Andrew's git / gitweb.git / run-command.c
?
run-command.con commit Merge branch 'jk/more-comments-on-textconv' (3ed26a4)
   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_to_clean *next;
  25};
  26static struct child_to_clean *children_to_clean;
  27static int installed_child_cleanup_handler;
  28
  29static void cleanup_children(int sig, int in_signal)
  30{
  31        while (children_to_clean) {
  32                struct child_to_clean *p = children_to_clean;
  33                children_to_clean = p->next;
  34                kill(p->pid, sig);
  35                if (!in_signal)
  36                        free(p);
  37        }
  38}
  39
  40static void cleanup_children_on_signal(int sig)
  41{
  42        cleanup_children(sig, 1);
  43        sigchain_pop(sig);
  44        raise(sig);
  45}
  46
  47static void cleanup_children_on_exit(void)
  48{
  49        cleanup_children(SIGTERM, 0);
  50}
  51
  52static void mark_child_for_cleanup(pid_t pid)
  53{
  54        struct child_to_clean *p = xmalloc(sizeof(*p));
  55        p->pid = pid;
  56        p->next = children_to_clean;
  57        children_to_clean = p;
  58
  59        if (!installed_child_cleanup_handler) {
  60                atexit(cleanup_children_on_exit);
  61                sigchain_push_common(cleanup_children_on_signal);
  62                installed_child_cleanup_handler = 1;
  63        }
  64}
  65
  66static void clear_child_for_cleanup(pid_t pid)
  67{
  68        struct child_to_clean **pp;
  69
  70        for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
  71                struct child_to_clean *clean_me = *pp;
  72
  73                if (clean_me->pid == pid) {
  74                        *pp = clean_me->next;
  75                        free(clean_me);
  76                        return;
  77                }
  78        }
  79}
  80
  81static inline void close_pair(int fd[2])
  82{
  83        close(fd[0]);
  84        close(fd[1]);
  85}
  86
  87#ifndef GIT_WINDOWS_NATIVE
  88static inline void dup_devnull(int to)
  89{
  90        int fd = open("/dev/null", O_RDWR);
  91        if (fd < 0)
  92                die_errno(_("open /dev/null failed"));
  93        if (dup2(fd, to) < 0)
  94                die_errno(_("dup2(%d,%d) failed"), fd, to);
  95        close(fd);
  96}
  97#endif
  98
  99static char *locate_in_PATH(const char *file)
 100{
 101        const char *p = getenv("PATH");
 102        struct strbuf buf = STRBUF_INIT;
 103
 104        if (!p || !*p)
 105                return NULL;
 106
 107        while (1) {
 108                const char *end = strchrnul(p, ':');
 109
 110                strbuf_reset(&buf);
 111
 112                /* POSIX specifies an empty entry as the current directory. */
 113                if (end != p) {
 114                        strbuf_add(&buf, p, end - p);
 115                        strbuf_addch(&buf, '/');
 116                }
 117                strbuf_addstr(&buf, file);
 118
 119                if (!access(buf.buf, F_OK))
 120                        return strbuf_detach(&buf, NULL);
 121
 122                if (!*end)
 123                        break;
 124                p = end + 1;
 125        }
 126
 127        strbuf_release(&buf);
 128        return NULL;
 129}
 130
 131static int exists_in_PATH(const char *file)
 132{
 133        char *r = locate_in_PATH(file);
 134        free(r);
 135        return r != NULL;
 136}
 137
 138int sane_execvp(const char *file, char * const argv[])
 139{
 140        if (!execvp(file, argv))
 141                return 0; /* cannot happen ;-) */
 142
 143        /*
 144         * When a command can't be found because one of the directories
 145         * listed in $PATH is unsearchable, execvp reports EACCES, but
 146         * careful usability testing (read: analysis of occasional bug
 147         * reports) reveals that "No such file or directory" is more
 148         * intuitive.
 149         *
 150         * We avoid commands with "/", because execvp will not do $PATH
 151         * lookups in that case.
 152         *
 153         * The reassignment of EACCES to errno looks like a no-op below,
 154         * but we need to protect against exists_in_PATH overwriting errno.
 155         */
 156        if (errno == EACCES && !strchr(file, '/'))
 157                errno = exists_in_PATH(file) ? EACCES : ENOENT;
 158        else if (errno == ENOTDIR && !strchr(file, '/'))
 159                errno = ENOENT;
 160        return -1;
 161}
 162
 163static const char **prepare_shell_cmd(const char **argv)
 164{
 165        int argc, nargc = 0;
 166        const char **nargv;
 167
 168        for (argc = 0; argv[argc]; argc++)
 169                ; /* just counting */
 170        /* +1 for NULL, +3 for "sh -c" plus extra $0 */
 171        nargv = xmalloc(sizeof(*nargv) * (argc + 1 + 3));
 172
 173        if (argc < 1)
 174                die("BUG: shell command is empty");
 175
 176        if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
 177#ifndef GIT_WINDOWS_NATIVE
 178                nargv[nargc++] = SHELL_PATH;
 179#else
 180                nargv[nargc++] = "sh";
 181#endif
 182                nargv[nargc++] = "-c";
 183
 184                if (argc < 2)
 185                        nargv[nargc++] = argv[0];
 186                else {
 187                        struct strbuf arg0 = STRBUF_INIT;
 188                        strbuf_addf(&arg0, "%s \"$@\"", argv[0]);
 189                        nargv[nargc++] = strbuf_detach(&arg0, NULL);
 190                }
 191        }
 192
 193        for (argc = 0; argv[argc]; argc++)
 194                nargv[nargc++] = argv[argc];
 195        nargv[nargc] = NULL;
 196
 197        return nargv;
 198}
 199
 200#ifndef GIT_WINDOWS_NATIVE
 201static int execv_shell_cmd(const char **argv)
 202{
 203        const char **nargv = prepare_shell_cmd(argv);
 204        trace_argv_printf(nargv, "trace: exec:");
 205        sane_execvp(nargv[0], (char **)nargv);
 206        free(nargv);
 207        return -1;
 208}
 209#endif
 210
 211#ifndef GIT_WINDOWS_NATIVE
 212static int child_notifier = -1;
 213
 214static void notify_parent(void)
 215{
 216        /*
 217         * execvp failed.  If possible, we'd like to let start_command
 218         * know, so failures like ENOENT can be handled right away; but
 219         * otherwise, finish_command will still report the error.
 220         */
 221        xwrite(child_notifier, "", 1);
 222}
 223#endif
 224
 225static inline void set_cloexec(int fd)
 226{
 227        int flags = fcntl(fd, F_GETFD);
 228        if (flags >= 0)
 229                fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
 230}
 231
 232static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
 233{
 234        int status, code = -1;
 235        pid_t waiting;
 236        int failed_errno = 0;
 237
 238        while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
 239                ;       /* nothing */
 240        if (in_signal)
 241                return 0;
 242
 243        if (waiting < 0) {
 244                failed_errno = errno;
 245                error("waitpid for %s failed: %s", argv0, strerror(errno));
 246        } else if (waiting != pid) {
 247                error("waitpid is confused (%s)", argv0);
 248        } else if (WIFSIGNALED(status)) {
 249                code = WTERMSIG(status);
 250                if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
 251                        error("%s died of signal %d", argv0, code);
 252                /*
 253                 * This return value is chosen so that code & 0xff
 254                 * mimics the exit code that a POSIX shell would report for
 255                 * a program that died from this signal.
 256                 */
 257                code += 128;
 258        } else if (WIFEXITED(status)) {
 259                code = WEXITSTATUS(status);
 260                /*
 261                 * Convert special exit code when execvp failed.
 262                 */
 263                if (code == 127) {
 264                        code = -1;
 265                        failed_errno = ENOENT;
 266                }
 267        } else {
 268                error("waitpid is confused (%s)", argv0);
 269        }
 270
 271        clear_child_for_cleanup(pid);
 272
 273        errno = failed_errno;
 274        return code;
 275}
 276
 277int start_command(struct child_process *cmd)
 278{
 279        int need_in, need_out, need_err;
 280        int fdin[2], fdout[2], fderr[2];
 281        int failed_errno;
 282        char *str;
 283
 284        if (!cmd->argv)
 285                cmd->argv = cmd->args.argv;
 286        if (!cmd->env)
 287                cmd->env = cmd->env_array.argv;
 288
 289        /*
 290         * In case of errors we must keep the promise to close FDs
 291         * that have been passed in via ->in and ->out.
 292         */
 293
 294        need_in = !cmd->no_stdin && cmd->in < 0;
 295        if (need_in) {
 296                if (pipe(fdin) < 0) {
 297                        failed_errno = errno;
 298                        if (cmd->out > 0)
 299                                close(cmd->out);
 300                        str = "standard input";
 301                        goto fail_pipe;
 302                }
 303                cmd->in = fdin[1];
 304        }
 305
 306        need_out = !cmd->no_stdout
 307                && !cmd->stdout_to_stderr
 308                && cmd->out < 0;
 309        if (need_out) {
 310                if (pipe(fdout) < 0) {
 311                        failed_errno = errno;
 312                        if (need_in)
 313                                close_pair(fdin);
 314                        else if (cmd->in)
 315                                close(cmd->in);
 316                        str = "standard output";
 317                        goto fail_pipe;
 318                }
 319                cmd->out = fdout[0];
 320        }
 321
 322        need_err = !cmd->no_stderr && cmd->err < 0;
 323        if (need_err) {
 324                if (pipe(fderr) < 0) {
 325                        failed_errno = errno;
 326                        if (need_in)
 327                                close_pair(fdin);
 328                        else if (cmd->in)
 329                                close(cmd->in);
 330                        if (need_out)
 331                                close_pair(fdout);
 332                        else if (cmd->out)
 333                                close(cmd->out);
 334                        str = "standard error";
 335fail_pipe:
 336                        error("cannot create %s pipe for %s: %s",
 337                                str, cmd->argv[0], strerror(failed_errno));
 338                        child_process_clear(cmd);
 339                        errno = failed_errno;
 340                        return -1;
 341                }
 342                cmd->err = fderr[0];
 343        }
 344
 345        trace_argv_printf(cmd->argv, "trace: run_command:");
 346        fflush(NULL);
 347
 348#ifndef GIT_WINDOWS_NATIVE
 349{
 350        int notify_pipe[2];
 351        if (pipe(notify_pipe))
 352                notify_pipe[0] = notify_pipe[1] = -1;
 353
 354        cmd->pid = fork();
 355        failed_errno = errno;
 356        if (!cmd->pid) {
 357                /*
 358                 * Redirect the channel to write syscall error messages to
 359                 * before redirecting the process's stderr so that all die()
 360                 * in subsequent call paths use the parent's stderr.
 361                 */
 362                if (cmd->no_stderr || need_err) {
 363                        int child_err = dup(2);
 364                        set_cloexec(child_err);
 365                        set_error_handle(fdopen(child_err, "w"));
 366                }
 367
 368                close(notify_pipe[0]);
 369                set_cloexec(notify_pipe[1]);
 370                child_notifier = notify_pipe[1];
 371                atexit(notify_parent);
 372
 373                if (cmd->no_stdin)
 374                        dup_devnull(0);
 375                else if (need_in) {
 376                        dup2(fdin[0], 0);
 377                        close_pair(fdin);
 378                } else if (cmd->in) {
 379                        dup2(cmd->in, 0);
 380                        close(cmd->in);
 381                }
 382
 383                if (cmd->no_stderr)
 384                        dup_devnull(2);
 385                else if (need_err) {
 386                        dup2(fderr[1], 2);
 387                        close_pair(fderr);
 388                } else if (cmd->err > 1) {
 389                        dup2(cmd->err, 2);
 390                        close(cmd->err);
 391                }
 392
 393                if (cmd->no_stdout)
 394                        dup_devnull(1);
 395                else if (cmd->stdout_to_stderr)
 396                        dup2(2, 1);
 397                else if (need_out) {
 398                        dup2(fdout[1], 1);
 399                        close_pair(fdout);
 400                } else if (cmd->out > 1) {
 401                        dup2(cmd->out, 1);
 402                        close(cmd->out);
 403                }
 404
 405                if (cmd->dir && chdir(cmd->dir))
 406                        die_errno("exec '%s': cd to '%s' failed", cmd->argv[0],
 407                            cmd->dir);
 408                if (cmd->env) {
 409                        for (; *cmd->env; cmd->env++) {
 410                                if (strchr(*cmd->env, '='))
 411                                        putenv((char *)*cmd->env);
 412                                else
 413                                        unsetenv(*cmd->env);
 414                        }
 415                }
 416                if (cmd->git_cmd)
 417                        execv_git_cmd(cmd->argv);
 418                else if (cmd->use_shell)
 419                        execv_shell_cmd(cmd->argv);
 420                else
 421                        sane_execvp(cmd->argv[0], (char *const*) cmd->argv);
 422                if (errno == ENOENT) {
 423                        if (!cmd->silent_exec_failure)
 424                                error("cannot run %s: %s", cmd->argv[0],
 425                                        strerror(ENOENT));
 426                        exit(127);
 427                } else {
 428                        die_errno("cannot exec '%s'", cmd->argv[0]);
 429                }
 430        }
 431        if (cmd->pid < 0)
 432                error("cannot fork() for %s: %s", cmd->argv[0],
 433                        strerror(errno));
 434        else if (cmd->clean_on_exit)
 435                mark_child_for_cleanup(cmd->pid);
 436
 437        /*
 438         * Wait for child's execvp. If the execvp succeeds (or if fork()
 439         * failed), EOF is seen immediately by the parent. Otherwise, the
 440         * child process sends a single byte.
 441         * Note that use of this infrastructure is completely advisory,
 442         * therefore, we keep error checks minimal.
 443         */
 444        close(notify_pipe[1]);
 445        if (read(notify_pipe[0], &notify_pipe[1], 1) == 1) {
 446                /*
 447                 * At this point we know that fork() succeeded, but execvp()
 448                 * failed. Errors have been reported to our stderr.
 449                 */
 450                wait_or_whine(cmd->pid, cmd->argv[0], 0);
 451                failed_errno = errno;
 452                cmd->pid = -1;
 453        }
 454        close(notify_pipe[0]);
 455}
 456#else
 457{
 458        int fhin = 0, fhout = 1, fherr = 2;
 459        const char **sargv = cmd->argv;
 460
 461        if (cmd->no_stdin)
 462                fhin = open("/dev/null", O_RDWR);
 463        else if (need_in)
 464                fhin = dup(fdin[0]);
 465        else if (cmd->in)
 466                fhin = dup(cmd->in);
 467
 468        if (cmd->no_stderr)
 469                fherr = open("/dev/null", O_RDWR);
 470        else if (need_err)
 471                fherr = dup(fderr[1]);
 472        else if (cmd->err > 2)
 473                fherr = dup(cmd->err);
 474
 475        if (cmd->no_stdout)
 476                fhout = open("/dev/null", O_RDWR);
 477        else if (cmd->stdout_to_stderr)
 478                fhout = dup(fherr);
 479        else if (need_out)
 480                fhout = dup(fdout[1]);
 481        else if (cmd->out > 1)
 482                fhout = dup(cmd->out);
 483
 484        if (cmd->git_cmd)
 485                cmd->argv = prepare_git_cmd(cmd->argv);
 486        else if (cmd->use_shell)
 487                cmd->argv = prepare_shell_cmd(cmd->argv);
 488
 489        cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
 490                        cmd->dir, fhin, fhout, fherr);
 491        failed_errno = errno;
 492        if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
 493                error("cannot spawn %s: %s", cmd->argv[0], strerror(errno));
 494        if (cmd->clean_on_exit && cmd->pid >= 0)
 495                mark_child_for_cleanup(cmd->pid);
 496
 497        if (cmd->git_cmd)
 498                free(cmd->argv);
 499
 500        cmd->argv = sargv;
 501        if (fhin != 0)
 502                close(fhin);
 503        if (fhout != 1)
 504                close(fhout);
 505        if (fherr != 2)
 506                close(fherr);
 507}
 508#endif
 509
 510        if (cmd->pid < 0) {
 511                if (need_in)
 512                        close_pair(fdin);
 513                else if (cmd->in)
 514                        close(cmd->in);
 515                if (need_out)
 516                        close_pair(fdout);
 517                else if (cmd->out)
 518                        close(cmd->out);
 519                if (need_err)
 520                        close_pair(fderr);
 521                else if (cmd->err)
 522                        close(cmd->err);
 523                child_process_clear(cmd);
 524                errno = failed_errno;
 525                return -1;
 526        }
 527
 528        if (need_in)
 529                close(fdin[0]);
 530        else if (cmd->in)
 531                close(cmd->in);
 532
 533        if (need_out)
 534                close(fdout[1]);
 535        else if (cmd->out)
 536                close(cmd->out);
 537
 538        if (need_err)
 539                close(fderr[1]);
 540        else if (cmd->err)
 541                close(cmd->err);
 542
 543        return 0;
 544}
 545
 546int finish_command(struct child_process *cmd)
 547{
 548        int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
 549        child_process_clear(cmd);
 550        return ret;
 551}
 552
 553int finish_command_in_signal(struct child_process *cmd)
 554{
 555        return wait_or_whine(cmd->pid, cmd->argv[0], 1);
 556}
 557
 558
 559int run_command(struct child_process *cmd)
 560{
 561        int code;
 562
 563        if (cmd->out < 0 || cmd->err < 0)
 564                die("BUG: run_command with a pipe can cause deadlock");
 565
 566        code = start_command(cmd);
 567        if (code)
 568                return code;
 569        return finish_command(cmd);
 570}
 571
 572int run_command_v_opt(const char **argv, int opt)
 573{
 574        return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
 575}
 576
 577int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
 578{
 579        struct child_process cmd = CHILD_PROCESS_INIT;
 580        cmd.argv = argv;
 581        cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
 582        cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
 583        cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
 584        cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
 585        cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
 586        cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
 587        cmd.dir = dir;
 588        cmd.env = env;
 589        return run_command(&cmd);
 590}
 591
 592#ifndef NO_PTHREADS
 593static pthread_t main_thread;
 594static int main_thread_set;
 595static pthread_key_t async_key;
 596static pthread_key_t async_die_counter;
 597
 598static void *run_thread(void *data)
 599{
 600        struct async *async = data;
 601        intptr_t ret;
 602
 603        pthread_setspecific(async_key, async);
 604        ret = async->proc(async->proc_in, async->proc_out, async->data);
 605        return (void *)ret;
 606}
 607
 608static NORETURN void die_async(const char *err, va_list params)
 609{
 610        vreportf("fatal: ", err, params);
 611
 612        if (in_async()) {
 613                struct async *async = pthread_getspecific(async_key);
 614                if (async->proc_in >= 0)
 615                        close(async->proc_in);
 616                if (async->proc_out >= 0)
 617                        close(async->proc_out);
 618                pthread_exit((void *)128);
 619        }
 620
 621        exit(128);
 622}
 623
 624static int async_die_is_recursing(void)
 625{
 626        void *ret = pthread_getspecific(async_die_counter);
 627        pthread_setspecific(async_die_counter, (void *)1);
 628        return ret != NULL;
 629}
 630
 631int in_async(void)
 632{
 633        if (!main_thread_set)
 634                return 0; /* no asyncs started yet */
 635        return !pthread_equal(main_thread, pthread_self());
 636}
 637
 638#else
 639
 640static struct {
 641        void (**handlers)(void);
 642        size_t nr;
 643        size_t alloc;
 644} git_atexit_hdlrs;
 645
 646static int git_atexit_installed;
 647
 648static void git_atexit_dispatch(void)
 649{
 650        size_t i;
 651
 652        for (i=git_atexit_hdlrs.nr ; i ; i--)
 653                git_atexit_hdlrs.handlers[i-1]();
 654}
 655
 656static void git_atexit_clear(void)
 657{
 658        free(git_atexit_hdlrs.handlers);
 659        memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
 660        git_atexit_installed = 0;
 661}
 662
 663#undef atexit
 664int git_atexit(void (*handler)(void))
 665{
 666        ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
 667        git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
 668        if (!git_atexit_installed) {
 669                if (atexit(&git_atexit_dispatch))
 670                        return -1;
 671                git_atexit_installed = 1;
 672        }
 673        return 0;
 674}
 675#define atexit git_atexit
 676
 677static int process_is_async;
 678int in_async(void)
 679{
 680        return process_is_async;
 681}
 682
 683#endif
 684
 685int start_async(struct async *async)
 686{
 687        int need_in, need_out;
 688        int fdin[2], fdout[2];
 689        int proc_in, proc_out;
 690
 691        need_in = async->in < 0;
 692        if (need_in) {
 693                if (pipe(fdin) < 0) {
 694                        if (async->out > 0)
 695                                close(async->out);
 696                        return error("cannot create pipe: %s", strerror(errno));
 697                }
 698                async->in = fdin[1];
 699        }
 700
 701        need_out = async->out < 0;
 702        if (need_out) {
 703                if (pipe(fdout) < 0) {
 704                        if (need_in)
 705                                close_pair(fdin);
 706                        else if (async->in)
 707                                close(async->in);
 708                        return error("cannot create pipe: %s", strerror(errno));
 709                }
 710                async->out = fdout[0];
 711        }
 712
 713        if (need_in)
 714                proc_in = fdin[0];
 715        else if (async->in)
 716                proc_in = async->in;
 717        else
 718                proc_in = -1;
 719
 720        if (need_out)
 721                proc_out = fdout[1];
 722        else if (async->out)
 723                proc_out = async->out;
 724        else
 725                proc_out = -1;
 726
 727#ifdef NO_PTHREADS
 728        /* Flush stdio before fork() to avoid cloning buffers */
 729        fflush(NULL);
 730
 731        async->pid = fork();
 732        if (async->pid < 0) {
 733                error("fork (async) failed: %s", strerror(errno));
 734                goto error;
 735        }
 736        if (!async->pid) {
 737                if (need_in)
 738                        close(fdin[1]);
 739                if (need_out)
 740                        close(fdout[0]);
 741                git_atexit_clear();
 742                process_is_async = 1;
 743                exit(!!async->proc(proc_in, proc_out, async->data));
 744        }
 745
 746        mark_child_for_cleanup(async->pid);
 747
 748        if (need_in)
 749                close(fdin[0]);
 750        else if (async->in)
 751                close(async->in);
 752
 753        if (need_out)
 754                close(fdout[1]);
 755        else if (async->out)
 756                close(async->out);
 757#else
 758        if (!main_thread_set) {
 759                /*
 760                 * We assume that the first time that start_async is called
 761                 * it is from the main thread.
 762                 */
 763                main_thread_set = 1;
 764                main_thread = pthread_self();
 765                pthread_key_create(&async_key, NULL);
 766                pthread_key_create(&async_die_counter, NULL);
 767                set_die_routine(die_async);
 768                set_die_is_recursing_routine(async_die_is_recursing);
 769        }
 770
 771        if (proc_in >= 0)
 772                set_cloexec(proc_in);
 773        if (proc_out >= 0)
 774                set_cloexec(proc_out);
 775        async->proc_in = proc_in;
 776        async->proc_out = proc_out;
 777        {
 778                int err = pthread_create(&async->tid, NULL, run_thread, async);
 779                if (err) {
 780                        error("cannot create thread: %s", strerror(err));
 781                        goto error;
 782                }
 783        }
 784#endif
 785        return 0;
 786
 787error:
 788        if (need_in)
 789                close_pair(fdin);
 790        else if (async->in)
 791                close(async->in);
 792
 793        if (need_out)
 794                close_pair(fdout);
 795        else if (async->out)
 796                close(async->out);
 797        return -1;
 798}
 799
 800int finish_async(struct async *async)
 801{
 802#ifdef NO_PTHREADS
 803        return wait_or_whine(async->pid, "child process", 0);
 804#else
 805        void *ret = (void *)(intptr_t)(-1);
 806
 807        if (pthread_join(async->tid, &ret))
 808                error("pthread_join failed");
 809        return (int)(intptr_t)ret;
 810#endif
 811}
 812
 813const char *find_hook(const char *name)
 814{
 815        static struct strbuf path = STRBUF_INIT;
 816
 817        strbuf_reset(&path);
 818        strbuf_git_path(&path, "hooks/%s", name);
 819        if (access(path.buf, X_OK) < 0)
 820                return NULL;
 821        return path.buf;
 822}
 823
 824int run_hook_ve(const char *const *env, const char *name, va_list args)
 825{
 826        struct child_process hook = CHILD_PROCESS_INIT;
 827        const char *p;
 828
 829        p = find_hook(name);
 830        if (!p)
 831                return 0;
 832
 833        argv_array_push(&hook.args, p);
 834        while ((p = va_arg(args, const char *)))
 835                argv_array_push(&hook.args, p);
 836        hook.env = env;
 837        hook.no_stdin = 1;
 838        hook.stdout_to_stderr = 1;
 839
 840        return run_command(&hook);
 841}
 842
 843int run_hook_le(const char *const *env, const char *name, ...)
 844{
 845        va_list args;
 846        int ret;
 847
 848        va_start(args, name);
 849        ret = run_hook_ve(env, name, args);
 850        va_end(args);
 851
 852        return ret;
 853}
 854
 855int capture_command(struct child_process *cmd, struct strbuf *buf, size_t hint)
 856{
 857        cmd->out = -1;
 858        if (start_command(cmd) < 0)
 859                return -1;
 860
 861        if (strbuf_read(buf, cmd->out, hint) < 0) {
 862                close(cmd->out);
 863                finish_command(cmd); /* throw away exit code */
 864                return -1;
 865        }
 866
 867        close(cmd->out);
 868        return finish_command(cmd);
 869}
 870
 871enum child_state {
 872        GIT_CP_FREE,
 873        GIT_CP_WORKING,
 874        GIT_CP_WAIT_CLEANUP,
 875};
 876
 877struct parallel_processes {
 878        void *data;
 879
 880        int max_processes;
 881        int nr_processes;
 882
 883        get_next_task_fn get_next_task;
 884        start_failure_fn start_failure;
 885        task_finished_fn task_finished;
 886
 887        struct {
 888                enum child_state state;
 889                struct child_process process;
 890                struct strbuf err;
 891                void *data;
 892        } *children;
 893        /*
 894         * The struct pollfd is logically part of *children,
 895         * but the system call expects it as its own array.
 896         */
 897        struct pollfd *pfd;
 898
 899        unsigned shutdown : 1;
 900
 901        int output_owner;
 902        struct strbuf buffered_output; /* of finished children */
 903};
 904
 905static int default_start_failure(struct child_process *cp,
 906                                 struct strbuf *err,
 907                                 void *pp_cb,
 908                                 void *pp_task_cb)
 909{
 910        int i;
 911
 912        strbuf_addstr(err, "Starting a child failed:");
 913        for (i = 0; cp->argv[i]; i++)
 914                strbuf_addf(err, " %s", cp->argv[i]);
 915
 916        return 0;
 917}
 918
 919static int default_task_finished(int result,
 920                                 struct child_process *cp,
 921                                 struct strbuf *err,
 922                                 void *pp_cb,
 923                                 void *pp_task_cb)
 924{
 925        int i;
 926
 927        if (!result)
 928                return 0;
 929
 930        strbuf_addf(err, "A child failed with return code %d:", result);
 931        for (i = 0; cp->argv[i]; i++)
 932                strbuf_addf(err, " %s", cp->argv[i]);
 933
 934        return 0;
 935}
 936
 937static void kill_children(struct parallel_processes *pp, int signo)
 938{
 939        int i, n = pp->max_processes;
 940
 941        for (i = 0; i < n; i++)
 942                if (pp->children[i].state == GIT_CP_WORKING)
 943                        kill(pp->children[i].process.pid, signo);
 944}
 945
 946static struct parallel_processes *pp_for_signal;
 947
 948static void handle_children_on_signal(int signo)
 949{
 950        kill_children(pp_for_signal, signo);
 951        sigchain_pop(signo);
 952        raise(signo);
 953}
 954
 955static void pp_init(struct parallel_processes *pp,
 956                    int n,
 957                    get_next_task_fn get_next_task,
 958                    start_failure_fn start_failure,
 959                    task_finished_fn task_finished,
 960                    void *data)
 961{
 962        int i;
 963
 964        if (n < 1)
 965                n = online_cpus();
 966
 967        pp->max_processes = n;
 968
 969        trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
 970
 971        pp->data = data;
 972        if (!get_next_task)
 973                die("BUG: you need to specify a get_next_task function");
 974        pp->get_next_task = get_next_task;
 975
 976        pp->start_failure = start_failure ? start_failure : default_start_failure;
 977        pp->task_finished = task_finished ? task_finished : default_task_finished;
 978
 979        pp->nr_processes = 0;
 980        pp->output_owner = 0;
 981        pp->shutdown = 0;
 982        pp->children = xcalloc(n, sizeof(*pp->children));
 983        pp->pfd = xcalloc(n, sizeof(*pp->pfd));
 984        strbuf_init(&pp->buffered_output, 0);
 985
 986        for (i = 0; i < n; i++) {
 987                strbuf_init(&pp->children[i].err, 0);
 988                child_process_init(&pp->children[i].process);
 989                pp->pfd[i].events = POLLIN | POLLHUP;
 990                pp->pfd[i].fd = -1;
 991        }
 992
 993        pp_for_signal = pp;
 994        sigchain_push_common(handle_children_on_signal);
 995}
 996
 997static void pp_cleanup(struct parallel_processes *pp)
 998{
 999        int i;
1000
1001        trace_printf("run_processes_parallel: done");
1002        for (i = 0; i < pp->max_processes; i++) {
1003                strbuf_release(&pp->children[i].err);
1004                child_process_clear(&pp->children[i].process);
1005        }
1006
1007        free(pp->children);
1008        free(pp->pfd);
1009
1010        /*
1011         * When get_next_task added messages to the buffer in its last
1012         * iteration, the buffered output is non empty.
1013         */
1014        fputs(pp->buffered_output.buf, stderr);
1015        strbuf_release(&pp->buffered_output);
1016
1017        sigchain_pop_common();
1018}
1019
1020/* returns
1021 *  0 if a new task was started.
1022 *  1 if no new jobs was started (get_next_task ran out of work, non critical
1023 *    problem with starting a new command)
1024 * <0 no new job was started, user wishes to shutdown early. Use negative code
1025 *    to signal the children.
1026 */
1027static int pp_start_one(struct parallel_processes *pp)
1028{
1029        int i, code;
1030
1031        for (i = 0; i < pp->max_processes; i++)
1032                if (pp->children[i].state == GIT_CP_FREE)
1033                        break;
1034        if (i == pp->max_processes)
1035                die("BUG: bookkeeping is hard");
1036
1037        code = pp->get_next_task(&pp->children[i].process,
1038                                 &pp->children[i].err,
1039                                 pp->data,
1040                                 &pp->children[i].data);
1041        if (!code) {
1042                strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1043                strbuf_reset(&pp->children[i].err);
1044                return 1;
1045        }
1046        pp->children[i].process.err = -1;
1047        pp->children[i].process.stdout_to_stderr = 1;
1048        pp->children[i].process.no_stdin = 1;
1049
1050        if (start_command(&pp->children[i].process)) {
1051                code = pp->start_failure(&pp->children[i].process,
1052                                         &pp->children[i].err,
1053                                         pp->data,
1054                                         &pp->children[i].data);
1055                strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1056                strbuf_reset(&pp->children[i].err);
1057                if (code)
1058                        pp->shutdown = 1;
1059                return code;
1060        }
1061
1062        pp->nr_processes++;
1063        pp->children[i].state = GIT_CP_WORKING;
1064        pp->pfd[i].fd = pp->children[i].process.err;
1065        return 0;
1066}
1067
1068static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1069{
1070        int i;
1071
1072        while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1073                if (errno == EINTR)
1074                        continue;
1075                pp_cleanup(pp);
1076                die_errno("poll");
1077        }
1078
1079        /* Buffer output from all pipes. */
1080        for (i = 0; i < pp->max_processes; i++) {
1081                if (pp->children[i].state == GIT_CP_WORKING &&
1082                    pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1083                        int n = strbuf_read_once(&pp->children[i].err,
1084                                                 pp->children[i].process.err, 0);
1085                        if (n == 0) {
1086                                close(pp->children[i].process.err);
1087                                pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1088                        } else if (n < 0)
1089                                if (errno != EAGAIN)
1090                                        die_errno("read");
1091                }
1092        }
1093}
1094
1095static void pp_output(struct parallel_processes *pp)
1096{
1097        int i = pp->output_owner;
1098        if (pp->children[i].state == GIT_CP_WORKING &&
1099            pp->children[i].err.len) {
1100                fputs(pp->children[i].err.buf, stderr);
1101                strbuf_reset(&pp->children[i].err);
1102        }
1103}
1104
1105static int pp_collect_finished(struct parallel_processes *pp)
1106{
1107        int i, code;
1108        int n = pp->max_processes;
1109        int result = 0;
1110
1111        while (pp->nr_processes > 0) {
1112                for (i = 0; i < pp->max_processes; i++)
1113                        if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1114                                break;
1115                if (i == pp->max_processes)
1116                        break;
1117
1118                code = finish_command(&pp->children[i].process);
1119
1120                code = pp->task_finished(code, &pp->children[i].process,
1121                                         &pp->children[i].err, pp->data,
1122                                         &pp->children[i].data);
1123
1124                if (code)
1125                        result = code;
1126                if (code < 0)
1127                        break;
1128
1129                pp->nr_processes--;
1130                pp->children[i].state = GIT_CP_FREE;
1131                pp->pfd[i].fd = -1;
1132                child_process_init(&pp->children[i].process);
1133
1134                if (i != pp->output_owner) {
1135                        strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1136                        strbuf_reset(&pp->children[i].err);
1137                } else {
1138                        fputs(pp->children[i].err.buf, stderr);
1139                        strbuf_reset(&pp->children[i].err);
1140
1141                        /* Output all other finished child processes */
1142                        fputs(pp->buffered_output.buf, stderr);
1143                        strbuf_reset(&pp->buffered_output);
1144
1145                        /*
1146                         * Pick next process to output live.
1147                         * NEEDSWORK:
1148                         * For now we pick it randomly by doing a round
1149                         * robin. Later we may want to pick the one with
1150                         * the most output or the longest or shortest
1151                         * running process time.
1152                         */
1153                        for (i = 0; i < n; i++)
1154                                if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1155                                        break;
1156                        pp->output_owner = (pp->output_owner + i) % n;
1157                }
1158        }
1159        return result;
1160}
1161
1162int run_processes_parallel(int n,
1163                           get_next_task_fn get_next_task,
1164                           start_failure_fn start_failure,
1165                           task_finished_fn task_finished,
1166                           void *pp_cb)
1167{
1168        int i, code;
1169        int output_timeout = 100;
1170        int spawn_cap = 4;
1171        struct parallel_processes pp;
1172
1173        pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1174        while (1) {
1175                for (i = 0;
1176                    i < spawn_cap && !pp.shutdown &&
1177                    pp.nr_processes < pp.max_processes;
1178                    i++) {
1179                        code = pp_start_one(&pp);
1180                        if (!code)
1181                                continue;
1182                        if (code < 0) {
1183                                pp.shutdown = 1;
1184                                kill_children(&pp, -code);
1185                        }
1186                        break;
1187                }
1188                if (!pp.nr_processes)
1189                        break;
1190                pp_buffer_stderr(&pp, output_timeout);
1191                pp_output(&pp);
1192                code = pp_collect_finished(&pp);
1193                if (code) {
1194                        pp.shutdown = 1;
1195                        if (code < 0)
1196                                kill_children(&pp, -code);
1197                }
1198        }
1199
1200        pp_cleanup(&pp);
1201        return 0;
1202}