pack-objects: fix early eviction for max depth delta objects
[git/trast.git] / builtin-pack-objects.c
blobe20851e1c9c9ad517314794b7c8d4e3615afa9c5
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 long z_delta_size; /* delta data size (compressed) */
47 unsigned int hash; /* name hint hash */
48 enum object_type type;
49 enum object_type in_pack_type; /* could be delta */
50 unsigned char in_pack_header_size;
51 unsigned char preferred_base; /* we do not pack this, but is available
52 * to be used as the base object to delta
53 * objects against.
55 unsigned char no_try_delta;
59 * Objects we are going to pack are collected in objects array (dynamically
60 * expanded). nr_objects & nr_alloc controls this array. They are stored
61 * in the order we see -- typically rev-list --objects order that gives us
62 * nice "minimum seek" order.
64 static struct object_entry *objects;
65 static struct pack_idx_entry **written_list;
66 static uint32_t nr_objects, nr_alloc, nr_result, nr_written;
68 static int non_empty;
69 static int reuse_delta = 1, reuse_object = 1;
70 static int keep_unreachable, include_tag;
71 static int local;
72 static int incremental;
73 static int allow_ofs_delta;
74 static const char *base_name;
75 static int progress = 1;
76 static int window = 10;
77 static uint32_t pack_size_limit, pack_size_limit_cfg;
78 static int depth = 50;
79 static int delta_search_threads = 1;
80 static int pack_to_stdout;
81 static int num_preferred_base;
82 static struct progress *progress_state;
83 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
84 static int pack_compression_seen;
86 static unsigned long delta_cache_size = 0;
87 static unsigned long max_delta_cache_size = 0;
88 static unsigned long cache_max_small_delta_size = 1000;
90 static unsigned long window_memory_limit = 0;
93 * The object names in objects array are hashed with this hashtable,
94 * to help looking up the entry by object name.
95 * This hashtable is built after all the objects are seen.
97 static int *object_ix;
98 static int object_ix_hashsz;
101 * stats
103 static uint32_t written, written_delta;
104 static uint32_t reused, reused_delta;
107 static void *get_delta(struct object_entry *entry)
109 unsigned long size, base_size, delta_size;
110 void *buf, *base_buf, *delta_buf;
111 enum object_type type;
113 buf = read_sha1_file(entry->idx.sha1, &type, &size);
114 if (!buf)
115 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
116 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
117 if (!base_buf)
118 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
119 delta_buf = diff_delta(base_buf, base_size,
120 buf, size, &delta_size, 0);
121 if (!delta_buf || delta_size != entry->delta_size)
122 die("delta size changed");
123 free(buf);
124 free(base_buf);
125 return delta_buf;
128 static unsigned long do_compress(void **pptr, unsigned long size)
130 z_stream stream;
131 void *in, *out;
132 unsigned long maxsize;
134 memset(&stream, 0, sizeof(stream));
135 deflateInit(&stream, pack_compression_level);
136 maxsize = deflateBound(&stream, size);
138 in = *pptr;
139 out = xmalloc(maxsize);
140 *pptr = out;
142 stream.next_in = in;
143 stream.avail_in = size;
144 stream.next_out = out;
145 stream.avail_out = maxsize;
146 while (deflate(&stream, Z_FINISH) == Z_OK)
147 ; /* nothing */
148 deflateEnd(&stream);
150 free(in);
151 return stream.total_out;
155 * The per-object header is a pretty dense thing, which is
156 * - first byte: low four bits are "size", then three bits of "type",
157 * and the high bit is "size continues".
158 * - each byte afterwards: low seven bits are size continuation,
159 * with the high bit being "size continues"
161 static int encode_header(enum object_type type, unsigned long size, unsigned char *hdr)
163 int n = 1;
164 unsigned char c;
166 if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
167 die("bad type %d", type);
169 c = (type << 4) | (size & 15);
170 size >>= 4;
171 while (size) {
172 *hdr++ = c | 0x80;
173 c = size & 0x7f;
174 size >>= 7;
175 n++;
177 *hdr = c;
178 return n;
182 * we are going to reuse the existing object data as is. make
183 * sure it is not corrupt.
185 static int check_pack_inflate(struct packed_git *p,
186 struct pack_window **w_curs,
187 off_t offset,
188 off_t len,
189 unsigned long expect)
191 z_stream stream;
192 unsigned char fakebuf[4096], *in;
193 int st;
195 memset(&stream, 0, sizeof(stream));
196 inflateInit(&stream);
197 do {
198 in = use_pack(p, w_curs, offset, &stream.avail_in);
199 stream.next_in = in;
200 stream.next_out = fakebuf;
201 stream.avail_out = sizeof(fakebuf);
202 st = inflate(&stream, Z_FINISH);
203 offset += stream.next_in - in;
204 } while (st == Z_OK || st == Z_BUF_ERROR);
205 inflateEnd(&stream);
206 return (st == Z_STREAM_END &&
207 stream.total_out == expect &&
208 stream.total_in == len) ? 0 : -1;
211 static int check_pack_crc(struct packed_git *p, struct pack_window **w_curs,
212 off_t offset, off_t len, unsigned int nr)
214 const uint32_t *index_crc;
215 uint32_t data_crc = crc32(0, Z_NULL, 0);
217 do {
218 unsigned int avail;
219 void *data = use_pack(p, w_curs, offset, &avail);
220 if (avail > len)
221 avail = len;
222 data_crc = crc32(data_crc, data, avail);
223 offset += avail;
224 len -= avail;
225 } while (len);
227 index_crc = p->index_data;
228 index_crc += 2 + 256 + p->num_objects * (20/4) + nr;
230 return data_crc != ntohl(*index_crc);
233 static void copy_pack_data(struct sha1file *f,
234 struct packed_git *p,
235 struct pack_window **w_curs,
236 off_t offset,
237 off_t len)
239 unsigned char *in;
240 unsigned int avail;
242 while (len) {
243 in = use_pack(p, w_curs, offset, &avail);
244 if (avail > len)
245 avail = (unsigned int)len;
246 sha1write(f, in, avail);
247 offset += avail;
248 len -= avail;
252 static unsigned long write_object(struct sha1file *f,
253 struct object_entry *entry,
254 off_t write_offset)
256 unsigned long size, limit, datalen;
257 void *buf;
258 unsigned char header[10], dheader[10];
259 unsigned hdrlen;
260 enum object_type type;
261 int usable_delta, to_reuse;
263 if (!pack_to_stdout)
264 crc32_begin(f);
266 type = entry->type;
268 /* write limit if limited packsize and not first object */
269 limit = pack_size_limit && nr_written ?
270 pack_size_limit - write_offset : 0;
272 if (!entry->delta)
273 usable_delta = 0; /* no delta */
274 else if (!pack_size_limit)
275 usable_delta = 1; /* unlimited packfile */
276 else if (entry->delta->idx.offset == (off_t)-1)
277 usable_delta = 0; /* base was written to another pack */
278 else if (entry->delta->idx.offset)
279 usable_delta = 1; /* base already exists in this pack */
280 else
281 usable_delta = 0; /* base could end up in another pack */
283 if (!reuse_object)
284 to_reuse = 0; /* explicit */
285 else if (!entry->in_pack)
286 to_reuse = 0; /* can't reuse what we don't have */
287 else if (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA)
288 /* check_object() decided it for us ... */
289 to_reuse = usable_delta;
290 /* ... but pack split may override that */
291 else if (type != entry->in_pack_type)
292 to_reuse = 0; /* pack has delta which is unusable */
293 else if (entry->delta)
294 to_reuse = 0; /* we want to pack afresh */
295 else
296 to_reuse = 1; /* we have it in-pack undeltified,
297 * and we do not need to deltify it.
300 if (!to_reuse) {
301 if (!usable_delta) {
302 buf = read_sha1_file(entry->idx.sha1, &type, &size);
303 if (!buf)
304 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
306 * make sure no cached delta data remains from a
307 * previous attempt before a pack split occured.
309 free(entry->delta_data);
310 entry->delta_data = NULL;
311 entry->z_delta_size = 0;
312 } else if (entry->delta_data) {
313 size = entry->delta_size;
314 buf = entry->delta_data;
315 entry->delta_data = NULL;
316 type = (allow_ofs_delta && entry->delta->idx.offset) ?
317 OBJ_OFS_DELTA : OBJ_REF_DELTA;
318 } else {
319 buf = get_delta(entry);
320 size = entry->delta_size;
321 type = (allow_ofs_delta && entry->delta->idx.offset) ?
322 OBJ_OFS_DELTA : OBJ_REF_DELTA;
325 if (entry->z_delta_size)
326 datalen = entry->z_delta_size;
327 else
328 datalen = do_compress(&buf, size);
331 * The object header is a byte of 'type' followed by zero or
332 * more bytes of length.
334 hdrlen = encode_header(type, size, header);
336 if (type == OBJ_OFS_DELTA) {
338 * Deltas with relative base contain an additional
339 * encoding of the relative offset for the delta
340 * base from this object's position in the pack.
342 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
343 unsigned pos = sizeof(dheader) - 1;
344 dheader[pos] = ofs & 127;
345 while (ofs >>= 7)
346 dheader[--pos] = 128 | (--ofs & 127);
347 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
348 free(buf);
349 return 0;
351 sha1write(f, header, hdrlen);
352 sha1write(f, dheader + pos, sizeof(dheader) - pos);
353 hdrlen += sizeof(dheader) - pos;
354 } else if (type == OBJ_REF_DELTA) {
356 * Deltas with a base reference contain
357 * an additional 20 bytes for the base sha1.
359 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
360 free(buf);
361 return 0;
363 sha1write(f, header, hdrlen);
364 sha1write(f, entry->delta->idx.sha1, 20);
365 hdrlen += 20;
366 } else {
367 if (limit && hdrlen + datalen + 20 >= limit) {
368 free(buf);
369 return 0;
371 sha1write(f, header, hdrlen);
373 sha1write(f, buf, datalen);
374 free(buf);
376 else {
377 struct packed_git *p = entry->in_pack;
378 struct pack_window *w_curs = NULL;
379 struct revindex_entry *revidx;
380 off_t offset;
382 if (entry->delta) {
383 type = (allow_ofs_delta && entry->delta->idx.offset) ?
384 OBJ_OFS_DELTA : OBJ_REF_DELTA;
385 reused_delta++;
387 hdrlen = encode_header(type, entry->size, header);
388 offset = entry->in_pack_offset;
389 revidx = find_pack_revindex(p, offset);
390 datalen = revidx[1].offset - offset;
391 if (!pack_to_stdout && p->index_version > 1 &&
392 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr))
393 die("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
394 offset += entry->in_pack_header_size;
395 datalen -= entry->in_pack_header_size;
396 if (type == OBJ_OFS_DELTA) {
397 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
398 unsigned pos = sizeof(dheader) - 1;
399 dheader[pos] = ofs & 127;
400 while (ofs >>= 7)
401 dheader[--pos] = 128 | (--ofs & 127);
402 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit)
403 return 0;
404 sha1write(f, header, hdrlen);
405 sha1write(f, dheader + pos, sizeof(dheader) - pos);
406 hdrlen += sizeof(dheader) - pos;
407 } else if (type == OBJ_REF_DELTA) {
408 if (limit && hdrlen + 20 + datalen + 20 >= limit)
409 return 0;
410 sha1write(f, header, hdrlen);
411 sha1write(f, entry->delta->idx.sha1, 20);
412 hdrlen += 20;
413 } else {
414 if (limit && hdrlen + datalen + 20 >= limit)
415 return 0;
416 sha1write(f, header, hdrlen);
419 if (!pack_to_stdout && p->index_version == 1 &&
420 check_pack_inflate(p, &w_curs, offset, datalen, entry->size))
421 die("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
422 copy_pack_data(f, p, &w_curs, offset, datalen);
423 unuse_pack(&w_curs);
424 reused++;
426 if (usable_delta)
427 written_delta++;
428 written++;
429 if (!pack_to_stdout)
430 entry->idx.crc32 = crc32_end(f);
431 return hdrlen + datalen;
434 static off_t write_one(struct sha1file *f,
435 struct object_entry *e,
436 off_t offset)
438 unsigned long size;
440 /* offset is non zero if object is written already. */
441 if (e->idx.offset || e->preferred_base)
442 return offset;
444 /* if we are deltified, write out base object first. */
445 if (e->delta) {
446 offset = write_one(f, e->delta, offset);
447 if (!offset)
448 return 0;
451 e->idx.offset = offset;
452 size = write_object(f, e, offset);
453 if (!size) {
454 e->idx.offset = 0;
455 return 0;
457 written_list[nr_written++] = &e->idx;
459 /* make sure off_t is sufficiently large not to wrap */
460 if (offset > offset + size)
461 die("pack too large for current definition of off_t");
462 return offset + size;
465 /* forward declaration for write_pack_file */
466 static int adjust_perm(const char *path, mode_t mode);
468 static void write_pack_file(void)
470 uint32_t i = 0, j;
471 struct sha1file *f;
472 off_t offset, offset_one, last_obj_offset = 0;
473 struct pack_header hdr;
474 uint32_t nr_remaining = nr_result;
475 time_t last_mtime = 0;
477 if (progress > pack_to_stdout)
478 progress_state = start_progress("Writing objects", nr_result);
479 written_list = xmalloc(nr_objects * sizeof(*written_list));
481 do {
482 unsigned char sha1[20];
483 char *pack_tmp_name = NULL;
485 if (pack_to_stdout) {
486 f = sha1fd_throughput(1, "<stdout>", progress_state);
487 } else {
488 char tmpname[PATH_MAX];
489 int fd;
490 snprintf(tmpname, sizeof(tmpname),
491 "%s/tmp_pack_XXXXXX", get_object_directory());
492 fd = xmkstemp(tmpname);
493 pack_tmp_name = xstrdup(tmpname);
494 f = sha1fd(fd, pack_tmp_name);
497 hdr.hdr_signature = htonl(PACK_SIGNATURE);
498 hdr.hdr_version = htonl(PACK_VERSION);
499 hdr.hdr_entries = htonl(nr_remaining);
500 sha1write(f, &hdr, sizeof(hdr));
501 offset = sizeof(hdr);
502 nr_written = 0;
503 for (; i < nr_objects; i++) {
504 last_obj_offset = offset;
505 offset_one = write_one(f, objects + i, offset);
506 if (!offset_one)
507 break;
508 offset = offset_one;
509 display_progress(progress_state, written);
513 * Did we write the wrong # entries in the header?
514 * If so, rewrite it like in fast-import
516 if (pack_to_stdout || nr_written == nr_remaining) {
517 sha1close(f, sha1, 1);
518 } else {
519 int fd = sha1close(f, NULL, 0);
520 fixup_pack_header_footer(fd, sha1, pack_tmp_name, nr_written);
521 close(fd);
524 if (!pack_to_stdout) {
525 mode_t mode = umask(0);
526 struct stat st;
527 char *idx_tmp_name, tmpname[PATH_MAX];
529 umask(mode);
530 mode = 0444 & ~mode;
532 idx_tmp_name = write_idx_file(NULL, written_list,
533 nr_written, sha1);
535 snprintf(tmpname, sizeof(tmpname), "%s-%s.pack",
536 base_name, sha1_to_hex(sha1));
537 if (adjust_perm(pack_tmp_name, mode))
538 die("unable to make temporary pack file readable: %s",
539 strerror(errno));
540 if (rename(pack_tmp_name, tmpname))
541 die("unable to rename temporary pack file: %s",
542 strerror(errno));
545 * Packs are runtime accessed in their mtime
546 * order since newer packs are more likely to contain
547 * younger objects. So if we are creating multiple
548 * packs then we should modify the mtime of later ones
549 * to preserve this property.
551 if (stat(tmpname, &st) < 0) {
552 warning("failed to stat %s: %s",
553 tmpname, strerror(errno));
554 } else if (!last_mtime) {
555 last_mtime = st.st_mtime;
556 } else {
557 struct utimbuf utb;
558 utb.actime = st.st_atime;
559 utb.modtime = --last_mtime;
560 if (utime(tmpname, &utb) < 0)
561 warning("failed utime() on %s: %s",
562 tmpname, strerror(errno));
565 snprintf(tmpname, sizeof(tmpname), "%s-%s.idx",
566 base_name, sha1_to_hex(sha1));
567 if (adjust_perm(idx_tmp_name, mode))
568 die("unable to make temporary index file readable: %s",
569 strerror(errno));
570 if (rename(idx_tmp_name, tmpname))
571 die("unable to rename temporary index file: %s",
572 strerror(errno));
574 free(idx_tmp_name);
575 free(pack_tmp_name);
576 puts(sha1_to_hex(sha1));
579 /* mark written objects as written to previous pack */
580 for (j = 0; j < nr_written; j++) {
581 written_list[j]->offset = (off_t)-1;
583 nr_remaining -= nr_written;
584 } while (nr_remaining && i < nr_objects);
586 free(written_list);
587 stop_progress(&progress_state);
588 if (written != nr_result)
589 die("wrote %u objects while expecting %u", written, nr_result);
591 * We have scanned through [0 ... i). Since we have written
592 * the correct number of objects, the remaining [i ... nr_objects)
593 * items must be either already written (due to out-of-order delta base)
594 * or a preferred base. Count those which are neither and complain if any.
596 for (j = 0; i < nr_objects; i++) {
597 struct object_entry *e = objects + i;
598 j += !e->idx.offset && !e->preferred_base;
600 if (j)
601 die("wrote %u objects as expected but %u unwritten", written, j);
604 static int locate_object_entry_hash(const unsigned char *sha1)
606 int i;
607 unsigned int ui;
608 memcpy(&ui, sha1, sizeof(unsigned int));
609 i = ui % object_ix_hashsz;
610 while (0 < object_ix[i]) {
611 if (!hashcmp(sha1, objects[object_ix[i] - 1].idx.sha1))
612 return i;
613 if (++i == object_ix_hashsz)
614 i = 0;
616 return -1 - i;
619 static struct object_entry *locate_object_entry(const unsigned char *sha1)
621 int i;
623 if (!object_ix_hashsz)
624 return NULL;
626 i = locate_object_entry_hash(sha1);
627 if (0 <= i)
628 return &objects[object_ix[i]-1];
629 return NULL;
632 static void rehash_objects(void)
634 uint32_t i;
635 struct object_entry *oe;
637 object_ix_hashsz = nr_objects * 3;
638 if (object_ix_hashsz < 1024)
639 object_ix_hashsz = 1024;
640 object_ix = xrealloc(object_ix, sizeof(int) * object_ix_hashsz);
641 memset(object_ix, 0, sizeof(int) * object_ix_hashsz);
642 for (i = 0, oe = objects; i < nr_objects; i++, oe++) {
643 int ix = locate_object_entry_hash(oe->idx.sha1);
644 if (0 <= ix)
645 continue;
646 ix = -1 - ix;
647 object_ix[ix] = i + 1;
651 static unsigned name_hash(const char *name)
653 unsigned char c;
654 unsigned hash = 0;
656 if (!name)
657 return 0;
660 * This effectively just creates a sortable number from the
661 * last sixteen non-whitespace characters. Last characters
662 * count "most", so things that end in ".c" sort together.
664 while ((c = *name++) != 0) {
665 if (isspace(c))
666 continue;
667 hash = (hash >> 2) + (c << 24);
669 return hash;
672 static void setup_delta_attr_check(struct git_attr_check *check)
674 static struct git_attr *attr_delta;
676 if (!attr_delta)
677 attr_delta = git_attr("delta", 5);
679 check[0].attr = attr_delta;
682 static int no_try_delta(const char *path)
684 struct git_attr_check check[1];
686 setup_delta_attr_check(check);
687 if (git_checkattr(path, ARRAY_SIZE(check), check))
688 return 0;
689 if (ATTR_FALSE(check->value))
690 return 1;
691 return 0;
694 static int add_object_entry(const unsigned char *sha1, enum object_type type,
695 const char *name, int exclude)
697 struct object_entry *entry;
698 struct packed_git *p, *found_pack = NULL;
699 off_t found_offset = 0;
700 int ix;
701 unsigned hash = name_hash(name);
703 ix = nr_objects ? locate_object_entry_hash(sha1) : -1;
704 if (ix >= 0) {
705 if (exclude) {
706 entry = objects + object_ix[ix] - 1;
707 if (!entry->preferred_base)
708 nr_result--;
709 entry->preferred_base = 1;
711 return 0;
714 for (p = packed_git; p; p = p->next) {
715 off_t offset = find_pack_entry_one(sha1, p);
716 if (offset) {
717 if (!found_pack) {
718 found_offset = offset;
719 found_pack = p;
721 if (exclude)
722 break;
723 if (incremental)
724 return 0;
725 if (local && !p->pack_local)
726 return 0;
730 if (nr_objects >= nr_alloc) {
731 nr_alloc = (nr_alloc + 1024) * 3 / 2;
732 objects = xrealloc(objects, nr_alloc * sizeof(*entry));
735 entry = objects + nr_objects++;
736 memset(entry, 0, sizeof(*entry));
737 hashcpy(entry->idx.sha1, sha1);
738 entry->hash = hash;
739 if (type)
740 entry->type = type;
741 if (exclude)
742 entry->preferred_base = 1;
743 else
744 nr_result++;
745 if (found_pack) {
746 entry->in_pack = found_pack;
747 entry->in_pack_offset = found_offset;
750 if (object_ix_hashsz * 3 <= nr_objects * 4)
751 rehash_objects();
752 else
753 object_ix[-1 - ix] = nr_objects;
755 display_progress(progress_state, nr_objects);
757 if (name && no_try_delta(name))
758 entry->no_try_delta = 1;
760 return 1;
763 struct pbase_tree_cache {
764 unsigned char sha1[20];
765 int ref;
766 int temporary;
767 void *tree_data;
768 unsigned long tree_size;
771 static struct pbase_tree_cache *(pbase_tree_cache[256]);
772 static int pbase_tree_cache_ix(const unsigned char *sha1)
774 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
776 static int pbase_tree_cache_ix_incr(int ix)
778 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
781 static struct pbase_tree {
782 struct pbase_tree *next;
783 /* This is a phony "cache" entry; we are not
784 * going to evict it nor find it through _get()
785 * mechanism -- this is for the toplevel node that
786 * would almost always change with any commit.
788 struct pbase_tree_cache pcache;
789 } *pbase_tree;
791 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
793 struct pbase_tree_cache *ent, *nent;
794 void *data;
795 unsigned long size;
796 enum object_type type;
797 int neigh;
798 int my_ix = pbase_tree_cache_ix(sha1);
799 int available_ix = -1;
801 /* pbase-tree-cache acts as a limited hashtable.
802 * your object will be found at your index or within a few
803 * slots after that slot if it is cached.
805 for (neigh = 0; neigh < 8; neigh++) {
806 ent = pbase_tree_cache[my_ix];
807 if (ent && !hashcmp(ent->sha1, sha1)) {
808 ent->ref++;
809 return ent;
811 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
812 ((0 <= available_ix) &&
813 (!ent && pbase_tree_cache[available_ix])))
814 available_ix = my_ix;
815 if (!ent)
816 break;
817 my_ix = pbase_tree_cache_ix_incr(my_ix);
820 /* Did not find one. Either we got a bogus request or
821 * we need to read and perhaps cache.
823 data = read_sha1_file(sha1, &type, &size);
824 if (!data)
825 return NULL;
826 if (type != OBJ_TREE) {
827 free(data);
828 return NULL;
831 /* We need to either cache or return a throwaway copy */
833 if (available_ix < 0)
834 ent = NULL;
835 else {
836 ent = pbase_tree_cache[available_ix];
837 my_ix = available_ix;
840 if (!ent) {
841 nent = xmalloc(sizeof(*nent));
842 nent->temporary = (available_ix < 0);
844 else {
845 /* evict and reuse */
846 free(ent->tree_data);
847 nent = ent;
849 hashcpy(nent->sha1, sha1);
850 nent->tree_data = data;
851 nent->tree_size = size;
852 nent->ref = 1;
853 if (!nent->temporary)
854 pbase_tree_cache[my_ix] = nent;
855 return nent;
858 static void pbase_tree_put(struct pbase_tree_cache *cache)
860 if (!cache->temporary) {
861 cache->ref--;
862 return;
864 free(cache->tree_data);
865 free(cache);
868 static int name_cmp_len(const char *name)
870 int i;
871 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
873 return i;
876 static void add_pbase_object(struct tree_desc *tree,
877 const char *name,
878 int cmplen,
879 const char *fullname)
881 struct name_entry entry;
882 int cmp;
884 while (tree_entry(tree,&entry)) {
885 if (S_ISGITLINK(entry.mode))
886 continue;
887 cmp = tree_entry_len(entry.path, entry.sha1) != cmplen ? 1 :
888 memcmp(name, entry.path, cmplen);
889 if (cmp > 0)
890 continue;
891 if (cmp < 0)
892 return;
893 if (name[cmplen] != '/') {
894 add_object_entry(entry.sha1,
895 object_type(entry.mode),
896 fullname, 1);
897 return;
899 if (S_ISDIR(entry.mode)) {
900 struct tree_desc sub;
901 struct pbase_tree_cache *tree;
902 const char *down = name+cmplen+1;
903 int downlen = name_cmp_len(down);
905 tree = pbase_tree_get(entry.sha1);
906 if (!tree)
907 return;
908 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
910 add_pbase_object(&sub, down, downlen, fullname);
911 pbase_tree_put(tree);
916 static unsigned *done_pbase_paths;
917 static int done_pbase_paths_num;
918 static int done_pbase_paths_alloc;
919 static int done_pbase_path_pos(unsigned hash)
921 int lo = 0;
922 int hi = done_pbase_paths_num;
923 while (lo < hi) {
924 int mi = (hi + lo) / 2;
925 if (done_pbase_paths[mi] == hash)
926 return mi;
927 if (done_pbase_paths[mi] < hash)
928 hi = mi;
929 else
930 lo = mi + 1;
932 return -lo-1;
935 static int check_pbase_path(unsigned hash)
937 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
938 if (0 <= pos)
939 return 1;
940 pos = -pos - 1;
941 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
942 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
943 done_pbase_paths = xrealloc(done_pbase_paths,
944 done_pbase_paths_alloc *
945 sizeof(unsigned));
947 done_pbase_paths_num++;
948 if (pos < done_pbase_paths_num)
949 memmove(done_pbase_paths + pos + 1,
950 done_pbase_paths + pos,
951 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
952 done_pbase_paths[pos] = hash;
953 return 0;
956 static void add_preferred_base_object(const char *name)
958 struct pbase_tree *it;
959 int cmplen;
960 unsigned hash = name_hash(name);
962 if (!num_preferred_base || check_pbase_path(hash))
963 return;
965 cmplen = name_cmp_len(name);
966 for (it = pbase_tree; it; it = it->next) {
967 if (cmplen == 0) {
968 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
970 else {
971 struct tree_desc tree;
972 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
973 add_pbase_object(&tree, name, cmplen, name);
978 static void add_preferred_base(unsigned char *sha1)
980 struct pbase_tree *it;
981 void *data;
982 unsigned long size;
983 unsigned char tree_sha1[20];
985 if (window <= num_preferred_base++)
986 return;
988 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
989 if (!data)
990 return;
992 for (it = pbase_tree; it; it = it->next) {
993 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
994 free(data);
995 return;
999 it = xcalloc(1, sizeof(*it));
1000 it->next = pbase_tree;
1001 pbase_tree = it;
1003 hashcpy(it->pcache.sha1, tree_sha1);
1004 it->pcache.tree_data = data;
1005 it->pcache.tree_size = size;
1008 static void check_object(struct object_entry *entry)
1010 if (entry->in_pack) {
1011 struct packed_git *p = entry->in_pack;
1012 struct pack_window *w_curs = NULL;
1013 const unsigned char *base_ref = NULL;
1014 struct object_entry *base_entry;
1015 unsigned long used, used_0;
1016 unsigned int avail;
1017 off_t ofs;
1018 unsigned char *buf, c;
1020 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1023 * We want in_pack_type even if we do not reuse delta
1024 * since non-delta representations could still be reused.
1026 used = unpack_object_header_gently(buf, avail,
1027 &entry->in_pack_type,
1028 &entry->size);
1031 * Determine if this is a delta and if so whether we can
1032 * reuse it or not. Otherwise let's find out as cheaply as
1033 * possible what the actual type and size for this object is.
1035 switch (entry->in_pack_type) {
1036 default:
1037 /* Not a delta hence we've already got all we need. */
1038 entry->type = entry->in_pack_type;
1039 entry->in_pack_header_size = used;
1040 unuse_pack(&w_curs);
1041 return;
1042 case OBJ_REF_DELTA:
1043 if (reuse_delta && !entry->preferred_base)
1044 base_ref = use_pack(p, &w_curs,
1045 entry->in_pack_offset + used, NULL);
1046 entry->in_pack_header_size = used + 20;
1047 break;
1048 case OBJ_OFS_DELTA:
1049 buf = use_pack(p, &w_curs,
1050 entry->in_pack_offset + used, NULL);
1051 used_0 = 0;
1052 c = buf[used_0++];
1053 ofs = c & 127;
1054 while (c & 128) {
1055 ofs += 1;
1056 if (!ofs || MSB(ofs, 7))
1057 die("delta base offset overflow in pack for %s",
1058 sha1_to_hex(entry->idx.sha1));
1059 c = buf[used_0++];
1060 ofs = (ofs << 7) + (c & 127);
1062 if (ofs >= entry->in_pack_offset)
1063 die("delta base offset out of bound for %s",
1064 sha1_to_hex(entry->idx.sha1));
1065 ofs = entry->in_pack_offset - ofs;
1066 if (reuse_delta && !entry->preferred_base) {
1067 struct revindex_entry *revidx;
1068 revidx = find_pack_revindex(p, ofs);
1069 base_ref = nth_packed_object_sha1(p, revidx->nr);
1071 entry->in_pack_header_size = used + used_0;
1072 break;
1075 if (base_ref && (base_entry = locate_object_entry(base_ref))) {
1077 * If base_ref was set above that means we wish to
1078 * reuse delta data, and we even found that base
1079 * in the list of objects we want to pack. Goodie!
1081 * Depth value does not matter - find_deltas() will
1082 * never consider reused delta as the base object to
1083 * deltify other objects against, in order to avoid
1084 * circular deltas.
1086 entry->type = entry->in_pack_type;
1087 entry->delta = base_entry;
1088 entry->delta_sibling = base_entry->delta_child;
1089 base_entry->delta_child = entry;
1090 unuse_pack(&w_curs);
1091 return;
1094 if (entry->type) {
1096 * This must be a delta and we already know what the
1097 * final object type is. Let's extract the actual
1098 * object size from the delta header.
1100 entry->size = get_size_from_delta(p, &w_curs,
1101 entry->in_pack_offset + entry->in_pack_header_size);
1102 unuse_pack(&w_curs);
1103 return;
1107 * No choice but to fall back to the recursive delta walk
1108 * with sha1_object_info() to find about the object type
1109 * at this point...
1111 unuse_pack(&w_curs);
1114 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1115 if (entry->type < 0)
1116 die("unable to get type of object %s",
1117 sha1_to_hex(entry->idx.sha1));
1120 static int pack_offset_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 /* avoid filesystem trashing with loose objects */
1126 if (!a->in_pack && !b->in_pack)
1127 return hashcmp(a->idx.sha1, b->idx.sha1);
1129 if (a->in_pack < b->in_pack)
1130 return -1;
1131 if (a->in_pack > b->in_pack)
1132 return 1;
1133 return a->in_pack_offset < b->in_pack_offset ? -1 :
1134 (a->in_pack_offset > b->in_pack_offset);
1137 static void get_object_details(void)
1139 uint32_t i;
1140 struct object_entry **sorted_by_offset;
1142 sorted_by_offset = xcalloc(nr_objects, sizeof(struct object_entry *));
1143 for (i = 0; i < nr_objects; i++)
1144 sorted_by_offset[i] = objects + i;
1145 qsort(sorted_by_offset, nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1147 init_pack_revindex();
1149 for (i = 0; i < nr_objects; i++)
1150 check_object(sorted_by_offset[i]);
1152 free(sorted_by_offset);
1156 * We search for deltas in a list sorted by type, by filename hash, and then
1157 * by size, so that we see progressively smaller and smaller files.
1158 * That's because we prefer deltas to be from the bigger file
1159 * to the smaller -- deletes are potentially cheaper, but perhaps
1160 * more importantly, the bigger file is likely the more recent
1161 * one. The deepest deltas are therefore the oldest objects which are
1162 * less susceptible to be accessed often.
1164 static int type_size_sort(const void *_a, const void *_b)
1166 const struct object_entry *a = *(struct object_entry **)_a;
1167 const struct object_entry *b = *(struct object_entry **)_b;
1169 if (a->type > b->type)
1170 return -1;
1171 if (a->type < b->type)
1172 return 1;
1173 if (a->hash > b->hash)
1174 return -1;
1175 if (a->hash < b->hash)
1176 return 1;
1177 if (a->preferred_base > b->preferred_base)
1178 return -1;
1179 if (a->preferred_base < b->preferred_base)
1180 return 1;
1181 if (a->size > b->size)
1182 return -1;
1183 if (a->size < b->size)
1184 return 1;
1185 return a < b ? -1 : (a > b); /* newest first */
1188 struct unpacked {
1189 struct object_entry *entry;
1190 void *data;
1191 struct delta_index *index;
1192 unsigned depth;
1195 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1196 unsigned long delta_size)
1198 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1199 return 0;
1201 if (delta_size < cache_max_small_delta_size)
1202 return 1;
1204 /* cache delta, if objects are large enough compared to delta size */
1205 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1206 return 1;
1208 return 0;
1211 #ifdef THREADED_DELTA_SEARCH
1213 static pthread_mutex_t read_mutex = PTHREAD_MUTEX_INITIALIZER;
1214 #define read_lock() pthread_mutex_lock(&read_mutex)
1215 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1217 static pthread_mutex_t cache_mutex = PTHREAD_MUTEX_INITIALIZER;
1218 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1219 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1221 static pthread_mutex_t progress_mutex = PTHREAD_MUTEX_INITIALIZER;
1222 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1223 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1225 #else
1227 #define read_lock() (void)0
1228 #define read_unlock() (void)0
1229 #define cache_lock() (void)0
1230 #define cache_unlock() (void)0
1231 #define progress_lock() (void)0
1232 #define progress_unlock() (void)0
1234 #endif
1236 static int try_delta(struct unpacked *trg, struct unpacked *src,
1237 unsigned max_depth, unsigned long *mem_usage)
1239 struct object_entry *trg_entry = trg->entry;
1240 struct object_entry *src_entry = src->entry;
1241 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1242 unsigned ref_depth;
1243 enum object_type type;
1244 void *delta_buf;
1246 /* Don't bother doing diffs between different types */
1247 if (trg_entry->type != src_entry->type)
1248 return -1;
1251 * We do not bother to try a delta that we discarded
1252 * on an earlier try, but only when reusing delta data.
1254 if (reuse_delta && trg_entry->in_pack &&
1255 trg_entry->in_pack == src_entry->in_pack &&
1256 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1257 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1258 return 0;
1260 /* Let's not bust the allowed depth. */
1261 if (src->depth >= max_depth)
1262 return 0;
1264 /* Now some size filtering heuristics. */
1265 trg_size = trg_entry->size;
1266 if (!trg_entry->delta) {
1267 max_size = trg_size/2 - 20;
1268 ref_depth = 1;
1269 } else {
1270 max_size = trg_entry->delta_size;
1271 ref_depth = trg->depth;
1273 max_size = max_size * (max_depth - src->depth) /
1274 (max_depth - ref_depth + 1);
1275 if (max_size == 0)
1276 return 0;
1277 src_size = src_entry->size;
1278 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1279 if (sizediff >= max_size)
1280 return 0;
1281 if (trg_size < src_size / 32)
1282 return 0;
1284 /* Load data if not already done */
1285 if (!trg->data) {
1286 read_lock();
1287 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1288 read_unlock();
1289 if (!trg->data)
1290 die("object %s cannot be read",
1291 sha1_to_hex(trg_entry->idx.sha1));
1292 if (sz != trg_size)
1293 die("object %s inconsistent object length (%lu vs %lu)",
1294 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1295 *mem_usage += sz;
1297 if (!src->data) {
1298 read_lock();
1299 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1300 read_unlock();
1301 if (!src->data)
1302 die("object %s cannot be read",
1303 sha1_to_hex(src_entry->idx.sha1));
1304 if (sz != src_size)
1305 die("object %s inconsistent object length (%lu vs %lu)",
1306 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1307 *mem_usage += sz;
1309 if (!src->index) {
1310 src->index = create_delta_index(src->data, src_size);
1311 if (!src->index) {
1312 static int warned = 0;
1313 if (!warned++)
1314 warning("suboptimal pack - out of memory");
1315 return 0;
1317 *mem_usage += sizeof_delta_index(src->index);
1320 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1321 if (!delta_buf)
1322 return 0;
1324 if (trg_entry->delta) {
1325 /* Prefer only shallower same-sized deltas. */
1326 if (delta_size == trg_entry->delta_size &&
1327 src->depth + 1 >= trg->depth) {
1328 free(delta_buf);
1329 return 0;
1334 * Handle memory allocation outside of the cache
1335 * accounting lock. Compiler will optimize the strangeness
1336 * away when THREADED_DELTA_SEARCH is not defined.
1338 free(trg_entry->delta_data);
1339 cache_lock();
1340 if (trg_entry->delta_data) {
1341 delta_cache_size -= trg_entry->delta_size;
1342 trg_entry->delta_data = NULL;
1344 if (delta_cacheable(src_size, trg_size, delta_size)) {
1345 delta_cache_size += delta_size;
1346 cache_unlock();
1347 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1348 } else {
1349 cache_unlock();
1350 free(delta_buf);
1353 trg_entry->delta = src_entry;
1354 trg_entry->delta_size = delta_size;
1355 trg->depth = src->depth + 1;
1357 return 1;
1360 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1362 struct object_entry *child = me->delta_child;
1363 unsigned int m = n;
1364 while (child) {
1365 unsigned int c = check_delta_limit(child, n + 1);
1366 if (m < c)
1367 m = c;
1368 child = child->delta_sibling;
1370 return m;
1373 static unsigned long free_unpacked(struct unpacked *n)
1375 unsigned long freed_mem = sizeof_delta_index(n->index);
1376 free_delta_index(n->index);
1377 n->index = NULL;
1378 if (n->data) {
1379 freed_mem += n->entry->size;
1380 free(n->data);
1381 n->data = NULL;
1383 n->entry = NULL;
1384 n->depth = 0;
1385 return freed_mem;
1388 static void find_deltas(struct object_entry **list, unsigned *list_size,
1389 int window, int depth, unsigned *processed)
1391 uint32_t i, idx = 0, count = 0;
1392 unsigned int array_size = window * sizeof(struct unpacked);
1393 struct unpacked *array;
1394 unsigned long mem_usage = 0;
1396 array = xmalloc(array_size);
1397 memset(array, 0, array_size);
1399 for (;;) {
1400 struct object_entry *entry = *list++;
1401 struct unpacked *n = array + idx;
1402 int j, max_depth, best_base = -1;
1404 progress_lock();
1405 if (!*list_size) {
1406 progress_unlock();
1407 break;
1409 (*list_size)--;
1410 if (!entry->preferred_base) {
1411 (*processed)++;
1412 display_progress(progress_state, *processed);
1414 progress_unlock();
1416 mem_usage -= free_unpacked(n);
1417 n->entry = entry;
1419 while (window_memory_limit &&
1420 mem_usage > window_memory_limit &&
1421 count > 1) {
1422 uint32_t tail = (idx + window - count) % window;
1423 mem_usage -= free_unpacked(array + tail);
1424 count--;
1427 /* We do not compute delta to *create* objects we are not
1428 * going to pack.
1430 if (entry->preferred_base)
1431 goto next;
1434 * If the current object is at pack edge, take the depth the
1435 * objects that depend on the current object into account
1436 * otherwise they would become too deep.
1438 max_depth = depth;
1439 if (entry->delta_child) {
1440 max_depth -= check_delta_limit(entry, 0);
1441 if (max_depth <= 0)
1442 goto next;
1445 j = window;
1446 while (--j > 0) {
1447 int ret;
1448 uint32_t other_idx = idx + j;
1449 struct unpacked *m;
1450 if (other_idx >= window)
1451 other_idx -= window;
1452 m = array + other_idx;
1453 if (!m->entry)
1454 break;
1455 ret = try_delta(n, m, max_depth, &mem_usage);
1456 if (ret < 0)
1457 break;
1458 else if (ret > 0)
1459 best_base = other_idx;
1463 * If we decided to cache the delta data, then it is best
1464 * to compress it right away. First because we have to do
1465 * it anyway, and doing it here while we're threaded will
1466 * save a lot of time in the non threaded write phase,
1467 * as well as allow for caching more deltas within
1468 * the same cache size limit.
1469 * ...
1470 * But only if not writing to stdout, since in that case
1471 * the network is most likely throttling writes anyway,
1472 * and therefore it is best to go to the write phase ASAP
1473 * instead, as we can afford spending more time compressing
1474 * between writes at that moment.
1476 if (entry->delta_data && !pack_to_stdout) {
1477 entry->z_delta_size = do_compress(&entry->delta_data,
1478 entry->delta_size);
1479 cache_lock();
1480 delta_cache_size -= entry->delta_size;
1481 delta_cache_size += entry->z_delta_size;
1482 cache_unlock();
1485 /* if we made n a delta, and if n is already at max
1486 * depth, leaving it in the window is pointless. we
1487 * should evict it first.
1489 if (entry->delta && max_depth <= n->depth)
1490 continue;
1493 * Move the best delta base up in the window, after the
1494 * currently deltified object, to keep it longer. It will
1495 * be the first base object to be attempted next.
1497 if (entry->delta) {
1498 struct unpacked swap = array[best_base];
1499 int dist = (window + idx - best_base) % window;
1500 int dst = best_base;
1501 while (dist--) {
1502 int src = (dst + 1) % window;
1503 array[dst] = array[src];
1504 dst = src;
1506 array[dst] = swap;
1509 next:
1510 idx++;
1511 if (count + 1 < window)
1512 count++;
1513 if (idx >= window)
1514 idx = 0;
1517 for (i = 0; i < window; ++i) {
1518 free_delta_index(array[i].index);
1519 free(array[i].data);
1521 free(array);
1524 #ifdef THREADED_DELTA_SEARCH
1527 * The main thread waits on the condition that (at least) one of the workers
1528 * has stopped working (which is indicated in the .working member of
1529 * struct thread_params).
1530 * When a work thread has completed its work, it sets .working to 0 and
1531 * signals the main thread and waits on the condition that .data_ready
1532 * becomes 1.
1535 struct thread_params {
1536 pthread_t thread;
1537 struct object_entry **list;
1538 unsigned list_size;
1539 unsigned remaining;
1540 int window;
1541 int depth;
1542 int working;
1543 int data_ready;
1544 pthread_mutex_t mutex;
1545 pthread_cond_t cond;
1546 unsigned *processed;
1549 static pthread_cond_t progress_cond = PTHREAD_COND_INITIALIZER;
1551 static void *threaded_find_deltas(void *arg)
1553 struct thread_params *me = arg;
1555 while (me->remaining) {
1556 find_deltas(me->list, &me->remaining,
1557 me->window, me->depth, me->processed);
1559 progress_lock();
1560 me->working = 0;
1561 pthread_cond_signal(&progress_cond);
1562 progress_unlock();
1565 * We must not set ->data_ready before we wait on the
1566 * condition because the main thread may have set it to 1
1567 * before we get here. In order to be sure that new
1568 * work is available if we see 1 in ->data_ready, it
1569 * was initialized to 0 before this thread was spawned
1570 * and we reset it to 0 right away.
1572 pthread_mutex_lock(&me->mutex);
1573 while (!me->data_ready)
1574 pthread_cond_wait(&me->cond, &me->mutex);
1575 me->data_ready = 0;
1576 pthread_mutex_unlock(&me->mutex);
1578 /* leave ->working 1 so that this doesn't get more work assigned */
1579 return NULL;
1582 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1583 int window, int depth, unsigned *processed)
1585 struct thread_params p[delta_search_threads];
1586 int i, ret, active_threads = 0;
1588 if (delta_search_threads <= 1) {
1589 find_deltas(list, &list_size, window, depth, processed);
1590 return;
1593 /* Partition the work amongst work threads. */
1594 for (i = 0; i < delta_search_threads; i++) {
1595 unsigned sub_size = list_size / (delta_search_threads - i);
1597 p[i].window = window;
1598 p[i].depth = depth;
1599 p[i].processed = processed;
1600 p[i].working = 1;
1601 p[i].data_ready = 0;
1603 /* try to split chunks on "path" boundaries */
1604 while (sub_size && sub_size < list_size &&
1605 list[sub_size]->hash &&
1606 list[sub_size]->hash == list[sub_size-1]->hash)
1607 sub_size++;
1609 p[i].list = list;
1610 p[i].list_size = sub_size;
1611 p[i].remaining = sub_size;
1613 list += sub_size;
1614 list_size -= sub_size;
1617 /* Start work threads. */
1618 for (i = 0; i < delta_search_threads; i++) {
1619 if (!p[i].list_size)
1620 continue;
1621 pthread_mutex_init(&p[i].mutex, NULL);
1622 pthread_cond_init(&p[i].cond, NULL);
1623 ret = pthread_create(&p[i].thread, NULL,
1624 threaded_find_deltas, &p[i]);
1625 if (ret)
1626 die("unable to create thread: %s", strerror(ret));
1627 active_threads++;
1631 * Now let's wait for work completion. Each time a thread is done
1632 * with its work, we steal half of the remaining work from the
1633 * thread with the largest number of unprocessed objects and give
1634 * it to that newly idle thread. This ensure good load balancing
1635 * until the remaining object list segments are simply too short
1636 * to be worth splitting anymore.
1638 while (active_threads) {
1639 struct thread_params *target = NULL;
1640 struct thread_params *victim = NULL;
1641 unsigned sub_size = 0;
1643 progress_lock();
1644 for (;;) {
1645 for (i = 0; !target && i < delta_search_threads; i++)
1646 if (!p[i].working)
1647 target = &p[i];
1648 if (target)
1649 break;
1650 pthread_cond_wait(&progress_cond, &progress_mutex);
1653 for (i = 0; i < delta_search_threads; i++)
1654 if (p[i].remaining > 2*window &&
1655 (!victim || victim->remaining < p[i].remaining))
1656 victim = &p[i];
1657 if (victim) {
1658 sub_size = victim->remaining / 2;
1659 list = victim->list + victim->list_size - sub_size;
1660 while (sub_size && list[0]->hash &&
1661 list[0]->hash == list[-1]->hash) {
1662 list++;
1663 sub_size--;
1665 if (!sub_size) {
1667 * It is possible for some "paths" to have
1668 * so many objects that no hash boundary
1669 * might be found. Let's just steal the
1670 * exact half in that case.
1672 sub_size = victim->remaining / 2;
1673 list -= sub_size;
1675 target->list = list;
1676 victim->list_size -= sub_size;
1677 victim->remaining -= sub_size;
1679 target->list_size = sub_size;
1680 target->remaining = sub_size;
1681 target->working = 1;
1682 progress_unlock();
1684 pthread_mutex_lock(&target->mutex);
1685 target->data_ready = 1;
1686 pthread_cond_signal(&target->cond);
1687 pthread_mutex_unlock(&target->mutex);
1689 if (!sub_size) {
1690 pthread_join(target->thread, NULL);
1691 pthread_cond_destroy(&target->cond);
1692 pthread_mutex_destroy(&target->mutex);
1693 active_threads--;
1698 #else
1699 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
1700 #endif
1702 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1704 unsigned char peeled[20];
1706 if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1707 !peel_ref(path, peeled) && /* peelable? */
1708 !is_null_sha1(peeled) && /* annotated tag? */
1709 locate_object_entry(peeled)) /* object packed? */
1710 add_object_entry(sha1, OBJ_TAG, NULL, 0);
1711 return 0;
1714 static void prepare_pack(int window, int depth)
1716 struct object_entry **delta_list;
1717 uint32_t i, n, nr_deltas;
1719 get_object_details();
1721 if (!nr_objects || !window || !depth)
1722 return;
1724 delta_list = xmalloc(nr_objects * sizeof(*delta_list));
1725 nr_deltas = n = 0;
1727 for (i = 0; i < nr_objects; i++) {
1728 struct object_entry *entry = objects + i;
1730 if (entry->delta)
1731 /* This happens if we decided to reuse existing
1732 * delta from a pack. "reuse_delta &&" is implied.
1734 continue;
1736 if (entry->size < 50)
1737 continue;
1739 if (entry->no_try_delta)
1740 continue;
1742 if (!entry->preferred_base)
1743 nr_deltas++;
1745 delta_list[n++] = entry;
1748 if (nr_deltas && n > 1) {
1749 unsigned nr_done = 0;
1750 if (progress)
1751 progress_state = start_progress("Compressing objects",
1752 nr_deltas);
1753 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1754 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1755 stop_progress(&progress_state);
1756 if (nr_done != nr_deltas)
1757 die("inconsistency with delta count");
1759 free(delta_list);
1762 static int git_pack_config(const char *k, const char *v)
1764 if(!strcmp(k, "pack.window")) {
1765 window = git_config_int(k, v);
1766 return 0;
1768 if (!strcmp(k, "pack.windowmemory")) {
1769 window_memory_limit = git_config_ulong(k, v);
1770 return 0;
1772 if (!strcmp(k, "pack.depth")) {
1773 depth = git_config_int(k, v);
1774 return 0;
1776 if (!strcmp(k, "pack.compression")) {
1777 int level = git_config_int(k, v);
1778 if (level == -1)
1779 level = Z_DEFAULT_COMPRESSION;
1780 else if (level < 0 || level > Z_BEST_COMPRESSION)
1781 die("bad pack compression level %d", level);
1782 pack_compression_level = level;
1783 pack_compression_seen = 1;
1784 return 0;
1786 if (!strcmp(k, "pack.deltacachesize")) {
1787 max_delta_cache_size = git_config_int(k, v);
1788 return 0;
1790 if (!strcmp(k, "pack.deltacachelimit")) {
1791 cache_max_small_delta_size = git_config_int(k, v);
1792 return 0;
1794 if (!strcmp(k, "pack.threads")) {
1795 delta_search_threads = git_config_int(k, v);
1796 if (delta_search_threads < 0)
1797 die("invalid number of threads specified (%d)",
1798 delta_search_threads);
1799 #ifndef THREADED_DELTA_SEARCH
1800 if (delta_search_threads != 1)
1801 warning("no threads support, ignoring %s", k);
1802 #endif
1803 return 0;
1805 if (!strcmp(k, "pack.indexversion")) {
1806 pack_idx_default_version = git_config_int(k, v);
1807 if (pack_idx_default_version > 2)
1808 die("bad pack.indexversion=%d", pack_idx_default_version);
1809 return 0;
1811 if (!strcmp(k, "pack.packsizelimit")) {
1812 pack_size_limit_cfg = git_config_ulong(k, v);
1813 return 0;
1815 return git_default_config(k, v);
1818 static void read_object_list_from_stdin(void)
1820 char line[40 + 1 + PATH_MAX + 2];
1821 unsigned char sha1[20];
1823 for (;;) {
1824 if (!fgets(line, sizeof(line), stdin)) {
1825 if (feof(stdin))
1826 break;
1827 if (!ferror(stdin))
1828 die("fgets returned NULL, not EOF, not error!");
1829 if (errno != EINTR)
1830 die("fgets: %s", strerror(errno));
1831 clearerr(stdin);
1832 continue;
1834 if (line[0] == '-') {
1835 if (get_sha1_hex(line+1, sha1))
1836 die("expected edge sha1, got garbage:\n %s",
1837 line);
1838 add_preferred_base(sha1);
1839 continue;
1841 if (get_sha1_hex(line, sha1))
1842 die("expected sha1, got garbage:\n %s", line);
1844 add_preferred_base_object(line+41);
1845 add_object_entry(sha1, 0, line+41, 0);
1849 #define OBJECT_ADDED (1u<<20)
1851 static void show_commit(struct commit *commit)
1853 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
1854 commit->object.flags |= OBJECT_ADDED;
1857 static void show_object(struct object_array_entry *p)
1859 add_preferred_base_object(p->name);
1860 add_object_entry(p->item->sha1, p->item->type, p->name, 0);
1861 p->item->flags |= OBJECT_ADDED;
1864 static void show_edge(struct commit *commit)
1866 add_preferred_base(commit->object.sha1);
1869 struct in_pack_object {
1870 off_t offset;
1871 struct object *object;
1874 struct in_pack {
1875 int alloc;
1876 int nr;
1877 struct in_pack_object *array;
1880 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
1882 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
1883 in_pack->array[in_pack->nr].object = object;
1884 in_pack->nr++;
1888 * Compare the objects in the offset order, in order to emulate the
1889 * "git-rev-list --objects" output that produced the pack originally.
1891 static int ofscmp(const void *a_, const void *b_)
1893 struct in_pack_object *a = (struct in_pack_object *)a_;
1894 struct in_pack_object *b = (struct in_pack_object *)b_;
1896 if (a->offset < b->offset)
1897 return -1;
1898 else if (a->offset > b->offset)
1899 return 1;
1900 else
1901 return hashcmp(a->object->sha1, b->object->sha1);
1904 static void add_objects_in_unpacked_packs(struct rev_info *revs)
1906 struct packed_git *p;
1907 struct in_pack in_pack;
1908 uint32_t i;
1910 memset(&in_pack, 0, sizeof(in_pack));
1912 for (p = packed_git; p; p = p->next) {
1913 const unsigned char *sha1;
1914 struct object *o;
1916 for (i = 0; i < revs->num_ignore_packed; i++) {
1917 if (matches_pack_name(p, revs->ignore_packed[i]))
1918 break;
1920 if (revs->num_ignore_packed <= i)
1921 continue;
1922 if (open_pack_index(p))
1923 die("cannot open pack index");
1925 ALLOC_GROW(in_pack.array,
1926 in_pack.nr + p->num_objects,
1927 in_pack.alloc);
1929 for (i = 0; i < p->num_objects; i++) {
1930 sha1 = nth_packed_object_sha1(p, i);
1931 o = lookup_unknown_object(sha1);
1932 if (!(o->flags & OBJECT_ADDED))
1933 mark_in_pack_object(o, p, &in_pack);
1934 o->flags |= OBJECT_ADDED;
1938 if (in_pack.nr) {
1939 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
1940 ofscmp);
1941 for (i = 0; i < in_pack.nr; i++) {
1942 struct object *o = in_pack.array[i].object;
1943 add_object_entry(o->sha1, o->type, "", 0);
1946 free(in_pack.array);
1949 static void get_object_list(int ac, const char **av)
1951 struct rev_info revs;
1952 char line[1000];
1953 int flags = 0;
1955 init_revisions(&revs, NULL);
1956 save_commit_buffer = 0;
1957 setup_revisions(ac, av, &revs, NULL);
1959 while (fgets(line, sizeof(line), stdin) != NULL) {
1960 int len = strlen(line);
1961 if (len && line[len - 1] == '\n')
1962 line[--len] = 0;
1963 if (!len)
1964 break;
1965 if (*line == '-') {
1966 if (!strcmp(line, "--not")) {
1967 flags ^= UNINTERESTING;
1968 continue;
1970 die("not a rev '%s'", line);
1972 if (handle_revision_arg(line, &revs, flags, 1))
1973 die("bad revision '%s'", line);
1976 if (prepare_revision_walk(&revs))
1977 die("revision walk setup failed");
1978 mark_edges_uninteresting(revs.commits, &revs, show_edge);
1979 traverse_commit_list(&revs, show_commit, show_object);
1981 if (keep_unreachable)
1982 add_objects_in_unpacked_packs(&revs);
1985 static int adjust_perm(const char *path, mode_t mode)
1987 if (chmod(path, mode))
1988 return -1;
1989 return adjust_shared_perm(path);
1992 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
1994 int use_internal_rev_list = 0;
1995 int thin = 0;
1996 uint32_t i;
1997 const char **rp_av;
1998 int rp_ac_alloc = 64;
1999 int rp_ac;
2001 rp_av = xcalloc(rp_ac_alloc, sizeof(*rp_av));
2003 rp_av[0] = "pack-objects";
2004 rp_av[1] = "--objects"; /* --thin will make it --objects-edge */
2005 rp_ac = 2;
2007 git_config(git_pack_config);
2008 if (!pack_compression_seen && core_compression_seen)
2009 pack_compression_level = core_compression_level;
2011 progress = isatty(2);
2012 for (i = 1; i < argc; i++) {
2013 const char *arg = argv[i];
2015 if (*arg != '-')
2016 break;
2018 if (!strcmp("--non-empty", arg)) {
2019 non_empty = 1;
2020 continue;
2022 if (!strcmp("--local", arg)) {
2023 local = 1;
2024 continue;
2026 if (!strcmp("--incremental", arg)) {
2027 incremental = 1;
2028 continue;
2030 if (!prefixcmp(arg, "--compression=")) {
2031 char *end;
2032 int level = strtoul(arg+14, &end, 0);
2033 if (!arg[14] || *end)
2034 usage(pack_usage);
2035 if (level == -1)
2036 level = Z_DEFAULT_COMPRESSION;
2037 else if (level < 0 || level > Z_BEST_COMPRESSION)
2038 die("bad pack compression level %d", level);
2039 pack_compression_level = level;
2040 continue;
2042 if (!prefixcmp(arg, "--max-pack-size=")) {
2043 char *end;
2044 pack_size_limit_cfg = 0;
2045 pack_size_limit = strtoul(arg+16, &end, 0) * 1024 * 1024;
2046 if (!arg[16] || *end)
2047 usage(pack_usage);
2048 continue;
2050 if (!prefixcmp(arg, "--window=")) {
2051 char *end;
2052 window = strtoul(arg+9, &end, 0);
2053 if (!arg[9] || *end)
2054 usage(pack_usage);
2055 continue;
2057 if (!prefixcmp(arg, "--window-memory=")) {
2058 if (!git_parse_ulong(arg+16, &window_memory_limit))
2059 usage(pack_usage);
2060 continue;
2062 if (!prefixcmp(arg, "--threads=")) {
2063 char *end;
2064 delta_search_threads = strtoul(arg+10, &end, 0);
2065 if (!arg[10] || *end || delta_search_threads < 0)
2066 usage(pack_usage);
2067 #ifndef THREADED_DELTA_SEARCH
2068 if (delta_search_threads != 1)
2069 warning("no threads support, "
2070 "ignoring %s", arg);
2071 #endif
2072 continue;
2074 if (!prefixcmp(arg, "--depth=")) {
2075 char *end;
2076 depth = strtoul(arg+8, &end, 0);
2077 if (!arg[8] || *end)
2078 usage(pack_usage);
2079 continue;
2081 if (!strcmp("--progress", arg)) {
2082 progress = 1;
2083 continue;
2085 if (!strcmp("--all-progress", arg)) {
2086 progress = 2;
2087 continue;
2089 if (!strcmp("-q", arg)) {
2090 progress = 0;
2091 continue;
2093 if (!strcmp("--no-reuse-delta", arg)) {
2094 reuse_delta = 0;
2095 continue;
2097 if (!strcmp("--no-reuse-object", arg)) {
2098 reuse_object = reuse_delta = 0;
2099 continue;
2101 if (!strcmp("--delta-base-offset", arg)) {
2102 allow_ofs_delta = 1;
2103 continue;
2105 if (!strcmp("--stdout", arg)) {
2106 pack_to_stdout = 1;
2107 continue;
2109 if (!strcmp("--revs", arg)) {
2110 use_internal_rev_list = 1;
2111 continue;
2113 if (!strcmp("--keep-unreachable", arg)) {
2114 keep_unreachable = 1;
2115 continue;
2117 if (!strcmp("--include-tag", arg)) {
2118 include_tag = 1;
2119 continue;
2121 if (!strcmp("--unpacked", arg) ||
2122 !prefixcmp(arg, "--unpacked=") ||
2123 !strcmp("--reflog", arg) ||
2124 !strcmp("--all", arg)) {
2125 use_internal_rev_list = 1;
2126 if (rp_ac >= rp_ac_alloc - 1) {
2127 rp_ac_alloc = alloc_nr(rp_ac_alloc);
2128 rp_av = xrealloc(rp_av,
2129 rp_ac_alloc * sizeof(*rp_av));
2131 rp_av[rp_ac++] = arg;
2132 continue;
2134 if (!strcmp("--thin", arg)) {
2135 use_internal_rev_list = 1;
2136 thin = 1;
2137 rp_av[1] = "--objects-edge";
2138 continue;
2140 if (!prefixcmp(arg, "--index-version=")) {
2141 char *c;
2142 pack_idx_default_version = strtoul(arg + 16, &c, 10);
2143 if (pack_idx_default_version > 2)
2144 die("bad %s", arg);
2145 if (*c == ',')
2146 pack_idx_off32_limit = strtoul(c+1, &c, 0);
2147 if (*c || pack_idx_off32_limit & 0x80000000)
2148 die("bad %s", arg);
2149 continue;
2151 usage(pack_usage);
2154 /* Traditionally "pack-objects [options] base extra" failed;
2155 * we would however want to take refs parameter that would
2156 * have been given to upstream rev-list ourselves, which means
2157 * we somehow want to say what the base name is. So the
2158 * syntax would be:
2160 * pack-objects [options] base <refs...>
2162 * in other words, we would treat the first non-option as the
2163 * base_name and send everything else to the internal revision
2164 * walker.
2167 if (!pack_to_stdout)
2168 base_name = argv[i++];
2170 if (pack_to_stdout != !base_name)
2171 usage(pack_usage);
2173 if (!pack_to_stdout && !pack_size_limit)
2174 pack_size_limit = pack_size_limit_cfg;
2176 if (pack_to_stdout && pack_size_limit)
2177 die("--max-pack-size cannot be used to build a pack for transfer.");
2179 if (!pack_to_stdout && thin)
2180 die("--thin cannot be used to build an indexable pack.");
2182 #ifdef THREADED_DELTA_SEARCH
2183 if (!delta_search_threads) /* --threads=0 means autodetect */
2184 delta_search_threads = online_cpus();
2185 #endif
2187 prepare_packed_git();
2189 if (progress)
2190 progress_state = start_progress("Counting objects", 0);
2191 if (!use_internal_rev_list)
2192 read_object_list_from_stdin();
2193 else {
2194 rp_av[rp_ac] = NULL;
2195 get_object_list(rp_ac, rp_av);
2197 if (include_tag && nr_result)
2198 for_each_ref(add_ref_tag, NULL);
2199 stop_progress(&progress_state);
2201 if (non_empty && !nr_result)
2202 return 0;
2203 if (nr_result)
2204 prepare_pack(window, depth);
2205 write_pack_file();
2206 if (progress)
2207 fprintf(stderr, "Total %u (delta %u), reused %u (delta %u)\n",
2208 written, written_delta, reused, reused_delta);
2209 return 0;