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 struct object_id *current_bad_oid;
20 static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
21 static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
23 static const char *term_bad;
24 static const char *term_good;
26 /* Remember to update object flag allocation in object.h */
27 #define COUNTED (1u<<16)
30 * This is a truly stupid algorithm, but it's only
31 * used for bisection, and we just don't care enough.
33 * We care just barely enough to avoid recursing for
36 static int count_distance(struct commit_list *entry)
41 struct commit *commit = entry->item;
42 struct commit_list *p;
44 if (commit->object.flags & (UNINTERESTING | COUNTED))
46 if (!(commit->object.flags & TREESAME))
48 commit->object.flags |= COUNTED;
54 nr += count_distance(p);
63 static void clear_distance(struct commit_list *list)
66 struct commit *commit = list->item;
67 commit->object.flags &= ~COUNTED;
72 #define DEBUG_BISECT 0
74 static inline int weight(struct commit_list *elem)
76 return *((int*)(elem->item->util));
79 static inline void weight_set(struct commit_list *elem, int weight)
81 *((int*)(elem->item->util)) = weight;
84 static int count_interesting_parents(struct commit *commit)
86 struct commit_list *p;
89 for (count = 0, p = commit->parents; p; p = p->next) {
90 if (p->item->object.flags & UNINTERESTING)
97 static inline int halfway(struct commit_list *p, int nr)
100 * Don't short-cut something we are not going to return!
102 if (p->item->object.flags & TREESAME)
107 * 2 and 3 are halfway of 5.
108 * 3 is halfway of 6 but 2 and 4 are not.
110 switch (2 * weight(p) - nr) {
111 case -1: case 0: case 1:
119 #define show_list(a,b,c,d) do { ; } while (0)
121 static void show_list(const char *debug, int counted, int nr,
122 struct commit_list *list)
124 struct commit_list *p;
126 fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
128 for (p = list; p; p = p->next) {
129 struct commit_list *pp;
130 struct commit *commit = p->item;
131 unsigned flags = commit->object.flags;
132 enum object_type type;
134 char *buf = read_sha1_file(commit->object.sha1, &type, &size);
135 const char *subject_start;
138 fprintf(stderr, "%c%c%c ",
139 (flags & TREESAME) ? ' ' : 'T',
140 (flags & UNINTERESTING) ? 'U' : ' ',
141 (flags & COUNTED) ? 'C' : ' ');
143 fprintf(stderr, "%3d", weight(p));
145 fprintf(stderr, "---");
146 fprintf(stderr, " %.*s", 8, sha1_to_hex(commit->object.sha1));
147 for (pp = commit->parents; pp; pp = pp->next)
148 fprintf(stderr, " %.*s", 8,
149 sha1_to_hex(pp->item->object.sha1));
151 subject_len = find_commit_subject(buf, &subject_start);
153 fprintf(stderr, " %.*s", subject_len, subject_start);
154 fprintf(stderr, "\n");
157 #endif /* DEBUG_BISECT */
159 static struct commit_list *best_bisection(struct commit_list *list, int nr)
161 struct commit_list *p, *best;
162 int best_distance = -1;
165 for (p = list; p; p = p->next) {
167 unsigned flags = p->item->object.flags;
169 if (flags & TREESAME)
171 distance = weight(p);
172 if (nr - distance < distance)
173 distance = nr - distance;
174 if (distance > best_distance) {
176 best_distance = distance;
184 struct commit *commit;
188 static int compare_commit_dist(const void *a_, const void *b_)
190 struct commit_dist *a, *b;
192 a = (struct commit_dist *)a_;
193 b = (struct commit_dist *)b_;
194 if (a->distance != b->distance)
195 return b->distance - a->distance; /* desc sort */
196 return oidcmp(&a->commit->object.oid, &b->commit->object.oid);
199 static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
201 struct commit_list *p;
202 struct commit_dist *array = xcalloc(nr, sizeof(*array));
205 for (p = list, cnt = 0; p; p = p->next) {
207 unsigned flags = p->item->object.flags;
209 if (flags & TREESAME)
211 distance = weight(p);
212 if (nr - distance < distance)
213 distance = nr - distance;
214 array[cnt].commit = p->item;
215 array[cnt].distance = distance;
218 qsort(array, cnt, sizeof(*array), compare_commit_dist);
219 for (p = list, i = 0; i < cnt; i++) {
220 char buf[100]; /* enough for dist=%d */
221 struct object *obj = &(array[i].commit->object);
223 snprintf(buf, sizeof(buf), "dist=%d", array[i].distance);
224 add_name_decoration(DECORATION_NONE, buf, obj);
226 p->item = array[i].commit;
236 * zero or positive weight is the number of interesting commits it can
237 * reach, including itself. Especially, weight = 0 means it does not
238 * reach any tree-changing commits (e.g. just above uninteresting one
239 * but traversal is with pathspec).
241 * weight = -1 means it has one parent and its distance is yet to
244 * weight = -2 means it has more than one parent and its distance is
245 * unknown. After running count_distance() first, they will get zero
246 * or positive distance.
248 static struct commit_list *do_find_bisection(struct commit_list *list,
249 int nr, int *weights,
253 struct commit_list *p;
257 for (n = 0, p = list; p; p = p->next) {
258 struct commit *commit = p->item;
259 unsigned flags = commit->object.flags;
261 p->item->util = &weights[n++];
262 switch (count_interesting_parents(commit)) {
264 if (!(flags & TREESAME)) {
267 show_list("bisection 2 count one",
271 * otherwise, it is known not to reach any
272 * tree-changing commit and gets weight 0.
284 show_list("bisection 2 initialize", counted, nr, list);
287 * If you have only one parent in the resulting set
288 * then you can reach one commit more than that parent
289 * can reach. So we do not have to run the expensive
290 * count_distance() for single strand of pearls.
292 * However, if you have more than one parents, you cannot
293 * just add their distance and one for yourself, since
294 * they usually reach the same ancestor and you would
295 * end up counting them twice that way.
297 * So we will first count distance of merges the usual
298 * way, and then fill the blanks using cheaper algorithm.
300 for (p = list; p; p = p->next) {
301 if (p->item->object.flags & UNINTERESTING)
305 weight_set(p, count_distance(p));
306 clear_distance(list);
308 /* Does it happen to be at exactly half-way? */
309 if (!find_all && halfway(p, nr))
314 show_list("bisection 2 count_distance", counted, nr, list);
316 while (counted < nr) {
317 for (p = list; p; p = p->next) {
318 struct commit_list *q;
319 unsigned flags = p->item->object.flags;
323 for (q = p->item->parents; q; q = q->next) {
324 if (q->item->object.flags & UNINTERESTING)
333 * weight for p is unknown but q is known.
334 * add one for p itself if p is to be counted,
335 * otherwise inherit it from q directly.
337 if (!(flags & TREESAME)) {
338 weight_set(p, weight(q)+1);
340 show_list("bisection 2 count one",
344 weight_set(p, weight(q));
346 /* Does it happen to be at exactly half-way? */
347 if (!find_all && halfway(p, nr))
352 show_list("bisection 2 counted all", counted, nr, list);
355 return best_bisection(list, nr);
357 return best_bisection_sorted(list, nr);
360 struct commit_list *find_bisection(struct commit_list *list,
361 int *reaches, int *all,
365 struct commit_list *p, *best, *next, *last;
368 show_list("bisection 2 entry", 0, 0, list);
371 * Count the number of total and tree-changing items on the
372 * list, while reversing the list.
374 for (nr = on_list = 0, last = NULL, p = list;
377 unsigned flags = p->item->object.flags;
380 if (flags & UNINTERESTING)
384 if (!(flags & TREESAME))
389 show_list("bisection 2 sorted", 0, nr, list);
392 weights = xcalloc(on_list, sizeof(*weights));
394 /* Do the real work of finding bisection commit. */
395 best = do_find_bisection(list, nr, weights, find_all);
399 *reaches = weight(best);
405 static int register_ref(const char *refname, const struct object_id *oid,
406 int flags, void *cb_data)
408 struct strbuf good_prefix = STRBUF_INIT;
409 strbuf_addstr(&good_prefix, term_good);
410 strbuf_addstr(&good_prefix, "-");
412 if (!strcmp(refname, term_bad)) {
413 current_bad_oid = xmalloc(sizeof(*current_bad_oid));
414 oidcpy(current_bad_oid, oid);
415 } else if (starts_with(refname, good_prefix.buf)) {
416 sha1_array_append(&good_revs, oid->hash);
417 } else if (starts_with(refname, "skip-")) {
418 sha1_array_append(&skipped_revs, oid->hash);
421 strbuf_release(&good_prefix);
426 static int read_bisect_refs(void)
428 return for_each_ref_in("refs/bisect/", register_ref, NULL);
431 static GIT_PATH_FUNC(git_path_bisect_names, "BISECT_NAMES")
432 static GIT_PATH_FUNC(git_path_bisect_expected_rev, "BISECT_EXPECTED_REV")
434 static void read_bisect_paths(struct argv_array *array)
436 struct strbuf str = STRBUF_INIT;
437 const char *filename = git_path_bisect_names();
438 FILE *fp = fopen(filename, "r");
441 die_errno("Could not open file '%s'", filename);
443 while (strbuf_getline_lf(&str, fp) != EOF) {
445 if (sq_dequote_to_argv_array(str.buf, array))
446 die("Badly quoted content in file '%s': %s",
450 strbuf_release(&str);
454 static char *join_sha1_array_hex(struct sha1_array *array, char delim)
456 struct strbuf joined_hexs = STRBUF_INIT;
459 for (i = 0; i < array->nr; i++) {
460 strbuf_addstr(&joined_hexs, sha1_to_hex(array->sha1[i]));
461 if (i + 1 < array->nr)
462 strbuf_addch(&joined_hexs, delim);
465 return strbuf_detach(&joined_hexs, NULL);
469 * In this function, passing a not NULL skipped_first is very special.
470 * It means that we want to know if the first commit in the list is
471 * skipped because we will want to test a commit away from it if it is
473 * So if the first commit is skipped, we cannot take the shortcut to
474 * just "return list" when we find the first non skipped commit, we
475 * have to return a fully filtered list.
477 * We use (*skipped_first == -1) to mean "it has been found that the
478 * first commit is not skipped". In this case *skipped_first is set back
479 * to 0 just before the function returns.
481 struct commit_list *filter_skipped(struct commit_list *list,
482 struct commit_list **tried,
487 struct commit_list *filtered = NULL, **f = &filtered;
496 if (!skipped_revs.nr)
500 struct commit_list *next = list->next;
502 if (0 <= sha1_array_lookup(&skipped_revs,
503 list->item->object.oid.hash)) {
504 if (skipped_first && !*skipped_first)
506 /* Move current to tried list */
511 if (!skipped_first || !*skipped_first)
513 } else if (skipped_first && !*skipped_first) {
514 /* This means we know it's not skipped */
517 /* Move current to filtered list */
526 if (skipped_first && *skipped_first == -1)
532 #define PRN_MODULO 32768
535 * This is a pseudo random number generator based on "man 3 rand".
536 * It is not used properly because the seed is the argument and it
537 * is increased by one between each call, but that should not matter
538 * for this application.
540 static unsigned get_prn(unsigned count) {
541 count = count * 1103515245 + 12345;
542 return (count/65536) % PRN_MODULO;
546 * Custom integer square root from
547 * http://en.wikipedia.org/wiki/Integer_square_root
549 static int sqrti(int val)
557 float y = (x + (float)val / x) / 2;
558 d = (y > x) ? y - x : x - y;
565 static struct commit_list *skip_away(struct commit_list *list, int count)
567 struct commit_list *cur, *previous;
570 prn = get_prn(count);
571 index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
576 for (i = 0; cur; cur = cur->next, i++) {
578 if (oidcmp(&cur->item->object.oid, current_bad_oid))
590 static struct commit_list *managed_skipped(struct commit_list *list,
591 struct commit_list **tried)
593 int count, skipped_first;
597 if (!skipped_revs.nr)
600 list = filter_skipped(list, tried, 0, &count, &skipped_first);
605 return skip_away(list, count);
608 static void bisect_rev_setup(struct rev_info *revs, const char *prefix,
609 const char *bad_format, const char *good_format,
612 struct argv_array rev_argv = ARGV_ARRAY_INIT;
615 init_revisions(revs, prefix);
617 revs->commit_format = CMIT_FMT_UNSPECIFIED;
619 /* rev_argv.argv[0] will be ignored by setup_revisions */
620 argv_array_push(&rev_argv, "bisect_rev_setup");
621 argv_array_pushf(&rev_argv, bad_format, oid_to_hex(current_bad_oid));
622 for (i = 0; i < good_revs.nr; i++)
623 argv_array_pushf(&rev_argv, good_format,
624 sha1_to_hex(good_revs.sha1[i]));
625 argv_array_push(&rev_argv, "--");
627 read_bisect_paths(&rev_argv);
629 setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
630 /* XXX leak rev_argv, as "revs" may still be pointing to it */
633 static void bisect_common(struct rev_info *revs)
635 if (prepare_revision_walk(revs))
636 die("revision walk setup failed");
637 if (revs->tree_objects)
638 mark_edges_uninteresting(revs, NULL);
641 static void exit_if_skipped_commits(struct commit_list *tried,
642 const struct object_id *bad)
647 printf("There are only 'skip'ped commits left to test.\n"
648 "The first %s commit could be any of:\n", term_bad);
649 print_commit_list(tried, "%s\n", "%s\n");
651 printf("%s\n", oid_to_hex(bad));
652 printf("We cannot bisect more!\n");
656 static int is_expected_rev(const struct object_id *oid)
658 const char *filename = git_path_bisect_expected_rev();
660 struct strbuf str = STRBUF_INIT;
664 if (stat(filename, &st) || !S_ISREG(st.st_mode))
667 fp = fopen(filename, "r");
671 if (strbuf_getline_lf(&str, fp) != EOF)
672 res = !strcmp(str.buf, oid_to_hex(oid));
674 strbuf_release(&str);
680 static int bisect_checkout(const unsigned char *bisect_rev, int no_checkout)
682 char bisect_rev_hex[GIT_SHA1_HEXSZ + 1];
684 memcpy(bisect_rev_hex, sha1_to_hex(bisect_rev), GIT_SHA1_HEXSZ + 1);
685 update_ref(NULL, "BISECT_EXPECTED_REV", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
687 argv_checkout[2] = bisect_rev_hex;
689 update_ref(NULL, "BISECT_HEAD", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
692 res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
697 argv_show_branch[1] = bisect_rev_hex;
698 return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
701 static struct commit *get_commit_reference(const unsigned char *sha1)
703 struct commit *r = lookup_commit_reference(sha1);
705 die("Not a valid commit name %s", sha1_to_hex(sha1));
709 static struct commit **get_bad_and_good_commits(int *rev_nr)
711 int len = 1 + good_revs.nr;
712 struct commit **rev = xmalloc(len * sizeof(*rev));
715 rev[n++] = get_commit_reference(current_bad_oid->hash);
716 for (i = 0; i < good_revs.nr; i++)
717 rev[n++] = get_commit_reference(good_revs.sha1[i]);
723 static void handle_bad_merge_base(void)
725 if (is_expected_rev(current_bad_oid)) {
726 char *bad_hex = oid_to_hex(current_bad_oid);
727 char *good_hex = join_sha1_array_hex(&good_revs, ' ');
728 if (!strcmp(term_bad, "bad") && !strcmp(term_good, "good")) {
729 fprintf(stderr, "The merge base %s is bad.\n"
730 "This means the bug has been fixed "
731 "between %s and [%s].\n",
732 bad_hex, bad_hex, good_hex);
733 } else if (!strcmp(term_bad, "new") && !strcmp(term_good, "old")) {
734 fprintf(stderr, "The merge base %s is new.\n"
735 "The property has changed "
736 "between %s and [%s].\n",
737 bad_hex, bad_hex, good_hex);
739 fprintf(stderr, "The merge base %s is %s.\n"
740 "This means the first '%s' commit is "
741 "between %s and [%s].\n",
742 bad_hex, term_bad, term_good, bad_hex, good_hex);
747 fprintf(stderr, "Some %s revs are not ancestor of the %s rev.\n"
748 "git bisect cannot work properly in this case.\n"
749 "Maybe you mistook %s and %s revs?\n",
750 term_good, term_bad, term_good, term_bad);
754 static void handle_skipped_merge_base(const unsigned char *mb)
756 char *mb_hex = sha1_to_hex(mb);
757 char *bad_hex = sha1_to_hex(current_bad_oid->hash);
758 char *good_hex = join_sha1_array_hex(&good_revs, ' ');
760 warning("the merge base between %s and [%s] "
762 "So we cannot be sure the first %s commit is "
763 "between %s and %s.\n"
764 "We continue anyway.",
765 bad_hex, good_hex, term_bad, mb_hex, bad_hex);
770 * "check_merge_bases" checks that merge bases are not "bad" (or "new").
772 * - If one is "bad" (or "new"), it means the user assumed something wrong
773 * and we must exit with a non 0 error code.
774 * - If one is "good" (or "old"), that's good, we have nothing to do.
775 * - If one is "skipped", we can't know but we should warn.
776 * - If we don't know, we should check it out and ask the user to test.
778 static void check_merge_bases(int no_checkout)
780 struct commit_list *result;
782 struct commit **rev = get_bad_and_good_commits(&rev_nr);
784 result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1);
786 for (; result; result = result->next) {
787 const unsigned char *mb = result->item->object.oid.hash;
788 if (!hashcmp(mb, current_bad_oid->hash)) {
789 handle_bad_merge_base();
790 } else if (0 <= sha1_array_lookup(&good_revs, mb)) {
792 } else if (0 <= sha1_array_lookup(&skipped_revs, mb)) {
793 handle_skipped_merge_base(mb);
795 printf("Bisecting: a merge base must be tested\n");
796 exit(bisect_checkout(mb, no_checkout));
801 free_commit_list(result);
804 static int check_ancestors(const char *prefix)
806 struct rev_info revs;
807 struct object_array pending_copy;
810 bisect_rev_setup(&revs, prefix, "^%s", "%s", 0);
812 /* Save pending objects, so they can be cleaned up later. */
813 pending_copy = revs.pending;
814 revs.leak_pending = 1;
817 * bisect_common calls prepare_revision_walk right away, which
818 * (together with .leak_pending = 1) makes us the sole owner of
819 * the list of pending objects.
821 bisect_common(&revs);
822 res = (revs.commits != NULL);
824 /* Clean up objects used, as they will be reused. */
825 clear_commit_marks_for_object_array(&pending_copy, ALL_REV_FLAGS);
826 free(pending_copy.objects);
832 * "check_good_are_ancestors_of_bad" checks that all "good" revs are
833 * ancestor of the "bad" rev.
835 * If that's not the case, we need to check the merge bases.
836 * If a merge base must be tested by the user, its source code will be
837 * checked out to be tested by the user and we will exit.
839 static void check_good_are_ancestors_of_bad(const char *prefix, int no_checkout)
841 char *filename = git_pathdup("BISECT_ANCESTORS_OK");
845 if (!current_bad_oid)
846 die("a %s revision is needed", term_bad);
848 /* Check if file BISECT_ANCESTORS_OK exists. */
849 if (!stat(filename, &st) && S_ISREG(st.st_mode))
852 /* Bisecting with no good rev is ok. */
853 if (good_revs.nr == 0)
856 /* Check if all good revs are ancestor of the bad rev. */
857 if (check_ancestors(prefix))
858 check_merge_bases(no_checkout);
860 /* Create file BISECT_ANCESTORS_OK. */
861 fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
863 warning("could not create file '%s': %s",
864 filename, strerror(errno));
872 * This does "git diff-tree --pretty COMMIT" without one fork+exec.
874 static void show_diff_tree(const char *prefix, struct commit *commit)
879 init_revisions(&opt, prefix);
880 git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
884 /* This is what "--pretty" does */
885 opt.verbose_header = 1;
886 opt.use_terminator = 0;
887 opt.commit_format = CMIT_FMT_DEFAULT;
890 if (!opt.diffopt.output_format)
891 opt.diffopt.output_format = DIFF_FORMAT_RAW;
893 log_tree_commit(&opt, commit);
897 * The terms used for this bisect session are stored in BISECT_TERMS.
898 * We read them and store them to adapt the messages accordingly.
899 * Default is bad/good.
901 void read_bisect_terms(const char **read_bad, const char **read_good)
903 struct strbuf str = STRBUF_INIT;
904 const char *filename = git_path("BISECT_TERMS");
905 FILE *fp = fopen(filename, "r");
908 if (errno == ENOENT) {
913 die("could not read file '%s': %s", filename,
917 strbuf_getline_lf(&str, fp);
918 *read_bad = strbuf_detach(&str, NULL);
919 strbuf_getline_lf(&str, fp);
920 *read_good = strbuf_detach(&str, NULL);
922 strbuf_release(&str);
927 * We use the convention that exiting with an exit code 10 means that
928 * the bisection process finished successfully.
929 * In this case the calling shell script should exit 0.
931 * If no_checkout is non-zero, the bisection process does not
932 * checkout the trial commit but instead simply updates BISECT_HEAD.
934 int bisect_next_all(const char *prefix, int no_checkout)
936 struct rev_info revs;
937 struct commit_list *tried;
938 int reaches = 0, all = 0, nr, steps;
939 const unsigned char *bisect_rev;
941 read_bisect_terms(&term_bad, &term_good);
942 if (read_bisect_refs())
943 die("reading bisect refs failed");
945 check_good_are_ancestors_of_bad(prefix, no_checkout);
947 bisect_rev_setup(&revs, prefix, "%s", "^%s", 1);
950 bisect_common(&revs);
952 revs.commits = find_bisection(revs.commits, &reaches, &all,
954 revs.commits = managed_skipped(revs.commits, &tried);
958 * We should exit here only if the "bad"
959 * commit is also a "skip" commit.
961 exit_if_skipped_commits(tried, NULL);
963 printf("%s was both %s and %s\n",
964 oid_to_hex(current_bad_oid),
971 fprintf(stderr, "No testable commit found.\n"
972 "Maybe you started with bad path parameters?\n");
976 bisect_rev = revs.commits->item->object.oid.hash;
978 if (!hashcmp(bisect_rev, current_bad_oid->hash)) {
979 exit_if_skipped_commits(tried, current_bad_oid);
980 printf("%s is the first %s commit\n", sha1_to_hex(bisect_rev),
982 show_diff_tree(prefix, revs.commits->item);
983 /* This means the bisection process succeeded. */
987 nr = all - reaches - 1;
988 steps = estimate_bisect_steps(all);
989 printf("Bisecting: %d revision%s left to test after this "
990 "(roughly %d step%s)\n", nr, (nr == 1 ? "" : "s"),
991 steps, (steps == 1 ? "" : "s"));
993 return bisect_checkout(bisect_rev, no_checkout);
996 static inline int log2i(int n)
1000 for (; n > 1; n >>= 1)
1006 static inline int exp2i(int n)
1012 * Estimate the number of bisect steps left (after the current step)
1014 * For any x between 0 included and 2^n excluded, the probability for
1015 * n - 1 steps left looks like:
1017 * P(2^n + x) == (2^n - x) / (2^n + x)
1019 * and P(2^n + x) < 0.5 means 2^n < 3x
1021 int estimate_bisect_steps(int all)
1032 return (e < 3 * x) ? n : n - 1;