]> git.ipfire.org Git - thirdparty/git.git/blob - builtin-pack-objects.c
basic threaded delta search
[thirdparty/git.git] / builtin-pack-objects.c
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 "csum-file.h"
12 #include "tree-walk.h"
13 #include "diff.h"
14 #include "revision.h"
15 #include "list-objects.h"
16 #include "progress.h"
17
18 #ifdef THREADED_DELTA_SEARCH
19 #include <pthread.h>
20 #endif
21
22 static const char pack_usage[] = "\
23 git-pack-objects [{ -q | --progress | --all-progress }] \n\
24 [--max-pack-size=N] [--local] [--incremental] \n\
25 [--window=N] [--window-memory=N] [--depth=N] \n\
26 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
27 [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
28 [--stdout | base-name] [<ref-list | <object-list]";
29
30 struct object_entry {
31 struct pack_idx_entry idx;
32 unsigned long size; /* uncompressed size */
33 struct packed_git *in_pack; /* already in pack */
34 off_t in_pack_offset;
35 struct object_entry *delta; /* delta base object */
36 struct object_entry *delta_child; /* deltified objects who bases me */
37 struct object_entry *delta_sibling; /* other deltified objects who
38 * uses the same base as me
39 */
40 void *delta_data; /* cached delta (uncompressed) */
41 unsigned long delta_size; /* delta data size (uncompressed) */
42 unsigned int hash; /* name hint hash */
43 enum object_type type;
44 enum object_type in_pack_type; /* could be delta */
45 unsigned char in_pack_header_size;
46 unsigned char preferred_base; /* we do not pack this, but is available
47 * to be used as the base object to delta
48 * objects against.
49 */
50 unsigned char no_try_delta;
51 };
52
53 /*
54 * Objects we are going to pack are collected in objects array (dynamically
55 * expanded). nr_objects & nr_alloc controls this array. They are stored
56 * in the order we see -- typically rev-list --objects order that gives us
57 * nice "minimum seek" order.
58 */
59 static struct object_entry *objects;
60 static struct object_entry **written_list;
61 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
62
63 static int non_empty;
64 static int no_reuse_delta, no_reuse_object;
65 static int local;
66 static int incremental;
67 static int allow_ofs_delta;
68 static const char *pack_tmp_name, *idx_tmp_name;
69 static char tmpname[PATH_MAX];
70 static const char *base_name;
71 static int progress = 1;
72 static int window = 10;
73 static uint32_t pack_size_limit;
74 static int depth = 50;
75 static int pack_to_stdout;
76 static int num_preferred_base;
77 static struct progress progress_state;
78 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
79 static int pack_compression_seen;
80
81 static unsigned long delta_cache_size = 0;
82 static unsigned long max_delta_cache_size = 0;
83 static unsigned long cache_max_small_delta_size = 1000;
84
85 static unsigned long window_memory_limit = 0;
86
87 /*
88 * The object names in objects array are hashed with this hashtable,
89 * to help looking up the entry by object name.
90 * This hashtable is built after all the objects are seen.
91 */
92 static int *object_ix;
93 static int object_ix_hashsz;
94
95 /*
96 * Pack index for existing packs give us easy access to the offsets into
97 * corresponding pack file where each object's data starts, but the entries
98 * do not store the size of the compressed representation (uncompressed
99 * size is easily available by examining the pack entry header). It is
100 * also rather expensive to find the sha1 for an object given its offset.
101 *
102 * We build a hashtable of existing packs (pack_revindex), and keep reverse
103 * index here -- pack index file is sorted by object name mapping to offset;
104 * this pack_revindex[].revindex array is a list of offset/index_nr pairs
105 * ordered by offset, so if you know the offset of an object, next offset
106 * is where its packed representation ends and the index_nr can be used to
107 * get the object sha1 from the main index.
108 */
109 struct revindex_entry {
110 off_t offset;
111 unsigned int nr;
112 };
113 struct pack_revindex {
114 struct packed_git *p;
115 struct revindex_entry *revindex;
116 };
117 static struct pack_revindex *pack_revindex;
118 static int pack_revindex_hashsz;
119
120 /*
121 * stats
122 */
123 static uint32_t written, written_delta;
124 static uint32_t reused, reused_delta;
125
126 static int pack_revindex_ix(struct packed_git *p)
127 {
128 unsigned long ui = (unsigned long)p;
129 int i;
130
131 ui = ui ^ (ui >> 16); /* defeat structure alignment */
132 i = (int)(ui % pack_revindex_hashsz);
133 while (pack_revindex[i].p) {
134 if (pack_revindex[i].p == p)
135 return i;
136 if (++i == pack_revindex_hashsz)
137 i = 0;
138 }
139 return -1 - i;
140 }
141
142 static void prepare_pack_ix(void)
143 {
144 int num;
145 struct packed_git *p;
146 for (num = 0, p = packed_git; p; p = p->next)
147 num++;
148 if (!num)
149 return;
150 pack_revindex_hashsz = num * 11;
151 pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
152 for (p = packed_git; p; p = p->next) {
153 num = pack_revindex_ix(p);
154 num = - 1 - num;
155 pack_revindex[num].p = p;
156 }
157 /* revindex elements are lazily initialized */
158 }
159
160 static int cmp_offset(const void *a_, const void *b_)
161 {
162 const struct revindex_entry *a = a_;
163 const struct revindex_entry *b = b_;
164 return (a->offset < b->offset) ? -1 : (a->offset > b->offset) ? 1 : 0;
165 }
166
167 /*
168 * Ordered list of offsets of objects in the pack.
169 */
170 static void prepare_pack_revindex(struct pack_revindex *rix)
171 {
172 struct packed_git *p = rix->p;
173 int num_ent = p->num_objects;
174 int i;
175 const char *index = p->index_data;
176
177 rix->revindex = xmalloc(sizeof(*rix->revindex) * (num_ent + 1));
178 index += 4 * 256;
179
180 if (p->index_version > 1) {
181 const uint32_t *off_32 =
182 (uint32_t *)(index + 8 + p->num_objects * (20 + 4));
183 const uint32_t *off_64 = off_32 + p->num_objects;
184 for (i = 0; i < num_ent; i++) {
185 uint32_t off = ntohl(*off_32++);
186 if (!(off & 0x80000000)) {
187 rix->revindex[i].offset = off;
188 } else {
189 rix->revindex[i].offset =
190 ((uint64_t)ntohl(*off_64++)) << 32;
191 rix->revindex[i].offset |=
192 ntohl(*off_64++);
193 }
194 rix->revindex[i].nr = i;
195 }
196 } else {
197 for (i = 0; i < num_ent; i++) {
198 uint32_t hl = *((uint32_t *)(index + 24 * i));
199 rix->revindex[i].offset = ntohl(hl);
200 rix->revindex[i].nr = i;
201 }
202 }
203
204 /* This knows the pack format -- the 20-byte trailer
205 * follows immediately after the last object data.
206 */
207 rix->revindex[num_ent].offset = p->pack_size - 20;
208 rix->revindex[num_ent].nr = -1;
209 qsort(rix->revindex, num_ent, sizeof(*rix->revindex), cmp_offset);
210 }
211
212 static struct revindex_entry * find_packed_object(struct packed_git *p,
213 off_t ofs)
214 {
215 int num;
216 int lo, hi;
217 struct pack_revindex *rix;
218 struct revindex_entry *revindex;
219 num = pack_revindex_ix(p);
220 if (num < 0)
221 die("internal error: pack revindex uninitialized");
222 rix = &pack_revindex[num];
223 if (!rix->revindex)
224 prepare_pack_revindex(rix);
225 revindex = rix->revindex;
226 lo = 0;
227 hi = p->num_objects + 1;
228 do {
229 int mi = (lo + hi) / 2;
230 if (revindex[mi].offset == ofs) {
231 return revindex + mi;
232 }
233 else if (ofs < revindex[mi].offset)
234 hi = mi;
235 else
236 lo = mi + 1;
237 } while (lo < hi);
238 die("internal error: pack revindex corrupt");
239 }
240
241 static const unsigned char *find_packed_object_name(struct packed_git *p,
242 off_t ofs)
243 {
244 struct revindex_entry *entry = find_packed_object(p, ofs);
245 return nth_packed_object_sha1(p, entry->nr);
246 }
247
248 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
249 {
250 unsigned long othersize, delta_size;
251 enum object_type type;
252 void *otherbuf = read_sha1_file(entry->delta->idx.sha1, &type, &othersize);
253 void *delta_buf;
254
255 if (!otherbuf)
256 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
257 delta_buf = diff_delta(otherbuf, othersize,
258 buf, size, &delta_size, 0);
259 if (!delta_buf || delta_size != entry->delta_size)
260 die("delta size changed");
261 free(buf);
262 free(otherbuf);
263 return delta_buf;
264 }
265
266 /*
267 * The per-object header is a pretty dense thing, which is
268 * - first byte: low four bits are "size", then three bits of "type",
269 * and the high bit is "size continues".
270 * - each byte afterwards: low seven bits are size continuation,
271 * with the high bit being "size continues"
272 */
273 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
274 {
275 int n = 1;
276 unsigned char c;
277
278 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
279 die("bad type %d", type);
280
281 c = (type << 4) | (size & 15);
282 size >>= 4;
283 while (size) {
284 *hdr++ = c | 0x80;
285 c = size & 0x7f;
286 size >>= 7;
287 n++;
288 }
289 *hdr = c;
290 return n;
291 }
292
293 /*
294 * we are going to reuse the existing object data as is. make
295 * sure it is not corrupt.
296 */
297 static int check_pack_inflate(struct packed_git *p,
298 struct pack_window **w_curs,
299 off_t offset,
300 off_t len,
301 unsigned long expect)
302 {
303 z_stream stream;
304 unsigned char fakebuf[4096], *in;
305 int st;
306
307 memset(&stream, 0, sizeof(stream));
308 inflateInit(&stream);
309 do {
310 in = use_pack(p, w_curs, offset, &stream.avail_in);
311 stream.next_in = in;
312 stream.next_out = fakebuf;
313 stream.avail_out = sizeof(fakebuf);
314 st = inflate(&stream, Z_FINISH);
315 offset += stream.next_in - in;
316 } while (st == Z_OK || st == Z_BUF_ERROR);
317 inflateEnd(&stream);
318 return (st == Z_STREAM_END &&
319 stream.total_out == expect &&
320 stream.total_in == len) ? 0 : -1;
321 }
322
323 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
324 off_t offset, off_t len, unsigned int nr)
325 {
326 const uint32_t *index_crc;
327 uint32_t data_crc = crc32(0, Z_NULL, 0);
328
329 do {
330 unsigned int avail;
331 void *data = use_pack(p, w_curs, offset, &avail);
332 if (avail > len)
333 avail = len;
334 data_crc = crc32(data_crc, data, avail);
335 offset += avail;
336 len -= avail;
337 } while (len);
338
339 index_crc = p->index_data;
340 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
341
342 return data_crc != ntohl(*index_crc);
343 }
344
345 static void copy_pack_data(struct sha1file *f,
346 struct packed_git *p,
347 struct pack_window **w_curs,
348 off_t offset,
349 off_t len)
350 {
351 unsigned char *in;
352 unsigned int avail;
353
354 while (len) {
355 in = use_pack(p, w_curs, offset, &avail);
356 if (avail > len)
357 avail = (unsigned int)len;
358 sha1write(f, in, avail);
359 offset += avail;
360 len -= avail;
361 }
362 }
363
364 static unsigned long write_object(struct sha1file *f,
365 struct object_entry *entry,
366 off_t write_offset)
367 {
368 unsigned long size;
369 enum object_type type;
370 void *buf;
371 unsigned char header[10];
372 unsigned char dheader[10];
373 unsigned hdrlen;
374 off_t datalen;
375 enum object_type obj_type;
376 int to_reuse = 0;
377 /* write limit if limited packsize and not first object */
378 unsigned long limit = pack_size_limit && nr_written ?
379 pack_size_limit - write_offset : 0;
380 /* no if no delta */
381 int usable_delta = !entry->delta ? 0 :
382 /* yes if unlimited packfile */
383 !pack_size_limit ? 1 :
384 /* no if base written to previous pack */
385 entry->delta->idx.offset == (off_t)-1 ? 0 :
386 /* otherwise double-check written to this
387 * pack, like we do below
388 */
389 entry->delta->idx.offset ? 1 : 0;
390
391 if (!pack_to_stdout)
392 crc32_begin(f);
393
394 obj_type = entry->type;
395 if (no_reuse_object)
396 to_reuse = 0; /* explicit */
397 else if (!entry->in_pack)
398 to_reuse = 0; /* can't reuse what we don't have */
399 else if (obj_type == OBJ_REF_DELTA || obj_type == OBJ_OFS_DELTA)
400 /* check_object() decided it for us ... */
401 to_reuse = usable_delta;
402 /* ... but pack split may override that */
403 else if (obj_type != entry->in_pack_type)
404 to_reuse = 0; /* pack has delta which is unusable */
405 else if (entry->delta)
406 to_reuse = 0; /* we want to pack afresh */
407 else
408 to_reuse = 1; /* we have it in-pack undeltified,
409 * and we do not need to deltify it.
410 */
411
412 if (!to_reuse) {
413 z_stream stream;
414 unsigned long maxsize;
415 void *out;
416 if (!usable_delta) {
417 buf = read_sha1_file(entry->idx.sha1, &obj_type, &size);
418 if (!buf)
419 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
420 } else if (entry->delta_data) {
421 size = entry->delta_size;
422 buf = entry->delta_data;
423 entry->delta_data = NULL;
424 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
425 OBJ_OFS_DELTA : OBJ_REF_DELTA;
426 } else {
427 buf = read_sha1_file(entry->idx.sha1, &type, &size);
428 if (!buf)
429 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
430 buf = delta_against(buf, size, entry);
431 size = entry->delta_size;
432 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
433 OBJ_OFS_DELTA : OBJ_REF_DELTA;
434 }
435 /* compress the data to store and put compressed length in datalen */
436 memset(&stream, 0, sizeof(stream));
437 deflateInit(&stream, pack_compression_level);
438 maxsize = deflateBound(&stream, size);
439 out = xmalloc(maxsize);
440 /* Compress it */
441 stream.next_in = buf;
442 stream.avail_in = size;
443 stream.next_out = out;
444 stream.avail_out = maxsize;
445 while (deflate(&stream, Z_FINISH) == Z_OK)
446 /* nothing */;
447 deflateEnd(&stream);
448 datalen = stream.total_out;
449 deflateEnd(&stream);
450 /*
451 * The object header is a byte of 'type' followed by zero or
452 * more bytes of length.
453 */
454 hdrlen = encode_header(obj_type, size, header);
455
456 if (obj_type == OBJ_OFS_DELTA) {
457 /*
458 * Deltas with relative base contain an additional
459 * encoding of the relative offset for the delta
460 * base from this object's position in the pack.
461 */
462 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
463 unsigned pos = sizeof(dheader) - 1;
464 dheader[pos] = ofs & 127;
465 while (ofs >>= 7)
466 dheader[--pos] = 128 | (--ofs & 127);
467 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
468 free(out);
469 free(buf);
470 return 0;
471 }
472 sha1write(f, header, hdrlen);
473 sha1write(f, dheader + pos, sizeof(dheader) - pos);
474 hdrlen += sizeof(dheader) - pos;
475 } else if (obj_type == OBJ_REF_DELTA) {
476 /*
477 * Deltas with a base reference contain
478 * an additional 20 bytes for the base sha1.
479 */
480 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
481 free(out);
482 free(buf);
483 return 0;
484 }
485 sha1write(f, header, hdrlen);
486 sha1write(f, entry->delta->idx.sha1, 20);
487 hdrlen += 20;
488 } else {
489 if (limit && hdrlen + datalen + 20 >= limit) {
490 free(out);
491 free(buf);
492 return 0;
493 }
494 sha1write(f, header, hdrlen);
495 }
496 sha1write(f, out, datalen);
497 free(out);
498 free(buf);
499 }
500 else {
501 struct packed_git *p = entry->in_pack;
502 struct pack_window *w_curs = NULL;
503 struct revindex_entry *revidx;
504 off_t offset;
505
506 if (entry->delta) {
507 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
508 OBJ_OFS_DELTA : OBJ_REF_DELTA;
509 reused_delta++;
510 }
511 hdrlen = encode_header(obj_type, entry->size, header);
512 offset = entry->in_pack_offset;
513 revidx = find_packed_object(p, offset);
514 datalen = revidx[1].offset - offset;
515 if (!pack_to_stdout && p->index_version > 1 &&
516 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
517 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
518 offset += entry->in_pack_header_size;
519 datalen -= entry->in_pack_header_size;
520 if (obj_type == OBJ_OFS_DELTA) {
521 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
522 unsigned pos = sizeof(dheader) - 1;
523 dheader[pos] = ofs & 127;
524 while (ofs >>= 7)
525 dheader[--pos] = 128 | (--ofs & 127);
526 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
527 return 0;
528 sha1write(f, header, hdrlen);
529 sha1write(f, dheader + pos, sizeof(dheader) - pos);
530 hdrlen += sizeof(dheader) - pos;
531 } else if (obj_type == OBJ_REF_DELTA) {
532 if (limit && hdrlen + 20 + datalen + 20 >= limit)
533 return 0;
534 sha1write(f, header, hdrlen);
535 sha1write(f, entry->delta->idx.sha1, 20);
536 hdrlen += 20;
537 } else {
538 if (limit && hdrlen + datalen + 20 >= limit)
539 return 0;
540 sha1write(f, header, hdrlen);
541 }
542
543 if (!pack_to_stdout && p->index_version == 1 &&
544 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
545 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
546 copy_pack_data(f, p, &w_curs, offset, datalen);
547 unuse_pack(&w_curs);
548 reused++;
549 }
550 if (usable_delta)
551 written_delta++;
552 written++;
553 if (!pack_to_stdout)
554 entry->idx.crc32 = crc32_end(f);
555 return hdrlen + datalen;
556 }
557
558 static off_t write_one(struct sha1file *f,
559 struct object_entry *e,
560 off_t offset)
561 {
562 unsigned long size;
563
564 /* offset is non zero if object is written already. */
565 if (e->idx.offset || e->preferred_base)
566 return offset;
567
568 /* if we are deltified, write out base object first. */
569 if (e->delta) {
570 offset = write_one(f, e->delta, offset);
571 if (!offset)
572 return 0;
573 }
574
575 e->idx.offset = offset;
576 size = write_object(f, e, offset);
577 if (!size) {
578 e->idx.offset = 0;
579 return 0;
580 }
581 written_list[nr_written++] = e;
582
583 /* make sure off_t is sufficiently large not to wrap */
584 if (offset > offset + size)
585 die("pack too large for current definition of off_t");
586 return offset + size;
587 }
588
589 static int open_object_dir_tmp(const char *path)
590 {
591 snprintf(tmpname, sizeof(tmpname), "%s/%s", get_object_directory(), path);
592 return xmkstemp(tmpname);
593 }
594
595 /* forward declaration for write_pack_file */
596 static int adjust_perm(const char *path, mode_t mode);
597
598 static void write_pack_file(void)
599 {
600 uint32_t i = 0, j;
601 struct sha1file *f;
602 off_t offset, offset_one, last_obj_offset = 0;
603 struct pack_header hdr;
604 int do_progress = progress >> pack_to_stdout;
605 uint32_t nr_remaining = nr_result;
606
607 if (do_progress)
608 start_progress(&progress_state, "Writing %u objects...", "", nr_result);
609 written_list = xmalloc(nr_objects * sizeof(struct object_entry *));
610
611 do {
612 unsigned char sha1[20];
613
614 if (pack_to_stdout) {
615 f = sha1fd(1, "<stdout>");
616 } else {
617 int fd = open_object_dir_tmp("tmp_pack_XXXXXX");
618 pack_tmp_name = xstrdup(tmpname);
619 f = sha1fd(fd, pack_tmp_name);
620 }
621
622 hdr.hdr_signature = htonl(PACK_SIGNATURE);
623 hdr.hdr_version = htonl(PACK_VERSION);
624 hdr.hdr_entries = htonl(nr_remaining);
625 sha1write(f, &hdr, sizeof(hdr));
626 offset = sizeof(hdr);
627 nr_written = 0;
628 for (; i < nr_objects; i++) {
629 last_obj_offset = offset;
630 offset_one = write_one(f, objects + i, offset);
631 if (!offset_one)
632 break;
633 offset = offset_one;
634 if (do_progress)
635 display_progress(&progress_state, written);
636 }
637
638 /*
639 * Did we write the wrong # entries in the header?
640 * If so, rewrite it like in fast-import
641 */
642 if (pack_to_stdout || nr_written == nr_remaining) {
643 sha1close(f, sha1, 1);
644 } else {
645 sha1close(f, sha1, 0);
646 fixup_pack_header_footer(f->fd, sha1, pack_tmp_name, nr_written);
647 close(f->fd);
648 }
649
650 if (!pack_to_stdout) {
651 mode_t mode = umask(0);
652
653 umask(mode);
654 mode = 0444 & ~mode;
655
656 idx_tmp_name = write_idx_file(NULL,
657 (struct pack_idx_entry **) written_list, nr_written, sha1);
658 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
659 base_name, sha1_to_hex(sha1));
660 if (adjust_perm(pack_tmp_name, mode))
661 die("unable to make temporary pack file readable: %s",
662 strerror(errno));
663 if (rename(pack_tmp_name, tmpname))
664 die("unable to rename temporary pack file: %s",
665 strerror(errno));
666 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
667 base_name, sha1_to_hex(sha1));
668 if (adjust_perm(idx_tmp_name, mode))
669 die("unable to make temporary index file readable: %s",
670 strerror(errno));
671 if (rename(idx_tmp_name, tmpname))
672 die("unable to rename temporary index file: %s",
673 strerror(errno));
674 puts(sha1_to_hex(sha1));
675 }
676
677 /* mark written objects as written to previous pack */
678 for (j = 0; j < nr_written; j++) {
679 written_list[j]->idx.offset = (off_t)-1;
680 }
681 nr_remaining -= nr_written;
682 } while (nr_remaining && i < nr_objects);
683
684 free(written_list);
685 if (do_progress)
686 stop_progress(&progress_state);
687 if (written != nr_result)
688 die("wrote %u objects while expecting %u", written, nr_result);
689 /*
690 * We have scanned through [0 ... i). Since we have written
691 * the correct number of objects, the remaining [i ... nr_objects)
692 * items must be either already written (due to out-of-order delta base)
693 * or a preferred base. Count those which are neither and complain if any.
694 */
695 for (j = 0; i < nr_objects; i++) {
696 struct object_entry *e = objects + i;
697 j += !e->idx.offset && !e->preferred_base;
698 }
699 if (j)
700 die("wrote %u objects as expected but %u unwritten", written, j);
701 }
702
703 static int locate_object_entry_hash(const unsigned char *sha1)
704 {
705 int i;
706 unsigned int ui;
707 memcpy(&ui, sha1, sizeof(unsigned int));
708 i = ui % object_ix_hashsz;
709 while (0 < object_ix[i]) {
710 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
711 return i;
712 if (++i == object_ix_hashsz)
713 i = 0;
714 }
715 return -1 - i;
716 }
717
718 static struct object_entry *locate_object_entry(const unsigned char *sha1)
719 {
720 int i;
721
722 if (!object_ix_hashsz)
723 return NULL;
724
725 i = locate_object_entry_hash(sha1);
726 if (0 <= i)
727 return &objects[object_ix[i]-1];
728 return NULL;
729 }
730
731 static void rehash_objects(void)
732 {
733 uint32_t i;
734 struct object_entry *oe;
735
736 object_ix_hashsz = nr_objects * 3;
737 if (object_ix_hashsz < 1024)
738 object_ix_hashsz = 1024;
739 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
740 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
741 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
742 int ix = locate_object_entry_hash(oe->idx.sha1);
743 if (0 <= ix)
744 continue;
745 ix = -1 - ix;
746 object_ix[ix] = i + 1;
747 }
748 }
749
750 static unsigned name_hash(const char *name)
751 {
752 unsigned char c;
753 unsigned hash = 0;
754
755 if (!name)
756 return 0;
757
758 /*
759 * This effectively just creates a sortable number from the
760 * last sixteen non-whitespace characters. Last characters
761 * count "most", so things that end in ".c" sort together.
762 */
763 while ((c = *name++) != 0) {
764 if (isspace(c))
765 continue;
766 hash = (hash >> 2) + (c << 24);
767 }
768 return hash;
769 }
770
771 static void setup_delta_attr_check(struct git_attr_check *check)
772 {
773 static struct git_attr *attr_delta;
774
775 if (!attr_delta)
776 attr_delta = git_attr("delta", 5);
777
778 check[0].attr = attr_delta;
779 }
780
781 static int no_try_delta(const char *path)
782 {
783 struct git_attr_check check[1];
784
785 setup_delta_attr_check(check);
786 if (git_checkattr(path, ARRAY_SIZE(check), check))
787 return 0;
788 if (ATTR_FALSE(check->value))
789 return 1;
790 return 0;
791 }
792
793 static int add_object_entry(const unsigned char *sha1, enum object_type type,
794 const char *name, int exclude)
795 {
796 struct object_entry *entry;
797 struct packed_git *p, *found_pack = NULL;
798 off_t found_offset = 0;
799 int ix;
800 unsigned hash = name_hash(name);
801
802 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
803 if (ix >= 0) {
804 if (exclude) {
805 entry = objects + object_ix[ix] - 1;
806 if (!entry->preferred_base)
807 nr_result--;
808 entry->preferred_base = 1;
809 }
810 return 0;
811 }
812
813 for (p = packed_git; p; p = p->next) {
814 off_t offset = find_pack_entry_one(sha1, p);
815 if (offset) {
816 if (!found_pack) {
817 found_offset = offset;
818 found_pack = p;
819 }
820 if (exclude)
821 break;
822 if (incremental)
823 return 0;
824 if (local && !p->pack_local)
825 return 0;
826 }
827 }
828
829 if (nr_objects >= nr_alloc) {
830 nr_alloc = (nr_alloc + 1024) * 3 / 2;
831 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
832 }
833
834 entry = objects + nr_objects++;
835 memset(entry, 0, sizeof(*entry));
836 hashcpy(entry->idx.sha1, sha1);
837 entry->hash = hash;
838 if (type)
839 entry->type = type;
840 if (exclude)
841 entry->preferred_base = 1;
842 else
843 nr_result++;
844 if (found_pack) {
845 entry->in_pack = found_pack;
846 entry->in_pack_offset = found_offset;
847 }
848
849 if (object_ix_hashsz * 3 <= nr_objects * 4)
850 rehash_objects();
851 else
852 object_ix[-1 - ix] = nr_objects;
853
854 if (progress)
855 display_progress(&progress_state, nr_objects);
856
857 if (name && no_try_delta(name))
858 entry->no_try_delta = 1;
859
860 return 1;
861 }
862
863 struct pbase_tree_cache {
864 unsigned char sha1[20];
865 int ref;
866 int temporary;
867 void *tree_data;
868 unsigned long tree_size;
869 };
870
871 static struct pbase_tree_cache *(pbase_tree_cache[256]);
872 static int pbase_tree_cache_ix(const unsigned char *sha1)
873 {
874 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
875 }
876 static int pbase_tree_cache_ix_incr(int ix)
877 {
878 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
879 }
880
881 static struct pbase_tree {
882 struct pbase_tree *next;
883 /* This is a phony "cache" entry; we are not
884 * going to evict it nor find it through _get()
885 * mechanism -- this is for the toplevel node that
886 * would almost always change with any commit.
887 */
888 struct pbase_tree_cache pcache;
889 } *pbase_tree;
890
891 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
892 {
893 struct pbase_tree_cache *ent, *nent;
894 void *data;
895 unsigned long size;
896 enum object_type type;
897 int neigh;
898 int my_ix = pbase_tree_cache_ix(sha1);
899 int available_ix = -1;
900
901 /* pbase-tree-cache acts as a limited hashtable.
902 * your object will be found at your index or within a few
903 * slots after that slot if it is cached.
904 */
905 for (neigh = 0; neigh < 8; neigh++) {
906 ent = pbase_tree_cache[my_ix];
907 if (ent && !hashcmp(ent->sha1, sha1)) {
908 ent->ref++;
909 return ent;
910 }
911 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
912 ((0 <= available_ix) &&
913 (!ent && pbase_tree_cache[available_ix])))
914 available_ix = my_ix;
915 if (!ent)
916 break;
917 my_ix = pbase_tree_cache_ix_incr(my_ix);
918 }
919
920 /* Did not find one. Either we got a bogus request or
921 * we need to read and perhaps cache.
922 */
923 data = read_sha1_file(sha1, &type, &size);
924 if (!data)
925 return NULL;
926 if (type != OBJ_TREE) {
927 free(data);
928 return NULL;
929 }
930
931 /* We need to either cache or return a throwaway copy */
932
933 if (available_ix < 0)
934 ent = NULL;
935 else {
936 ent = pbase_tree_cache[available_ix];
937 my_ix = available_ix;
938 }
939
940 if (!ent) {
941 nent = xmalloc(sizeof(*nent));
942 nent->temporary = (available_ix < 0);
943 }
944 else {
945 /* evict and reuse */
946 free(ent->tree_data);
947 nent = ent;
948 }
949 hashcpy(nent->sha1, sha1);
950 nent->tree_data = data;
951 nent->tree_size = size;
952 nent->ref = 1;
953 if (!nent->temporary)
954 pbase_tree_cache[my_ix] = nent;
955 return nent;
956 }
957
958 static void pbase_tree_put(struct pbase_tree_cache *cache)
959 {
960 if (!cache->temporary) {
961 cache->ref--;
962 return;
963 }
964 free(cache->tree_data);
965 free(cache);
966 }
967
968 static int name_cmp_len(const char *name)
969 {
970 int i;
971 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
972 ;
973 return i;
974 }
975
976 static void add_pbase_object(struct tree_desc *tree,
977 const char *name,
978 int cmplen,
979 const char *fullname)
980 {
981 struct name_entry entry;
982 int cmp;
983
984 while (tree_entry(tree,&entry)) {
985 if (S_ISGITLINK(entry.mode))
986 continue;
987 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
988 memcmp(name, entry.path, cmplen);
989 if (cmp > 0)
990 continue;
991 if (cmp < 0)
992 return;
993 if (name[cmplen] != '/') {
994 add_object_entry(entry.sha1,
995 S_ISDIR(entry.mode) ? OBJ_TREE : OBJ_BLOB,
996 fullname, 1);
997 return;
998 }
999 if (S_ISDIR(entry.mode)) {
1000 struct tree_desc sub;
1001 struct pbase_tree_cache *tree;
1002 const char *down = name+cmplen+1;
1003 int downlen = name_cmp_len(down);
1004
1005 tree = pbase_tree_get(entry.sha1);
1006 if (!tree)
1007 return;
1008 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1009
1010 add_pbase_object(&sub, down, downlen, fullname);
1011 pbase_tree_put(tree);
1012 }
1013 }
1014 }
1015
1016 static unsigned *done_pbase_paths;
1017 static int done_pbase_paths_num;
1018 static int done_pbase_paths_alloc;
1019 static int done_pbase_path_pos(unsigned hash)
1020 {
1021 int lo = 0;
1022 int hi = done_pbase_paths_num;
1023 while (lo < hi) {
1024 int mi = (hi + lo) / 2;
1025 if (done_pbase_paths[mi] == hash)
1026 return mi;
1027 if (done_pbase_paths[mi] < hash)
1028 hi = mi;
1029 else
1030 lo = mi + 1;
1031 }
1032 return -lo-1;
1033 }
1034
1035 static int check_pbase_path(unsigned hash)
1036 {
1037 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1038 if (0 <= pos)
1039 return 1;
1040 pos = -pos - 1;
1041 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1042 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1043 done_pbase_paths = xrealloc(done_pbase_paths,
1044 done_pbase_paths_alloc *
1045 sizeof(unsigned));
1046 }
1047 done_pbase_paths_num++;
1048 if (pos < done_pbase_paths_num)
1049 memmove(done_pbase_paths + pos + 1,
1050 done_pbase_paths + pos,
1051 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1052 done_pbase_paths[pos] = hash;
1053 return 0;
1054 }
1055
1056 static void add_preferred_base_object(const char *name)
1057 {
1058 struct pbase_tree *it;
1059 int cmplen;
1060 unsigned hash = name_hash(name);
1061
1062 if (!num_preferred_base || check_pbase_path(hash))
1063 return;
1064
1065 cmplen = name_cmp_len(name);
1066 for (it = pbase_tree; it; it = it->next) {
1067 if (cmplen == 0) {
1068 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1069 }
1070 else {
1071 struct tree_desc tree;
1072 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1073 add_pbase_object(&tree, name, cmplen, name);
1074 }
1075 }
1076 }
1077
1078 static void add_preferred_base(unsigned char *sha1)
1079 {
1080 struct pbase_tree *it;
1081 void *data;
1082 unsigned long size;
1083 unsigned char tree_sha1[20];
1084
1085 if (window <= num_preferred_base++)
1086 return;
1087
1088 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1089 if (!data)
1090 return;
1091
1092 for (it = pbase_tree; it; it = it->next) {
1093 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1094 free(data);
1095 return;
1096 }
1097 }
1098
1099 it = xcalloc(1, sizeof(*it));
1100 it->next = pbase_tree;
1101 pbase_tree = it;
1102
1103 hashcpy(it->pcache.sha1, tree_sha1);
1104 it->pcache.tree_data = data;
1105 it->pcache.tree_size = size;
1106 }
1107
1108 static void check_object(struct object_entry *entry)
1109 {
1110 if (entry->in_pack) {
1111 struct packed_git *p = entry->in_pack;
1112 struct pack_window *w_curs = NULL;
1113 const unsigned char *base_ref = NULL;
1114 struct object_entry *base_entry;
1115 unsigned long used, used_0;
1116 unsigned int avail;
1117 off_t ofs;
1118 unsigned char *buf, c;
1119
1120 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1121
1122 /*
1123 * We want in_pack_type even if we do not reuse delta
1124 * since non-delta representations could still be reused.
1125 */
1126 used = unpack_object_header_gently(buf, avail,
1127 &entry->in_pack_type,
1128 &entry->size);
1129
1130 /*
1131 * Determine if this is a delta and if so whether we can
1132 * reuse it or not. Otherwise let's find out as cheaply as
1133 * possible what the actual type and size for this object is.
1134 */
1135 switch (entry->in_pack_type) {
1136 default:
1137 /* Not a delta hence we've already got all we need. */
1138 entry->type = entry->in_pack_type;
1139 entry->in_pack_header_size = used;
1140 unuse_pack(&w_curs);
1141 return;
1142 case OBJ_REF_DELTA:
1143 if (!no_reuse_delta && !entry->preferred_base)
1144 base_ref = use_pack(p, &w_curs,
1145 entry->in_pack_offset + used, NULL);
1146 entry->in_pack_header_size = used + 20;
1147 break;
1148 case OBJ_OFS_DELTA:
1149 buf = use_pack(p, &w_curs,
1150 entry->in_pack_offset + used, NULL);
1151 used_0 = 0;
1152 c = buf[used_0++];
1153 ofs = c & 127;
1154 while (c & 128) {
1155 ofs += 1;
1156 if (!ofs || MSB(ofs, 7))
1157 die("delta base offset overflow in pack for %s",
1158 sha1_to_hex(entry->idx.sha1));
1159 c = buf[used_0++];
1160 ofs = (ofs << 7) + (c & 127);
1161 }
1162 if (ofs >= entry->in_pack_offset)
1163 die("delta base offset out of bound for %s",
1164 sha1_to_hex(entry->idx.sha1));
1165 ofs = entry->in_pack_offset - ofs;
1166 if (!no_reuse_delta && !entry->preferred_base)
1167 base_ref = find_packed_object_name(p, ofs);
1168 entry->in_pack_header_size = used + used_0;
1169 break;
1170 }
1171
1172 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1173 /*
1174 * If base_ref was set above that means we wish to
1175 * reuse delta data, and we even found that base
1176 * in the list of objects we want to pack. Goodie!
1177 *
1178 * Depth value does not matter - find_deltas() will
1179 * never consider reused delta as the base object to
1180 * deltify other objects against, in order to avoid
1181 * circular deltas.
1182 */
1183 entry->type = entry->in_pack_type;
1184 entry->delta = base_entry;
1185 entry->delta_sibling = base_entry->delta_child;
1186 base_entry->delta_child = entry;
1187 unuse_pack(&w_curs);
1188 return;
1189 }
1190
1191 if (entry->type) {
1192 /*
1193 * This must be a delta and we already know what the
1194 * final object type is. Let's extract the actual
1195 * object size from the delta header.
1196 */
1197 entry->size = get_size_from_delta(p, &w_curs,
1198 entry->in_pack_offset + entry->in_pack_header_size);
1199 unuse_pack(&w_curs);
1200 return;
1201 }
1202
1203 /*
1204 * No choice but to fall back to the recursive delta walk
1205 * with sha1_object_info() to find about the object type
1206 * at this point...
1207 */
1208 unuse_pack(&w_curs);
1209 }
1210
1211 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1212 if (entry->type < 0)
1213 die("unable to get type of object %s",
1214 sha1_to_hex(entry->idx.sha1));
1215 }
1216
1217 static int pack_offset_sort(const void *_a, const void *_b)
1218 {
1219 const struct object_entry *a = *(struct object_entry **)_a;
1220 const struct object_entry *b = *(struct object_entry **)_b;
1221
1222 /* avoid filesystem trashing with loose objects */
1223 if (!a->in_pack && !b->in_pack)
1224 return hashcmp(a->idx.sha1, b->idx.sha1);
1225
1226 if (a->in_pack < b->in_pack)
1227 return -1;
1228 if (a->in_pack > b->in_pack)
1229 return 1;
1230 return a->in_pack_offset < b->in_pack_offset ? -1 :
1231 (a->in_pack_offset > b->in_pack_offset);
1232 }
1233
1234 static void get_object_details(void)
1235 {
1236 uint32_t i;
1237 struct object_entry **sorted_by_offset;
1238
1239 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1240 for (i = 0; i < nr_objects; i++)
1241 sorted_by_offset[i] = objects + i;
1242 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1243
1244 prepare_pack_ix();
1245 for (i = 0; i < nr_objects; i++)
1246 check_object(sorted_by_offset[i]);
1247 free(sorted_by_offset);
1248 }
1249
1250 static int type_size_sort(const void *_a, const void *_b)
1251 {
1252 const struct object_entry *a = *(struct object_entry **)_a;
1253 const struct object_entry *b = *(struct object_entry **)_b;
1254
1255 if (a->type < b->type)
1256 return -1;
1257 if (a->type > b->type)
1258 return 1;
1259 if (a->hash < b->hash)
1260 return -1;
1261 if (a->hash > b->hash)
1262 return 1;
1263 if (a->preferred_base < b->preferred_base)
1264 return -1;
1265 if (a->preferred_base > b->preferred_base)
1266 return 1;
1267 if (a->size < b->size)
1268 return -1;
1269 if (a->size > b->size)
1270 return 1;
1271 return a > b ? -1 : (a < b); /* newest last */
1272 }
1273
1274 struct unpacked {
1275 struct object_entry *entry;
1276 void *data;
1277 struct delta_index *index;
1278 unsigned depth;
1279 };
1280
1281 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1282 unsigned long delta_size)
1283 {
1284 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1285 return 0;
1286
1287 if (delta_size < cache_max_small_delta_size)
1288 return 1;
1289
1290 /* cache delta, if objects are large enough compared to delta size */
1291 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1292 return 1;
1293
1294 return 0;
1295 }
1296
1297 #ifdef THREADED_DELTA_SEARCH
1298
1299 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1300 #define read_lock() pthread_mutex_lock(&read_mutex)
1301 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1302
1303 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1304 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1305 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1306
1307 #else
1308
1309 #define read_lock() 0
1310 #define read_unlock() 0
1311 #define progress_lock() 0
1312 #define progress_unlock() 0
1313
1314 #endif
1315
1316 /*
1317 * We search for deltas _backwards_ in a list sorted by type and
1318 * by size, so that we see progressively smaller and smaller files.
1319 * That's because we prefer deltas to be from the bigger file
1320 * to the smaller - deletes are potentially cheaper, but perhaps
1321 * more importantly, the bigger file is likely the more recent
1322 * one.
1323 */
1324 static int try_delta(struct unpacked *trg, struct unpacked *src,
1325 unsigned max_depth, unsigned long *mem_usage)
1326 {
1327 struct object_entry *trg_entry = trg->entry;
1328 struct object_entry *src_entry = src->entry;
1329 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1330 unsigned ref_depth;
1331 enum object_type type;
1332 void *delta_buf;
1333
1334 /* Don't bother doing diffs between different types */
1335 if (trg_entry->type != src_entry->type)
1336 return -1;
1337
1338 /*
1339 * We do not bother to try a delta that we discarded
1340 * on an earlier try, but only when reusing delta data.
1341 */
1342 if (!no_reuse_delta && trg_entry->in_pack &&
1343 trg_entry->in_pack == src_entry->in_pack &&
1344 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1345 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1346 return 0;
1347
1348 /* Let's not bust the allowed depth. */
1349 if (src->depth >= max_depth)
1350 return 0;
1351
1352 /* Now some size filtering heuristics. */
1353 trg_size = trg_entry->size;
1354 if (!trg_entry->delta) {
1355 max_size = trg_size/2 - 20;
1356 ref_depth = 1;
1357 } else {
1358 max_size = trg_entry->delta_size;
1359 ref_depth = trg->depth;
1360 }
1361 max_size = max_size * (max_depth - src->depth) /
1362 (max_depth - ref_depth + 1);
1363 if (max_size == 0)
1364 return 0;
1365 src_size = src_entry->size;
1366 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1367 if (sizediff >= max_size)
1368 return 0;
1369 if (trg_size < src_size / 32)
1370 return 0;
1371
1372 /* Load data if not already done */
1373 if (!trg->data) {
1374 read_lock();
1375 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1376 read_unlock();
1377 if (!trg->data)
1378 die("object %s cannot be read",
1379 sha1_to_hex(trg_entry->idx.sha1));
1380 if (sz != trg_size)
1381 die("object %s inconsistent object length (%lu vs %lu)",
1382 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1383 *mem_usage += sz;
1384 }
1385 if (!src->data) {
1386 read_lock();
1387 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1388 read_unlock();
1389 if (!src->data)
1390 die("object %s cannot be read",
1391 sha1_to_hex(src_entry->idx.sha1));
1392 if (sz != src_size)
1393 die("object %s inconsistent object length (%lu vs %lu)",
1394 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1395 *mem_usage += sz;
1396 }
1397 if (!src->index) {
1398 src->index = create_delta_index(src->data, src_size);
1399 if (!src->index) {
1400 static int warned = 0;
1401 if (!warned++)
1402 warning("suboptimal pack - out of memory");
1403 return 0;
1404 }
1405 *mem_usage += sizeof_delta_index(src->index);
1406 }
1407
1408 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1409 if (!delta_buf)
1410 return 0;
1411
1412 if (trg_entry->delta) {
1413 /* Prefer only shallower same-sized deltas. */
1414 if (delta_size == trg_entry->delta_size &&
1415 src->depth + 1 >= trg->depth) {
1416 free(delta_buf);
1417 return 0;
1418 }
1419 }
1420
1421 trg_entry->delta = src_entry;
1422 trg_entry->delta_size = delta_size;
1423 trg->depth = src->depth + 1;
1424
1425 if (trg_entry->delta_data) {
1426 delta_cache_size -= trg_entry->delta_size;
1427 free(trg_entry->delta_data);
1428 trg_entry->delta_data = NULL;
1429 }
1430
1431 if (delta_cacheable(src_size, trg_size, delta_size)) {
1432 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1433 delta_cache_size += trg_entry->delta_size;
1434 } else
1435 free(delta_buf);
1436 return 1;
1437 }
1438
1439 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1440 {
1441 struct object_entry *child = me->delta_child;
1442 unsigned int m = n;
1443 while (child) {
1444 unsigned int c = check_delta_limit(child, n + 1);
1445 if (m < c)
1446 m = c;
1447 child = child->delta_sibling;
1448 }
1449 return m;
1450 }
1451
1452 static unsigned long free_unpacked(struct unpacked *n)
1453 {
1454 unsigned long freed_mem = sizeof_delta_index(n->index);
1455 free_delta_index(n->index);
1456 n->index = NULL;
1457 if (n->data) {
1458 freed_mem += n->entry->size;
1459 free(n->data);
1460 n->data = NULL;
1461 }
1462 n->entry = NULL;
1463 n->depth = 0;
1464 return freed_mem;
1465 }
1466
1467 static void find_deltas(struct object_entry **list, unsigned list_size,
1468 int window, int depth, unsigned *processed)
1469 {
1470 uint32_t i = list_size, idx = 0, count = 0;
1471 unsigned int array_size = window * sizeof(struct unpacked);
1472 struct unpacked *array;
1473 unsigned long mem_usage = 0;
1474
1475 array = xmalloc(array_size);
1476 memset(array, 0, array_size);
1477
1478 do {
1479 struct object_entry *entry = list[--i];
1480 struct unpacked *n = array + idx;
1481 int j, max_depth, best_base = -1;
1482
1483 mem_usage -= free_unpacked(n);
1484 n->entry = entry;
1485
1486 while (window_memory_limit &&
1487 mem_usage > window_memory_limit &&
1488 count > 1) {
1489 uint32_t tail = (idx + window - count) % window;
1490 mem_usage -= free_unpacked(array + tail);
1491 count--;
1492 }
1493
1494 /* We do not compute delta to *create* objects we are not
1495 * going to pack.
1496 */
1497 if (entry->preferred_base)
1498 goto next;
1499
1500 progress_lock();
1501 (*processed)++;
1502 if (progress)
1503 display_progress(&progress_state, *processed);
1504 progress_unlock();
1505
1506 /*
1507 * If the current object is at pack edge, take the depth the
1508 * objects that depend on the current object into account
1509 * otherwise they would become too deep.
1510 */
1511 max_depth = depth;
1512 if (entry->delta_child) {
1513 max_depth -= check_delta_limit(entry, 0);
1514 if (max_depth <= 0)
1515 goto next;
1516 }
1517
1518 j = window;
1519 while (--j > 0) {
1520 int ret;
1521 uint32_t other_idx = idx + j;
1522 struct unpacked *m;
1523 if (other_idx >= window)
1524 other_idx -= window;
1525 m = array + other_idx;
1526 if (!m->entry)
1527 break;
1528 ret = try_delta(n, m, max_depth, &mem_usage);
1529 if (ret < 0)
1530 break;
1531 else if (ret > 0)
1532 best_base = other_idx;
1533 }
1534
1535 /* if we made n a delta, and if n is already at max
1536 * depth, leaving it in the window is pointless. we
1537 * should evict it first.
1538 */
1539 if (entry->delta && depth <= n->depth)
1540 continue;
1541
1542 /*
1543 * Move the best delta base up in the window, after the
1544 * currently deltified object, to keep it longer. It will
1545 * be the first base object to be attempted next.
1546 */
1547 if (entry->delta) {
1548 struct unpacked swap = array[best_base];
1549 int dist = (window + idx - best_base) % window;
1550 int dst = best_base;
1551 while (dist--) {
1552 int src = (dst + 1) % window;
1553 array[dst] = array[src];
1554 dst = src;
1555 }
1556 array[dst] = swap;
1557 }
1558
1559 next:
1560 idx++;
1561 if (count + 1 < window)
1562 count++;
1563 if (idx >= window)
1564 idx = 0;
1565 } while (i > 0);
1566
1567 for (i = 0; i < window; ++i) {
1568 free_delta_index(array[i].index);
1569 free(array[i].data);
1570 }
1571 free(array);
1572 }
1573
1574 #ifdef THREADED_DELTA_SEARCH
1575
1576 struct thread_params {
1577 pthread_t thread;
1578 struct object_entry **list;
1579 unsigned list_size;
1580 int window;
1581 int depth;
1582 unsigned *processed;
1583 };
1584
1585 static void *threaded_find_deltas(void *arg)
1586 {
1587 struct thread_params *p = arg;
1588 if (p->list_size)
1589 find_deltas(p->list, p->list_size,
1590 p->window, p->depth, p->processed);
1591 return NULL;
1592 }
1593
1594 #define NR_THREADS 8
1595
1596 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1597 int window, int depth, unsigned *processed)
1598 {
1599 struct thread_params p[NR_THREADS];
1600 int i, ret;
1601
1602 for (i = 0; i < NR_THREADS; i++) {
1603 unsigned sublist_size = list_size / (NR_THREADS - i);
1604 p[i].list = list;
1605 p[i].list_size = sublist_size;
1606 p[i].window = window;
1607 p[i].depth = depth;
1608 p[i].processed = processed;
1609 ret = pthread_create(&p[i].thread, NULL,
1610 threaded_find_deltas, &p[i]);
1611 if (ret)
1612 die("unable to create thread: %s", strerror(ret));
1613 list += sublist_size;
1614 list_size -= sublist_size;
1615 }
1616
1617 for (i = 0; i < NR_THREADS; i++) {
1618 pthread_join(p[i].thread, NULL);
1619 }
1620 }
1621
1622 #else
1623 #define ll_find_deltas find_deltas
1624 #endif
1625
1626 static void prepare_pack(int window, int depth)
1627 {
1628 struct object_entry **delta_list;
1629 uint32_t i, n, nr_deltas;
1630
1631 get_object_details();
1632
1633 if (!nr_objects || !window || !depth)
1634 return;
1635
1636 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1637 nr_deltas = n = 0;
1638
1639 for (i = 0; i < nr_objects; i++) {
1640 struct object_entry *entry = objects + i;
1641
1642 if (entry->delta)
1643 /* This happens if we decided to reuse existing
1644 * delta from a pack. "!no_reuse_delta &&" is implied.
1645 */
1646 continue;
1647
1648 if (entry->size < 50)
1649 continue;
1650
1651 if (entry->no_try_delta)
1652 continue;
1653
1654 if (!entry->preferred_base)
1655 nr_deltas++;
1656
1657 delta_list[n++] = entry;
1658 }
1659
1660 if (nr_deltas) {
1661 unsigned nr_done = 0;
1662 if (progress)
1663 start_progress(&progress_state,
1664 "Deltifying %u objects...", "",
1665 nr_deltas);
1666 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1667 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1668 if (progress)
1669 stop_progress(&progress_state);
1670 if (nr_done != nr_deltas)
1671 die("inconsistency with delta count");
1672 }
1673 free(delta_list);
1674 }
1675
1676 static int git_pack_config(const char *k, const char *v)
1677 {
1678 if(!strcmp(k, "pack.window")) {
1679 window = git_config_int(k, v);
1680 return 0;
1681 }
1682 if (!strcmp(k, "pack.windowmemory")) {
1683 window_memory_limit = git_config_ulong(k, v);
1684 return 0;
1685 }
1686 if (!strcmp(k, "pack.depth")) {
1687 depth = git_config_int(k, v);
1688 return 0;
1689 }
1690 if (!strcmp(k, "pack.compression")) {
1691 int level = git_config_int(k, v);
1692 if (level == -1)
1693 level = Z_DEFAULT_COMPRESSION;
1694 else if (level < 0 || level > Z_BEST_COMPRESSION)
1695 die("bad pack compression level %d", level);
1696 pack_compression_level = level;
1697 pack_compression_seen = 1;
1698 return 0;
1699 }
1700 if (!strcmp(k, "pack.deltacachesize")) {
1701 max_delta_cache_size = git_config_int(k, v);
1702 return 0;
1703 }
1704 if (!strcmp(k, "pack.deltacachelimit")) {
1705 cache_max_small_delta_size = git_config_int(k, v);
1706 return 0;
1707 }
1708 return git_default_config(k, v);
1709 }
1710
1711 static void read_object_list_from_stdin(void)
1712 {
1713 char line[40 + 1 + PATH_MAX + 2];
1714 unsigned char sha1[20];
1715
1716 for (;;) {
1717 if (!fgets(line, sizeof(line), stdin)) {
1718 if (feof(stdin))
1719 break;
1720 if (!ferror(stdin))
1721 die("fgets returned NULL, not EOF, not error!");
1722 if (errno != EINTR)
1723 die("fgets: %s", strerror(errno));
1724 clearerr(stdin);
1725 continue;
1726 }
1727 if (line[0] == '-') {
1728 if (get_sha1_hex(line+1, sha1))
1729 die("expected edge sha1, got garbage:\n %s",
1730 line);
1731 add_preferred_base(sha1);
1732 continue;
1733 }
1734 if (get_sha1_hex(line, sha1))
1735 die("expected sha1, got garbage:\n %s", line);
1736
1737 add_preferred_base_object(line+41);
1738 add_object_entry(sha1, 0, line+41, 0);
1739 }
1740 }
1741
1742 static void show_commit(struct commit *commit)
1743 {
1744 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1745 }
1746
1747 static void show_object(struct object_array_entry *p)
1748 {
1749 add_preferred_base_object(p->name);
1750 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1751 }
1752
1753 static void show_edge(struct commit *commit)
1754 {
1755 add_preferred_base(commit->object.sha1);
1756 }
1757
1758 static void get_object_list(int ac, const char **av)
1759 {
1760 struct rev_info revs;
1761 char line[1000];
1762 int flags = 0;
1763
1764 init_revisions(&revs, NULL);
1765 save_commit_buffer = 0;
1766 track_object_refs = 0;
1767 setup_revisions(ac, av, &revs, NULL);
1768
1769 while (fgets(line, sizeof(line), stdin) != NULL) {
1770 int len = strlen(line);
1771 if (line[len - 1] == '\n')
1772 line[--len] = 0;
1773 if (!len)
1774 break;
1775 if (*line == '-') {
1776 if (!strcmp(line, "--not")) {
1777 flags ^= UNINTERESTING;
1778 continue;
1779 }
1780 die("not a rev '%s'", line);
1781 }
1782 if (handle_revision_arg(line, &revs, flags, 1))
1783 die("bad revision '%s'", line);
1784 }
1785
1786 prepare_revision_walk(&revs);
1787 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1788 traverse_commit_list(&revs, show_commit, show_object);
1789 }
1790
1791 static int adjust_perm(const char *path, mode_t mode)
1792 {
1793 if (chmod(path, mode))
1794 return -1;
1795 return adjust_shared_perm(path);
1796 }
1797
1798 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1799 {
1800 int use_internal_rev_list = 0;
1801 int thin = 0;
1802 uint32_t i;
1803 const char **rp_av;
1804 int rp_ac_alloc = 64;
1805 int rp_ac;
1806
1807 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
1808
1809 rp_av[0] = "pack-objects";
1810 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
1811 rp_ac = 2;
1812
1813 git_config(git_pack_config);
1814 if (!pack_compression_seen && core_compression_seen)
1815 pack_compression_level = core_compression_level;
1816
1817 progress = isatty(2);
1818 for (i = 1; i < argc; i++) {
1819 const char *arg = argv[i];
1820
1821 if (*arg != '-')
1822 break;
1823
1824 if (!strcmp("--non-empty", arg)) {
1825 non_empty = 1;
1826 continue;
1827 }
1828 if (!strcmp("--local", arg)) {
1829 local = 1;
1830 continue;
1831 }
1832 if (!strcmp("--incremental", arg)) {
1833 incremental = 1;
1834 continue;
1835 }
1836 if (!prefixcmp(arg, "--compression=")) {
1837 char *end;
1838 int level = strtoul(arg+14, &end, 0);
1839 if (!arg[14] || *end)
1840 usage(pack_usage);
1841 if (level == -1)
1842 level = Z_DEFAULT_COMPRESSION;
1843 else if (level < 0 || level > Z_BEST_COMPRESSION)
1844 die("bad pack compression level %d", level);
1845 pack_compression_level = level;
1846 continue;
1847 }
1848 if (!prefixcmp(arg, "--max-pack-size=")) {
1849 char *end;
1850 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
1851 if (!arg[16] || *end)
1852 usage(pack_usage);
1853 continue;
1854 }
1855 if (!prefixcmp(arg, "--window=")) {
1856 char *end;
1857 window = strtoul(arg+9, &end, 0);
1858 if (!arg[9] || *end)
1859 usage(pack_usage);
1860 continue;
1861 }
1862 if (!prefixcmp(arg, "--window-memory=")) {
1863 if (!git_parse_ulong(arg+16, &window_memory_limit))
1864 usage(pack_usage);
1865 continue;
1866 }
1867 if (!prefixcmp(arg, "--depth=")) {
1868 char *end;
1869 depth = strtoul(arg+8, &end, 0);
1870 if (!arg[8] || *end)
1871 usage(pack_usage);
1872 continue;
1873 }
1874 if (!strcmp("--progress", arg)) {
1875 progress = 1;
1876 continue;
1877 }
1878 if (!strcmp("--all-progress", arg)) {
1879 progress = 2;
1880 continue;
1881 }
1882 if (!strcmp("-q", arg)) {
1883 progress = 0;
1884 continue;
1885 }
1886 if (!strcmp("--no-reuse-delta", arg)) {
1887 no_reuse_delta = 1;
1888 continue;
1889 }
1890 if (!strcmp("--no-reuse-object", arg)) {
1891 no_reuse_object = no_reuse_delta = 1;
1892 continue;
1893 }
1894 if (!strcmp("--delta-base-offset", arg)) {
1895 allow_ofs_delta = 1;
1896 continue;
1897 }
1898 if (!strcmp("--stdout", arg)) {
1899 pack_to_stdout = 1;
1900 continue;
1901 }
1902 if (!strcmp("--revs", arg)) {
1903 use_internal_rev_list = 1;
1904 continue;
1905 }
1906 if (!strcmp("--unpacked", arg) ||
1907 !prefixcmp(arg, "--unpacked=") ||
1908 !strcmp("--reflog", arg) ||
1909 !strcmp("--all", arg)) {
1910 use_internal_rev_list = 1;
1911 if (rp_ac >= rp_ac_alloc - 1) {
1912 rp_ac_alloc = alloc_nr(rp_ac_alloc);
1913 rp_av = xrealloc(rp_av,
1914 rp_ac_alloc * sizeof(*rp_av));
1915 }
1916 rp_av[rp_ac++] = arg;
1917 continue;
1918 }
1919 if (!strcmp("--thin", arg)) {
1920 use_internal_rev_list = 1;
1921 thin = 1;
1922 rp_av[1] = "--objects-edge";
1923 continue;
1924 }
1925 if (!prefixcmp(arg, "--index-version=")) {
1926 char *c;
1927 pack_idx_default_version = strtoul(arg + 16, &c, 10);
1928 if (pack_idx_default_version > 2)
1929 die("bad %s", arg);
1930 if (*c == ',')
1931 pack_idx_off32_limit = strtoul(c+1, &c, 0);
1932 if (*c || pack_idx_off32_limit & 0x80000000)
1933 die("bad %s", arg);
1934 continue;
1935 }
1936 usage(pack_usage);
1937 }
1938
1939 /* Traditionally "pack-objects [options] base extra" failed;
1940 * we would however want to take refs parameter that would
1941 * have been given to upstream rev-list ourselves, which means
1942 * we somehow want to say what the base name is. So the
1943 * syntax would be:
1944 *
1945 * pack-objects [options] base <refs...>
1946 *
1947 * in other words, we would treat the first non-option as the
1948 * base_name and send everything else to the internal revision
1949 * walker.
1950 */
1951
1952 if (!pack_to_stdout)
1953 base_name = argv[i++];
1954
1955 if (pack_to_stdout != !base_name)
1956 usage(pack_usage);
1957
1958 if (pack_to_stdout && pack_size_limit)
1959 die("--max-pack-size cannot be used to build a pack for transfer.");
1960
1961 if (!pack_to_stdout && thin)
1962 die("--thin cannot be used to build an indexable pack.");
1963
1964 prepare_packed_git();
1965
1966 if (progress)
1967 start_progress(&progress_state, "Generating pack...",
1968 "Counting objects: ", 0);
1969 if (!use_internal_rev_list)
1970 read_object_list_from_stdin();
1971 else {
1972 rp_av[rp_ac] = NULL;
1973 get_object_list(rp_ac, rp_av);
1974 }
1975 if (progress) {
1976 stop_progress(&progress_state);
1977 fprintf(stderr, "Done counting %u objects.\n", nr_objects);
1978 }
1979
1980 if (non_empty && !nr_result)
1981 return 0;
1982 if (progress && (nr_objects != nr_result))
1983 fprintf(stderr, "Result has %u objects.\n", nr_result);
1984 if (nr_result)
1985 prepare_pack(window, depth);
1986 write_pack_file();
1987 if (progress)
1988 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
1989 written, written_delta, reused, reused_delta);
1990 return 0;
1991 }