11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
16 #include "list-objects.h"
17 #include "pack-objects.h"
20 #include "streaming.h"
21 #include "thread-utils.h"
22 #include "pack-bitmap.h"
24 static const char *pack_usage[] = {
25 N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
26 N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
31 * Objects we are going to pack are collected in the `to_pack` structure.
32 * It contains an array (dynamically expanded) of the object data, and a map
33 * that can resolve SHA1s to their position in the array.
35 static struct packing_data to_pack;
37 static struct pack_idx_entry **written_list;
38 static uint32_t nr_result, nr_written;
41 static int reuse_delta = 1, reuse_object = 1;
42 static int keep_unreachable, unpack_unreachable, include_tag;
43 static unsigned long unpack_unreachable_expiration;
45 static int incremental;
46 static int ignore_packed_keep;
47 static int allow_ofs_delta;
48 static struct pack_idx_option pack_idx_opts;
49 static const char *base_name;
50 static int progress = 1;
51 static int window = 10;
52 static unsigned long pack_size_limit;
53 static int depth = 50;
54 static int delta_search_threads;
55 static int pack_to_stdout;
56 static int num_preferred_base;
57 static struct progress *progress_state;
58 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
59 static int pack_compression_seen;
61 static struct packed_git *reuse_packfile;
62 static uint32_t reuse_packfile_objects;
63 static off_t reuse_packfile_offset;
65 static int use_bitmap_index = 1;
66 static int write_bitmap_index;
67 static uint16_t write_bitmap_options;
69 static unsigned long delta_cache_size = 0;
70 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
71 static unsigned long cache_max_small_delta_size = 1000;
73 static unsigned long window_memory_limit = 0;
78 static uint32_t written, written_delta;
79 static uint32_t reused, reused_delta;
84 static struct commit **indexed_commits;
85 static unsigned int indexed_commits_nr;
86 static unsigned int indexed_commits_alloc;
88 static void index_commit_for_bitmap(struct commit *commit)
90 if (indexed_commits_nr >= indexed_commits_alloc) {
91 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
92 indexed_commits = xrealloc(indexed_commits,
93 indexed_commits_alloc * sizeof(struct commit *));
96 indexed_commits[indexed_commits_nr++] = commit;
99 static void *get_delta(struct object_entry *entry)
101 unsigned long size, base_size, delta_size;
102 void *buf, *base_buf, *delta_buf;
103 enum object_type type;
105 buf = read_sha1_file(entry->idx.sha1, &type, &size);
107 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
108 base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
110 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
111 delta_buf = diff_delta(base_buf, base_size,
112 buf, size, &delta_size, 0);
113 if (!delta_buf || delta_size != entry->delta_size)
114 die("delta size changed");
120 static unsigned long do_compress(void **pptr, unsigned long size)
124 unsigned long maxsize;
126 memset(&stream, 0, sizeof(stream));
127 git_deflate_init(&stream, pack_compression_level);
128 maxsize = git_deflate_bound(&stream, size);
131 out = xmalloc(maxsize);
135 stream.avail_in = size;
136 stream.next_out = out;
137 stream.avail_out = maxsize;
138 while (git_deflate(&stream, Z_FINISH) == Z_OK)
140 git_deflate_end(&stream);
143 return stream.total_out;
146 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
147 const unsigned char *sha1)
150 unsigned char ibuf[1024 * 16];
151 unsigned char obuf[1024 * 16];
152 unsigned long olen = 0;
154 memset(&stream, 0, sizeof(stream));
155 git_deflate_init(&stream, pack_compression_level);
160 readlen = read_istream(st, ibuf, sizeof(ibuf));
162 die(_("unable to read %s"), sha1_to_hex(sha1));
164 stream.next_in = ibuf;
165 stream.avail_in = readlen;
166 while ((stream.avail_in || readlen == 0) &&
167 (zret == Z_OK || zret == Z_BUF_ERROR)) {
168 stream.next_out = obuf;
169 stream.avail_out = sizeof(obuf);
170 zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
171 sha1write(f, obuf, stream.next_out - obuf);
172 olen += stream.next_out - obuf;
175 die(_("deflate error (%d)"), zret);
177 if (zret != Z_STREAM_END)
178 die(_("deflate error (%d)"), zret);
182 git_deflate_end(&stream);
187 * we are going to reuse the existing object data as is. make
188 * sure it is not corrupt.
190 static int check_pack_inflate(struct packed_git *p,
191 struct pack_window **w_curs,
194 unsigned long expect)
197 unsigned char fakebuf[4096], *in;
200 memset(&stream, 0, sizeof(stream));
201 git_inflate_init(&stream);
203 in = use_pack(p, w_curs, offset, &stream.avail_in);
205 stream.next_out = fakebuf;
206 stream.avail_out = sizeof(fakebuf);
207 st = git_inflate(&stream, Z_FINISH);
208 offset += stream.next_in - in;
209 } while (st == Z_OK || st == Z_BUF_ERROR);
210 git_inflate_end(&stream);
211 return (st == Z_STREAM_END &&
212 stream.total_out == expect &&
213 stream.total_in == len) ? 0 : -1;
216 static void copy_pack_data(struct sha1file *f,
217 struct packed_git *p,
218 struct pack_window **w_curs,
226 in = use_pack(p, w_curs, offset, &avail);
228 avail = (unsigned long)len;
229 sha1write(f, in, avail);
235 /* Return 0 if we will bust the pack-size limit */
236 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
237 unsigned long limit, int usable_delta)
239 unsigned long size, datalen;
240 unsigned char header[10], dheader[10];
242 enum object_type type;
244 struct git_istream *st = NULL;
247 if (entry->type == OBJ_BLOB &&
248 entry->size > big_file_threshold &&
249 (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
252 buf = read_sha1_file(entry->idx.sha1, &type, &size);
254 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
257 * make sure no cached delta data remains from a
258 * previous attempt before a pack split occurred.
260 free(entry->delta_data);
261 entry->delta_data = NULL;
262 entry->z_delta_size = 0;
263 } else if (entry->delta_data) {
264 size = entry->delta_size;
265 buf = entry->delta_data;
266 entry->delta_data = NULL;
267 type = (allow_ofs_delta && entry->delta->idx.offset) ?
268 OBJ_OFS_DELTA : OBJ_REF_DELTA;
270 buf = get_delta(entry);
271 size = entry->delta_size;
272 type = (allow_ofs_delta && entry->delta->idx.offset) ?
273 OBJ_OFS_DELTA : OBJ_REF_DELTA;
276 if (st) /* large blob case, just assume we don't compress well */
278 else if (entry->z_delta_size)
279 datalen = entry->z_delta_size;
281 datalen = do_compress(&buf, size);
284 * The object header is a byte of 'type' followed by zero or
285 * more bytes of length.
287 hdrlen = encode_in_pack_object_header(type, size, header);
289 if (type == OBJ_OFS_DELTA) {
291 * Deltas with relative base contain an additional
292 * encoding of the relative offset for the delta
293 * base from this object's position in the pack.
295 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
296 unsigned pos = sizeof(dheader) - 1;
297 dheader[pos] = ofs & 127;
299 dheader[--pos] = 128 | (--ofs & 127);
300 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
306 sha1write(f, header, hdrlen);
307 sha1write(f, dheader + pos, sizeof(dheader) - pos);
308 hdrlen += sizeof(dheader) - pos;
309 } else if (type == OBJ_REF_DELTA) {
311 * Deltas with a base reference contain
312 * an additional 20 bytes for the base sha1.
314 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
320 sha1write(f, header, hdrlen);
321 sha1write(f, entry->delta->idx.sha1, 20);
324 if (limit && hdrlen + datalen + 20 >= limit) {
330 sha1write(f, header, hdrlen);
333 datalen = write_large_blob_data(st, f, entry->idx.sha1);
336 sha1write(f, buf, datalen);
340 return hdrlen + datalen;
343 /* Return 0 if we will bust the pack-size limit */
344 static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
345 unsigned long limit, int usable_delta)
347 struct packed_git *p = entry->in_pack;
348 struct pack_window *w_curs = NULL;
349 struct revindex_entry *revidx;
351 enum object_type type = entry->type;
352 unsigned long datalen;
353 unsigned char header[10], dheader[10];
357 type = (allow_ofs_delta && entry->delta->idx.offset) ?
358 OBJ_OFS_DELTA : OBJ_REF_DELTA;
359 hdrlen = encode_in_pack_object_header(type, entry->size, header);
361 offset = entry->in_pack_offset;
362 revidx = find_pack_revindex(p, offset);
363 datalen = revidx[1].offset - offset;
364 if (!pack_to_stdout && p->index_version > 1 &&
365 check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
366 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
368 return write_no_reuse_object(f, entry, limit, usable_delta);
371 offset += entry->in_pack_header_size;
372 datalen -= entry->in_pack_header_size;
374 if (!pack_to_stdout && p->index_version == 1 &&
375 check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
376 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
378 return write_no_reuse_object(f, entry, limit, usable_delta);
381 if (type == OBJ_OFS_DELTA) {
382 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
383 unsigned pos = sizeof(dheader) - 1;
384 dheader[pos] = ofs & 127;
386 dheader[--pos] = 128 | (--ofs & 127);
387 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
391 sha1write(f, header, hdrlen);
392 sha1write(f, dheader + pos, sizeof(dheader) - pos);
393 hdrlen += sizeof(dheader) - pos;
395 } else if (type == OBJ_REF_DELTA) {
396 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
400 sha1write(f, header, hdrlen);
401 sha1write(f, entry->delta->idx.sha1, 20);
405 if (limit && hdrlen + datalen + 20 >= limit) {
409 sha1write(f, header, hdrlen);
411 copy_pack_data(f, p, &w_curs, offset, datalen);
414 return hdrlen + datalen;
417 /* Return 0 if we will bust the pack-size limit */
418 static unsigned long write_object(struct sha1file *f,
419 struct object_entry *entry,
422 unsigned long limit, len;
423 int usable_delta, to_reuse;
428 /* apply size limit if limited packsize and not first object */
429 if (!pack_size_limit || !nr_written)
431 else if (pack_size_limit <= write_offset)
433 * the earlier object did not fit the limit; avoid
434 * mistaking this with unlimited (i.e. limit = 0).
438 limit = pack_size_limit - write_offset;
441 usable_delta = 0; /* no delta */
442 else if (!pack_size_limit)
443 usable_delta = 1; /* unlimited packfile */
444 else if (entry->delta->idx.offset == (off_t)-1)
445 usable_delta = 0; /* base was written to another pack */
446 else if (entry->delta->idx.offset)
447 usable_delta = 1; /* base already exists in this pack */
449 usable_delta = 0; /* base could end up in another pack */
452 to_reuse = 0; /* explicit */
453 else if (!entry->in_pack)
454 to_reuse = 0; /* can't reuse what we don't have */
455 else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
456 /* check_object() decided it for us ... */
457 to_reuse = usable_delta;
458 /* ... but pack split may override that */
459 else if (entry->type != entry->in_pack_type)
460 to_reuse = 0; /* pack has delta which is unusable */
461 else if (entry->delta)
462 to_reuse = 0; /* we want to pack afresh */
464 to_reuse = 1; /* we have it in-pack undeltified,
465 * and we do not need to deltify it.
469 len = write_no_reuse_object(f, entry, limit, usable_delta);
471 len = write_reuse_object(f, entry, limit, usable_delta);
479 entry->idx.crc32 = crc32_end(f);
483 enum write_one_status {
484 WRITE_ONE_SKIP = -1, /* already written */
485 WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
486 WRITE_ONE_WRITTEN = 1, /* normal */
487 WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
490 static enum write_one_status write_one(struct sha1file *f,
491 struct object_entry *e,
498 * we set offset to 1 (which is an impossible value) to mark
499 * the fact that this object is involved in "write its base
500 * first before writing a deltified object" recursion.
502 recursing = (e->idx.offset == 1);
504 warning("recursive delta detected for object %s",
505 sha1_to_hex(e->idx.sha1));
506 return WRITE_ONE_RECURSIVE;
507 } else if (e->idx.offset || e->preferred_base) {
508 /* offset is non zero if object is written already. */
509 return WRITE_ONE_SKIP;
512 /* if we are deltified, write out base object first. */
514 e->idx.offset = 1; /* now recurse */
515 switch (write_one(f, e->delta, offset)) {
516 case WRITE_ONE_RECURSIVE:
517 /* we cannot depend on this one */
522 case WRITE_ONE_BREAK:
523 e->idx.offset = recursing;
524 return WRITE_ONE_BREAK;
528 e->idx.offset = *offset;
529 size = write_object(f, e, *offset);
531 e->idx.offset = recursing;
532 return WRITE_ONE_BREAK;
534 written_list[nr_written++] = &e->idx;
536 /* make sure off_t is sufficiently large not to wrap */
537 if (signed_add_overflows(*offset, size))
538 die("pack too large for current definition of off_t");
540 return WRITE_ONE_WRITTEN;
543 static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
546 unsigned char peeled[20];
547 struct object_entry *entry = packlist_find(&to_pack, sha1, NULL);
551 if (!peel_ref(path, peeled)) {
552 entry = packlist_find(&to_pack, peeled, NULL);
559 static inline void add_to_write_order(struct object_entry **wo,
561 struct object_entry *e)
569 static void add_descendants_to_write_order(struct object_entry **wo,
571 struct object_entry *e)
573 int add_to_order = 1;
576 struct object_entry *s;
577 /* add this node... */
578 add_to_write_order(wo, endp, e);
579 /* all its siblings... */
580 for (s = e->delta_sibling; s; s = s->delta_sibling) {
581 add_to_write_order(wo, endp, s);
584 /* drop down a level to add left subtree nodes if possible */
585 if (e->delta_child) {
590 /* our sibling might have some children, it is next */
591 if (e->delta_sibling) {
592 e = e->delta_sibling;
595 /* go back to our parent node */
597 while (e && !e->delta_sibling) {
598 /* we're on the right side of a subtree, keep
599 * going up until we can go right again */
603 /* done- we hit our original root node */
606 /* pass it off to sibling at this level */
607 e = e->delta_sibling;
612 static void add_family_to_write_order(struct object_entry **wo,
614 struct object_entry *e)
616 struct object_entry *root;
618 for (root = e; root->delta; root = root->delta)
620 add_descendants_to_write_order(wo, endp, root);
623 static struct object_entry **compute_write_order(void)
625 unsigned int i, wo_end, last_untagged;
627 struct object_entry **wo = xmalloc(to_pack.nr_objects * sizeof(*wo));
628 struct object_entry *objects = to_pack.objects;
630 for (i = 0; i < to_pack.nr_objects; i++) {
631 objects[i].tagged = 0;
632 objects[i].filled = 0;
633 objects[i].delta_child = NULL;
634 objects[i].delta_sibling = NULL;
638 * Fully connect delta_child/delta_sibling network.
639 * Make sure delta_sibling is sorted in the original
642 for (i = to_pack.nr_objects; i > 0;) {
643 struct object_entry *e = &objects[--i];
646 /* Mark me as the first child */
647 e->delta_sibling = e->delta->delta_child;
648 e->delta->delta_child = e;
652 * Mark objects that are at the tip of tags.
654 for_each_tag_ref(mark_tagged, NULL);
657 * Give the objects in the original recency order until
658 * we see a tagged tip.
660 for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
661 if (objects[i].tagged)
663 add_to_write_order(wo, &wo_end, &objects[i]);
668 * Then fill all the tagged tips.
670 for (; i < to_pack.nr_objects; i++) {
671 if (objects[i].tagged)
672 add_to_write_order(wo, &wo_end, &objects[i]);
676 * And then all remaining commits and tags.
678 for (i = last_untagged; i < to_pack.nr_objects; i++) {
679 if (objects[i].type != OBJ_COMMIT &&
680 objects[i].type != OBJ_TAG)
682 add_to_write_order(wo, &wo_end, &objects[i]);
686 * And then all the trees.
688 for (i = last_untagged; i < to_pack.nr_objects; i++) {
689 if (objects[i].type != OBJ_TREE)
691 add_to_write_order(wo, &wo_end, &objects[i]);
695 * Finally all the rest in really tight order
697 for (i = last_untagged; i < to_pack.nr_objects; i++) {
698 if (!objects[i].filled)
699 add_family_to_write_order(wo, &wo_end, &objects[i]);
702 if (wo_end != to_pack.nr_objects)
703 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
708 static off_t write_reused_pack(struct sha1file *f)
710 unsigned char buffer[8192];
714 if (!is_pack_valid(reuse_packfile))
715 die("packfile is invalid: %s", reuse_packfile->pack_name);
717 fd = git_open_noatime(reuse_packfile->pack_name);
719 die_errno("unable to open packfile for reuse: %s",
720 reuse_packfile->pack_name);
722 if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
723 die_errno("unable to seek in reused packfile");
725 if (reuse_packfile_offset < 0)
726 reuse_packfile_offset = reuse_packfile->pack_size - 20;
728 to_write = reuse_packfile_offset - sizeof(struct pack_header);
731 int read_pack = xread(fd, buffer, sizeof(buffer));
734 die_errno("unable to read from reused packfile");
736 if (read_pack > to_write)
737 read_pack = to_write;
739 sha1write(f, buffer, read_pack);
740 to_write -= read_pack;
744 written += reuse_packfile_objects;
745 return reuse_packfile_offset - sizeof(struct pack_header);
748 static void write_pack_file(void)
753 uint32_t nr_remaining = nr_result;
754 time_t last_mtime = 0;
755 struct object_entry **write_order;
757 if (progress > pack_to_stdout)
758 progress_state = start_progress("Writing objects", nr_result);
759 written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
760 write_order = compute_write_order();
763 unsigned char sha1[20];
764 char *pack_tmp_name = NULL;
767 f = sha1fd_throughput(1, "<stdout>", progress_state);
769 f = create_tmp_packfile(&pack_tmp_name);
771 offset = write_pack_header(f, nr_remaining);
773 if (reuse_packfile) {
775 assert(pack_to_stdout);
777 packfile_size = write_reused_pack(f);
778 offset += packfile_size;
782 for (; i < to_pack.nr_objects; i++) {
783 struct object_entry *e = write_order[i];
784 if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
786 display_progress(progress_state, written);
790 * Did we write the wrong # entries in the header?
791 * If so, rewrite it like in fast-import
793 if (pack_to_stdout) {
794 sha1close(f, sha1, CSUM_CLOSE);
795 } else if (nr_written == nr_remaining) {
796 sha1close(f, sha1, CSUM_FSYNC);
798 int fd = sha1close(f, sha1, 0);
799 fixup_pack_header_footer(fd, sha1, pack_tmp_name,
800 nr_written, sha1, offset);
804 if (!pack_to_stdout) {
806 char tmpname[PATH_MAX];
809 * Packs are runtime accessed in their mtime
810 * order since newer packs are more likely to contain
811 * younger objects. So if we are creating multiple
812 * packs then we should modify the mtime of later ones
813 * to preserve this property.
815 if (stat(pack_tmp_name, &st) < 0) {
816 warning("failed to stat %s: %s",
817 pack_tmp_name, strerror(errno));
818 } else if (!last_mtime) {
819 last_mtime = st.st_mtime;
822 utb.actime = st.st_atime;
823 utb.modtime = --last_mtime;
824 if (utime(pack_tmp_name, &utb) < 0)
825 warning("failed utime() on %s: %s",
826 tmpname, strerror(errno));
829 /* Enough space for "-<sha-1>.pack"? */
830 if (sizeof(tmpname) <= strlen(base_name) + 50)
831 die("pack base name '%s' too long", base_name);
832 snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
834 if (write_bitmap_index) {
835 bitmap_writer_set_checksum(sha1);
836 bitmap_writer_build_type_index(written_list, nr_written);
839 finish_tmp_packfile(tmpname, pack_tmp_name,
840 written_list, nr_written,
841 &pack_idx_opts, sha1);
843 if (write_bitmap_index) {
844 char *end_of_name_prefix = strrchr(tmpname, 0);
845 sprintf(end_of_name_prefix, "%s.bitmap", sha1_to_hex(sha1));
847 stop_progress(&progress_state);
849 bitmap_writer_show_progress(progress);
850 bitmap_writer_reuse_bitmaps(&to_pack);
851 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
852 bitmap_writer_build(&to_pack);
853 bitmap_writer_finish(written_list, nr_written,
854 tmpname, write_bitmap_options);
855 write_bitmap_index = 0;
859 puts(sha1_to_hex(sha1));
862 /* mark written objects as written to previous pack */
863 for (j = 0; j < nr_written; j++) {
864 written_list[j]->offset = (off_t)-1;
866 nr_remaining -= nr_written;
867 } while (nr_remaining && i < to_pack.nr_objects);
871 stop_progress(&progress_state);
872 if (written != nr_result)
873 die("wrote %"PRIu32" objects while expecting %"PRIu32,
877 static void setup_delta_attr_check(struct git_attr_check *check)
879 static struct git_attr *attr_delta;
882 attr_delta = git_attr("delta");
884 check[0].attr = attr_delta;
887 static int no_try_delta(const char *path)
889 struct git_attr_check check[1];
891 setup_delta_attr_check(check);
892 if (git_check_attr(path, ARRAY_SIZE(check), check))
894 if (ATTR_FALSE(check->value))
900 * When adding an object, check whether we have already added it
901 * to our packing list. If so, we can skip. However, if we are
902 * being asked to excludei t, but the previous mention was to include
903 * it, make sure to adjust its flags and tweak our numbers accordingly.
905 * As an optimization, we pass out the index position where we would have
906 * found the item, since that saves us from having to look it up again a
907 * few lines later when we want to add the new entry.
909 static int have_duplicate_entry(const unsigned char *sha1,
913 struct object_entry *entry;
915 entry = packlist_find(&to_pack, sha1, index_pos);
920 if (!entry->preferred_base)
922 entry->preferred_base = 1;
929 * Check whether we want the object in the pack (e.g., we do not want
930 * objects found in non-local stores if the "--local" option was used).
932 * As a side effect of this check, we will find the packed version of this
933 * object, if any. We therefore pass out the pack information to avoid having
934 * to look it up again later.
936 static int want_object_in_pack(const unsigned char *sha1,
938 struct packed_git **found_pack,
941 struct packed_git *p;
943 if (!exclude && local && has_loose_object_nonlocal(sha1))
949 for (p = packed_git; p; p = p->next) {
950 off_t offset = find_pack_entry_one(sha1, p);
953 if (!is_pack_valid(p)) {
954 warning("packfile %s cannot be accessed", p->pack_name);
957 *found_offset = offset;
964 if (local && !p->pack_local)
966 if (ignore_packed_keep && p->pack_local && p->pack_keep)
974 static void create_object_entry(const unsigned char *sha1,
975 enum object_type type,
980 struct packed_git *found_pack,
983 struct object_entry *entry;
985 entry = packlist_alloc(&to_pack, sha1, index_pos);
990 entry->preferred_base = 1;
994 entry->in_pack = found_pack;
995 entry->in_pack_offset = found_offset;
998 entry->no_try_delta = no_try_delta;
1001 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1002 const char *name, int exclude)
1004 struct packed_git *found_pack;
1008 if (have_duplicate_entry(sha1, exclude, &index_pos))
1011 if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset))
1014 create_object_entry(sha1, type, pack_name_hash(name),
1015 exclude, name && no_try_delta(name),
1016 index_pos, found_pack, found_offset);
1018 display_progress(progress_state, to_pack.nr_objects);
1022 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1023 enum object_type type,
1024 int flags, uint32_t name_hash,
1025 struct packed_git *pack, off_t offset)
1029 if (have_duplicate_entry(sha1, 0, &index_pos))
1032 create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1034 display_progress(progress_state, to_pack.nr_objects);
1038 struct pbase_tree_cache {
1039 unsigned char sha1[20];
1043 unsigned long tree_size;
1046 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1047 static int pbase_tree_cache_ix(const unsigned char *sha1)
1049 return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1051 static int pbase_tree_cache_ix_incr(int ix)
1053 return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1056 static struct pbase_tree {
1057 struct pbase_tree *next;
1058 /* This is a phony "cache" entry; we are not
1059 * going to evict it nor find it through _get()
1060 * mechanism -- this is for the toplevel node that
1061 * would almost always change with any commit.
1063 struct pbase_tree_cache pcache;
1066 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1068 struct pbase_tree_cache *ent, *nent;
1071 enum object_type type;
1073 int my_ix = pbase_tree_cache_ix(sha1);
1074 int available_ix = -1;
1076 /* pbase-tree-cache acts as a limited hashtable.
1077 * your object will be found at your index or within a few
1078 * slots after that slot if it is cached.
1080 for (neigh = 0; neigh < 8; neigh++) {
1081 ent = pbase_tree_cache[my_ix];
1082 if (ent && !hashcmp(ent->sha1, sha1)) {
1086 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1087 ((0 <= available_ix) &&
1088 (!ent && pbase_tree_cache[available_ix])))
1089 available_ix = my_ix;
1092 my_ix = pbase_tree_cache_ix_incr(my_ix);
1095 /* Did not find one. Either we got a bogus request or
1096 * we need to read and perhaps cache.
1098 data = read_sha1_file(sha1, &type, &size);
1101 if (type != OBJ_TREE) {
1106 /* We need to either cache or return a throwaway copy */
1108 if (available_ix < 0)
1111 ent = pbase_tree_cache[available_ix];
1112 my_ix = available_ix;
1116 nent = xmalloc(sizeof(*nent));
1117 nent->temporary = (available_ix < 0);
1120 /* evict and reuse */
1121 free(ent->tree_data);
1124 hashcpy(nent->sha1, sha1);
1125 nent->tree_data = data;
1126 nent->tree_size = size;
1128 if (!nent->temporary)
1129 pbase_tree_cache[my_ix] = nent;
1133 static void pbase_tree_put(struct pbase_tree_cache *cache)
1135 if (!cache->temporary) {
1139 free(cache->tree_data);
1143 static int name_cmp_len(const char *name)
1146 for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1151 static void add_pbase_object(struct tree_desc *tree,
1154 const char *fullname)
1156 struct name_entry entry;
1159 while (tree_entry(tree,&entry)) {
1160 if (S_ISGITLINK(entry.mode))
1162 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1163 memcmp(name, entry.path, cmplen);
1168 if (name[cmplen] != '/') {
1169 add_object_entry(entry.sha1,
1170 object_type(entry.mode),
1174 if (S_ISDIR(entry.mode)) {
1175 struct tree_desc sub;
1176 struct pbase_tree_cache *tree;
1177 const char *down = name+cmplen+1;
1178 int downlen = name_cmp_len(down);
1180 tree = pbase_tree_get(entry.sha1);
1183 init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1185 add_pbase_object(&sub, down, downlen, fullname);
1186 pbase_tree_put(tree);
1191 static unsigned *done_pbase_paths;
1192 static int done_pbase_paths_num;
1193 static int done_pbase_paths_alloc;
1194 static int done_pbase_path_pos(unsigned hash)
1197 int hi = done_pbase_paths_num;
1199 int mi = (hi + lo) / 2;
1200 if (done_pbase_paths[mi] == hash)
1202 if (done_pbase_paths[mi] < hash)
1210 static int check_pbase_path(unsigned hash)
1212 int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1216 if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1217 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1218 done_pbase_paths = xrealloc(done_pbase_paths,
1219 done_pbase_paths_alloc *
1222 done_pbase_paths_num++;
1223 if (pos < done_pbase_paths_num)
1224 memmove(done_pbase_paths + pos + 1,
1225 done_pbase_paths + pos,
1226 (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1227 done_pbase_paths[pos] = hash;
1231 static void add_preferred_base_object(const char *name)
1233 struct pbase_tree *it;
1235 unsigned hash = pack_name_hash(name);
1237 if (!num_preferred_base || check_pbase_path(hash))
1240 cmplen = name_cmp_len(name);
1241 for (it = pbase_tree; it; it = it->next) {
1243 add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1246 struct tree_desc tree;
1247 init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1248 add_pbase_object(&tree, name, cmplen, name);
1253 static void add_preferred_base(unsigned char *sha1)
1255 struct pbase_tree *it;
1258 unsigned char tree_sha1[20];
1260 if (window <= num_preferred_base++)
1263 data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1267 for (it = pbase_tree; it; it = it->next) {
1268 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1274 it = xcalloc(1, sizeof(*it));
1275 it->next = pbase_tree;
1278 hashcpy(it->pcache.sha1, tree_sha1);
1279 it->pcache.tree_data = data;
1280 it->pcache.tree_size = size;
1283 static void cleanup_preferred_base(void)
1285 struct pbase_tree *it;
1291 struct pbase_tree *this = it;
1293 free(this->pcache.tree_data);
1297 for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1298 if (!pbase_tree_cache[i])
1300 free(pbase_tree_cache[i]->tree_data);
1301 free(pbase_tree_cache[i]);
1302 pbase_tree_cache[i] = NULL;
1305 free(done_pbase_paths);
1306 done_pbase_paths = NULL;
1307 done_pbase_paths_num = done_pbase_paths_alloc = 0;
1310 static void check_object(struct object_entry *entry)
1312 if (entry->in_pack) {
1313 struct packed_git *p = entry->in_pack;
1314 struct pack_window *w_curs = NULL;
1315 const unsigned char *base_ref = NULL;
1316 struct object_entry *base_entry;
1317 unsigned long used, used_0;
1318 unsigned long avail;
1320 unsigned char *buf, c;
1322 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1325 * We want in_pack_type even if we do not reuse delta
1326 * since non-delta representations could still be reused.
1328 used = unpack_object_header_buffer(buf, avail,
1329 &entry->in_pack_type,
1335 * Determine if this is a delta and if so whether we can
1336 * reuse it or not. Otherwise let's find out as cheaply as
1337 * possible what the actual type and size for this object is.
1339 switch (entry->in_pack_type) {
1341 /* Not a delta hence we've already got all we need. */
1342 entry->type = entry->in_pack_type;
1343 entry->in_pack_header_size = used;
1344 if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1346 unuse_pack(&w_curs);
1349 if (reuse_delta && !entry->preferred_base)
1350 base_ref = use_pack(p, &w_curs,
1351 entry->in_pack_offset + used, NULL);
1352 entry->in_pack_header_size = used + 20;
1355 buf = use_pack(p, &w_curs,
1356 entry->in_pack_offset + used, NULL);
1362 if (!ofs || MSB(ofs, 7)) {
1363 error("delta base offset overflow in pack for %s",
1364 sha1_to_hex(entry->idx.sha1));
1368 ofs = (ofs << 7) + (c & 127);
1370 ofs = entry->in_pack_offset - ofs;
1371 if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1372 error("delta base offset out of bound for %s",
1373 sha1_to_hex(entry->idx.sha1));
1376 if (reuse_delta && !entry->preferred_base) {
1377 struct revindex_entry *revidx;
1378 revidx = find_pack_revindex(p, ofs);
1381 base_ref = nth_packed_object_sha1(p, revidx->nr);
1383 entry->in_pack_header_size = used + used_0;
1387 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1389 * If base_ref was set above that means we wish to
1390 * reuse delta data, and we even found that base
1391 * in the list of objects we want to pack. Goodie!
1393 * Depth value does not matter - find_deltas() will
1394 * never consider reused delta as the base object to
1395 * deltify other objects against, in order to avoid
1398 entry->type = entry->in_pack_type;
1399 entry->delta = base_entry;
1400 entry->delta_size = entry->size;
1401 entry->delta_sibling = base_entry->delta_child;
1402 base_entry->delta_child = entry;
1403 unuse_pack(&w_curs);
1409 * This must be a delta and we already know what the
1410 * final object type is. Let's extract the actual
1411 * object size from the delta header.
1413 entry->size = get_size_from_delta(p, &w_curs,
1414 entry->in_pack_offset + entry->in_pack_header_size);
1415 if (entry->size == 0)
1417 unuse_pack(&w_curs);
1422 * No choice but to fall back to the recursive delta walk
1423 * with sha1_object_info() to find about the object type
1427 unuse_pack(&w_curs);
1430 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1432 * The error condition is checked in prepare_pack(). This is
1433 * to permit a missing preferred base object to be ignored
1434 * as a preferred base. Doing so can result in a larger
1435 * pack file, but the transfer will still take place.
1439 static int pack_offset_sort(const void *_a, const void *_b)
1441 const struct object_entry *a = *(struct object_entry **)_a;
1442 const struct object_entry *b = *(struct object_entry **)_b;
1444 /* avoid filesystem trashing with loose objects */
1445 if (!a->in_pack && !b->in_pack)
1446 return hashcmp(a->idx.sha1, b->idx.sha1);
1448 if (a->in_pack < b->in_pack)
1450 if (a->in_pack > b->in_pack)
1452 return a->in_pack_offset < b->in_pack_offset ? -1 :
1453 (a->in_pack_offset > b->in_pack_offset);
1456 static void get_object_details(void)
1459 struct object_entry **sorted_by_offset;
1461 sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1462 for (i = 0; i < to_pack.nr_objects; i++)
1463 sorted_by_offset[i] = to_pack.objects + i;
1464 qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1466 for (i = 0; i < to_pack.nr_objects; i++) {
1467 struct object_entry *entry = sorted_by_offset[i];
1468 check_object(entry);
1469 if (big_file_threshold < entry->size)
1470 entry->no_try_delta = 1;
1473 free(sorted_by_offset);
1477 * We search for deltas in a list sorted by type, by filename hash, and then
1478 * by size, so that we see progressively smaller and smaller files.
1479 * That's because we prefer deltas to be from the bigger file
1480 * to the smaller -- deletes are potentially cheaper, but perhaps
1481 * more importantly, the bigger file is likely the more recent
1482 * one. The deepest deltas are therefore the oldest objects which are
1483 * less susceptible to be accessed often.
1485 static int type_size_sort(const void *_a, const void *_b)
1487 const struct object_entry *a = *(struct object_entry **)_a;
1488 const struct object_entry *b = *(struct object_entry **)_b;
1490 if (a->type > b->type)
1492 if (a->type < b->type)
1494 if (a->hash > b->hash)
1496 if (a->hash < b->hash)
1498 if (a->preferred_base > b->preferred_base)
1500 if (a->preferred_base < b->preferred_base)
1502 if (a->size > b->size)
1504 if (a->size < b->size)
1506 return a < b ? -1 : (a > b); /* newest first */
1510 struct object_entry *entry;
1512 struct delta_index *index;
1516 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1517 unsigned long delta_size)
1519 if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1522 if (delta_size < cache_max_small_delta_size)
1525 /* cache delta, if objects are large enough compared to delta size */
1526 if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1534 static pthread_mutex_t read_mutex;
1535 #define read_lock() pthread_mutex_lock(&read_mutex)
1536 #define read_unlock() pthread_mutex_unlock(&read_mutex)
1538 static pthread_mutex_t cache_mutex;
1539 #define cache_lock() pthread_mutex_lock(&cache_mutex)
1540 #define cache_unlock() pthread_mutex_unlock(&cache_mutex)
1542 static pthread_mutex_t progress_mutex;
1543 #define progress_lock() pthread_mutex_lock(&progress_mutex)
1544 #define progress_unlock() pthread_mutex_unlock(&progress_mutex)
1548 #define read_lock() (void)0
1549 #define read_unlock() (void)0
1550 #define cache_lock() (void)0
1551 #define cache_unlock() (void)0
1552 #define progress_lock() (void)0
1553 #define progress_unlock() (void)0
1557 static int try_delta(struct unpacked *trg, struct unpacked *src,
1558 unsigned max_depth, unsigned long *mem_usage)
1560 struct object_entry *trg_entry = trg->entry;
1561 struct object_entry *src_entry = src->entry;
1562 unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1564 enum object_type type;
1567 /* Don't bother doing diffs between different types */
1568 if (trg_entry->type != src_entry->type)
1572 * We do not bother to try a delta that we discarded on an
1573 * earlier try, but only when reusing delta data. Note that
1574 * src_entry that is marked as the preferred_base should always
1575 * be considered, as even if we produce a suboptimal delta against
1576 * it, we will still save the transfer cost, as we already know
1577 * the other side has it and we won't send src_entry at all.
1579 if (reuse_delta && trg_entry->in_pack &&
1580 trg_entry->in_pack == src_entry->in_pack &&
1581 !src_entry->preferred_base &&
1582 trg_entry->in_pack_type != OBJ_REF_DELTA &&
1583 trg_entry->in_pack_type != OBJ_OFS_DELTA)
1586 /* Let's not bust the allowed depth. */
1587 if (src->depth >= max_depth)
1590 /* Now some size filtering heuristics. */
1591 trg_size = trg_entry->size;
1592 if (!trg_entry->delta) {
1593 max_size = trg_size/2 - 20;
1596 max_size = trg_entry->delta_size;
1597 ref_depth = trg->depth;
1599 max_size = (uint64_t)max_size * (max_depth - src->depth) /
1600 (max_depth - ref_depth + 1);
1603 src_size = src_entry->size;
1604 sizediff = src_size < trg_size ? trg_size - src_size : 0;
1605 if (sizediff >= max_size)
1607 if (trg_size < src_size / 32)
1610 /* Load data if not already done */
1613 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1616 die("object %s cannot be read",
1617 sha1_to_hex(trg_entry->idx.sha1));
1619 die("object %s inconsistent object length (%lu vs %lu)",
1620 sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1625 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1628 if (src_entry->preferred_base) {
1629 static int warned = 0;
1631 warning("object %s cannot be read",
1632 sha1_to_hex(src_entry->idx.sha1));
1634 * Those objects are not included in the
1635 * resulting pack. Be resilient and ignore
1636 * them if they can't be read, in case the
1637 * pack could be created nevertheless.
1641 die("object %s cannot be read",
1642 sha1_to_hex(src_entry->idx.sha1));
1645 die("object %s inconsistent object length (%lu vs %lu)",
1646 sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1650 src->index = create_delta_index(src->data, src_size);
1652 static int warned = 0;
1654 warning("suboptimal pack - out of memory");
1657 *mem_usage += sizeof_delta_index(src->index);
1660 delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1664 if (trg_entry->delta) {
1665 /* Prefer only shallower same-sized deltas. */
1666 if (delta_size == trg_entry->delta_size &&
1667 src->depth + 1 >= trg->depth) {
1674 * Handle memory allocation outside of the cache
1675 * accounting lock. Compiler will optimize the strangeness
1676 * away when NO_PTHREADS is defined.
1678 free(trg_entry->delta_data);
1680 if (trg_entry->delta_data) {
1681 delta_cache_size -= trg_entry->delta_size;
1682 trg_entry->delta_data = NULL;
1684 if (delta_cacheable(src_size, trg_size, delta_size)) {
1685 delta_cache_size += delta_size;
1687 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1693 trg_entry->delta = src_entry;
1694 trg_entry->delta_size = delta_size;
1695 trg->depth = src->depth + 1;
1700 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1702 struct object_entry *child = me->delta_child;
1705 unsigned int c = check_delta_limit(child, n + 1);
1708 child = child->delta_sibling;
1713 static unsigned long free_unpacked(struct unpacked *n)
1715 unsigned long freed_mem = sizeof_delta_index(n->index);
1716 free_delta_index(n->index);
1719 freed_mem += n->entry->size;
1728 static void find_deltas(struct object_entry **list, unsigned *list_size,
1729 int window, int depth, unsigned *processed)
1731 uint32_t i, idx = 0, count = 0;
1732 struct unpacked *array;
1733 unsigned long mem_usage = 0;
1735 array = xcalloc(window, sizeof(struct unpacked));
1738 struct object_entry *entry;
1739 struct unpacked *n = array + idx;
1740 int j, max_depth, best_base = -1;
1749 if (!entry->preferred_base) {
1751 display_progress(progress_state, *processed);
1755 mem_usage -= free_unpacked(n);
1758 while (window_memory_limit &&
1759 mem_usage > window_memory_limit &&
1761 uint32_t tail = (idx + window - count) % window;
1762 mem_usage -= free_unpacked(array + tail);
1766 /* We do not compute delta to *create* objects we are not
1769 if (entry->preferred_base)
1773 * If the current object is at pack edge, take the depth the
1774 * objects that depend on the current object into account
1775 * otherwise they would become too deep.
1778 if (entry->delta_child) {
1779 max_depth -= check_delta_limit(entry, 0);
1787 uint32_t other_idx = idx + j;
1789 if (other_idx >= window)
1790 other_idx -= window;
1791 m = array + other_idx;
1794 ret = try_delta(n, m, max_depth, &mem_usage);
1798 best_base = other_idx;
1802 * If we decided to cache the delta data, then it is best
1803 * to compress it right away. First because we have to do
1804 * it anyway, and doing it here while we're threaded will
1805 * save a lot of time in the non threaded write phase,
1806 * as well as allow for caching more deltas within
1807 * the same cache size limit.
1809 * But only if not writing to stdout, since in that case
1810 * the network is most likely throttling writes anyway,
1811 * and therefore it is best to go to the write phase ASAP
1812 * instead, as we can afford spending more time compressing
1813 * between writes at that moment.
1815 if (entry->delta_data && !pack_to_stdout) {
1816 entry->z_delta_size = do_compress(&entry->delta_data,
1819 delta_cache_size -= entry->delta_size;
1820 delta_cache_size += entry->z_delta_size;
1824 /* if we made n a delta, and if n is already at max
1825 * depth, leaving it in the window is pointless. we
1826 * should evict it first.
1828 if (entry->delta && max_depth <= n->depth)
1832 * Move the best delta base up in the window, after the
1833 * currently deltified object, to keep it longer. It will
1834 * be the first base object to be attempted next.
1837 struct unpacked swap = array[best_base];
1838 int dist = (window + idx - best_base) % window;
1839 int dst = best_base;
1841 int src = (dst + 1) % window;
1842 array[dst] = array[src];
1850 if (count + 1 < window)
1856 for (i = 0; i < window; ++i) {
1857 free_delta_index(array[i].index);
1858 free(array[i].data);
1865 static void try_to_free_from_threads(size_t size)
1868 release_pack_memory(size);
1872 static try_to_free_t old_try_to_free_routine;
1875 * The main thread waits on the condition that (at least) one of the workers
1876 * has stopped working (which is indicated in the .working member of
1877 * struct thread_params).
1878 * When a work thread has completed its work, it sets .working to 0 and
1879 * signals the main thread and waits on the condition that .data_ready
1883 struct thread_params {
1885 struct object_entry **list;
1892 pthread_mutex_t mutex;
1893 pthread_cond_t cond;
1894 unsigned *processed;
1897 static pthread_cond_t progress_cond;
1900 * Mutex and conditional variable can't be statically-initialized on Windows.
1902 static void init_threaded_search(void)
1904 init_recursive_mutex(&read_mutex);
1905 pthread_mutex_init(&cache_mutex, NULL);
1906 pthread_mutex_init(&progress_mutex, NULL);
1907 pthread_cond_init(&progress_cond, NULL);
1908 old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1911 static void cleanup_threaded_search(void)
1913 set_try_to_free_routine(old_try_to_free_routine);
1914 pthread_cond_destroy(&progress_cond);
1915 pthread_mutex_destroy(&read_mutex);
1916 pthread_mutex_destroy(&cache_mutex);
1917 pthread_mutex_destroy(&progress_mutex);
1920 static void *threaded_find_deltas(void *arg)
1922 struct thread_params *me = arg;
1924 while (me->remaining) {
1925 find_deltas(me->list, &me->remaining,
1926 me->window, me->depth, me->processed);
1930 pthread_cond_signal(&progress_cond);
1934 * We must not set ->data_ready before we wait on the
1935 * condition because the main thread may have set it to 1
1936 * before we get here. In order to be sure that new
1937 * work is available if we see 1 in ->data_ready, it
1938 * was initialized to 0 before this thread was spawned
1939 * and we reset it to 0 right away.
1941 pthread_mutex_lock(&me->mutex);
1942 while (!me->data_ready)
1943 pthread_cond_wait(&me->cond, &me->mutex);
1945 pthread_mutex_unlock(&me->mutex);
1947 /* leave ->working 1 so that this doesn't get more work assigned */
1951 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1952 int window, int depth, unsigned *processed)
1954 struct thread_params *p;
1955 int i, ret, active_threads = 0;
1957 init_threaded_search();
1959 if (!delta_search_threads) /* --threads=0 means autodetect */
1960 delta_search_threads = online_cpus();
1961 if (delta_search_threads <= 1) {
1962 find_deltas(list, &list_size, window, depth, processed);
1963 cleanup_threaded_search();
1966 if (progress > pack_to_stdout)
1967 fprintf(stderr, "Delta compression using up to %d threads.\n",
1968 delta_search_threads);
1969 p = xcalloc(delta_search_threads, sizeof(*p));
1971 /* Partition the work amongst work threads. */
1972 for (i = 0; i < delta_search_threads; i++) {
1973 unsigned sub_size = list_size / (delta_search_threads - i);
1975 /* don't use too small segments or no deltas will be found */
1976 if (sub_size < 2*window && i+1 < delta_search_threads)
1979 p[i].window = window;
1981 p[i].processed = processed;
1983 p[i].data_ready = 0;
1985 /* try to split chunks on "path" boundaries */
1986 while (sub_size && sub_size < list_size &&
1987 list[sub_size]->hash &&
1988 list[sub_size]->hash == list[sub_size-1]->hash)
1992 p[i].list_size = sub_size;
1993 p[i].remaining = sub_size;
1996 list_size -= sub_size;
1999 /* Start work threads. */
2000 for (i = 0; i < delta_search_threads; i++) {
2001 if (!p[i].list_size)
2003 pthread_mutex_init(&p[i].mutex, NULL);
2004 pthread_cond_init(&p[i].cond, NULL);
2005 ret = pthread_create(&p[i].thread, NULL,
2006 threaded_find_deltas, &p[i]);
2008 die("unable to create thread: %s", strerror(ret));
2013 * Now let's wait for work completion. Each time a thread is done
2014 * with its work, we steal half of the remaining work from the
2015 * thread with the largest number of unprocessed objects and give
2016 * it to that newly idle thread. This ensure good load balancing
2017 * until the remaining object list segments are simply too short
2018 * to be worth splitting anymore.
2020 while (active_threads) {
2021 struct thread_params *target = NULL;
2022 struct thread_params *victim = NULL;
2023 unsigned sub_size = 0;
2027 for (i = 0; !target && i < delta_search_threads; i++)
2032 pthread_cond_wait(&progress_cond, &progress_mutex);
2035 for (i = 0; i < delta_search_threads; i++)
2036 if (p[i].remaining > 2*window &&
2037 (!victim || victim->remaining < p[i].remaining))
2040 sub_size = victim->remaining / 2;
2041 list = victim->list + victim->list_size - sub_size;
2042 while (sub_size && list[0]->hash &&
2043 list[0]->hash == list[-1]->hash) {
2049 * It is possible for some "paths" to have
2050 * so many objects that no hash boundary
2051 * might be found. Let's just steal the
2052 * exact half in that case.
2054 sub_size = victim->remaining / 2;
2057 target->list = list;
2058 victim->list_size -= sub_size;
2059 victim->remaining -= sub_size;
2061 target->list_size = sub_size;
2062 target->remaining = sub_size;
2063 target->working = 1;
2066 pthread_mutex_lock(&target->mutex);
2067 target->data_ready = 1;
2068 pthread_cond_signal(&target->cond);
2069 pthread_mutex_unlock(&target->mutex);
2072 pthread_join(target->thread, NULL);
2073 pthread_cond_destroy(&target->cond);
2074 pthread_mutex_destroy(&target->mutex);
2078 cleanup_threaded_search();
2083 #define ll_find_deltas(l, s, w, d, p) find_deltas(l, &s, w, d, p)
2086 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2088 unsigned char peeled[20];
2090 if (starts_with(path, "refs/tags/") && /* is a tag? */
2091 !peel_ref(path, peeled) && /* peelable? */
2092 packlist_find(&to_pack, peeled, NULL)) /* object packed? */
2093 add_object_entry(sha1, OBJ_TAG, NULL, 0);
2097 static void prepare_pack(int window, int depth)
2099 struct object_entry **delta_list;
2100 uint32_t i, nr_deltas;
2103 get_object_details();
2106 * If we're locally repacking then we need to be doubly careful
2107 * from now on in order to make sure no stealth corruption gets
2108 * propagated to the new pack. Clients receiving streamed packs
2109 * should validate everything they get anyway so no need to incur
2110 * the additional cost here in that case.
2112 if (!pack_to_stdout)
2113 do_check_packed_object_crc = 1;
2115 if (!to_pack.nr_objects || !window || !depth)
2118 delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2121 for (i = 0; i < to_pack.nr_objects; i++) {
2122 struct object_entry *entry = to_pack.objects + i;
2125 /* This happens if we decided to reuse existing
2126 * delta from a pack. "reuse_delta &&" is implied.
2130 if (entry->size < 50)
2133 if (entry->no_try_delta)
2136 if (!entry->preferred_base) {
2138 if (entry->type < 0)
2139 die("unable to get type of object %s",
2140 sha1_to_hex(entry->idx.sha1));
2142 if (entry->type < 0) {
2144 * This object is not found, but we
2145 * don't have to include it anyway.
2151 delta_list[n++] = entry;
2154 if (nr_deltas && n > 1) {
2155 unsigned nr_done = 0;
2157 progress_state = start_progress("Compressing objects",
2159 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2160 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2161 stop_progress(&progress_state);
2162 if (nr_done != nr_deltas)
2163 die("inconsistency with delta count");
2168 static int git_pack_config(const char *k, const char *v, void *cb)
2170 if (!strcmp(k, "pack.window")) {
2171 window = git_config_int(k, v);
2174 if (!strcmp(k, "pack.windowmemory")) {
2175 window_memory_limit = git_config_ulong(k, v);
2178 if (!strcmp(k, "pack.depth")) {
2179 depth = git_config_int(k, v);
2182 if (!strcmp(k, "pack.compression")) {
2183 int level = git_config_int(k, v);
2185 level = Z_DEFAULT_COMPRESSION;
2186 else if (level < 0 || level > Z_BEST_COMPRESSION)
2187 die("bad pack compression level %d", level);
2188 pack_compression_level = level;
2189 pack_compression_seen = 1;
2192 if (!strcmp(k, "pack.deltacachesize")) {
2193 max_delta_cache_size = git_config_int(k, v);
2196 if (!strcmp(k, "pack.deltacachelimit")) {
2197 cache_max_small_delta_size = git_config_int(k, v);
2200 if (!strcmp(k, "pack.writebitmaps")) {
2201 write_bitmap_index = git_config_bool(k, v);
2204 if (!strcmp(k, "pack.writebitmaphashcache")) {
2205 if (git_config_bool(k, v))
2206 write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2208 write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2210 if (!strcmp(k, "pack.usebitmaps")) {
2211 use_bitmap_index = git_config_bool(k, v);
2214 if (!strcmp(k, "pack.threads")) {
2215 delta_search_threads = git_config_int(k, v);
2216 if (delta_search_threads < 0)
2217 die("invalid number of threads specified (%d)",
2218 delta_search_threads);
2220 if (delta_search_threads != 1)
2221 warning("no threads support, ignoring %s", k);
2225 if (!strcmp(k, "pack.indexversion")) {
2226 pack_idx_opts.version = git_config_int(k, v);
2227 if (pack_idx_opts.version > 2)
2228 die("bad pack.indexversion=%"PRIu32,
2229 pack_idx_opts.version);
2232 return git_default_config(k, v, cb);
2235 static void read_object_list_from_stdin(void)
2237 char line[40 + 1 + PATH_MAX + 2];
2238 unsigned char sha1[20];
2241 if (!fgets(line, sizeof(line), stdin)) {
2245 die("fgets returned NULL, not EOF, not error!");
2251 if (line[0] == '-') {
2252 if (get_sha1_hex(line+1, sha1))
2253 die("expected edge sha1, got garbage:\n %s",
2255 add_preferred_base(sha1);
2258 if (get_sha1_hex(line, sha1))
2259 die("expected sha1, got garbage:\n %s", line);
2261 add_preferred_base_object(line+41);
2262 add_object_entry(sha1, 0, line+41, 0);
2266 #define OBJECT_ADDED (1u<<20)
2268 static void show_commit(struct commit *commit, void *data)
2270 add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2271 commit->object.flags |= OBJECT_ADDED;
2273 if (write_bitmap_index)
2274 index_commit_for_bitmap(commit);
2277 static void show_object(struct object *obj,
2278 const struct name_path *path, const char *last,
2281 char *name = path_name(path, last);
2283 add_preferred_base_object(name);
2284 add_object_entry(obj->sha1, obj->type, name, 0);
2285 obj->flags |= OBJECT_ADDED;
2288 * We will have generated the hash from the name,
2289 * but not saved a pointer to it - we can free it
2294 static void show_edge(struct commit *commit)
2296 add_preferred_base(commit->object.sha1);
2299 struct in_pack_object {
2301 struct object *object;
2307 struct in_pack_object *array;
2310 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2312 in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2313 in_pack->array[in_pack->nr].object = object;
2318 * Compare the objects in the offset order, in order to emulate the
2319 * "git rev-list --objects" output that produced the pack originally.
2321 static int ofscmp(const void *a_, const void *b_)
2323 struct in_pack_object *a = (struct in_pack_object *)a_;
2324 struct in_pack_object *b = (struct in_pack_object *)b_;
2326 if (a->offset < b->offset)
2328 else if (a->offset > b->offset)
2331 return hashcmp(a->object->sha1, b->object->sha1);
2334 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2336 struct packed_git *p;
2337 struct in_pack in_pack;
2340 memset(&in_pack, 0, sizeof(in_pack));
2342 for (p = packed_git; p; p = p->next) {
2343 const unsigned char *sha1;
2346 if (!p->pack_local || p->pack_keep)
2348 if (open_pack_index(p))
2349 die("cannot open pack index");
2351 ALLOC_GROW(in_pack.array,
2352 in_pack.nr + p->num_objects,
2355 for (i = 0; i < p->num_objects; i++) {
2356 sha1 = nth_packed_object_sha1(p, i);
2357 o = lookup_unknown_object(sha1);
2358 if (!(o->flags & OBJECT_ADDED))
2359 mark_in_pack_object(o, p, &in_pack);
2360 o->flags |= OBJECT_ADDED;
2365 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2367 for (i = 0; i < in_pack.nr; i++) {
2368 struct object *o = in_pack.array[i].object;
2369 add_object_entry(o->sha1, o->type, "", 0);
2372 free(in_pack.array);
2375 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2377 static struct packed_git *last_found = (void *)1;
2378 struct packed_git *p;
2380 p = (last_found != (void *)1) ? last_found : packed_git;
2383 if ((!p->pack_local || p->pack_keep) &&
2384 find_pack_entry_one(sha1, p)) {
2388 if (p == last_found)
2392 if (p == last_found)
2398 static void loosen_unused_packed_objects(struct rev_info *revs)
2400 struct packed_git *p;
2402 const unsigned char *sha1;
2404 for (p = packed_git; p; p = p->next) {
2405 if (!p->pack_local || p->pack_keep)
2408 if (unpack_unreachable_expiration &&
2409 p->mtime < unpack_unreachable_expiration)
2412 if (open_pack_index(p))
2413 die("cannot open pack index");
2415 for (i = 0; i < p->num_objects; i++) {
2416 sha1 = nth_packed_object_sha1(p, i);
2417 if (!packlist_find(&to_pack, sha1, NULL) &&
2418 !has_sha1_pack_kept_or_nonlocal(sha1))
2419 if (force_object_loose(sha1, p->mtime))
2420 die("unable to force loose object");
2425 static int get_object_list_from_bitmap(struct rev_info *revs)
2427 if (prepare_bitmap_walk(revs) < 0)
2430 if (!reuse_partial_packfile_from_bitmap(
2432 &reuse_packfile_objects,
2433 &reuse_packfile_offset)) {
2434 assert(reuse_packfile_objects);
2435 nr_result += reuse_packfile_objects;
2438 fprintf(stderr, "Reusing existing pack: %d, done.\n",
2439 reuse_packfile_objects);
2444 traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2448 static void get_object_list(int ac, const char **av)
2450 struct rev_info revs;
2454 init_revisions(&revs, NULL);
2455 save_commit_buffer = 0;
2456 setup_revisions(ac, av, &revs, NULL);
2458 while (fgets(line, sizeof(line), stdin) != NULL) {
2459 int len = strlen(line);
2460 if (len && line[len - 1] == '\n')
2465 if (!strcmp(line, "--not")) {
2466 flags ^= UNINTERESTING;
2467 write_bitmap_index = 0;
2470 die("not a rev '%s'", line);
2472 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2473 die("bad revision '%s'", line);
2476 if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2479 if (prepare_revision_walk(&revs))
2480 die("revision walk setup failed");
2481 mark_edges_uninteresting(&revs, show_edge);
2482 traverse_commit_list(&revs, show_commit, show_object, NULL);
2484 if (keep_unreachable)
2485 add_objects_in_unpacked_packs(&revs);
2486 if (unpack_unreachable)
2487 loosen_unused_packed_objects(&revs);
2490 static int option_parse_index_version(const struct option *opt,
2491 const char *arg, int unset)
2494 const char *val = arg;
2495 pack_idx_opts.version = strtoul(val, &c, 10);
2496 if (pack_idx_opts.version > 2)
2497 die(_("unsupported index version %s"), val);
2498 if (*c == ',' && c[1])
2499 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2500 if (*c || pack_idx_opts.off32_limit & 0x80000000)
2501 die(_("bad index version '%s'"), val);
2505 static int option_parse_unpack_unreachable(const struct option *opt,
2506 const char *arg, int unset)
2509 unpack_unreachable = 0;
2510 unpack_unreachable_expiration = 0;
2513 unpack_unreachable = 1;
2515 unpack_unreachable_expiration = approxidate(arg);
2520 static int option_parse_ulong(const struct option *opt,
2521 const char *arg, int unset)
2524 die(_("option %s does not accept negative form"),
2527 if (!git_parse_ulong(arg, opt->value))
2528 die(_("unable to parse value '%s' for option %s"),
2529 arg, opt->long_name);
2533 #define OPT_ULONG(s, l, v, h) \
2534 { OPTION_CALLBACK, (s), (l), (v), "n", (h), \
2535 PARSE_OPT_NONEG, option_parse_ulong }
2537 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2539 int use_internal_rev_list = 0;
2541 int all_progress_implied = 0;
2542 const char *rp_av[6];
2544 int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2545 struct option pack_objects_options[] = {
2546 OPT_SET_INT('q', "quiet", &progress,
2547 N_("do not show progress meter"), 0),
2548 OPT_SET_INT(0, "progress", &progress,
2549 N_("show progress meter"), 1),
2550 OPT_SET_INT(0, "all-progress", &progress,
2551 N_("show progress meter during object writing phase"), 2),
2552 OPT_BOOL(0, "all-progress-implied",
2553 &all_progress_implied,
2554 N_("similar to --all-progress when progress meter is shown")),
2555 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2556 N_("write the pack index file in the specified idx format version"),
2557 0, option_parse_index_version },
2558 OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2559 N_("maximum size of each output pack file")),
2560 OPT_BOOL(0, "local", &local,
2561 N_("ignore borrowed objects from alternate object store")),
2562 OPT_BOOL(0, "incremental", &incremental,
2563 N_("ignore packed objects")),
2564 OPT_INTEGER(0, "window", &window,
2565 N_("limit pack window by objects")),
2566 OPT_ULONG(0, "window-memory", &window_memory_limit,
2567 N_("limit pack window by memory in addition to object limit")),
2568 OPT_INTEGER(0, "depth", &depth,
2569 N_("maximum length of delta chain allowed in the resulting pack")),
2570 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2571 N_("reuse existing deltas")),
2572 OPT_BOOL(0, "reuse-object", &reuse_object,
2573 N_("reuse existing objects")),
2574 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2575 N_("use OFS_DELTA objects")),
2576 OPT_INTEGER(0, "threads", &delta_search_threads,
2577 N_("use threads when searching for best delta matches")),
2578 OPT_BOOL(0, "non-empty", &non_empty,
2579 N_("do not create an empty pack output")),
2580 OPT_BOOL(0, "revs", &use_internal_rev_list,
2581 N_("read revision arguments from standard input")),
2582 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2583 N_("limit the objects to those that are not yet packed"),
2584 PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2585 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2586 N_("include objects reachable from any reference"),
2587 PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2588 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2589 N_("include objects referred by reflog entries"),
2590 PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2591 OPT_BOOL(0, "stdout", &pack_to_stdout,
2592 N_("output pack to stdout")),
2593 OPT_BOOL(0, "include-tag", &include_tag,
2594 N_("include tag objects that refer to objects to be packed")),
2595 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2596 N_("keep unreachable objects")),
2597 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2598 N_("unpack unreachable objects newer than <time>"),
2599 PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2600 OPT_BOOL(0, "thin", &thin,
2601 N_("create thin packs")),
2602 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2603 N_("ignore packs that have companion .keep file")),
2604 OPT_INTEGER(0, "compression", &pack_compression_level,
2605 N_("pack compression level")),
2606 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2607 N_("do not hide commits by grafts"), 0),
2608 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2609 N_("use a bitmap index if available to speed up counting objects")),
2610 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2611 N_("write a bitmap index together with the pack index")),
2615 read_replace_refs = 0;
2617 reset_pack_idx_option(&pack_idx_opts);
2618 git_config(git_pack_config, NULL);
2619 if (!pack_compression_seen && core_compression_seen)
2620 pack_compression_level = core_compression_level;
2622 progress = isatty(2);
2623 argc = parse_options(argc, argv, prefix, pack_objects_options,
2627 base_name = argv[0];
2630 if (pack_to_stdout != !base_name || argc)
2631 usage_with_options(pack_usage, pack_objects_options);
2633 rp_av[rp_ac++] = "pack-objects";
2635 use_internal_rev_list = 1;
2636 rp_av[rp_ac++] = "--objects-edge";
2638 rp_av[rp_ac++] = "--objects";
2641 use_internal_rev_list = 1;
2642 rp_av[rp_ac++] = "--all";
2644 if (rev_list_reflog) {
2645 use_internal_rev_list = 1;
2646 rp_av[rp_ac++] = "--reflog";
2648 if (rev_list_unpacked) {
2649 use_internal_rev_list = 1;
2650 rp_av[rp_ac++] = "--unpacked";
2655 if (pack_compression_level == -1)
2656 pack_compression_level = Z_DEFAULT_COMPRESSION;
2657 else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2658 die("bad pack compression level %d", pack_compression_level);
2660 if (delta_search_threads != 1)
2661 warning("no threads support, ignoring --threads");
2663 if (!pack_to_stdout && !pack_size_limit)
2664 pack_size_limit = pack_size_limit_cfg;
2665 if (pack_to_stdout && pack_size_limit)
2666 die("--max-pack-size cannot be used to build a pack for transfer.");
2667 if (pack_size_limit && pack_size_limit < 1024*1024) {
2668 warning("minimum pack size limit is 1 MiB");
2669 pack_size_limit = 1024*1024;
2672 if (!pack_to_stdout && thin)
2673 die("--thin cannot be used to build an indexable pack.");
2675 if (keep_unreachable && unpack_unreachable)
2676 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2678 if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2679 use_bitmap_index = 0;
2681 if (pack_to_stdout || !rev_list_all)
2682 write_bitmap_index = 0;
2684 if (progress && all_progress_implied)
2687 prepare_packed_git();
2690 progress_state = start_progress("Counting objects", 0);
2691 if (!use_internal_rev_list)
2692 read_object_list_from_stdin();
2694 rp_av[rp_ac] = NULL;
2695 get_object_list(rp_ac, rp_av);
2697 cleanup_preferred_base();
2698 if (include_tag && nr_result)
2699 for_each_ref(add_ref_tag, NULL);
2700 stop_progress(&progress_state);
2702 if (non_empty && !nr_result)
2705 prepare_pack(window, depth);
2708 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2709 " reused %"PRIu32" (delta %"PRIu32")\n",
2710 written, written_delta, reused, reused_delta);