bisect.con commit sequencer: refactor check_todo_list() to work on a todo_list (6ca89c6)
   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#include "commit-slab.h"
  16#include "commit-reach.h"
  17#include "object-store.h"
  18
  19static struct oid_array good_revs;
  20static struct oid_array skipped_revs;
  21
  22static struct object_id *current_bad_oid;
  23
  24static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
  25static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
  26
  27static const char *term_bad;
  28static const char *term_good;
  29
  30/* Remember to update object flag allocation in object.h */
  31#define COUNTED         (1u<<16)
  32
  33/*
  34 * This is a truly stupid algorithm, but it's only
  35 * used for bisection, and we just don't care enough.
  36 *
  37 * We care just barely enough to avoid recursing for
  38 * non-merge entries.
  39 */
  40static int count_distance(struct commit_list *entry)
  41{
  42        int nr = 0;
  43
  44        while (entry) {
  45                struct commit *commit = entry->item;
  46                struct commit_list *p;
  47
  48                if (commit->object.flags & (UNINTERESTING | COUNTED))
  49                        break;
  50                if (!(commit->object.flags & TREESAME))
  51                        nr++;
  52                commit->object.flags |= COUNTED;
  53                p = commit->parents;
  54                entry = p;
  55                if (p) {
  56                        p = p->next;
  57                        while (p) {
  58                                nr += count_distance(p);
  59                                p = p->next;
  60                        }
  61                }
  62        }
  63
  64        return nr;
  65}
  66
  67static void clear_distance(struct commit_list *list)
  68{
  69        while (list) {
  70                struct commit *commit = list->item;
  71                commit->object.flags &= ~COUNTED;
  72                list = list->next;
  73        }
  74}
  75
  76define_commit_slab(commit_weight, int *);
  77static struct commit_weight commit_weight;
  78
  79#define DEBUG_BISECT 0
  80
  81static inline int weight(struct commit_list *elem)
  82{
  83        return **commit_weight_at(&commit_weight, elem->item);
  84}
  85
  86static inline void weight_set(struct commit_list *elem, int weight)
  87{
  88        **commit_weight_at(&commit_weight, elem->item) = weight;
  89}
  90
  91static int count_interesting_parents(struct commit *commit)
  92{
  93        struct commit_list *p;
  94        int count;
  95
  96        for (count = 0, p = commit->parents; p; p = p->next) {
  97                if (p->item->object.flags & UNINTERESTING)
  98                        continue;
  99                count++;
 100        }
 101        return count;
 102}
 103
 104static inline int halfway(struct commit_list *p, int nr)
 105{
 106        /*
 107         * Don't short-cut something we are not going to return!
 108         */
 109        if (p->item->object.flags & TREESAME)
 110                return 0;
 111        if (DEBUG_BISECT)
 112                return 0;
 113        /*
 114         * 2 and 3 are halfway of 5.
 115         * 3 is halfway of 6 but 2 and 4 are not.
 116         */
 117        switch (2 * weight(p) - nr) {
 118        case -1: case 0: case 1:
 119                return 1;
 120        default:
 121                return 0;
 122        }
 123}
 124
 125static void show_list(const char *debug, int counted, int nr,
 126                      struct commit_list *list)
 127{
 128        struct commit_list *p;
 129
 130        if (!DEBUG_BISECT)
 131                return;
 132
 133        fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
 134
 135        for (p = list; p; p = p->next) {
 136                struct commit_list *pp;
 137                struct commit *commit = p->item;
 138                unsigned flags = commit->object.flags;
 139                enum object_type type;
 140                unsigned long size;
 141                char *buf = read_object_file(&commit->object.oid, &type,
 142                                             &size);
 143                const char *subject_start;
 144                int subject_len;
 145
 146                fprintf(stderr, "%c%c%c ",
 147                        (flags & TREESAME) ? ' ' : 'T',
 148                        (flags & UNINTERESTING) ? 'U' : ' ',
 149                        (flags & COUNTED) ? 'C' : ' ');
 150                if (*commit_weight_at(&commit_weight, p->item))
 151                        fprintf(stderr, "%3d", weight(p));
 152                else
 153                        fprintf(stderr, "---");
 154                fprintf(stderr, " %.*s", 8, oid_to_hex(&commit->object.oid));
 155                for (pp = commit->parents; pp; pp = pp->next)
 156                        fprintf(stderr, " %.*s", 8,
 157                                oid_to_hex(&pp->item->object.oid));
 158
 159                subject_len = find_commit_subject(buf, &subject_start);
 160                if (subject_len)
 161                        fprintf(stderr, " %.*s", subject_len, subject_start);
 162                fprintf(stderr, "\n");
 163        }
 164}
 165
 166static struct commit_list *best_bisection(struct commit_list *list, int nr)
 167{
 168        struct commit_list *p, *best;
 169        int best_distance = -1;
 170
 171        best = list;
 172        for (p = list; p; p = p->next) {
 173                int distance;
 174                unsigned flags = p->item->object.flags;
 175
 176                if (flags & TREESAME)
 177                        continue;
 178                distance = weight(p);
 179                if (nr - distance < distance)
 180                        distance = nr - distance;
 181                if (distance > best_distance) {
 182                        best = p;
 183                        best_distance = distance;
 184                }
 185        }
 186
 187        return best;
 188}
 189
 190struct commit_dist {
 191        struct commit *commit;
 192        int distance;
 193};
 194
 195static int compare_commit_dist(const void *a_, const void *b_)
 196{
 197        struct commit_dist *a, *b;
 198
 199        a = (struct commit_dist *)a_;
 200        b = (struct commit_dist *)b_;
 201        if (a->distance != b->distance)
 202                return b->distance - a->distance; /* desc sort */
 203        return oidcmp(&a->commit->object.oid, &b->commit->object.oid);
 204}
 205
 206static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
 207{
 208        struct commit_list *p;
 209        struct commit_dist *array = xcalloc(nr, sizeof(*array));
 210        struct strbuf buf = STRBUF_INIT;
 211        int cnt, i;
 212
 213        for (p = list, cnt = 0; p; p = p->next) {
 214                int distance;
 215                unsigned flags = p->item->object.flags;
 216
 217                if (flags & TREESAME)
 218                        continue;
 219                distance = weight(p);
 220                if (nr - distance < distance)
 221                        distance = nr - distance;
 222                array[cnt].commit = p->item;
 223                array[cnt].distance = distance;
 224                cnt++;
 225        }
 226        QSORT(array, cnt, compare_commit_dist);
 227        for (p = list, i = 0; i < cnt; i++) {
 228                struct object *obj = &(array[i].commit->object);
 229
 230                strbuf_reset(&buf);
 231                strbuf_addf(&buf, "dist=%d", array[i].distance);
 232                add_name_decoration(DECORATION_NONE, buf.buf, obj);
 233
 234                p->item = array[i].commit;
 235                if (i < cnt - 1)
 236                        p = p->next;
 237        }
 238        if (p) {
 239                free_commit_list(p->next);
 240                p->next = NULL;
 241        }
 242        strbuf_release(&buf);
 243        free(array);
 244        return list;
 245}
 246
 247/*
 248 * zero or positive weight is the number of interesting commits it can
 249 * reach, including itself.  Especially, weight = 0 means it does not
 250 * reach any tree-changing commits (e.g. just above uninteresting one
 251 * but traversal is with pathspec).
 252 *
 253 * weight = -1 means it has one parent and its distance is yet to
 254 * be computed.
 255 *
 256 * weight = -2 means it has more than one parent and its distance is
 257 * unknown.  After running count_distance() first, they will get zero
 258 * or positive distance.
 259 */
 260static struct commit_list *do_find_bisection(struct commit_list *list,
 261                                             int nr, int *weights,
 262                                             int find_all)
 263{
 264        int n, counted;
 265        struct commit_list *p;
 266
 267        counted = 0;
 268
 269        for (n = 0, p = list; p; p = p->next) {
 270                struct commit *commit = p->item;
 271                unsigned flags = commit->object.flags;
 272
 273                *commit_weight_at(&commit_weight, p->item) = &weights[n++];
 274                switch (count_interesting_parents(commit)) {
 275                case 0:
 276                        if (!(flags & TREESAME)) {
 277                                weight_set(p, 1);
 278                                counted++;
 279                                show_list("bisection 2 count one",
 280                                          counted, nr, list);
 281                        }
 282                        /*
 283                         * otherwise, it is known not to reach any
 284                         * tree-changing commit and gets weight 0.
 285                         */
 286                        break;
 287                case 1:
 288                        weight_set(p, -1);
 289                        break;
 290                default:
 291                        weight_set(p, -2);
 292                        break;
 293                }
 294        }
 295
 296        show_list("bisection 2 initialize", counted, nr, list);
 297
 298        /*
 299         * If you have only one parent in the resulting set
 300         * then you can reach one commit more than that parent
 301         * can reach.  So we do not have to run the expensive
 302         * count_distance() for single strand of pearls.
 303         *
 304         * However, if you have more than one parents, you cannot
 305         * just add their distance and one for yourself, since
 306         * they usually reach the same ancestor and you would
 307         * end up counting them twice that way.
 308         *
 309         * So we will first count distance of merges the usual
 310         * way, and then fill the blanks using cheaper algorithm.
 311         */
 312        for (p = list; p; p = p->next) {
 313                if (p->item->object.flags & UNINTERESTING)
 314                        continue;
 315                if (weight(p) != -2)
 316                        continue;
 317                weight_set(p, count_distance(p));
 318                clear_distance(list);
 319
 320                /* Does it happen to be at exactly half-way? */
 321                if (!find_all && halfway(p, nr))
 322                        return p;
 323                counted++;
 324        }
 325
 326        show_list("bisection 2 count_distance", counted, nr, list);
 327
 328        while (counted < nr) {
 329                for (p = list; p; p = p->next) {
 330                        struct commit_list *q;
 331                        unsigned flags = p->item->object.flags;
 332
 333                        if (0 <= weight(p))
 334                                continue;
 335                        for (q = p->item->parents; q; q = q->next) {
 336                                if (q->item->object.flags & UNINTERESTING)
 337                                        continue;
 338                                if (0 <= weight(q))
 339                                        break;
 340                        }
 341                        if (!q)
 342                                continue;
 343
 344                        /*
 345                         * weight for p is unknown but q is known.
 346                         * add one for p itself if p is to be counted,
 347                         * otherwise inherit it from q directly.
 348                         */
 349                        if (!(flags & TREESAME)) {
 350                                weight_set(p, weight(q)+1);
 351                                counted++;
 352                                show_list("bisection 2 count one",
 353                                          counted, nr, list);
 354                        }
 355                        else
 356                                weight_set(p, weight(q));
 357
 358                        /* Does it happen to be at exactly half-way? */
 359                        if (!find_all && halfway(p, nr))
 360                                return p;
 361                }
 362        }
 363
 364        show_list("bisection 2 counted all", counted, nr, list);
 365
 366        if (!find_all)
 367                return best_bisection(list, nr);
 368        else
 369                return best_bisection_sorted(list, nr);
 370}
 371
 372void find_bisection(struct commit_list **commit_list, int *reaches,
 373                    int *all, int find_all)
 374{
 375        int nr, on_list;
 376        struct commit_list *list, *p, *best, *next, *last;
 377        int *weights;
 378
 379        show_list("bisection 2 entry", 0, 0, *commit_list);
 380        init_commit_weight(&commit_weight);
 381
 382        /*
 383         * Count the number of total and tree-changing items on the
 384         * list, while reversing the list.
 385         */
 386        for (nr = on_list = 0, last = NULL, p = *commit_list;
 387             p;
 388             p = next) {
 389                unsigned flags = p->item->object.flags;
 390
 391                next = p->next;
 392                if (flags & UNINTERESTING) {
 393                        free(p);
 394                        continue;
 395                }
 396                p->next = last;
 397                last = p;
 398                if (!(flags & TREESAME))
 399                        nr++;
 400                on_list++;
 401        }
 402        list = last;
 403        show_list("bisection 2 sorted", 0, nr, list);
 404
 405        *all = nr;
 406        weights = xcalloc(on_list, sizeof(*weights));
 407
 408        /* Do the real work of finding bisection commit. */
 409        best = do_find_bisection(list, nr, weights, find_all);
 410        if (best) {
 411                if (!find_all) {
 412                        list->item = best->item;
 413                        free_commit_list(list->next);
 414                        best = list;
 415                        best->next = NULL;
 416                }
 417                *reaches = weight(best);
 418        }
 419        free(weights);
 420        *commit_list = best;
 421        clear_commit_weight(&commit_weight);
 422}
 423
 424static int register_ref(const char *refname, const struct object_id *oid,
 425                        int flags, void *cb_data)
 426{
 427        struct strbuf good_prefix = STRBUF_INIT;
 428        strbuf_addstr(&good_prefix, term_good);
 429        strbuf_addstr(&good_prefix, "-");
 430
 431        if (!strcmp(refname, term_bad)) {
 432                current_bad_oid = xmalloc(sizeof(*current_bad_oid));
 433                oidcpy(current_bad_oid, oid);
 434        } else if (starts_with(refname, good_prefix.buf)) {
 435                oid_array_append(&good_revs, oid);
 436        } else if (starts_with(refname, "skip-")) {
 437                oid_array_append(&skipped_revs, oid);
 438        }
 439
 440        strbuf_release(&good_prefix);
 441
 442        return 0;
 443}
 444
 445static int read_bisect_refs(void)
 446{
 447        return for_each_ref_in("refs/bisect/", register_ref, NULL);
 448}
 449
 450static GIT_PATH_FUNC(git_path_bisect_names, "BISECT_NAMES")
 451static GIT_PATH_FUNC(git_path_bisect_expected_rev, "BISECT_EXPECTED_REV")
 452static GIT_PATH_FUNC(git_path_bisect_ancestors_ok, "BISECT_ANCESTORS_OK")
 453static GIT_PATH_FUNC(git_path_bisect_run, "BISECT_RUN")
 454static GIT_PATH_FUNC(git_path_bisect_start, "BISECT_START")
 455static GIT_PATH_FUNC(git_path_bisect_log, "BISECT_LOG")
 456static GIT_PATH_FUNC(git_path_bisect_terms, "BISECT_TERMS")
 457static GIT_PATH_FUNC(git_path_head_name, "head-name")
 458
 459static void read_bisect_paths(struct argv_array *array)
 460{
 461        struct strbuf str = STRBUF_INIT;
 462        const char *filename = git_path_bisect_names();
 463        FILE *fp = xfopen(filename, "r");
 464
 465        while (strbuf_getline_lf(&str, fp) != EOF) {
 466                strbuf_trim(&str);
 467                if (sq_dequote_to_argv_array(str.buf, array))
 468                        die(_("Badly quoted content in file '%s': %s"),
 469                            filename, str.buf);
 470        }
 471
 472        strbuf_release(&str);
 473        fclose(fp);
 474}
 475
 476static char *join_sha1_array_hex(struct oid_array *array, char delim)
 477{
 478        struct strbuf joined_hexs = STRBUF_INIT;
 479        int i;
 480
 481        for (i = 0; i < array->nr; i++) {
 482                strbuf_addstr(&joined_hexs, oid_to_hex(array->oid + i));
 483                if (i + 1 < array->nr)
 484                        strbuf_addch(&joined_hexs, delim);
 485        }
 486
 487        return strbuf_detach(&joined_hexs, NULL);
 488}
 489
 490/*
 491 * In this function, passing a not NULL skipped_first is very special.
 492 * It means that we want to know if the first commit in the list is
 493 * skipped because we will want to test a commit away from it if it is
 494 * indeed skipped.
 495 * So if the first commit is skipped, we cannot take the shortcut to
 496 * just "return list" when we find the first non skipped commit, we
 497 * have to return a fully filtered list.
 498 *
 499 * We use (*skipped_first == -1) to mean "it has been found that the
 500 * first commit is not skipped". In this case *skipped_first is set back
 501 * to 0 just before the function returns.
 502 */
 503struct commit_list *filter_skipped(struct commit_list *list,
 504                                   struct commit_list **tried,
 505                                   int show_all,
 506                                   int *count,
 507                                   int *skipped_first)
 508{
 509        struct commit_list *filtered = NULL, **f = &filtered;
 510
 511        *tried = NULL;
 512
 513        if (skipped_first)
 514                *skipped_first = 0;
 515        if (count)
 516                *count = 0;
 517
 518        if (!skipped_revs.nr)
 519                return list;
 520
 521        while (list) {
 522                struct commit_list *next = list->next;
 523                list->next = NULL;
 524                if (0 <= oid_array_lookup(&skipped_revs, &list->item->object.oid)) {
 525                        if (skipped_first && !*skipped_first)
 526                                *skipped_first = 1;
 527                        /* Move current to tried list */
 528                        *tried = list;
 529                        tried = &list->next;
 530                } else {
 531                        if (!show_all) {
 532                                if (!skipped_first || !*skipped_first)
 533                                        return list;
 534                        } else if (skipped_first && !*skipped_first) {
 535                                /* This means we know it's not skipped */
 536                                *skipped_first = -1;
 537                        }
 538                        /* Move current to filtered list */
 539                        *f = list;
 540                        f = &list->next;
 541                        if (count)
 542                                (*count)++;
 543                }
 544                list = next;
 545        }
 546
 547        if (skipped_first && *skipped_first == -1)
 548                *skipped_first = 0;
 549
 550        return filtered;
 551}
 552
 553#define PRN_MODULO 32768
 554
 555/*
 556 * This is a pseudo random number generator based on "man 3 rand".
 557 * It is not used properly because the seed is the argument and it
 558 * is increased by one between each call, but that should not matter
 559 * for this application.
 560 */
 561static unsigned get_prn(unsigned count) {
 562        count = count * 1103515245 + 12345;
 563        return (count/65536) % PRN_MODULO;
 564}
 565
 566/*
 567 * Custom integer square root from
 568 * https://en.wikipedia.org/wiki/Integer_square_root
 569 */
 570static int sqrti(int val)
 571{
 572        float d, x = val;
 573
 574        if (val == 0)
 575                return 0;
 576
 577        do {
 578                float y = (x + (float)val / x) / 2;
 579                d = (y > x) ? y - x : x - y;
 580                x = y;
 581        } while (d >= 0.5);
 582
 583        return (int)x;
 584}
 585
 586static struct commit_list *skip_away(struct commit_list *list, int count)
 587{
 588        struct commit_list *cur, *previous;
 589        int prn, index, i;
 590
 591        prn = get_prn(count);
 592        index = (count * prn / PRN_MODULO) * sqrti(prn) / sqrti(PRN_MODULO);
 593
 594        cur = list;
 595        previous = NULL;
 596
 597        for (i = 0; cur; cur = cur->next, i++) {
 598                if (i == index) {
 599                        if (!oideq(&cur->item->object.oid, current_bad_oid))
 600                                return cur;
 601                        if (previous)
 602                                return previous;
 603                        return list;
 604                }
 605                previous = cur;
 606        }
 607
 608        return list;
 609}
 610
 611static struct commit_list *managed_skipped(struct commit_list *list,
 612                                           struct commit_list **tried)
 613{
 614        int count, skipped_first;
 615
 616        *tried = NULL;
 617
 618        if (!skipped_revs.nr)
 619                return list;
 620
 621        list = filter_skipped(list, tried, 0, &count, &skipped_first);
 622
 623        if (!skipped_first)
 624                return list;
 625
 626        return skip_away(list, count);
 627}
 628
 629static void bisect_rev_setup(struct repository *r, struct rev_info *revs,
 630                             const char *prefix,
 631                             const char *bad_format, const char *good_format,
 632                             int read_paths)
 633{
 634        struct argv_array rev_argv = ARGV_ARRAY_INIT;
 635        int i;
 636
 637        repo_init_revisions(r, revs, prefix);
 638        revs->abbrev = 0;
 639        revs->commit_format = CMIT_FMT_UNSPECIFIED;
 640
 641        /* rev_argv.argv[0] will be ignored by setup_revisions */
 642        argv_array_push(&rev_argv, "bisect_rev_setup");
 643        argv_array_pushf(&rev_argv, bad_format, oid_to_hex(current_bad_oid));
 644        for (i = 0; i < good_revs.nr; i++)
 645                argv_array_pushf(&rev_argv, good_format,
 646                                 oid_to_hex(good_revs.oid + i));
 647        argv_array_push(&rev_argv, "--");
 648        if (read_paths)
 649                read_bisect_paths(&rev_argv);
 650
 651        setup_revisions(rev_argv.argc, rev_argv.argv, revs, NULL);
 652        /* XXX leak rev_argv, as "revs" may still be pointing to it */
 653}
 654
 655static void bisect_common(struct rev_info *revs)
 656{
 657        if (prepare_revision_walk(revs))
 658                die("revision walk setup failed");
 659        if (revs->tree_objects)
 660                mark_edges_uninteresting(revs, NULL);
 661}
 662
 663static void exit_if_skipped_commits(struct commit_list *tried,
 664                                    const struct object_id *bad)
 665{
 666        if (!tried)
 667                return;
 668
 669        printf("There are only 'skip'ped commits left to test.\n"
 670               "The first %s commit could be any of:\n", term_bad);
 671
 672        for ( ; tried; tried = tried->next)
 673                printf("%s\n", oid_to_hex(&tried->item->object.oid));
 674
 675        if (bad)
 676                printf("%s\n", oid_to_hex(bad));
 677        printf(_("We cannot bisect more!\n"));
 678        exit(2);
 679}
 680
 681static int is_expected_rev(const struct object_id *oid)
 682{
 683        const char *filename = git_path_bisect_expected_rev();
 684        struct stat st;
 685        struct strbuf str = STRBUF_INIT;
 686        FILE *fp;
 687        int res = 0;
 688
 689        if (stat(filename, &st) || !S_ISREG(st.st_mode))
 690                return 0;
 691
 692        fp = fopen_or_warn(filename, "r");
 693        if (!fp)
 694                return 0;
 695
 696        if (strbuf_getline_lf(&str, fp) != EOF)
 697                res = !strcmp(str.buf, oid_to_hex(oid));
 698
 699        strbuf_release(&str);
 700        fclose(fp);
 701
 702        return res;
 703}
 704
 705static int bisect_checkout(const struct object_id *bisect_rev, int no_checkout)
 706{
 707        char bisect_rev_hex[GIT_MAX_HEXSZ + 1];
 708
 709        memcpy(bisect_rev_hex, oid_to_hex(bisect_rev), GIT_SHA1_HEXSZ + 1);
 710        update_ref(NULL, "BISECT_EXPECTED_REV", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
 711
 712        argv_checkout[2] = bisect_rev_hex;
 713        if (no_checkout) {
 714                update_ref(NULL, "BISECT_HEAD", bisect_rev, NULL, 0,
 715                           UPDATE_REFS_DIE_ON_ERR);
 716        } else {
 717                int res;
 718                res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
 719                if (res)
 720                        exit(res);
 721        }
 722
 723        argv_show_branch[1] = bisect_rev_hex;
 724        return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
 725}
 726
 727static struct commit *get_commit_reference(struct repository *r,
 728                                           const struct object_id *oid)
 729{
 730        struct commit *c = lookup_commit_reference(r, oid);
 731        if (!c)
 732                die(_("Not a valid commit name %s"), oid_to_hex(oid));
 733        return c;
 734}
 735
 736static struct commit **get_bad_and_good_commits(struct repository *r,
 737                                                int *rev_nr)
 738{
 739        struct commit **rev;
 740        int i, n = 0;
 741
 742        ALLOC_ARRAY(rev, 1 + good_revs.nr);
 743        rev[n++] = get_commit_reference(r, current_bad_oid);
 744        for (i = 0; i < good_revs.nr; i++)
 745                rev[n++] = get_commit_reference(r, good_revs.oid + i);
 746        *rev_nr = n;
 747
 748        return rev;
 749}
 750
 751static void handle_bad_merge_base(void)
 752{
 753        if (is_expected_rev(current_bad_oid)) {
 754                char *bad_hex = oid_to_hex(current_bad_oid);
 755                char *good_hex = join_sha1_array_hex(&good_revs, ' ');
 756                if (!strcmp(term_bad, "bad") && !strcmp(term_good, "good")) {
 757                        fprintf(stderr, _("The merge base %s is bad.\n"
 758                                "This means the bug has been fixed "
 759                                "between %s and [%s].\n"),
 760                                bad_hex, bad_hex, good_hex);
 761                } else if (!strcmp(term_bad, "new") && !strcmp(term_good, "old")) {
 762                        fprintf(stderr, _("The merge base %s is new.\n"
 763                                "The property has changed "
 764                                "between %s and [%s].\n"),
 765                                bad_hex, bad_hex, good_hex);
 766                } else {
 767                        fprintf(stderr, _("The merge base %s is %s.\n"
 768                                "This means the first '%s' commit is "
 769                                "between %s and [%s].\n"),
 770                                bad_hex, term_bad, term_good, bad_hex, good_hex);
 771                }
 772                exit(3);
 773        }
 774
 775        fprintf(stderr, _("Some %s revs are not ancestors of the %s rev.\n"
 776                "git bisect cannot work properly in this case.\n"
 777                "Maybe you mistook %s and %s revs?\n"),
 778                term_good, term_bad, term_good, term_bad);
 779        exit(1);
 780}
 781
 782static void handle_skipped_merge_base(const struct object_id *mb)
 783{
 784        char *mb_hex = oid_to_hex(mb);
 785        char *bad_hex = oid_to_hex(current_bad_oid);
 786        char *good_hex = join_sha1_array_hex(&good_revs, ' ');
 787
 788        warning(_("the merge base between %s and [%s] "
 789                "must be skipped.\n"
 790                "So we cannot be sure the first %s commit is "
 791                "between %s and %s.\n"
 792                "We continue anyway."),
 793                bad_hex, good_hex, term_bad, mb_hex, bad_hex);
 794        free(good_hex);
 795}
 796
 797/*
 798 * "check_merge_bases" checks that merge bases are not "bad" (or "new").
 799 *
 800 * - If one is "bad" (or "new"), it means the user assumed something wrong
 801 * and we must exit with a non 0 error code.
 802 * - If one is "good" (or "old"), that's good, we have nothing to do.
 803 * - If one is "skipped", we can't know but we should warn.
 804 * - If we don't know, we should check it out and ask the user to test.
 805 */
 806static void check_merge_bases(int rev_nr, struct commit **rev, int no_checkout)
 807{
 808        struct commit_list *result;
 809
 810        result = get_merge_bases_many(rev[0], rev_nr - 1, rev + 1);
 811
 812        for (; result; result = result->next) {
 813                const struct object_id *mb = &result->item->object.oid;
 814                if (oideq(mb, current_bad_oid)) {
 815                        handle_bad_merge_base();
 816                } else if (0 <= oid_array_lookup(&good_revs, mb)) {
 817                        continue;
 818                } else if (0 <= oid_array_lookup(&skipped_revs, mb)) {
 819                        handle_skipped_merge_base(mb);
 820                } else {
 821                        printf(_("Bisecting: a merge base must be tested\n"));
 822                        exit(bisect_checkout(mb, no_checkout));
 823                }
 824        }
 825
 826        free_commit_list(result);
 827}
 828
 829static int check_ancestors(struct repository *r, int rev_nr,
 830                           struct commit **rev, const char *prefix)
 831{
 832        struct rev_info revs;
 833        int res;
 834
 835        bisect_rev_setup(r, &revs, prefix, "^%s", "%s", 0);
 836
 837        bisect_common(&revs);
 838        res = (revs.commits != NULL);
 839
 840        /* Clean up objects used, as they will be reused. */
 841        clear_commit_marks_many(rev_nr, rev, ALL_REV_FLAGS);
 842
 843        return res;
 844}
 845
 846/*
 847 * "check_good_are_ancestors_of_bad" checks that all "good" revs are
 848 * ancestor of the "bad" rev.
 849 *
 850 * If that's not the case, we need to check the merge bases.
 851 * If a merge base must be tested by the user, its source code will be
 852 * checked out to be tested by the user and we will exit.
 853 */
 854static void check_good_are_ancestors_of_bad(struct repository *r,
 855                                            const char *prefix,
 856                                            int no_checkout)
 857{
 858        char *filename = git_pathdup("BISECT_ANCESTORS_OK");
 859        struct stat st;
 860        int fd, rev_nr;
 861        struct commit **rev;
 862
 863        if (!current_bad_oid)
 864                die(_("a %s revision is needed"), term_bad);
 865
 866        /* Check if file BISECT_ANCESTORS_OK exists. */
 867        if (!stat(filename, &st) && S_ISREG(st.st_mode))
 868                goto done;
 869
 870        /* Bisecting with no good rev is ok. */
 871        if (good_revs.nr == 0)
 872                goto done;
 873
 874        /* Check if all good revs are ancestor of the bad rev. */
 875        rev = get_bad_and_good_commits(r, &rev_nr);
 876        if (check_ancestors(r, rev_nr, rev, prefix))
 877                check_merge_bases(rev_nr, rev, no_checkout);
 878        free(rev);
 879
 880        /* Create file BISECT_ANCESTORS_OK. */
 881        fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
 882        if (fd < 0)
 883                warning_errno(_("could not create file '%s'"),
 884                              filename);
 885        else
 886                close(fd);
 887 done:
 888        free(filename);
 889}
 890
 891/*
 892 * This does "git diff-tree --pretty COMMIT" without one fork+exec.
 893 */
 894static void show_diff_tree(struct repository *r,
 895                           const char *prefix,
 896                           struct commit *commit)
 897{
 898        struct rev_info opt;
 899
 900        /* diff-tree init */
 901        repo_init_revisions(r, &opt, prefix);
 902        git_config(git_diff_basic_config, NULL); /* no "diff" UI options */
 903        opt.abbrev = 0;
 904        opt.diff = 1;
 905
 906        /* This is what "--pretty" does */
 907        opt.verbose_header = 1;
 908        opt.use_terminator = 0;
 909        opt.commit_format = CMIT_FMT_DEFAULT;
 910
 911        /* diff-tree init */
 912        if (!opt.diffopt.output_format)
 913                opt.diffopt.output_format = DIFF_FORMAT_RAW;
 914
 915        setup_revisions(0, NULL, &opt, NULL);
 916        log_tree_commit(&opt, commit);
 917}
 918
 919/*
 920 * The terms used for this bisect session are stored in BISECT_TERMS.
 921 * We read them and store them to adapt the messages accordingly.
 922 * Default is bad/good.
 923 */
 924void read_bisect_terms(const char **read_bad, const char **read_good)
 925{
 926        struct strbuf str = STRBUF_INIT;
 927        const char *filename = git_path_bisect_terms();
 928        FILE *fp = fopen(filename, "r");
 929
 930        if (!fp) {
 931                if (errno == ENOENT) {
 932                        *read_bad = "bad";
 933                        *read_good = "good";
 934                        return;
 935                } else {
 936                        die_errno(_("could not read file '%s'"), filename);
 937                }
 938        } else {
 939                strbuf_getline_lf(&str, fp);
 940                *read_bad = strbuf_detach(&str, NULL);
 941                strbuf_getline_lf(&str, fp);
 942                *read_good = strbuf_detach(&str, NULL);
 943        }
 944        strbuf_release(&str);
 945        fclose(fp);
 946}
 947
 948/*
 949 * We use the convention that exiting with an exit code 10 means that
 950 * the bisection process finished successfully.
 951 * In this case the calling shell script should exit 0.
 952 *
 953 * If no_checkout is non-zero, the bisection process does not
 954 * checkout the trial commit but instead simply updates BISECT_HEAD.
 955 */
 956int bisect_next_all(struct repository *r, const char *prefix, int no_checkout)
 957{
 958        struct rev_info revs;
 959        struct commit_list *tried;
 960        int reaches = 0, all = 0, nr, steps;
 961        struct object_id *bisect_rev;
 962        char *steps_msg;
 963
 964        read_bisect_terms(&term_bad, &term_good);
 965        if (read_bisect_refs())
 966                die(_("reading bisect refs failed"));
 967
 968        check_good_are_ancestors_of_bad(r, prefix, no_checkout);
 969
 970        bisect_rev_setup(r, &revs, prefix, "%s", "^%s", 1);
 971        revs.limited = 1;
 972
 973        bisect_common(&revs);
 974
 975        find_bisection(&revs.commits, &reaches, &all, !!skipped_revs.nr);
 976        revs.commits = managed_skipped(revs.commits, &tried);
 977
 978        if (!revs.commits) {
 979                /*
 980                 * We should exit here only if the "bad"
 981                 * commit is also a "skip" commit.
 982                 */
 983                exit_if_skipped_commits(tried, NULL);
 984
 985                printf(_("%s was both %s and %s\n"),
 986                       oid_to_hex(current_bad_oid),
 987                       term_good,
 988                       term_bad);
 989                exit(1);
 990        }
 991
 992        if (!all) {
 993                fprintf(stderr, _("No testable commit found.\n"
 994                        "Maybe you started with bad path parameters?\n"));
 995                exit(4);
 996        }
 997
 998        bisect_rev = &revs.commits->item->object.oid;
 999
1000        if (oideq(bisect_rev, current_bad_oid)) {
1001                exit_if_skipped_commits(tried, current_bad_oid);
1002                printf("%s is the first %s commit\n", oid_to_hex(bisect_rev),
1003                        term_bad);
1004                show_diff_tree(r, prefix, revs.commits->item);
1005                /* This means the bisection process succeeded. */
1006                exit(10);
1007        }
1008
1009        nr = all - reaches - 1;
1010        steps = estimate_bisect_steps(all);
1011
1012        steps_msg = xstrfmt(Q_("(roughly %d step)", "(roughly %d steps)",
1013                  steps), steps);
1014        /*
1015         * TRANSLATORS: the last %s will be replaced with "(roughly %d
1016         * steps)" translation.
1017         */
1018        printf(Q_("Bisecting: %d revision left to test after this %s\n",
1019                  "Bisecting: %d revisions left to test after this %s\n",
1020                  nr), nr, steps_msg);
1021        free(steps_msg);
1022
1023        return bisect_checkout(bisect_rev, no_checkout);
1024}
1025
1026static inline int log2i(int n)
1027{
1028        int log2 = 0;
1029
1030        for (; n > 1; n >>= 1)
1031                log2++;
1032
1033        return log2;
1034}
1035
1036static inline int exp2i(int n)
1037{
1038        return 1 << n;
1039}
1040
1041/*
1042 * Estimate the number of bisect steps left (after the current step)
1043 *
1044 * For any x between 0 included and 2^n excluded, the probability for
1045 * n - 1 steps left looks like:
1046 *
1047 * P(2^n + x) == (2^n - x) / (2^n + x)
1048 *
1049 * and P(2^n + x) < 0.5 means 2^n < 3x
1050 */
1051int estimate_bisect_steps(int all)
1052{
1053        int n, x, e;
1054
1055        if (all < 3)
1056                return 0;
1057
1058        n = log2i(all);
1059        e = exp2i(n);
1060        x = all - e;
1061
1062        return (e < 3 * x) ? n : n - 1;
1063}
1064
1065static int mark_for_removal(const char *refname, const struct object_id *oid,
1066                            int flag, void *cb_data)
1067{
1068        struct string_list *refs = cb_data;
1069        char *ref = xstrfmt("refs/bisect%s", refname);
1070        string_list_append(refs, ref);
1071        return 0;
1072}
1073
1074int bisect_clean_state(void)
1075{
1076        int result = 0;
1077
1078        /* There may be some refs packed during bisection */
1079        struct string_list refs_for_removal = STRING_LIST_INIT_NODUP;
1080        for_each_ref_in("refs/bisect", mark_for_removal, (void *) &refs_for_removal);
1081        string_list_append(&refs_for_removal, xstrdup("BISECT_HEAD"));
1082        result = delete_refs("bisect: remove", &refs_for_removal, REF_NO_DEREF);
1083        refs_for_removal.strdup_strings = 1;
1084        string_list_clear(&refs_for_removal, 0);
1085        unlink_or_warn(git_path_bisect_expected_rev());
1086        unlink_or_warn(git_path_bisect_ancestors_ok());
1087        unlink_or_warn(git_path_bisect_log());
1088        unlink_or_warn(git_path_bisect_names());
1089        unlink_or_warn(git_path_bisect_run());
1090        unlink_or_warn(git_path_bisect_terms());
1091        /* Cleanup head-name if it got left by an old version of git-bisect */
1092        unlink_or_warn(git_path_head_name());
1093        /*
1094         * Cleanup BISECT_START last to support the --no-checkout option
1095         * introduced in the commit 4796e823a.
1096         */
1097        unlink_or_warn(git_path_bisect_start());
1098
1099        return result;
1100}