bisect.con commit Merge branch 'jk/long-error-messages' (1302c9f)
   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
  15static struct sha1_array good_revs;
  16static struct sha1_array skipped_revs;
  17
  18static struct object_id *current_bad_oid;
  19
  20static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
  21static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
  22static const char *argv_update_ref[] = {"update-ref", "--no-deref", "BISECT_HEAD", NULL, NULL};
  23
  24static const char *term_bad;
  25static 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 */
  37static 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
  64static 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
  75static inline int weight(struct commit_list *elem)
  76{
  77        return *((int*)(elem->item->util));
  78}
  79
  80static inline void weight_set(struct commit_list *elem, int weight)
  81{
  82        *((int*)(elem->item->util)) = weight;
  83}
  84
  85static 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
  98static 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
 122static 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_sha1_file(commit->object.sha1, &type, &size);
 136                const char *subject_start;
 137                int subject_len;
 138
 139                fprintf(stderr, "%c%c%c ",
 140                        (flags & TREESAME) ? ' ' : 'T',
 141                        (flags & UNINTERESTING) ? 'U' : ' ',
 142                        (flags & COUNTED) ? 'C' : ' ');
 143                if (commit->util)
 144                        fprintf(stderr, "%3d", weight(p));
 145                else
 146                        fprintf(stderr, "---");
 147                fprintf(stderr, " %.*s", 8, sha1_to_hex(commit->object.sha1));
 148                for (pp = commit->parents; pp; pp = pp->next)
 149                        fprintf(stderr, " %.*s", 8,
 150                                sha1_to_hex(pp->item->object.sha1));
 151
 152                subject_len = find_commit_subject(buf, &subject_start);
 153                if (subject_len)
 154                        fprintf(stderr, " %.*s", subject_len, subject_start);
 155                fprintf(stderr, "\n");
 156        }
 157}
 158#endif /* DEBUG_BISECT */
 159
 160static struct commit_list *best_bisection(struct commit_list *list, int nr)
 161{
 162        struct commit_list *p, *best;
 163        int best_distance = -1;
 164
 165        best = list;
 166        for (p = list; p; p = p->next) {
 167                int distance;
 168                unsigned flags = p->item->object.flags;
 169
 170                if (flags & TREESAME)
 171                        continue;
 172                distance = weight(p);
 173                if (nr - distance < distance)
 174                        distance = nr - distance;
 175                if (distance > best_distance) {
 176                        best = p;
 177                        best_distance = distance;
 178                }
 179        }
 180
 181        return best;
 182}
 183
 184struct commit_dist {
 185        struct commit *commit;
 186        int distance;
 187};
 188
 189static int compare_commit_dist(const void *a_, const void *b_)
 190{
 191        struct commit_dist *a, *b;
 192
 193        a = (struct commit_dist *)a_;
 194        b = (struct commit_dist *)b_;
 195        if (a->distance != b->distance)
 196                return b->distance - a->distance; /* desc sort */
 197        return hashcmp(a->commit->object.sha1, b->commit->object.sha1);
 198}
 199
 200static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
 201{
 202        struct commit_list *p;
 203        struct commit_dist *array = xcalloc(nr, sizeof(*array));
 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, sizeof(*array), compare_commit_dist);
 220        for (p = list, i = 0; i < cnt; i++) {
 221                char buf[100]; /* enough for dist=%d */
 222                struct object *obj = &(array[i].commit->object);
 223
 224                snprintf(buf, sizeof(buf), "dist=%d", array[i].distance);
 225                add_name_decoration(DECORATION_NONE, buf, obj);
 226
 227                p->item = array[i].commit;
 228                p = p->next;
 229        }
 230        if (p)
 231                p->next = NULL;
 232        free(array);
 233        return list;
 234}
 235
 236/*
 237 * zero or positive weight is the number of interesting commits it can
 238 * reach, including itself.  Especially, weight = 0 means it does not
 239 * reach any tree-changing commits (e.g. just above uninteresting one
 240 * but traversal is with pathspec).
 241 *
 242 * weight = -1 means it has one parent and its distance is yet to
 243 * be computed.
 244 *
 245 * weight = -2 means it has more than one parent and its distance is
 246 * unknown.  After running count_distance() first, they will get zero
 247 * or positive distance.
 248 */
 249static struct commit_list *do_find_bisection(struct commit_list *list,
 250                                             int nr, int *weights,
 251                                             int find_all)
 252{
 253        int n, counted;
 254        struct commit_list *p;
 255
 256        counted = 0;
 257
 258        for (n = 0, p = list; p; p = p->next) {
 259                struct commit *commit = p->item;
 260                unsigned flags = commit->object.flags;
 261
 262                p->item->util = &weights[n++];
 263                switch (count_interesting_parents(commit)) {
 264                case 0:
 265                        if (!(flags & TREESAME)) {
 266                                weight_set(p, 1);
 267                                counted++;
 268                                show_list("bisection 2 count one",
 269                                          counted, nr, list);
 270                        }
 271                        /*
 272                         * otherwise, it is known not to reach any
 273                         * tree-changing commit and gets weight 0.
 274                         */
 275                        break;
 276                case 1:
 277                        weight_set(p, -1);
 278                        break;
 279                default:
 280                        weight_set(p, -2);
 281                        break;
 282                }
 283        }
 284
 285        show_list("bisection 2 initialize", counted, nr, list);
 286
 287        /*
 288         * If you have only one parent in the resulting set
 289         * then you can reach one commit more than that parent
 290         * can reach.  So we do not have to run the expensive
 291         * count_distance() for single strand of pearls.
 292         *
 293         * However, if you have more than one parents, you cannot
 294         * just add their distance and one for yourself, since
 295         * they usually reach the same ancestor and you would
 296         * end up counting them twice that way.
 297         *
 298         * So we will first count distance of merges the usual
 299         * way, and then fill the blanks using cheaper algorithm.
 300         */
 301        for (p = list; p; p = p->next) {
 302                if (p->item->object.flags & UNINTERESTING)
 303                        continue;
 304                if (weight(p) != -2)
 305                        continue;
 306                weight_set(p, count_distance(p));
 307                clear_distance(list);
 308
 309                /* Does it happen to be at exactly half-way? */
 310                if (!find_all && halfway(p, nr))
 311                        return p;
 312                counted++;
 313        }
 314
 315        show_list("bisection 2 count_distance", counted, nr, list);
 316
 317        while (counted < nr) {
 318                for (p = list; p; p = p->next) {
 319                        struct commit_list *q;
 320                        unsigned flags = p->item->object.flags;
 321
 322                        if (0 <= weight(p))
 323                                continue;
 324                        for (q = p->item->parents; q; q = q->next) {
 325                                if (q->item->object.flags & UNINTERESTING)
 326                                        continue;
 327                                if (0 <= weight(q))
 328                                        break;
 329                        }
 330                        if (!q)
 331                                continue;
 332
 333                        /*
 334                         * weight for p is unknown but q is known.
 335                         * add one for p itself if p is to be counted,
 336                         * otherwise inherit it from q directly.
 337                         */
 338                        if (!(flags & TREESAME)) {
 339                                weight_set(p, weight(q)+1);
 340                                counted++;
 341                                show_list("bisection 2 count one",
 342                                          counted, nr, list);
 343                        }
 344                        else
 345                                weight_set(p, weight(q));
 346
 347                        /* Does it happen to be at exactly half-way? */
 348                        if (!find_all && halfway(p, nr))
 349                                return p;
 350                }
 351        }
 352
 353        show_list("bisection 2 counted all", counted, nr, list);
 354
 355        if (!find_all)
 356                return best_bisection(list, nr);
 357        else
 358                return best_bisection_sorted(list, nr);
 359}
 360
 361struct commit_list *find_bisection(struct commit_list *list,
 362                                          int *reaches, int *all,
 363                                          int find_all)
 364{
 365        int nr, on_list;
 366        struct commit_list *p, *best, *next, *last;
 367        int *weights;
 368
 369        show_list("bisection 2 entry", 0, 0, list);
 370
 371        /*
 372         * Count the number of total and tree-changing items on the
 373         * list, while reversing the list.
 374         */
 375        for (nr = on_list = 0, last = NULL, p = list;
 376             p;
 377             p = next) {
 378                unsigned flags = p->item->object.flags;
 379
 380                next = p->next;
 381                if (flags & UNINTERESTING)
 382                        continue;
 383                p->next = last;
 384                last = p;
 385                if (!(flags & TREESAME))
 386                        nr++;
 387                on_list++;
 388        }
 389        list = last;
 390        show_list("bisection 2 sorted", 0, nr, list);
 391
 392        *all = nr;
 393        weights = xcalloc(on_list, sizeof(*weights));
 394
 395        /* Do the real work of finding bisection commit. */
 396        best = do_find_bisection(list, nr, weights, find_all);
 397        if (best) {
 398                if (!find_all)
 399                        best->next = NULL;
 400                *reaches = weight(best);
 401        }
 402        free(weights);
 403        return best;
 404}
 405
 406static int register_ref(const char *refname, const struct object_id *oid,
 407                        int flags, void *cb_data)
 408{
 409        struct strbuf good_prefix = STRBUF_INIT;
 410        strbuf_addstr(&good_prefix, term_good);
 411        strbuf_addstr(&good_prefix, "-");
 412
 413        if (!strcmp(refname, term_bad)) {
 414                current_bad_oid = xmalloc(sizeof(*current_bad_oid));
 415                oidcpy(current_bad_oid, oid);
 416        } else if (starts_with(refname, good_prefix.buf)) {
 417                sha1_array_append(&good_revs, oid->hash);
 418        } else if (starts_with(refname, "skip-")) {
 419                sha1_array_append(&skipped_revs, oid->hash);
 420        }
 421
 422        strbuf_release(&good_prefix);
 423
 424        return 0;
 425}
 426
 427static int read_bisect_refs(void)
 428{
 429        return for_each_ref_in("refs/bisect/", register_ref, NULL);
 430}
 431
 432static GIT_PATH_FUNC(git_path_bisect_names, "BISECT_NAMES")
 433static GIT_PATH_FUNC(git_path_bisect_expected_rev, "BISECT_EXPECTED_REV")
 434
 435static void read_bisect_paths(struct argv_array *array)
 436{
 437        struct strbuf str = STRBUF_INIT;
 438        const char *filename = git_path_bisect_names();
 439        FILE *fp = fopen(filename, "r");
 440
 441        if (!fp)
 442                die_errno("Could not open file '%s'", filename);
 443
 444        while (strbuf_getline(&str, fp, '\n') != EOF) {
 445                strbuf_trim(&str);
 446                if (sq_dequote_to_argv_array(str.buf, array))
 447                        die("Badly quoted content in file '%s': %s",
 448                            filename, str.buf);
 449        }
 450
 451        strbuf_release(&str);
 452        fclose(fp);
 453}
 454
 455static char *join_sha1_array_hex(struct sha1_array *array, char delim)
 456{
 457        struct strbuf joined_hexs = STRBUF_INIT;
 458        int i;
 459
 460        for (i = 0; i < array->nr; i++) {
 461                strbuf_addstr(&joined_hexs, sha1_to_hex(array->sha1[i]));
 462                if (i + 1 < array->nr)
 463                        strbuf_addch(&joined_hexs, delim);
 464        }
 465
 466        return strbuf_detach(&joined_hexs, NULL);
 467}
 468
 469/*
 470 * In this function, passing a not NULL skipped_first is very special.
 471 * It means that we want to know if the first commit in the list is
 472 * skipped because we will want to test a commit away from it if it is
 473 * indeed skipped.
 474 * So if the first commit is skipped, we cannot take the shortcut to
 475 * just "return list" when we find the first non skipped commit, we
 476 * have to return a fully filtered list.
 477 *
 478 * We use (*skipped_first == -1) to mean "it has been found that the
 479 * first commit is not skipped". In this case *skipped_first is set back
 480 * to 0 just before the function returns.
 481 */
 482struct commit_list *filter_skipped(struct commit_list *list,
 483                                   struct commit_list **tried,
 484                                   int show_all,
 485                                   int *count,
 486                                   int *skipped_first)
 487{
 488        struct commit_list *filtered = NULL, **f = &filtered;
 489
 490        *tried = NULL;
 491
 492        if (skipped_first)
 493                *skipped_first = 0;
 494        if (count)
 495                *count = 0;
 496
 497        if (!skipped_revs.nr)
 498                return list;
 499
 500        while (list) {
 501                struct commit_list *next = list->next;
 502                list->next = NULL;
 503                if (0 <= sha1_array_lookup(&skipped_revs,
 504                                           list->item->object.sha1)) {
 505                        if (skipped_first && !*skipped_first)
 506                                *skipped_first = 1;
 507                        /* Move current to tried list */
 508                        *tried = list;
 509                        tried = &list->next;
 510                } else {
 511                        if (!show_all) {
 512                                if (!skipped_first || !*skipped_first)
 513                                        return list;
 514                        } else if (skipped_first && !*skipped_first) {
 515                                /* This means we know it's not skipped */
 516                                *skipped_first = -1;
 517                        }
 518                        /* Move current to filtered list */
 519                        *f = list;
 520                        f = &list->next;
 521                        if (count)
 522                                (*count)++;
 523                }
 524                list = next;
 525        }
 526
 527        if (skipped_first && *skipped_first == -1)
 528                *skipped_first = 0;
 529
 530        return filtered;
 531}
 532
 533#define PRN_MODULO 32768
 534
 535/*
 536 * This is a pseudo random number generator based on "man 3 rand".
 537 * It is not used properly because the seed is the argument and it
 538 * is increased by one between each call, but that should not matter
 539 * for this application.
 540 */
 541static unsigned get_prn(unsigned count) {
 542        count = count * 1103515245 + 12345;
 543        return (count/65536) % PRN_MODULO;
 544}
 545
 546/*
 547 * Custom integer square root from
 548 * http://en.wikipedia.org/wiki/Integer_square_root
 549 */
 550static int sqrti(int val)
 551{
 552        float d, x = val;
 553
 554        if (val == 0)
 555                return 0;
 556
 557        do {
 558                float y = (x + (float)val / x) / 2;
 559                d = (y > x) ? y - x : x - y;
 560                x = y;
 561        } while (d >= 0.5);
 562
 563        return (int)x;
 564}
 565
 566static struct commit_list *skip_away(struct commit_list *list, int count)
 567{
 568        struct commit_list *cur, *previous;
 569        int prn, index, i;
 570
 571        prn = get_prn(count);
 572        index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
 573
 574        cur = list;
 575        previous = NULL;
 576
 577        for (i = 0; cur; cur = cur->next, i++) {
 578                if (i == index) {
 579                        if (hashcmp(cur->item->object.sha1, current_bad_oid->hash))
 580                                return cur;
 581                        if (previous)
 582                                return previous;
 583                        return list;
 584                }
 585                previous = cur;
 586        }
 587
 588        return list;
 589}
 590
 591static struct commit_list *managed_skipped(struct commit_list *list,
 592                                           struct commit_list **tried)
 593{
 594        int count, skipped_first;
 595
 596        *tried = NULL;
 597
 598        if (!skipped_revs.nr)
 599                return list;
 600
 601        list = filter_skipped(list, tried, 0, &count, &skipped_first);
 602
 603        if (!skipped_first)
 604                return list;
 605
 606        return skip_away(list, count);
 607}
 608
 609static void bisect_rev_setup(struct rev_info *revs, const char *prefix,
 610                             const char *bad_format, const char *good_format,
 611                             int read_paths)
 612{
 613        struct argv_array rev_argv = ARGV_ARRAY_INIT;
 614        int i;
 615
 616        init_revisions(revs, prefix);
 617        revs->abbrev = 0;
 618        revs->commit_format = CMIT_FMT_UNSPECIFIED;
 619
 620        /* rev_argv.argv[0] will be ignored by setup_revisions */
 621        argv_array_push(&rev_argv, "bisect_rev_setup");
 622        argv_array_pushf(&rev_argv, bad_format, oid_to_hex(current_bad_oid));
 623        for (i = 0; i < good_revs.nr; i++)
 624                argv_array_pushf(&rev_argv, good_format,
 625                                 sha1_to_hex(good_revs.sha1[i]));
 626        argv_array_push(&rev_argv, "--");
 627        if (read_paths)
 628                read_bisect_paths(&rev_argv);
 629
 630        setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
 631        /* XXX leak rev_argv, as "revs" may still be pointing to it */
 632}
 633
 634static void bisect_common(struct rev_info *revs)
 635{
 636        if (prepare_revision_walk(revs))
 637                die("revision walk setup failed");
 638        if (revs->tree_objects)
 639                mark_edges_uninteresting(revs, NULL);
 640}
 641
 642static void exit_if_skipped_commits(struct commit_list *tried,
 643                                    const struct object_id *bad)
 644{
 645        if (!tried)
 646                return;
 647
 648        printf("There are only 'skip'ped commits left to test.\n"
 649               "The first %s commit could be any of:\n", term_bad);
 650        print_commit_list(tried, "%s\n", "%s\n");
 651        if (bad)
 652                printf("%s\n", oid_to_hex(bad));
 653        printf("We cannot bisect more!\n");
 654        exit(2);
 655}
 656
 657static int is_expected_rev(const struct object_id *oid)
 658{
 659        const char *filename = git_path_bisect_expected_rev();
 660        struct stat st;
 661        struct strbuf str = STRBUF_INIT;
 662        FILE *fp;
 663        int res = 0;
 664
 665        if (stat(filename, &st) || !S_ISREG(st.st_mode))
 666                return 0;
 667
 668        fp = fopen(filename, "r");
 669        if (!fp)
 670                return 0;
 671
 672        if (strbuf_getline(&str, fp, '\n') != EOF)
 673                res = !strcmp(str.buf, oid_to_hex(oid));
 674
 675        strbuf_release(&str);
 676        fclose(fp);
 677
 678        return res;
 679}
 680
 681static void mark_expected_rev(char *bisect_rev_hex)
 682{
 683        int len = strlen(bisect_rev_hex);
 684        const char *filename = git_path("BISECT_EXPECTED_REV");
 685        int fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
 686
 687        if (fd < 0)
 688                die_errno("could not create file '%s'", filename);
 689
 690        bisect_rev_hex[len] = '\n';
 691        write_or_die(fd, bisect_rev_hex, len + 1);
 692        bisect_rev_hex[len] = '\0';
 693
 694        if (close(fd) < 0)
 695                die("closing file %s: %s", filename, strerror(errno));
 696}
 697
 698static int bisect_checkout(char *bisect_rev_hex, int no_checkout)
 699{
 700
 701        mark_expected_rev(bisect_rev_hex);
 702
 703        argv_checkout[2] = bisect_rev_hex;
 704        if (no_checkout) {
 705                argv_update_ref[3] = bisect_rev_hex;
 706                if (run_command_v_opt(argv_update_ref, RUN_GIT_CMD))
 707                        die("update-ref --no-deref HEAD failed on %s",
 708                            bisect_rev_hex);
 709        } else {
 710                int res;
 711                res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
 712                if (res)
 713                        exit(res);
 714        }
 715
 716        argv_show_branch[1] = bisect_rev_hex;
 717        return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
 718}
 719
 720static struct commit *get_commit_reference(const unsigned char *sha1)
 721{
 722        struct commit *r = lookup_commit_reference(sha1);
 723        if (!r)
 724                die("Not a valid commit name %s", sha1_to_hex(sha1));
 725        return r;
 726}
 727
 728static struct commit **get_bad_and_good_commits(int *rev_nr)
 729{
 730        int len = 1 + good_revs.nr;
 731        struct commit **rev = xmalloc(len * sizeof(*rev));
 732        int i, n = 0;
 733
 734        rev[n++] = get_commit_reference(current_bad_oid->hash);
 735        for (i = 0; i < good_revs.nr; i++)
 736                rev[n++] = get_commit_reference(good_revs.sha1[i]);
 737        *rev_nr = n;
 738
 739        return rev;
 740}
 741
 742static void handle_bad_merge_base(void)
 743{
 744        if (is_expected_rev(current_bad_oid)) {
 745                char *bad_hex = oid_to_hex(current_bad_oid);
 746                char *good_hex = join_sha1_array_hex(&good_revs, ' ');
 747                if (!strcmp(term_bad, "bad") && !strcmp(term_good, "good")) {
 748                        fprintf(stderr, "The merge base %s is bad.\n"
 749                                "This means the bug has been fixed "
 750                                "between %s and [%s].\n",
 751                                bad_hex, bad_hex, good_hex);
 752                } else {
 753                        fprintf(stderr, "The merge base %s is %s.\n"
 754                                "This means the first '%s' commit is "
 755                                "between %s and [%s].\n",
 756                                bad_hex, term_bad, term_good, bad_hex, good_hex);
 757                }
 758                exit(3);
 759        }
 760
 761        fprintf(stderr, "Some %s revs are not ancestor of the %s rev.\n"
 762                "git bisect cannot work properly in this case.\n"
 763                "Maybe you mistook %s and %s revs?\n",
 764                term_good, term_bad, term_good, term_bad);
 765        exit(1);
 766}
 767
 768static void handle_skipped_merge_base(const unsigned char *mb)
 769{
 770        char *mb_hex = sha1_to_hex(mb);
 771        char *bad_hex = sha1_to_hex(current_bad_oid->hash);
 772        char *good_hex = join_sha1_array_hex(&good_revs, ' ');
 773
 774        warning("the merge base between %s and [%s] "
 775                "must be skipped.\n"
 776                "So we cannot be sure the first %s commit is "
 777                "between %s and %s.\n"
 778                "We continue anyway.",
 779                bad_hex, good_hex, term_bad, mb_hex, bad_hex);
 780        free(good_hex);
 781}
 782
 783/*
 784 * "check_merge_bases" checks that merge bases are not "bad".
 785 *
 786 * - If one is "bad", it means the user assumed something wrong
 787 * and we must exit with a non 0 error code.
 788 * - If one is "good", that's good, we have nothing to do.
 789 * - If one is "skipped", we can't know but we should warn.
 790 * - If we don't know, we should check it out and ask the user to test.
 791 */
 792static void check_merge_bases(int no_checkout)
 793{
 794        struct commit_list *result;
 795        int rev_nr;
 796        struct commit **rev = get_bad_and_good_commits(&rev_nr);
 797
 798        result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1);
 799
 800        for (; result; result = result->next) {
 801                const unsigned char *mb = result->item->object.sha1;
 802                if (!hashcmp(mb, current_bad_oid->hash)) {
 803                        handle_bad_merge_base();
 804                } else if (0 <= sha1_array_lookup(&good_revs, mb)) {
 805                        continue;
 806                } else if (0 <= sha1_array_lookup(&skipped_revs, mb)) {
 807                        handle_skipped_merge_base(mb);
 808                } else {
 809                        printf("Bisecting: a merge base must be tested\n");
 810                        exit(bisect_checkout(sha1_to_hex(mb), no_checkout));
 811                }
 812        }
 813
 814        free(rev);
 815        free_commit_list(result);
 816}
 817
 818static int check_ancestors(const char *prefix)
 819{
 820        struct rev_info revs;
 821        struct object_array pending_copy;
 822        int res;
 823
 824        bisect_rev_setup(&revs, prefix, "^%s", "%s", 0);
 825
 826        /* Save pending objects, so they can be cleaned up later. */
 827        pending_copy = revs.pending;
 828        revs.leak_pending = 1;
 829
 830        /*
 831         * bisect_common calls prepare_revision_walk right away, which
 832         * (together with .leak_pending = 1) makes us the sole owner of
 833         * the list of pending objects.
 834         */
 835        bisect_common(&revs);
 836        res = (revs.commits != NULL);
 837
 838        /* Clean up objects used, as they will be reused. */
 839        clear_commit_marks_for_object_array(&pending_copy, ALL_REV_FLAGS);
 840        free(pending_copy.objects);
 841
 842        return res;
 843}
 844
 845/*
 846 * "check_good_are_ancestors_of_bad" checks that all "good" revs are
 847 * ancestor of the "bad" rev.
 848 *
 849 * If that's not the case, we need to check the merge bases.
 850 * If a merge base must be tested by the user, its source code will be
 851 * checked out to be tested by the user and we will exit.
 852 */
 853static void check_good_are_ancestors_of_bad(const char *prefix, int no_checkout)
 854{
 855        char *filename = git_pathdup("BISECT_ANCESTORS_OK");
 856        struct stat st;
 857        int fd;
 858
 859        if (!current_bad_oid)
 860                die("a %s revision is needed", term_bad);
 861
 862        /* Check if file BISECT_ANCESTORS_OK exists. */
 863        if (!stat(filename, &st) && S_ISREG(st.st_mode))
 864                goto done;
 865
 866        /* Bisecting with no good rev is ok. */
 867        if (good_revs.nr == 0)
 868                goto done;
 869
 870        /* Check if all good revs are ancestor of the bad rev. */
 871        if (check_ancestors(prefix))
 872                check_merge_bases(no_checkout);
 873
 874        /* Create file BISECT_ANCESTORS_OK. */
 875        fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
 876        if (fd < 0)
 877                warning("could not create file '%s': %s",
 878                        filename, strerror(errno));
 879        else
 880                close(fd);
 881 done:
 882        free(filename);
 883}
 884
 885/*
 886 * This does "git diff-tree --pretty COMMIT" without one fork+exec.
 887 */
 888static void show_diff_tree(const char *prefix, struct commit *commit)
 889{
 890        struct rev_info opt;
 891
 892        /* diff-tree init */
 893        init_revisions(&opt, prefix);
 894        git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
 895        opt.abbrev = 0;
 896        opt.diff = 1;
 897
 898        /* This is what "--pretty" does */
 899        opt.verbose_header = 1;
 900        opt.use_terminator = 0;
 901        opt.commit_format = CMIT_FMT_DEFAULT;
 902
 903        /* diff-tree init */
 904        if (!opt.diffopt.output_format)
 905                opt.diffopt.output_format = DIFF_FORMAT_RAW;
 906
 907        log_tree_commit(&opt, commit);
 908}
 909
 910/*
 911 * The terms used for this bisect session are stored in BISECT_TERMS.
 912 * We read them and store them to adapt the messages accordingly.
 913 * Default is bad/good.
 914 */
 915void read_bisect_terms(const char **read_bad, const char **read_good)
 916{
 917        struct strbuf str = STRBUF_INIT;
 918        const char *filename = git_path("BISECT_TERMS");
 919        FILE *fp = fopen(filename, "r");
 920
 921        if (!fp) {
 922                if (errno == ENOENT) {
 923                        *read_bad = "bad";
 924                        *read_good = "good";
 925                        return;
 926                } else {
 927                        die("could not read file '%s': %s", filename,
 928                                strerror(errno));
 929                }
 930        } else {
 931                strbuf_getline(&str, fp, '\n');
 932                *read_bad = strbuf_detach(&str, NULL);
 933                strbuf_getline(&str, fp, '\n');
 934                *read_good = strbuf_detach(&str, NULL);
 935        }
 936        strbuf_release(&str);
 937        fclose(fp);
 938}
 939
 940/*
 941 * We use the convention that exiting with an exit code 10 means that
 942 * the bisection process finished successfully.
 943 * In this case the calling shell script should exit 0.
 944 *
 945 * If no_checkout is non-zero, the bisection process does not
 946 * checkout the trial commit but instead simply updates BISECT_HEAD.
 947 */
 948int bisect_next_all(const char *prefix, int no_checkout)
 949{
 950        struct rev_info revs;
 951        struct commit_list *tried;
 952        int reaches = 0, all = 0, nr, steps;
 953        const unsigned char *bisect_rev;
 954        char bisect_rev_hex[GIT_SHA1_HEXSZ + 1];
 955
 956        read_bisect_terms(&term_bad, &term_good);
 957        if (read_bisect_refs())
 958                die("reading bisect refs failed");
 959
 960        check_good_are_ancestors_of_bad(prefix, no_checkout);
 961
 962        bisect_rev_setup(&revs, prefix, "%s", "^%s", 1);
 963        revs.limited = 1;
 964
 965        bisect_common(&revs);
 966
 967        revs.commits = find_bisection(revs.commits, &reaches, &all,
 968                                       !!skipped_revs.nr);
 969        revs.commits = managed_skipped(revs.commits, &tried);
 970
 971        if (!revs.commits) {
 972                /*
 973                 * We should exit here only if the "bad"
 974                 * commit is also a "skip" commit.
 975                 */
 976                exit_if_skipped_commits(tried, NULL);
 977
 978                printf("%s was both %s and %s\n",
 979                       oid_to_hex(current_bad_oid),
 980                       term_good,
 981                       term_bad);
 982                exit(1);
 983        }
 984
 985        if (!all) {
 986                fprintf(stderr, "No testable commit found.\n"
 987                        "Maybe you started with bad path parameters?\n");
 988                exit(4);
 989        }
 990
 991        bisect_rev = revs.commits->item->object.sha1;
 992        memcpy(bisect_rev_hex, sha1_to_hex(bisect_rev), GIT_SHA1_HEXSZ + 1);
 993
 994        if (!hashcmp(bisect_rev, current_bad_oid->hash)) {
 995                exit_if_skipped_commits(tried, current_bad_oid);
 996                printf("%s is the first %s commit\n", bisect_rev_hex,
 997                        term_bad);
 998                show_diff_tree(prefix, revs.commits->item);
 999                /* This means the bisection process succeeded. */
1000                exit(10);
1001        }
1002
1003        nr = all - reaches - 1;
1004        steps = estimate_bisect_steps(all);
1005        printf("Bisecting: %d revision%s left to test after this "
1006               "(roughly %d step%s)\n", nr, (nr == 1 ? "" : "s"),
1007               steps, (steps == 1 ? "" : "s"));
1008
1009        return bisect_checkout(bisect_rev_hex, no_checkout);
1010}
1011
1012static inline int log2i(int n)
1013{
1014        int log2 = 0;
1015
1016        for (; n > 1; n >>= 1)
1017                log2++;
1018
1019        return log2;
1020}
1021
1022static inline int exp2i(int n)
1023{
1024        return 1 << n;
1025}
1026
1027/*
1028 * Estimate the number of bisect steps left (after the current step)
1029 *
1030 * For any x between 0 included and 2^n excluded, the probability for
1031 * n - 1 steps left looks like:
1032 *
1033 * P(2^n + x) == (2^n - x) / (2^n + x)
1034 *
1035 * and P(2^n + x) < 0.5 means 2^n < 3x
1036 */
1037int estimate_bisect_steps(int all)
1038{
1039        int n, x, e;
1040
1041        if (all < 3)
1042                return 0;
1043
1044        n = log2i(all);
1045        e = exp2i(n);
1046        x = all - e;
1047
1048        return (e < 3 * x) ? n : n - 1;
1049}