t4200: use cut instead of sed
[git/peff.git] / builtin-pack-objects.c
blobf504cff7566159332c89f830a871da0ea05868fe
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "progress.h"
18 #include "refs.h"
20 #ifdef THREADED_DELTA_SEARCH
21 #include "thread-utils.h"
22 #include <pthread.h>
23 #endif
25 static const char pack_usage[] = "\
26 git-pack-objects [{ -q | --progress | --all-progress }] \n\
27 [--max-pack-size=N] [--local] [--incremental] \n\
28 [--window=N] [--window-memory=N] [--depth=N] \n\
29 [--no-reuse-delta] [--no-reuse-object] [--delta-base-offset] \n\
30 [--threads=N] [--non-empty] [--revs [--unpacked | --all]*] [--reflog] \n\
31 [--stdout | base-name] [--include-tag] [--keep-unreachable] \n\
32 [<ref-list | <object-list]";
34 struct object_entry {
35 struct pack_idx_entry idx;
36 unsigned long size; /* uncompressed size */
37 struct packed_git *in_pack; /* already in pack */
38 off_t in_pack_offset;
39 struct object_entry *delta; /* delta base object */
40 struct object_entry *delta_child; /* deltified objects who bases me */
41 struct object_entry *delta_sibling; /* other deltified objects who
42 * uses the same base as me
44 void *delta_data; /* cached delta (uncompressed) */
45 unsigned long delta_size; /* delta data size (uncompressed) */
46 unsigned int hash; /* name hint hash */
47 enum object_type type;
48 enum object_type in_pack_type; /* could be delta */
49 unsigned char in_pack_header_size;
50 unsigned char preferred_base; /* we do not pack this, but is available
51 * to be used as the base object to delta
52 * objects against.
54 unsigned char no_try_delta;
58 * Objects we are going to pack are collected in objects array (dynamically
59 * expanded). nr_objects & nr_alloc controls this array. They are stored
60 * in the order we see -- typically rev-list --objects order that gives us
61 * nice "minimum seek" order.
63 static struct object_entry *objects;
64 static struct pack_idx_entry **written_list;
65 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
67 static int non_empty;
68 static int no_reuse_delta, no_reuse_object, keep_unreachable, include_tag;
69 static int local;
70 static int incremental;
71 static int allow_ofs_delta;
72 static const char *base_name;
73 static int progress = 1;
74 static int window = 10;
75 static uint32_t pack_size_limit, pack_size_limit_cfg;
76 static int depth = 50;
77 static int delta_search_threads = 1;
78 static int pack_to_stdout;
79 static int num_preferred_base;
80 static struct progress *progress_state;
81 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
82 static int pack_compression_seen;
84 static unsigned long delta_cache_size = 0;
85 static unsigned long max_delta_cache_size = 0;
86 static unsigned long cache_max_small_delta_size = 1000;
88 static unsigned long window_memory_limit = 0;
91 * The object names in objects array are hashed with this hashtable,
92 * to help looking up the entry by object name.
93 * This hashtable is built after all the objects are seen.
95 static int *object_ix;
96 static int object_ix_hashsz;
99 * stats
101 static uint32_t written, written_delta;
102 static uint32_t reused, reused_delta;
105 static void *delta_against(void *buf, unsigned long size, struct object_entry *entry)
107 unsigned long othersize, delta_size;
108 enum object_type type;
109 void *otherbuf = read_sha1_file(entry->delta->idx.sha1, &type, &othersize);
110 void *delta_buf;
112 if (!otherbuf)
113 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
114 delta_buf = diff_delta(otherbuf, othersize,
115 buf, size, &delta_size, 0);
116 if (!delta_buf || delta_size != entry->delta_size)
117 die("delta size changed");
118 free(buf);
119 free(otherbuf);
120 return delta_buf;
124 * The per-object header is a pretty dense thing, which is
125 * - first byte: low four bits are "size", then three bits of "type",
126 * and the high bit is "size continues".
127 * - each byte afterwards: low seven bits are size continuation,
128 * with the high bit being "size continues"
130 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
132 int n = 1;
133 unsigned char c;
135 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
136 die("bad type %d", type);
138 c = (type << 4) | (size & 15);
139 size >>= 4;
140 while (size) {
141 *hdr++ = c | 0x80;
142 c = size & 0x7f;
143 size >>= 7;
144 n++;
146 *hdr = c;
147 return n;
151 * we are going to reuse the existing object data as is. make
152 * sure it is not corrupt.
154 static int check_pack_inflate(struct packed_git *p,
155 struct pack_window **w_curs,
156 off_t offset,
157 off_t len,
158 unsigned long expect)
160 z_stream stream;
161 unsigned char fakebuf[4096], *in;
162 int st;
164 memset(&stream, 0, sizeof(stream));
165 inflateInit(&stream);
166 do {
167 in = use_pack(p, w_curs, offset, &stream.avail_in);
168 stream.next_in = in;
169 stream.next_out = fakebuf;
170 stream.avail_out = sizeof(fakebuf);
171 st = inflate(&stream, Z_FINISH);
172 offset += stream.next_in - in;
173 } while (st == Z_OK || st == Z_BUF_ERROR);
174 inflateEnd(&stream);
175 return (st == Z_STREAM_END &&
176 stream.total_out == expect &&
177 stream.total_in == len) ? 0 : -1;
180 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
181 off_t offset, off_t len, unsigned int nr)
183 const uint32_t *index_crc;
184 uint32_t data_crc = crc32(0, Z_NULL, 0);
186 do {
187 unsigned int avail;
188 void *data = use_pack(p, w_curs, offset, &avail);
189 if (avail > len)
190 avail = len;
191 data_crc = crc32(data_crc, data, avail);
192 offset += avail;
193 len -= avail;
194 } while (len);
196 index_crc = p->index_data;
197 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
199 return data_crc != ntohl(*index_crc);
202 static void copy_pack_data(struct sha1file *f,
203 struct packed_git *p,
204 struct pack_window **w_curs,
205 off_t offset,
206 off_t len)
208 unsigned char *in;
209 unsigned int avail;
211 while (len) {
212 in = use_pack(p, w_curs, offset, &avail);
213 if (avail > len)
214 avail = (unsigned int)len;
215 sha1write(f, in, avail);
216 offset += avail;
217 len -= avail;
221 static unsigned long write_object(struct sha1file *f,
222 struct object_entry *entry,
223 off_t write_offset)
225 unsigned long size;
226 enum object_type type;
227 void *buf;
228 unsigned char header[10];
229 unsigned char dheader[10];
230 unsigned hdrlen;
231 off_t datalen;
232 enum object_type obj_type;
233 int to_reuse = 0;
234 /* write limit if limited packsize and not first object */
235 unsigned long limit = pack_size_limit && nr_written ?
236 pack_size_limit - write_offset : 0;
237 /* no if no delta */
238 int usable_delta = !entry->delta ? 0 :
239 /* yes if unlimited packfile */
240 !pack_size_limit ? 1 :
241 /* no if base written to previous pack */
242 entry->delta->idx.offset == (off_t)-1 ? 0 :
243 /* otherwise double-check written to this
244 * pack, like we do below
246 entry->delta->idx.offset ? 1 : 0;
248 if (!pack_to_stdout)
249 crc32_begin(f);
251 obj_type = entry->type;
252 if (no_reuse_object)
253 to_reuse = 0; /* explicit */
254 else if (!entry->in_pack)
255 to_reuse = 0; /* can't reuse what we don't have */
256 else if (obj_type == OBJ_REF_DELTA || obj_type == OBJ_OFS_DELTA)
257 /* check_object() decided it for us ... */
258 to_reuse = usable_delta;
259 /* ... but pack split may override that */
260 else if (obj_type != entry->in_pack_type)
261 to_reuse = 0; /* pack has delta which is unusable */
262 else if (entry->delta)
263 to_reuse = 0; /* we want to pack afresh */
264 else
265 to_reuse = 1; /* we have it in-pack undeltified,
266 * and we do not need to deltify it.
269 if (!to_reuse) {
270 z_stream stream;
271 unsigned long maxsize;
272 void *out;
273 if (!usable_delta) {
274 buf = read_sha1_file(entry->idx.sha1, &obj_type, &size);
275 if (!buf)
276 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
277 } else if (entry->delta_data) {
278 size = entry->delta_size;
279 buf = entry->delta_data;
280 entry->delta_data = NULL;
281 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
282 OBJ_OFS_DELTA : OBJ_REF_DELTA;
283 } else {
284 buf = read_sha1_file(entry->idx.sha1, &type, &size);
285 if (!buf)
286 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
287 buf = delta_against(buf, size, entry);
288 size = entry->delta_size;
289 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
290 OBJ_OFS_DELTA : OBJ_REF_DELTA;
292 /* compress the data to store and put compressed length in datalen */
293 memset(&stream, 0, sizeof(stream));
294 deflateInit(&stream, pack_compression_level);
295 maxsize = deflateBound(&stream, size);
296 out = xmalloc(maxsize);
297 /* Compress it */
298 stream.next_in = buf;
299 stream.avail_in = size;
300 stream.next_out = out;
301 stream.avail_out = maxsize;
302 while (deflate(&stream, Z_FINISH) == Z_OK)
303 /* nothing */;
304 deflateEnd(&stream);
305 datalen = stream.total_out;
308 * The object header is a byte of 'type' followed by zero or
309 * more bytes of length.
311 hdrlen = encode_header(obj_type, size, header);
313 if (obj_type == OBJ_OFS_DELTA) {
315 * Deltas with relative base contain an additional
316 * encoding of the relative offset for the delta
317 * base from this object's position in the pack.
319 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
320 unsigned pos = sizeof(dheader) - 1;
321 dheader[pos] = ofs & 127;
322 while (ofs >>= 7)
323 dheader[--pos] = 128 | (--ofs & 127);
324 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
325 free(out);
326 free(buf);
327 return 0;
329 sha1write(f, header, hdrlen);
330 sha1write(f, dheader + pos, sizeof(dheader) - pos);
331 hdrlen += sizeof(dheader) - pos;
332 } else if (obj_type == OBJ_REF_DELTA) {
334 * Deltas with a base reference contain
335 * an additional 20 bytes for the base sha1.
337 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
338 free(out);
339 free(buf);
340 return 0;
342 sha1write(f, header, hdrlen);
343 sha1write(f, entry->delta->idx.sha1, 20);
344 hdrlen += 20;
345 } else {
346 if (limit && hdrlen + datalen + 20 >= limit) {
347 free(out);
348 free(buf);
349 return 0;
351 sha1write(f, header, hdrlen);
353 sha1write(f, out, datalen);
354 free(out);
355 free(buf);
357 else {
358 struct packed_git *p = entry->in_pack;
359 struct pack_window *w_curs = NULL;
360 struct revindex_entry *revidx;
361 off_t offset;
363 if (entry->delta) {
364 obj_type = (allow_ofs_delta && entry->delta->idx.offset) ?
365 OBJ_OFS_DELTA : OBJ_REF_DELTA;
366 reused_delta++;
368 hdrlen = encode_header(obj_type, entry->size, header);
369 offset = entry->in_pack_offset;
370 revidx = find_pack_revindex(p, offset);
371 datalen = revidx[1].offset - offset;
372 if (!pack_to_stdout && p->index_version > 1 &&
373 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
374 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
375 offset += entry->in_pack_header_size;
376 datalen -= entry->in_pack_header_size;
377 if (obj_type == OBJ_OFS_DELTA) {
378 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
379 unsigned pos = sizeof(dheader) - 1;
380 dheader[pos] = ofs & 127;
381 while (ofs >>= 7)
382 dheader[--pos] = 128 | (--ofs & 127);
383 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
384 return 0;
385 sha1write(f, header, hdrlen);
386 sha1write(f, dheader + pos, sizeof(dheader) - pos);
387 hdrlen += sizeof(dheader) - pos;
388 } else if (obj_type == OBJ_REF_DELTA) {
389 if (limit && hdrlen + 20 + datalen + 20 >= limit)
390 return 0;
391 sha1write(f, header, hdrlen);
392 sha1write(f, entry->delta->idx.sha1, 20);
393 hdrlen += 20;
394 } else {
395 if (limit && hdrlen + datalen + 20 >= limit)
396 return 0;
397 sha1write(f, header, hdrlen);
400 if (!pack_to_stdout && p->index_version == 1 &&
401 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
402 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
403 copy_pack_data(f, p, &w_curs, offset, datalen);
404 unuse_pack(&w_curs);
405 reused++;
407 if (usable_delta)
408 written_delta++;
409 written++;
410 if (!pack_to_stdout)
411 entry->idx.crc32 = crc32_end(f);
412 return hdrlen + datalen;
415 static off_t write_one(struct sha1file *f,
416 struct object_entry *e,
417 off_t offset)
419 unsigned long size;
421 /* offset is non zero if object is written already. */
422 if (e->idx.offset || e->preferred_base)
423 return offset;
425 /* if we are deltified, write out base object first. */
426 if (e->delta) {
427 offset = write_one(f, e->delta, offset);
428 if (!offset)
429 return 0;
432 e->idx.offset = offset;
433 size = write_object(f, e, offset);
434 if (!size) {
435 e->idx.offset = 0;
436 return 0;
438 written_list[nr_written++] = &e->idx;
440 /* make sure off_t is sufficiently large not to wrap */
441 if (offset > offset + size)
442 die("pack too large for current definition of off_t");
443 return offset + size;
446 /* forward declaration for write_pack_file */
447 static int adjust_perm(const char *path, mode_t mode);
449 static void write_pack_file(void)
451 uint32_t i = 0, j;
452 struct sha1file *f;
453 off_t offset, offset_one, last_obj_offset = 0;
454 struct pack_header hdr;
455 int do_progress = progress >> pack_to_stdout;
456 uint32_t nr_remaining = nr_result;
458 if (do_progress)
459 progress_state = start_progress("Writing objects", nr_result);
460 written_list = xmalloc(nr_objects * sizeof(*written_list));
462 do {
463 unsigned char sha1[20];
464 char *pack_tmp_name = NULL;
466 if (pack_to_stdout) {
467 f = sha1fd_throughput(1, "<stdout>", progress_state);
468 } else {
469 char tmpname[PATH_MAX];
470 int fd;
471 snprintf(tmpname, sizeof(tmpname),
472 "%s/tmp_pack_XXXXXX", get_object_directory());
473 fd = xmkstemp(tmpname);
474 pack_tmp_name = xstrdup(tmpname);
475 f = sha1fd(fd, pack_tmp_name);
478 hdr.hdr_signature = htonl(PACK_SIGNATURE);
479 hdr.hdr_version = htonl(PACK_VERSION);
480 hdr.hdr_entries = htonl(nr_remaining);
481 sha1write(f, &hdr, sizeof(hdr));
482 offset = sizeof(hdr);
483 nr_written = 0;
484 for (; i < nr_objects; i++) {
485 last_obj_offset = offset;
486 offset_one = write_one(f, objects + i, offset);
487 if (!offset_one)
488 break;
489 offset = offset_one;
490 display_progress(progress_state, written);
494 * Did we write the wrong # entries in the header?
495 * If so, rewrite it like in fast-import
497 if (pack_to_stdout || nr_written == nr_remaining) {
498 sha1close(f, sha1, 1);
499 } else {
500 int fd = sha1close(f, NULL, 0);
501 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
502 close(fd);
505 if (!pack_to_stdout) {
506 mode_t mode = umask(0);
507 char *idx_tmp_name, tmpname[PATH_MAX];
509 umask(mode);
510 mode = 0444 & ~mode;
512 idx_tmp_name = write_idx_file(NULL, written_list,
513 nr_written, sha1);
514 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
515 base_name, sha1_to_hex(sha1));
516 if (adjust_perm(pack_tmp_name, mode))
517 die("unable to make temporary pack file readable: %s",
518 strerror(errno));
519 if (rename(pack_tmp_name, tmpname))
520 die("unable to rename temporary pack file: %s",
521 strerror(errno));
522 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
523 base_name, sha1_to_hex(sha1));
524 if (adjust_perm(idx_tmp_name, mode))
525 die("unable to make temporary index file readable: %s",
526 strerror(errno));
527 if (rename(idx_tmp_name, tmpname))
528 die("unable to rename temporary index file: %s",
529 strerror(errno));
530 free(idx_tmp_name);
531 free(pack_tmp_name);
532 puts(sha1_to_hex(sha1));
535 /* mark written objects as written to previous pack */
536 for (j = 0; j < nr_written; j++) {
537 written_list[j]->offset = (off_t)-1;
539 nr_remaining -= nr_written;
540 } while (nr_remaining && i < nr_objects);
542 free(written_list);
543 stop_progress(&progress_state);
544 if (written != nr_result)
545 die("wrote %u objects while expecting %u", written, nr_result);
547 * We have scanned through [0 ... i). Since we have written
548 * the correct number of objects, the remaining [i ... nr_objects)
549 * items must be either already written (due to out-of-order delta base)
550 * or a preferred base. Count those which are neither and complain if any.
552 for (j = 0; i < nr_objects; i++) {
553 struct object_entry *e = objects + i;
554 j += !e->idx.offset && !e->preferred_base;
556 if (j)
557 die("wrote %u objects as expected but %u unwritten", written, j);
560 static int locate_object_entry_hash(const unsigned char *sha1)
562 int i;
563 unsigned int ui;
564 memcpy(&ui, sha1, sizeof(unsigned int));
565 i = ui % object_ix_hashsz;
566 while (0 < object_ix[i]) {
567 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
568 return i;
569 if (++i == object_ix_hashsz)
570 i = 0;
572 return -1 - i;
575 static struct object_entry *locate_object_entry(const unsigned char *sha1)
577 int i;
579 if (!object_ix_hashsz)
580 return NULL;
582 i = locate_object_entry_hash(sha1);
583 if (0 <= i)
584 return &objects[object_ix[i]-1];
585 return NULL;
588 static void rehash_objects(void)
590 uint32_t i;
591 struct object_entry *oe;
593 object_ix_hashsz = nr_objects * 3;
594 if (object_ix_hashsz < 1024)
595 object_ix_hashsz = 1024;
596 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
597 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
598 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
599 int ix = locate_object_entry_hash(oe->idx.sha1);
600 if (0 <= ix)
601 continue;
602 ix = -1 - ix;
603 object_ix[ix] = i + 1;
607 static unsigned name_hash(const char *name)
609 unsigned char c;
610 unsigned hash = 0;
612 if (!name)
613 return 0;
616 * This effectively just creates a sortable number from the
617 * last sixteen non-whitespace characters. Last characters
618 * count "most", so things that end in ".c" sort together.
620 while ((c = *name++) != 0) {
621 if (isspace(c))
622 continue;
623 hash = (hash >> 2) + (c << 24);
625 return hash;
628 static void setup_delta_attr_check(struct git_attr_check *check)
630 static struct git_attr *attr_delta;
632 if (!attr_delta)
633 attr_delta = git_attr("delta", 5);
635 check[0].attr = attr_delta;
638 static int no_try_delta(const char *path)
640 struct git_attr_check check[1];
642 setup_delta_attr_check(check);
643 if (git_checkattr(path, ARRAY_SIZE(check), check))
644 return 0;
645 if (ATTR_FALSE(check->value))
646 return 1;
647 return 0;
650 static int add_object_entry(const unsigned char *sha1, enum object_type type,
651 const char *name, int exclude)
653 struct object_entry *entry;
654 struct packed_git *p, *found_pack = NULL;
655 off_t found_offset = 0;
656 int ix;
657 unsigned hash = name_hash(name);
659 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
660 if (ix >= 0) {
661 if (exclude) {
662 entry = objects + object_ix[ix] - 1;
663 if (!entry->preferred_base)
664 nr_result--;
665 entry->preferred_base = 1;
667 return 0;
670 for (p = packed_git; p; p = p->next) {
671 off_t offset = find_pack_entry_one(sha1, p);
672 if (offset) {
673 if (!found_pack) {
674 found_offset = offset;
675 found_pack = p;
677 if (exclude)
678 break;
679 if (incremental)
680 return 0;
681 if (local && !p->pack_local)
682 return 0;
686 if (nr_objects >= nr_alloc) {
687 nr_alloc = (nr_alloc + 1024) * 3 / 2;
688 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
691 entry = objects + nr_objects++;
692 memset(entry, 0, sizeof(*entry));
693 hashcpy(entry->idx.sha1, sha1);
694 entry->hash = hash;
695 if (type)
696 entry->type = type;
697 if (exclude)
698 entry->preferred_base = 1;
699 else
700 nr_result++;
701 if (found_pack) {
702 entry->in_pack = found_pack;
703 entry->in_pack_offset = found_offset;
706 if (object_ix_hashsz * 3 <= nr_objects * 4)
707 rehash_objects();
708 else
709 object_ix[-1 - ix] = nr_objects;
711 display_progress(progress_state, nr_objects);
713 if (name && no_try_delta(name))
714 entry->no_try_delta = 1;
716 return 1;
719 struct pbase_tree_cache {
720 unsigned char sha1[20];
721 int ref;
722 int temporary;
723 void *tree_data;
724 unsigned long tree_size;
727 static struct pbase_tree_cache *(pbase_tree_cache[256]);
728 static int pbase_tree_cache_ix(const unsigned char *sha1)
730 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
732 static int pbase_tree_cache_ix_incr(int ix)
734 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
737 static struct pbase_tree {
738 struct pbase_tree *next;
739 /* This is a phony "cache" entry; we are not
740 * going to evict it nor find it through _get()
741 * mechanism -- this is for the toplevel node that
742 * would almost always change with any commit.
744 struct pbase_tree_cache pcache;
745 } *pbase_tree;
747 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
749 struct pbase_tree_cache *ent, *nent;
750 void *data;
751 unsigned long size;
752 enum object_type type;
753 int neigh;
754 int my_ix = pbase_tree_cache_ix(sha1);
755 int available_ix = -1;
757 /* pbase-tree-cache acts as a limited hashtable.
758 * your object will be found at your index or within a few
759 * slots after that slot if it is cached.
761 for (neigh = 0; neigh < 8; neigh++) {
762 ent = pbase_tree_cache[my_ix];
763 if (ent && !hashcmp(ent->sha1, sha1)) {
764 ent->ref++;
765 return ent;
767 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
768 ((0 <= available_ix) &&
769 (!ent && pbase_tree_cache[available_ix])))
770 available_ix = my_ix;
771 if (!ent)
772 break;
773 my_ix = pbase_tree_cache_ix_incr(my_ix);
776 /* Did not find one. Either we got a bogus request or
777 * we need to read and perhaps cache.
779 data = read_sha1_file(sha1, &type, &size);
780 if (!data)
781 return NULL;
782 if (type != OBJ_TREE) {
783 free(data);
784 return NULL;
787 /* We need to either cache or return a throwaway copy */
789 if (available_ix < 0)
790 ent = NULL;
791 else {
792 ent = pbase_tree_cache[available_ix];
793 my_ix = available_ix;
796 if (!ent) {
797 nent = xmalloc(sizeof(*nent));
798 nent->temporary = (available_ix < 0);
800 else {
801 /* evict and reuse */
802 free(ent->tree_data);
803 nent = ent;
805 hashcpy(nent->sha1, sha1);
806 nent->tree_data = data;
807 nent->tree_size = size;
808 nent->ref = 1;
809 if (!nent->temporary)
810 pbase_tree_cache[my_ix] = nent;
811 return nent;
814 static void pbase_tree_put(struct pbase_tree_cache *cache)
816 if (!cache->temporary) {
817 cache->ref--;
818 return;
820 free(cache->tree_data);
821 free(cache);
824 static int name_cmp_len(const char *name)
826 int i;
827 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
829 return i;
832 static void add_pbase_object(struct tree_desc *tree,
833 const char *name,
834 int cmplen,
835 const char *fullname)
837 struct name_entry entry;
838 int cmp;
840 while (tree_entry(tree,&entry)) {
841 if (S_ISGITLINK(entry.mode))
842 continue;
843 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
844 memcmp(name, entry.path, cmplen);
845 if (cmp > 0)
846 continue;
847 if (cmp < 0)
848 return;
849 if (name[cmplen] != '/') {
850 add_object_entry(entry.sha1,
851 object_type(entry.mode),
852 fullname, 1);
853 return;
855 if (S_ISDIR(entry.mode)) {
856 struct tree_desc sub;
857 struct pbase_tree_cache *tree;
858 const char *down = name+cmplen+1;
859 int downlen = name_cmp_len(down);
861 tree = pbase_tree_get(entry.sha1);
862 if (!tree)
863 return;
864 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
866 add_pbase_object(&sub, down, downlen, fullname);
867 pbase_tree_put(tree);
872 static unsigned *done_pbase_paths;
873 static int done_pbase_paths_num;
874 static int done_pbase_paths_alloc;
875 static int done_pbase_path_pos(unsigned hash)
877 int lo = 0;
878 int hi = done_pbase_paths_num;
879 while (lo < hi) {
880 int mi = (hi + lo) / 2;
881 if (done_pbase_paths[mi] == hash)
882 return mi;
883 if (done_pbase_paths[mi] < hash)
884 hi = mi;
885 else
886 lo = mi + 1;
888 return -lo-1;
891 static int check_pbase_path(unsigned hash)
893 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
894 if (0 <= pos)
895 return 1;
896 pos = -pos - 1;
897 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
898 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
899 done_pbase_paths = xrealloc(done_pbase_paths,
900 done_pbase_paths_alloc *
901 sizeof(unsigned));
903 done_pbase_paths_num++;
904 if (pos < done_pbase_paths_num)
905 memmove(done_pbase_paths + pos + 1,
906 done_pbase_paths + pos,
907 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
908 done_pbase_paths[pos] = hash;
909 return 0;
912 static void add_preferred_base_object(const char *name)
914 struct pbase_tree *it;
915 int cmplen;
916 unsigned hash = name_hash(name);
918 if (!num_preferred_base || check_pbase_path(hash))
919 return;
921 cmplen = name_cmp_len(name);
922 for (it = pbase_tree; it; it = it->next) {
923 if (cmplen == 0) {
924 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
926 else {
927 struct tree_desc tree;
928 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
929 add_pbase_object(&tree, name, cmplen, name);
934 static void add_preferred_base(unsigned char *sha1)
936 struct pbase_tree *it;
937 void *data;
938 unsigned long size;
939 unsigned char tree_sha1[20];
941 if (window <= num_preferred_base++)
942 return;
944 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
945 if (!data)
946 return;
948 for (it = pbase_tree; it; it = it->next) {
949 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
950 free(data);
951 return;
955 it = xcalloc(1, sizeof(*it));
956 it->next = pbase_tree;
957 pbase_tree = it;
959 hashcpy(it->pcache.sha1, tree_sha1);
960 it->pcache.tree_data = data;
961 it->pcache.tree_size = size;
964 static void check_object(struct object_entry *entry)
966 if (entry->in_pack) {
967 struct packed_git *p = entry->in_pack;
968 struct pack_window *w_curs = NULL;
969 const unsigned char *base_ref = NULL;
970 struct object_entry *base_entry;
971 unsigned long used, used_0;
972 unsigned int avail;
973 off_t ofs;
974 unsigned char *buf, c;
976 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
979 * We want in_pack_type even if we do not reuse delta
980 * since non-delta representations could still be reused.
982 used = unpack_object_header_gently(buf, avail,
983 &entry->in_pack_type,
984 &entry->size);
987 * Determine if this is a delta and if so whether we can
988 * reuse it or not. Otherwise let's find out as cheaply as
989 * possible what the actual type and size for this object is.
991 switch (entry->in_pack_type) {
992 default:
993 /* Not a delta hence we've already got all we need. */
994 entry->type = entry->in_pack_type;
995 entry->in_pack_header_size = used;
996 unuse_pack(&w_curs);
997 return;
998 case OBJ_REF_DELTA:
999 if (!no_reuse_delta && !entry->preferred_base)
1000 base_ref = use_pack(p, &w_curs,
1001 entry->in_pack_offset + used, NULL);
1002 entry->in_pack_header_size = used + 20;
1003 break;
1004 case OBJ_OFS_DELTA:
1005 buf = use_pack(p, &w_curs,
1006 entry->in_pack_offset + used, NULL);
1007 used_0 = 0;
1008 c = buf[used_0++];
1009 ofs = c & 127;
1010 while (c & 128) {
1011 ofs += 1;
1012 if (!ofs || MSB(ofs, 7))
1013 die("delta base offset overflow in pack for %s",
1014 sha1_to_hex(entry->idx.sha1));
1015 c = buf[used_0++];
1016 ofs = (ofs << 7) + (c & 127);
1018 if (ofs >= entry->in_pack_offset)
1019 die("delta base offset out of bound for %s",
1020 sha1_to_hex(entry->idx.sha1));
1021 ofs = entry->in_pack_offset - ofs;
1022 if (!no_reuse_delta && !entry->preferred_base) {
1023 struct revindex_entry *revidx;
1024 revidx = find_pack_revindex(p, ofs);
1025 base_ref = nth_packed_object_sha1(p, revidx->nr);
1027 entry->in_pack_header_size = used + used_0;
1028 break;
1031 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1033 * If base_ref was set above that means we wish to
1034 * reuse delta data, and we even found that base
1035 * in the list of objects we want to pack. Goodie!
1037 * Depth value does not matter - find_deltas() will
1038 * never consider reused delta as the base object to
1039 * deltify other objects against, in order to avoid
1040 * circular deltas.
1042 entry->type = entry->in_pack_type;
1043 entry->delta = base_entry;
1044 entry->delta_sibling = base_entry->delta_child;
1045 base_entry->delta_child = entry;
1046 unuse_pack(&w_curs);
1047 return;
1050 if (entry->type) {
1052 * This must be a delta and we already know what the
1053 * final object type is. Let's extract the actual
1054 * object size from the delta header.
1056 entry->size = get_size_from_delta(p, &w_curs,
1057 entry->in_pack_offset + entry->in_pack_header_size);
1058 unuse_pack(&w_curs);
1059 return;
1063 * No choice but to fall back to the recursive delta walk
1064 * with sha1_object_info() to find about the object type
1065 * at this point...
1067 unuse_pack(&w_curs);
1070 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1071 if (entry->type < 0)
1072 die("unable to get type of object %s",
1073 sha1_to_hex(entry->idx.sha1));
1076 static int pack_offset_sort(const void *_a, const void *_b)
1078 const struct object_entry *a = *(struct object_entry **)_a;
1079 const struct object_entry *b = *(struct object_entry **)_b;
1081 /* avoid filesystem trashing with loose objects */
1082 if (!a->in_pack && !b->in_pack)
1083 return hashcmp(a->idx.sha1, b->idx.sha1);
1085 if (a->in_pack < b->in_pack)
1086 return -1;
1087 if (a->in_pack > b->in_pack)
1088 return 1;
1089 return a->in_pack_offset < b->in_pack_offset ? -1 :
1090 (a->in_pack_offset > b->in_pack_offset);
1093 static void get_object_details(void)
1095 uint32_t i;
1096 struct object_entry **sorted_by_offset;
1098 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1099 for (i = 0; i < nr_objects; i++)
1100 sorted_by_offset[i] = objects + i;
1101 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1103 init_pack_revindex();
1105 for (i = 0; i < nr_objects; i++)
1106 check_object(sorted_by_offset[i]);
1108 free(sorted_by_offset);
1112 * We search for deltas in a list sorted by type, by filename hash, and then
1113 * by size, so that we see progressively smaller and smaller files.
1114 * That's because we prefer deltas to be from the bigger file
1115 * to the smaller -- deletes are potentially cheaper, but perhaps
1116 * more importantly, the bigger file is likely the more recent
1117 * one. The deepest deltas are therefore the oldest objects which are
1118 * less susceptible to be accessed often.
1120 static int type_size_sort(const void *_a, const void *_b)
1122 const struct object_entry *a = *(struct object_entry **)_a;
1123 const struct object_entry *b = *(struct object_entry **)_b;
1125 if (a->type > b->type)
1126 return -1;
1127 if (a->type < b->type)
1128 return 1;
1129 if (a->hash > b->hash)
1130 return -1;
1131 if (a->hash < b->hash)
1132 return 1;
1133 if (a->preferred_base > b->preferred_base)
1134 return -1;
1135 if (a->preferred_base < b->preferred_base)
1136 return 1;
1137 if (a->size > b->size)
1138 return -1;
1139 if (a->size < b->size)
1140 return 1;
1141 return a < b ? -1 : (a > b); /* newest first */
1144 struct unpacked {
1145 struct object_entry *entry;
1146 void *data;
1147 struct delta_index *index;
1148 unsigned depth;
1151 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1152 unsigned long delta_size)
1154 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1155 return 0;
1157 if (delta_size < cache_max_small_delta_size)
1158 return 1;
1160 /* cache delta, if objects are large enough compared to delta size */
1161 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1162 return 1;
1164 return 0;
1167 #ifdef THREADED_DELTA_SEARCH
1169 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1170 #define read_lock() pthread_mutex_lock(&read_mutex)
1171 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1173 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1174 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1175 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1177 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1178 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1179 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1181 #else
1183 #define read_lock() (void)0
1184 #define read_unlock() (void)0
1185 #define cache_lock() (void)0
1186 #define cache_unlock() (void)0
1187 #define progress_lock() (void)0
1188 #define progress_unlock() (void)0
1190 #endif
1192 static int try_delta(struct unpacked *trg, struct unpacked *src,
1193 unsigned max_depth, unsigned long *mem_usage)
1195 struct object_entry *trg_entry = trg->entry;
1196 struct object_entry *src_entry = src->entry;
1197 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1198 unsigned ref_depth;
1199 enum object_type type;
1200 void *delta_buf;
1202 /* Don't bother doing diffs between different types */
1203 if (trg_entry->type != src_entry->type)
1204 return -1;
1207 * We do not bother to try a delta that we discarded
1208 * on an earlier try, but only when reusing delta data.
1210 if (!no_reuse_delta && trg_entry->in_pack &&
1211 trg_entry->in_pack == src_entry->in_pack &&
1212 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1213 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1214 return 0;
1216 /* Let's not bust the allowed depth. */
1217 if (src->depth >= max_depth)
1218 return 0;
1220 /* Now some size filtering heuristics. */
1221 trg_size = trg_entry->size;
1222 if (!trg_entry->delta) {
1223 max_size = trg_size/2 - 20;
1224 ref_depth = 1;
1225 } else {
1226 max_size = trg_entry->delta_size;
1227 ref_depth = trg->depth;
1229 max_size = max_size * (max_depth - src->depth) /
1230 (max_depth - ref_depth + 1);
1231 if (max_size == 0)
1232 return 0;
1233 src_size = src_entry->size;
1234 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1235 if (sizediff >= max_size)
1236 return 0;
1237 if (trg_size < src_size / 32)
1238 return 0;
1240 /* Load data if not already done */
1241 if (!trg->data) {
1242 read_lock();
1243 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1244 read_unlock();
1245 if (!trg->data)
1246 die("object %s cannot be read",
1247 sha1_to_hex(trg_entry->idx.sha1));
1248 if (sz != trg_size)
1249 die("object %s inconsistent object length (%lu vs %lu)",
1250 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1251 *mem_usage += sz;
1253 if (!src->data) {
1254 read_lock();
1255 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1256 read_unlock();
1257 if (!src->data)
1258 die("object %s cannot be read",
1259 sha1_to_hex(src_entry->idx.sha1));
1260 if (sz != src_size)
1261 die("object %s inconsistent object length (%lu vs %lu)",
1262 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1263 *mem_usage += sz;
1265 if (!src->index) {
1266 src->index = create_delta_index(src->data, src_size);
1267 if (!src->index) {
1268 static int warned = 0;
1269 if (!warned++)
1270 warning("suboptimal pack - out of memory");
1271 return 0;
1273 *mem_usage += sizeof_delta_index(src->index);
1276 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1277 if (!delta_buf)
1278 return 0;
1280 if (trg_entry->delta) {
1281 /* Prefer only shallower same-sized deltas. */
1282 if (delta_size == trg_entry->delta_size &&
1283 src->depth + 1 >= trg->depth) {
1284 free(delta_buf);
1285 return 0;
1290 * Handle memory allocation outside of the cache
1291 * accounting lock. Compiler will optimize the strangeness
1292 * away when THREADED_DELTA_SEARCH is not defined.
1294 free(trg_entry->delta_data);
1295 cache_lock();
1296 if (trg_entry->delta_data) {
1297 delta_cache_size -= trg_entry->delta_size;
1298 trg_entry->delta_data = NULL;
1300 if (delta_cacheable(src_size, trg_size, delta_size)) {
1301 delta_cache_size += delta_size;
1302 cache_unlock();
1303 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1304 } else {
1305 cache_unlock();
1306 free(delta_buf);
1309 trg_entry->delta = src_entry;
1310 trg_entry->delta_size = delta_size;
1311 trg->depth = src->depth + 1;
1313 return 1;
1316 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1318 struct object_entry *child = me->delta_child;
1319 unsigned int m = n;
1320 while (child) {
1321 unsigned int c = check_delta_limit(child, n + 1);
1322 if (m < c)
1323 m = c;
1324 child = child->delta_sibling;
1326 return m;
1329 static unsigned long free_unpacked(struct unpacked *n)
1331 unsigned long freed_mem = sizeof_delta_index(n->index);
1332 free_delta_index(n->index);
1333 n->index = NULL;
1334 if (n->data) {
1335 freed_mem += n->entry->size;
1336 free(n->data);
1337 n->data = NULL;
1339 n->entry = NULL;
1340 n->depth = 0;
1341 return freed_mem;
1344 static void find_deltas(struct object_entry **list, unsigned *list_size,
1345 int window, int depth, unsigned *processed)
1347 uint32_t i, idx = 0, count = 0;
1348 unsigned int array_size = window * sizeof(struct unpacked);
1349 struct unpacked *array;
1350 unsigned long mem_usage = 0;
1352 array = xmalloc(array_size);
1353 memset(array, 0, array_size);
1355 for (;;) {
1356 struct object_entry *entry = *list++;
1357 struct unpacked *n = array + idx;
1358 int j, max_depth, best_base = -1;
1360 progress_lock();
1361 if (!*list_size) {
1362 progress_unlock();
1363 break;
1365 (*list_size)--;
1366 if (!entry->preferred_base) {
1367 (*processed)++;
1368 display_progress(progress_state, *processed);
1370 progress_unlock();
1372 mem_usage -= free_unpacked(n);
1373 n->entry = entry;
1375 while (window_memory_limit &&
1376 mem_usage > window_memory_limit &&
1377 count > 1) {
1378 uint32_t tail = (idx + window - count) % window;
1379 mem_usage -= free_unpacked(array + tail);
1380 count--;
1383 /* We do not compute delta to *create* objects we are not
1384 * going to pack.
1386 if (entry->preferred_base)
1387 goto next;
1390 * If the current object is at pack edge, take the depth the
1391 * objects that depend on the current object into account
1392 * otherwise they would become too deep.
1394 max_depth = depth;
1395 if (entry->delta_child) {
1396 max_depth -= check_delta_limit(entry, 0);
1397 if (max_depth <= 0)
1398 goto next;
1401 j = window;
1402 while (--j > 0) {
1403 int ret;
1404 uint32_t other_idx = idx + j;
1405 struct unpacked *m;
1406 if (other_idx >= window)
1407 other_idx -= window;
1408 m = array + other_idx;
1409 if (!m->entry)
1410 break;
1411 ret = try_delta(n, m, max_depth, &mem_usage);
1412 if (ret < 0)
1413 break;
1414 else if (ret > 0)
1415 best_base = other_idx;
1418 /* if we made n a delta, and if n is already at max
1419 * depth, leaving it in the window is pointless. we
1420 * should evict it first.
1422 if (entry->delta && depth <= n->depth)
1423 continue;
1426 * Move the best delta base up in the window, after the
1427 * currently deltified object, to keep it longer. It will
1428 * be the first base object to be attempted next.
1430 if (entry->delta) {
1431 struct unpacked swap = array[best_base];
1432 int dist = (window + idx - best_base) % window;
1433 int dst = best_base;
1434 while (dist--) {
1435 int src = (dst + 1) % window;
1436 array[dst] = array[src];
1437 dst = src;
1439 array[dst] = swap;
1442 next:
1443 idx++;
1444 if (count + 1 < window)
1445 count++;
1446 if (idx >= window)
1447 idx = 0;
1450 for (i = 0; i < window; ++i) {
1451 free_delta_index(array[i].index);
1452 free(array[i].data);
1454 free(array);
1457 #ifdef THREADED_DELTA_SEARCH
1460 * The main thread waits on the condition that (at least) one of the workers
1461 * has stopped working (which is indicated in the .working member of
1462 * struct thread_params).
1463 * When a work thread has completed its work, it sets .working to 0 and
1464 * signals the main thread and waits on the condition that .data_ready
1465 * becomes 1.
1468 struct thread_params {
1469 pthread_t thread;
1470 struct object_entry **list;
1471 unsigned list_size;
1472 unsigned remaining;
1473 int window;
1474 int depth;
1475 int working;
1476 int data_ready;
1477 pthread_mutex_t mutex;
1478 pthread_cond_t cond;
1479 unsigned *processed;
1482 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1484 static void *threaded_find_deltas(void *arg)
1486 struct thread_params *me = arg;
1488 while (me->remaining) {
1489 find_deltas(me->list, &me->remaining,
1490 me->window, me->depth, me->processed);
1492 progress_lock();
1493 me->working = 0;
1494 pthread_cond_signal(&progress_cond);
1495 progress_unlock();
1498 * We must not set ->data_ready before we wait on the
1499 * condition because the main thread may have set it to 1
1500 * before we get here. In order to be sure that new
1501 * work is available if we see 1 in ->data_ready, it
1502 * was initialized to 0 before this thread was spawned
1503 * and we reset it to 0 right away.
1505 pthread_mutex_lock(&me->mutex);
1506 while (!me->data_ready)
1507 pthread_cond_wait(&me->cond, &me->mutex);
1508 me->data_ready = 0;
1509 pthread_mutex_unlock(&me->mutex);
1511 /* leave ->working 1 so that this doesn't get more work assigned */
1512 return NULL;
1515 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1516 int window, int depth, unsigned *processed)
1518 struct thread_params p[delta_search_threads];
1519 int i, ret, active_threads = 0;
1521 if (delta_search_threads <= 1) {
1522 find_deltas(list, &list_size, window, depth, processed);
1523 return;
1526 /* Partition the work amongst work threads. */
1527 for (i = 0; i < delta_search_threads; i++) {
1528 unsigned sub_size = list_size / (delta_search_threads - i);
1530 p[i].window = window;
1531 p[i].depth = depth;
1532 p[i].processed = processed;
1533 p[i].working = 1;
1534 p[i].data_ready = 0;
1536 /* try to split chunks on "path" boundaries */
1537 while (sub_size && sub_size < list_size &&
1538 list[sub_size]->hash &&
1539 list[sub_size]->hash == list[sub_size-1]->hash)
1540 sub_size++;
1542 p[i].list = list;
1543 p[i].list_size = sub_size;
1544 p[i].remaining = sub_size;
1546 list += sub_size;
1547 list_size -= sub_size;
1550 /* Start work threads. */
1551 for (i = 0; i < delta_search_threads; i++) {
1552 if (!p[i].list_size)
1553 continue;
1554 pthread_mutex_init(&p[i].mutex, NULL);
1555 pthread_cond_init(&p[i].cond, NULL);
1556 ret = pthread_create(&p[i].thread, NULL,
1557 threaded_find_deltas, &p[i]);
1558 if (ret)
1559 die("unable to create thread: %s", strerror(ret));
1560 active_threads++;
1564 * Now let's wait for work completion. Each time a thread is done
1565 * with its work, we steal half of the remaining work from the
1566 * thread with the largest number of unprocessed objects and give
1567 * it to that newly idle thread. This ensure good load balancing
1568 * until the remaining object list segments are simply too short
1569 * to be worth splitting anymore.
1571 while (active_threads) {
1572 struct thread_params *target = NULL;
1573 struct thread_params *victim = NULL;
1574 unsigned sub_size = 0;
1576 progress_lock();
1577 for (;;) {
1578 for (i = 0; !target && i < delta_search_threads; i++)
1579 if (!p[i].working)
1580 target = &p[i];
1581 if (target)
1582 break;
1583 pthread_cond_wait(&progress_cond, &progress_mutex);
1586 for (i = 0; i < delta_search_threads; i++)
1587 if (p[i].remaining > 2*window &&
1588 (!victim || victim->remaining < p[i].remaining))
1589 victim = &p[i];
1590 if (victim) {
1591 sub_size = victim->remaining / 2;
1592 list = victim->list + victim->list_size - sub_size;
1593 while (sub_size && list[0]->hash &&
1594 list[0]->hash == list[-1]->hash) {
1595 list++;
1596 sub_size--;
1598 if (!sub_size) {
1600 * It is possible for some "paths" to have
1601 * so many objects that no hash boundary
1602 * might be found. Let's just steal the
1603 * exact half in that case.
1605 sub_size = victim->remaining / 2;
1606 list -= sub_size;
1608 target->list = list;
1609 victim->list_size -= sub_size;
1610 victim->remaining -= sub_size;
1612 target->list_size = sub_size;
1613 target->remaining = sub_size;
1614 target->working = 1;
1615 progress_unlock();
1617 pthread_mutex_lock(&target->mutex);
1618 target->data_ready = 1;
1619 pthread_cond_signal(&target->cond);
1620 pthread_mutex_unlock(&target->mutex);
1622 if (!sub_size) {
1623 pthread_join(target->thread, NULL);
1624 pthread_cond_destroy(&target->cond);
1625 pthread_mutex_destroy(&target->mutex);
1626 active_threads--;
1631 #else
1632 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1633 #endif
1635 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1637 unsigned char peeled[20];
1639 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1640 !peel_ref(path, peeled) && /* peelable? */
1641 !is_null_sha1(peeled) && /* annotated tag? */
1642 locate_object_entry(peeled)) /* object packed? */
1643 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1644 return 0;
1647 static void prepare_pack(int window, int depth)
1649 struct object_entry **delta_list;
1650 uint32_t i, n, nr_deltas;
1652 get_object_details();
1654 if (!nr_objects || !window || !depth)
1655 return;
1657 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1658 nr_deltas = n = 0;
1660 for (i = 0; i < nr_objects; i++) {
1661 struct object_entry *entry = objects + i;
1663 if (entry->delta)
1664 /* This happens if we decided to reuse existing
1665 * delta from a pack. "!no_reuse_delta &&" is implied.
1667 continue;
1669 if (entry->size < 50)
1670 continue;
1672 if (entry->no_try_delta)
1673 continue;
1675 if (!entry->preferred_base)
1676 nr_deltas++;
1678 delta_list[n++] = entry;
1681 if (nr_deltas && n > 1) {
1682 unsigned nr_done = 0;
1683 if (progress)
1684 progress_state = start_progress("Compressing objects",
1685 nr_deltas);
1686 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1687 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1688 stop_progress(&progress_state);
1689 if (nr_done != nr_deltas)
1690 die("inconsistency with delta count");
1692 free(delta_list);
1695 static int git_pack_config(const char *k, const char *v)
1697 if(!strcmp(k, "pack.window")) {
1698 window = git_config_int(k, v);
1699 return 0;
1701 if (!strcmp(k, "pack.windowmemory")) {
1702 window_memory_limit = git_config_ulong(k, v);
1703 return 0;
1705 if (!strcmp(k, "pack.depth")) {
1706 depth = git_config_int(k, v);
1707 return 0;
1709 if (!strcmp(k, "pack.compression")) {
1710 int level = git_config_int(k, v);
1711 if (level == -1)
1712 level = Z_DEFAULT_COMPRESSION;
1713 else if (level < 0 || level > Z_BEST_COMPRESSION)
1714 die("bad pack compression level %d", level);
1715 pack_compression_level = level;
1716 pack_compression_seen = 1;
1717 return 0;
1719 if (!strcmp(k, "pack.deltacachesize")) {
1720 max_delta_cache_size = git_config_int(k, v);
1721 return 0;
1723 if (!strcmp(k, "pack.deltacachelimit")) {
1724 cache_max_small_delta_size = git_config_int(k, v);
1725 return 0;
1727 if (!strcmp(k, "pack.threads")) {
1728 delta_search_threads = git_config_int(k, v);
1729 if (delta_search_threads < 0)
1730 die("invalid number of threads specified (%d)",
1731 delta_search_threads);
1732 #ifndef THREADED_DELTA_SEARCH
1733 if (delta_search_threads != 1)
1734 warning("no threads support, ignoring %s", k);
1735 #endif
1736 return 0;
1738 if (!strcmp(k, "pack.indexversion")) {
1739 pack_idx_default_version = git_config_int(k, v);
1740 if (pack_idx_default_version > 2)
1741 die("bad pack.indexversion=%d", pack_idx_default_version);
1742 return 0;
1744 if (!strcmp(k, "pack.packsizelimit")) {
1745 pack_size_limit_cfg = git_config_ulong(k, v);
1746 return 0;
1748 return git_default_config(k, v);
1751 static void read_object_list_from_stdin(void)
1753 char line[40 + 1 + PATH_MAX + 2];
1754 unsigned char sha1[20];
1756 for (;;) {
1757 if (!fgets(line, sizeof(line), stdin)) {
1758 if (feof(stdin))
1759 break;
1760 if (!ferror(stdin))
1761 die("fgets returned NULL, not EOF, not error!");
1762 if (errno != EINTR)
1763 die("fgets: %s", strerror(errno));
1764 clearerr(stdin);
1765 continue;
1767 if (line[0] == '-') {
1768 if (get_sha1_hex(line+1, sha1))
1769 die("expected edge sha1, got garbage:\n %s",
1770 line);
1771 add_preferred_base(sha1);
1772 continue;
1774 if (get_sha1_hex(line, sha1))
1775 die("expected sha1, got garbage:\n %s", line);
1777 add_preferred_base_object(line+41);
1778 add_object_entry(sha1, 0, line+41, 0);
1782 #define OBJECT_ADDED (1u<<20)
1784 static void show_commit(struct commit *commit)
1786 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1787 commit->object.flags |= OBJECT_ADDED;
1790 static void show_object(struct object_array_entry *p)
1792 add_preferred_base_object(p->name);
1793 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1794 p->item->flags |= OBJECT_ADDED;
1797 static void show_edge(struct commit *commit)
1799 add_preferred_base(commit->object.sha1);
1802 struct in_pack_object {
1803 off_t offset;
1804 struct object *object;
1807 struct in_pack {
1808 int alloc;
1809 int nr;
1810 struct in_pack_object *array;
1813 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1815 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1816 in_pack->array[in_pack->nr].object = object;
1817 in_pack->nr++;
1821 * Compare the objects in the offset order, in order to emulate the
1822 * "git-rev-list --objects" output that produced the pack originally.
1824 static int ofscmp(const void *a_, const void *b_)
1826 struct in_pack_object *a = (struct in_pack_object *)a_;
1827 struct in_pack_object *b = (struct in_pack_object *)b_;
1829 if (a->offset < b->offset)
1830 return -1;
1831 else if (a->offset > b->offset)
1832 return 1;
1833 else
1834 return hashcmp(a->object->sha1, b->object->sha1);
1837 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1839 struct packed_git *p;
1840 struct in_pack in_pack;
1841 uint32_t i;
1843 memset(&in_pack, 0, sizeof(in_pack));
1845 for (p = packed_git; p; p = p->next) {
1846 const unsigned char *sha1;
1847 struct object *o;
1849 for (i = 0; i < revs->num_ignore_packed; i++) {
1850 if (matches_pack_name(p, revs->ignore_packed[i]))
1851 break;
1853 if (revs->num_ignore_packed <= i)
1854 continue;
1855 if (open_pack_index(p))
1856 die("cannot open pack index");
1858 ALLOC_GROW(in_pack.array,
1859 in_pack.nr + p->num_objects,
1860 in_pack.alloc);
1862 for (i = 0; i < p->num_objects; i++) {
1863 sha1 = nth_packed_object_sha1(p, i);
1864 o = lookup_unknown_object(sha1);
1865 if (!(o->flags & OBJECT_ADDED))
1866 mark_in_pack_object(o, p, &in_pack);
1867 o->flags |= OBJECT_ADDED;
1871 if (in_pack.nr) {
1872 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1873 ofscmp);
1874 for (i = 0; i < in_pack.nr; i++) {
1875 struct object *o = in_pack.array[i].object;
1876 add_object_entry(o->sha1, o->type, "", 0);
1879 free(in_pack.array);
1882 static void get_object_list(int ac, const char **av)
1884 struct rev_info revs;
1885 char line[1000];
1886 int flags = 0;
1888 init_revisions(&revs, NULL);
1889 save_commit_buffer = 0;
1890 setup_revisions(ac, av, &revs, NULL);
1892 while (fgets(line, sizeof(line), stdin) != NULL) {
1893 int len = strlen(line);
1894 if (len && line[len - 1] == '\n')
1895 line[--len] = 0;
1896 if (!len)
1897 break;
1898 if (*line == '-') {
1899 if (!strcmp(line, "--not")) {
1900 flags ^= UNINTERESTING;
1901 continue;
1903 die("not a rev '%s'", line);
1905 if (handle_revision_arg(line, &revs, flags, 1))
1906 die("bad revision '%s'", line);
1909 if (prepare_revision_walk(&revs))
1910 die("revision walk setup failed");
1911 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1912 traverse_commit_list(&revs, show_commit, show_object);
1914 if (keep_unreachable)
1915 add_objects_in_unpacked_packs(&revs);
1918 static int adjust_perm(const char *path, mode_t mode)
1920 if (chmod(path, mode))
1921 return -1;
1922 return adjust_shared_perm(path);
1925 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1927 int use_internal_rev_list = 0;
1928 int thin = 0;
1929 uint32_t i;
1930 const char **rp_av;
1931 int rp_ac_alloc = 64;
1932 int rp_ac;
1934 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
1936 rp_av[0] = "pack-objects";
1937 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
1938 rp_ac = 2;
1940 git_config(git_pack_config);
1941 if (!pack_compression_seen && core_compression_seen)
1942 pack_compression_level = core_compression_level;
1944 progress = isatty(2);
1945 for (i = 1; i < argc; i++) {
1946 const char *arg = argv[i];
1948 if (*arg != '-')
1949 break;
1951 if (!strcmp("--non-empty", arg)) {
1952 non_empty = 1;
1953 continue;
1955 if (!strcmp("--local", arg)) {
1956 local = 1;
1957 continue;
1959 if (!strcmp("--incremental", arg)) {
1960 incremental = 1;
1961 continue;
1963 if (!prefixcmp(arg, "--compression=")) {
1964 char *end;
1965 int level = strtoul(arg+14, &end, 0);
1966 if (!arg[14] || *end)
1967 usage(pack_usage);
1968 if (level == -1)
1969 level = Z_DEFAULT_COMPRESSION;
1970 else if (level < 0 || level > Z_BEST_COMPRESSION)
1971 die("bad pack compression level %d", level);
1972 pack_compression_level = level;
1973 continue;
1975 if (!prefixcmp(arg, "--max-pack-size=")) {
1976 char *end;
1977 pack_size_limit_cfg = 0;
1978 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
1979 if (!arg[16] || *end)
1980 usage(pack_usage);
1981 continue;
1983 if (!prefixcmp(arg, "--window=")) {
1984 char *end;
1985 window = strtoul(arg+9, &end, 0);
1986 if (!arg[9] || *end)
1987 usage(pack_usage);
1988 continue;
1990 if (!prefixcmp(arg, "--window-memory=")) {
1991 if (!git_parse_ulong(arg+16, &window_memory_limit))
1992 usage(pack_usage);
1993 continue;
1995 if (!prefixcmp(arg, "--threads=")) {
1996 char *end;
1997 delta_search_threads = strtoul(arg+10, &end, 0);
1998 if (!arg[10] || *end || delta_search_threads < 0)
1999 usage(pack_usage);
2000 #ifndef THREADED_DELTA_SEARCH
2001 if (delta_search_threads != 1)
2002 warning("no threads support, "
2003 "ignoring %s", arg);
2004 #endif
2005 continue;
2007 if (!prefixcmp(arg, "--depth=")) {
2008 char *end;
2009 depth = strtoul(arg+8, &end, 0);
2010 if (!arg[8] || *end)
2011 usage(pack_usage);
2012 continue;
2014 if (!strcmp("--progress", arg)) {
2015 progress = 1;
2016 continue;
2018 if (!strcmp("--all-progress", arg)) {
2019 progress = 2;
2020 continue;
2022 if (!strcmp("-q", arg)) {
2023 progress = 0;
2024 continue;
2026 if (!strcmp("--no-reuse-delta", arg)) {
2027 no_reuse_delta = 1;
2028 continue;
2030 if (!strcmp("--no-reuse-object", arg)) {
2031 no_reuse_object = no_reuse_delta = 1;
2032 continue;
2034 if (!strcmp("--delta-base-offset", arg)) {
2035 allow_ofs_delta = 1;
2036 continue;
2038 if (!strcmp("--stdout", arg)) {
2039 pack_to_stdout = 1;
2040 continue;
2042 if (!strcmp("--revs", arg)) {
2043 use_internal_rev_list = 1;
2044 continue;
2046 if (!strcmp("--keep-unreachable", arg)) {
2047 keep_unreachable = 1;
2048 continue;
2050 if (!strcmp("--include-tag", arg)) {
2051 include_tag = 1;
2052 continue;
2054 if (!strcmp("--unpacked", arg) ||
2055 !prefixcmp(arg, "--unpacked=") ||
2056 !strcmp("--reflog", arg) ||
2057 !strcmp("--all", arg)) {
2058 use_internal_rev_list = 1;
2059 if (rp_ac >= rp_ac_alloc - 1) {
2060 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2061 rp_av = xrealloc(rp_av,
2062 rp_ac_alloc * sizeof(*rp_av));
2064 rp_av[rp_ac++] = arg;
2065 continue;
2067 if (!strcmp("--thin", arg)) {
2068 use_internal_rev_list = 1;
2069 thin = 1;
2070 rp_av[1] = "--objects-edge";
2071 continue;
2073 if (!prefixcmp(arg, "--index-version=")) {
2074 char *c;
2075 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2076 if (pack_idx_default_version > 2)
2077 die("bad %s", arg);
2078 if (*c == ',')
2079 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2080 if (*c || pack_idx_off32_limit & 0x80000000)
2081 die("bad %s", arg);
2082 continue;
2084 usage(pack_usage);
2087 /* Traditionally "pack-objects [options] base extra" failed;
2088 * we would however want to take refs parameter that would
2089 * have been given to upstream rev-list ourselves, which means
2090 * we somehow want to say what the base name is. So the
2091 * syntax would be:
2093 * pack-objects [options] base <refs...>
2095 * in other words, we would treat the first non-option as the
2096 * base_name and send everything else to the internal revision
2097 * walker.
2100 if (!pack_to_stdout)
2101 base_name = argv[i++];
2103 if (pack_to_stdout != !base_name)
2104 usage(pack_usage);
2106 if (!pack_to_stdout && !pack_size_limit)
2107 pack_size_limit = pack_size_limit_cfg;
2109 if (pack_to_stdout && pack_size_limit)
2110 die("--max-pack-size cannot be used to build a pack for transfer.");
2112 if (!pack_to_stdout && thin)
2113 die("--thin cannot be used to build an indexable pack.");
2115 #ifdef THREADED_DELTA_SEARCH
2116 if (!delta_search_threads) /* --threads=0 means autodetect */
2117 delta_search_threads = online_cpus();
2118 #endif
2120 prepare_packed_git();
2122 if (progress)
2123 progress_state = start_progress("Counting objects", 0);
2124 if (!use_internal_rev_list)
2125 read_object_list_from_stdin();
2126 else {
2127 rp_av[rp_ac] = NULL;
2128 get_object_list(rp_ac, rp_av);
2130 if (include_tag && nr_result)
2131 for_each_ref(add_ref_tag, NULL);
2132 stop_progress(&progress_state);
2134 if (non_empty && !nr_result)
2135 return 0;
2136 if (nr_result)
2137 prepare_pack(window, depth);
2138 write_pack_file();
2139 if (progress)
2140 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2141 written, written_delta, reused, reused_delta);
2142 return 0;