Merge branch 'mm/ls-files-s-doc' into next
[git] / bisect.c
1 #include "cache.h"
2 #include "commit.h"
3 #include "diff.h"
4 #include "revision.h"
5 #include "refs.h"
6 #include "list-objects.h"
7 #include "quote.h"
8 #include "sha1-lookup.h"
9 #include "run-command.h"
10 #include "log-tree.h"
11 #include "bisect.h"
12 #include "sha1-array.h"
13 #include "argv-array.h"
14
15 static struct sha1_array good_revs;
16 static struct sha1_array skipped_revs;
17
18 static struct object_id *current_bad_oid;
19
20 static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
21 static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
22
23 static const char *term_bad;
24 static const char *term_good;
25
26 /* Remember to update object flag allocation in object.h */
27 #define COUNTED         (1u<<16)
28
29 /*
30  * This is a truly stupid algorithm, but it's only
31  * used for bisection, and we just don't care enough.
32  *
33  * We care just barely enough to avoid recursing for
34  * non-merge entries.
35  */
36 static int count_distance(struct commit_list *entry)
37 {
38         int nr = 0;
39
40         while (entry) {
41                 struct commit *commit = entry->item;
42                 struct commit_list *p;
43
44                 if (commit->object.flags & (UNINTERESTING | COUNTED))
45                         break;
46                 if (!(commit->object.flags & TREESAME))
47                         nr++;
48                 commit->object.flags |= COUNTED;
49                 p = commit->parents;
50                 entry = p;
51                 if (p) {
52                         p = p->next;
53                         while (p) {
54                                 nr += count_distance(p);
55                                 p = p->next;
56                         }
57                 }
58         }
59
60         return nr;
61 }
62
63 static void clear_distance(struct commit_list *list)
64 {
65         while (list) {
66                 struct commit *commit = list->item;
67                 commit->object.flags &= ~COUNTED;
68                 list = list->next;
69         }
70 }
71
72 #define DEBUG_BISECT 0
73
74 static inline int weight(struct commit_list *elem)
75 {
76         return *((int*)(elem->item->util));
77 }
78
79 static inline void weight_set(struct commit_list *elem, int weight)
80 {
81         *((int*)(elem->item->util)) = weight;
82 }
83
84 static int count_interesting_parents(struct commit *commit)
85 {
86         struct commit_list *p;
87         int count;
88
89         for (count = 0, p = commit->parents; p; p = p->next) {
90                 if (p->item->object.flags & UNINTERESTING)
91                         continue;
92                 count++;
93         }
94         return count;
95 }
96
97 static inline int halfway(struct commit_list *p, int nr)
98 {
99         /*
100          * Don't short-cut something we are not going to return!
101          */
102         if (p->item->object.flags & TREESAME)
103                 return 0;
104         if (DEBUG_BISECT)
105                 return 0;
106         /*
107          * 2 and 3 are halfway of 5.
108          * 3 is halfway of 6 but 2 and 4 are not.
109          */
110         switch (2 * weight(p) - nr) {
111         case -1: case 0: case 1:
112                 return 1;
113         default:
114                 return 0;
115         }
116 }
117
118 #if !DEBUG_BISECT
119 #define show_list(a,b,c,d) do { ; } while (0)
120 #else
121 static void show_list(const char *debug, int counted, int nr,
122                       struct commit_list *list)
123 {
124         struct commit_list *p;
125
126         fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
127
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;
133                 unsigned long size;
134                 char *buf = read_sha1_file(commit->object.sha1, &type, &size);
135                 const char *subject_start;
136                 int subject_len;
137
138                 fprintf(stderr, "%c%c%c ",
139                         (flags & TREESAME) ? ' ' : 'T',
140                         (flags & UNINTERESTING) ? 'U' : ' ',
141                         (flags & COUNTED) ? 'C' : ' ');
142                 if (commit->util)
143                         fprintf(stderr, "%3d", weight(p));
144                 else
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));
150
151                 subject_len = find_commit_subject(buf, &subject_start);
152                 if (subject_len)
153                         fprintf(stderr, " %.*s", subject_len, subject_start);
154                 fprintf(stderr, "\n");
155         }
156 }
157 #endif /* DEBUG_BISECT */
158
159 static struct commit_list *best_bisection(struct commit_list *list, int nr)
160 {
161         struct commit_list *p, *best;
162         int best_distance = -1;
163
164         best = list;
165         for (p = list; p; p = p->next) {
166                 int distance;
167                 unsigned flags = p->item->object.flags;
168
169                 if (flags & TREESAME)
170                         continue;
171                 distance = weight(p);
172                 if (nr - distance < distance)
173                         distance = nr - distance;
174                 if (distance > best_distance) {
175                         best = p;
176                         best_distance = distance;
177                 }
178         }
179
180         return best;
181 }
182
183 struct commit_dist {
184         struct commit *commit;
185         int distance;
186 };
187
188 static int compare_commit_dist(const void *a_, const void *b_)
189 {
190         struct commit_dist *a, *b;
191
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);
197 }
198
199 static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
200 {
201         struct commit_list *p;
202         struct commit_dist *array = xcalloc(nr, sizeof(*array));
203         struct strbuf buf = STRBUF_INIT;
204         int cnt, i;
205
206         for (p = list, cnt = 0; p; p = p->next) {
207                 int distance;
208                 unsigned flags = p->item->object.flags;
209
210                 if (flags & TREESAME)
211                         continue;
212                 distance = weight(p);
213                 if (nr - distance < distance)
214                         distance = nr - distance;
215                 array[cnt].commit = p->item;
216                 array[cnt].distance = distance;
217                 cnt++;
218         }
219         QSORT(array, cnt, compare_commit_dist);
220         for (p = list, i = 0; i < cnt; i++) {
221                 struct object *obj = &(array[i].commit->object);
222
223                 strbuf_reset(&buf);
224                 strbuf_addf(&buf, "dist=%d", array[i].distance);
225                 add_name_decoration(DECORATION_NONE, buf.buf, obj);
226
227                 p->item = array[i].commit;
228                 p = p->next;
229         }
230         if (p)
231                 p->next = NULL;
232         strbuf_release(&buf);
233         free(array);
234         return list;
235 }
236
237 /*
238  * zero or positive weight is the number of interesting commits it can
239  * reach, including itself.  Especially, weight = 0 means it does not
240  * reach any tree-changing commits (e.g. just above uninteresting one
241  * but traversal is with pathspec).
242  *
243  * weight = -1 means it has one parent and its distance is yet to
244  * be computed.
245  *
246  * weight = -2 means it has more than one parent and its distance is
247  * unknown.  After running count_distance() first, they will get zero
248  * or positive distance.
249  */
250 static struct commit_list *do_find_bisection(struct commit_list *list,
251                                              int nr, int *weights,
252                                              int find_all)
253 {
254         int n, counted;
255         struct commit_list *p;
256
257         counted = 0;
258
259         for (n = 0, p = list; p; p = p->next) {
260                 struct commit *commit = p->item;
261                 unsigned flags = commit->object.flags;
262
263                 p->item->util = &weights[n++];
264                 switch (count_interesting_parents(commit)) {
265                 case 0:
266                         if (!(flags & TREESAME)) {
267                                 weight_set(p, 1);
268                                 counted++;
269                                 show_list("bisection 2 count one",
270                                           counted, nr, list);
271                         }
272                         /*
273                          * otherwise, it is known not to reach any
274                          * tree-changing commit and gets weight 0.
275                          */
276                         break;
277                 case 1:
278                         weight_set(p, -1);
279                         break;
280                 default:
281                         weight_set(p, -2);
282                         break;
283                 }
284         }
285
286         show_list("bisection 2 initialize", counted, nr, list);
287
288         /*
289          * If you have only one parent in the resulting set
290          * then you can reach one commit more than that parent
291          * can reach.  So we do not have to run the expensive
292          * count_distance() for single strand of pearls.
293          *
294          * However, if you have more than one parents, you cannot
295          * just add their distance and one for yourself, since
296          * they usually reach the same ancestor and you would
297          * end up counting them twice that way.
298          *
299          * So we will first count distance of merges the usual
300          * way, and then fill the blanks using cheaper algorithm.
301          */
302         for (p = list; p; p = p->next) {
303                 if (p->item->object.flags & UNINTERESTING)
304                         continue;
305                 if (weight(p) != -2)
306                         continue;
307                 weight_set(p, count_distance(p));
308                 clear_distance(list);
309
310                 /* Does it happen to be at exactly half-way? */
311                 if (!find_all && halfway(p, nr))
312                         return p;
313                 counted++;
314         }
315
316         show_list("bisection 2 count_distance", counted, nr, list);
317
318         while (counted < nr) {
319                 for (p = list; p; p = p->next) {
320                         struct commit_list *q;
321                         unsigned flags = p->item->object.flags;
322
323                         if (0 <= weight(p))
324                                 continue;
325                         for (q = p->item->parents; q; q = q->next) {
326                                 if (q->item->object.flags & UNINTERESTING)
327                                         continue;
328                                 if (0 <= weight(q))
329                                         break;
330                         }
331                         if (!q)
332                                 continue;
333
334                         /*
335                          * weight for p is unknown but q is known.
336                          * add one for p itself if p is to be counted,
337                          * otherwise inherit it from q directly.
338                          */
339                         if (!(flags & TREESAME)) {
340                                 weight_set(p, weight(q)+1);
341                                 counted++;
342                                 show_list("bisection 2 count one",
343                                           counted, nr, list);
344                         }
345                         else
346                                 weight_set(p, weight(q));
347
348                         /* Does it happen to be at exactly half-way? */
349                         if (!find_all && halfway(p, nr))
350                                 return p;
351                 }
352         }
353
354         show_list("bisection 2 counted all", counted, nr, list);
355
356         if (!find_all)
357                 return best_bisection(list, nr);
358         else
359                 return best_bisection_sorted(list, nr);
360 }
361
362 struct commit_list *find_bisection(struct commit_list *list,
363                                           int *reaches, int *all,
364                                           int find_all)
365 {
366         int nr, on_list;
367         struct commit_list *p, *best, *next, *last;
368         int *weights;
369
370         show_list("bisection 2 entry", 0, 0, list);
371
372         /*
373          * Count the number of total and tree-changing items on the
374          * list, while reversing the list.
375          */
376         for (nr = on_list = 0, last = NULL, p = list;
377              p;
378              p = next) {
379                 unsigned flags = p->item->object.flags;
380
381                 next = p->next;
382                 if (flags & UNINTERESTING)
383                         continue;
384                 p->next = last;
385                 last = p;
386                 if (!(flags & TREESAME))
387                         nr++;
388                 on_list++;
389         }
390         list = last;
391         show_list("bisection 2 sorted", 0, nr, list);
392
393         *all = nr;
394         weights = xcalloc(on_list, sizeof(*weights));
395
396         /* Do the real work of finding bisection commit. */
397         best = do_find_bisection(list, nr, weights, find_all);
398         if (best) {
399                 if (!find_all)
400                         best->next = NULL;
401                 *reaches = weight(best);
402         }
403         free(weights);
404         return best;
405 }
406
407 static int register_ref(const char *refname, const struct object_id *oid,
408                         int flags, void *cb_data)
409 {
410         struct strbuf good_prefix = STRBUF_INIT;
411         strbuf_addstr(&good_prefix, term_good);
412         strbuf_addstr(&good_prefix, "-");
413
414         if (!strcmp(refname, term_bad)) {
415                 current_bad_oid = xmalloc(sizeof(*current_bad_oid));
416                 oidcpy(current_bad_oid, oid);
417         } else if (starts_with(refname, good_prefix.buf)) {
418                 sha1_array_append(&good_revs, oid->hash);
419         } else if (starts_with(refname, "skip-")) {
420                 sha1_array_append(&skipped_revs, oid->hash);
421         }
422
423         strbuf_release(&good_prefix);
424
425         return 0;
426 }
427
428 static int read_bisect_refs(void)
429 {
430         return for_each_ref_in("refs/bisect/", register_ref, NULL);
431 }
432
433 static GIT_PATH_FUNC(git_path_bisect_names, "BISECT_NAMES")
434 static GIT_PATH_FUNC(git_path_bisect_expected_rev, "BISECT_EXPECTED_REV")
435
436 static void read_bisect_paths(struct argv_array *array)
437 {
438         struct strbuf str = STRBUF_INIT;
439         const char *filename = git_path_bisect_names();
440         FILE *fp = fopen(filename, "r");
441
442         if (!fp)
443                 die_errno(_("Could not open file '%s'"), filename);
444
445         while (strbuf_getline_lf(&str, fp) != EOF) {
446                 strbuf_trim(&str);
447                 if (sq_dequote_to_argv_array(str.buf, array))
448                         die(_("Badly quoted content in file '%s': %s"),
449                             filename, str.buf);
450         }
451
452         strbuf_release(&str);
453         fclose(fp);
454 }
455
456 static char *join_sha1_array_hex(struct sha1_array *array, char delim)
457 {
458         struct strbuf joined_hexs = STRBUF_INIT;
459         int i;
460
461         for (i = 0; i < array->nr; i++) {
462                 strbuf_addstr(&joined_hexs, sha1_to_hex(array->sha1[i]));
463                 if (i + 1 < array->nr)
464                         strbuf_addch(&joined_hexs, delim);
465         }
466
467         return strbuf_detach(&joined_hexs, NULL);
468 }
469
470 /*
471  * In this function, passing a not NULL skipped_first is very special.
472  * It means that we want to know if the first commit in the list is
473  * skipped because we will want to test a commit away from it if it is
474  * indeed skipped.
475  * So if the first commit is skipped, we cannot take the shortcut to
476  * just "return list" when we find the first non skipped commit, we
477  * have to return a fully filtered list.
478  *
479  * We use (*skipped_first == -1) to mean "it has been found that the
480  * first commit is not skipped". In this case *skipped_first is set back
481  * to 0 just before the function returns.
482  */
483 struct commit_list *filter_skipped(struct commit_list *list,
484                                    struct commit_list **tried,
485                                    int show_all,
486                                    int *count,
487                                    int *skipped_first)
488 {
489         struct commit_list *filtered = NULL, **f = &filtered;
490
491         *tried = NULL;
492
493         if (skipped_first)
494                 *skipped_first = 0;
495         if (count)
496                 *count = 0;
497
498         if (!skipped_revs.nr)
499                 return list;
500
501         while (list) {
502                 struct commit_list *next = list->next;
503                 list->next = NULL;
504                 if (0 <= sha1_array_lookup(&skipped_revs,
505                                            list->item->object.oid.hash)) {
506                         if (skipped_first && !*skipped_first)
507                                 *skipped_first = 1;
508                         /* Move current to tried list */
509                         *tried = list;
510                         tried = &list->next;
511                 } else {
512                         if (!show_all) {
513                                 if (!skipped_first || !*skipped_first)
514                                         return list;
515                         } else if (skipped_first && !*skipped_first) {
516                                 /* This means we know it's not skipped */
517                                 *skipped_first = -1;
518                         }
519                         /* Move current to filtered list */
520                         *f = list;
521                         f = &list->next;
522                         if (count)
523                                 (*count)++;
524                 }
525                 list = next;
526         }
527
528         if (skipped_first && *skipped_first == -1)
529                 *skipped_first = 0;
530
531         return filtered;
532 }
533
534 #define PRN_MODULO 32768
535
536 /*
537  * This is a pseudo random number generator based on "man 3 rand".
538  * It is not used properly because the seed is the argument and it
539  * is increased by one between each call, but that should not matter
540  * for this application.
541  */
542 static unsigned get_prn(unsigned count) {
543         count = count * 1103515245 + 12345;
544         return (count/65536) % PRN_MODULO;
545 }
546
547 /*
548  * Custom integer square root from
549  * http://en.wikipedia.org/wiki/Integer_square_root
550  */
551 static int sqrti(int val)
552 {
553         float d, x = val;
554
555         if (val == 0)
556                 return 0;
557
558         do {
559                 float y = (x + (float)val / x) / 2;
560                 d = (y > x) ? y - x : x - y;
561                 x = y;
562         } while (d >= 0.5);
563
564         return (int)x;
565 }
566
567 static struct commit_list *skip_away(struct commit_list *list, int count)
568 {
569         struct commit_list *cur, *previous;
570         int prn, index, i;
571
572         prn = get_prn(count);
573         index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
574
575         cur = list;
576         previous = NULL;
577
578         for (i = 0; cur; cur = cur->next, i++) {
579                 if (i == index) {
580                         if (oidcmp(&cur->item->object.oid, current_bad_oid))
581                                 return cur;
582                         if (previous)
583                                 return previous;
584                         return list;
585                 }
586                 previous = cur;
587         }
588
589         return list;
590 }
591
592 static struct commit_list *managed_skipped(struct commit_list *list,
593                                            struct commit_list **tried)
594 {
595         int count, skipped_first;
596
597         *tried = NULL;
598
599         if (!skipped_revs.nr)
600                 return list;
601
602         list = filter_skipped(list, tried, 0, &count, &skipped_first);
603
604         if (!skipped_first)
605                 return list;
606
607         return skip_away(list, count);
608 }
609
610 static void bisect_rev_setup(struct rev_info *revs, const char *prefix,
611                              const char *bad_format, const char *good_format,
612                              int read_paths)
613 {
614         struct argv_array rev_argv = ARGV_ARRAY_INIT;
615         int i;
616
617         init_revisions(revs, prefix);
618         revs->abbrev = 0;
619         revs->commit_format = CMIT_FMT_UNSPECIFIED;
620
621         /* rev_argv.argv[0] will be ignored by setup_revisions */
622         argv_array_push(&rev_argv, "bisect_rev_setup");
623         argv_array_pushf(&rev_argv, bad_format, oid_to_hex(current_bad_oid));
624         for (i = 0; i < good_revs.nr; i++)
625                 argv_array_pushf(&rev_argv, good_format,
626                                  sha1_to_hex(good_revs.sha1[i]));
627         argv_array_push(&rev_argv, "--");
628         if (read_paths)
629                 read_bisect_paths(&rev_argv);
630
631         setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
632         /* XXX leak rev_argv, as "revs" may still be pointing to it */
633 }
634
635 static void bisect_common(struct rev_info *revs)
636 {
637         if (prepare_revision_walk(revs))
638                 die("revision walk setup failed");
639         if (revs->tree_objects)
640                 mark_edges_uninteresting(revs, NULL);
641 }
642
643 static void exit_if_skipped_commits(struct commit_list *tried,
644                                     const struct object_id *bad)
645 {
646         if (!tried)
647                 return;
648
649         printf("There are only 'skip'ped commits left to test.\n"
650                "The first %s commit could be any of:\n", term_bad);
651
652         for ( ; tried; tried = tried->next)
653                 printf("%s\n", oid_to_hex(&tried->item->object.oid));
654
655         if (bad)
656                 printf("%s\n", oid_to_hex(bad));
657         printf(_("We cannot bisect more!\n"));
658         exit(2);
659 }
660
661 static int is_expected_rev(const struct object_id *oid)
662 {
663         const char *filename = git_path_bisect_expected_rev();
664         struct stat st;
665         struct strbuf str = STRBUF_INIT;
666         FILE *fp;
667         int res = 0;
668
669         if (stat(filename, &st) || !S_ISREG(st.st_mode))
670                 return 0;
671
672         fp = fopen(filename, "r");
673         if (!fp)
674                 return 0;
675
676         if (strbuf_getline_lf(&str, fp) != EOF)
677                 res = !strcmp(str.buf, oid_to_hex(oid));
678
679         strbuf_release(&str);
680         fclose(fp);
681
682         return res;
683 }
684
685 static int bisect_checkout(const unsigned char *bisect_rev, int no_checkout)
686 {
687         char bisect_rev_hex[GIT_SHA1_HEXSZ + 1];
688
689         memcpy(bisect_rev_hex, sha1_to_hex(bisect_rev), GIT_SHA1_HEXSZ + 1);
690         update_ref(NULL, "BISECT_EXPECTED_REV", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
691
692         argv_checkout[2] = bisect_rev_hex;
693         if (no_checkout) {
694                 update_ref(NULL, "BISECT_HEAD", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
695         } else {
696                 int res;
697                 res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
698                 if (res)
699                         exit(res);
700         }
701
702         argv_show_branch[1] = bisect_rev_hex;
703         return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
704 }
705
706 static struct commit *get_commit_reference(const unsigned char *sha1)
707 {
708         struct commit *r = lookup_commit_reference(sha1);
709         if (!r)
710                 die(_("Not a valid commit name %s"), sha1_to_hex(sha1));
711         return r;
712 }
713
714 static struct commit **get_bad_and_good_commits(int *rev_nr)
715 {
716         struct commit **rev;
717         int i, n = 0;
718
719         ALLOC_ARRAY(rev, 1 + good_revs.nr);
720         rev[n++] = get_commit_reference(current_bad_oid->hash);
721         for (i = 0; i < good_revs.nr; i++)
722                 rev[n++] = get_commit_reference(good_revs.sha1[i]);
723         *rev_nr = n;
724
725         return rev;
726 }
727
728 static void handle_bad_merge_base(void)
729 {
730         if (is_expected_rev(current_bad_oid)) {
731                 char *bad_hex = oid_to_hex(current_bad_oid);
732                 char *good_hex = join_sha1_array_hex(&good_revs, ' ');
733                 if (!strcmp(term_bad, "bad") && !strcmp(term_good, "good")) {
734                         fprintf(stderr, _("The merge base %s is bad.\n"
735                                 "This means the bug has been fixed "
736                                 "between %s and [%s].\n"),
737                                 bad_hex, bad_hex, good_hex);
738                 } else if (!strcmp(term_bad, "new") && !strcmp(term_good, "old")) {
739                         fprintf(stderr, _("The merge base %s is new.\n"
740                                 "The property has changed "
741                                 "between %s and [%s].\n"),
742                                 bad_hex, bad_hex, good_hex);
743                 } else {
744                         fprintf(stderr, _("The merge base %s is %s.\n"
745                                 "This means the first '%s' commit is "
746                                 "between %s and [%s].\n"),
747                                 bad_hex, term_bad, term_good, bad_hex, good_hex);
748                 }
749                 exit(3);
750         }
751
752         fprintf(stderr, _("Some %s revs are not ancestors of the %s rev.\n"
753                 "git bisect cannot work properly in this case.\n"
754                 "Maybe you mistook %s and %s revs?\n"),
755                 term_good, term_bad, term_good, term_bad);
756         exit(1);
757 }
758
759 static void handle_skipped_merge_base(const unsigned char *mb)
760 {
761         char *mb_hex = sha1_to_hex(mb);
762         char *bad_hex = oid_to_hex(current_bad_oid);
763         char *good_hex = join_sha1_array_hex(&good_revs, ' ');
764
765         warning(_("the merge base between %s and [%s] "
766                 "must be skipped.\n"
767                 "So we cannot be sure the first %s commit is "
768                 "between %s and %s.\n"
769                 "We continue anyway."),
770                 bad_hex, good_hex, term_bad, mb_hex, bad_hex);
771         free(good_hex);
772 }
773
774 /*
775  * "check_merge_bases" checks that merge bases are not "bad" (or "new").
776  *
777  * - If one is "bad" (or "new"), it means the user assumed something wrong
778  * and we must exit with a non 0 error code.
779  * - If one is "good" (or "old"), that's good, we have nothing to do.
780  * - If one is "skipped", we can't know but we should warn.
781  * - If we don't know, we should check it out and ask the user to test.
782  */
783 static void check_merge_bases(int no_checkout)
784 {
785         struct commit_list *result;
786         int rev_nr;
787         struct commit **rev = get_bad_and_good_commits(&rev_nr);
788
789         result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1);
790
791         for (; result; result = result->next) {
792                 const unsigned char *mb = result->item->object.oid.hash;
793                 if (!hashcmp(mb, current_bad_oid->hash)) {
794                         handle_bad_merge_base();
795                 } else if (0 <= sha1_array_lookup(&good_revs, mb)) {
796                         continue;
797                 } else if (0 <= sha1_array_lookup(&skipped_revs, mb)) {
798                         handle_skipped_merge_base(mb);
799                 } else {
800                         printf(_("Bisecting: a merge base must be tested\n"));
801                         exit(bisect_checkout(mb, no_checkout));
802                 }
803         }
804
805         free(rev);
806         free_commit_list(result);
807 }
808
809 static int check_ancestors(const char *prefix)
810 {
811         struct rev_info revs;
812         struct object_array pending_copy;
813         int res;
814
815         bisect_rev_setup(&revs, prefix, "^%s", "%s", 0);
816
817         /* Save pending objects, so they can be cleaned up later. */
818         pending_copy = revs.pending;
819         revs.leak_pending = 1;
820
821         /*
822          * bisect_common calls prepare_revision_walk right away, which
823          * (together with .leak_pending = 1) makes us the sole owner of
824          * the list of pending objects.
825          */
826         bisect_common(&revs);
827         res = (revs.commits != NULL);
828
829         /* Clean up objects used, as they will be reused. */
830         clear_commit_marks_for_object_array(&pending_copy, ALL_REV_FLAGS);
831         free(pending_copy.objects);
832
833         return res;
834 }
835
836 /*
837  * "check_good_are_ancestors_of_bad" checks that all "good" revs are
838  * ancestor of the "bad" rev.
839  *
840  * If that's not the case, we need to check the merge bases.
841  * If a merge base must be tested by the user, its source code will be
842  * checked out to be tested by the user and we will exit.
843  */
844 static void check_good_are_ancestors_of_bad(const char *prefix, int no_checkout)
845 {
846         char *filename = git_pathdup("BISECT_ANCESTORS_OK");
847         struct stat st;
848         int fd;
849
850         if (!current_bad_oid)
851                 die(_("a %s revision is needed"), term_bad);
852
853         /* Check if file BISECT_ANCESTORS_OK exists. */
854         if (!stat(filename, &st) && S_ISREG(st.st_mode))
855                 goto done;
856
857         /* Bisecting with no good rev is ok. */
858         if (good_revs.nr == 0)
859                 goto done;
860
861         /* Check if all good revs are ancestor of the bad rev. */
862         if (check_ancestors(prefix))
863                 check_merge_bases(no_checkout);
864
865         /* Create file BISECT_ANCESTORS_OK. */
866         fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
867         if (fd < 0)
868                 warning_errno(_("could not create file '%s'"),
869                               filename);
870         else
871                 close(fd);
872  done:
873         free(filename);
874 }
875
876 /*
877  * This does "git diff-tree --pretty COMMIT" without one fork+exec.
878  */
879 static void show_diff_tree(const char *prefix, struct commit *commit)
880 {
881         struct rev_info opt;
882
883         /* diff-tree init */
884         init_revisions(&opt, prefix);
885         git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
886         opt.abbrev = 0;
887         opt.diff = 1;
888
889         /* This is what "--pretty" does */
890         opt.verbose_header = 1;
891         opt.use_terminator = 0;
892         opt.commit_format = CMIT_FMT_DEFAULT;
893
894         /* diff-tree init */
895         if (!opt.diffopt.output_format)
896                 opt.diffopt.output_format = DIFF_FORMAT_RAW;
897
898         setup_revisions(0, NULL, &opt, NULL);
899         log_tree_commit(&opt, commit);
900 }
901
902 /*
903  * The terms used for this bisect session are stored in BISECT_TERMS.
904  * We read them and store them to adapt the messages accordingly.
905  * Default is bad/good.
906  */
907 void read_bisect_terms(const char **read_bad, const char **read_good)
908 {
909         struct strbuf str = STRBUF_INIT;
910         const char *filename = git_path("BISECT_TERMS");
911         FILE *fp = fopen(filename, "r");
912
913         if (!fp) {
914                 if (errno == ENOENT) {
915                         *read_bad = "bad";
916                         *read_good = "good";
917                         return;
918                 } else {
919                         die_errno(_("could not read file '%s'"), filename);
920                 }
921         } else {
922                 strbuf_getline_lf(&str, fp);
923                 *read_bad = strbuf_detach(&str, NULL);
924                 strbuf_getline_lf(&str, fp);
925                 *read_good = strbuf_detach(&str, NULL);
926         }
927         strbuf_release(&str);
928         fclose(fp);
929 }
930
931 /*
932  * We use the convention that exiting with an exit code 10 means that
933  * the bisection process finished successfully.
934  * In this case the calling shell script should exit 0.
935  *
936  * If no_checkout is non-zero, the bisection process does not
937  * checkout the trial commit but instead simply updates BISECT_HEAD.
938  */
939 int bisect_next_all(const char *prefix, int no_checkout)
940 {
941         struct rev_info revs;
942         struct commit_list *tried;
943         int reaches = 0, all = 0, nr, steps;
944         const unsigned char *bisect_rev;
945         char *steps_msg;
946
947         read_bisect_terms(&term_bad, &term_good);
948         if (read_bisect_refs())
949                 die(_("reading bisect refs failed"));
950
951         check_good_are_ancestors_of_bad(prefix, no_checkout);
952
953         bisect_rev_setup(&revs, prefix, "%s", "^%s", 1);
954         revs.limited = 1;
955
956         bisect_common(&revs);
957
958         revs.commits = find_bisection(revs.commits, &reaches, &all,
959                                        !!skipped_revs.nr);
960         revs.commits = managed_skipped(revs.commits, &tried);
961
962         if (!revs.commits) {
963                 /*
964                  * We should exit here only if the "bad"
965                  * commit is also a "skip" commit.
966                  */
967                 exit_if_skipped_commits(tried, NULL);
968
969                 printf(_("%s was both %s and %s\n"),
970                        oid_to_hex(current_bad_oid),
971                        term_good,
972                        term_bad);
973                 exit(1);
974         }
975
976         if (!all) {
977                 fprintf(stderr, _("No testable commit found.\n"
978                         "Maybe you started with bad path parameters?\n"));
979                 exit(4);
980         }
981
982         bisect_rev = revs.commits->item->object.oid.hash;
983
984         if (!hashcmp(bisect_rev, current_bad_oid->hash)) {
985                 exit_if_skipped_commits(tried, current_bad_oid);
986                 printf("%s is the first %s commit\n", sha1_to_hex(bisect_rev),
987                         term_bad);
988                 show_diff_tree(prefix, revs.commits->item);
989                 /* This means the bisection process succeeded. */
990                 exit(10);
991         }
992
993         nr = all - reaches - 1;
994         steps = estimate_bisect_steps(all);
995
996         steps_msg = xstrfmt(Q_("(roughly %d step)", "(roughly %d steps)",
997                   steps), steps);
998         /* TRANSLATORS: the last %s will be replaced with
999            "(roughly %d steps)" translation */
1000         printf(Q_("Bisecting: %d revision left to test after this %s\n",
1001                   "Bisecting: %d revisions left to test after this %s\n",
1002                   nr), nr, steps_msg);
1003         free(steps_msg);
1004
1005         return bisect_checkout(bisect_rev, no_checkout);
1006 }
1007
1008 static inline int log2i(int n)
1009 {
1010         int log2 = 0;
1011
1012         for (; n > 1; n >>= 1)
1013                 log2++;
1014
1015         return log2;
1016 }
1017
1018 static inline int exp2i(int n)
1019 {
1020         return 1 << n;
1021 }
1022
1023 /*
1024  * Estimate the number of bisect steps left (after the current step)
1025  *
1026  * For any x between 0 included and 2^n excluded, the probability for
1027  * n - 1 steps left looks like:
1028  *
1029  * P(2^n + x) == (2^n - x) / (2^n + x)
1030  *
1031  * and P(2^n + x) < 0.5 means 2^n < 3x
1032  */
1033 int estimate_bisect_steps(int all)
1034 {
1035         int n, x, e;
1036
1037         if (all < 3)
1038                 return 0;
1039
1040         n = log2i(all);
1041         e = exp2i(n);
1042         x = all - e;
1043
1044         return (e < 3 * x) ? n : n - 1;
1045 }