Merge branch 'ld/git-p4-updates'
[git] / bisect.c
1 #include "cache.h"
2 #include "config.h"
3 #include "commit.h"
4 #include "diff.h"
5 #include "revision.h"
6 #include "refs.h"
7 #include "list-objects.h"
8 #include "quote.h"
9 #include "sha1-lookup.h"
10 #include "run-command.h"
11 #include "log-tree.h"
12 #include "bisect.h"
13 #include "sha1-array.h"
14 #include "argv-array.h"
15
16 static struct oid_array good_revs;
17 static struct oid_array skipped_revs;
18
19 static struct object_id *current_bad_oid;
20
21 static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
22 static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
23
24 static const char *term_bad;
25 static const char *term_good;
26
27 /* Remember to update object flag allocation in object.h */
28 #define COUNTED         (1u<<16)
29
30 /*
31  * This is a truly stupid algorithm, but it's only
32  * used for bisection, and we just don't care enough.
33  *
34  * We care just barely enough to avoid recursing for
35  * non-merge entries.
36  */
37 static int count_distance(struct commit_list *entry)
38 {
39         int nr = 0;
40
41         while (entry) {
42                 struct commit *commit = entry->item;
43                 struct commit_list *p;
44
45                 if (commit->object.flags & (UNINTERESTING | COUNTED))
46                         break;
47                 if (!(commit->object.flags & TREESAME))
48                         nr++;
49                 commit->object.flags |= COUNTED;
50                 p = commit->parents;
51                 entry = p;
52                 if (p) {
53                         p = p->next;
54                         while (p) {
55                                 nr += count_distance(p);
56                                 p = p->next;
57                         }
58                 }
59         }
60
61         return nr;
62 }
63
64 static void clear_distance(struct commit_list *list)
65 {
66         while (list) {
67                 struct commit *commit = list->item;
68                 commit->object.flags &= ~COUNTED;
69                 list = list->next;
70         }
71 }
72
73 #define DEBUG_BISECT 0
74
75 static inline int weight(struct commit_list *elem)
76 {
77         return *((int*)(elem->item->util));
78 }
79
80 static inline void weight_set(struct commit_list *elem, int weight)
81 {
82         *((int*)(elem->item->util)) = weight;
83 }
84
85 static int count_interesting_parents(struct commit *commit)
86 {
87         struct commit_list *p;
88         int count;
89
90         for (count = 0, p = commit->parents; p; p = p->next) {
91                 if (p->item->object.flags & UNINTERESTING)
92                         continue;
93                 count++;
94         }
95         return count;
96 }
97
98 static inline int halfway(struct commit_list *p, int nr)
99 {
100         /*
101          * Don't short-cut something we are not going to return!
102          */
103         if (p->item->object.flags & TREESAME)
104                 return 0;
105         if (DEBUG_BISECT)
106                 return 0;
107         /*
108          * 2 and 3 are halfway of 5.
109          * 3 is halfway of 6 but 2 and 4 are not.
110          */
111         switch (2 * weight(p) - nr) {
112         case -1: case 0: case 1:
113                 return 1;
114         default:
115                 return 0;
116         }
117 }
118
119 #if !DEBUG_BISECT
120 #define show_list(a,b,c,d) do { ; } while (0)
121 #else
122 static void show_list(const char *debug, int counted, int nr,
123                       struct commit_list *list)
124 {
125         struct commit_list *p;
126
127         fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
128
129         for (p = list; p; p = p->next) {
130                 struct commit_list *pp;
131                 struct commit *commit = p->item;
132                 unsigned flags = commit->object.flags;
133                 enum object_type type;
134                 unsigned long size;
135                 char *buf = read_object_file(&commit->object.oid, &type,
136                                              &size);
137                 const char *subject_start;
138                 int subject_len;
139
140                 fprintf(stderr, "%c%c%c ",
141                         (flags & TREESAME) ? ' ' : 'T',
142                         (flags & UNINTERESTING) ? 'U' : ' ',
143                         (flags & COUNTED) ? 'C' : ' ');
144                 if (commit->util)
145                         fprintf(stderr, "%3d", weight(p));
146                 else
147                         fprintf(stderr, "---");
148                 fprintf(stderr, " %.*s", 8, oid_to_hex(&commit->object.oid));
149                 for (pp = commit->parents; pp; pp = pp->next)
150                         fprintf(stderr, " %.*s", 8,
151                                 oid_to_hex(&pp->item->object.oid));
152
153                 subject_len = find_commit_subject(buf, &subject_start);
154                 if (subject_len)
155                         fprintf(stderr, " %.*s", subject_len, subject_start);
156                 fprintf(stderr, "\n");
157         }
158 }
159 #endif /* DEBUG_BISECT */
160
161 static struct commit_list *best_bisection(struct commit_list *list, int nr)
162 {
163         struct commit_list *p, *best;
164         int best_distance = -1;
165
166         best = list;
167         for (p = list; p; p = p->next) {
168                 int distance;
169                 unsigned flags = p->item->object.flags;
170
171                 if (flags & TREESAME)
172                         continue;
173                 distance = weight(p);
174                 if (nr - distance < distance)
175                         distance = nr - distance;
176                 if (distance > best_distance) {
177                         best = p;
178                         best_distance = distance;
179                 }
180         }
181
182         return best;
183 }
184
185 struct commit_dist {
186         struct commit *commit;
187         int distance;
188 };
189
190 static int compare_commit_dist(const void *a_, const void *b_)
191 {
192         struct commit_dist *a, *b;
193
194         a = (struct commit_dist *)a_;
195         b = (struct commit_dist *)b_;
196         if (a->distance != b->distance)
197                 return b->distance - a->distance; /* desc sort */
198         return oidcmp(&a->commit->object.oid, &b->commit->object.oid);
199 }
200
201 static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
202 {
203         struct commit_list *p;
204         struct commit_dist *array = xcalloc(nr, sizeof(*array));
205         struct strbuf buf = STRBUF_INIT;
206         int cnt, i;
207
208         for (p = list, cnt = 0; p; p = p->next) {
209                 int distance;
210                 unsigned flags = p->item->object.flags;
211
212                 if (flags & TREESAME)
213                         continue;
214                 distance = weight(p);
215                 if (nr - distance < distance)
216                         distance = nr - distance;
217                 array[cnt].commit = p->item;
218                 array[cnt].distance = distance;
219                 cnt++;
220         }
221         QSORT(array, cnt, compare_commit_dist);
222         for (p = list, i = 0; i < cnt; i++) {
223                 struct object *obj = &(array[i].commit->object);
224
225                 strbuf_reset(&buf);
226                 strbuf_addf(&buf, "dist=%d", array[i].distance);
227                 add_name_decoration(DECORATION_NONE, buf.buf, obj);
228
229                 p->item = array[i].commit;
230                 if (i < cnt - 1)
231                         p = p->next;
232         }
233         if (p) {
234                 free_commit_list(p->next);
235                 p->next = NULL;
236         }
237         strbuf_release(&buf);
238         free(array);
239         return list;
240 }
241
242 /*
243  * zero or positive weight is the number of interesting commits it can
244  * reach, including itself.  Especially, weight = 0 means it does not
245  * reach any tree-changing commits (e.g. just above uninteresting one
246  * but traversal is with pathspec).
247  *
248  * weight = -1 means it has one parent and its distance is yet to
249  * be computed.
250  *
251  * weight = -2 means it has more than one parent and its distance is
252  * unknown.  After running count_distance() first, they will get zero
253  * or positive distance.
254  */
255 static struct commit_list *do_find_bisection(struct commit_list *list,
256                                              int nr, int *weights,
257                                              int find_all)
258 {
259         int n, counted;
260         struct commit_list *p;
261
262         counted = 0;
263
264         for (n = 0, p = list; p; p = p->next) {
265                 struct commit *commit = p->item;
266                 unsigned flags = commit->object.flags;
267
268                 p->item->util = &weights[n++];
269                 switch (count_interesting_parents(commit)) {
270                 case 0:
271                         if (!(flags & TREESAME)) {
272                                 weight_set(p, 1);
273                                 counted++;
274                                 show_list("bisection 2 count one",
275                                           counted, nr, list);
276                         }
277                         /*
278                          * otherwise, it is known not to reach any
279                          * tree-changing commit and gets weight 0.
280                          */
281                         break;
282                 case 1:
283                         weight_set(p, -1);
284                         break;
285                 default:
286                         weight_set(p, -2);
287                         break;
288                 }
289         }
290
291         show_list("bisection 2 initialize", counted, nr, list);
292
293         /*
294          * If you have only one parent in the resulting set
295          * then you can reach one commit more than that parent
296          * can reach.  So we do not have to run the expensive
297          * count_distance() for single strand of pearls.
298          *
299          * However, if you have more than one parents, you cannot
300          * just add their distance and one for yourself, since
301          * they usually reach the same ancestor and you would
302          * end up counting them twice that way.
303          *
304          * So we will first count distance of merges the usual
305          * way, and then fill the blanks using cheaper algorithm.
306          */
307         for (p = list; p; p = p->next) {
308                 if (p->item->object.flags & UNINTERESTING)
309                         continue;
310                 if (weight(p) != -2)
311                         continue;
312                 weight_set(p, count_distance(p));
313                 clear_distance(list);
314
315                 /* Does it happen to be at exactly half-way? */
316                 if (!find_all && halfway(p, nr))
317                         return p;
318                 counted++;
319         }
320
321         show_list("bisection 2 count_distance", counted, nr, list);
322
323         while (counted < nr) {
324                 for (p = list; p; p = p->next) {
325                         struct commit_list *q;
326                         unsigned flags = p->item->object.flags;
327
328                         if (0 <= weight(p))
329                                 continue;
330                         for (q = p->item->parents; q; q = q->next) {
331                                 if (q->item->object.flags & UNINTERESTING)
332                                         continue;
333                                 if (0 <= weight(q))
334                                         break;
335                         }
336                         if (!q)
337                                 continue;
338
339                         /*
340                          * weight for p is unknown but q is known.
341                          * add one for p itself if p is to be counted,
342                          * otherwise inherit it from q directly.
343                          */
344                         if (!(flags & TREESAME)) {
345                                 weight_set(p, weight(q)+1);
346                                 counted++;
347                                 show_list("bisection 2 count one",
348                                           counted, nr, list);
349                         }
350                         else
351                                 weight_set(p, weight(q));
352
353                         /* Does it happen to be at exactly half-way? */
354                         if (!find_all && halfway(p, nr))
355                                 return p;
356                 }
357         }
358
359         show_list("bisection 2 counted all", counted, nr, list);
360
361         if (!find_all)
362                 return best_bisection(list, nr);
363         else
364                 return best_bisection_sorted(list, nr);
365 }
366
367 void find_bisection(struct commit_list **commit_list, int *reaches,
368                     int *all, int find_all)
369 {
370         int nr, on_list;
371         struct commit_list *list, *p, *best, *next, *last;
372         int *weights;
373
374         show_list("bisection 2 entry", 0, 0, *commit_list);
375
376         /*
377          * Count the number of total and tree-changing items on the
378          * list, while reversing the list.
379          */
380         for (nr = on_list = 0, last = NULL, p = *commit_list;
381              p;
382              p = next) {
383                 unsigned flags = p->item->object.flags;
384
385                 next = p->next;
386                 if (flags & UNINTERESTING) {
387                         free(p);
388                         continue;
389                 }
390                 p->next = last;
391                 last = p;
392                 if (!(flags & TREESAME))
393                         nr++;
394                 on_list++;
395         }
396         list = last;
397         show_list("bisection 2 sorted", 0, nr, list);
398
399         *all = nr;
400         weights = xcalloc(on_list, sizeof(*weights));
401
402         /* Do the real work of finding bisection commit. */
403         best = do_find_bisection(list, nr, weights, find_all);
404         if (best) {
405                 if (!find_all) {
406                         list->item = best->item;
407                         free_commit_list(list->next);
408                         best = list;
409                         best->next = NULL;
410                 }
411                 *reaches = weight(best);
412         }
413         free(weights);
414         *commit_list = best;
415 }
416
417 static int register_ref(const char *refname, const struct object_id *oid,
418                         int flags, void *cb_data)
419 {
420         struct strbuf good_prefix = STRBUF_INIT;
421         strbuf_addstr(&good_prefix, term_good);
422         strbuf_addstr(&good_prefix, "-");
423
424         if (!strcmp(refname, term_bad)) {
425                 current_bad_oid = xmalloc(sizeof(*current_bad_oid));
426                 oidcpy(current_bad_oid, oid);
427         } else if (starts_with(refname, good_prefix.buf)) {
428                 oid_array_append(&good_revs, oid);
429         } else if (starts_with(refname, "skip-")) {
430                 oid_array_append(&skipped_revs, oid);
431         }
432
433         strbuf_release(&good_prefix);
434
435         return 0;
436 }
437
438 static int read_bisect_refs(void)
439 {
440         return for_each_ref_in("refs/bisect/", register_ref, NULL);
441 }
442
443 static GIT_PATH_FUNC(git_path_bisect_names, "BISECT_NAMES")
444 static GIT_PATH_FUNC(git_path_bisect_expected_rev, "BISECT_EXPECTED_REV")
445 static GIT_PATH_FUNC(git_path_bisect_ancestors_ok, "BISECT_ANCESTORS_OK")
446 static GIT_PATH_FUNC(git_path_bisect_run, "BISECT_RUN")
447 static GIT_PATH_FUNC(git_path_bisect_start, "BISECT_START")
448 static GIT_PATH_FUNC(git_path_bisect_log, "BISECT_LOG")
449 static GIT_PATH_FUNC(git_path_bisect_terms, "BISECT_TERMS")
450 static GIT_PATH_FUNC(git_path_head_name, "head-name")
451
452 static void read_bisect_paths(struct argv_array *array)
453 {
454         struct strbuf str = STRBUF_INIT;
455         const char *filename = git_path_bisect_names();
456         FILE *fp = xfopen(filename, "r");
457
458         while (strbuf_getline_lf(&str, fp) != EOF) {
459                 strbuf_trim(&str);
460                 if (sq_dequote_to_argv_array(str.buf, array))
461                         die(_("Badly quoted content in file '%s': %s"),
462                             filename, str.buf);
463         }
464
465         strbuf_release(&str);
466         fclose(fp);
467 }
468
469 static char *join_sha1_array_hex(struct oid_array *array, char delim)
470 {
471         struct strbuf joined_hexs = STRBUF_INIT;
472         int i;
473
474         for (i = 0; i < array->nr; i++) {
475                 strbuf_addstr(&joined_hexs, oid_to_hex(array->oid + i));
476                 if (i + 1 < array->nr)
477                         strbuf_addch(&joined_hexs, delim);
478         }
479
480         return strbuf_detach(&joined_hexs, NULL);
481 }
482
483 /*
484  * In this function, passing a not NULL skipped_first is very special.
485  * It means that we want to know if the first commit in the list is
486  * skipped because we will want to test a commit away from it if it is
487  * indeed skipped.
488  * So if the first commit is skipped, we cannot take the shortcut to
489  * just "return list" when we find the first non skipped commit, we
490  * have to return a fully filtered list.
491  *
492  * We use (*skipped_first == -1) to mean "it has been found that the
493  * first commit is not skipped". In this case *skipped_first is set back
494  * to 0 just before the function returns.
495  */
496 struct commit_list *filter_skipped(struct commit_list *list,
497                                    struct commit_list **tried,
498                                    int show_all,
499                                    int *count,
500                                    int *skipped_first)
501 {
502         struct commit_list *filtered = NULL, **f = &filtered;
503
504         *tried = NULL;
505
506         if (skipped_first)
507                 *skipped_first = 0;
508         if (count)
509                 *count = 0;
510
511         if (!skipped_revs.nr)
512                 return list;
513
514         while (list) {
515                 struct commit_list *next = list->next;
516                 list->next = NULL;
517                 if (0 <= oid_array_lookup(&skipped_revs, &list->item->object.oid)) {
518                         if (skipped_first && !*skipped_first)
519                                 *skipped_first = 1;
520                         /* Move current to tried list */
521                         *tried = list;
522                         tried = &list->next;
523                 } else {
524                         if (!show_all) {
525                                 if (!skipped_first || !*skipped_first)
526                                         return list;
527                         } else if (skipped_first && !*skipped_first) {
528                                 /* This means we know it's not skipped */
529                                 *skipped_first = -1;
530                         }
531                         /* Move current to filtered list */
532                         *f = list;
533                         f = &list->next;
534                         if (count)
535                                 (*count)++;
536                 }
537                 list = next;
538         }
539
540         if (skipped_first && *skipped_first == -1)
541                 *skipped_first = 0;
542
543         return filtered;
544 }
545
546 #define PRN_MODULO 32768
547
548 /*
549  * This is a pseudo random number generator based on "man 3 rand".
550  * It is not used properly because the seed is the argument and it
551  * is increased by one between each call, but that should not matter
552  * for this application.
553  */
554 static unsigned get_prn(unsigned count) {
555         count = count * 1103515245 + 12345;
556         return (count/65536) % PRN_MODULO;
557 }
558
559 /*
560  * Custom integer square root from
561  * https://en.wikipedia.org/wiki/Integer_square_root
562  */
563 static int sqrti(int val)
564 {
565         float d, x = val;
566
567         if (val == 0)
568                 return 0;
569
570         do {
571                 float y = (x + (float)val / x) / 2;
572                 d = (y > x) ? y - x : x - y;
573                 x = y;
574         } while (d >= 0.5);
575
576         return (int)x;
577 }
578
579 static struct commit_list *skip_away(struct commit_list *list, int count)
580 {
581         struct commit_list *cur, *previous;
582         int prn, index, i;
583
584         prn = get_prn(count);
585         index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
586
587         cur = list;
588         previous = NULL;
589
590         for (i = 0; cur; cur = cur->next, i++) {
591                 if (i == index) {
592                         if (oidcmp(&cur->item->object.oid, current_bad_oid))
593                                 return cur;
594                         if (previous)
595                                 return previous;
596                         return list;
597                 }
598                 previous = cur;
599         }
600
601         return list;
602 }
603
604 static struct commit_list *managed_skipped(struct commit_list *list,
605                                            struct commit_list **tried)
606 {
607         int count, skipped_first;
608
609         *tried = NULL;
610
611         if (!skipped_revs.nr)
612                 return list;
613
614         list = filter_skipped(list, tried, 0, &count, &skipped_first);
615
616         if (!skipped_first)
617                 return list;
618
619         return skip_away(list, count);
620 }
621
622 static void bisect_rev_setup(struct rev_info *revs, const char *prefix,
623                              const char *bad_format, const char *good_format,
624                              int read_paths)
625 {
626         struct argv_array rev_argv = ARGV_ARRAY_INIT;
627         int i;
628
629         init_revisions(revs, prefix);
630         revs->abbrev = 0;
631         revs->commit_format = CMIT_FMT_UNSPECIFIED;
632
633         /* rev_argv.argv[0] will be ignored by setup_revisions */
634         argv_array_push(&rev_argv, "bisect_rev_setup");
635         argv_array_pushf(&rev_argv, bad_format, oid_to_hex(current_bad_oid));
636         for (i = 0; i < good_revs.nr; i++)
637                 argv_array_pushf(&rev_argv, good_format,
638                                  oid_to_hex(good_revs.oid + i));
639         argv_array_push(&rev_argv, "--");
640         if (read_paths)
641                 read_bisect_paths(&rev_argv);
642
643         setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
644         /* XXX leak rev_argv, as "revs" may still be pointing to it */
645 }
646
647 static void bisect_common(struct rev_info *revs)
648 {
649         if (prepare_revision_walk(revs))
650                 die("revision walk setup failed");
651         if (revs->tree_objects)
652                 mark_edges_uninteresting(revs, NULL);
653 }
654
655 static void exit_if_skipped_commits(struct commit_list *tried,
656                                     const struct object_id *bad)
657 {
658         if (!tried)
659                 return;
660
661         printf("There are only 'skip'ped commits left to test.\n"
662                "The first %s commit could be any of:\n", term_bad);
663
664         for ( ; tried; tried = tried->next)
665                 printf("%s\n", oid_to_hex(&tried->item->object.oid));
666
667         if (bad)
668                 printf("%s\n", oid_to_hex(bad));
669         printf(_("We cannot bisect more!\n"));
670         exit(2);
671 }
672
673 static int is_expected_rev(const struct object_id *oid)
674 {
675         const char *filename = git_path_bisect_expected_rev();
676         struct stat st;
677         struct strbuf str = STRBUF_INIT;
678         FILE *fp;
679         int res = 0;
680
681         if (stat(filename, &st) || !S_ISREG(st.st_mode))
682                 return 0;
683
684         fp = fopen_or_warn(filename, "r");
685         if (!fp)
686                 return 0;
687
688         if (strbuf_getline_lf(&str, fp) != EOF)
689                 res = !strcmp(str.buf, oid_to_hex(oid));
690
691         strbuf_release(&str);
692         fclose(fp);
693
694         return res;
695 }
696
697 static int bisect_checkout(const struct object_id *bisect_rev, int no_checkout)
698 {
699         char bisect_rev_hex[GIT_MAX_HEXSZ + 1];
700
701         memcpy(bisect_rev_hex, oid_to_hex(bisect_rev), GIT_SHA1_HEXSZ + 1);
702         update_ref(NULL, "BISECT_EXPECTED_REV", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
703
704         argv_checkout[2] = bisect_rev_hex;
705         if (no_checkout) {
706                 update_ref(NULL, "BISECT_HEAD", bisect_rev, NULL, 0,
707                            UPDATE_REFS_DIE_ON_ERR);
708         } else {
709                 int res;
710                 res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
711                 if (res)
712                         exit(res);
713         }
714
715         argv_show_branch[1] = bisect_rev_hex;
716         return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
717 }
718
719 static struct commit *get_commit_reference(const struct object_id *oid)
720 {
721         struct commit *r = lookup_commit_reference(oid);
722         if (!r)
723                 die(_("Not a valid commit name %s"), oid_to_hex(oid));
724         return r;
725 }
726
727 static struct commit **get_bad_and_good_commits(int *rev_nr)
728 {
729         struct commit **rev;
730         int i, n = 0;
731
732         ALLOC_ARRAY(rev, 1 + good_revs.nr);
733         rev[n++] = get_commit_reference(current_bad_oid);
734         for (i = 0; i < good_revs.nr; i++)
735                 rev[n++] = get_commit_reference(good_revs.oid + i);
736         *rev_nr = n;
737
738         return rev;
739 }
740
741 static void handle_bad_merge_base(void)
742 {
743         if (is_expected_rev(current_bad_oid)) {
744                 char *bad_hex = oid_to_hex(current_bad_oid);
745                 char *good_hex = join_sha1_array_hex(&good_revs, ' ');
746                 if (!strcmp(term_bad, "bad") && !strcmp(term_good, "good")) {
747                         fprintf(stderr, _("The merge base %s is bad.\n"
748                                 "This means the bug has been fixed "
749                                 "between %s and [%s].\n"),
750                                 bad_hex, bad_hex, good_hex);
751                 } else if (!strcmp(term_bad, "new") && !strcmp(term_good, "old")) {
752                         fprintf(stderr, _("The merge base %s is new.\n"
753                                 "The property has changed "
754                                 "between %s and [%s].\n"),
755                                 bad_hex, bad_hex, good_hex);
756                 } else {
757                         fprintf(stderr, _("The merge base %s is %s.\n"
758                                 "This means the first '%s' commit is "
759                                 "between %s and [%s].\n"),
760                                 bad_hex, term_bad, term_good, bad_hex, good_hex);
761                 }
762                 exit(3);
763         }
764
765         fprintf(stderr, _("Some %s revs are not ancestors of the %s rev.\n"
766                 "git bisect cannot work properly in this case.\n"
767                 "Maybe you mistook %s and %s revs?\n"),
768                 term_good, term_bad, term_good, term_bad);
769         exit(1);
770 }
771
772 static void handle_skipped_merge_base(const struct object_id *mb)
773 {
774         char *mb_hex = oid_to_hex(mb);
775         char *bad_hex = oid_to_hex(current_bad_oid);
776         char *good_hex = join_sha1_array_hex(&good_revs, ' ');
777
778         warning(_("the merge base between %s and [%s] "
779                 "must be skipped.\n"
780                 "So we cannot be sure the first %s commit is "
781                 "between %s and %s.\n"
782                 "We continue anyway."),
783                 bad_hex, good_hex, term_bad, mb_hex, bad_hex);
784         free(good_hex);
785 }
786
787 /*
788  * "check_merge_bases" checks that merge bases are not "bad" (or "new").
789  *
790  * - If one is "bad" (or "new"), it means the user assumed something wrong
791  * and we must exit with a non 0 error code.
792  * - If one is "good" (or "old"), that's good, we have nothing to do.
793  * - If one is "skipped", we can't know but we should warn.
794  * - If we don't know, we should check it out and ask the user to test.
795  */
796 static void check_merge_bases(int rev_nr, struct commit **rev, int no_checkout)
797 {
798         struct commit_list *result;
799
800         result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1);
801
802         for (; result; result = result->next) {
803                 const struct object_id *mb = &result->item->object.oid;
804                 if (!oidcmp(mb, current_bad_oid)) {
805                         handle_bad_merge_base();
806                 } else if (0 <= oid_array_lookup(&good_revs, mb)) {
807                         continue;
808                 } else if (0 <= oid_array_lookup(&skipped_revs, mb)) {
809                         handle_skipped_merge_base(mb);
810                 } else {
811                         printf(_("Bisecting: a merge base must be tested\n"));
812                         exit(bisect_checkout(mb, no_checkout));
813                 }
814         }
815
816         free_commit_list(result);
817 }
818
819 static int check_ancestors(int rev_nr, struct commit **rev, const char *prefix)
820 {
821         struct rev_info revs;
822         int res;
823
824         bisect_rev_setup(&revs, prefix, "^%s", "%s", 0);
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_many(rev_nr, rev, ALL_REV_FLAGS);
831
832         return res;
833 }
834
835 /*
836  * "check_good_are_ancestors_of_bad" checks that all "good" revs are
837  * ancestor of the "bad" rev.
838  *
839  * If that's not the case, we need to check the merge bases.
840  * If a merge base must be tested by the user, its source code will be
841  * checked out to be tested by the user and we will exit.
842  */
843 static void check_good_are_ancestors_of_bad(const char *prefix, int no_checkout)
844 {
845         char *filename = git_pathdup("BISECT_ANCESTORS_OK");
846         struct stat st;
847         int fd, rev_nr;
848         struct commit **rev;
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         rev = get_bad_and_good_commits(&rev_nr);
863         if (check_ancestors(rev_nr, rev, prefix))
864                 check_merge_bases(rev_nr, rev, no_checkout);
865         free(rev);
866
867         /* Create file BISECT_ANCESTORS_OK. */
868         fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
869         if (fd < 0)
870                 warning_errno(_("could not create file '%s'"),
871                               filename);
872         else
873                 close(fd);
874  done:
875         free(filename);
876 }
877
878 /*
879  * This does "git diff-tree --pretty COMMIT" without one fork+exec.
880  */
881 static void show_diff_tree(const char *prefix, struct commit *commit)
882 {
883         struct rev_info opt;
884
885         /* diff-tree init */
886         init_revisions(&opt, prefix);
887         git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
888         opt.abbrev = 0;
889         opt.diff = 1;
890
891         /* This is what "--pretty" does */
892         opt.verbose_header = 1;
893         opt.use_terminator = 0;
894         opt.commit_format = CMIT_FMT_DEFAULT;
895
896         /* diff-tree init */
897         if (!opt.diffopt.output_format)
898                 opt.diffopt.output_format = DIFF_FORMAT_RAW;
899
900         setup_revisions(0, NULL, &opt, NULL);
901         log_tree_commit(&opt, commit);
902 }
903
904 /*
905  * The terms used for this bisect session are stored in BISECT_TERMS.
906  * We read them and store them to adapt the messages accordingly.
907  * Default is bad/good.
908  */
909 void read_bisect_terms(const char **read_bad, const char **read_good)
910 {
911         struct strbuf str = STRBUF_INIT;
912         const char *filename = git_path_bisect_terms();
913         FILE *fp = fopen(filename, "r");
914
915         if (!fp) {
916                 if (errno == ENOENT) {
917                         *read_bad = "bad";
918                         *read_good = "good";
919                         return;
920                 } else {
921                         die_errno(_("could not read file '%s'"), filename);
922                 }
923         } else {
924                 strbuf_getline_lf(&str, fp);
925                 *read_bad = strbuf_detach(&str, NULL);
926                 strbuf_getline_lf(&str, fp);
927                 *read_good = strbuf_detach(&str, NULL);
928         }
929         strbuf_release(&str);
930         fclose(fp);
931 }
932
933 /*
934  * We use the convention that exiting with an exit code 10 means that
935  * the bisection process finished successfully.
936  * In this case the calling shell script should exit 0.
937  *
938  * If no_checkout is non-zero, the bisection process does not
939  * checkout the trial commit but instead simply updates BISECT_HEAD.
940  */
941 int bisect_next_all(const char *prefix, int no_checkout)
942 {
943         struct rev_info revs;
944         struct commit_list *tried;
945         int reaches = 0, all = 0, nr, steps;
946         struct object_id *bisect_rev;
947         char *steps_msg;
948
949         read_bisect_terms(&term_bad, &term_good);
950         if (read_bisect_refs())
951                 die(_("reading bisect refs failed"));
952
953         check_good_are_ancestors_of_bad(prefix, no_checkout);
954
955         bisect_rev_setup(&revs, prefix, "%s", "^%s", 1);
956         revs.limited = 1;
957
958         bisect_common(&revs);
959
960         find_bisection(&revs.commits, &reaches, &all, !!skipped_revs.nr);
961         revs.commits = managed_skipped(revs.commits, &tried);
962
963         if (!revs.commits) {
964                 /*
965                  * We should exit here only if the "bad"
966                  * commit is also a "skip" commit.
967                  */
968                 exit_if_skipped_commits(tried, NULL);
969
970                 printf(_("%s was both %s and %s\n"),
971                        oid_to_hex(current_bad_oid),
972                        term_good,
973                        term_bad);
974                 exit(1);
975         }
976
977         if (!all) {
978                 fprintf(stderr, _("No testable commit found.\n"
979                         "Maybe you started with bad path parameters?\n"));
980                 exit(4);
981         }
982
983         bisect_rev = &revs.commits->item->object.oid;
984
985         if (!oidcmp(bisect_rev, current_bad_oid)) {
986                 exit_if_skipped_commits(tried, current_bad_oid);
987                 printf("%s is the first %s commit\n", oid_to_hex(bisect_rev),
988                         term_bad);
989                 show_diff_tree(prefix, revs.commits->item);
990                 /* This means the bisection process succeeded. */
991                 exit(10);
992         }
993
994         nr = all - reaches - 1;
995         steps = estimate_bisect_steps(all);
996
997         steps_msg = xstrfmt(Q_("(roughly %d step)", "(roughly %d steps)",
998                   steps), steps);
999         /*
1000          * TRANSLATORS: the last %s will be replaced with "(roughly %d
1001          * steps)" translation.
1002          */
1003         printf(Q_("Bisecting: %d revision left to test after this %s\n",
1004                   "Bisecting: %d revisions left to test after this %s\n",
1005                   nr), nr, steps_msg);
1006         free(steps_msg);
1007
1008         return bisect_checkout(bisect_rev, no_checkout);
1009 }
1010
1011 static inline int log2i(int n)
1012 {
1013         int log2 = 0;
1014
1015         for (; n > 1; n >>= 1)
1016                 log2++;
1017
1018         return log2;
1019 }
1020
1021 static inline int exp2i(int n)
1022 {
1023         return 1 << n;
1024 }
1025
1026 /*
1027  * Estimate the number of bisect steps left (after the current step)
1028  *
1029  * For any x between 0 included and 2^n excluded, the probability for
1030  * n - 1 steps left looks like:
1031  *
1032  * P(2^n + x) == (2^n - x) / (2^n + x)
1033  *
1034  * and P(2^n + x) < 0.5 means 2^n < 3x
1035  */
1036 int estimate_bisect_steps(int all)
1037 {
1038         int n, x, e;
1039
1040         if (all < 3)
1041                 return 0;
1042
1043         n = log2i(all);
1044         e = exp2i(n);
1045         x = all - e;
1046
1047         return (e < 3 * x) ? n : n - 1;
1048 }
1049
1050 static int mark_for_removal(const char *refname, const struct object_id *oid,
1051                             int flag, void *cb_data)
1052 {
1053         struct string_list *refs = cb_data;
1054         char *ref = xstrfmt("refs/bisect%s", refname);
1055         string_list_append(refs, ref);
1056         return 0;
1057 }
1058
1059 int bisect_clean_state(void)
1060 {
1061         int result = 0;
1062
1063         /* There may be some refs packed during bisection */
1064         struct string_list refs_for_removal = STRING_LIST_INIT_NODUP;
1065         for_each_ref_in("refs/bisect", mark_for_removal, (void *) &refs_for_removal);
1066         string_list_append(&refs_for_removal, xstrdup("BISECT_HEAD"));
1067         result = delete_refs("bisect: remove", &refs_for_removal, REF_NO_DEREF);
1068         refs_for_removal.strdup_strings = 1;
1069         string_list_clear(&refs_for_removal, 0);
1070         unlink_or_warn(git_path_bisect_expected_rev());
1071         unlink_or_warn(git_path_bisect_ancestors_ok());
1072         unlink_or_warn(git_path_bisect_log());
1073         unlink_or_warn(git_path_bisect_names());
1074         unlink_or_warn(git_path_bisect_run());
1075         unlink_or_warn(git_path_bisect_terms());
1076         /* Cleanup head-name if it got left by an old version of git-bisect */
1077         unlink_or_warn(git_path_head_name());
1078         /*
1079          * Cleanup BISECT_START last to support the --no-checkout option
1080          * introduced in the commit 4796e823a.
1081          */
1082         unlink_or_warn(git_path_bisect_start());
1083
1084         return result;
1085 }