run-command.con commit run-command: prepare child environment before forking (ae25394)
   1#include "cache.h"
   2#include "run-command.h"
   3#include "exec_cmd.h"
   4#include "sigchain.h"
   5#include "argv-array.h"
   6#include "thread-utils.h"
   7#include "strbuf.h"
   8
   9void child_process_init(struct child_process *child)
  10{
  11        memset(child, 0, sizeof(*child));
  12        argv_array_init(&child->args);
  13        argv_array_init(&child->env_array);
  14}
  15
  16void child_process_clear(struct child_process *child)
  17{
  18        argv_array_clear(&child->args);
  19        argv_array_clear(&child->env_array);
  20}
  21
  22struct child_to_clean {
  23        pid_t pid;
  24        struct child_process *process;
  25        struct child_to_clean *next;
  26};
  27static struct child_to_clean *children_to_clean;
  28static int installed_child_cleanup_handler;
  29
  30static void cleanup_children(int sig, int in_signal)
  31{
  32        struct child_to_clean *children_to_wait_for = NULL;
  33
  34        while (children_to_clean) {
  35                struct child_to_clean *p = children_to_clean;
  36                children_to_clean = p->next;
  37
  38                if (p->process && !in_signal) {
  39                        struct child_process *process = p->process;
  40                        if (process->clean_on_exit_handler) {
  41                                trace_printf(
  42                                        "trace: run_command: running exit handler for pid %"
  43                                        PRIuMAX, (uintmax_t)p->pid
  44                                );
  45                                process->clean_on_exit_handler(process);
  46                        }
  47                }
  48
  49                kill(p->pid, sig);
  50
  51                if (p->process && p->process->wait_after_clean) {
  52                        p->next = children_to_wait_for;
  53                        children_to_wait_for = p;
  54                } else {
  55                        if (!in_signal)
  56                                free(p);
  57                }
  58        }
  59
  60        while (children_to_wait_for) {
  61                struct child_to_clean *p = children_to_wait_for;
  62                children_to_wait_for = p->next;
  63
  64                while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
  65                        ; /* spin waiting for process exit or error */
  66
  67                if (!in_signal)
  68                        free(p);
  69        }
  70}
  71
  72static void cleanup_children_on_signal(int sig)
  73{
  74        cleanup_children(sig, 1);
  75        sigchain_pop(sig);
  76        raise(sig);
  77}
  78
  79static void cleanup_children_on_exit(void)
  80{
  81        cleanup_children(SIGTERM, 0);
  82}
  83
  84static void mark_child_for_cleanup(pid_t pid, struct child_process *process)
  85{
  86        struct child_to_clean *p = xmalloc(sizeof(*p));
  87        p->pid = pid;
  88        p->process = process;
  89        p->next = children_to_clean;
  90        children_to_clean = p;
  91
  92        if (!installed_child_cleanup_handler) {
  93                atexit(cleanup_children_on_exit);
  94                sigchain_push_common(cleanup_children_on_signal);
  95                installed_child_cleanup_handler = 1;
  96        }
  97}
  98
  99static void clear_child_for_cleanup(pid_t pid)
 100{
 101        struct child_to_clean **pp;
 102
 103        for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
 104                struct child_to_clean *clean_me = *pp;
 105
 106                if (clean_me->pid == pid) {
 107                        *pp = clean_me->next;
 108                        free(clean_me);
 109                        return;
 110                }
 111        }
 112}
 113
 114static inline void close_pair(int fd[2])
 115{
 116        close(fd[0]);
 117        close(fd[1]);
 118}
 119
 120#ifndef GIT_WINDOWS_NATIVE
 121static inline void dup_devnull(int to)
 122{
 123        int fd = open("/dev/null", O_RDWR);
 124        if (fd < 0)
 125                die_errno(_("open /dev/null failed"));
 126        if (dup2(fd, to) < 0)
 127                die_errno(_("dup2(%d,%d) failed"), fd, to);
 128        close(fd);
 129}
 130#endif
 131
 132static char *locate_in_PATH(const char *file)
 133{
 134        const char *p = getenv("PATH");
 135        struct strbuf buf = STRBUF_INIT;
 136
 137        if (!p || !*p)
 138                return NULL;
 139
 140        while (1) {
 141                const char *end = strchrnul(p, ':');
 142
 143                strbuf_reset(&buf);
 144
 145                /* POSIX specifies an empty entry as the current directory. */
 146                if (end != p) {
 147                        strbuf_add(&buf, p, end - p);
 148                        strbuf_addch(&buf, '/');
 149                }
 150                strbuf_addstr(&buf, file);
 151
 152                if (!access(buf.buf, F_OK))
 153                        return strbuf_detach(&buf, NULL);
 154
 155                if (!*end)
 156                        break;
 157                p = end + 1;
 158        }
 159
 160        strbuf_release(&buf);
 161        return NULL;
 162}
 163
 164static int exists_in_PATH(const char *file)
 165{
 166        char *r = locate_in_PATH(file);
 167        free(r);
 168        return r != NULL;
 169}
 170
 171int sane_execvp(const char *file, char * const argv[])
 172{
 173        if (!execvp(file, argv))
 174                return 0; /* cannot happen ;-) */
 175
 176        /*
 177         * When a command can't be found because one of the directories
 178         * listed in $PATH is unsearchable, execvp reports EACCES, but
 179         * careful usability testing (read: analysis of occasional bug
 180         * reports) reveals that "No such file or directory" is more
 181         * intuitive.
 182         *
 183         * We avoid commands with "/", because execvp will not do $PATH
 184         * lookups in that case.
 185         *
 186         * The reassignment of EACCES to errno looks like a no-op below,
 187         * but we need to protect against exists_in_PATH overwriting errno.
 188         */
 189        if (errno == EACCES && !strchr(file, '/'))
 190                errno = exists_in_PATH(file) ? EACCES : ENOENT;
 191        else if (errno == ENOTDIR && !strchr(file, '/'))
 192                errno = ENOENT;
 193        return -1;
 194}
 195
 196static const char **prepare_shell_cmd(struct argv_array *out, const char **argv)
 197{
 198        if (!argv[0])
 199                die("BUG: shell command is empty");
 200
 201        if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
 202#ifndef GIT_WINDOWS_NATIVE
 203                argv_array_push(out, SHELL_PATH);
 204#else
 205                argv_array_push(out, "sh");
 206#endif
 207                argv_array_push(out, "-c");
 208
 209                /*
 210                 * If we have no extra arguments, we do not even need to
 211                 * bother with the "$@" magic.
 212                 */
 213                if (!argv[1])
 214                        argv_array_push(out, argv[0]);
 215                else
 216                        argv_array_pushf(out, "%s \"$@\"", argv[0]);
 217        }
 218
 219        argv_array_pushv(out, argv);
 220        return out->argv;
 221}
 222
 223#ifndef GIT_WINDOWS_NATIVE
 224static int child_notifier = -1;
 225
 226static void notify_parent(void)
 227{
 228        /*
 229         * execvp failed.  If possible, we'd like to let start_command
 230         * know, so failures like ENOENT can be handled right away; but
 231         * otherwise, finish_command will still report the error.
 232         */
 233        xwrite(child_notifier, "", 1);
 234}
 235
 236static void prepare_cmd(struct argv_array *out, const struct child_process *cmd)
 237{
 238        if (!cmd->argv[0])
 239                die("BUG: command is empty");
 240
 241        /*
 242         * Add SHELL_PATH so in the event exec fails with ENOEXEC we can
 243         * attempt to interpret the command with 'sh'.
 244         */
 245        argv_array_push(out, SHELL_PATH);
 246
 247        if (cmd->git_cmd) {
 248                argv_array_push(out, "git");
 249                argv_array_pushv(out, cmd->argv);
 250        } else if (cmd->use_shell) {
 251                prepare_shell_cmd(out, cmd->argv);
 252        } else {
 253                argv_array_pushv(out, cmd->argv);
 254        }
 255
 256        /*
 257         * If there are no '/' characters in the command then perform a path
 258         * lookup and use the resolved path as the command to exec.  If there
 259         * are no '/' characters or if the command wasn't found in the path,
 260         * have exec attempt to invoke the command directly.
 261         */
 262        if (!strchr(out->argv[1], '/')) {
 263                char *program = locate_in_PATH(out->argv[1]);
 264                if (program) {
 265                        free((char *)out->argv[1]);
 266                        out->argv[1] = program;
 267                }
 268        }
 269}
 270
 271static char **prep_childenv(const char *const *deltaenv)
 272{
 273        extern char **environ;
 274        char **childenv;
 275        struct string_list env = STRING_LIST_INIT_DUP;
 276        struct strbuf key = STRBUF_INIT;
 277        const char *const *p;
 278        int i;
 279
 280        /* Construct a sorted string list consisting of the current environ */
 281        for (p = (const char *const *) environ; p && *p; p++) {
 282                const char *equals = strchr(*p, '=');
 283
 284                if (equals) {
 285                        strbuf_reset(&key);
 286                        strbuf_add(&key, *p, equals - *p);
 287                        string_list_append(&env, key.buf)->util = (void *) *p;
 288                } else {
 289                        string_list_append(&env, *p)->util = (void *) *p;
 290                }
 291        }
 292        string_list_sort(&env);
 293
 294        /* Merge in 'deltaenv' with the current environ */
 295        for (p = deltaenv; p && *p; p++) {
 296                const char *equals = strchr(*p, '=');
 297
 298                if (equals) {
 299                        /* ('key=value'), insert or replace entry */
 300                        strbuf_reset(&key);
 301                        strbuf_add(&key, *p, equals - *p);
 302                        string_list_insert(&env, key.buf)->util = (void *) *p;
 303                } else {
 304                        /* otherwise ('key') remove existing entry */
 305                        string_list_remove(&env, *p, 0);
 306                }
 307        }
 308
 309        /* Create an array of 'char *' to be used as the childenv */
 310        childenv = xmalloc((env.nr + 1) * sizeof(char *));
 311        for (i = 0; i < env.nr; i++)
 312                childenv[i] = env.items[i].util;
 313        childenv[env.nr] = NULL;
 314
 315        string_list_clear(&env, 0);
 316        strbuf_release(&key);
 317        return childenv;
 318}
 319#endif
 320
 321static inline void set_cloexec(int fd)
 322{
 323        int flags = fcntl(fd, F_GETFD);
 324        if (flags >= 0)
 325                fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
 326}
 327
 328static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
 329{
 330        int status, code = -1;
 331        pid_t waiting;
 332        int failed_errno = 0;
 333
 334        while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
 335                ;       /* nothing */
 336        if (in_signal)
 337                return 0;
 338
 339        if (waiting < 0) {
 340                failed_errno = errno;
 341                error_errno("waitpid for %s failed", argv0);
 342        } else if (waiting != pid) {
 343                error("waitpid is confused (%s)", argv0);
 344        } else if (WIFSIGNALED(status)) {
 345                code = WTERMSIG(status);
 346                if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
 347                        error("%s died of signal %d", argv0, code);
 348                /*
 349                 * This return value is chosen so that code & 0xff
 350                 * mimics the exit code that a POSIX shell would report for
 351                 * a program that died from this signal.
 352                 */
 353                code += 128;
 354        } else if (WIFEXITED(status)) {
 355                code = WEXITSTATUS(status);
 356                /*
 357                 * Convert special exit code when execvp failed.
 358                 */
 359                if (code == 127) {
 360                        code = -1;
 361                        failed_errno = ENOENT;
 362                }
 363        } else {
 364                error("waitpid is confused (%s)", argv0);
 365        }
 366
 367        clear_child_for_cleanup(pid);
 368
 369        errno = failed_errno;
 370        return code;
 371}
 372
 373int start_command(struct child_process *cmd)
 374{
 375        int need_in, need_out, need_err;
 376        int fdin[2], fdout[2], fderr[2];
 377        int failed_errno;
 378        char *str;
 379
 380        if (!cmd->argv)
 381                cmd->argv = cmd->args.argv;
 382        if (!cmd->env)
 383                cmd->env = cmd->env_array.argv;
 384
 385        /*
 386         * In case of errors we must keep the promise to close FDs
 387         * that have been passed in via ->in and ->out.
 388         */
 389
 390        need_in = !cmd->no_stdin && cmd->in < 0;
 391        if (need_in) {
 392                if (pipe(fdin) < 0) {
 393                        failed_errno = errno;
 394                        if (cmd->out > 0)
 395                                close(cmd->out);
 396                        str = "standard input";
 397                        goto fail_pipe;
 398                }
 399                cmd->in = fdin[1];
 400        }
 401
 402        need_out = !cmd->no_stdout
 403                && !cmd->stdout_to_stderr
 404                && cmd->out < 0;
 405        if (need_out) {
 406                if (pipe(fdout) < 0) {
 407                        failed_errno = errno;
 408                        if (need_in)
 409                                close_pair(fdin);
 410                        else if (cmd->in)
 411                                close(cmd->in);
 412                        str = "standard output";
 413                        goto fail_pipe;
 414                }
 415                cmd->out = fdout[0];
 416        }
 417
 418        need_err = !cmd->no_stderr && cmd->err < 0;
 419        if (need_err) {
 420                if (pipe(fderr) < 0) {
 421                        failed_errno = errno;
 422                        if (need_in)
 423                                close_pair(fdin);
 424                        else if (cmd->in)
 425                                close(cmd->in);
 426                        if (need_out)
 427                                close_pair(fdout);
 428                        else if (cmd->out)
 429                                close(cmd->out);
 430                        str = "standard error";
 431fail_pipe:
 432                        error("cannot create %s pipe for %s: %s",
 433                                str, cmd->argv[0], strerror(failed_errno));
 434                        child_process_clear(cmd);
 435                        errno = failed_errno;
 436                        return -1;
 437                }
 438                cmd->err = fderr[0];
 439        }
 440
 441        trace_argv_printf(cmd->argv, "trace: run_command:");
 442        fflush(NULL);
 443
 444#ifndef GIT_WINDOWS_NATIVE
 445{
 446        int notify_pipe[2];
 447        char **childenv;
 448        struct argv_array argv = ARGV_ARRAY_INIT;
 449
 450        if (pipe(notify_pipe))
 451                notify_pipe[0] = notify_pipe[1] = -1;
 452
 453        prepare_cmd(&argv, cmd);
 454        childenv = prep_childenv(cmd->env);
 455
 456        cmd->pid = fork();
 457        failed_errno = errno;
 458        if (!cmd->pid) {
 459                /*
 460                 * Redirect the channel to write syscall error messages to
 461                 * before redirecting the process's stderr so that all die()
 462                 * in subsequent call paths use the parent's stderr.
 463                 */
 464                if (cmd->no_stderr || need_err) {
 465                        int child_err = dup(2);
 466                        set_cloexec(child_err);
 467                        set_error_handle(fdopen(child_err, "w"));
 468                }
 469
 470                close(notify_pipe[0]);
 471                set_cloexec(notify_pipe[1]);
 472                child_notifier = notify_pipe[1];
 473                atexit(notify_parent);
 474
 475                if (cmd->no_stdin)
 476                        dup_devnull(0);
 477                else if (need_in) {
 478                        dup2(fdin[0], 0);
 479                        close_pair(fdin);
 480                } else if (cmd->in) {
 481                        dup2(cmd->in, 0);
 482                        close(cmd->in);
 483                }
 484
 485                if (cmd->no_stderr)
 486                        dup_devnull(2);
 487                else if (need_err) {
 488                        dup2(fderr[1], 2);
 489                        close_pair(fderr);
 490                } else if (cmd->err > 1) {
 491                        dup2(cmd->err, 2);
 492                        close(cmd->err);
 493                }
 494
 495                if (cmd->no_stdout)
 496                        dup_devnull(1);
 497                else if (cmd->stdout_to_stderr)
 498                        dup2(2, 1);
 499                else if (need_out) {
 500                        dup2(fdout[1], 1);
 501                        close_pair(fdout);
 502                } else if (cmd->out > 1) {
 503                        dup2(cmd->out, 1);
 504                        close(cmd->out);
 505                }
 506
 507                if (cmd->dir && chdir(cmd->dir))
 508                        die_errno("exec '%s': cd to '%s' failed", cmd->argv[0],
 509                            cmd->dir);
 510
 511                /*
 512                 * Attempt to exec using the command and arguments starting at
 513                 * argv.argv[1].  argv.argv[0] contains SHELL_PATH which will
 514                 * be used in the event exec failed with ENOEXEC at which point
 515                 * we will try to interpret the command using 'sh'.
 516                 */
 517                execve(argv.argv[1], (char *const *) argv.argv + 1,
 518                       (char *const *) childenv);
 519                if (errno == ENOEXEC)
 520                        execve(argv.argv[0], (char *const *) argv.argv,
 521                               (char *const *) childenv);
 522
 523                if (errno == ENOENT) {
 524                        if (!cmd->silent_exec_failure)
 525                                error("cannot run %s: %s", cmd->argv[0],
 526                                        strerror(ENOENT));
 527                        exit(127);
 528                } else {
 529                        die_errno("cannot exec '%s'", cmd->argv[0]);
 530                }
 531        }
 532        if (cmd->pid < 0)
 533                error_errno("cannot fork() for %s", cmd->argv[0]);
 534        else if (cmd->clean_on_exit)
 535                mark_child_for_cleanup(cmd->pid, cmd);
 536
 537        /*
 538         * Wait for child's exec. If the exec succeeds (or if fork()
 539         * failed), EOF is seen immediately by the parent. Otherwise, the
 540         * child process sends a single byte.
 541         * Note that use of this infrastructure is completely advisory,
 542         * therefore, we keep error checks minimal.
 543         */
 544        close(notify_pipe[1]);
 545        if (read(notify_pipe[0], &notify_pipe[1], 1) == 1) {
 546                /*
 547                 * At this point we know that fork() succeeded, but exec()
 548                 * failed. Errors have been reported to our stderr.
 549                 */
 550                wait_or_whine(cmd->pid, cmd->argv[0], 0);
 551                failed_errno = errno;
 552                cmd->pid = -1;
 553        }
 554        close(notify_pipe[0]);
 555
 556        argv_array_clear(&argv);
 557        free(childenv);
 558}
 559#else
 560{
 561        int fhin = 0, fhout = 1, fherr = 2;
 562        const char **sargv = cmd->argv;
 563        struct argv_array nargv = ARGV_ARRAY_INIT;
 564
 565        if (cmd->no_stdin)
 566                fhin = open("/dev/null", O_RDWR);
 567        else if (need_in)
 568                fhin = dup(fdin[0]);
 569        else if (cmd->in)
 570                fhin = dup(cmd->in);
 571
 572        if (cmd->no_stderr)
 573                fherr = open("/dev/null", O_RDWR);
 574        else if (need_err)
 575                fherr = dup(fderr[1]);
 576        else if (cmd->err > 2)
 577                fherr = dup(cmd->err);
 578
 579        if (cmd->no_stdout)
 580                fhout = open("/dev/null", O_RDWR);
 581        else if (cmd->stdout_to_stderr)
 582                fhout = dup(fherr);
 583        else if (need_out)
 584                fhout = dup(fdout[1]);
 585        else if (cmd->out > 1)
 586                fhout = dup(cmd->out);
 587
 588        if (cmd->git_cmd)
 589                cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
 590        else if (cmd->use_shell)
 591                cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
 592
 593        cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
 594                        cmd->dir, fhin, fhout, fherr);
 595        failed_errno = errno;
 596        if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
 597                error_errno("cannot spawn %s", cmd->argv[0]);
 598        if (cmd->clean_on_exit && cmd->pid >= 0)
 599                mark_child_for_cleanup(cmd->pid, cmd);
 600
 601        argv_array_clear(&nargv);
 602        cmd->argv = sargv;
 603        if (fhin != 0)
 604                close(fhin);
 605        if (fhout != 1)
 606                close(fhout);
 607        if (fherr != 2)
 608                close(fherr);
 609}
 610#endif
 611
 612        if (cmd->pid < 0) {
 613                if (need_in)
 614                        close_pair(fdin);
 615                else if (cmd->in)
 616                        close(cmd->in);
 617                if (need_out)
 618                        close_pair(fdout);
 619                else if (cmd->out)
 620                        close(cmd->out);
 621                if (need_err)
 622                        close_pair(fderr);
 623                else if (cmd->err)
 624                        close(cmd->err);
 625                child_process_clear(cmd);
 626                errno = failed_errno;
 627                return -1;
 628        }
 629
 630        if (need_in)
 631                close(fdin[0]);
 632        else if (cmd->in)
 633                close(cmd->in);
 634
 635        if (need_out)
 636                close(fdout[1]);
 637        else if (cmd->out)
 638                close(cmd->out);
 639
 640        if (need_err)
 641                close(fderr[1]);
 642        else if (cmd->err)
 643                close(cmd->err);
 644
 645        return 0;
 646}
 647
 648int finish_command(struct child_process *cmd)
 649{
 650        int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
 651        child_process_clear(cmd);
 652        return ret;
 653}
 654
 655int finish_command_in_signal(struct child_process *cmd)
 656{
 657        return wait_or_whine(cmd->pid, cmd->argv[0], 1);
 658}
 659
 660
 661int run_command(struct child_process *cmd)
 662{
 663        int code;
 664
 665        if (cmd->out < 0 || cmd->err < 0)
 666                die("BUG: run_command with a pipe can cause deadlock");
 667
 668        code = start_command(cmd);
 669        if (code)
 670                return code;
 671        return finish_command(cmd);
 672}
 673
 674int run_command_v_opt(const char **argv, int opt)
 675{
 676        return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
 677}
 678
 679int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
 680{
 681        struct child_process cmd = CHILD_PROCESS_INIT;
 682        cmd.argv = argv;
 683        cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
 684        cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
 685        cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
 686        cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
 687        cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
 688        cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
 689        cmd.dir = dir;
 690        cmd.env = env;
 691        return run_command(&cmd);
 692}
 693
 694#ifndef NO_PTHREADS
 695static pthread_t main_thread;
 696static int main_thread_set;
 697static pthread_key_t async_key;
 698static pthread_key_t async_die_counter;
 699
 700static void *run_thread(void *data)
 701{
 702        struct async *async = data;
 703        intptr_t ret;
 704
 705        if (async->isolate_sigpipe) {
 706                sigset_t mask;
 707                sigemptyset(&mask);
 708                sigaddset(&mask, SIGPIPE);
 709                if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
 710                        ret = error("unable to block SIGPIPE in async thread");
 711                        return (void *)ret;
 712                }
 713        }
 714
 715        pthread_setspecific(async_key, async);
 716        ret = async->proc(async->proc_in, async->proc_out, async->data);
 717        return (void *)ret;
 718}
 719
 720static NORETURN void die_async(const char *err, va_list params)
 721{
 722        vreportf("fatal: ", err, params);
 723
 724        if (in_async()) {
 725                struct async *async = pthread_getspecific(async_key);
 726                if (async->proc_in >= 0)
 727                        close(async->proc_in);
 728                if (async->proc_out >= 0)
 729                        close(async->proc_out);
 730                pthread_exit((void *)128);
 731        }
 732
 733        exit(128);
 734}
 735
 736static int async_die_is_recursing(void)
 737{
 738        void *ret = pthread_getspecific(async_die_counter);
 739        pthread_setspecific(async_die_counter, (void *)1);
 740        return ret != NULL;
 741}
 742
 743int in_async(void)
 744{
 745        if (!main_thread_set)
 746                return 0; /* no asyncs started yet */
 747        return !pthread_equal(main_thread, pthread_self());
 748}
 749
 750static void NORETURN async_exit(int code)
 751{
 752        pthread_exit((void *)(intptr_t)code);
 753}
 754
 755#else
 756
 757static struct {
 758        void (**handlers)(void);
 759        size_t nr;
 760        size_t alloc;
 761} git_atexit_hdlrs;
 762
 763static int git_atexit_installed;
 764
 765static void git_atexit_dispatch(void)
 766{
 767        size_t i;
 768
 769        for (i=git_atexit_hdlrs.nr ; i ; i--)
 770                git_atexit_hdlrs.handlers[i-1]();
 771}
 772
 773static void git_atexit_clear(void)
 774{
 775        free(git_atexit_hdlrs.handlers);
 776        memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
 777        git_atexit_installed = 0;
 778}
 779
 780#undef atexit
 781int git_atexit(void (*handler)(void))
 782{
 783        ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
 784        git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
 785        if (!git_atexit_installed) {
 786                if (atexit(&git_atexit_dispatch))
 787                        return -1;
 788                git_atexit_installed = 1;
 789        }
 790        return 0;
 791}
 792#define atexit git_atexit
 793
 794static int process_is_async;
 795int in_async(void)
 796{
 797        return process_is_async;
 798}
 799
 800static void NORETURN async_exit(int code)
 801{
 802        exit(code);
 803}
 804
 805#endif
 806
 807void check_pipe(int err)
 808{
 809        if (err == EPIPE) {
 810                if (in_async())
 811                        async_exit(141);
 812
 813                signal(SIGPIPE, SIG_DFL);
 814                raise(SIGPIPE);
 815                /* Should never happen, but just in case... */
 816                exit(141);
 817        }
 818}
 819
 820int start_async(struct async *async)
 821{
 822        int need_in, need_out;
 823        int fdin[2], fdout[2];
 824        int proc_in, proc_out;
 825
 826        need_in = async->in < 0;
 827        if (need_in) {
 828                if (pipe(fdin) < 0) {
 829                        if (async->out > 0)
 830                                close(async->out);
 831                        return error_errno("cannot create pipe");
 832                }
 833                async->in = fdin[1];
 834        }
 835
 836        need_out = async->out < 0;
 837        if (need_out) {
 838                if (pipe(fdout) < 0) {
 839                        if (need_in)
 840                                close_pair(fdin);
 841                        else if (async->in)
 842                                close(async->in);
 843                        return error_errno("cannot create pipe");
 844                }
 845                async->out = fdout[0];
 846        }
 847
 848        if (need_in)
 849                proc_in = fdin[0];
 850        else if (async->in)
 851                proc_in = async->in;
 852        else
 853                proc_in = -1;
 854
 855        if (need_out)
 856                proc_out = fdout[1];
 857        else if (async->out)
 858                proc_out = async->out;
 859        else
 860                proc_out = -1;
 861
 862#ifdef NO_PTHREADS
 863        /* Flush stdio before fork() to avoid cloning buffers */
 864        fflush(NULL);
 865
 866        async->pid = fork();
 867        if (async->pid < 0) {
 868                error_errno("fork (async) failed");
 869                goto error;
 870        }
 871        if (!async->pid) {
 872                if (need_in)
 873                        close(fdin[1]);
 874                if (need_out)
 875                        close(fdout[0]);
 876                git_atexit_clear();
 877                process_is_async = 1;
 878                exit(!!async->proc(proc_in, proc_out, async->data));
 879        }
 880
 881        mark_child_for_cleanup(async->pid, NULL);
 882
 883        if (need_in)
 884                close(fdin[0]);
 885        else if (async->in)
 886                close(async->in);
 887
 888        if (need_out)
 889                close(fdout[1]);
 890        else if (async->out)
 891                close(async->out);
 892#else
 893        if (!main_thread_set) {
 894                /*
 895                 * We assume that the first time that start_async is called
 896                 * it is from the main thread.
 897                 */
 898                main_thread_set = 1;
 899                main_thread = pthread_self();
 900                pthread_key_create(&async_key, NULL);
 901                pthread_key_create(&async_die_counter, NULL);
 902                set_die_routine(die_async);
 903                set_die_is_recursing_routine(async_die_is_recursing);
 904        }
 905
 906        if (proc_in >= 0)
 907                set_cloexec(proc_in);
 908        if (proc_out >= 0)
 909                set_cloexec(proc_out);
 910        async->proc_in = proc_in;
 911        async->proc_out = proc_out;
 912        {
 913                int err = pthread_create(&async->tid, NULL, run_thread, async);
 914                if (err) {
 915                        error_errno("cannot create thread");
 916                        goto error;
 917                }
 918        }
 919#endif
 920        return 0;
 921
 922error:
 923        if (need_in)
 924                close_pair(fdin);
 925        else if (async->in)
 926                close(async->in);
 927
 928        if (need_out)
 929                close_pair(fdout);
 930        else if (async->out)
 931                close(async->out);
 932        return -1;
 933}
 934
 935int finish_async(struct async *async)
 936{
 937#ifdef NO_PTHREADS
 938        return wait_or_whine(async->pid, "child process", 0);
 939#else
 940        void *ret = (void *)(intptr_t)(-1);
 941
 942        if (pthread_join(async->tid, &ret))
 943                error("pthread_join failed");
 944        return (int)(intptr_t)ret;
 945#endif
 946}
 947
 948const char *find_hook(const char *name)
 949{
 950        static struct strbuf path = STRBUF_INIT;
 951
 952        strbuf_reset(&path);
 953        strbuf_git_path(&path, "hooks/%s", name);
 954        if (access(path.buf, X_OK) < 0) {
 955#ifdef STRIP_EXTENSION
 956                strbuf_addstr(&path, STRIP_EXTENSION);
 957                if (access(path.buf, X_OK) >= 0)
 958                        return path.buf;
 959#endif
 960                return NULL;
 961        }
 962        return path.buf;
 963}
 964
 965int run_hook_ve(const char *const *env, const char *name, va_list args)
 966{
 967        struct child_process hook = CHILD_PROCESS_INIT;
 968        const char *p;
 969
 970        p = find_hook(name);
 971        if (!p)
 972                return 0;
 973
 974        argv_array_push(&hook.args, p);
 975        while ((p = va_arg(args, const char *)))
 976                argv_array_push(&hook.args, p);
 977        hook.env = env;
 978        hook.no_stdin = 1;
 979        hook.stdout_to_stderr = 1;
 980
 981        return run_command(&hook);
 982}
 983
 984int run_hook_le(const char *const *env, const char *name, ...)
 985{
 986        va_list args;
 987        int ret;
 988
 989        va_start(args, name);
 990        ret = run_hook_ve(env, name, args);
 991        va_end(args);
 992
 993        return ret;
 994}
 995
 996struct io_pump {
 997        /* initialized by caller */
 998        int fd;
 999        int type; /* POLLOUT or POLLIN */
1000        union {
1001                struct {
1002                        const char *buf;
1003                        size_t len;
1004                } out;
1005                struct {
1006                        struct strbuf *buf;
1007                        size_t hint;
1008                } in;
1009        } u;
1010
1011        /* returned by pump_io */
1012        int error; /* 0 for success, otherwise errno */
1013
1014        /* internal use */
1015        struct pollfd *pfd;
1016};
1017
1018static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
1019{
1020        int pollsize = 0;
1021        int i;
1022
1023        for (i = 0; i < nr; i++) {
1024                struct io_pump *io = &slots[i];
1025                if (io->fd < 0)
1026                        continue;
1027                pfd[pollsize].fd = io->fd;
1028                pfd[pollsize].events = io->type;
1029                io->pfd = &pfd[pollsize++];
1030        }
1031
1032        if (!pollsize)
1033                return 0;
1034
1035        if (poll(pfd, pollsize, -1) < 0) {
1036                if (errno == EINTR)
1037                        return 1;
1038                die_errno("poll failed");
1039        }
1040
1041        for (i = 0; i < nr; i++) {
1042                struct io_pump *io = &slots[i];
1043
1044                if (io->fd < 0)
1045                        continue;
1046
1047                if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
1048                        continue;
1049
1050                if (io->type == POLLOUT) {
1051                        ssize_t len = xwrite(io->fd,
1052                                             io->u.out.buf, io->u.out.len);
1053                        if (len < 0) {
1054                                io->error = errno;
1055                                close(io->fd);
1056                                io->fd = -1;
1057                        } else {
1058                                io->u.out.buf += len;
1059                                io->u.out.len -= len;
1060                                if (!io->u.out.len) {
1061                                        close(io->fd);
1062                                        io->fd = -1;
1063                                }
1064                        }
1065                }
1066
1067                if (io->type == POLLIN) {
1068                        ssize_t len = strbuf_read_once(io->u.in.buf,
1069                                                       io->fd, io->u.in.hint);
1070                        if (len < 0)
1071                                io->error = errno;
1072                        if (len <= 0) {
1073                                close(io->fd);
1074                                io->fd = -1;
1075                        }
1076                }
1077        }
1078
1079        return 1;
1080}
1081
1082static int pump_io(struct io_pump *slots, int nr)
1083{
1084        struct pollfd *pfd;
1085        int i;
1086
1087        for (i = 0; i < nr; i++)
1088                slots[i].error = 0;
1089
1090        ALLOC_ARRAY(pfd, nr);
1091        while (pump_io_round(slots, nr, pfd))
1092                ; /* nothing */
1093        free(pfd);
1094
1095        /* There may be multiple errno values, so just pick the first. */
1096        for (i = 0; i < nr; i++) {
1097                if (slots[i].error) {
1098                        errno = slots[i].error;
1099                        return -1;
1100                }
1101        }
1102        return 0;
1103}
1104
1105
1106int pipe_command(struct child_process *cmd,
1107                 const char *in, size_t in_len,
1108                 struct strbuf *out, size_t out_hint,
1109                 struct strbuf *err, size_t err_hint)
1110{
1111        struct io_pump io[3];
1112        int nr = 0;
1113
1114        if (in)
1115                cmd->in = -1;
1116        if (out)
1117                cmd->out = -1;
1118        if (err)
1119                cmd->err = -1;
1120
1121        if (start_command(cmd) < 0)
1122                return -1;
1123
1124        if (in) {
1125                io[nr].fd = cmd->in;
1126                io[nr].type = POLLOUT;
1127                io[nr].u.out.buf = in;
1128                io[nr].u.out.len = in_len;
1129                nr++;
1130        }
1131        if (out) {
1132                io[nr].fd = cmd->out;
1133                io[nr].type = POLLIN;
1134                io[nr].u.in.buf = out;
1135                io[nr].u.in.hint = out_hint;
1136                nr++;
1137        }
1138        if (err) {
1139                io[nr].fd = cmd->err;
1140                io[nr].type = POLLIN;
1141                io[nr].u.in.buf = err;
1142                io[nr].u.in.hint = err_hint;
1143                nr++;
1144        }
1145
1146        if (pump_io(io, nr) < 0) {
1147                finish_command(cmd); /* throw away exit code */
1148                return -1;
1149        }
1150
1151        return finish_command(cmd);
1152}
1153
1154enum child_state {
1155        GIT_CP_FREE,
1156        GIT_CP_WORKING,
1157        GIT_CP_WAIT_CLEANUP,
1158};
1159
1160struct parallel_processes {
1161        void *data;
1162
1163        int max_processes;
1164        int nr_processes;
1165
1166        get_next_task_fn get_next_task;
1167        start_failure_fn start_failure;
1168        task_finished_fn task_finished;
1169
1170        struct {
1171                enum child_state state;
1172                struct child_process process;
1173                struct strbuf err;
1174                void *data;
1175        } *children;
1176        /*
1177         * The struct pollfd is logically part of *children,
1178         * but the system call expects it as its own array.
1179         */
1180        struct pollfd *pfd;
1181
1182        unsigned shutdown : 1;
1183
1184        int output_owner;
1185        struct strbuf buffered_output; /* of finished children */
1186};
1187
1188static int default_start_failure(struct strbuf *out,
1189                                 void *pp_cb,
1190                                 void *pp_task_cb)
1191{
1192        return 0;
1193}
1194
1195static int default_task_finished(int result,
1196                                 struct strbuf *out,
1197                                 void *pp_cb,
1198                                 void *pp_task_cb)
1199{
1200        return 0;
1201}
1202
1203static void kill_children(struct parallel_processes *pp, int signo)
1204{
1205        int i, n = pp->max_processes;
1206
1207        for (i = 0; i < n; i++)
1208                if (pp->children[i].state == GIT_CP_WORKING)
1209                        kill(pp->children[i].process.pid, signo);
1210}
1211
1212static struct parallel_processes *pp_for_signal;
1213
1214static void handle_children_on_signal(int signo)
1215{
1216        kill_children(pp_for_signal, signo);
1217        sigchain_pop(signo);
1218        raise(signo);
1219}
1220
1221static void pp_init(struct parallel_processes *pp,
1222                    int n,
1223                    get_next_task_fn get_next_task,
1224                    start_failure_fn start_failure,
1225                    task_finished_fn task_finished,
1226                    void *data)
1227{
1228        int i;
1229
1230        if (n < 1)
1231                n = online_cpus();
1232
1233        pp->max_processes = n;
1234
1235        trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
1236
1237        pp->data = data;
1238        if (!get_next_task)
1239                die("BUG: you need to specify a get_next_task function");
1240        pp->get_next_task = get_next_task;
1241
1242        pp->start_failure = start_failure ? start_failure : default_start_failure;
1243        pp->task_finished = task_finished ? task_finished : default_task_finished;
1244
1245        pp->nr_processes = 0;
1246        pp->output_owner = 0;
1247        pp->shutdown = 0;
1248        pp->children = xcalloc(n, sizeof(*pp->children));
1249        pp->pfd = xcalloc(n, sizeof(*pp->pfd));
1250        strbuf_init(&pp->buffered_output, 0);
1251
1252        for (i = 0; i < n; i++) {
1253                strbuf_init(&pp->children[i].err, 0);
1254                child_process_init(&pp->children[i].process);
1255                pp->pfd[i].events = POLLIN | POLLHUP;
1256                pp->pfd[i].fd = -1;
1257        }
1258
1259        pp_for_signal = pp;
1260        sigchain_push_common(handle_children_on_signal);
1261}
1262
1263static void pp_cleanup(struct parallel_processes *pp)
1264{
1265        int i;
1266
1267        trace_printf("run_processes_parallel: done");
1268        for (i = 0; i < pp->max_processes; i++) {
1269                strbuf_release(&pp->children[i].err);
1270                child_process_clear(&pp->children[i].process);
1271        }
1272
1273        free(pp->children);
1274        free(pp->pfd);
1275
1276        /*
1277         * When get_next_task added messages to the buffer in its last
1278         * iteration, the buffered output is non empty.
1279         */
1280        strbuf_write(&pp->buffered_output, stderr);
1281        strbuf_release(&pp->buffered_output);
1282
1283        sigchain_pop_common();
1284}
1285
1286/* returns
1287 *  0 if a new task was started.
1288 *  1 if no new jobs was started (get_next_task ran out of work, non critical
1289 *    problem with starting a new command)
1290 * <0 no new job was started, user wishes to shutdown early. Use negative code
1291 *    to signal the children.
1292 */
1293static int pp_start_one(struct parallel_processes *pp)
1294{
1295        int i, code;
1296
1297        for (i = 0; i < pp->max_processes; i++)
1298                if (pp->children[i].state == GIT_CP_FREE)
1299                        break;
1300        if (i == pp->max_processes)
1301                die("BUG: bookkeeping is hard");
1302
1303        code = pp->get_next_task(&pp->children[i].process,
1304                                 &pp->children[i].err,
1305                                 pp->data,
1306                                 &pp->children[i].data);
1307        if (!code) {
1308                strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1309                strbuf_reset(&pp->children[i].err);
1310                return 1;
1311        }
1312        pp->children[i].process.err = -1;
1313        pp->children[i].process.stdout_to_stderr = 1;
1314        pp->children[i].process.no_stdin = 1;
1315
1316        if (start_command(&pp->children[i].process)) {
1317                code = pp->start_failure(&pp->children[i].err,
1318                                         pp->data,
1319                                         &pp->children[i].data);
1320                strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1321                strbuf_reset(&pp->children[i].err);
1322                if (code)
1323                        pp->shutdown = 1;
1324                return code;
1325        }
1326
1327        pp->nr_processes++;
1328        pp->children[i].state = GIT_CP_WORKING;
1329        pp->pfd[i].fd = pp->children[i].process.err;
1330        return 0;
1331}
1332
1333static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1334{
1335        int i;
1336
1337        while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1338                if (errno == EINTR)
1339                        continue;
1340                pp_cleanup(pp);
1341                die_errno("poll");
1342        }
1343
1344        /* Buffer output from all pipes. */
1345        for (i = 0; i < pp->max_processes; i++) {
1346                if (pp->children[i].state == GIT_CP_WORKING &&
1347                    pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1348                        int n = strbuf_read_once(&pp->children[i].err,
1349                                                 pp->children[i].process.err, 0);
1350                        if (n == 0) {
1351                                close(pp->children[i].process.err);
1352                                pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1353                        } else if (n < 0)
1354                                if (errno != EAGAIN)
1355                                        die_errno("read");
1356                }
1357        }
1358}
1359
1360static void pp_output(struct parallel_processes *pp)
1361{
1362        int i = pp->output_owner;
1363        if (pp->children[i].state == GIT_CP_WORKING &&
1364            pp->children[i].err.len) {
1365                strbuf_write(&pp->children[i].err, stderr);
1366                strbuf_reset(&pp->children[i].err);
1367        }
1368}
1369
1370static int pp_collect_finished(struct parallel_processes *pp)
1371{
1372        int i, code;
1373        int n = pp->max_processes;
1374        int result = 0;
1375
1376        while (pp->nr_processes > 0) {
1377                for (i = 0; i < pp->max_processes; i++)
1378                        if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1379                                break;
1380                if (i == pp->max_processes)
1381                        break;
1382
1383                code = finish_command(&pp->children[i].process);
1384
1385                code = pp->task_finished(code,
1386                                         &pp->children[i].err, pp->data,
1387                                         &pp->children[i].data);
1388
1389                if (code)
1390                        result = code;
1391                if (code < 0)
1392                        break;
1393
1394                pp->nr_processes--;
1395                pp->children[i].state = GIT_CP_FREE;
1396                pp->pfd[i].fd = -1;
1397                child_process_init(&pp->children[i].process);
1398
1399                if (i != pp->output_owner) {
1400                        strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1401                        strbuf_reset(&pp->children[i].err);
1402                } else {
1403                        strbuf_write(&pp->children[i].err, stderr);
1404                        strbuf_reset(&pp->children[i].err);
1405
1406                        /* Output all other finished child processes */
1407                        strbuf_write(&pp->buffered_output, stderr);
1408                        strbuf_reset(&pp->buffered_output);
1409
1410                        /*
1411                         * Pick next process to output live.
1412                         * NEEDSWORK:
1413                         * For now we pick it randomly by doing a round
1414                         * robin. Later we may want to pick the one with
1415                         * the most output or the longest or shortest
1416                         * running process time.
1417                         */
1418                        for (i = 0; i < n; i++)
1419                                if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1420                                        break;
1421                        pp->output_owner = (pp->output_owner + i) % n;
1422                }
1423        }
1424        return result;
1425}
1426
1427int run_processes_parallel(int n,
1428                           get_next_task_fn get_next_task,
1429                           start_failure_fn start_failure,
1430                           task_finished_fn task_finished,
1431                           void *pp_cb)
1432{
1433        int i, code;
1434        int output_timeout = 100;
1435        int spawn_cap = 4;
1436        struct parallel_processes pp;
1437
1438        pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1439        while (1) {
1440                for (i = 0;
1441                    i < spawn_cap && !pp.shutdown &&
1442                    pp.nr_processes < pp.max_processes;
1443                    i++) {
1444                        code = pp_start_one(&pp);
1445                        if (!code)
1446                                continue;
1447                        if (code < 0) {
1448                                pp.shutdown = 1;
1449                                kill_children(&pp, -code);
1450                        }
1451                        break;
1452                }
1453                if (!pp.nr_processes)
1454                        break;
1455                pp_buffer_stderr(&pp, output_timeout);
1456                pp_output(&pp);
1457                code = pp_collect_finished(&pp);
1458                if (code) {
1459                        pp.shutdown = 1;
1460                        if (code < 0)
1461                                kill_children(&pp, -code);
1462                }
1463        }
1464
1465        pp_cleanup(&pp);
1466        return 0;
1467}