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
22static const char pack_usage[] = "\
23git-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 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
28 [--stdout | base-name] [--keep-unreachable] [<ref-list | <object-list]";
29
30struct 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 */
59static struct object_entry *objects;
60static struct object_entry **written_list;
61static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
62
63static int non_empty;
64static int no_reuse_delta, no_reuse_object, keep_unreachable;
65static int local;
66static int incremental;
67static int allow_ofs_delta;
68static const char *base_name;
69static int progress = 1;
70static int window = 10;
71static uint32_t pack_size_limit;
72static int depth = 50;
73static int delta_search_threads = 1;
74static int pack_to_stdout;
75static int num_preferred_base;
76static struct progress *progress_state;
77static int pack_compression_level = Z_DEFAULT_COMPRESSION;
78static int pack_compression_seen;
79
80static unsigned long delta_cache_size = 0;
81static unsigned long max_delta_cache_size = 0;
82static unsigned long cache_max_small_delta_size = 1000;
83
84static unsigned long window_memory_limit = 0;
85
86/*
87 * The object names in objects array are hashed with this hashtable,
88 * to help looking up the entry by object name.
89 * This hashtable is built after all the objects are seen.
90 */
91static int *object_ix;
92static int object_ix_hashsz;
93
94/*
95 * Pack index for existing packs give us easy access to the offsets into
96 * corresponding pack file where each object's data starts, but the entries
97 * do not store the size of the compressed representation (uncompressed
98 * size is easily available by examining the pack entry header). It is
99 * also rather expensive to find the sha1 for an object given its offset.
100 *
101 * We build a hashtable of existing packs (pack_revindex), and keep reverse
102 * index here -- pack index file is sorted by object name mapping to offset;
103 * this pack_revindex[].revindex array is a list of offset/index_nr pairs
104 * ordered by offset, so if you know the offset of an object, next offset
105 * is where its packed representation ends and the index_nr can be used to
106 * get the object sha1 from the main index.
107 */
108struct revindex_entry {
109 off_t offset;
110 unsigned int nr;
111};
112struct pack_revindex {
113 struct packed_git *p;
114 struct revindex_entry *revindex;
115};
116static struct pack_revindex *pack_revindex;
117static int pack_revindex_hashsz;
118
119/*
120 * stats
121 */
122static uint32_t written, written_delta;
123static uint32_t reused, reused_delta;
124
125static int pack_revindex_ix(struct packed_git *p)
126{
127 unsigned long ui = (unsigned long)p;
128 int i;
129
130 ui = ui ^ (ui >> 16); /* defeat structure alignment */
131 i = (int)(ui % pack_revindex_hashsz);
132 while (pack_revindex[i].p) {
133 if (pack_revindex[i].p == p)
134 return i;
135 if (++i == pack_revindex_hashsz)
136 i = 0;
137 }
138 return -1 - i;
139}
140
141static void prepare_pack_ix(void)
142{
143 int num;
144 struct packed_git *p;
145 for (num = 0, p = packed_git; p; p = p->next)
146 num++;
147 if (!num)
148 return;
149 pack_revindex_hashsz = num * 11;
150 pack_revindex = xcalloc(sizeof(*pack_revindex), pack_revindex_hashsz);
151 for (p = packed_git; p; p = p->next) {
152 num = pack_revindex_ix(p);
153 num = - 1 - num;
154 pack_revindex[num].p = p;
155 }
156 /* revindex elements are lazily initialized */
157}
158
159static int cmp_offset(const void *a_, const void *b_)
160{
161 const struct revindex_entry *a = a_;
162 const struct revindex_entry *b = b_;
163 return (a->offset < b->offset) ? -1 : (a->offset > b->offset) ? 1 : 0;
164}
165
166/*
167 * Ordered list of offsets of objects in the pack.
168 */
169static void prepare_pack_revindex(struct pack_revindex *rix)
170{
171 struct packed_git *p = rix->p;
172 int num_ent = p->num_objects;
173 int i;
174 const char *index = p->index_data;
175
176 rix->revindex = xmalloc(sizeof(*rix->revindex) * (num_ent + 1));
177 index += 4 * 256;
178
179 if (p->index_version > 1) {
180 const uint32_t *off_32 =
181 (uint32_t *)(index + 8 + p->num_objects * (20 + 4));
182 const uint32_t *off_64 = off_32 + p->num_objects;
183 for (i = 0; i < num_ent; i++) {
184 uint32_t off = ntohl(*off_32++);
185 if (!(off & 0x80000000)) {
186 rix->revindex[i].offset = off;
187 } else {
188 rix->revindex[i].offset =
189 ((uint64_t)ntohl(*off_64++)) << 32;
190 rix->revindex[i].offset |=
191 ntohl(*off_64++);
192 }
193 rix->revindex[i].nr = i;
194 }
195 } else {
196 for (i = 0; i < num_ent; i++) {
197 uint32_t hl = *((uint32_t *)(index + 24 * i));
198 rix->revindex[i].offset = ntohl(hl);
199 rix->revindex[i].nr = i;
200 }
201 }
202
203 /* This knows the pack format -- the 20-byte trailer
204 * follows immediately after the last object data.
205 */
206 rix->revindex[num_ent].offset = p->pack_size - 20;
207 rix->revindex[num_ent].nr = -1;
208 qsort(rix->revindex, num_ent, sizeof(*rix->revindex), cmp_offset);
209}
210
211static struct revindex_entry * find_packed_object(struct packed_git *p,
212 off_t ofs)
213{
214 int num;
215 int lo, hi;
216 struct pack_revindex *rix;
217 struct revindex_entry *revindex;
218 num = pack_revindex_ix(p);
219 if (num < 0)
220 die("internal error: pack revindex uninitialized");
221 rix = &pack_revindex[num];
222 if (!rix->revindex)
223 prepare_pack_revindex(rix);
224 revindex = rix->revindex;
225 lo = 0;
226 hi = p->num_objects + 1;
227 do {
228 int mi = (lo + hi) / 2;
229 if (revindex[mi].offset == ofs) {
230 return revindex + mi;
231 }
232 else if (ofs < revindex[mi].offset)
233 hi = mi;
234 else
235 lo = mi + 1;
236 } while (lo < hi);
237 die("internal error: pack revindex corrupt");
238}
239
240static const unsigned char *find_packed_object_name(struct packed_git *p,
241 off_t ofs)
242{
243 struct revindex_entry *entry = find_packed_object(p, ofs);
244 return nth_packed_object_sha1(p, entry->nr);
245}
246
247static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
248{
249 unsigned long othersize, delta_size;
250 enum object_type type;
251 void *otherbuf = read_sha1_file(entry->delta->idx.sha1, &type, &othersize);
252 void *delta_buf;
253
254 if (!otherbuf)
255 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
256 delta_buf = diff_delta(otherbuf, othersize,
257 buf, size, &delta_size, 0);
258 if (!delta_buf || delta_size != entry->delta_size)
259 die("delta size changed");
260 free(buf);
261 free(otherbuf);
262 return delta_buf;
263}
264
265/*
266 * The per-object header is a pretty dense thing, which is
267 * - first byte: low four bits are "size", then three bits of "type",
268 * and the high bit is "size continues".
269 * - each byte afterwards: low seven bits are size continuation,
270 * with the high bit being "size continues"
271 */
272static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
273{
274 int n = 1;
275 unsigned char c;
276
277 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
278 die("bad type %d", type);
279
280 c = (type << 4) | (size & 15);
281 size >>= 4;
282 while (size) {
283 *hdr++ = c | 0x80;
284 c = size & 0x7f;
285 size >>= 7;
286 n++;
287 }
288 *hdr = c;
289 return n;
290}
291
292/*
293 * we are going to reuse the existing object data as is. make
294 * sure it is not corrupt.
295 */
296static int check_pack_inflate(struct packed_git *p,
297 struct pack_window **w_curs,
298 off_t offset,
299 off_t len,
300 unsigned long expect)
301{
302 z_stream stream;
303 unsigned char fakebuf[4096], *in;
304 int st;
305
306 memset(&stream, 0, sizeof(stream));
307 inflateInit(&stream);
308 do {
309 in = use_pack(p, w_curs, offset, &stream.avail_in);
310 stream.next_in = in;
311 stream.next_out = fakebuf;
312 stream.avail_out = sizeof(fakebuf);
313 st = inflate(&stream, Z_FINISH);
314 offset += stream.next_in - in;
315 } while (st == Z_OK || st == Z_BUF_ERROR);
316 inflateEnd(&stream);
317 return (st == Z_STREAM_END &&
318 stream.total_out == expect &&
319 stream.total_in == len) ? 0 : -1;
320}
321
322static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
323 off_t offset, off_t len, unsigned int nr)
324{
325 const uint32_t *index_crc;
326 uint32_t data_crc = crc32(0, Z_NULL, 0);
327
328 do {
329 unsigned int avail;
330 void *data = use_pack(p, w_curs, offset, &avail);
331 if (avail > len)
332 avail = len;
333 data_crc = crc32(data_crc, data, avail);
334 offset += avail;
335 len -= avail;
336 } while (len);
337
338 index_crc = p->index_data;
339 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
340
341 return data_crc != ntohl(*index_crc);
342}
343
344static void copy_pack_data(struct sha1file *f,
345 struct packed_git *p,
346 struct pack_window **w_curs,
347 off_t offset,
348 off_t len)
349{
350 unsigned char *in;
351 unsigned int avail;
352
353 while (len) {
354 in = use_pack(p, w_curs, offset, &avail);
355 if (avail > len)
356 avail = (unsigned int)len;
357 sha1write(f, in, avail);
358 offset += avail;
359 len -= avail;
360 }
361}
362
363static unsigned long write_object(struct sha1file *f,
364 struct object_entry *entry,
365 off_t write_offset)
366{
367 unsigned long size;
368 enum object_type type;
369 void *buf;
370 unsigned char header[10];
371 unsigned char dheader[10];
372 unsigned hdrlen;
373 off_t datalen;
374 enum object_type obj_type;
375 int to_reuse = 0;
376 /* write limit if limited packsize and not first object */
377 unsigned long limit = pack_size_limit && nr_written ?
378 pack_size_limit - write_offset : 0;
379 /* no if no delta */
380 int usable_delta = !entry->delta ? 0 :
381 /* yes if unlimited packfile */
382 !pack_size_limit ? 1 :
383 /* no if base written to previous pack */
384 entry->delta->idx.offset == (off_t)-1 ? 0 :
385 /* otherwise double-check written to this
386 * pack, like we do below
387 */
388 entry->delta->idx.offset ? 1 : 0;
389
390 if (!pack_to_stdout)
391 crc32_begin(f);
392
393 obj_type = entry->type;
394 if (no_reuse_object)
395 to_reuse = 0; /* explicit */
396 else if (!entry->in_pack)
397 to_reuse = 0; /* can't reuse what we don't have */
398 else if (obj_type == OBJ_REF_DELTA || obj_type == OBJ_OFS_DELTA)
399 /* check_object() decided it for us ... */
400 to_reuse = usable_delta;
401 /* ... but pack split may override that */
402 else if (obj_type != entry->in_pack_type)
403 to_reuse = 0; /* pack has delta which is unusable */
404 else if (entry->delta)
405 to_reuse = 0; /* we want to pack afresh */
406 else
407 to_reuse = 1; /* we have it in-pack undeltified,
408 * and we do not need to deltify it.
409 */
410
411 if (!to_reuse) {
412 z_stream stream;
413 unsigned long maxsize;
414 void *out;
415 if (!usable_delta) {
416 buf = read_sha1_file(entry->idx.sha1, &obj_type, &size);
417 if (!buf)
418 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
419 } else if (entry->delta_data) {
420 size = entry->delta_size;
421 buf = entry->delta_data;
422 entry->delta_data = NULL;
423 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
424 OBJ_OFS_DELTA : OBJ_REF_DELTA;
425 } else {
426 buf = read_sha1_file(entry->idx.sha1, &type, &size);
427 if (!buf)
428 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
429 buf = delta_against(buf, size, entry);
430 size = entry->delta_size;
431 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
432 OBJ_OFS_DELTA : OBJ_REF_DELTA;
433 }
434 /* compress the data to store and put compressed length in datalen */
435 memset(&stream, 0, sizeof(stream));
436 deflateInit(&stream, pack_compression_level);
437 maxsize = deflateBound(&stream, size);
438 out = xmalloc(maxsize);
439 /* Compress it */
440 stream.next_in = buf;
441 stream.avail_in = size;
442 stream.next_out = out;
443 stream.avail_out = maxsize;
444 while (deflate(&stream, Z_FINISH) == Z_OK)
445 /* nothing */;
446 deflateEnd(&stream);
447 datalen = stream.total_out;
448 deflateEnd(&stream);
449 /*
450 * The object header is a byte of 'type' followed by zero or
451 * more bytes of length.
452 */
453 hdrlen = encode_header(obj_type, size, header);
454
455 if (obj_type == OBJ_OFS_DELTA) {
456 /*
457 * Deltas with relative base contain an additional
458 * encoding of the relative offset for the delta
459 * base from this object's position in the pack.
460 */
461 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
462 unsigned pos = sizeof(dheader) - 1;
463 dheader[pos] = ofs & 127;
464 while (ofs >>= 7)
465 dheader[--pos] = 128 | (--ofs & 127);
466 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
467 free(out);
468 free(buf);
469 return 0;
470 }
471 sha1write(f, header, hdrlen);
472 sha1write(f, dheader + pos, sizeof(dheader) - pos);
473 hdrlen += sizeof(dheader) - pos;
474 } else if (obj_type == OBJ_REF_DELTA) {
475 /*
476 * Deltas with a base reference contain
477 * an additional 20 bytes for the base sha1.
478 */
479 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
480 free(out);
481 free(buf);
482 return 0;
483 }
484 sha1write(f, header, hdrlen);
485 sha1write(f, entry->delta->idx.sha1, 20);
486 hdrlen += 20;
487 } else {
488 if (limit && hdrlen + datalen + 20 >= limit) {
489 free(out);
490 free(buf);
491 return 0;
492 }
493 sha1write(f, header, hdrlen);
494 }
495 sha1write(f, out, datalen);
496 free(out);
497 free(buf);
498 }
499 else {
500 struct packed_git *p = entry->in_pack;
501 struct pack_window *w_curs = NULL;
502 struct revindex_entry *revidx;
503 off_t offset;
504
505 if (entry->delta) {
506 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
507 OBJ_OFS_DELTA : OBJ_REF_DELTA;
508 reused_delta++;
509 }
510 hdrlen = encode_header(obj_type, entry->size, header);
511 offset = entry->in_pack_offset;
512 revidx = find_packed_object(p, offset);
513 datalen = revidx[1].offset - offset;
514 if (!pack_to_stdout && p->index_version > 1 &&
515 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
516 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
517 offset += entry->in_pack_header_size;
518 datalen -= entry->in_pack_header_size;
519 if (obj_type == OBJ_OFS_DELTA) {
520 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
521 unsigned pos = sizeof(dheader) - 1;
522 dheader[pos] = ofs & 127;
523 while (ofs >>= 7)
524 dheader[--pos] = 128 | (--ofs & 127);
525 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
526 return 0;
527 sha1write(f, header, hdrlen);
528 sha1write(f, dheader + pos, sizeof(dheader) - pos);
529 hdrlen += sizeof(dheader) - pos;
530 } else if (obj_type == OBJ_REF_DELTA) {
531 if (limit && hdrlen + 20 + datalen + 20 >= limit)
532 return 0;
533 sha1write(f, header, hdrlen);
534 sha1write(f, entry->delta->idx.sha1, 20);
535 hdrlen += 20;
536 } else {
537 if (limit && hdrlen + datalen + 20 >= limit)
538 return 0;
539 sha1write(f, header, hdrlen);
540 }
541
542 if (!pack_to_stdout && p->index_version == 1 &&
543 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
544 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
545 copy_pack_data(f, p, &w_curs, offset, datalen);
546 unuse_pack(&w_curs);
547 reused++;
548 }
549 if (usable_delta)
550 written_delta++;
551 written++;
552 if (!pack_to_stdout)
553 entry->idx.crc32 = crc32_end(f);
554 return hdrlen + datalen;
555}
556
557static off_t write_one(struct sha1file *f,
558 struct object_entry *e,
559 off_t offset)
560{
561 unsigned long size;
562
563 /* offset is non zero if object is written already. */
564 if (e->idx.offset || e->preferred_base)
565 return offset;
566
567 /* if we are deltified, write out base object first. */
568 if (e->delta) {
569 offset = write_one(f, e->delta, offset);
570 if (!offset)
571 return 0;
572 }
573
574 e->idx.offset = offset;
575 size = write_object(f, e, offset);
576 if (!size) {
577 e->idx.offset = 0;
578 return 0;
579 }
580 written_list[nr_written++] = e;
581
582 /* make sure off_t is sufficiently large not to wrap */
583 if (offset > offset + size)
584 die("pack too large for current definition of off_t");
585 return offset + size;
586}
587
588/* forward declaration for write_pack_file */
589static int adjust_perm(const char *path, mode_t mode);
590
591static void write_pack_file(void)
592{
593 uint32_t i = 0, j;
594 struct sha1file *f;
595 off_t offset, offset_one, last_obj_offset = 0;
596 struct pack_header hdr;
597 int do_progress = progress >> pack_to_stdout;
598 uint32_t nr_remaining = nr_result;
599
600 if (do_progress)
601 progress_state = start_progress("Writing objects", nr_result);
602 written_list = xmalloc(nr_objects * sizeof(struct object_entry *));
603
604 do {
605 unsigned char sha1[20];
606 char *pack_tmp_name = NULL;
607
608 if (pack_to_stdout) {
609 f = sha1fd_throughput(1, "<stdout>", progress_state);
610 } else {
611 char tmpname[PATH_MAX];
612 int fd;
613 snprintf(tmpname, sizeof(tmpname),
614 "%s/tmp_pack_XXXXXX", get_object_directory());
615 fd = xmkstemp(tmpname);
616 pack_tmp_name = xstrdup(tmpname);
617 f = sha1fd(fd, pack_tmp_name);
618 }
619
620 hdr.hdr_signature = htonl(PACK_SIGNATURE);
621 hdr.hdr_version = htonl(PACK_VERSION);
622 hdr.hdr_entries = htonl(nr_remaining);
623 sha1write(f, &hdr, sizeof(hdr));
624 offset = sizeof(hdr);
625 nr_written = 0;
626 for (; i < nr_objects; i++) {
627 last_obj_offset = offset;
628 offset_one = write_one(f, objects + i, offset);
629 if (!offset_one)
630 break;
631 offset = offset_one;
632 display_progress(progress_state, written);
633 }
634
635 /*
636 * Did we write the wrong # entries in the header?
637 * If so, rewrite it like in fast-import
638 */
639 if (pack_to_stdout || nr_written == nr_remaining) {
640 sha1close(f, sha1, 1);
641 } else {
642 int fd = sha1close(f, NULL, 0);
643 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
644 close(fd);
645 }
646
647 if (!pack_to_stdout) {
648 mode_t mode = umask(0);
649 char *idx_tmp_name, tmpname[PATH_MAX];
650
651 umask(mode);
652 mode = 0444 & ~mode;
653
654 idx_tmp_name = write_idx_file(NULL,
655 (struct pack_idx_entry **) written_list,
656 nr_written, sha1);
657 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
658 base_name, sha1_to_hex(sha1));
659 if (adjust_perm(pack_tmp_name, mode))
660 die("unable to make temporary pack file readable: %s",
661 strerror(errno));
662 if (rename(pack_tmp_name, tmpname))
663 die("unable to rename temporary pack file: %s",
664 strerror(errno));
665 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
666 base_name, sha1_to_hex(sha1));
667 if (adjust_perm(idx_tmp_name, mode))
668 die("unable to make temporary index file readable: %s",
669 strerror(errno));
670 if (rename(idx_tmp_name, tmpname))
671 die("unable to rename temporary index file: %s",
672 strerror(errno));
673 free(idx_tmp_name);
674 free(pack_tmp_name);
675 puts(sha1_to_hex(sha1));
676 }
677
678 /* mark written objects as written to previous pack */
679 for (j = 0; j < nr_written; j++) {
680 written_list[j]->idx.offset = (off_t)-1;
681 }
682 nr_remaining -= nr_written;
683 } while (nr_remaining && i < nr_objects);
684
685 free(written_list);
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
703static 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
718static 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
731static 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
750static 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
771static 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
781static 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
793static 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 display_progress(progress_state, nr_objects);
855
856 if (name && no_try_delta(name))
857 entry->no_try_delta = 1;
858
859 return 1;
860}
861
862struct pbase_tree_cache {
863 unsigned char sha1[20];
864 int ref;
865 int temporary;
866 void *tree_data;
867 unsigned long tree_size;
868};
869
870static struct pbase_tree_cache *(pbase_tree_cache[256]);
871static int pbase_tree_cache_ix(const unsigned char *sha1)
872{
873 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
874}
875static int pbase_tree_cache_ix_incr(int ix)
876{
877 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
878}
879
880static struct pbase_tree {
881 struct pbase_tree *next;
882 /* This is a phony "cache" entry; we are not
883 * going to evict it nor find it through _get()
884 * mechanism -- this is for the toplevel node that
885 * would almost always change with any commit.
886 */
887 struct pbase_tree_cache pcache;
888} *pbase_tree;
889
890static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
891{
892 struct pbase_tree_cache *ent, *nent;
893 void *data;
894 unsigned long size;
895 enum object_type type;
896 int neigh;
897 int my_ix = pbase_tree_cache_ix(sha1);
898 int available_ix = -1;
899
900 /* pbase-tree-cache acts as a limited hashtable.
901 * your object will be found at your index or within a few
902 * slots after that slot if it is cached.
903 */
904 for (neigh = 0; neigh < 8; neigh++) {
905 ent = pbase_tree_cache[my_ix];
906 if (ent && !hashcmp(ent->sha1, sha1)) {
907 ent->ref++;
908 return ent;
909 }
910 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
911 ((0 <= available_ix) &&
912 (!ent && pbase_tree_cache[available_ix])))
913 available_ix = my_ix;
914 if (!ent)
915 break;
916 my_ix = pbase_tree_cache_ix_incr(my_ix);
917 }
918
919 /* Did not find one. Either we got a bogus request or
920 * we need to read and perhaps cache.
921 */
922 data = read_sha1_file(sha1, &type, &size);
923 if (!data)
924 return NULL;
925 if (type != OBJ_TREE) {
926 free(data);
927 return NULL;
928 }
929
930 /* We need to either cache or return a throwaway copy */
931
932 if (available_ix < 0)
933 ent = NULL;
934 else {
935 ent = pbase_tree_cache[available_ix];
936 my_ix = available_ix;
937 }
938
939 if (!ent) {
940 nent = xmalloc(sizeof(*nent));
941 nent->temporary = (available_ix < 0);
942 }
943 else {
944 /* evict and reuse */
945 free(ent->tree_data);
946 nent = ent;
947 }
948 hashcpy(nent->sha1, sha1);
949 nent->tree_data = data;
950 nent->tree_size = size;
951 nent->ref = 1;
952 if (!nent->temporary)
953 pbase_tree_cache[my_ix] = nent;
954 return nent;
955}
956
957static void pbase_tree_put(struct pbase_tree_cache *cache)
958{
959 if (!cache->temporary) {
960 cache->ref--;
961 return;
962 }
963 free(cache->tree_data);
964 free(cache);
965}
966
967static int name_cmp_len(const char *name)
968{
969 int i;
970 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
971 ;
972 return i;
973}
974
975static void add_pbase_object(struct tree_desc *tree,
976 const char *name,
977 int cmplen,
978 const char *fullname)
979{
980 struct name_entry entry;
981 int cmp;
982
983 while (tree_entry(tree,&entry)) {
984 if (S_ISGITLINK(entry.mode))
985 continue;
986 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
987 memcmp(name, entry.path, cmplen);
988 if (cmp > 0)
989 continue;
990 if (cmp < 0)
991 return;
992 if (name[cmplen] != '/') {
993 add_object_entry(entry.sha1,
994 S_ISDIR(entry.mode) ? OBJ_TREE : OBJ_BLOB,
995 fullname, 1);
996 return;
997 }
998 if (S_ISDIR(entry.mode)) {
999 struct tree_desc sub;
1000 struct pbase_tree_cache *tree;
1001 const char *down = name+cmplen+1;
1002 int downlen = name_cmp_len(down);
1003
1004 tree = pbase_tree_get(entry.sha1);
1005 if (!tree)
1006 return;
1007 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1008
1009 add_pbase_object(&sub, down, downlen, fullname);
1010 pbase_tree_put(tree);
1011 }
1012 }
1013}
1014
1015static unsigned *done_pbase_paths;
1016static int done_pbase_paths_num;
1017static int done_pbase_paths_alloc;
1018static int done_pbase_path_pos(unsigned hash)
1019{
1020 int lo = 0;
1021 int hi = done_pbase_paths_num;
1022 while (lo < hi) {
1023 int mi = (hi + lo) / 2;
1024 if (done_pbase_paths[mi] == hash)
1025 return mi;
1026 if (done_pbase_paths[mi] < hash)
1027 hi = mi;
1028 else
1029 lo = mi + 1;
1030 }
1031 return -lo-1;
1032}
1033
1034static int check_pbase_path(unsigned hash)
1035{
1036 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1037 if (0 <= pos)
1038 return 1;
1039 pos = -pos - 1;
1040 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1041 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1042 done_pbase_paths = xrealloc(done_pbase_paths,
1043 done_pbase_paths_alloc *
1044 sizeof(unsigned));
1045 }
1046 done_pbase_paths_num++;
1047 if (pos < done_pbase_paths_num)
1048 memmove(done_pbase_paths + pos + 1,
1049 done_pbase_paths + pos,
1050 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1051 done_pbase_paths[pos] = hash;
1052 return 0;
1053}
1054
1055static void add_preferred_base_object(const char *name)
1056{
1057 struct pbase_tree *it;
1058 int cmplen;
1059 unsigned hash = name_hash(name);
1060
1061 if (!num_preferred_base || check_pbase_path(hash))
1062 return;
1063
1064 cmplen = name_cmp_len(name);
1065 for (it = pbase_tree; it; it = it->next) {
1066 if (cmplen == 0) {
1067 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1068 }
1069 else {
1070 struct tree_desc tree;
1071 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1072 add_pbase_object(&tree, name, cmplen, name);
1073 }
1074 }
1075}
1076
1077static void add_preferred_base(unsigned char *sha1)
1078{
1079 struct pbase_tree *it;
1080 void *data;
1081 unsigned long size;
1082 unsigned char tree_sha1[20];
1083
1084 if (window <= num_preferred_base++)
1085 return;
1086
1087 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1088 if (!data)
1089 return;
1090
1091 for (it = pbase_tree; it; it = it->next) {
1092 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1093 free(data);
1094 return;
1095 }
1096 }
1097
1098 it = xcalloc(1, sizeof(*it));
1099 it->next = pbase_tree;
1100 pbase_tree = it;
1101
1102 hashcpy(it->pcache.sha1, tree_sha1);
1103 it->pcache.tree_data = data;
1104 it->pcache.tree_size = size;
1105}
1106
1107static void check_object(struct object_entry *entry)
1108{
1109 if (entry->in_pack) {
1110 struct packed_git *p = entry->in_pack;
1111 struct pack_window *w_curs = NULL;
1112 const unsigned char *base_ref = NULL;
1113 struct object_entry *base_entry;
1114 unsigned long used, used_0;
1115 unsigned int avail;
1116 off_t ofs;
1117 unsigned char *buf, c;
1118
1119 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1120
1121 /*
1122 * We want in_pack_type even if we do not reuse delta
1123 * since non-delta representations could still be reused.
1124 */
1125 used = unpack_object_header_gently(buf, avail,
1126 &entry->in_pack_type,
1127 &entry->size);
1128
1129 /*
1130 * Determine if this is a delta and if so whether we can
1131 * reuse it or not. Otherwise let's find out as cheaply as
1132 * possible what the actual type and size for this object is.
1133 */
1134 switch (entry->in_pack_type) {
1135 default:
1136 /* Not a delta hence we've already got all we need. */
1137 entry->type = entry->in_pack_type;
1138 entry->in_pack_header_size = used;
1139 unuse_pack(&w_curs);
1140 return;
1141 case OBJ_REF_DELTA:
1142 if (!no_reuse_delta && !entry->preferred_base)
1143 base_ref = use_pack(p, &w_curs,
1144 entry->in_pack_offset + used, NULL);
1145 entry->in_pack_header_size = used + 20;
1146 break;
1147 case OBJ_OFS_DELTA:
1148 buf = use_pack(p, &w_curs,
1149 entry->in_pack_offset + used, NULL);
1150 used_0 = 0;
1151 c = buf[used_0++];
1152 ofs = c & 127;
1153 while (c & 128) {
1154 ofs += 1;
1155 if (!ofs || MSB(ofs, 7))
1156 die("delta base offset overflow in pack for %s",
1157 sha1_to_hex(entry->idx.sha1));
1158 c = buf[used_0++];
1159 ofs = (ofs << 7) + (c & 127);
1160 }
1161 if (ofs >= entry->in_pack_offset)
1162 die("delta base offset out of bound for %s",
1163 sha1_to_hex(entry->idx.sha1));
1164 ofs = entry->in_pack_offset - ofs;
1165 if (!no_reuse_delta && !entry->preferred_base)
1166 base_ref = find_packed_object_name(p, ofs);
1167 entry->in_pack_header_size = used + used_0;
1168 break;
1169 }
1170
1171 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1172 /*
1173 * If base_ref was set above that means we wish to
1174 * reuse delta data, and we even found that base
1175 * in the list of objects we want to pack. Goodie!
1176 *
1177 * Depth value does not matter - find_deltas() will
1178 * never consider reused delta as the base object to
1179 * deltify other objects against, in order to avoid
1180 * circular deltas.
1181 */
1182 entry->type = entry->in_pack_type;
1183 entry->delta = base_entry;
1184 entry->delta_sibling = base_entry->delta_child;
1185 base_entry->delta_child = entry;
1186 unuse_pack(&w_curs);
1187 return;
1188 }
1189
1190 if (entry->type) {
1191 /*
1192 * This must be a delta and we already know what the
1193 * final object type is. Let's extract the actual
1194 * object size from the delta header.
1195 */
1196 entry->size = get_size_from_delta(p, &w_curs,
1197 entry->in_pack_offset + entry->in_pack_header_size);
1198 unuse_pack(&w_curs);
1199 return;
1200 }
1201
1202 /*
1203 * No choice but to fall back to the recursive delta walk
1204 * with sha1_object_info() to find about the object type
1205 * at this point...
1206 */
1207 unuse_pack(&w_curs);
1208 }
1209
1210 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1211 if (entry->type < 0)
1212 die("unable to get type of object %s",
1213 sha1_to_hex(entry->idx.sha1));
1214}
1215
1216static int pack_offset_sort(const void *_a, const void *_b)
1217{
1218 const struct object_entry *a = *(struct object_entry **)_a;
1219 const struct object_entry *b = *(struct object_entry **)_b;
1220
1221 /* avoid filesystem trashing with loose objects */
1222 if (!a->in_pack && !b->in_pack)
1223 return hashcmp(a->idx.sha1, b->idx.sha1);
1224
1225 if (a->in_pack < b->in_pack)
1226 return -1;
1227 if (a->in_pack > b->in_pack)
1228 return 1;
1229 return a->in_pack_offset < b->in_pack_offset ? -1 :
1230 (a->in_pack_offset > b->in_pack_offset);
1231}
1232
1233static void get_object_details(void)
1234{
1235 uint32_t i;
1236 struct object_entry **sorted_by_offset;
1237
1238 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1239 for (i = 0; i < nr_objects; i++)
1240 sorted_by_offset[i] = objects + i;
1241 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1242
1243 prepare_pack_ix();
1244 for (i = 0; i < nr_objects; i++)
1245 check_object(sorted_by_offset[i]);
1246 free(sorted_by_offset);
1247}
1248
1249static int type_size_sort(const void *_a, const void *_b)
1250{
1251 const struct object_entry *a = *(struct object_entry **)_a;
1252 const struct object_entry *b = *(struct object_entry **)_b;
1253
1254 if (a->type < b->type)
1255 return -1;
1256 if (a->type > b->type)
1257 return 1;
1258 if (a->hash < b->hash)
1259 return -1;
1260 if (a->hash > b->hash)
1261 return 1;
1262 if (a->preferred_base < b->preferred_base)
1263 return -1;
1264 if (a->preferred_base > b->preferred_base)
1265 return 1;
1266 if (a->size < b->size)
1267 return -1;
1268 if (a->size > b->size)
1269 return 1;
1270 return a > b ? -1 : (a < b); /* newest last */
1271}
1272
1273struct unpacked {
1274 struct object_entry *entry;
1275 void *data;
1276 struct delta_index *index;
1277 unsigned depth;
1278};
1279
1280static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1281 unsigned long delta_size)
1282{
1283 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1284 return 0;
1285
1286 if (delta_size < cache_max_small_delta_size)
1287 return 1;
1288
1289 /* cache delta, if objects are large enough compared to delta size */
1290 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1291 return 1;
1292
1293 return 0;
1294}
1295
1296#ifdef THREADED_DELTA_SEARCH
1297
1298static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1299#define read_lock() pthread_mutex_lock(&read_mutex)
1300#define read_unlock() pthread_mutex_unlock(&read_mutex)
1301
1302static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1303#define cache_lock() pthread_mutex_lock(&cache_mutex)
1304#define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1305
1306static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1307#define progress_lock() pthread_mutex_lock(&progress_mutex)
1308#define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1309
1310#else
1311
1312#define read_lock() (void)0
1313#define read_unlock() (void)0
1314#define cache_lock() (void)0
1315#define cache_unlock() (void)0
1316#define progress_lock() (void)0
1317#define progress_unlock() (void)0
1318
1319#endif
1320
1321/*
1322 * We search for deltas _backwards_ in a list sorted by type and
1323 * by size, so that we see progressively smaller and smaller files.
1324 * That's because we prefer deltas to be from the bigger file
1325 * to the smaller - deletes are potentially cheaper, but perhaps
1326 * more importantly, the bigger file is likely the more recent
1327 * one.
1328 */
1329static int try_delta(struct unpacked *trg, struct unpacked *src,
1330 unsigned max_depth, unsigned long *mem_usage)
1331{
1332 struct object_entry *trg_entry = trg->entry;
1333 struct object_entry *src_entry = src->entry;
1334 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1335 unsigned ref_depth;
1336 enum object_type type;
1337 void *delta_buf;
1338
1339 /* Don't bother doing diffs between different types */
1340 if (trg_entry->type != src_entry->type)
1341 return -1;
1342
1343 /*
1344 * We do not bother to try a delta that we discarded
1345 * on an earlier try, but only when reusing delta data.
1346 */
1347 if (!no_reuse_delta && trg_entry->in_pack &&
1348 trg_entry->in_pack == src_entry->in_pack &&
1349 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1350 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1351 return 0;
1352
1353 /* Let's not bust the allowed depth. */
1354 if (src->depth >= max_depth)
1355 return 0;
1356
1357 /* Now some size filtering heuristics. */
1358 trg_size = trg_entry->size;
1359 if (!trg_entry->delta) {
1360 max_size = trg_size/2 - 20;
1361 ref_depth = 1;
1362 } else {
1363 max_size = trg_entry->delta_size;
1364 ref_depth = trg->depth;
1365 }
1366 max_size = max_size * (max_depth - src->depth) /
1367 (max_depth - ref_depth + 1);
1368 if (max_size == 0)
1369 return 0;
1370 src_size = src_entry->size;
1371 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1372 if (sizediff >= max_size)
1373 return 0;
1374 if (trg_size < src_size / 32)
1375 return 0;
1376
1377 /* Load data if not already done */
1378 if (!trg->data) {
1379 read_lock();
1380 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1381 read_unlock();
1382 if (!trg->data)
1383 die("object %s cannot be read",
1384 sha1_to_hex(trg_entry->idx.sha1));
1385 if (sz != trg_size)
1386 die("object %s inconsistent object length (%lu vs %lu)",
1387 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1388 *mem_usage += sz;
1389 }
1390 if (!src->data) {
1391 read_lock();
1392 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1393 read_unlock();
1394 if (!src->data)
1395 die("object %s cannot be read",
1396 sha1_to_hex(src_entry->idx.sha1));
1397 if (sz != src_size)
1398 die("object %s inconsistent object length (%lu vs %lu)",
1399 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1400 *mem_usage += sz;
1401 }
1402 if (!src->index) {
1403 src->index = create_delta_index(src->data, src_size);
1404 if (!src->index) {
1405 static int warned = 0;
1406 if (!warned++)
1407 warning("suboptimal pack - out of memory");
1408 return 0;
1409 }
1410 *mem_usage += sizeof_delta_index(src->index);
1411 }
1412
1413 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1414 if (!delta_buf)
1415 return 0;
1416
1417 if (trg_entry->delta) {
1418 /* Prefer only shallower same-sized deltas. */
1419 if (delta_size == trg_entry->delta_size &&
1420 src->depth + 1 >= trg->depth) {
1421 free(delta_buf);
1422 return 0;
1423 }
1424 }
1425
1426 trg_entry->delta = src_entry;
1427 trg_entry->delta_size = delta_size;
1428 trg->depth = src->depth + 1;
1429
1430 /*
1431 * Handle memory allocation outside of the cache
1432 * accounting lock. Compiler will optimize the strangeness
1433 * away when THREADED_DELTA_SEARCH is not defined.
1434 */
1435 if (trg_entry->delta_data)
1436 free(trg_entry->delta_data);
1437 cache_lock();
1438 if (trg_entry->delta_data) {
1439 delta_cache_size -= trg_entry->delta_size;
1440 trg_entry->delta_data = NULL;
1441 }
1442 if (delta_cacheable(src_size, trg_size, delta_size)) {
1443 delta_cache_size += trg_entry->delta_size;
1444 cache_unlock();
1445 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1446 } else {
1447 cache_unlock();
1448 free(delta_buf);
1449 }
1450
1451 return 1;
1452}
1453
1454static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1455{
1456 struct object_entry *child = me->delta_child;
1457 unsigned int m = n;
1458 while (child) {
1459 unsigned int c = check_delta_limit(child, n + 1);
1460 if (m < c)
1461 m = c;
1462 child = child->delta_sibling;
1463 }
1464 return m;
1465}
1466
1467static unsigned long free_unpacked(struct unpacked *n)
1468{
1469 unsigned long freed_mem = sizeof_delta_index(n->index);
1470 free_delta_index(n->index);
1471 n->index = NULL;
1472 if (n->data) {
1473 freed_mem += n->entry->size;
1474 free(n->data);
1475 n->data = NULL;
1476 }
1477 n->entry = NULL;
1478 n->depth = 0;
1479 return freed_mem;
1480}
1481
1482static void find_deltas(struct object_entry **list, unsigned list_size,
1483 int window, int depth, unsigned *processed)
1484{
1485 uint32_t i = list_size, idx = 0, count = 0;
1486 unsigned int array_size = window * sizeof(struct unpacked);
1487 struct unpacked *array;
1488 unsigned long mem_usage = 0;
1489
1490 array = xmalloc(array_size);
1491 memset(array, 0, array_size);
1492
1493 do {
1494 struct object_entry *entry = list[--i];
1495 struct unpacked *n = array + idx;
1496 int j, max_depth, best_base = -1;
1497
1498 mem_usage -= free_unpacked(n);
1499 n->entry = entry;
1500
1501 while (window_memory_limit &&
1502 mem_usage > window_memory_limit &&
1503 count > 1) {
1504 uint32_t tail = (idx + window - count) % window;
1505 mem_usage -= free_unpacked(array + tail);
1506 count--;
1507 }
1508
1509 /* We do not compute delta to *create* objects we are not
1510 * going to pack.
1511 */
1512 if (entry->preferred_base)
1513 goto next;
1514
1515 progress_lock();
1516 (*processed)++;
1517 display_progress(progress_state, *processed);
1518 progress_unlock();
1519
1520 /*
1521 * If the current object is at pack edge, take the depth the
1522 * objects that depend on the current object into account
1523 * otherwise they would become too deep.
1524 */
1525 max_depth = depth;
1526 if (entry->delta_child) {
1527 max_depth -= check_delta_limit(entry, 0);
1528 if (max_depth <= 0)
1529 goto next;
1530 }
1531
1532 j = window;
1533 while (--j > 0) {
1534 int ret;
1535 uint32_t other_idx = idx + j;
1536 struct unpacked *m;
1537 if (other_idx >= window)
1538 other_idx -= window;
1539 m = array + other_idx;
1540 if (!m->entry)
1541 break;
1542 ret = try_delta(n, m, max_depth, &mem_usage);
1543 if (ret < 0)
1544 break;
1545 else if (ret > 0)
1546 best_base = other_idx;
1547 }
1548
1549 /* if we made n a delta, and if n is already at max
1550 * depth, leaving it in the window is pointless. we
1551 * should evict it first.
1552 */
1553 if (entry->delta && depth <= n->depth)
1554 continue;
1555
1556 /*
1557 * Move the best delta base up in the window, after the
1558 * currently deltified object, to keep it longer. It will
1559 * be the first base object to be attempted next.
1560 */
1561 if (entry->delta) {
1562 struct unpacked swap = array[best_base];
1563 int dist = (window + idx - best_base) % window;
1564 int dst = best_base;
1565 while (dist--) {
1566 int src = (dst + 1) % window;
1567 array[dst] = array[src];
1568 dst = src;
1569 }
1570 array[dst] = swap;
1571 }
1572
1573 next:
1574 idx++;
1575 if (count + 1 < window)
1576 count++;
1577 if (idx >= window)
1578 idx = 0;
1579 } while (i > 0);
1580
1581 for (i = 0; i < window; ++i) {
1582 free_delta_index(array[i].index);
1583 free(array[i].data);
1584 }
1585 free(array);
1586}
1587
1588#ifdef THREADED_DELTA_SEARCH
1589
1590struct thread_params {
1591 pthread_t thread;
1592 struct object_entry **list;
1593 unsigned list_size;
1594 int window;
1595 int depth;
1596 unsigned *processed;
1597};
1598
1599static pthread_mutex_t data_request = PTHREAD_MUTEX_INITIALIZER;
1600static pthread_mutex_t data_ready = PTHREAD_MUTEX_INITIALIZER;
1601static pthread_mutex_t data_provider = PTHREAD_MUTEX_INITIALIZER;
1602static struct thread_params *data_requester;
1603
1604static void *threaded_find_deltas(void *arg)
1605{
1606 struct thread_params *me = arg;
1607
1608 for (;;) {
1609 pthread_mutex_lock(&data_request);
1610 data_requester = me;
1611 pthread_mutex_unlock(&data_provider);
1612 pthread_mutex_lock(&data_ready);
1613 pthread_mutex_unlock(&data_request);
1614
1615 if (!me->list_size)
1616 return NULL;
1617
1618 find_deltas(me->list, me->list_size,
1619 me->window, me->depth, me->processed);
1620 }
1621}
1622
1623static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1624 int window, int depth, unsigned *processed)
1625{
1626 struct thread_params *target, p[delta_search_threads];
1627 int i, ret;
1628 unsigned chunk_size;
1629
1630 if (delta_search_threads <= 1) {
1631 find_deltas(list, list_size, window, depth, processed);
1632 return;
1633 }
1634
1635 pthread_mutex_lock(&data_provider);
1636 pthread_mutex_lock(&data_ready);
1637
1638 for (i = 0; i < delta_search_threads; i++) {
1639 p[i].window = window;
1640 p[i].depth = depth;
1641 p[i].processed = processed;
1642 ret = pthread_create(&p[i].thread, NULL,
1643 threaded_find_deltas, &p[i]);
1644 if (ret)
1645 die("unable to create thread: %s", strerror(ret));
1646 }
1647
1648 /* this should be auto-tuned somehow */
1649 chunk_size = window * 1000;
1650
1651 do {
1652 unsigned sublist_size = chunk_size;
1653 if (sublist_size > list_size)
1654 sublist_size = list_size;
1655
1656 /* try to split chunks on "path" boundaries */
1657 while (sublist_size < list_size && list[sublist_size]->hash &&
1658 list[sublist_size]->hash == list[sublist_size-1]->hash)
1659 sublist_size++;
1660
1661 pthread_mutex_lock(&data_provider);
1662 target = data_requester;
1663 target->list = list;
1664 target->list_size = sublist_size;
1665 pthread_mutex_unlock(&data_ready);
1666
1667 list += sublist_size;
1668 list_size -= sublist_size;
1669 if (!sublist_size) {
1670 pthread_join(target->thread, NULL);
1671 i--;
1672 }
1673 } while (i);
1674}
1675
1676#else
1677#define ll_find_deltas find_deltas
1678#endif
1679
1680static void prepare_pack(int window, int depth)
1681{
1682 struct object_entry **delta_list;
1683 uint32_t i, n, nr_deltas;
1684
1685 get_object_details();
1686
1687 if (!nr_objects || !window || !depth)
1688 return;
1689
1690 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1691 nr_deltas = n = 0;
1692
1693 for (i = 0; i < nr_objects; i++) {
1694 struct object_entry *entry = objects + i;
1695
1696 if (entry->delta)
1697 /* This happens if we decided to reuse existing
1698 * delta from a pack. "!no_reuse_delta &&" is implied.
1699 */
1700 continue;
1701
1702 if (entry->size < 50)
1703 continue;
1704
1705 if (entry->no_try_delta)
1706 continue;
1707
1708 if (!entry->preferred_base)
1709 nr_deltas++;
1710
1711 delta_list[n++] = entry;
1712 }
1713
1714 if (nr_deltas && n > 1) {
1715 unsigned nr_done = 0;
1716 if (progress)
1717 progress_state = start_progress("Compressing objects",
1718 nr_deltas);
1719 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1720 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1721 stop_progress(&progress_state);
1722 if (nr_done != nr_deltas)
1723 die("inconsistency with delta count");
1724 }
1725 free(delta_list);
1726}
1727
1728static int git_pack_config(const char *k, const char *v)
1729{
1730 if(!strcmp(k, "pack.window")) {
1731 window = git_config_int(k, v);
1732 return 0;
1733 }
1734 if (!strcmp(k, "pack.windowmemory")) {
1735 window_memory_limit = git_config_ulong(k, v);
1736 return 0;
1737 }
1738 if (!strcmp(k, "pack.depth")) {
1739 depth = git_config_int(k, v);
1740 return 0;
1741 }
1742 if (!strcmp(k, "pack.compression")) {
1743 int level = git_config_int(k, v);
1744 if (level == -1)
1745 level = Z_DEFAULT_COMPRESSION;
1746 else if (level < 0 || level > Z_BEST_COMPRESSION)
1747 die("bad pack compression level %d", level);
1748 pack_compression_level = level;
1749 pack_compression_seen = 1;
1750 return 0;
1751 }
1752 if (!strcmp(k, "pack.deltacachesize")) {
1753 max_delta_cache_size = git_config_int(k, v);
1754 return 0;
1755 }
1756 if (!strcmp(k, "pack.deltacachelimit")) {
1757 cache_max_small_delta_size = git_config_int(k, v);
1758 return 0;
1759 }
1760 if (!strcmp(k, "pack.threads")) {
1761 delta_search_threads = git_config_int(k, v);
1762 if (delta_search_threads < 1)
1763 die("invalid number of threads specified (%d)",
1764 delta_search_threads);
1765#ifndef THREADED_DELTA_SEARCH
1766 if (delta_search_threads > 1)
1767 warning("no threads support, ignoring %s", k);
1768#endif
1769 return 0;
1770 }
1771 return git_default_config(k, v);
1772}
1773
1774static void read_object_list_from_stdin(void)
1775{
1776 char line[40 + 1 + PATH_MAX + 2];
1777 unsigned char sha1[20];
1778
1779 for (;;) {
1780 if (!fgets(line, sizeof(line), stdin)) {
1781 if (feof(stdin))
1782 break;
1783 if (!ferror(stdin))
1784 die("fgets returned NULL, not EOF, not error!");
1785 if (errno != EINTR)
1786 die("fgets: %s", strerror(errno));
1787 clearerr(stdin);
1788 continue;
1789 }
1790 if (line[0] == '-') {
1791 if (get_sha1_hex(line+1, sha1))
1792 die("expected edge sha1, got garbage:\n %s",
1793 line);
1794 add_preferred_base(sha1);
1795 continue;
1796 }
1797 if (get_sha1_hex(line, sha1))
1798 die("expected sha1, got garbage:\n %s", line);
1799
1800 add_preferred_base_object(line+41);
1801 add_object_entry(sha1, 0, line+41, 0);
1802 }
1803}
1804
1805#define OBJECT_ADDED (1u<<20)
1806
1807static void show_commit(struct commit *commit)
1808{
1809 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1810 commit->object.flags |= OBJECT_ADDED;
1811}
1812
1813static void show_object(struct object_array_entry *p)
1814{
1815 add_preferred_base_object(p->name);
1816 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1817 p->item->flags |= OBJECT_ADDED;
1818}
1819
1820static void show_edge(struct commit *commit)
1821{
1822 add_preferred_base(commit->object.sha1);
1823}
1824
1825struct in_pack_object {
1826 off_t offset;
1827 struct object *object;
1828};
1829
1830struct in_pack {
1831 int alloc;
1832 int nr;
1833 struct in_pack_object *array;
1834};
1835
1836static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1837{
1838 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1839 in_pack->array[in_pack->nr].object = object;
1840 in_pack->nr++;
1841}
1842
1843/*
1844 * Compare the objects in the offset order, in order to emulate the
1845 * "git-rev-list --objects" output that produced the pack originally.
1846 */
1847static int ofscmp(const void *a_, const void *b_)
1848{
1849 struct in_pack_object *a = (struct in_pack_object *)a_;
1850 struct in_pack_object *b = (struct in_pack_object *)b_;
1851
1852 if (a->offset < b->offset)
1853 return -1;
1854 else if (a->offset > b->offset)
1855 return 1;
1856 else
1857 return hashcmp(a->object->sha1, b->object->sha1);
1858}
1859
1860static void add_objects_in_unpacked_packs(struct rev_info *revs)
1861{
1862 struct packed_git *p;
1863 struct in_pack in_pack;
1864 uint32_t i;
1865
1866 memset(&in_pack, 0, sizeof(in_pack));
1867
1868 for (p = packed_git; p; p = p->next) {
1869 const unsigned char *sha1;
1870 struct object *o;
1871
1872 for (i = 0; i < revs->num_ignore_packed; i++) {
1873 if (matches_pack_name(p, revs->ignore_packed[i]))
1874 break;
1875 }
1876 if (revs->num_ignore_packed <= i)
1877 continue;
1878 if (open_pack_index(p))
1879 die("cannot open pack index");
1880
1881 ALLOC_GROW(in_pack.array,
1882 in_pack.nr + p->num_objects,
1883 in_pack.alloc);
1884
1885 for (i = 0; i < p->num_objects; i++) {
1886 sha1 = nth_packed_object_sha1(p, i);
1887 o = lookup_unknown_object(sha1);
1888 if (!(o->flags & OBJECT_ADDED))
1889 mark_in_pack_object(o, p, &in_pack);
1890 o->flags |= OBJECT_ADDED;
1891 }
1892 }
1893
1894 if (in_pack.nr) {
1895 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1896 ofscmp);
1897 for (i = 0; i < in_pack.nr; i++) {
1898 struct object *o = in_pack.array[i].object;
1899 add_object_entry(o->sha1, o->type, "", 0);
1900 }
1901 }
1902 free(in_pack.array);
1903}
1904
1905static void get_object_list(int ac, const char **av)
1906{
1907 struct rev_info revs;
1908 char line[1000];
1909 int flags = 0;
1910
1911 init_revisions(&revs, NULL);
1912 save_commit_buffer = 0;
1913 track_object_refs = 0;
1914 setup_revisions(ac, av, &revs, NULL);
1915
1916 while (fgets(line, sizeof(line), stdin) != NULL) {
1917 int len = strlen(line);
1918 if (line[len - 1] == '\n')
1919 line[--len] = 0;
1920 if (!len)
1921 break;
1922 if (*line == '-') {
1923 if (!strcmp(line, "--not")) {
1924 flags ^= UNINTERESTING;
1925 continue;
1926 }
1927 die("not a rev '%s'", line);
1928 }
1929 if (handle_revision_arg(line, &revs, flags, 1))
1930 die("bad revision '%s'", line);
1931 }
1932
1933 prepare_revision_walk(&revs);
1934 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1935 traverse_commit_list(&revs, show_commit, show_object);
1936
1937 if (keep_unreachable)
1938 add_objects_in_unpacked_packs(&revs);
1939}
1940
1941static int adjust_perm(const char *path, mode_t mode)
1942{
1943 if (chmod(path, mode))
1944 return -1;
1945 return adjust_shared_perm(path);
1946}
1947
1948int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1949{
1950 int use_internal_rev_list = 0;
1951 int thin = 0;
1952 uint32_t i;
1953 const char **rp_av;
1954 int rp_ac_alloc = 64;
1955 int rp_ac;
1956
1957 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
1958
1959 rp_av[0] = "pack-objects";
1960 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
1961 rp_ac = 2;
1962
1963 git_config(git_pack_config);
1964 if (!pack_compression_seen && core_compression_seen)
1965 pack_compression_level = core_compression_level;
1966
1967 progress = isatty(2);
1968 for (i = 1; i < argc; i++) {
1969 const char *arg = argv[i];
1970
1971 if (*arg != '-')
1972 break;
1973
1974 if (!strcmp("--non-empty", arg)) {
1975 non_empty = 1;
1976 continue;
1977 }
1978 if (!strcmp("--local", arg)) {
1979 local = 1;
1980 continue;
1981 }
1982 if (!strcmp("--incremental", arg)) {
1983 incremental = 1;
1984 continue;
1985 }
1986 if (!prefixcmp(arg, "--compression=")) {
1987 char *end;
1988 int level = strtoul(arg+14, &end, 0);
1989 if (!arg[14] || *end)
1990 usage(pack_usage);
1991 if (level == -1)
1992 level = Z_DEFAULT_COMPRESSION;
1993 else if (level < 0 || level > Z_BEST_COMPRESSION)
1994 die("bad pack compression level %d", level);
1995 pack_compression_level = level;
1996 continue;
1997 }
1998 if (!prefixcmp(arg, "--max-pack-size=")) {
1999 char *end;
2000 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2001 if (!arg[16] || *end)
2002 usage(pack_usage);
2003 continue;
2004 }
2005 if (!prefixcmp(arg, "--window=")) {
2006 char *end;
2007 window = strtoul(arg+9, &end, 0);
2008 if (!arg[9] || *end)
2009 usage(pack_usage);
2010 continue;
2011 }
2012 if (!prefixcmp(arg, "--window-memory=")) {
2013 if (!git_parse_ulong(arg+16, &window_memory_limit))
2014 usage(pack_usage);
2015 continue;
2016 }
2017 if (!prefixcmp(arg, "--threads=")) {
2018 char *end;
2019 delta_search_threads = strtoul(arg+10, &end, 0);
2020 if (!arg[10] || *end || delta_search_threads < 1)
2021 usage(pack_usage);
2022#ifndef THREADED_DELTA_SEARCH
2023 if (delta_search_threads > 1)
2024 warning("no threads support, "
2025 "ignoring %s", arg);
2026#endif
2027 continue;
2028 }
2029 if (!prefixcmp(arg, "--depth=")) {
2030 char *end;
2031 depth = strtoul(arg+8, &end, 0);
2032 if (!arg[8] || *end)
2033 usage(pack_usage);
2034 continue;
2035 }
2036 if (!strcmp("--progress", arg)) {
2037 progress = 1;
2038 continue;
2039 }
2040 if (!strcmp("--all-progress", arg)) {
2041 progress = 2;
2042 continue;
2043 }
2044 if (!strcmp("-q", arg)) {
2045 progress = 0;
2046 continue;
2047 }
2048 if (!strcmp("--no-reuse-delta", arg)) {
2049 no_reuse_delta = 1;
2050 continue;
2051 }
2052 if (!strcmp("--no-reuse-object", arg)) {
2053 no_reuse_object = no_reuse_delta = 1;
2054 continue;
2055 }
2056 if (!strcmp("--delta-base-offset", arg)) {
2057 allow_ofs_delta = 1;
2058 continue;
2059 }
2060 if (!strcmp("--stdout", arg)) {
2061 pack_to_stdout = 1;
2062 continue;
2063 }
2064 if (!strcmp("--revs", arg)) {
2065 use_internal_rev_list = 1;
2066 continue;
2067 }
2068 if (!strcmp("--keep-unreachable", arg)) {
2069 keep_unreachable = 1;
2070 continue;
2071 }
2072 if (!strcmp("--unpacked", arg) ||
2073 !prefixcmp(arg, "--unpacked=") ||
2074 !strcmp("--reflog", arg) ||
2075 !strcmp("--all", arg)) {
2076 use_internal_rev_list = 1;
2077 if (rp_ac >= rp_ac_alloc - 1) {
2078 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2079 rp_av = xrealloc(rp_av,
2080 rp_ac_alloc * sizeof(*rp_av));
2081 }
2082 rp_av[rp_ac++] = arg;
2083 continue;
2084 }
2085 if (!strcmp("--thin", arg)) {
2086 use_internal_rev_list = 1;
2087 thin = 1;
2088 rp_av[1] = "--objects-edge";
2089 continue;
2090 }
2091 if (!prefixcmp(arg, "--index-version=")) {
2092 char *c;
2093 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2094 if (pack_idx_default_version > 2)
2095 die("bad %s", arg);
2096 if (*c == ',')
2097 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2098 if (*c || pack_idx_off32_limit & 0x80000000)
2099 die("bad %s", arg);
2100 continue;
2101 }
2102 usage(pack_usage);
2103 }
2104
2105 /* Traditionally "pack-objects [options] base extra" failed;
2106 * we would however want to take refs parameter that would
2107 * have been given to upstream rev-list ourselves, which means
2108 * we somehow want to say what the base name is. So the
2109 * syntax would be:
2110 *
2111 * pack-objects [options] base <refs...>
2112 *
2113 * in other words, we would treat the first non-option as the
2114 * base_name and send everything else to the internal revision
2115 * walker.
2116 */
2117
2118 if (!pack_to_stdout)
2119 base_name = argv[i++];
2120
2121 if (pack_to_stdout != !base_name)
2122 usage(pack_usage);
2123
2124 if (pack_to_stdout && pack_size_limit)
2125 die("--max-pack-size cannot be used to build a pack for transfer.");
2126
2127 if (!pack_to_stdout && thin)
2128 die("--thin cannot be used to build an indexable pack.");
2129
2130 prepare_packed_git();
2131
2132 if (progress)
2133 progress_state = start_progress("Counting objects", 0);
2134 if (!use_internal_rev_list)
2135 read_object_list_from_stdin();
2136 else {
2137 rp_av[rp_ac] = NULL;
2138 get_object_list(rp_ac, rp_av);
2139 }
2140 stop_progress(&progress_state);
2141
2142 if (non_empty && !nr_result)
2143 return 0;
2144 if (nr_result)
2145 prepare_pack(window, depth);
2146 write_pack_file();
2147 if (progress)
2148 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2149 written, written_delta, reused, reused_delta);
2150 return 0;
2151}