bisect.con commit Merge branch 'mh/tidy-ref-update-flags' (a972229)
   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
  16static struct oid_array good_revs;
  17static struct oid_array skipped_revs;
  18
  19static struct object_id *current_bad_oid;
  20
  21static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
  22static const char *argv_show_branch[] = {"show-branch", 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 oidcmp(&a->commit->object.oid, &b->commit->object.oid);
 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        struct strbuf buf = STRBUF_INIT;
 205        int cnt, i;
 206
 207        for (p = list, cnt = 0; p; p = p->next) {
 208                int distance;
 209                unsigned flags = p->item->object.flags;
 210
 211                if (flags & TREESAME)
 212                        continue;
 213                distance = weight(p);
 214                if (nr - distance < distance)
 215                        distance = nr - distance;
 216                array[cnt].commit = p->item;
 217                array[cnt].distance = distance;
 218                cnt++;
 219        }
 220        QSORT(array, cnt, compare_commit_dist);
 221        for (p = list, i = 0; i < cnt; i++) {
 222                struct object *obj = &(array[i].commit->object);
 223
 224                strbuf_reset(&buf);
 225                strbuf_addf(&buf, "dist=%d", array[i].distance);
 226                add_name_decoration(DECORATION_NONE, buf.buf, obj);
 227
 228                p->item = array[i].commit;
 229                if (i < cnt - 1)
 230                        p = p->next;
 231        }
 232        free_commit_list(p->next);
 233        p->next = NULL;
 234        strbuf_release(&buf);
 235        free(array);
 236        return list;
 237}
 238
 239/*
 240 * zero or positive weight is the number of interesting commits it can
 241 * reach, including itself.  Especially, weight = 0 means it does not
 242 * reach any tree-changing commits (e.g. just above uninteresting one
 243 * but traversal is with pathspec).
 244 *
 245 * weight = -1 means it has one parent and its distance is yet to
 246 * be computed.
 247 *
 248 * weight = -2 means it has more than one parent and its distance is
 249 * unknown.  After running count_distance() first, they will get zero
 250 * or positive distance.
 251 */
 252static struct commit_list *do_find_bisection(struct commit_list *list,
 253                                             int nr, int *weights,
 254                                             int find_all)
 255{
 256        int n, counted;
 257        struct commit_list *p;
 258
 259        counted = 0;
 260
 261        for (n = 0, p = list; p; p = p->next) {
 262                struct commit *commit = p->item;
 263                unsigned flags = commit->object.flags;
 264
 265                p->item->util = &weights[n++];
 266                switch (count_interesting_parents(commit)) {
 267                case 0:
 268                        if (!(flags & TREESAME)) {
 269                                weight_set(p, 1);
 270                                counted++;
 271                                show_list("bisection 2 count one",
 272                                          counted, nr, list);
 273                        }
 274                        /*
 275                         * otherwise, it is known not to reach any
 276                         * tree-changing commit and gets weight 0.
 277                         */
 278                        break;
 279                case 1:
 280                        weight_set(p, -1);
 281                        break;
 282                default:
 283                        weight_set(p, -2);
 284                        break;
 285                }
 286        }
 287
 288        show_list("bisection 2 initialize", counted, nr, list);
 289
 290        /*
 291         * If you have only one parent in the resulting set
 292         * then you can reach one commit more than that parent
 293         * can reach.  So we do not have to run the expensive
 294         * count_distance() for single strand of pearls.
 295         *
 296         * However, if you have more than one parents, you cannot
 297         * just add their distance and one for yourself, since
 298         * they usually reach the same ancestor and you would
 299         * end up counting them twice that way.
 300         *
 301         * So we will first count distance of merges the usual
 302         * way, and then fill the blanks using cheaper algorithm.
 303         */
 304        for (p = list; p; p = p->next) {
 305                if (p->item->object.flags & UNINTERESTING)
 306                        continue;
 307                if (weight(p) != -2)
 308                        continue;
 309                weight_set(p, count_distance(p));
 310                clear_distance(list);
 311
 312                /* Does it happen to be at exactly half-way? */
 313                if (!find_all && halfway(p, nr))
 314                        return p;
 315                counted++;
 316        }
 317
 318        show_list("bisection 2 count_distance", counted, nr, list);
 319
 320        while (counted < nr) {
 321                for (p = list; p; p = p->next) {
 322                        struct commit_list *q;
 323                        unsigned flags = p->item->object.flags;
 324
 325                        if (0 <= weight(p))
 326                                continue;
 327                        for (q = p->item->parents; q; q = q->next) {
 328                                if (q->item->object.flags & UNINTERESTING)
 329                                        continue;
 330                                if (0 <= weight(q))
 331                                        break;
 332                        }
 333                        if (!q)
 334                                continue;
 335
 336                        /*
 337                         * weight for p is unknown but q is known.
 338                         * add one for p itself if p is to be counted,
 339                         * otherwise inherit it from q directly.
 340                         */
 341                        if (!(flags & TREESAME)) {
 342                                weight_set(p, weight(q)+1);
 343                                counted++;
 344                                show_list("bisection 2 count one",
 345                                          counted, nr, list);
 346                        }
 347                        else
 348                                weight_set(p, weight(q));
 349
 350                        /* Does it happen to be at exactly half-way? */
 351                        if (!find_all && halfway(p, nr))
 352                                return p;
 353                }
 354        }
 355
 356        show_list("bisection 2 counted all", counted, nr, list);
 357
 358        if (!find_all)
 359                return best_bisection(list, nr);
 360        else
 361                return best_bisection_sorted(list, nr);
 362}
 363
 364void find_bisection(struct commit_list **commit_list, int *reaches,
 365                    int *all, int find_all)
 366{
 367        int nr, on_list;
 368        struct commit_list *list, *p, *best, *next, *last;
 369        int *weights;
 370
 371        show_list("bisection 2 entry", 0, 0, *commit_list);
 372
 373        /*
 374         * Count the number of total and tree-changing items on the
 375         * list, while reversing the list.
 376         */
 377        for (nr = on_list = 0, last = NULL, p = *commit_list;
 378             p;
 379             p = next) {
 380                unsigned flags = p->item->object.flags;
 381
 382                next = p->next;
 383                if (flags & UNINTERESTING) {
 384                        free(p);
 385                        continue;
 386                }
 387                p->next = last;
 388                last = p;
 389                if (!(flags & TREESAME))
 390                        nr++;
 391                on_list++;
 392        }
 393        list = last;
 394        show_list("bisection 2 sorted", 0, nr, list);
 395
 396        *all = nr;
 397        weights = xcalloc(on_list, sizeof(*weights));
 398
 399        /* Do the real work of finding bisection commit. */
 400        best = do_find_bisection(list, nr, weights, find_all);
 401        if (best) {
 402                if (!find_all) {
 403                        list->item = best->item;
 404                        free_commit_list(list->next);
 405                        best = list;
 406                        best->next = NULL;
 407                }
 408                *reaches = weight(best);
 409        }
 410        free(weights);
 411        *commit_list = best;
 412}
 413
 414static int register_ref(const char *refname, const struct object_id *oid,
 415                        int flags, void *cb_data)
 416{
 417        struct strbuf good_prefix = STRBUF_INIT;
 418        strbuf_addstr(&good_prefix, term_good);
 419        strbuf_addstr(&good_prefix, "-");
 420
 421        if (!strcmp(refname, term_bad)) {
 422                current_bad_oid = xmalloc(sizeof(*current_bad_oid));
 423                oidcpy(current_bad_oid, oid);
 424        } else if (starts_with(refname, good_prefix.buf)) {
 425                oid_array_append(&good_revs, oid);
 426        } else if (starts_with(refname, "skip-")) {
 427                oid_array_append(&skipped_revs, oid);
 428        }
 429
 430        strbuf_release(&good_prefix);
 431
 432        return 0;
 433}
 434
 435static int read_bisect_refs(void)
 436{
 437        return for_each_ref_in("refs/bisect/", register_ref, NULL);
 438}
 439
 440static GIT_PATH_FUNC(git_path_bisect_names, "BISECT_NAMES")
 441static GIT_PATH_FUNC(git_path_bisect_expected_rev, "BISECT_EXPECTED_REV")
 442static GIT_PATH_FUNC(git_path_bisect_ancestors_ok, "BISECT_ANCESTORS_OK")
 443static GIT_PATH_FUNC(git_path_bisect_run, "BISECT_RUN")
 444static GIT_PATH_FUNC(git_path_bisect_start, "BISECT_START")
 445static GIT_PATH_FUNC(git_path_bisect_log, "BISECT_LOG")
 446static GIT_PATH_FUNC(git_path_bisect_terms, "BISECT_TERMS")
 447static GIT_PATH_FUNC(git_path_head_name, "head-name")
 448
 449static void read_bisect_paths(struct argv_array *array)
 450{
 451        struct strbuf str = STRBUF_INIT;
 452        const char *filename = git_path_bisect_names();
 453        FILE *fp = xfopen(filename, "r");
 454
 455        while (strbuf_getline_lf(&str, fp) != EOF) {
 456                strbuf_trim(&str);
 457                if (sq_dequote_to_argv_array(str.buf, array))
 458                        die(_("Badly quoted content in file '%s': %s"),
 459                            filename, str.buf);
 460        }
 461
 462        strbuf_release(&str);
 463        fclose(fp);
 464}
 465
 466static char *join_sha1_array_hex(struct oid_array *array, char delim)
 467{
 468        struct strbuf joined_hexs = STRBUF_INIT;
 469        int i;
 470
 471        for (i = 0; i < array->nr; i++) {
 472                strbuf_addstr(&joined_hexs, oid_to_hex(array->oid + i));
 473                if (i + 1 < array->nr)
 474                        strbuf_addch(&joined_hexs, delim);
 475        }
 476
 477        return strbuf_detach(&joined_hexs, NULL);
 478}
 479
 480/*
 481 * In this function, passing a not NULL skipped_first is very special.
 482 * It means that we want to know if the first commit in the list is
 483 * skipped because we will want to test a commit away from it if it is
 484 * indeed skipped.
 485 * So if the first commit is skipped, we cannot take the shortcut to
 486 * just "return list" when we find the first non skipped commit, we
 487 * have to return a fully filtered list.
 488 *
 489 * We use (*skipped_first == -1) to mean "it has been found that the
 490 * first commit is not skipped". In this case *skipped_first is set back
 491 * to 0 just before the function returns.
 492 */
 493struct commit_list *filter_skipped(struct commit_list *list,
 494                                   struct commit_list **tried,
 495                                   int show_all,
 496                                   int *count,
 497                                   int *skipped_first)
 498{
 499        struct commit_list *filtered = NULL, **f = &filtered;
 500
 501        *tried = NULL;
 502
 503        if (skipped_first)
 504                *skipped_first = 0;
 505        if (count)
 506                *count = 0;
 507
 508        if (!skipped_revs.nr)
 509                return list;
 510
 511        while (list) {
 512                struct commit_list *next = list->next;
 513                list->next = NULL;
 514                if (0 <= oid_array_lookup(&skipped_revs, &list->item->object.oid)) {
 515                        if (skipped_first && !*skipped_first)
 516                                *skipped_first = 1;
 517                        /* Move current to tried list */
 518                        *tried = list;
 519                        tried = &list->next;
 520                } else {
 521                        if (!show_all) {
 522                                if (!skipped_first || !*skipped_first)
 523                                        return list;
 524                        } else if (skipped_first && !*skipped_first) {
 525                                /* This means we know it's not skipped */
 526                                *skipped_first = -1;
 527                        }
 528                        /* Move current to filtered list */
 529                        *f = list;
 530                        f = &list->next;
 531                        if (count)
 532                                (*count)++;
 533                }
 534                list = next;
 535        }
 536
 537        if (skipped_first && *skipped_first == -1)
 538                *skipped_first = 0;
 539
 540        return filtered;
 541}
 542
 543#define PRN_MODULO 32768
 544
 545/*
 546 * This is a pseudo random number generator based on "man 3 rand".
 547 * It is not used properly because the seed is the argument and it
 548 * is increased by one between each call, but that should not matter
 549 * for this application.
 550 */
 551static unsigned get_prn(unsigned count) {
 552        count = count * 1103515245 + 12345;
 553        return (count/65536) % PRN_MODULO;
 554}
 555
 556/*
 557 * Custom integer square root from
 558 * https://en.wikipedia.org/wiki/Integer_square_root
 559 */
 560static int sqrti(int val)
 561{
 562        float d, x = val;
 563
 564        if (val == 0)
 565                return 0;
 566
 567        do {
 568                float y = (x + (float)val / x) / 2;
 569                d = (y > x) ? y - x : x - y;
 570                x = y;
 571        } while (d >= 0.5);
 572
 573        return (int)x;
 574}
 575
 576static struct commit_list *skip_away(struct commit_list *list, int count)
 577{
 578        struct commit_list *cur, *previous;
 579        int prn, index, i;
 580
 581        prn = get_prn(count);
 582        index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
 583
 584        cur = list;
 585        previous = NULL;
 586
 587        for (i = 0; cur; cur = cur->next, i++) {
 588                if (i == index) {
 589                        if (oidcmp(&cur->item->object.oid, current_bad_oid))
 590                                return cur;
 591                        if (previous)
 592                                return previous;
 593                        return list;
 594                }
 595                previous = cur;
 596        }
 597
 598        return list;
 599}
 600
 601static struct commit_list *managed_skipped(struct commit_list *list,
 602                                           struct commit_list **tried)
 603{
 604        int count, skipped_first;
 605
 606        *tried = NULL;
 607
 608        if (!skipped_revs.nr)
 609                return list;
 610
 611        list = filter_skipped(list, tried, 0, &count, &skipped_first);
 612
 613        if (!skipped_first)
 614                return list;
 615
 616        return skip_away(list, count);
 617}
 618
 619static void bisect_rev_setup(struct rev_info *revs, const char *prefix,
 620                             const char *bad_format, const char *good_format,
 621                             int read_paths)
 622{
 623        struct argv_array rev_argv = ARGV_ARRAY_INIT;
 624        int i;
 625
 626        init_revisions(revs, prefix);
 627        revs->abbrev = 0;
 628        revs->commit_format = CMIT_FMT_UNSPECIFIED;
 629
 630        /* rev_argv.argv[0] will be ignored by setup_revisions */
 631        argv_array_push(&rev_argv, "bisect_rev_setup");
 632        argv_array_pushf(&rev_argv, bad_format, oid_to_hex(current_bad_oid));
 633        for (i = 0; i < good_revs.nr; i++)
 634                argv_array_pushf(&rev_argv, good_format,
 635                                 oid_to_hex(good_revs.oid + i));
 636        argv_array_push(&rev_argv, "--");
 637        if (read_paths)
 638                read_bisect_paths(&rev_argv);
 639
 640        setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
 641        /* XXX leak rev_argv, as "revs" may still be pointing to it */
 642}
 643
 644static void bisect_common(struct rev_info *revs)
 645{
 646        if (prepare_revision_walk(revs))
 647                die("revision walk setup failed");
 648        if (revs->tree_objects)
 649                mark_edges_uninteresting(revs, NULL);
 650}
 651
 652static void exit_if_skipped_commits(struct commit_list *tried,
 653                                    const struct object_id *bad)
 654{
 655        if (!tried)
 656                return;
 657
 658        printf("There are only 'skip'ped commits left to test.\n"
 659               "The first %s commit could be any of:\n", term_bad);
 660
 661        for ( ; tried; tried = tried->next)
 662                printf("%s\n", oid_to_hex(&tried->item->object.oid));
 663
 664        if (bad)
 665                printf("%s\n", oid_to_hex(bad));
 666        printf(_("We cannot bisect more!\n"));
 667        exit(2);
 668}
 669
 670static int is_expected_rev(const struct object_id *oid)
 671{
 672        const char *filename = git_path_bisect_expected_rev();
 673        struct stat st;
 674        struct strbuf str = STRBUF_INIT;
 675        FILE *fp;
 676        int res = 0;
 677
 678        if (stat(filename, &st) || !S_ISREG(st.st_mode))
 679                return 0;
 680
 681        fp = fopen_or_warn(filename, "r");
 682        if (!fp)
 683                return 0;
 684
 685        if (strbuf_getline_lf(&str, fp) != EOF)
 686                res = !strcmp(str.buf, oid_to_hex(oid));
 687
 688        strbuf_release(&str);
 689        fclose(fp);
 690
 691        return res;
 692}
 693
 694static int bisect_checkout(const struct object_id *bisect_rev, int no_checkout)
 695{
 696        char bisect_rev_hex[GIT_MAX_HEXSZ + 1];
 697
 698        memcpy(bisect_rev_hex, oid_to_hex(bisect_rev), GIT_SHA1_HEXSZ + 1);
 699        update_ref(NULL, "BISECT_EXPECTED_REV", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
 700
 701        argv_checkout[2] = bisect_rev_hex;
 702        if (no_checkout) {
 703                update_ref(NULL, "BISECT_HEAD", bisect_rev, NULL, 0,
 704                           UPDATE_REFS_DIE_ON_ERR);
 705        } else {
 706                int res;
 707                res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
 708                if (res)
 709                        exit(res);
 710        }
 711
 712        argv_show_branch[1] = bisect_rev_hex;
 713        return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
 714}
 715
 716static struct commit *get_commit_reference(const struct object_id *oid)
 717{
 718        struct commit *r = lookup_commit_reference(oid);
 719        if (!r)
 720                die(_("Not a valid commit name %s"), oid_to_hex(oid));
 721        return r;
 722}
 723
 724static struct commit **get_bad_and_good_commits(int *rev_nr)
 725{
 726        struct commit **rev;
 727        int i, n = 0;
 728
 729        ALLOC_ARRAY(rev, 1 + good_revs.nr);
 730        rev[n++] = get_commit_reference(current_bad_oid);
 731        for (i = 0; i < good_revs.nr; i++)
 732                rev[n++] = get_commit_reference(good_revs.oid + i);
 733        *rev_nr = n;
 734
 735        return rev;
 736}
 737
 738static void handle_bad_merge_base(void)
 739{
 740        if (is_expected_rev(current_bad_oid)) {
 741                char *bad_hex = oid_to_hex(current_bad_oid);
 742                char *good_hex = join_sha1_array_hex(&good_revs, ' ');
 743                if (!strcmp(term_bad, "bad") && !strcmp(term_good, "good")) {
 744                        fprintf(stderr, _("The merge base %s is bad.\n"
 745                                "This means the bug has been fixed "
 746                                "between %s and [%s].\n"),
 747                                bad_hex, bad_hex, good_hex);
 748                } else if (!strcmp(term_bad, "new") && !strcmp(term_good, "old")) {
 749                        fprintf(stderr, _("The merge base %s is new.\n"
 750                                "The property has changed "
 751                                "between %s and [%s].\n"),
 752                                bad_hex, bad_hex, good_hex);
 753                } else {
 754                        fprintf(stderr, _("The merge base %s is %s.\n"
 755                                "This means the first '%s' commit is "
 756                                "between %s and [%s].\n"),
 757                                bad_hex, term_bad, term_good, bad_hex, good_hex);
 758                }
 759                exit(3);
 760        }
 761
 762        fprintf(stderr, _("Some %s revs are not ancestors of the %s rev.\n"
 763                "git bisect cannot work properly in this case.\n"
 764                "Maybe you mistook %s and %s revs?\n"),
 765                term_good, term_bad, term_good, term_bad);
 766        exit(1);
 767}
 768
 769static void handle_skipped_merge_base(const struct object_id *mb)
 770{
 771        char *mb_hex = oid_to_hex(mb);
 772        char *bad_hex = oid_to_hex(current_bad_oid);
 773        char *good_hex = join_sha1_array_hex(&good_revs, ' ');
 774
 775        warning(_("the merge base between %s and [%s] "
 776                "must be skipped.\n"
 777                "So we cannot be sure the first %s commit is "
 778                "between %s and %s.\n"
 779                "We continue anyway."),
 780                bad_hex, good_hex, term_bad, mb_hex, bad_hex);
 781        free(good_hex);
 782}
 783
 784/*
 785 * "check_merge_bases" checks that merge bases are not "bad" (or "new").
 786 *
 787 * - If one is "bad" (or "new"), it means the user assumed something wrong
 788 * and we must exit with a non 0 error code.
 789 * - If one is "good" (or "old"), that's good, we have nothing to do.
 790 * - If one is "skipped", we can't know but we should warn.
 791 * - If we don't know, we should check it out and ask the user to test.
 792 */
 793static void check_merge_bases(int no_checkout)
 794{
 795        struct commit_list *result;
 796        int rev_nr;
 797        struct commit **rev = get_bad_and_good_commits(&rev_nr);
 798
 799        result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1);
 800
 801        for (; result; result = result->next) {
 802                const struct object_id *mb = &result->item->object.oid;
 803                if (!oidcmp(mb, current_bad_oid)) {
 804                        handle_bad_merge_base();
 805                } else if (0 <= oid_array_lookup(&good_revs, mb)) {
 806                        continue;
 807                } else if (0 <= oid_array_lookup(&skipped_revs, mb)) {
 808                        handle_skipped_merge_base(mb);
 809                } else {
 810                        printf(_("Bisecting: a merge base must be tested\n"));
 811                        exit(bisect_checkout(mb, no_checkout));
 812                }
 813        }
 814
 815        free(rev);
 816        free_commit_list(result);
 817}
 818
 819static int check_ancestors(const char *prefix)
 820{
 821        struct rev_info revs;
 822        struct object_array pending_copy;
 823        int res;
 824
 825        bisect_rev_setup(&revs, prefix, "^%s", "%s", 0);
 826
 827        /* Save pending objects, so they can be cleaned up later. */
 828        pending_copy = revs.pending;
 829        revs.leak_pending = 1;
 830
 831        /*
 832         * bisect_common calls prepare_revision_walk right away, which
 833         * (together with .leak_pending = 1) makes us the sole owner of
 834         * the list of pending objects.
 835         */
 836        bisect_common(&revs);
 837        res = (revs.commits != NULL);
 838
 839        /* Clean up objects used, as they will be reused. */
 840        clear_commit_marks_for_object_array(&pending_copy, ALL_REV_FLAGS);
 841
 842        object_array_clear(&pending_copy);
 843
 844        return res;
 845}
 846
 847/*
 848 * "check_good_are_ancestors_of_bad" checks that all "good" revs are
 849 * ancestor of the "bad" rev.
 850 *
 851 * If that's not the case, we need to check the merge bases.
 852 * If a merge base must be tested by the user, its source code will be
 853 * checked out to be tested by the user and we will exit.
 854 */
 855static void check_good_are_ancestors_of_bad(const char *prefix, int no_checkout)
 856{
 857        char *filename = git_pathdup("BISECT_ANCESTORS_OK");
 858        struct stat st;
 859        int fd;
 860
 861        if (!current_bad_oid)
 862                die(_("a %s revision is needed"), term_bad);
 863
 864        /* Check if file BISECT_ANCESTORS_OK exists. */
 865        if (!stat(filename, &st) && S_ISREG(st.st_mode))
 866                goto done;
 867
 868        /* Bisecting with no good rev is ok. */
 869        if (good_revs.nr == 0)
 870                goto done;
 871
 872        /* Check if all good revs are ancestor of the bad rev. */
 873        if (check_ancestors(prefix))
 874                check_merge_bases(no_checkout);
 875
 876        /* Create file BISECT_ANCESTORS_OK. */
 877        fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
 878        if (fd < 0)
 879                warning_errno(_("could not create file '%s'"),
 880                              filename);
 881        else
 882                close(fd);
 883 done:
 884        free(filename);
 885}
 886
 887/*
 888 * This does "git diff-tree --pretty COMMIT" without one fork+exec.
 889 */
 890static void show_diff_tree(const char *prefix, struct commit *commit)
 891{
 892        struct rev_info opt;
 893
 894        /* diff-tree init */
 895        init_revisions(&opt, prefix);
 896        git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
 897        opt.abbrev = 0;
 898        opt.diff = 1;
 899
 900        /* This is what "--pretty" does */
 901        opt.verbose_header = 1;
 902        opt.use_terminator = 0;
 903        opt.commit_format = CMIT_FMT_DEFAULT;
 904
 905        /* diff-tree init */
 906        if (!opt.diffopt.output_format)
 907                opt.diffopt.output_format = DIFF_FORMAT_RAW;
 908
 909        setup_revisions(0, NULL, &opt, NULL);
 910        log_tree_commit(&opt, commit);
 911}
 912
 913/*
 914 * The terms used for this bisect session are stored in BISECT_TERMS.
 915 * We read them and store them to adapt the messages accordingly.
 916 * Default is bad/good.
 917 */
 918void read_bisect_terms(const char **read_bad, const char **read_good)
 919{
 920        struct strbuf str = STRBUF_INIT;
 921        const char *filename = git_path_bisect_terms();
 922        FILE *fp = fopen(filename, "r");
 923
 924        if (!fp) {
 925                if (errno == ENOENT) {
 926                        *read_bad = "bad";
 927                        *read_good = "good";
 928                        return;
 929                } else {
 930                        die_errno(_("could not read file '%s'"), filename);
 931                }
 932        } else {
 933                strbuf_getline_lf(&str, fp);
 934                *read_bad = strbuf_detach(&str, NULL);
 935                strbuf_getline_lf(&str, fp);
 936                *read_good = strbuf_detach(&str, NULL);
 937        }
 938        strbuf_release(&str);
 939        fclose(fp);
 940}
 941
 942/*
 943 * We use the convention that exiting with an exit code 10 means that
 944 * the bisection process finished successfully.
 945 * In this case the calling shell script should exit 0.
 946 *
 947 * If no_checkout is non-zero, the bisection process does not
 948 * checkout the trial commit but instead simply updates BISECT_HEAD.
 949 */
 950int bisect_next_all(const char *prefix, int no_checkout)
 951{
 952        struct rev_info revs;
 953        struct commit_list *tried;
 954        int reaches = 0, all = 0, nr, steps;
 955        struct object_id *bisect_rev;
 956        char *steps_msg;
 957
 958        read_bisect_terms(&term_bad, &term_good);
 959        if (read_bisect_refs())
 960                die(_("reading bisect refs failed"));
 961
 962        check_good_are_ancestors_of_bad(prefix, no_checkout);
 963
 964        bisect_rev_setup(&revs, prefix, "%s", "^%s", 1);
 965        revs.limited = 1;
 966
 967        bisect_common(&revs);
 968
 969        find_bisection(&revs.commits, &reaches, &all, !!skipped_revs.nr);
 970        revs.commits = managed_skipped(revs.commits, &tried);
 971
 972        if (!revs.commits) {
 973                /*
 974                 * We should exit here only if the "bad"
 975                 * commit is also a "skip" commit.
 976                 */
 977                exit_if_skipped_commits(tried, NULL);
 978
 979                printf(_("%s was both %s and %s\n"),
 980                       oid_to_hex(current_bad_oid),
 981                       term_good,
 982                       term_bad);
 983                exit(1);
 984        }
 985
 986        if (!all) {
 987                fprintf(stderr, _("No testable commit found.\n"
 988                        "Maybe you started with bad path parameters?\n"));
 989                exit(4);
 990        }
 991
 992        bisect_rev = &revs.commits->item->object.oid;
 993
 994        if (!oidcmp(bisect_rev, current_bad_oid)) {
 995                exit_if_skipped_commits(tried, current_bad_oid);
 996                printf("%s is the first %s commit\n", oid_to_hex(bisect_rev),
 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
1006        steps_msg = xstrfmt(Q_("(roughly %d step)", "(roughly %d steps)",
1007                  steps), steps);
1008        /*
1009         * TRANSLATORS: the last %s will be replaced with "(roughly %d
1010         * steps)" translation.
1011         */
1012        printf(Q_("Bisecting: %d revision left to test after this %s\n",
1013                  "Bisecting: %d revisions left to test after this %s\n",
1014                  nr), nr, steps_msg);
1015        free(steps_msg);
1016
1017        return bisect_checkout(bisect_rev, no_checkout);
1018}
1019
1020static inline int log2i(int n)
1021{
1022        int log2 = 0;
1023
1024        for (; n > 1; n >>= 1)
1025                log2++;
1026
1027        return log2;
1028}
1029
1030static inline int exp2i(int n)
1031{
1032        return 1 << n;
1033}
1034
1035/*
1036 * Estimate the number of bisect steps left (after the current step)
1037 *
1038 * For any x between 0 included and 2^n excluded, the probability for
1039 * n - 1 steps left looks like:
1040 *
1041 * P(2^n + x) == (2^n - x) / (2^n + x)
1042 *
1043 * and P(2^n + x) < 0.5 means 2^n < 3x
1044 */
1045int estimate_bisect_steps(int all)
1046{
1047        int n, x, e;
1048
1049        if (all < 3)
1050                return 0;
1051
1052        n = log2i(all);
1053        e = exp2i(n);
1054        x = all - e;
1055
1056        return (e < 3 * x) ? n : n - 1;
1057}
1058
1059static int mark_for_removal(const char *refname, const struct object_id *oid,
1060                            int flag, void *cb_data)
1061{
1062        struct string_list *refs = cb_data;
1063        char *ref = xstrfmt("refs/bisect%s", refname);
1064        string_list_append(refs, ref);
1065        return 0;
1066}
1067
1068int bisect_clean_state(void)
1069{
1070        int result = 0;
1071
1072        /* There may be some refs packed during bisection */
1073        struct string_list refs_for_removal = STRING_LIST_INIT_NODUP;
1074        for_each_ref_in("refs/bisect", mark_for_removal, (void *) &refs_for_removal);
1075        string_list_append(&refs_for_removal, xstrdup("BISECT_HEAD"));
1076        result = delete_refs("bisect: remove", &refs_for_removal, REF_NO_DEREF);
1077        refs_for_removal.strdup_strings = 1;
1078        string_list_clear(&refs_for_removal, 0);
1079        unlink_or_warn(git_path_bisect_expected_rev());
1080        unlink_or_warn(git_path_bisect_ancestors_ok());
1081        unlink_or_warn(git_path_bisect_log());
1082        unlink_or_warn(git_path_bisect_names());
1083        unlink_or_warn(git_path_bisect_run());
1084        unlink_or_warn(git_path_bisect_terms());
1085        /* Cleanup head-name if it got left by an old version of git-bisect */
1086        unlink_or_warn(git_path_head_name());
1087        /*
1088         * Cleanup BISECT_START last to support the --no-checkout option
1089         * introduced in the commit 4796e823a.
1090         */
1091        unlink_or_warn(git_path_bisect_start());
1092
1093        return result;
1094}