rerere.con commit rerere: plug memory leaks upon "rerere forget" failure (8f44961)
   1#include "cache.h"
   2#include "lockfile.h"
   3#include "string-list.h"
   4#include "rerere.h"
   5#include "xdiff-interface.h"
   6#include "dir.h"
   7#include "resolve-undo.h"
   8#include "ll-merge.h"
   9#include "attr.h"
  10#include "pathspec.h"
  11#include "sha1-lookup.h"
  12
  13#define RESOLVED 0
  14#define PUNTED 1
  15#define THREE_STAGED 2
  16void *RERERE_RESOLVED = &RERERE_RESOLVED;
  17
  18/* if rerere_enabled == -1, fall back to detection of .git/rr-cache */
  19static int rerere_enabled = -1;
  20
  21/* automatically update cleanly resolved paths to the index */
  22static int rerere_autoupdate;
  23
  24static char *merge_rr_path;
  25
  26static int rerere_dir_nr;
  27static int rerere_dir_alloc;
  28
  29#define RR_HAS_POSTIMAGE 1
  30#define RR_HAS_PREIMAGE 2
  31static struct rerere_dir {
  32        unsigned char sha1[20];
  33        int status_alloc, status_nr;
  34        unsigned char *status;
  35} **rerere_dir;
  36
  37static void free_rerere_dirs(void)
  38{
  39        int i;
  40        for (i = 0; i < rerere_dir_nr; i++) {
  41                free(rerere_dir[i]->status);
  42                free(rerere_dir[i]);
  43        }
  44        free(rerere_dir);
  45        rerere_dir_nr = rerere_dir_alloc = 0;
  46        rerere_dir = NULL;
  47}
  48
  49static void free_rerere_id(struct string_list_item *item)
  50{
  51        free(item->util);
  52}
  53
  54static const char *rerere_id_hex(const struct rerere_id *id)
  55{
  56        return sha1_to_hex(id->collection->sha1);
  57}
  58
  59static void fit_variant(struct rerere_dir *rr_dir, int variant)
  60{
  61        variant++;
  62        ALLOC_GROW(rr_dir->status, variant, rr_dir->status_alloc);
  63        if (rr_dir->status_nr < variant) {
  64                memset(rr_dir->status + rr_dir->status_nr,
  65                       '\0', variant - rr_dir->status_nr);
  66                rr_dir->status_nr = variant;
  67        }
  68}
  69
  70static void assign_variant(struct rerere_id *id)
  71{
  72        int variant;
  73        struct rerere_dir *rr_dir = id->collection;
  74
  75        variant = id->variant;
  76        if (variant < 0) {
  77                for (variant = 0; variant < rr_dir->status_nr; variant++)
  78                        if (!rr_dir->status[variant])
  79                                break;
  80        }
  81        fit_variant(rr_dir, variant);
  82        id->variant = variant;
  83}
  84
  85const char *rerere_path(const struct rerere_id *id, const char *file)
  86{
  87        if (!file)
  88                return git_path("rr-cache/%s", rerere_id_hex(id));
  89
  90        if (id->variant <= 0)
  91                return git_path("rr-cache/%s/%s", rerere_id_hex(id), file);
  92
  93        return git_path("rr-cache/%s/%s.%d",
  94                        rerere_id_hex(id), file, id->variant);
  95}
  96
  97static int is_rr_file(const char *name, const char *filename, int *variant)
  98{
  99        const char *suffix;
 100        char *ep;
 101
 102        if (!strcmp(name, filename)) {
 103                *variant = 0;
 104                return 1;
 105        }
 106        if (!skip_prefix(name, filename, &suffix) || *suffix != '.')
 107                return 0;
 108
 109        errno = 0;
 110        *variant = strtol(suffix + 1, &ep, 10);
 111        if (errno || *ep)
 112                return 0;
 113        return 1;
 114}
 115
 116static void scan_rerere_dir(struct rerere_dir *rr_dir)
 117{
 118        struct dirent *de;
 119        DIR *dir = opendir(git_path("rr-cache/%s", sha1_to_hex(rr_dir->sha1)));
 120
 121        if (!dir)
 122                return;
 123        while ((de = readdir(dir)) != NULL) {
 124                int variant;
 125
 126                if (is_rr_file(de->d_name, "postimage", &variant)) {
 127                        fit_variant(rr_dir, variant);
 128                        rr_dir->status[variant] |= RR_HAS_POSTIMAGE;
 129                } else if (is_rr_file(de->d_name, "preimage", &variant)) {
 130                        fit_variant(rr_dir, variant);
 131                        rr_dir->status[variant] |= RR_HAS_PREIMAGE;
 132                }
 133        }
 134        closedir(dir);
 135}
 136
 137static const unsigned char *rerere_dir_sha1(size_t i, void *table)
 138{
 139        struct rerere_dir **rr_dir = table;
 140        return rr_dir[i]->sha1;
 141}
 142
 143static struct rerere_dir *find_rerere_dir(const char *hex)
 144{
 145        unsigned char sha1[20];
 146        struct rerere_dir *rr_dir;
 147        int pos;
 148
 149        if (get_sha1_hex(hex, sha1))
 150                return NULL; /* BUG */
 151        pos = sha1_pos(sha1, rerere_dir, rerere_dir_nr, rerere_dir_sha1);
 152        if (pos < 0) {
 153                rr_dir = xmalloc(sizeof(*rr_dir));
 154                hashcpy(rr_dir->sha1, sha1);
 155                rr_dir->status = NULL;
 156                rr_dir->status_nr = 0;
 157                rr_dir->status_alloc = 0;
 158                pos = -1 - pos;
 159
 160                /* Make sure the array is big enough ... */
 161                ALLOC_GROW(rerere_dir, rerere_dir_nr + 1, rerere_dir_alloc);
 162                /* ... and add it in. */
 163                rerere_dir_nr++;
 164                memmove(rerere_dir + pos + 1, rerere_dir + pos,
 165                        (rerere_dir_nr - pos - 1) * sizeof(*rerere_dir));
 166                rerere_dir[pos] = rr_dir;
 167                scan_rerere_dir(rr_dir);
 168        }
 169        return rerere_dir[pos];
 170}
 171
 172static int has_rerere_resolution(const struct rerere_id *id)
 173{
 174        const int both = RR_HAS_POSTIMAGE|RR_HAS_PREIMAGE;
 175        int variant = id->variant;
 176
 177        if (variant < 0)
 178                return 0;
 179        return ((id->collection->status[variant] & both) == both);
 180}
 181
 182static struct rerere_id *new_rerere_id_hex(char *hex)
 183{
 184        struct rerere_id *id = xmalloc(sizeof(*id));
 185        id->collection = find_rerere_dir(hex);
 186        id->variant = -1; /* not known yet */
 187        return id;
 188}
 189
 190static struct rerere_id *new_rerere_id(unsigned char *sha1)
 191{
 192        return new_rerere_id_hex(sha1_to_hex(sha1));
 193}
 194
 195/*
 196 * $GIT_DIR/MERGE_RR file is a collection of records, each of which is
 197 * "conflict ID", a HT and pathname, terminated with a NUL, and is
 198 * used to keep track of the set of paths that "rerere" may need to
 199 * work on (i.e. what is left by the previous invocation of "git
 200 * rerere" during the current conflict resolution session).
 201 */
 202static void read_rr(struct string_list *rr)
 203{
 204        struct strbuf buf = STRBUF_INIT;
 205        FILE *in = fopen(merge_rr_path, "r");
 206
 207        if (!in)
 208                return;
 209        while (!strbuf_getwholeline(&buf, in, '\0')) {
 210                char *path;
 211                unsigned char sha1[20];
 212                struct rerere_id *id;
 213                int variant;
 214
 215                /* There has to be the hash, tab, path and then NUL */
 216                if (buf.len < 42 || get_sha1_hex(buf.buf, sha1))
 217                        die("corrupt MERGE_RR");
 218
 219                if (buf.buf[40] != '.') {
 220                        variant = 0;
 221                        path = buf.buf + 40;
 222                } else {
 223                        errno = 0;
 224                        variant = strtol(buf.buf + 41, &path, 10);
 225                        if (errno)
 226                                die("corrupt MERGE_RR");
 227                }
 228                if (*(path++) != '\t')
 229                        die("corrupt MERGE_RR");
 230                buf.buf[40] = '\0';
 231                id = new_rerere_id_hex(buf.buf);
 232                id->variant = variant;
 233                string_list_insert(rr, path)->util = id;
 234        }
 235        strbuf_release(&buf);
 236        fclose(in);
 237}
 238
 239static struct lock_file write_lock;
 240
 241static int write_rr(struct string_list *rr, int out_fd)
 242{
 243        int i;
 244        for (i = 0; i < rr->nr; i++) {
 245                struct strbuf buf = STRBUF_INIT;
 246                struct rerere_id *id;
 247
 248                assert(rr->items[i].util != RERERE_RESOLVED);
 249
 250                id = rr->items[i].util;
 251                if (!id)
 252                        continue;
 253                assert(id->variant >= 0);
 254                if (0 < id->variant)
 255                        strbuf_addf(&buf, "%s.%d\t%s%c",
 256                                    rerere_id_hex(id), id->variant,
 257                                    rr->items[i].string, 0);
 258                else
 259                        strbuf_addf(&buf, "%s\t%s%c",
 260                                    rerere_id_hex(id),
 261                                    rr->items[i].string, 0);
 262
 263                if (write_in_full(out_fd, buf.buf, buf.len) != buf.len)
 264                        die("unable to write rerere record");
 265
 266                strbuf_release(&buf);
 267        }
 268        if (commit_lock_file(&write_lock) != 0)
 269                die("unable to write rerere record");
 270        return 0;
 271}
 272
 273/*
 274 * "rerere" interacts with conflicted file contents using this I/O
 275 * abstraction.  It reads a conflicted contents from one place via
 276 * "getline()" method, and optionally can write it out after
 277 * normalizing the conflicted hunks to the "output".  Subclasses of
 278 * rerere_io embed this structure at the beginning of their own
 279 * rerere_io object.
 280 */
 281struct rerere_io {
 282        int (*getline)(struct strbuf *, struct rerere_io *);
 283        FILE *output;
 284        int wrerror;
 285        /* some more stuff */
 286};
 287
 288static void ferr_write(const void *p, size_t count, FILE *fp, int *err)
 289{
 290        if (!count || *err)
 291                return;
 292        if (fwrite(p, count, 1, fp) != 1)
 293                *err = errno;
 294}
 295
 296static inline void ferr_puts(const char *s, FILE *fp, int *err)
 297{
 298        ferr_write(s, strlen(s), fp, err);
 299}
 300
 301static void rerere_io_putstr(const char *str, struct rerere_io *io)
 302{
 303        if (io->output)
 304                ferr_puts(str, io->output, &io->wrerror);
 305}
 306
 307/*
 308 * Write a conflict marker to io->output (if defined).
 309 */
 310static void rerere_io_putconflict(int ch, int size, struct rerere_io *io)
 311{
 312        char buf[64];
 313
 314        while (size) {
 315                if (size <= sizeof(buf) - 2) {
 316                        memset(buf, ch, size);
 317                        buf[size] = '\n';
 318                        buf[size + 1] = '\0';
 319                        size = 0;
 320                } else {
 321                        int sz = sizeof(buf) - 1;
 322
 323                        /*
 324                         * Make sure we will not write everything out
 325                         * in this round by leaving at least 1 byte
 326                         * for the next round, giving the next round
 327                         * a chance to add the terminating LF.  Yuck.
 328                         */
 329                        if (size <= sz)
 330                                sz -= (sz - size) + 1;
 331                        memset(buf, ch, sz);
 332                        buf[sz] = '\0';
 333                        size -= sz;
 334                }
 335                rerere_io_putstr(buf, io);
 336        }
 337}
 338
 339static void rerere_io_putmem(const char *mem, size_t sz, struct rerere_io *io)
 340{
 341        if (io->output)
 342                ferr_write(mem, sz, io->output, &io->wrerror);
 343}
 344
 345/*
 346 * Subclass of rerere_io that reads from an on-disk file
 347 */
 348struct rerere_io_file {
 349        struct rerere_io io;
 350        FILE *input;
 351};
 352
 353/*
 354 * ... and its getline() method implementation
 355 */
 356static int rerere_file_getline(struct strbuf *sb, struct rerere_io *io_)
 357{
 358        struct rerere_io_file *io = (struct rerere_io_file *)io_;
 359        return strbuf_getwholeline(sb, io->input, '\n');
 360}
 361
 362/*
 363 * Require the exact number of conflict marker letters, no more, no
 364 * less, followed by SP or any whitespace
 365 * (including LF).
 366 */
 367static int is_cmarker(char *buf, int marker_char, int marker_size)
 368{
 369        int want_sp;
 370
 371        /*
 372         * The beginning of our version and the end of their version
 373         * always are labeled like "<<<<< ours" or ">>>>> theirs",
 374         * hence we set want_sp for them.  Note that the version from
 375         * the common ancestor in diff3-style output is not always
 376         * labelled (e.g. "||||| common" is often seen but "|||||"
 377         * alone is also valid), so we do not set want_sp.
 378         */
 379        want_sp = (marker_char == '<') || (marker_char == '>');
 380
 381        while (marker_size--)
 382                if (*buf++ != marker_char)
 383                        return 0;
 384        if (want_sp && *buf != ' ')
 385                return 0;
 386        return isspace(*buf);
 387}
 388
 389/*
 390 * Read contents a file with conflicts, normalize the conflicts
 391 * by (1) discarding the common ancestor version in diff3-style,
 392 * (2) reordering our side and their side so that whichever sorts
 393 * alphabetically earlier comes before the other one, while
 394 * computing the "conflict ID", which is just an SHA-1 hash of
 395 * one side of the conflict, NUL, the other side of the conflict,
 396 * and NUL concatenated together.
 397 *
 398 * Return the number of conflict hunks found.
 399 *
 400 * NEEDSWORK: the logic and theory of operation behind this conflict
 401 * normalization may deserve to be documented somewhere, perhaps in
 402 * Documentation/technical/rerere.txt.
 403 */
 404static int handle_path(unsigned char *sha1, struct rerere_io *io, int marker_size)
 405{
 406        git_SHA_CTX ctx;
 407        int hunk_no = 0;
 408        enum {
 409                RR_CONTEXT = 0, RR_SIDE_1, RR_SIDE_2, RR_ORIGINAL
 410        } hunk = RR_CONTEXT;
 411        struct strbuf one = STRBUF_INIT, two = STRBUF_INIT;
 412        struct strbuf buf = STRBUF_INIT;
 413
 414        if (sha1)
 415                git_SHA1_Init(&ctx);
 416
 417        while (!io->getline(&buf, io)) {
 418                if (is_cmarker(buf.buf, '<', marker_size)) {
 419                        if (hunk != RR_CONTEXT)
 420                                goto bad;
 421                        hunk = RR_SIDE_1;
 422                } else if (is_cmarker(buf.buf, '|', marker_size)) {
 423                        if (hunk != RR_SIDE_1)
 424                                goto bad;
 425                        hunk = RR_ORIGINAL;
 426                } else if (is_cmarker(buf.buf, '=', marker_size)) {
 427                        if (hunk != RR_SIDE_1 && hunk != RR_ORIGINAL)
 428                                goto bad;
 429                        hunk = RR_SIDE_2;
 430                } else if (is_cmarker(buf.buf, '>', marker_size)) {
 431                        if (hunk != RR_SIDE_2)
 432                                goto bad;
 433                        if (strbuf_cmp(&one, &two) > 0)
 434                                strbuf_swap(&one, &two);
 435                        hunk_no++;
 436                        hunk = RR_CONTEXT;
 437                        rerere_io_putconflict('<', marker_size, io);
 438                        rerere_io_putmem(one.buf, one.len, io);
 439                        rerere_io_putconflict('=', marker_size, io);
 440                        rerere_io_putmem(two.buf, two.len, io);
 441                        rerere_io_putconflict('>', marker_size, io);
 442                        if (sha1) {
 443                                git_SHA1_Update(&ctx, one.buf ? one.buf : "",
 444                                            one.len + 1);
 445                                git_SHA1_Update(&ctx, two.buf ? two.buf : "",
 446                                            two.len + 1);
 447                        }
 448                        strbuf_reset(&one);
 449                        strbuf_reset(&two);
 450                } else if (hunk == RR_SIDE_1)
 451                        strbuf_addbuf(&one, &buf);
 452                else if (hunk == RR_ORIGINAL)
 453                        ; /* discard */
 454                else if (hunk == RR_SIDE_2)
 455                        strbuf_addbuf(&two, &buf);
 456                else
 457                        rerere_io_putstr(buf.buf, io);
 458                continue;
 459        bad:
 460                hunk = 99; /* force error exit */
 461                break;
 462        }
 463        strbuf_release(&one);
 464        strbuf_release(&two);
 465        strbuf_release(&buf);
 466
 467        if (sha1)
 468                git_SHA1_Final(sha1, &ctx);
 469        if (hunk != RR_CONTEXT)
 470                return -1;
 471        return hunk_no;
 472}
 473
 474/*
 475 * Scan the path for conflicts, do the "handle_path()" thing above, and
 476 * return the number of conflict hunks found.
 477 */
 478static int handle_file(const char *path, unsigned char *sha1, const char *output)
 479{
 480        int hunk_no = 0;
 481        struct rerere_io_file io;
 482        int marker_size = ll_merge_marker_size(path);
 483
 484        memset(&io, 0, sizeof(io));
 485        io.io.getline = rerere_file_getline;
 486        io.input = fopen(path, "r");
 487        io.io.wrerror = 0;
 488        if (!io.input)
 489                return error("Could not open %s", path);
 490
 491        if (output) {
 492                io.io.output = fopen(output, "w");
 493                if (!io.io.output) {
 494                        fclose(io.input);
 495                        return error("Could not write %s", output);
 496                }
 497        }
 498
 499        hunk_no = handle_path(sha1, (struct rerere_io *)&io, marker_size);
 500
 501        fclose(io.input);
 502        if (io.io.wrerror)
 503                error("There were errors while writing %s (%s)",
 504                      path, strerror(io.io.wrerror));
 505        if (io.io.output && fclose(io.io.output))
 506                io.io.wrerror = error("Failed to flush %s: %s",
 507                                      path, strerror(errno));
 508
 509        if (hunk_no < 0) {
 510                if (output)
 511                        unlink_or_warn(output);
 512                return error("Could not parse conflict hunks in %s", path);
 513        }
 514        if (io.io.wrerror)
 515                return -1;
 516        return hunk_no;
 517}
 518
 519/*
 520 * Look at a cache entry at "i" and see if it is not conflicting,
 521 * conflicting and we are willing to handle, or conflicting and
 522 * we are unable to handle, and return the determination in *type.
 523 * Return the cache index to be looked at next, by skipping the
 524 * stages we have already looked at in this invocation of this
 525 * function.
 526 */
 527static int check_one_conflict(int i, int *type)
 528{
 529        const struct cache_entry *e = active_cache[i];
 530
 531        if (!ce_stage(e)) {
 532                *type = RESOLVED;
 533                return i + 1;
 534        }
 535
 536        *type = PUNTED;
 537        while (ce_stage(active_cache[i]) == 1)
 538                i++;
 539
 540        /* Only handle regular files with both stages #2 and #3 */
 541        if (i + 1 < active_nr) {
 542                const struct cache_entry *e2 = active_cache[i];
 543                const struct cache_entry *e3 = active_cache[i + 1];
 544                if (ce_stage(e2) == 2 &&
 545                    ce_stage(e3) == 3 &&
 546                    ce_same_name(e, e3) &&
 547                    S_ISREG(e2->ce_mode) &&
 548                    S_ISREG(e3->ce_mode))
 549                        *type = THREE_STAGED;
 550        }
 551
 552        /* Skip the entries with the same name */
 553        while (i < active_nr && ce_same_name(e, active_cache[i]))
 554                i++;
 555        return i;
 556}
 557
 558/*
 559 * Scan the index and find paths that have conflicts that rerere can
 560 * handle, i.e. the ones that has both stages #2 and #3.
 561 *
 562 * NEEDSWORK: we do not record or replay a previous "resolve by
 563 * deletion" for a delete-modify conflict, as that is inherently risky
 564 * without knowing what modification is being discarded.  The only
 565 * safe case, i.e. both side doing the deletion and modification that
 566 * are identical to the previous round, might want to be handled,
 567 * though.
 568 */
 569static int find_conflict(struct string_list *conflict)
 570{
 571        int i;
 572        if (read_cache() < 0)
 573                return error("Could not read index");
 574
 575        for (i = 0; i < active_nr;) {
 576                int conflict_type;
 577                const struct cache_entry *e = active_cache[i];
 578                i = check_one_conflict(i, &conflict_type);
 579                if (conflict_type == THREE_STAGED)
 580                        string_list_insert(conflict, (const char *)e->name);
 581        }
 582        return 0;
 583}
 584
 585/*
 586 * The merge_rr list is meant to hold outstanding conflicted paths
 587 * that rerere could handle.  Abuse the list by adding other types of
 588 * entries to allow the caller to show "rerere remaining".
 589 *
 590 * - Conflicted paths that rerere does not handle are added
 591 * - Conflicted paths that have been resolved are marked as such
 592 *   by storing RERERE_RESOLVED to .util field (where conflict ID
 593 *   is expected to be stored).
 594 *
 595 * Do *not* write MERGE_RR file out after calling this function.
 596 *
 597 * NEEDSWORK: we may want to fix the caller that implements "rerere
 598 * remaining" to do this without abusing merge_rr.
 599 */
 600int rerere_remaining(struct string_list *merge_rr)
 601{
 602        int i;
 603        if (read_cache() < 0)
 604                return error("Could not read index");
 605
 606        for (i = 0; i < active_nr;) {
 607                int conflict_type;
 608                const struct cache_entry *e = active_cache[i];
 609                i = check_one_conflict(i, &conflict_type);
 610                if (conflict_type == PUNTED)
 611                        string_list_insert(merge_rr, (const char *)e->name);
 612                else if (conflict_type == RESOLVED) {
 613                        struct string_list_item *it;
 614                        it = string_list_lookup(merge_rr, (const char *)e->name);
 615                        if (it != NULL) {
 616                                free_rerere_id(it);
 617                                it->util = RERERE_RESOLVED;
 618                        }
 619                }
 620        }
 621        return 0;
 622}
 623
 624/*
 625 * Try using the given conflict resolution "ID" to see
 626 * if that recorded conflict resolves cleanly what we
 627 * got in the "cur".
 628 */
 629static int try_merge(const struct rerere_id *id, const char *path,
 630                     mmfile_t *cur, mmbuffer_t *result)
 631{
 632        int ret;
 633        mmfile_t base = {NULL, 0}, other = {NULL, 0};
 634
 635        if (read_mmfile(&base, rerere_path(id, "preimage")) ||
 636            read_mmfile(&other, rerere_path(id, "postimage")))
 637                ret = 1;
 638        else
 639                /*
 640                 * A three-way merge. Note that this honors user-customizable
 641                 * low-level merge driver settings.
 642                 */
 643                ret = ll_merge(result, path, &base, NULL, cur, "", &other, "", NULL);
 644
 645        free(base.ptr);
 646        free(other.ptr);
 647
 648        return ret;
 649}
 650
 651/*
 652 * Find the conflict identified by "id"; the change between its
 653 * "preimage" (i.e. a previous contents with conflict markers) and its
 654 * "postimage" (i.e. the corresponding contents with conflicts
 655 * resolved) may apply cleanly to the contents stored in "path", i.e.
 656 * the conflict this time around.
 657 *
 658 * Returns 0 for successful replay of recorded resolution, or non-zero
 659 * for failure.
 660 */
 661static int merge(const struct rerere_id *id, const char *path)
 662{
 663        FILE *f;
 664        int ret;
 665        mmfile_t cur = {NULL, 0};
 666        mmbuffer_t result = {NULL, 0};
 667
 668        /*
 669         * Normalize the conflicts in path and write it out to
 670         * "thisimage" temporary file.
 671         */
 672        if ((handle_file(path, NULL, rerere_path(id, "thisimage")) < 0) ||
 673            read_mmfile(&cur, rerere_path(id, "thisimage"))) {
 674                ret = 1;
 675                goto out;
 676        }
 677
 678        ret = try_merge(id, path, &cur, &result);
 679        if (ret)
 680                goto out;
 681
 682        /*
 683         * A successful replay of recorded resolution.
 684         * Mark that "postimage" was used to help gc.
 685         */
 686        if (utime(rerere_path(id, "postimage"), NULL) < 0)
 687                warning("failed utime() on %s: %s",
 688                        rerere_path(id, "postimage"),
 689                        strerror(errno));
 690
 691        /* Update "path" with the resolution */
 692        f = fopen(path, "w");
 693        if (!f)
 694                return error("Could not open %s: %s", path,
 695                             strerror(errno));
 696        if (fwrite(result.ptr, result.size, 1, f) != 1)
 697                error("Could not write %s: %s", path, strerror(errno));
 698        if (fclose(f))
 699                return error("Writing %s failed: %s", path,
 700                             strerror(errno));
 701
 702out:
 703        free(cur.ptr);
 704        free(result.ptr);
 705
 706        return ret;
 707}
 708
 709static struct lock_file index_lock;
 710
 711static void update_paths(struct string_list *update)
 712{
 713        int i;
 714
 715        hold_locked_index(&index_lock, 1);
 716
 717        for (i = 0; i < update->nr; i++) {
 718                struct string_list_item *item = &update->items[i];
 719                if (add_file_to_cache(item->string, 0))
 720                        exit(128);
 721                fprintf(stderr, "Staged '%s' using previous resolution.\n",
 722                        item->string);
 723        }
 724
 725        if (active_cache_changed) {
 726                if (write_locked_index(&the_index, &index_lock, COMMIT_LOCK))
 727                        die("Unable to write new index file");
 728        } else
 729                rollback_lock_file(&index_lock);
 730}
 731
 732static void remove_variant(struct rerere_id *id)
 733{
 734        unlink_or_warn(rerere_path(id, "postimage"));
 735        unlink_or_warn(rerere_path(id, "preimage"));
 736        id->collection->status[id->variant] = 0;
 737}
 738
 739/*
 740 * The path indicated by rr_item may still have conflict for which we
 741 * have a recorded resolution, in which case replay it and optionally
 742 * update it.  Or it may have been resolved by the user and we may
 743 * only have the preimage for that conflict, in which case the result
 744 * needs to be recorded as a resolution in a postimage file.
 745 */
 746static void do_rerere_one_path(struct string_list_item *rr_item,
 747                               struct string_list *update)
 748{
 749        const char *path = rr_item->string;
 750        struct rerere_id *id = rr_item->util;
 751        struct rerere_dir *rr_dir = id->collection;
 752        int variant;
 753
 754        variant = id->variant;
 755
 756        /* Has the user resolved it already? */
 757        if (variant >= 0) {
 758                if (!handle_file(path, NULL, NULL)) {
 759                        copy_file(rerere_path(id, "postimage"), path, 0666);
 760                        id->collection->status[variant] |= RR_HAS_POSTIMAGE;
 761                        fprintf(stderr, "Recorded resolution for '%s'.\n", path);
 762                        free_rerere_id(rr_item);
 763                        rr_item->util = NULL;
 764                        return;
 765                }
 766                /*
 767                 * There may be other variants that can cleanly
 768                 * replay.  Try them and update the variant number for
 769                 * this one.
 770                 */
 771        }
 772
 773        /* Does any existing resolution apply cleanly? */
 774        for (variant = 0; variant < rr_dir->status_nr; variant++) {
 775                const int both = RR_HAS_PREIMAGE | RR_HAS_POSTIMAGE;
 776                struct rerere_id vid = *id;
 777
 778                if ((rr_dir->status[variant] & both) != both)
 779                        continue;
 780
 781                vid.variant = variant;
 782                if (merge(&vid, path))
 783                        continue; /* failed to replay */
 784
 785                /*
 786                 * If there already is a different variant that applies
 787                 * cleanly, there is no point maintaining our own variant.
 788                 */
 789                if (0 <= id->variant && id->variant != variant)
 790                        remove_variant(id);
 791
 792                if (rerere_autoupdate)
 793                        string_list_insert(update, path);
 794                else
 795                        fprintf(stderr,
 796                                "Resolved '%s' using previous resolution.\n",
 797                                path);
 798                free_rerere_id(rr_item);
 799                rr_item->util = NULL;
 800                return;
 801        }
 802
 803        /* None of the existing one applies; we need a new variant */
 804        assign_variant(id);
 805
 806        variant = id->variant;
 807        handle_file(path, NULL, rerere_path(id, "preimage"));
 808        if (id->collection->status[variant] & RR_HAS_POSTIMAGE) {
 809                const char *path = rerere_path(id, "postimage");
 810                if (unlink(path))
 811                        die_errno("cannot unlink stray '%s'", path);
 812                id->collection->status[variant] &= ~RR_HAS_POSTIMAGE;
 813        }
 814        id->collection->status[variant] |= RR_HAS_PREIMAGE;
 815        fprintf(stderr, "Recorded preimage for '%s'\n", path);
 816}
 817
 818static int do_plain_rerere(struct string_list *rr, int fd)
 819{
 820        struct string_list conflict = STRING_LIST_INIT_DUP;
 821        struct string_list update = STRING_LIST_INIT_DUP;
 822        int i;
 823
 824        find_conflict(&conflict);
 825
 826        /*
 827         * MERGE_RR records paths with conflicts immediately after
 828         * merge failed.  Some of the conflicted paths might have been
 829         * hand resolved in the working tree since then, but the
 830         * initial run would catch all and register their preimages.
 831         */
 832        for (i = 0; i < conflict.nr; i++) {
 833                struct rerere_id *id;
 834                unsigned char sha1[20];
 835                const char *path = conflict.items[i].string;
 836                int ret;
 837
 838                if (string_list_has_string(rr, path))
 839                        continue;
 840
 841                /*
 842                 * Ask handle_file() to scan and assign a
 843                 * conflict ID.  No need to write anything out
 844                 * yet.
 845                 */
 846                ret = handle_file(path, sha1, NULL);
 847                if (ret < 1)
 848                        continue;
 849
 850                id = new_rerere_id(sha1);
 851                string_list_insert(rr, path)->util = id;
 852
 853                /* Ensure that the directory exists. */
 854                mkdir_in_gitdir(rerere_path(id, NULL));
 855        }
 856
 857        for (i = 0; i < rr->nr; i++)
 858                do_rerere_one_path(&rr->items[i], &update);
 859
 860        if (update.nr)
 861                update_paths(&update);
 862
 863        return write_rr(rr, fd);
 864}
 865
 866static void git_rerere_config(void)
 867{
 868        git_config_get_bool("rerere.enabled", &rerere_enabled);
 869        git_config_get_bool("rerere.autoupdate", &rerere_autoupdate);
 870        git_config(git_default_config, NULL);
 871}
 872
 873static int is_rerere_enabled(void)
 874{
 875        const char *rr_cache;
 876        int rr_cache_exists;
 877
 878        if (!rerere_enabled)
 879                return 0;
 880
 881        rr_cache = git_path("rr-cache");
 882        rr_cache_exists = is_directory(rr_cache);
 883        if (rerere_enabled < 0)
 884                return rr_cache_exists;
 885
 886        if (!rr_cache_exists && mkdir_in_gitdir(rr_cache))
 887                die("Could not create directory %s", rr_cache);
 888        return 1;
 889}
 890
 891int setup_rerere(struct string_list *merge_rr, int flags)
 892{
 893        int fd;
 894
 895        git_rerere_config();
 896        if (!is_rerere_enabled())
 897                return -1;
 898
 899        if (flags & (RERERE_AUTOUPDATE|RERERE_NOAUTOUPDATE))
 900                rerere_autoupdate = !!(flags & RERERE_AUTOUPDATE);
 901        merge_rr_path = git_pathdup("MERGE_RR");
 902        fd = hold_lock_file_for_update(&write_lock, merge_rr_path,
 903                                       LOCK_DIE_ON_ERROR);
 904        read_rr(merge_rr);
 905        return fd;
 906}
 907
 908/*
 909 * The main entry point that is called internally from codepaths that
 910 * perform mergy operations, possibly leaving conflicted index entries
 911 * and working tree files.
 912 */
 913int rerere(int flags)
 914{
 915        struct string_list merge_rr = STRING_LIST_INIT_DUP;
 916        int fd, status;
 917
 918        fd = setup_rerere(&merge_rr, flags);
 919        if (fd < 0)
 920                return 0;
 921        status = do_plain_rerere(&merge_rr, fd);
 922        free_rerere_dirs();
 923        return status;
 924}
 925
 926/*
 927 * Subclass of rerere_io that reads from an in-core buffer that is a
 928 * strbuf
 929 */
 930struct rerere_io_mem {
 931        struct rerere_io io;
 932        struct strbuf input;
 933};
 934
 935/*
 936 * ... and its getline() method implementation
 937 */
 938static int rerere_mem_getline(struct strbuf *sb, struct rerere_io *io_)
 939{
 940        struct rerere_io_mem *io = (struct rerere_io_mem *)io_;
 941        char *ep;
 942        size_t len;
 943
 944        strbuf_release(sb);
 945        if (!io->input.len)
 946                return -1;
 947        ep = memchr(io->input.buf, '\n', io->input.len);
 948        if (!ep)
 949                ep = io->input.buf + io->input.len;
 950        else if (*ep == '\n')
 951                ep++;
 952        len = ep - io->input.buf;
 953        strbuf_add(sb, io->input.buf, len);
 954        strbuf_remove(&io->input, 0, len);
 955        return 0;
 956}
 957
 958static int handle_cache(const char *path, unsigned char *sha1, const char *output)
 959{
 960        mmfile_t mmfile[3] = {{NULL}};
 961        mmbuffer_t result = {NULL, 0};
 962        const struct cache_entry *ce;
 963        int pos, len, i, hunk_no;
 964        struct rerere_io_mem io;
 965        int marker_size = ll_merge_marker_size(path);
 966
 967        /*
 968         * Reproduce the conflicted merge in-core
 969         */
 970        len = strlen(path);
 971        pos = cache_name_pos(path, len);
 972        if (0 <= pos)
 973                return -1;
 974        pos = -pos - 1;
 975
 976        while (pos < active_nr) {
 977                enum object_type type;
 978                unsigned long size;
 979
 980                ce = active_cache[pos++];
 981                if (ce_namelen(ce) != len || memcmp(ce->name, path, len))
 982                        break;
 983                i = ce_stage(ce) - 1;
 984                if (!mmfile[i].ptr) {
 985                        mmfile[i].ptr = read_sha1_file(ce->sha1, &type, &size);
 986                        mmfile[i].size = size;
 987                }
 988        }
 989        for (i = 0; i < 3; i++)
 990                if (!mmfile[i].ptr && !mmfile[i].size)
 991                        mmfile[i].ptr = xstrdup("");
 992
 993        /*
 994         * NEEDSWORK: handle conflicts from merges with
 995         * merge.renormalize set, too?
 996         */
 997        ll_merge(&result, path, &mmfile[0], NULL,
 998                 &mmfile[1], "ours",
 999                 &mmfile[2], "theirs", NULL);
1000        for (i = 0; i < 3; i++)
1001                free(mmfile[i].ptr);
1002
1003        memset(&io, 0, sizeof(io));
1004        io.io.getline = rerere_mem_getline;
1005        if (output)
1006                io.io.output = fopen(output, "w");
1007        else
1008                io.io.output = NULL;
1009        strbuf_init(&io.input, 0);
1010        strbuf_attach(&io.input, result.ptr, result.size, result.size);
1011
1012        /*
1013         * Grab the conflict ID and optionally write the original
1014         * contents with conflict markers out.
1015         */
1016        hunk_no = handle_path(sha1, (struct rerere_io *)&io, marker_size);
1017        strbuf_release(&io.input);
1018        if (io.io.output)
1019                fclose(io.io.output);
1020        return hunk_no;
1021}
1022
1023static int rerere_forget_one_path(const char *path, struct string_list *rr)
1024{
1025        const char *filename;
1026        struct rerere_id *id;
1027        unsigned char sha1[20];
1028        int ret;
1029        struct string_list_item *item;
1030
1031        /*
1032         * Recreate the original conflict from the stages in the
1033         * index and compute the conflict ID
1034         */
1035        ret = handle_cache(path, sha1, NULL);
1036        if (ret < 1)
1037                return error("Could not parse conflict hunks in '%s'", path);
1038
1039        /* Nuke the recorded resolution for the conflict */
1040        id = new_rerere_id(sha1);
1041
1042        for (id->variant = 0;
1043             id->variant < id->collection->status_nr;
1044             id->variant++) {
1045                mmfile_t cur = { NULL, 0 };
1046                mmbuffer_t result = {NULL, 0};
1047                int cleanly_resolved;
1048
1049                if (!has_rerere_resolution(id))
1050                        continue;
1051
1052                handle_cache(path, sha1, rerere_path(id, "thisimage"));
1053                if (read_mmfile(&cur, rerere_path(id, "thisimage"))) {
1054                        free(cur.ptr);
1055                        error("Failed to update conflicted state in '%s'", path);
1056                        goto fail_exit;
1057                }
1058                cleanly_resolved = !try_merge(id, path, &cur, &result);
1059                free(result.ptr);
1060                free(cur.ptr);
1061                if (cleanly_resolved)
1062                        break;
1063        }
1064
1065        if (id->collection->status_nr <= id->variant) {
1066                error("no remembered resolution for '%s'", path);
1067                goto fail_exit;
1068        }
1069
1070        filename = rerere_path(id, "postimage");
1071        if (unlink(filename)) {
1072                if (errno == ENOENT)
1073                        error("no remembered resolution for %s", path);
1074                else
1075                        error("cannot unlink %s: %s", filename, strerror(errno));
1076                goto fail_exit;
1077        };
1078
1079        /*
1080         * Update the preimage so that the user can resolve the
1081         * conflict in the working tree, run us again to record
1082         * the postimage.
1083         */
1084        handle_cache(path, sha1, rerere_path(id, "preimage"));
1085        fprintf(stderr, "Updated preimage for '%s'\n", path);
1086
1087        /*
1088         * And remember that we can record resolution for this
1089         * conflict when the user is done.
1090         */
1091        item = string_list_insert(rr, path);
1092        free_rerere_id(item);
1093        item->util = id;
1094        fprintf(stderr, "Forgot resolution for %s\n", path);
1095        return 0;
1096
1097fail_exit:
1098        free(id);
1099        return -1;
1100}
1101
1102int rerere_forget(struct pathspec *pathspec)
1103{
1104        int i, fd;
1105        struct string_list conflict = STRING_LIST_INIT_DUP;
1106        struct string_list merge_rr = STRING_LIST_INIT_DUP;
1107
1108        if (read_cache() < 0)
1109                return error("Could not read index");
1110
1111        fd = setup_rerere(&merge_rr, RERERE_NOAUTOUPDATE);
1112
1113        /*
1114         * The paths may have been resolved (incorrectly);
1115         * recover the original conflicted state and then
1116         * find the conflicted paths.
1117         */
1118        unmerge_cache(pathspec);
1119        find_conflict(&conflict);
1120        for (i = 0; i < conflict.nr; i++) {
1121                struct string_list_item *it = &conflict.items[i];
1122                if (!match_pathspec(pathspec, it->string,
1123                                    strlen(it->string), 0, NULL, 0))
1124                        continue;
1125                rerere_forget_one_path(it->string, &merge_rr);
1126        }
1127        return write_rr(&merge_rr, fd);
1128}
1129
1130/*
1131 * Garbage collection support
1132 */
1133
1134static time_t rerere_created_at(struct rerere_id *id)
1135{
1136        struct stat st;
1137
1138        return stat(rerere_path(id, "preimage"), &st) ? (time_t) 0 : st.st_mtime;
1139}
1140
1141static time_t rerere_last_used_at(struct rerere_id *id)
1142{
1143        struct stat st;
1144
1145        return stat(rerere_path(id, "postimage"), &st) ? (time_t) 0 : st.st_mtime;
1146}
1147
1148/*
1149 * Remove the recorded resolution for a given conflict ID
1150 */
1151static void unlink_rr_item(struct rerere_id *id)
1152{
1153        unlink_or_warn(rerere_path(id, "thisimage"));
1154        remove_variant(id);
1155        id->collection->status[id->variant] = 0;
1156}
1157
1158static void prune_one(struct rerere_id *id, time_t now,
1159                      int cutoff_resolve, int cutoff_noresolve)
1160{
1161        time_t then;
1162        int cutoff;
1163
1164        then = rerere_last_used_at(id);
1165        if (then)
1166                cutoff = cutoff_resolve;
1167        else {
1168                then = rerere_created_at(id);
1169                if (!then)
1170                        return;
1171                cutoff = cutoff_noresolve;
1172        }
1173        if (then < now - cutoff * 86400)
1174                unlink_rr_item(id);
1175}
1176
1177void rerere_gc(struct string_list *rr)
1178{
1179        struct string_list to_remove = STRING_LIST_INIT_DUP;
1180        DIR *dir;
1181        struct dirent *e;
1182        int i;
1183        time_t now = time(NULL);
1184        int cutoff_noresolve = 15;
1185        int cutoff_resolve = 60;
1186
1187        git_config_get_int("gc.rerereresolved", &cutoff_resolve);
1188        git_config_get_int("gc.rerereunresolved", &cutoff_noresolve);
1189        git_config(git_default_config, NULL);
1190        dir = opendir(git_path("rr-cache"));
1191        if (!dir)
1192                die_errno("unable to open rr-cache directory");
1193        /* Collect stale conflict IDs ... */
1194        while ((e = readdir(dir))) {
1195                struct rerere_dir *rr_dir;
1196                struct rerere_id id;
1197                int now_empty;
1198
1199                if (is_dot_or_dotdot(e->d_name))
1200                        continue;
1201                rr_dir = find_rerere_dir(e->d_name);
1202                if (!rr_dir)
1203                        continue; /* or should we remove e->d_name? */
1204
1205                now_empty = 1;
1206                for (id.variant = 0, id.collection = rr_dir;
1207                     id.variant < id.collection->status_nr;
1208                     id.variant++) {
1209                        prune_one(&id, now, cutoff_resolve, cutoff_noresolve);
1210                        if (id.collection->status[id.variant])
1211                                now_empty = 0;
1212                }
1213                if (now_empty)
1214                        string_list_append(&to_remove, e->d_name);
1215        }
1216        closedir(dir);
1217
1218        /* ... and then remove the empty directories */
1219        for (i = 0; i < to_remove.nr; i++)
1220                rmdir(git_path("rr-cache/%s", to_remove.items[i].string));
1221        string_list_clear(&to_remove, 0);
1222}
1223
1224/*
1225 * During a conflict resolution, after "rerere" recorded the
1226 * preimages, abandon them if the user did not resolve them or
1227 * record their resolutions.  And drop $GIT_DIR/MERGE_RR.
1228 *
1229 * NEEDSWORK: shouldn't we be calling this from "reset --hard"?
1230 */
1231void rerere_clear(struct string_list *merge_rr)
1232{
1233        int i;
1234
1235        for (i = 0; i < merge_rr->nr; i++) {
1236                struct rerere_id *id = merge_rr->items[i].util;
1237                if (!has_rerere_resolution(id)) {
1238                        unlink_rr_item(id);
1239                        rmdir(rerere_path(id, NULL));
1240                }
1241        }
1242        unlink_or_warn(git_path("MERGE_RR"));
1243}