6 #include "list-objects.h"
8 #include "sha1-lookup.h"
9 #include "run-command.h"
12 #include "sha1-array.h"
13 #include "argv-array.h"
15 static struct sha1_array good_revs;
16 static struct sha1_array skipped_revs;
18 static unsigned char *current_bad_sha1;
20 static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
21 static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
22 static const char *argv_update_ref[] = {"update-ref", "--no-deref", "BISECT_HEAD", NULL, NULL};
24 /* bits #0-15 in revision.h */
26 #define COUNTED (1u<<16)
29 * This is a truly stupid algorithm, but it's only
30 * used for bisection, and we just don't care enough.
32 * We care just barely enough to avoid recursing for
35 static int count_distance(struct commit_list *entry)
40 struct commit *commit = entry->item;
41 struct commit_list *p;
43 if (commit->object.flags & (UNINTERESTING | COUNTED))
45 if (!(commit->object.flags & TREESAME))
47 commit->object.flags |= COUNTED;
53 nr += count_distance(p);
62 static void clear_distance(struct commit_list *list)
65 struct commit *commit = list->item;
66 commit->object.flags &= ~COUNTED;
71 #define DEBUG_BISECT 0
73 static inline int weight(struct commit_list *elem)
75 return *((int*)(elem->item->util));
78 static inline void weight_set(struct commit_list *elem, int weight)
80 *((int*)(elem->item->util)) = weight;
83 static int count_interesting_parents(struct commit *commit)
85 struct commit_list *p;
88 for (count = 0, p = commit->parents; p; p = p->next) {
89 if (p->item->object.flags & UNINTERESTING)
96 static inline int halfway(struct commit_list *p, int nr)
99 * Don't short-cut something we are not going to return!
101 if (p->item->object.flags & TREESAME)
106 * 2 and 3 are halfway of 5.
107 * 3 is halfway of 6 but 2 and 4 are not.
109 switch (2 * weight(p) - nr) {
110 case -1: case 0: case 1:
118 #define show_list(a,b,c,d) do { ; } while (0)
120 static void show_list(const char *debug, int counted, int nr,
121 struct commit_list *list)
123 struct commit_list *p;
125 fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
127 for (p = list; p; p = p->next) {
128 struct commit_list *pp;
129 struct commit *commit = p->item;
130 unsigned flags = commit->object.flags;
131 enum object_type type;
133 char *buf = read_sha1_file(commit->object.sha1, &type, &size);
134 const char *subject_start;
137 fprintf(stderr, "%c%c%c ",
138 (flags & TREESAME) ? ' ' : 'T',
139 (flags & UNINTERESTING) ? 'U' : ' ',
140 (flags & COUNTED) ? 'C' : ' ');
142 fprintf(stderr, "%3d", weight(p));
144 fprintf(stderr, "---");
145 fprintf(stderr, " %.*s", 8, sha1_to_hex(commit->object.sha1));
146 for (pp = commit->parents; pp; pp = pp->next)
147 fprintf(stderr, " %.*s", 8,
148 sha1_to_hex(pp->item->object.sha1));
150 subject_len = find_commit_subject(buf, &subject_start);
152 fprintf(stderr, " %.*s", subject_len, subject_start);
153 fprintf(stderr, "\n");
156 #endif /* DEBUG_BISECT */
158 static struct commit_list *best_bisection(struct commit_list *list, int nr)
160 struct commit_list *p, *best;
161 int best_distance = -1;
164 for (p = list; p; p = p->next) {
166 unsigned flags = p->item->object.flags;
168 if (flags & TREESAME)
170 distance = weight(p);
171 if (nr - distance < distance)
172 distance = nr - distance;
173 if (distance > best_distance) {
175 best_distance = distance;
183 struct commit *commit;
187 static int compare_commit_dist(const void *a_, const void *b_)
189 struct commit_dist *a, *b;
191 a = (struct commit_dist *)a_;
192 b = (struct commit_dist *)b_;
193 if (a->distance != b->distance)
194 return b->distance - a->distance; /* desc sort */
195 return hashcmp(a->commit->object.sha1, b->commit->object.sha1);
198 static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
200 struct commit_list *p;
201 struct commit_dist *array = xcalloc(nr, sizeof(*array));
204 for (p = list, cnt = 0; p; p = p->next) {
206 unsigned flags = p->item->object.flags;
208 if (flags & TREESAME)
210 distance = weight(p);
211 if (nr - distance < distance)
212 distance = nr - distance;
213 array[cnt].commit = p->item;
214 array[cnt].distance = distance;
217 qsort(array, cnt, sizeof(*array), compare_commit_dist);
218 for (p = list, i = 0; i < cnt; i++) {
219 char buf[100]; /* enough for dist=%d */
220 struct object *obj = &(array[i].commit->object);
222 snprintf(buf, sizeof(buf), "dist=%d", array[i].distance);
223 add_name_decoration(DECORATION_NONE, buf, obj);
225 p->item = array[i].commit;
235 * zero or positive weight is the number of interesting commits it can
236 * reach, including itself. Especially, weight = 0 means it does not
237 * reach any tree-changing commits (e.g. just above uninteresting one
238 * but traversal is with pathspec).
240 * weight = -1 means it has one parent and its distance is yet to
243 * weight = -2 means it has more than one parent and its distance is
244 * unknown. After running count_distance() first, they will get zero
245 * or positive distance.
247 static struct commit_list *do_find_bisection(struct commit_list *list,
248 int nr, int *weights,
252 struct commit_list *p;
256 for (n = 0, p = list; p; p = p->next) {
257 struct commit *commit = p->item;
258 unsigned flags = commit->object.flags;
260 p->item->util = &weights[n++];
261 switch (count_interesting_parents(commit)) {
263 if (!(flags & TREESAME)) {
266 show_list("bisection 2 count one",
270 * otherwise, it is known not to reach any
271 * tree-changing commit and gets weight 0.
283 show_list("bisection 2 initialize", counted, nr, list);
286 * If you have only one parent in the resulting set
287 * then you can reach one commit more than that parent
288 * can reach. So we do not have to run the expensive
289 * count_distance() for single strand of pearls.
291 * However, if you have more than one parents, you cannot
292 * just add their distance and one for yourself, since
293 * they usually reach the same ancestor and you would
294 * end up counting them twice that way.
296 * So we will first count distance of merges the usual
297 * way, and then fill the blanks using cheaper algorithm.
299 for (p = list; p; p = p->next) {
300 if (p->item->object.flags & UNINTERESTING)
304 weight_set(p, count_distance(p));
305 clear_distance(list);
307 /* Does it happen to be at exactly half-way? */
308 if (!find_all && halfway(p, nr))
313 show_list("bisection 2 count_distance", counted, nr, list);
315 while (counted < nr) {
316 for (p = list; p; p = p->next) {
317 struct commit_list *q;
318 unsigned flags = p->item->object.flags;
322 for (q = p->item->parents; q; q = q->next) {
323 if (q->item->object.flags & UNINTERESTING)
332 * weight for p is unknown but q is known.
333 * add one for p itself if p is to be counted,
334 * otherwise inherit it from q directly.
336 if (!(flags & TREESAME)) {
337 weight_set(p, weight(q)+1);
339 show_list("bisection 2 count one",
343 weight_set(p, weight(q));
345 /* Does it happen to be at exactly half-way? */
346 if (!find_all && halfway(p, nr))
351 show_list("bisection 2 counted all", counted, nr, list);
354 return best_bisection(list, nr);
356 return best_bisection_sorted(list, nr);
359 struct commit_list *find_bisection(struct commit_list *list,
360 int *reaches, int *all,
364 struct commit_list *p, *best, *next, *last;
367 show_list("bisection 2 entry", 0, 0, list);
370 * Count the number of total and tree-changing items on the
371 * list, while reversing the list.
373 for (nr = on_list = 0, last = NULL, p = list;
376 unsigned flags = p->item->object.flags;
379 if (flags & UNINTERESTING)
383 if (!(flags & TREESAME))
388 show_list("bisection 2 sorted", 0, nr, list);
391 weights = xcalloc(on_list, sizeof(*weights));
393 /* Do the real work of finding bisection commit. */
394 best = do_find_bisection(list, nr, weights, find_all);
398 *reaches = weight(best);
404 static int register_ref(const char *refname, const unsigned char *sha1,
405 int flags, void *cb_data)
407 if (!strcmp(refname, "bad")) {
408 current_bad_sha1 = xmalloc(20);
409 hashcpy(current_bad_sha1, sha1);
410 } else if (starts_with(refname, "good-")) {
411 sha1_array_append(&good_revs, sha1);
412 } else if (starts_with(refname, "skip-")) {
413 sha1_array_append(&skipped_revs, sha1);
419 static int read_bisect_refs(void)
421 return for_each_ref_in("refs/bisect/", register_ref, NULL);
424 static void read_bisect_paths(struct argv_array *array)
426 struct strbuf str = STRBUF_INIT;
427 const char *filename = git_path("BISECT_NAMES");
428 FILE *fp = fopen(filename, "r");
431 die_errno("Could not open file '%s'", filename);
433 while (strbuf_getline(&str, fp, '\n') != EOF) {
435 if (sq_dequote_to_argv_array(str.buf, array))
436 die("Badly quoted content in file '%s': %s",
440 strbuf_release(&str);
444 static char *join_sha1_array_hex(struct sha1_array *array, char delim)
446 struct strbuf joined_hexs = STRBUF_INIT;
449 for (i = 0; i < array->nr; i++) {
450 strbuf_addstr(&joined_hexs, sha1_to_hex(array->sha1[i]));
451 if (i + 1 < array->nr)
452 strbuf_addch(&joined_hexs, delim);
455 return strbuf_detach(&joined_hexs, NULL);
459 * In this function, passing a not NULL skipped_first is very special.
460 * It means that we want to know if the first commit in the list is
461 * skipped because we will want to test a commit away from it if it is
463 * So if the first commit is skipped, we cannot take the shortcut to
464 * just "return list" when we find the first non skipped commit, we
465 * have to return a fully filtered list.
467 * We use (*skipped_first == -1) to mean "it has been found that the
468 * first commit is not skipped". In this case *skipped_first is set back
469 * to 0 just before the function returns.
471 struct commit_list *filter_skipped(struct commit_list *list,
472 struct commit_list **tried,
477 struct commit_list *filtered = NULL, **f = &filtered;
486 if (!skipped_revs.nr)
490 struct commit_list *next = list->next;
492 if (0 <= sha1_array_lookup(&skipped_revs,
493 list->item->object.sha1)) {
494 if (skipped_first && !*skipped_first)
496 /* Move current to tried list */
501 if (!skipped_first || !*skipped_first)
503 } else if (skipped_first && !*skipped_first) {
504 /* This means we know it's not skipped */
507 /* Move current to filtered list */
516 if (skipped_first && *skipped_first == -1)
522 #define PRN_MODULO 32768
525 * This is a pseudo random number generator based on "man 3 rand".
526 * It is not used properly because the seed is the argument and it
527 * is increased by one between each call, but that should not matter
528 * for this application.
530 static unsigned get_prn(unsigned count) {
531 count = count * 1103515245 + 12345;
532 return (count/65536) % PRN_MODULO;
536 * Custom integer square root from
537 * http://en.wikipedia.org/wiki/Integer_square_root
539 static int sqrti(int val)
547 float y = (x + (float)val / x) / 2;
548 d = (y > x) ? y - x : x - y;
555 static struct commit_list *skip_away(struct commit_list *list, int count)
557 struct commit_list *cur, *previous;
560 prn = get_prn(count);
561 index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
566 for (i = 0; cur; cur = cur->next, i++) {
568 if (hashcmp(cur->item->object.sha1, current_bad_sha1))
580 static struct commit_list *managed_skipped(struct commit_list *list,
581 struct commit_list **tried)
583 int count, skipped_first;
587 if (!skipped_revs.nr)
590 list = filter_skipped(list, tried, 0, &count, &skipped_first);
595 return skip_away(list, count);
598 static void bisect_rev_setup(struct rev_info *revs, const char *prefix,
599 const char *bad_format, const char *good_format,
602 struct argv_array rev_argv = ARGV_ARRAY_INIT;
605 init_revisions(revs, prefix);
607 revs->commit_format = CMIT_FMT_UNSPECIFIED;
609 /* rev_argv.argv[0] will be ignored by setup_revisions */
610 argv_array_push(&rev_argv, "bisect_rev_setup");
611 argv_array_pushf(&rev_argv, bad_format, sha1_to_hex(current_bad_sha1));
612 for (i = 0; i < good_revs.nr; i++)
613 argv_array_pushf(&rev_argv, good_format,
614 sha1_to_hex(good_revs.sha1[i]));
615 argv_array_push(&rev_argv, "--");
617 read_bisect_paths(&rev_argv);
619 setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
620 /* XXX leak rev_argv, as "revs" may still be pointing to it */
623 static void bisect_common(struct rev_info *revs)
625 if (prepare_revision_walk(revs))
626 die("revision walk setup failed");
627 if (revs->tree_objects)
628 mark_edges_uninteresting(revs, NULL);
631 static void exit_if_skipped_commits(struct commit_list *tried,
632 const unsigned char *bad)
637 printf("There are only 'skip'ped commits left to test.\n"
638 "The first bad commit could be any of:\n");
639 print_commit_list(tried, "%s\n", "%s\n");
641 printf("%s\n", sha1_to_hex(bad));
642 printf("We cannot bisect more!\n");
646 static int is_expected_rev(const unsigned char *sha1)
648 const char *filename = git_path("BISECT_EXPECTED_REV");
650 struct strbuf str = STRBUF_INIT;
654 if (stat(filename, &st) || !S_ISREG(st.st_mode))
657 fp = fopen(filename, "r");
661 if (strbuf_getline(&str, fp, '\n') != EOF)
662 res = !strcmp(str.buf, sha1_to_hex(sha1));
664 strbuf_release(&str);
670 static void mark_expected_rev(char *bisect_rev_hex)
672 int len = strlen(bisect_rev_hex);
673 const char *filename = git_path("BISECT_EXPECTED_REV");
674 int fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
677 die_errno("could not create file '%s'", filename);
679 bisect_rev_hex[len] = '\n';
680 write_or_die(fd, bisect_rev_hex, len + 1);
681 bisect_rev_hex[len] = '\0';
684 die("closing file %s: %s", filename, strerror(errno));
687 static int bisect_checkout(char *bisect_rev_hex, int no_checkout)
691 mark_expected_rev(bisect_rev_hex);
693 argv_checkout[2] = bisect_rev_hex;
695 argv_update_ref[3] = bisect_rev_hex;
696 if (run_command_v_opt(argv_update_ref, RUN_GIT_CMD))
697 die("update-ref --no-deref HEAD failed on %s",
700 res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
705 argv_show_branch[1] = bisect_rev_hex;
706 return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
709 static struct commit *get_commit_reference(const unsigned char *sha1)
711 struct commit *r = lookup_commit_reference(sha1);
713 die("Not a valid commit name %s", sha1_to_hex(sha1));
717 static struct commit **get_bad_and_good_commits(int *rev_nr)
719 int len = 1 + good_revs.nr;
720 struct commit **rev = xmalloc(len * sizeof(*rev));
723 rev[n++] = get_commit_reference(current_bad_sha1);
724 for (i = 0; i < good_revs.nr; i++)
725 rev[n++] = get_commit_reference(good_revs.sha1[i]);
731 static void handle_bad_merge_base(void)
733 if (is_expected_rev(current_bad_sha1)) {
734 char *bad_hex = sha1_to_hex(current_bad_sha1);
735 char *good_hex = join_sha1_array_hex(&good_revs, ' ');
737 fprintf(stderr, "The merge base %s is bad.\n"
738 "This means the bug has been fixed "
739 "between %s and [%s].\n",
740 bad_hex, bad_hex, good_hex);
745 fprintf(stderr, "Some good revs are not ancestor of the bad rev.\n"
746 "git bisect cannot work properly in this case.\n"
747 "Maybe you mistake good and bad revs?\n");
751 static void handle_skipped_merge_base(const unsigned char *mb)
753 char *mb_hex = sha1_to_hex(mb);
754 char *bad_hex = sha1_to_hex(current_bad_sha1);
755 char *good_hex = join_sha1_array_hex(&good_revs, ' ');
757 warning("the merge base between %s and [%s] "
759 "So we cannot be sure the first bad commit is "
760 "between %s and %s.\n"
761 "We continue anyway.",
762 bad_hex, good_hex, mb_hex, bad_hex);
767 * "check_merge_bases" checks that merge bases are not "bad".
769 * - If one is "bad", it means the user assumed something wrong
770 * and we must exit with a non 0 error code.
771 * - If one is "good", that's good, we have nothing to do.
772 * - If one is "skipped", we can't know but we should warn.
773 * - If we don't know, we should check it out and ask the user to test.
775 static void check_merge_bases(int no_checkout)
777 struct commit_list *result;
779 struct commit **rev = get_bad_and_good_commits(&rev_nr);
781 result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1, 0);
783 for (; result; result = result->next) {
784 const unsigned char *mb = result->item->object.sha1;
785 if (!hashcmp(mb, current_bad_sha1)) {
786 handle_bad_merge_base();
787 } else if (0 <= sha1_array_lookup(&good_revs, mb)) {
789 } else if (0 <= sha1_array_lookup(&skipped_revs, mb)) {
790 handle_skipped_merge_base(mb);
792 printf("Bisecting: a merge base must be tested\n");
793 exit(bisect_checkout(sha1_to_hex(mb), no_checkout));
798 free_commit_list(result);
801 static int check_ancestors(const char *prefix)
803 struct rev_info revs;
804 struct object_array pending_copy;
807 bisect_rev_setup(&revs, prefix, "^%s", "%s", 0);
809 /* Save pending objects, so they can be cleaned up later. */
810 pending_copy = revs.pending;
811 revs.leak_pending = 1;
814 * bisect_common calls prepare_revision_walk right away, which
815 * (together with .leak_pending = 1) makes us the sole owner of
816 * the list of pending objects.
818 bisect_common(&revs);
819 res = (revs.commits != NULL);
821 /* Clean up objects used, as they will be reused. */
822 clear_commit_marks_for_object_array(&pending_copy, ALL_REV_FLAGS);
823 free(pending_copy.objects);
829 * "check_good_are_ancestors_of_bad" checks that all "good" revs are
830 * ancestor of the "bad" rev.
832 * If that's not the case, we need to check the merge bases.
833 * If a merge base must be tested by the user, its source code will be
834 * checked out to be tested by the user and we will exit.
836 static void check_good_are_ancestors_of_bad(const char *prefix, int no_checkout)
838 char *filename = git_pathdup("BISECT_ANCESTORS_OK");
842 if (!current_bad_sha1)
843 die("a bad revision is needed");
845 /* Check if file BISECT_ANCESTORS_OK exists. */
846 if (!stat(filename, &st) && S_ISREG(st.st_mode))
849 /* Bisecting with no good rev is ok. */
850 if (good_revs.nr == 0)
853 /* Check if all good revs are ancestor of the bad rev. */
854 if (check_ancestors(prefix))
855 check_merge_bases(no_checkout);
857 /* Create file BISECT_ANCESTORS_OK. */
858 fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
860 warning("could not create file '%s': %s",
861 filename, strerror(errno));
869 * This does "git diff-tree --pretty COMMIT" without one fork+exec.
871 static void show_diff_tree(const char *prefix, struct commit *commit)
876 init_revisions(&opt, prefix);
877 git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
881 /* This is what "--pretty" does */
882 opt.verbose_header = 1;
883 opt.use_terminator = 0;
884 opt.commit_format = CMIT_FMT_DEFAULT;
887 if (!opt.diffopt.output_format)
888 opt.diffopt.output_format = DIFF_FORMAT_RAW;
890 log_tree_commit(&opt, commit);
894 * We use the convention that exiting with an exit code 10 means that
895 * the bisection process finished successfully.
896 * In this case the calling shell script should exit 0.
898 * If no_checkout is non-zero, the bisection process does not
899 * checkout the trial commit but instead simply updates BISECT_HEAD.
901 int bisect_next_all(const char *prefix, int no_checkout)
903 struct rev_info revs;
904 struct commit_list *tried;
905 int reaches = 0, all = 0, nr, steps;
906 const unsigned char *bisect_rev;
907 char bisect_rev_hex[41];
909 if (read_bisect_refs())
910 die("reading bisect refs failed");
912 check_good_are_ancestors_of_bad(prefix, no_checkout);
914 bisect_rev_setup(&revs, prefix, "%s", "^%s", 1);
917 bisect_common(&revs);
919 revs.commits = find_bisection(revs.commits, &reaches, &all,
921 revs.commits = managed_skipped(revs.commits, &tried);
925 * We should exit here only if the "bad"
926 * commit is also a "skip" commit.
928 exit_if_skipped_commits(tried, NULL);
930 printf("%s was both good and bad\n",
931 sha1_to_hex(current_bad_sha1));
936 fprintf(stderr, "No testable commit found.\n"
937 "Maybe you started with bad path parameters?\n");
941 bisect_rev = revs.commits->item->object.sha1;
942 memcpy(bisect_rev_hex, sha1_to_hex(bisect_rev), 41);
944 if (!hashcmp(bisect_rev, current_bad_sha1)) {
945 exit_if_skipped_commits(tried, current_bad_sha1);
946 printf("%s is the first bad commit\n", bisect_rev_hex);
947 show_diff_tree(prefix, revs.commits->item);
948 /* This means the bisection process succeeded. */
952 nr = all - reaches - 1;
953 steps = estimate_bisect_steps(all);
954 printf("Bisecting: %d revision%s left to test after this "
955 "(roughly %d step%s)\n", nr, (nr == 1 ? "" : "s"),
956 steps, (steps == 1 ? "" : "s"));
958 return bisect_checkout(bisect_rev_hex, no_checkout);
961 static inline int log2i(int n)
965 for (; n > 1; n >>= 1)
971 static inline int exp2i(int n)
977 * Estimate the number of bisect steps left (after the current step)
979 * For any x between 0 included and 2^n excluded, the probability for
980 * n - 1 steps left looks like:
982 * P(2^n + x) == (2^n - x) / (2^n + x)
984 * and P(2^n + x) < 0.5 means 2^n < 3x
986 int estimate_bisect_steps(int all)
997 return (e < 3 * x) ? n : n - 1;