 return 1;
944 pos = -pos - 1;
945 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
946 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
947 done_pbase_paths = xrealloc(done_pbase_paths,
948 done_pbase_paths_alloc *
949 sizeof(unsigned));
950 }
951 done_pbase_paths_num++;
952 if (pos < done_pbase_paths_num)
953 memmove(done_pbase_paths + pos + 1,
954 done_pbase_paths + pos,
955 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
956 done_pbase_paths[pos] = hash;
957 return 0;
958 }
960 static void add_preferred_base_object(const char *name)
961 {
962 struct pbase_tree *it;
963 int cmplen;
964 unsigned hash = name_hash(name);
966 if (!num_preferred_base || check_pbase_path(hash))
967 return;
969 cmplen = name_cmp_len(name);
970 for (it = pbase_tree; it; it = it->next) {
971 if (cmplen == 0) {
972 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
973 }
974 else {
975 struct tree_desc tree;
976 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
977 add_pbase_object(&tree, name, cmplen, name);
978 }
979 }
980 }
982 static void add_preferred_base(unsigned char *sha1)
983 {
984 struct pbase_tree *it;
985 void *data;
986 unsigned long size;
987 unsigned char tree_sha1[20];
989 if (window <= num_preferred_base++)
990 return;
992 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
993 if (!data)
994 return;
996 for (it = pbase_tree; it; it = it->next) {
997 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
998 free(data);
999 return;
1000 }
1001 }
1003 it = xcalloc(1, sizeof(*it));
1004 it->next = pbase_tree;
1005 pbase_tree = it;
1007 hashcpy(it->pcache.sha1, tree_sha1);
1008 it->pcache.tree_data = data;
1009 it->pcache.tree_size = size;
1010 }
1012 static void check_object(struct object_entry *entry)
1013 {
1014 if (entry->in_pack) {
1015 struct packed_git *p = entry->in_pack;
1016 struct pack_window *w_curs = NULL;
1017 const unsigned char *base_ref = NULL;
1018 struct object_entry *base_entry;
1019 unsigned long used, used_0;
1020 unsigned int avail;
1021 off_t ofs;
1022 unsigned char *buf, c;
1024 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1026 /*
1027 * We want in_pack_type even if we do not reuse delta
1028 * since non-delta representations could still be reused.
1029 */
1030 used = unpack_object_header_gently(buf, avail,
1031 &entry->in_pack_type,
1032 &entry->size);
1034 /*
1035 * Determine if this is a delta and if so whether we can
1036 * reuse it or not. Otherwise let's find out as cheaply as
1037 * possible what the actual type and size for this object is.
1038 */
1039 switch (entry->in_pack_type) {
1040 default:
1041 /* Not a delta hence we've already got all we need. */
1042 entry->type = entry->in_pack_type;
1043 entry->in_pack_header_size = used;
1044 unuse_pack(&w_curs);
1045 return;
1046 case OBJ_REF_DELTA:
1047 if (reuse_delta && !entry->preferred_base)
1048 base_ref = use_pack(p, &w_curs,
1049 entry->in_pack_offset + used, NULL);
1050 entry->in_pack_header_size = used + 20;
1051 break;
1052 case OBJ_OFS_DELTA:
1053 buf = use_pack(p, &w_curs,
1054 entry->in_pack_offset + used, NULL);
1055 used_0 = 0;
1056 c = buf[used_0++];
1057 ofs = c & 127;
1058 while (c & 128) {
1059 ofs += 1;
1060 if (!ofs || MSB(ofs, 7))
1061 die("delta base offset overflow in pack for %s",
1062 sha1_to_hex(entry->idx.sha1));
1063 c = buf[used_0++];
1064 ofs = (ofs << 7) + (c & 127);
1065 }
1066 if (ofs >= entry->in_pack_offset)
1067 die("delta base offset out of bound for %s",
1068 sha1_to_hex(entry->idx.sha1));
1069 ofs = entry->in_pack_offset - ofs;
1070 if (reuse_delta && !entry->preferred_base) {
1071 struct revindex_entry *revidx;
1072 revidx = find_pack_revindex(p, ofs);
1073 base_ref = nth_packed_object_sha1(p, revidx->nr);
1074 }
1075 entry->in_pack_header_size = used + used_0;
1076 break;
1077 }
1079 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1080 /*
1081 * If base_ref was set above that means we wish to
1082 * reuse delta data, and we even found that base
1083 * in the list of objects we want to pack. Goodie!
1084 *
1085 * Depth value does not matter - find_deltas() will
1086 * never consider reused delta as the base object to
1087 * deltify other objects against, in order to avoid
1088 * circular deltas.
1089 */
1090 entry->type = entry->in_pack_type;
1091 entry->delta = base_entry;
1092 entry->delta_sibling = base_entry->delta_child;
1093 base_entry->delta_child = entry;
1094 unuse_pack(&w_curs);
1095 return;
1096 }
1098 if (entry->type) {
1099 /*
1100 * This must be a delta and we already know what the
1101 * final object type is. Let's extract the actual
1102 * object size from the delta header.
1103 */
1104 entry->size = get_size_from_delta(p, &w_curs,
1105 entry->in_pack_offset + entry->in_pack_header_size);
1106 unuse_pack(&w_curs);
1107 return;
1108 }
1110 /*
1111 * No choice but to fall back to the recursive delta walk
1112 * with sha1_object_info() to find about the object type
1113 * at this point...
1114 */
1115 unuse_pack(&w_curs);
1116 }
1118 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1119 if (entry->type < 0)
1120 die("unable to get type of object %s",
1121 sha1_to_hex(entry->idx.sha1));
1122 }
1124 static int pack_offset_sort(const void *_a, const void *_b)
1125 {
1126 const struct object_entry *a = *(struct object_entry **)_a;
1127 const struct object_entry *b = *(struct object_entry **)_b;
1129 /* avoid filesystem trashing with loose objects */
1130 if (!a->in_pack && !b->in_pack)
1131 return hashcmp(a->idx.sha1, b->idx.sha1);
1133 if (a->in_pack < b->in_pack)
1134 return -1;
1135 if (a->in_pack > b->in_pack)
1136 return 1;
1137 return a->in_pack_offset < b->in_pack_offset ? -1 :
1138 (a->in_pack_offset > b->in_pack_offset);
1139 }
1141 static void get_object_details(void)
1142 {
1143 uint32_t i;
1144 struct object_entry **sorted_by_offset;
1146 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1147 for (i = 0; i < nr_objects; i++)
1148 sorted_by_offset[i] = objects + i;
1149 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1151 for (i = 0; i < nr_objects; i++)
1152 check_object(sorted_by_offset[i]);
1154 free(sorted_by_offset);
1155 }
1157 /*
1158 * We search for deltas in a list sorted by type, by filename hash, and then
1159 * by size, so that we see progressively smaller and smaller files.
1160 * That's because we prefer deltas to be from the bigger file
1161 * to the smaller -- deletes are potentially cheaper, but perhaps
1162 * more importantly, the bigger file is likely the more recent
1163 * one. The deepest deltas are therefore the oldest objects which are
1164 * less susceptible to be accessed often.
1165 */
1166 static int type_size_sort(const void *_a, const void *_b)
1167 {
1168 const struct object_entry *a = *(struct object_entry **)_a;
1169 const struct object_entry *b = *(struct object_entry **)_b;
1171 if (a->type > b->type)
1172 return -1;
1173 if (a->type < b->type)
1174 return 1;
1175 if (a->hash > b->hash)
1176 return -1;
1177 if (a->hash < b->hash)
1178 return 1;
1179 if (a->preferred_base > b->preferred_base)
1180 return -1;
1181 if (a->preferred_base < b->preferred_base)
1182 return 1;
1183 if (a->size > b->size)
1184 return -1;
1185 if (a->size < b->size)
1186 return 1;
1187 return a < b ? -1 : (a > b); /* newest first */
1188 }
1190 struct unpacked {
1191 struct object_entry *entry;
1192 void *data;
1193 struct delta_index *index;
1194 unsigned depth;
1195 };
1197 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1198 unsigned long delta_size)
1199 {
1200 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1201 return 0;
1203 if (delta_size < cache_max_small_delta_size)
1204 return 1;
1206 /* cache delta, if objects are large enough compared to delta size */
1207 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1208 return 1;
1210 return 0;
1211 }
1213 #ifdef THREADED_DELTA_SEARCH
1215 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1216 #define read_lock() pthread_mutex_lock(&read_mutex)
1217 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1219 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1220 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1221 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1223 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1224 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1225 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1227 #else
1229 #define read_lock() (void)0
1230 #define read_unlock() (void)0
1231 #define cache_lock() (void)0
1232 #define cache_unlock() (void)0
1233 #define progress_lock() (void)0
1234 #define progress_unlock() (void)0
1236 #endif
1238 static int try_delta(struct unpacked *trg, struct unpacked *src,
1239 unsigned max_depth, unsigned long *mem_usage)
1240 {
1241 struct object_entry *trg_entry = trg->entry;
1242 struct object_entry *src_entry = src->entry;
1243 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1244 unsigned ref_depth;
1245 enum object_type type;
1246 void *delta_buf;
1248 /* Don't bother doing diffs between different types */
1249 if (trg_entry->type != src_entry->type)
1250 return -1;
1252 /*
1253 * We do not bother to try a delta that we discarded
1254 * on an earlier try, but only when reusing delta data.
1255 */
1256 if (reuse_delta && trg_entry->in_pack &&
1257 trg_entry->in_pack == src_entry->in_pack &&
1258 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1259 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1260 return 0;
1262 /* Let's not bust the allowed depth. */
1263 if (src->depth >= max_depth)
1264 return 0;
1266 /* Now some size filtering heuristics. */
1267 trg_size = trg_entry->size;
1268 if (!trg_entry->delta) {
1269 max_size = trg_size/2 - 20;
1270 ref_depth = 1;
1271 } else {
1272 max_size = trg_entry->delta_size;
1273 ref_depth = trg->depth;
1274 }
1275 max_size = max_size * (max_depth - src->depth) /
1276 (max_depth - ref_depth + 1);
1277 if (max_size == 0)
1278 return 0;
1279 src_size = src_entry->size;
1280 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1281 if (sizediff >= max_size)
1282 return 0;
1283 if (trg_size < src_size / 32)
1284 return 0;
1286 /* Load data if not already done */
1287 if (!trg->data) {
1288 read_lock();
1289 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1290 read_unlock();
1291 if (!trg->data)
1292 die("object %s cannot be read",
1293 sha1_to_hex(trg_entry->idx.sha1));
1294 if (sz != trg_size)
1295 die("object %s inconsistent object length (%lu vs %lu)",
1296 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1297 *mem_usage += sz;
1298 }
1299 if (!src->data) {
1300 read_lock();
1301 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1302 read_unlock();
1303 if (!src->data)
1304 die("object %s cannot be read",
1305 sha1_to_hex(src_entry->idx.sha1));
1306 if (sz != src_size)
1307 die("object %s inconsistent object length (%lu vs %lu)",
1308 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1309 *mem_usage += sz;
1310 }
1311 if (!src->index) {
1312 src->index = create_delta_index(src->data, src_size);
1313 if (!src->index) {
1314 static int warned = 0;
1315 if (!warned++)
1316 warning("suboptimal pack - out of memory");
1317 return 0;
1318 }
1319 *mem_usage += sizeof_delta_index(src->index);
1320 }
1322 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1323 if (!delta_buf)
1324 return 0;
1326 if (trg_entry->delta) {
1327 /* Prefer only shallower same-sized deltas. */
1328 if (delta_size == trg_entry->delta_size &&
1329 src->depth + 1 >= trg->depth) {
1330 free(delta_buf);
1331 return 0;
1332 }
1333 }
1335 /*
1336 * Handle memory allocation outside of the cache
1337 * accounting lock. Compiler will optimize the strangeness
1338 * away when THREADED_DELTA_SEARCH is not defined.
1339 */
1340 free(trg_entry->delta_data);
1341 cache_lock();
1342 if (trg_entry->delta_data) {
1343 delta_cache_size -= trg_entry->delta_size;
1344 trg_entry->delta_data = NULL;
1345 }
1346 if (delta_cacheable(src_size, trg_size, delta_size)) {
1347 delta_cache_size += delta_size;
1348 cache_unlock();
1349 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1350 } else {
1351 cache_unlock();
1352 free(delta_buf);
1353 }
1355 trg_entry->delta = src_entry;
1356 trg_entry->delta_size = delta_size;
1357 trg->depth = src->depth + 1;
1359 return 1;
1360 }
1362 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1363 {
1364 struct object_entry *child = me->delta_child;
1365 unsigned int m = n;
1366 while (child) {
1367 unsigned int c = check_delta_limit(child, n + 1);
1368 if (m < c)
1369 m = c;
1370 child = child->delta_sibling;
1371 }
1372 return m;
1373 }
1375 static unsigned long free_unpacked(struct unpacked *n)
1376 {
1377 unsigned long freed_mem = sizeof_delta_index(n->index);
1378 free_delta_index(n->index);
1379 n->index = NULL;
1380 if (n->data) {
1381 freed_mem += n->entry->size;
1382 free(n->data);
1383 n->data = NULL;
1384 }
1385 n->entry = NULL;
1386 n->depth = 0;
1387 return freed_mem;
1388 }
1390 static void find_deltas(struct object_entry **list, unsigned *list_size,
1391 int window, int depth, unsigned *processed)
1392 {
1393 uint32_t i, idx = 0, count = 0;
1394 unsigned int array_size = window * sizeof(struct unpacked);
1395 struct unpacked *array;
1396 unsigned long mem_usage = 0;
1398 array = xmalloc(array_size);
1399 memset(array, 0, array_size);
1401 for (;;) {
1402 struct object_entry *entry = *list++;
1403 struct unpacked *n = array + idx;
1404 int j, max_depth, best_base = -1;
1406 progress_lock();
1407 if (!*list_size) {
1408 progress_unlock();
1409 break;
1410 }
1411 (*list_size)--;
1412 if (!entry->preferred_base) {
1413 (*processed)++;
1414 display_progress(progress_state, *processed);
1415 }
1416 progress_unlock();
1418 mem_usage -= free_unpacked(n);
1419 n->entry = entry;
1421 while (window_memory_limit &&
1422 mem_usage > window_memory_limit &&
1423 count > 1) {
1424 uint32_t tail = (idx + window - count) % window;
1425 mem_usage -= free_unpacked(array + tail);
1426 count--;
1427 }
1429 /* We do not compute delta to *create* objects we are not
1430 * going to pack.
1431 */
1432 if (entry->preferred_base)
1433 goto next;
1435 /*
1436 * If the current object is at pack edge, take the depth the
1437 * objects that depend on the current object into account
1438 * otherwise they would become too deep.
1439 */
1440 max_depth = depth;
1441 if (entry->delta_child) {
1442 max_depth -= check_delta_limit(entry, 0);
1443 if (max_depth <= 0)
1444 goto next;
1445 }
1447 j = window;
1448 while (--j > 0) {
1449 int ret;
1450 uint32_t other_idx = idx + j;
1451 struct unpacked *m;
1452 if (other_idx >= window)
1453 other_idx -= window;
1454 m = array + other_idx;
1455 if (!m->entry)
1456 break;
1457 ret = try_delta(n, m, max_depth, &mem_usage);
1458 if (ret < 0)
1459 break;
1460 else if (ret > 0)
1461 best_base = other_idx;
1462 }
1464 /*
1465 * If we decided to cache the delta data, then it is best
1466 * to compress it right away. First because we have to do
1467 * it anyway, and doing it here while we're threaded will
1468 * save a lot of time in the non threaded write phase,
1469 * as well as allow for caching more deltas within
1470 * the same cache size limit.
1471 * ...
1472 * But only if not writing to stdout, since in that case
1473 * the network is most likely throttling writes anyway,
1474 * and therefore it is best to go to the write phase ASAP
1475 * instead, as we can afford spending more time compressing
1476 * between writes at that moment.
1477 */
1478 if (entry->delta_data && !pack_to_stdout) {
1479 entry->z_delta_size = do_compress(&entry->delta_data,
1480 entry->delta_size);
1481 cache_lock();
1482 delta_cache_size -= entry->delta_size;
1483 delta_cache_size += entry->z_delta_size;
1484 cache_unlock();
1485 }
1487 /* if we made n a delta, and if n is already at max
1488 * depth, leaving it in the window is pointless. we
1489 * should evict it first.
1490 */
1491 if (entry->delta && max_depth <= n->depth)
1492 continue;
1494 /*
1495 * Move the best delta base up in the window, after the
1496 * currently deltified object, to keep it longer. It will
1497 * be the first base object to be attempted next.
1498 */
1499 if (entry->delta) {
1500 struct unpacked swap = array[best_base];
1501 int dist = (window + idx - best_base) % window;
1502 int dst = best_base;
1503 while (dist--) {
1504 int src = (dst + 1) % window;
1505 array[dst] = array[src];
1506 dst = src;
1507 }
1508 array[dst] = swap;
1509 }
1511 next:
1512 idx++;
1513 if (count + 1 < window)
1514 count++;
1515 if (idx >= window)
1516 idx = 0;
1517 }
1519 for (i = 0; i < window; ++i) {
1520 free_delta_index(array[i].index);
1521 free(array[i].data);
1522 }
1523 free(array);
1524 }
1526 #ifdef THREADED_DELTA_SEARCH
1528 /*
1529 * The main thread waits on the condition that (at least) one of the workers
1530 * has stopped working (which is indicated in the .working member of
1531 * struct thread_params).
1532 * When a work thread has completed its work, it sets .working to 0 and
1533 * signals the main thread and waits on the condition that .data_ready
1534 * becomes 1.
1535 */
1537 struct thread_params {
1538 pthread_t thread;
1539 struct object_entry **list;
1540 unsigned list_size;
1541 unsigned remaining;
1542 int window;
1543 int depth;
1544 int working;
1545 int data_ready;
1546 pthread_mutex_t mutex;
1547 pthread_cond_t cond;
1548 unsigned *processed;
1549 };
1551 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1553 static void *threaded_find_deltas(void *arg)
1554 {
1555 struct thread_params *me = arg;
1557 while (me->remaining) {
1558 find_deltas(me->list, &me->remaining,
1559 me->window, me->depth, me->processed);
1561 progress_lock();
1562 me->working = 0;
1563 pthread_cond_signal(&progress_cond);
1564 progress_unlock();
1566 /*
1567 * We must not set ->data_ready before we wait on the
1568 * condition because the main thread may have set it to 1
1569 * before we get here. In order to be sure that new
1570 * work is available if we see 1 in ->data_ready, it
1571 * was initialized to 0 before this thread was spawned
1572 * and we reset it to 0 right away.
1573 */
1574 pthread_mutex_lock(&me->mutex);
1575 while (!me->data_ready)
1576 pthread_cond_wait(&me->cond, &me->mutex);
1577 me->data_ready = 0;
1578 pthread_mutex_unlock(&me->mutex);
1579 }
1580 /* leave ->working 1 so that this doesn't get more work assigned */
1581 return NULL;
1582 }
1584 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1585 int window, int depth, unsigned *processed)
1586 {
1587 struct thread_params p[delta_search_threads];
1588 int i, ret, active_threads = 0;
1590 if (delta_search_threads <= 1) {
1591 find_deltas(list, &list_size, window, depth, processed);
1592 return;
1593 }
1595 /* Partition the work amongst work threads. */
1596 for (i = 0; i < delta_search_threads; i++) {
1597 unsigned sub_size = list_size / (delta_search_threads - i);
1599 p[i].window = window;
1600 p[i].depth = depth;
1601 p[i].processed = processed;
1602 p[i].working = 1;
1603 p[i].data_ready = 0;
1605 /* try to split chunks on "path" boundaries */
1606 while (sub_size && sub_size < list_size &&
1607 list[sub_size]->hash &&
1608 list[sub_size]->hash == list[sub_size-1]->hash)
1609 sub_size++;
1611 p[i].list = list;
1612 p[i].list_size = sub_size;
1613 p[i].remaining = sub_size;
1615 list += sub_size;
1616 list_size -= sub_size;
1617 }
1619 /* Start work threads. */
1620 for (i = 0; i < delta_search_threads; i++) {
1621 if (!p[i].list_size)
1622 continue;
1623 pthread_mutex_init(&p[i].mutex, NULL);
1624 pthread_cond_init(&p[i].cond, NULL);
1625 ret = pthread_create(&p[i].thread, NULL,
1626 threaded_find_deltas, &p[i]);
1627 if (ret)
1628 die("unable to create thread: %s", strerror(ret));
1629 active_threads++;
1630 }
1632 /*
1633 * Now let's wait for work completion. Each time a thread is done
1634 * with its work, we steal half of the remaining work from the
1635 * thread with the largest number of unprocessed objects and give
1636 * it to that newly idle thread. This ensure good load balancing
1637 * until the remaining object list segments are simply too short
1638 * to be worth splitting anymore.
1639 */
1640 while (active_threads) {
1641 struct thread_params *target = NULL;
1642 struct thread_params *victim = NULL;
1643 unsigned sub_size = 0;
1645 progress_lock();
1646 for (;;) {
1647 for (i = 0; !target && i < delta_search_threads; i++)
1648 if (!p[i].working)
1649 target = &p[i];
1650 if (target)
1651 break;
1652 pthread_cond_wait(&progress_cond, &progress_mutex);
1653 }
1655 for (i = 0; i < delta_search_threads; i++)
1656 if (p[i].remaining > 2*window &&
1657 (!victim || victim->remaining < p[i].remaining))
1658 victim = &p[i];
1659 if (victim) {
1660 sub_size = victim->remaining / 2;
1661 list = victim->list + victim->list_size - sub_size;
1662 while (sub_size && list[0]->hash &&
1663 list[0]->hash == list[-1]->hash) {
1664 list++;
1665 sub_size--;
1666 }
1667 if (!sub_size) {
1668 /*
1669 * It is possible for some "paths" to have
1670 * so many objects that no hash boundary
1671 * might be found. Let's just steal the
1672 * exact half in that case.
1673 */
1674 sub_size = victim->remaining / 2;
1675 list -= sub_size;
1676 }
1677 target->list = list;
1678 victim->list_size -= sub_size;
1679 victim->remaining -= sub_size;
1680 }
1681 target->list_size = sub_size;
1682 target->remaining = sub_size;
1683 target->working = 1;
1684 progress_unlock();
1686 pthread_mutex_lock(&target->mutex);
1687 target->data_ready = 1;
1688 pthread_cond_signal(&target->cond);
1689 pthread_mutex_unlock(&target->mutex);
1691 if (!sub_size) {
1692 pthread_join(target->thread, NULL);
1693 pthread_cond_destroy(&target->cond);
1694 pthread_mutex_destroy(&target->mutex);
1695 active_threads--;
1696 }
1697 }
1698 }
1700 #else
1701 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1702 #endif
1704 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1705 {
1706 unsigned char peeled[20];
1708 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1709 !peel_ref(path, peeled) && /* peelable? */
1710 !is_null_sha1(peeled) && /* annotated tag? */
1711 locate_object_entry(peeled)) /* object packed? */
1712 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1713 return 0;
1714 }
1716 static void prepare_pack(int window, int depth)
1717 {
1718 struct object_entry **delta_list;
1719 uint32_t i, n, nr_deltas;
1721 get_object_details();
1723 if (!nr_objects || !window || !depth)
1724 return;
1726 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1727 nr_deltas = n = 0;
1729 for (i = 0; i < nr_objects; i++) {
1730 struct object_entry *entry = objects + i;
1732 if (entry->delta)
1733 /* This happens if we decided to reuse existing
1734 * delta from a pack. "reuse_delta &&" is implied.
1735 */
1736 continue;
1738 if (entry->size < 50)
1739 continue;
1741 if (entry->no_try_delta)
1742 continue;
1744 if (!entry->preferred_base)
1745 nr_deltas++;
1747 delta_list[n++] = entry;
1748 }
1750 if (nr_deltas && n > 1) {
1751 unsigned nr_done = 0;
1752 if (progress)
1753 progress_state = start_progress("Compressing objects",
1754 nr_deltas);
1755 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1756 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1757 stop_progress(&progress_state);
1758 if (nr_done != nr_deltas)
1759 die("inconsistency with delta count");
1760 }
1761 free(delta_list);
1762 }
1764 static int git_pack_config(const char *k, const char *v, void *cb)
1765 {
1766 if(!strcmp(k, "pack.window")) {
1767 window = git_config_int(k, v);
1768 return 0;
1769 }
1770 if (!strcmp(k, "pack.windowmemory")) {
1771 window_memory_limit = git_config_ulong(k, v);
1772 return 0;
1773 }
1774 if (!strcmp(k, "pack.depth")) {
1775 depth = git_config_int(k, v);
1776 return 0;
1777 }
1778 if (!strcmp(k, "pack.compression")) {
1779 int level = git_config_int(k, v);
1780 if (level == -1)
1781 level = Z_DEFAULT_COMPRESSION;
1782 else if (level < 0 || level > Z_BEST_COMPRESSION)
1783 die("bad pack compression level %d", level);
1784 pack_compression_level = level;
1785 pack_compression_seen = 1;
1786 return 0;
1787 }
1788 if (!strcmp(k, "pack.deltacachesize")) {
1789 max_delta_cache_size = git_config_int(k, v);
1790 return 0;
1791 }
1792 if (!strcmp(k, "pack.deltacachelimit")) {
1793 cache_max_small_delta_size = git_config_int(k, v);
1794 return 0;
1795 }
1796 if (!strcmp(k, "pack.threads")) {
1797 delta_search_threads = git_config_int(k, v);
1798 if (delta_search_threads < 0)
1799 die("invalid number of threads specified (%d)",
1800 delta_search_threads);
1801 #ifndef THREADED_DELTA_SEARCH
1802 if (delta_search_threads != 1)
1803 warning("no threads support, ignoring %s", k);
1804 #endif
1805 return 0;
1806 }
1807 if (!strcmp(k, "pack.indexversion")) {
1808 pack_idx_default_version = git_config_int(k, v);
1809 if (pack_idx_default_version > 2)
1810 die("bad pack.indexversion=%d", pack_idx_default_version);
1811 return 0;
1812 }
1813 if (!strcmp(k, "pack.packsizelimit")) {
1814 pack_size_limit_cfg = git_config_ulong(k, v);
1815 return 0;
1816 }
1817 return git_default_config(k, v, cb);
1818 }
1820 static void read_object_list_from_stdin(void)
1821 {
1822 char line[40 + 1 + PATH_MAX + 2];
1823 unsigned char sha1[20];
1825 for (;;) {
1826 if (!fgets(line, sizeof(line), stdin)) {
1827 if (feof(stdin))
1828 break;
1829 if (!ferror(stdin))
1830 die("fgets returned NULL, not EOF, not error!");
1831 if (errno != EINTR)
1832 die("fgets: %s", strerror(errno));
1833 clearerr(stdin);
1834 continue;
1835 }
1836 if (line[0] == '-') {
1837 if (get_sha1_hex(line+1, sha1))
1838 die("expected edge sha1, got garbage:\n %s",
1839 line);
1840 add_preferred_base(sha1);
1841 continue;
1842 }
1843 if (get_sha1_hex(line, sha1))
1844 die("expected sha1, got garbage:\n %s", line);
1846 add_preferred_base_object(line+41);
1847 add_object_entry(sha1, 0, line+41, 0);
1848 }
1849 }
1851 #define OBJECT_ADDED (1u<<20)
1853 static void show_commit(struct commit *commit)
1854 {
1855 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1856 commit->object.flags |= OBJECT_ADDED;
1857 }
1859 static void show_object(struct object_array_entry *p)
1860 {
1861 add_preferred_base_object(p->name);
1862 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1863 p->item->flags |= OBJECT_ADDED;
1864 }
1866 static void show_edge(struct commit *commit)
1867 {
1868 add_preferred_base(commit->object.sha1);
1869 }
1871 struct in_pack_object {
1872 off_t offset;
1873 struct object *object;
1874 };
1876 struct in_pack {
1877 int alloc;
1878 int nr;
1879 struct in_pack_object *array;
1880 };
1882 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1883 {
1884 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1885 in_pack->array[in_pack->nr].object = object;
1886 in_pack->nr++;
1887 }
1889 /*
1890 * Compare the objects in the offset order, in order to emulate the
1891 * "git-rev-list --objects" output that produced the pack originally.
1892 */
1893 static int ofscmp(const void *a_, const void *b_)
1894 {
1895 struct in_pack_object *a = (struct in_pack_object *)a_;
1896 struct in_pack_object *b = (struct in_pack_object *)b_;
1898 if (a->offset < b->offset)
1899 return -1;
1900 else if (a->offset > b->offset)
1901 return 1;
1902 else
1903 return hashcmp(a->object->sha1, b->object->sha1);
1904 }
1906 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1907 {
1908 struct packed_git *p;
1909 struct in_pack in_pack;
1910 uint32_t i;
1912 memset(&in_pack, 0, sizeof(in_pack));
1914 for (p = packed_git; p; p = p->next) {
1915 const unsigned char *sha1;
1916 struct object *o;
1918 for (i = 0; i < revs->num_ignore_packed; i++) {
1919 if (matches_pack_name(p, revs->ignore_packed[i]))
1920 break;
1921 }
1922 if (revs->num_ignore_packed <= i)
1923 continue;
1924 if (open_pack_index(p))
1925 die("cannot open pack index");
1927 ALLOC_GROW(in_pack.array,
1928 in_pack.nr + p->num_objects,
1929 in_pack.alloc);
1931 for (i = 0; i < p->num_objects; i++) {
1932 sha1 = nth_packed_object_sha1(p, i);
1933 o = lookup_unknown_object(sha1);
1934 if (!(o->flags & OBJECT_ADDED))
1935 mark_in_pack_object(o, p, &in_pack);
1936 o->flags |= OBJECT_ADDED;
1937 }
1938 }
1940 if (in_pack.nr) {
1941 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1942 ofscmp);
1943 for (i = 0; i < in_pack.nr; i++) {
1944 struct object *o = in_pack.array[i].object;
1945 add_object_entry(o->sha1, o->type, "", 0);
1946 }
1947 }
1948 free(in_pack.array);
1949 }
1951 static void loosen_unused_packed_objects(struct rev_info *revs)
1952 {
1953 struct packed_git *p;
1954 uint32_t i;
1955 const unsigned char *sha1;
1957 for (p = packed_git; p; p = p->next) {
1958 for (i = 0; i < revs->num_ignore_packed; i++) {
1959 if (matches_pack_name(p, revs->ignore_packed[i]))
1960 break;
1961 }
1962 if (revs->num_ignore_packed <= i)
1963 continue;
1965 if (open_pack_index(p))
1966 die("cannot open pack index");
1968 for (i = 0; i < p->num_objects; i++) {
1969 sha1 = nth_packed_object_sha1(p, i);
1970 if (!locate_object_entry(sha1))
1971 if (force_object_loose(sha1, p->mtime))
1972 die("unable to force loose object");
1973 }
1974 }
1975 }
1977 static void get_object_list(int ac, const char **av)
1978 {
1979 struct rev_info revs;
1980 char line[1000];
1981 int flags = 0;
1983 init_revisions(&revs, NULL);
1984 save_commit_buffer = 0;
1985 setup_revisions(ac, av, &revs, NULL);
1987 while (fgets(line, sizeof(line), stdin) != NULL) {
1988 int len = strlen(line);
1989 if (len && line[len - 1] == '\n')
1990 line[--len] = 0;
1991 if (!len)
1992 break;
1993 if (*line == '-') {
1994 if (!strcmp(line, "--not")) {
1995 flags ^= UNINTERESTING;
1996 continue;
1997 }
1998 die("not a rev '%s'", line);
1999 }
2000 if (handle_revision_arg(line, &revs, flags, 1))
2001 die("bad revision '%s'", line);
2002 }
2004 if (prepare_revision_walk(&revs))
2005 die("revision walk setup failed");
2006 mark_edges_uninteresting(revs.commits, &revs, show_edge);
2007 traverse_commit_list(&revs, show_commit, show_object);
2009 if (keep_unreachable)
2010 add_objects_in_unpacked_packs(&revs);
2011 if (unpack_unreachable)
2012 loosen_unused_packed_objects(&revs);
2013 }
2015 static int adjust_perm(const char *path, mode_t mode)
2016 {
2017 if (chmod(path, mode))
2018 return -1;
2019 return adjust_shared_perm(path);
2020 }
2022 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2023 {
2024 int use_internal_rev_list = 0;
2025 int thin = 0;
2026 uint32_t i;
2027 const char **rp_av;
2028 int rp_ac_alloc = 64;
2029 int rp_ac;
2031 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2033 rp_av[0] = "pack-objects";
2034 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2035 rp_ac = 2;
2037 git_config(git_pack_config, NULL);
2038 if (!pack_compression_seen && core_compression_seen)
2039 pack_compression_level = core_compression_level;
2041 progress = isatty(2);
2042 for (i = 1; i < argc; i++) {
2043 const char *arg = argv[i];
2045 if (*arg != '-')
2046 break;
2048 if (!strcmp("--non-empty", arg)) {
2049 non_empty = 1;
2050 continue;
2051 }
2052 if (!strcmp("--local", arg)) {
2053 local = 1;
2054 continue;
2055 }
2056 if (!strcmp("--incremental", arg)) {
2057 incremental = 1;
2058 continue;
2059 }
2060 if (!prefixcmp(arg, "--compression=")) {
2061 char *end;
2062 int level = strtoul(arg+14, &end, 0);
2063 if (!arg[14] || *end)
2064 usage(pack_usage);
2065 if (level == -1)
2066 level = Z_DEFAULT_COMPRESSION;
2067 else if (level < 0 || level > Z_BEST_COMPRESSION)
2068 die("bad pack compression level %d", level);
2069 pack_compression_level = level;
2070 continue;
2071 }
2072 if (!prefixcmp(arg, "--max-pack-size=")) {
2073 char *end;
2074 pack_size_limit_cfg = 0;
2075 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2076 if (!arg[16] || *end)
2077 usage(pack_usage);
2078 continue;
2079 }
2080 if (!prefixcmp(arg, "--window=")) {
2081 char *end;
2082 window = strtoul(arg+9, &end, 0);
2083 if (!arg[9] || *end)
2084 usage(pack_usage);
2085 continue;
2086 }
2087 if (!prefixcmp(arg, "--window-memory=")) {
2088 if (!git_parse_ulong(arg+16, &window_memory_limit))
2089 usage(pack_usage);
2090 continue;
2091 }
2092 if (!prefixcmp(arg, "--threads=")) {
2093 char *end;
2094 delta_search_threads = strtoul(arg+10, &end, 0);
2095 if (!arg[10] || *end || delta_search_threads < 0)
2096 usage(pack_usage);
2097 #ifndef THREADED_DELTA_SEARCH
2098 if (delta_search_threads != 1)
2099 warning("no threads support, "
2100 "ignoring %s", arg);
2101 #endif
2102 continue;
2103 }
2104 if (!prefixcmp(arg, "--depth=")) {
2105 char *end;
2106 depth = strtoul(arg+8, &end, 0);
2107 if (!arg[8] || *end)
2108 usage(pack_usage);
2109 continue;
2110 }
2111 if (!strcmp("--progress", arg)) {
2112 progress = 1;
2113 continue;
2114 }
2115 if (!strcmp("--all-progress", arg)) {
2116 progress = 2;
2117 continue;
2118 }
2119 if (!strcmp("-q", arg)) {
2120 progress = 0;
2121 continue;
2122 }
2123 if (!strcmp("--no-reuse-delta", arg)) {
2124 reuse_delta = 0;
2125 continue;
2126 }
2127 if (!strcmp("--no-reuse-object", arg)) {
2128 reuse_object = reuse_delta = 0;
2129 continue;
2130 }
2131 if (!strcmp("--delta-base-offset", arg)) {
2132 allow_ofs_delta = 1;
2133 continue;
2134 }
2135 if (!strcmp("--stdout", arg)) {
2136 pack_to_stdout = 1;
2137 continue;
2138 }
2139 if (!strcmp("--revs", arg)) {
2140 use_internal_rev_list = 1;
2141 continue;
2142 }
2143 if (!strcmp("--keep-unreachable", arg)) {
2144 keep_unreachable = 1;
2145 continue;
2146 }
2147 if (!strcmp("--unpack-unreachable", arg)) {
2148 unpack_unreachable = 1;
2149 continue;
2150 }
2151 if (!strcmp("--include-tag", arg)) {
2152 include_tag = 1;
2153 continue;
2154 }
2155 if (!strcmp("--unpacked", arg) ||
2156 !prefixcmp(arg, "--unpacked=") ||
2157 !strcmp("--reflog", arg) ||
2158 !strcmp("--all", arg)) {
2159 use_internal_rev_list = 1;
2160 if (rp_ac >= rp_ac_alloc - 1) {
2161 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2162 rp_av = xrealloc(rp_av,
2163 rp_ac_alloc * sizeof(*rp_av));
2164 }
2165 rp_av[rp_ac++] = arg;
2166 continue;
2167 }
2168 if (!strcmp("--thin", arg)) {
2169 use_internal_rev_list = 1;
2170 thin = 1;
2171 rp_av[1] = "--objects-edge";
2172 continue;
2173 }
2174 if (!prefixcmp(arg, "--index-version=")) {
2175 char *c;
2176 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2177 if (pack_idx_default_version > 2)
2178 die("bad %s", arg);
2179 if (*c == ',')
2180 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2181 if (*c || pack_idx_off32_limit & 0x80000000)
2182 die("bad %s", arg);
2183 continue;
2184 }
2185 usage(pack_usage);
2186 }
2188 /* Traditionally "pack-objects [options] base extra" failed;
2189 * we would however want to take refs parameter that would
2190 * have been given to upstream rev-list ourselves, which means
2191 * we somehow want to say what the base name is. So the
2192 * syntax would be:
2193 *
2194 * pack-objects [options] base <refs...>
2195 *
2196 * in other words, we would treat the first non-option as the
2197 * base_name and send everything else to the internal revision
2198 * walker.
2199 */
2201 if (!pack_to_stdout)
2202 base_name = argv[i++];
2204 if (pack_to_stdout != !base_name)
2205 usage(pack_usage);
2207 if (!pack_to_stdout && !pack_size_limit)
2208 pack_size_limit = pack_size_limit_cfg;
2210 if (pack_to_stdout && pack_size_limit)
2211 die("--max-pack-size cannot be used to build a pack for transfer.");
2213 if (!pack_to_stdout && thin)
2214 die("--thin cannot be used to build an indexable pack.");
2216 if (keep_unreachable && unpack_unreachable)
2217 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2219 #ifdef THREADED_DELTA_SEARCH
2220 if (!delta_search_threads) /* --threads=0 means autodetect */
2221 delta_search_threads = online_cpus();
2222 #endif
2224 prepare_packed_git();
2226 if (progress)
2227 progress_state = start_progress("Counting objects", 0);
2228 if (!use_internal_rev_list)
2229 read_object_list_from_stdin();
2230 else {
2231 rp_av[rp_ac] = NULL;
2232 get_object_list(rp_ac, rp_av);
2233 }
2234 if (include_tag && nr_result)
2235 for_each_ref(add_ref_tag, NULL);
2236 stop_progress(&progress_state);
2238 if (non_empty && !nr_result)
2239 return 0;
2240 if (nr_result)
2241 prepare_pack(window, depth);
2242 write_pack_file();
2243 if (progress)
2244 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2245 written, written_delta, reused, reused_delta);
2246 return 0;
2247 }