git.oblomov.eu Git - git/blob - run-command.c

   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
   9 void 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
  16 void child_process_clear(struct child_process *child)
  17 {
  18         argv_array_clear(&child->args);
  19         argv_array_clear(&child->env_array);
  20 }
  21
  22 struct child_to_clean {
  23         pid_t pid;
  24         struct child_to_clean *next;
  25 };
  26 static struct child_to_clean *children_to_clean;
  27 static int installed_child_cleanup_handler;
  28
  29 static 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
  40 static void cleanup_children_on_signal(int sig)
  41 {
  42         cleanup_children(sig, 1);
  43         sigchain_pop(sig);
  44         raise(sig);
  45 }
  46
  47 static void cleanup_children_on_exit(void)
  48 {
  49         cleanup_children(SIGTERM, 0);
  50 }
  51
  52 static 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
  66 static 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
  81 static inline void close_pair(int fd[2])
  82 {
  83         close(fd[0]);
  84         close(fd[1]);
  85 }
  86
  87 #ifndef GIT_WINDOWS_NATIVE
  88 static 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
  99 static 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
 131 static int exists_in_PATH(const char *file)
 132 {
 133         char *r = locate_in_PATH(file);
 134         free(r);
 135         return r != NULL;
 136 }
 137
 138 int 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 $PATH
 151          * lookups in that case.
 152          *
 153          * The reassignment of EACCES to errno looks like a no-op below,
 154          * but we need to protect against exists_in_PATH overwriting errno.
 155          */
 156         if (errno == EACCES && !strchr(file, '/'))
 157                 errno = exists_in_PATH(file) ? EACCES : ENOENT;
 158         else if (errno == ENOTDIR && !strchr(file, '/'))
 159                 errno = ENOENT;
 160         return -1;
 161 }
 162
 163 static const char **prepare_shell_cmd(struct argv_array *out, const char **argv)
 164 {
 165         if (!argv[0])
 166                 die("BUG: shell command is empty");
 167
 168         if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
 169 #ifndef GIT_WINDOWS_NATIVE
 170                 argv_array_push(out, SHELL_PATH);
 171 #else
 172                 argv_array_push(out, "sh");
 173 #endif
 174                 argv_array_push(out, "-c");
 175
 176                 /*
 177                  * If we have no extra arguments, we do not even need to
 178                  * bother with the "$@" magic.
 179                  */
 180                 if (!argv[1])
 181                         argv_array_push(out, argv[0]);
 182                 else
 183                         argv_array_pushf(out, "%s \"$@\"", argv[0]);
 184         }
 185
 186         argv_array_pushv(out, argv);
 187         return out->argv;
 188 }
 189
 190 #ifndef GIT_WINDOWS_NATIVE
 191 static int execv_shell_cmd(const char **argv)
 192 {
 193         struct argv_array nargv = ARGV_ARRAY_INIT;
 194         prepare_shell_cmd(&nargv, argv);
 195         trace_argv_printf(nargv.argv, "trace: exec:");
 196         sane_execvp(nargv.argv[0], (char **)nargv.argv);
 197         argv_array_clear(&nargv);
 198         return -1;
 199 }
 200 #endif
 201
 202 #ifndef GIT_WINDOWS_NATIVE
 203 static int child_notifier = -1;
 204
 205 static void notify_parent(void)
 206 {
 207         /*
 208          * execvp failed.  If possible, we'd like to let start_command
 209          * know, so failures like ENOENT can be handled right away; but
 210          * otherwise, finish_command will still report the error.
 211          */
 212         xwrite(child_notifier, "", 1);
 213 }
 214 #endif
 215
 216 static inline void set_cloexec(int fd)
 217 {
 218         int flags = fcntl(fd, F_GETFD);
 219         if (flags >= 0)
 220                 fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
 221 }
 222
 223 static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
 224 {
 225         int status, code = -1;
 226         pid_t waiting;
 227         int failed_errno = 0;
 228
 229         while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
 230                 ;       /* nothing */
 231         if (in_signal)
 232                 return 0;
 233
 234         if (waiting < 0) {
 235                 failed_errno = errno;
 236                 error_errno("waitpid for %s failed", argv0);
 237         } else if (waiting != pid) {
 238                 error("waitpid is confused (%s)", argv0);
 239         } else if (WIFSIGNALED(status)) {
 240                 code = WTERMSIG(status);
 241                 if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
 242                         error("%s died of signal %d", argv0, code);
 243                 /*
 244                  * This return value is chosen so that code & 0xff
 245                  * mimics the exit code that a POSIX shell would report for
 246                  * a program that died from this signal.
 247                  */
 248                 code += 128;
 249         } else if (WIFEXITED(status)) {
 250                 code = WEXITSTATUS(status);
 251                 /*
 252                  * Convert special exit code when execvp failed.
 253                  */
 254                 if (code == 127) {
 255                         code = -1;
 256                         failed_errno = ENOENT;
 257                 }
 258         } else {
 259                 error("waitpid is confused (%s)", argv0);
 260         }
 261
 262         clear_child_for_cleanup(pid);
 263
 264         errno = failed_errno;
 265         return code;
 266 }
 267
 268 int start_command(struct child_process *cmd)
 269 {
 270         int need_in, need_out, need_err;
 271         int fdin[2], fdout[2], fderr[2];
 272         int failed_errno;
 273         char *str;
 274
 275         if (!cmd->argv)
 276                 cmd->argv = cmd->args.argv;
 277         if (!cmd->env)
 278                 cmd->env = cmd->env_array.argv;
 279
 280         /*
 281          * In case of errors we must keep the promise to close FDs
 282          * that have been passed in via ->in and ->out.
 283          */
 284
 285         need_in = !cmd->no_stdin && cmd->in < 0;
 286         if (need_in) {
 287                 if (pipe(fdin) < 0) {
 288                         failed_errno = errno;
 289                         if (cmd->out > 0)
 290                                 close(cmd->out);
 291                         str = "standard input";
 292                         goto fail_pipe;
 293                 }
 294                 cmd->in = fdin[1];
 295         }
 296
 297         need_out = !cmd->no_stdout
 298                 && !cmd->stdout_to_stderr
 299                 && cmd->out < 0;
 300         if (need_out) {
 301                 if (pipe(fdout) < 0) {
 302                         failed_errno = errno;
 303                         if (need_in)
 304                                 close_pair(fdin);
 305                         else if (cmd->in)
 306                                 close(cmd->in);
 307                         str = "standard output";
 308                         goto fail_pipe;
 309                 }
 310                 cmd->out = fdout[0];
 311         }
 312
 313         need_err = !cmd->no_stderr && cmd->err < 0;
 314         if (need_err) {
 315                 if (pipe(fderr) < 0) {
 316                         failed_errno = errno;
 317                         if (need_in)
 318                                 close_pair(fdin);
 319                         else if (cmd->in)
 320                                 close(cmd->in);
 321                         if (need_out)
 322                                 close_pair(fdout);
 323                         else if (cmd->out)
 324                                 close(cmd->out);
 325                         str = "standard error";
 326 fail_pipe:
 327                         error("cannot create %s pipe for %s: %s",
 328                                 str, cmd->argv[0], strerror(failed_errno));
 329                         child_process_clear(cmd);
 330                         errno = failed_errno;
 331                         return -1;
 332                 }
 333                 cmd->err = fderr[0];
 334         }
 335
 336         trace_argv_printf(cmd->argv, "trace: run_command:");
 337         fflush(NULL);
 338
 339 #ifndef GIT_WINDOWS_NATIVE
 340 {
 341         int notify_pipe[2];
 342         if (pipe(notify_pipe))
 343                 notify_pipe[0] = notify_pipe[1] = -1;
 344
 345         cmd->pid = fork();
 346         failed_errno = errno;
 347         if (!cmd->pid) {
 348                 /*
 349                  * Redirect the channel to write syscall error messages to
 350                  * before redirecting the process's stderr so that all die()
 351                  * in subsequent call paths use the parent's stderr.
 352                  */
 353                 if (cmd->no_stderr || need_err) {
 354                         int child_err = dup(2);
 355                         set_cloexec(child_err);
 356                         set_error_handle(fdopen(child_err, "w"));
 357                 }
 358
 359                 close(notify_pipe[0]);
 360                 set_cloexec(notify_pipe[1]);
 361                 child_notifier = notify_pipe[1];
 362                 atexit(notify_parent);
 363
 364                 if (cmd->no_stdin)
 365                         dup_devnull(0);
 366                 else if (need_in) {
 367                         dup2(fdin[0], 0);
 368                         close_pair(fdin);
 369                 } else if (cmd->in) {
 370                         dup2(cmd->in, 0);
 371                         close(cmd->in);
 372                 }
 373
 374                 if (cmd->no_stderr)
 375                         dup_devnull(2);
 376                 else if (need_err) {
 377                         dup2(fderr[1], 2);
 378                         close_pair(fderr);
 379                 } else if (cmd->err > 1) {
 380                         dup2(cmd->err, 2);
 381                         close(cmd->err);
 382                 }
 383
 384                 if (cmd->no_stdout)
 385                         dup_devnull(1);
 386                 else if (cmd->stdout_to_stderr)
 387                         dup2(2, 1);
 388                 else if (need_out) {
 389                         dup2(fdout[1], 1);
 390                         close_pair(fdout);
 391                 } else if (cmd->out > 1) {
 392                         dup2(cmd->out, 1);
 393                         close(cmd->out);
 394                 }
 395
 396                 if (cmd->dir && chdir(cmd->dir))
 397                         die_errno("exec '%s': cd to '%s' failed", cmd->argv[0],
 398                             cmd->dir);
 399                 if (cmd->env) {
 400                         for (; *cmd->env; cmd->env++) {
 401                                 if (strchr(*cmd->env, '='))
 402                                         putenv((char *)*cmd->env);
 403                                 else
 404                                         unsetenv(*cmd->env);
 405                         }
 406                 }
 407                 if (cmd->git_cmd)
 408                         execv_git_cmd(cmd->argv);
 409                 else if (cmd->use_shell)
 410                         execv_shell_cmd(cmd->argv);
 411                 else
 412                         sane_execvp(cmd->argv[0], (char *const*) cmd->argv);
 413                 if (errno == ENOENT) {
 414                         if (!cmd->silent_exec_failure)
 415                                 error("cannot run %s: %s", cmd->argv[0],
 416                                         strerror(ENOENT));
 417                         exit(127);
 418                 } else {
 419                         die_errno("cannot exec '%s'", cmd->argv[0]);
 420                 }
 421         }
 422         if (cmd->pid < 0)
 423                 error_errno("cannot fork() for %s", cmd->argv[0]);
 424         else if (cmd->clean_on_exit)
 425                 mark_child_for_cleanup(cmd->pid);
 426
 427         /*
 428          * Wait for child's execvp. If the execvp succeeds (or if fork()
 429          * failed), EOF is seen immediately by the parent. Otherwise, the
 430          * child process sends a single byte.
 431          * Note that use of this infrastructure is completely advisory,
 432          * therefore, we keep error checks minimal.
 433          */
 434         close(notify_pipe[1]);
 435         if (read(notify_pipe[0], &notify_pipe[1], 1) == 1) {
 436                 /*
 437                  * At this point we know that fork() succeeded, but execvp()
 438                  * failed. Errors have been reported to our stderr.
 439                  */
 440                 wait_or_whine(cmd->pid, cmd->argv[0], 0);
 441                 failed_errno = errno;
 442                 cmd->pid = -1;
 443         }
 444         close(notify_pipe[0]);
 445 }
 446 #else
 447 {
 448         int fhin = 0, fhout = 1, fherr = 2;
 449         const char **sargv = cmd->argv;
 450         struct argv_array nargv = ARGV_ARRAY_INIT;
 451
 452         if (cmd->no_stdin)
 453                 fhin = open("/dev/null", O_RDWR);
 454         else if (need_in)
 455                 fhin = dup(fdin[0]);
 456         else if (cmd->in)
 457                 fhin = dup(cmd->in);
 458
 459         if (cmd->no_stderr)
 460                 fherr = open("/dev/null", O_RDWR);
 461         else if (need_err)
 462                 fherr = dup(fderr[1]);
 463         else if (cmd->err > 2)
 464                 fherr = dup(cmd->err);
 465
 466         if (cmd->no_stdout)
 467                 fhout = open("/dev/null", O_RDWR);
 468         else if (cmd->stdout_to_stderr)
 469                 fhout = dup(fherr);
 470         else if (need_out)
 471                 fhout = dup(fdout[1]);
 472         else if (cmd->out > 1)
 473                 fhout = dup(cmd->out);
 474
 475         if (cmd->git_cmd)
 476                 cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
 477         else if (cmd->use_shell)
 478                 cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
 479
 480         cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
 481                         cmd->dir, fhin, fhout, fherr);
 482         failed_errno = errno;
 483         if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
 484                 error_errno("cannot spawn %s", cmd->argv[0]);
 485         if (cmd->clean_on_exit && cmd->pid >= 0)
 486                 mark_child_for_cleanup(cmd->pid);
 487
 488         argv_array_clear(&nargv);
 489         cmd->argv = sargv;
 490         if (fhin != 0)
 491                 close(fhin);
 492         if (fhout != 1)
 493                 close(fhout);
 494         if (fherr != 2)
 495                 close(fherr);
 496 }
 497 #endif
 498
 499         if (cmd->pid < 0) {
 500                 if (need_in)
 501                         close_pair(fdin);
 502                 else if (cmd->in)
 503                         close(cmd->in);
 504                 if (need_out)
 505                         close_pair(fdout);
 506                 else if (cmd->out)
 507                         close(cmd->out);
 508                 if (need_err)
 509                         close_pair(fderr);
 510                 else if (cmd->err)
 511                         close(cmd->err);
 512                 child_process_clear(cmd);
 513                 errno = failed_errno;
 514                 return -1;
 515         }
 516
 517         if (need_in)
 518                 close(fdin[0]);
 519         else if (cmd->in)
 520                 close(cmd->in);
 521
 522         if (need_out)
 523                 close(fdout[1]);
 524         else if (cmd->out)
 525                 close(cmd->out);
 526
 527         if (need_err)
 528                 close(fderr[1]);
 529         else if (cmd->err)
 530                 close(cmd->err);
 531
 532         return 0;
 533 }
 534
 535 int finish_command(struct child_process *cmd)
 536 {
 537         int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
 538         child_process_clear(cmd);
 539         return ret;
 540 }
 541
 542 int finish_command_in_signal(struct child_process *cmd)
 543 {
 544         return wait_or_whine(cmd->pid, cmd->argv[0], 1);
 545 }
 546
 547
 548 int run_command(struct child_process *cmd)
 549 {
 550         int code;
 551
 552         if (cmd->out < 0 || cmd->err < 0)
 553                 die("BUG: run_command with a pipe can cause deadlock");
 554
 555         code = start_command(cmd);
 556         if (code)
 557                 return code;
 558         return finish_command(cmd);
 559 }
 560
 561 int run_command_v_opt(const char **argv, int opt)
 562 {
 563         return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
 564 }
 565
 566 int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
 567 {
 568         struct child_process cmd = CHILD_PROCESS_INIT;
 569         cmd.argv = argv;
 570         cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
 571         cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
 572         cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
 573         cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
 574         cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
 575         cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
 576         cmd.dir = dir;
 577         cmd.env = env;
 578         return run_command(&cmd);
 579 }
 580
 581 #ifndef NO_PTHREADS
 582 static pthread_t main_thread;
 583 static int main_thread_set;
 584 static pthread_key_t async_key;
 585 static pthread_key_t async_die_counter;
 586
 587 static void *run_thread(void *data)
 588 {
 589         struct async *async = data;
 590         intptr_t ret;
 591
 592         pthread_setspecific(async_key, async);
 593         ret = async->proc(async->proc_in, async->proc_out, async->data);
 594         return (void *)ret;
 595 }
 596
 597 static NORETURN void die_async(const char *err, va_list params)
 598 {
 599         vreportf("fatal: ", err, params);
 600
 601         if (in_async()) {
 602                 struct async *async = pthread_getspecific(async_key);
 603                 if (async->proc_in >= 0)
 604                         close(async->proc_in);
 605                 if (async->proc_out >= 0)
 606                         close(async->proc_out);
 607                 pthread_exit((void *)128);
 608         }
 609
 610         exit(128);
 611 }
 612
 613 static int async_die_is_recursing(void)
 614 {
 615         void *ret = pthread_getspecific(async_die_counter);
 616         pthread_setspecific(async_die_counter, (void *)1);
 617         return ret != NULL;
 618 }
 619
 620 int in_async(void)
 621 {
 622         if (!main_thread_set)
 623                 return 0; /* no asyncs started yet */
 624         return !pthread_equal(main_thread, pthread_self());
 625 }
 626
 627 void NORETURN async_exit(int code)
 628 {
 629         pthread_exit((void *)(intptr_t)code);
 630 }
 631
 632 #else
 633
 634 static struct {
 635         void (**handlers)(void);
 636         size_t nr;
 637         size_t alloc;
 638 } git_atexit_hdlrs;
 639
 640 static int git_atexit_installed;
 641
 642 static void git_atexit_dispatch(void)
 643 {
 644         size_t i;
 645
 646         for (i=git_atexit_hdlrs.nr ; i ; i--)
 647                 git_atexit_hdlrs.handlers[i-1]();
 648 }
 649
 650 static void git_atexit_clear(void)
 651 {
 652         free(git_atexit_hdlrs.handlers);
 653         memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
 654         git_atexit_installed = 0;
 655 }
 656
 657 #undef atexit
 658 int git_atexit(void (*handler)(void))
 659 {
 660         ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
 661         git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
 662         if (!git_atexit_installed) {
 663                 if (atexit(&git_atexit_dispatch))
 664                         return -1;
 665                 git_atexit_installed = 1;
 666         }
 667         return 0;
 668 }
 669 #define atexit git_atexit
 670
 671 static int process_is_async;
 672 int in_async(void)
 673 {
 674         return process_is_async;
 675 }
 676
 677 void NORETURN async_exit(int code)
 678 {
 679         exit(code);
 680 }
 681
 682 #endif
 683
 684 int start_async(struct async *async)
 685 {
 686         int need_in, need_out;
 687         int fdin[2], fdout[2];
 688         int proc_in, proc_out;
 689
 690         need_in = async->in < 0;
 691         if (need_in) {
 692                 if (pipe(fdin) < 0) {
 693                         if (async->out > 0)
 694                                 close(async->out);
 695                         return error_errno("cannot create pipe");
 696                 }
 697                 async->in = fdin[1];
 698         }
 699
 700         need_out = async->out < 0;
 701         if (need_out) {
 702                 if (pipe(fdout) < 0) {
 703                         if (need_in)
 704                                 close_pair(fdin);
 705                         else if (async->in)
 706                                 close(async->in);
 707                         return error_errno("cannot create pipe");
 708                 }
 709                 async->out = fdout[0];
 710         }
 711
 712         if (need_in)
 713                 proc_in = fdin[0];
 714         else if (async->in)
 715                 proc_in = async->in;
 716         else
 717                 proc_in = -1;
 718
 719         if (need_out)
 720                 proc_out = fdout[1];
 721         else if (async->out)
 722                 proc_out = async->out;
 723         else
 724                 proc_out = -1;
 725
 726 #ifdef NO_PTHREADS
 727         /* Flush stdio before fork() to avoid cloning buffers */
 728         fflush(NULL);
 729
 730         async->pid = fork();
 731         if (async->pid < 0) {
 732                 error_errno("fork (async) failed");
 733                 goto error;
 734         }
 735         if (!async->pid) {
 736                 if (need_in)
 737                         close(fdin[1]);
 738                 if (need_out)
 739                         close(fdout[0]);
 740                 git_atexit_clear();
 741                 process_is_async = 1;
 742                 exit(!!async->proc(proc_in, proc_out, async->data));
 743         }
 744
 745         mark_child_for_cleanup(async->pid);
 746
 747         if (need_in)
 748                 close(fdin[0]);
 749         else if (async->in)
 750                 close(async->in);
 751
 752         if (need_out)
 753                 close(fdout[1]);
 754         else if (async->out)
 755                 close(async->out);
 756 #else
 757         if (!main_thread_set) {
 758                 /*
 759                  * We assume that the first time that start_async is called
 760                  * it is from the main thread.
 761                  */
 762                 main_thread_set = 1;
 763                 main_thread = pthread_self();
 764                 pthread_key_create(&async_key, NULL);
 765                 pthread_key_create(&async_die_counter, NULL);
 766                 set_die_routine(die_async);
 767                 set_die_is_recursing_routine(async_die_is_recursing);
 768         }
 769
 770         if (proc_in >= 0)
 771                 set_cloexec(proc_in);
 772         if (proc_out >= 0)
 773                 set_cloexec(proc_out);
 774         async->proc_in = proc_in;
 775         async->proc_out = proc_out;
 776         {
 777                 int err = pthread_create(&async->tid, NULL, run_thread, async);
 778                 if (err) {
 779                         error_errno("cannot create thread");
 780                         goto error;
 781                 }
 782         }
 783 #endif
 784         return 0;
 785
 786 error:
 787         if (need_in)
 788                 close_pair(fdin);
 789         else if (async->in)
 790                 close(async->in);
 791
 792         if (need_out)
 793                 close_pair(fdout);
 794         else if (async->out)
 795                 close(async->out);
 796         return -1;
 797 }
 798
 799 int finish_async(struct async *async)
 800 {
 801 #ifdef NO_PTHREADS
 802         return wait_or_whine(async->pid, "child process", 0);
 803 #else
 804         void *ret = (void *)(intptr_t)(-1);
 805
 806         if (pthread_join(async->tid, &ret))
 807                 error("pthread_join failed");
 808         return (int)(intptr_t)ret;
 809 #endif
 810 }
 811
 812 const char *find_hook(const char *name)
 813 {
 814         static struct strbuf path = STRBUF_INIT;
 815
 816         strbuf_reset(&path);
 817         strbuf_git_path(&path, "hooks/%s", name);
 818         if (access(path.buf, X_OK) < 0)
 819                 return NULL;
 820         return path.buf;
 821 }
 822
 823 int run_hook_ve(const char *const *env, const char *name, va_list args)
 824 {
 825         struct child_process hook = CHILD_PROCESS_INIT;
 826         const char *p;
 827
 828         p = find_hook(name);
 829         if (!p)
 830                 return 0;
 831
 832         argv_array_push(&hook.args, p);
 833         while ((p = va_arg(args, const char *)))
 834                 argv_array_push(&hook.args, p);
 835         hook.env = env;
 836         hook.no_stdin = 1;
 837         hook.stdout_to_stderr = 1;
 838
 839         return run_command(&hook);
 840 }
 841
 842 int run_hook_le(const char *const *env, const char *name, ...)
 843 {
 844         va_list args;
 845         int ret;
 846
 847         va_start(args, name);
 848         ret = run_hook_ve(env, name, args);
 849         va_end(args);
 850
 851         return ret;
 852 }
 853
 854 int capture_command(struct child_process *cmd, struct strbuf *buf, size_t hint)
 855 {
 856         cmd->out = -1;
 857         if (start_command(cmd) < 0)
 858                 return -1;
 859
 860         if (strbuf_read(buf, cmd->out, hint) < 0) {
 861                 close(cmd->out);
 862                 finish_command(cmd); /* throw away exit code */
 863                 return -1;
 864         }
 865
 866         close(cmd->out);
 867         return finish_command(cmd);
 868 }
 869
 870 enum child_state {
 871         GIT_CP_FREE,
 872         GIT_CP_WORKING,
 873         GIT_CP_WAIT_CLEANUP,
 874 };
 875
 876 struct parallel_processes {
 877         void *data;
 878
 879         int max_processes;
 880         int nr_processes;
 881
 882         get_next_task_fn get_next_task;
 883         start_failure_fn start_failure;
 884         task_finished_fn task_finished;
 885
 886         struct {
 887                 enum child_state state;
 888                 struct child_process process;
 889                 struct strbuf err;
 890                 void *data;
 891         } *children;
 892         /*
 893          * The struct pollfd is logically part of *children,
 894          * but the system call expects it as its own array.
 895          */
 896         struct pollfd *pfd;
 897
 898         unsigned shutdown : 1;
 899
 900         int output_owner;
 901         struct strbuf buffered_output; /* of finished children */
 902 };
 903
 904 static int default_start_failure(struct strbuf *err,
 905                                  void *pp_cb,
 906                                  void *pp_task_cb)
 907 {
 908         return 0;
 909 }
 910
 911 static int default_task_finished(int result,
 912                                  struct strbuf *err,
 913                                  void *pp_cb,
 914                                  void *pp_task_cb)
 915 {
 916         return 0;
 917 }
 918
 919 static void kill_children(struct parallel_processes *pp, int signo)
 920 {
 921         int i, n = pp->max_processes;
 922
 923         for (i = 0; i < n; i++)
 924                 if (pp->children[i].state == GIT_CP_WORKING)
 925                         kill(pp->children[i].process.pid, signo);
 926 }
 927
 928 static struct parallel_processes *pp_for_signal;
 929
 930 static void handle_children_on_signal(int signo)
 931 {
 932         kill_children(pp_for_signal, signo);
 933         sigchain_pop(signo);
 934         raise(signo);
 935 }
 936
 937 static void pp_init(struct parallel_processes *pp,
 938                     int n,
 939                     get_next_task_fn get_next_task,
 940                     start_failure_fn start_failure,
 941                     task_finished_fn task_finished,
 942                     void *data)
 943 {
 944         int i;
 945
 946         if (n < 1)
 947                 n = online_cpus();
 948
 949         pp->max_processes = n;
 950
 951         trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
 952
 953         pp->data = data;
 954         if (!get_next_task)
 955                 die("BUG: you need to specify a get_next_task function");
 956         pp->get_next_task = get_next_task;
 957
 958         pp->start_failure = start_failure ? start_failure : default_start_failure;
 959         pp->task_finished = task_finished ? task_finished : default_task_finished;
 960
 961         pp->nr_processes = 0;
 962         pp->output_owner = 0;
 963         pp->shutdown = 0;
 964         pp->children = xcalloc(n, sizeof(*pp->children));
 965         pp->pfd = xcalloc(n, sizeof(*pp->pfd));
 966         strbuf_init(&pp->buffered_output, 0);
 967
 968         for (i = 0; i < n; i++) {
 969                 strbuf_init(&pp->children[i].err, 0);
 970                 child_process_init(&pp->children[i].process);
 971                 pp->pfd[i].events = POLLIN | POLLHUP;
 972                 pp->pfd[i].fd = -1;
 973         }
 974
 975         pp_for_signal = pp;
 976         sigchain_push_common(handle_children_on_signal);
 977 }
 978
 979 static void pp_cleanup(struct parallel_processes *pp)
 980 {
 981         int i;
 982
 983         trace_printf("run_processes_parallel: done");
 984         for (i = 0; i < pp->max_processes; i++) {
 985                 strbuf_release(&pp->children[i].err);
 986                 child_process_clear(&pp->children[i].process);
 987         }
 988
 989         free(pp->children);
 990         free(pp->pfd);
 991
 992         /*
 993          * When get_next_task added messages to the buffer in its last
 994          * iteration, the buffered output is non empty.
 995          */
 996         fputs(pp->buffered_output.buf, stderr);
 997         strbuf_release(&pp->buffered_output);
 998
 999         sigchain_pop_common();
1000 }
1001
1002 /* returns
1003  *  0 if a new task was started.
1004  *  1 if no new jobs was started (get_next_task ran out of work, non critical
1005  *    problem with starting a new command)
1006  * <0 no new job was started, user wishes to shutdown early. Use negative code
1007  *    to signal the children.
1008  */
1009 static int pp_start_one(struct parallel_processes *pp)
1010 {
1011         int i, code;
1012
1013         for (i = 0; i < pp->max_processes; i++)
1014                 if (pp->children[i].state == GIT_CP_FREE)
1015                         break;
1016         if (i == pp->max_processes)
1017                 die("BUG: bookkeeping is hard");
1018
1019         code = pp->get_next_task(&pp->children[i].process,
1020                                  &pp->children[i].err,
1021                                  pp->data,
1022                                  &pp->children[i].data);
1023         if (!code) {
1024                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1025                 strbuf_reset(&pp->children[i].err);
1026                 return 1;
1027         }
1028         pp->children[i].process.err = -1;
1029         pp->children[i].process.stdout_to_stderr = 1;
1030         pp->children[i].process.no_stdin = 1;
1031
1032         if (start_command(&pp->children[i].process)) {
1033                 code = pp->start_failure(&pp->children[i].err,
1034                                          pp->data,
1035                                          &pp->children[i].data);
1036                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1037                 strbuf_reset(&pp->children[i].err);
1038                 if (code)
1039                         pp->shutdown = 1;
1040                 return code;
1041         }
1042
1043         pp->nr_processes++;
1044         pp->children[i].state = GIT_CP_WORKING;
1045         pp->pfd[i].fd = pp->children[i].process.err;
1046         return 0;
1047 }
1048
1049 static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1050 {
1051         int i;
1052
1053         while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1054                 if (errno == EINTR)
1055                         continue;
1056                 pp_cleanup(pp);
1057                 die_errno("poll");
1058         }
1059
1060         /* Buffer output from all pipes. */
1061         for (i = 0; i < pp->max_processes; i++) {
1062                 if (pp->children[i].state == GIT_CP_WORKING &&
1063                     pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1064                         int n = strbuf_read_once(&pp->children[i].err,
1065                                                  pp->children[i].process.err, 0);
1066                         if (n == 0) {
1067                                 close(pp->children[i].process.err);
1068                                 pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1069                         } else if (n < 0)
1070                                 if (errno != EAGAIN)
1071                                         die_errno("read");
1072                 }
1073         }
1074 }
1075
1076 static void pp_output(struct parallel_processes *pp)
1077 {
1078         int i = pp->output_owner;
1079         if (pp->children[i].state == GIT_CP_WORKING &&
1080             pp->children[i].err.len) {
1081                 fputs(pp->children[i].err.buf, stderr);
1082                 strbuf_reset(&pp->children[i].err);
1083         }
1084 }
1085
1086 static int pp_collect_finished(struct parallel_processes *pp)
1087 {
1088         int i, code;
1089         int n = pp->max_processes;
1090         int result = 0;
1091
1092         while (pp->nr_processes > 0) {
1093                 for (i = 0; i < pp->max_processes; i++)
1094                         if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1095                                 break;
1096                 if (i == pp->max_processes)
1097                         break;
1098
1099                 code = finish_command(&pp->children[i].process);
1100
1101                 code = pp->task_finished(code,
1102                                          &pp->children[i].err, pp->data,
1103                                          &pp->children[i].data);
1104
1105                 if (code)
1106                         result = code;
1107                 if (code < 0)
1108                         break;
1109
1110                 pp->nr_processes--;
1111                 pp->children[i].state = GIT_CP_FREE;
1112                 pp->pfd[i].fd = -1;
1113                 child_process_init(&pp->children[i].process);
1114
1115                 if (i != pp->output_owner) {
1116                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1117                         strbuf_reset(&pp->children[i].err);
1118                 } else {
1119                         fputs(pp->children[i].err.buf, stderr);
1120                         strbuf_reset(&pp->children[i].err);
1121
1122                         /* Output all other finished child processes */
1123                         fputs(pp->buffered_output.buf, stderr);
1124                         strbuf_reset(&pp->buffered_output);
1125
1126                         /*
1127                          * Pick next process to output live.
1128                          * NEEDSWORK:
1129                          * For now we pick it randomly by doing a round
1130                          * robin. Later we may want to pick the one with
1131                          * the most output or the longest or shortest
1132                          * running process time.
1133                          */
1134                         for (i = 0; i < n; i++)
1135                                 if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1136                                         break;
1137                         pp->output_owner = (pp->output_owner + i) % n;
1138                 }
1139         }
1140         return result;
1141 }
1142
1143 int run_processes_parallel(int n,
1144                            get_next_task_fn get_next_task,
1145                            start_failure_fn start_failure,
1146                            task_finished_fn task_finished,
1147                            void *pp_cb)
1148 {
1149         int i, code;
1150         int output_timeout = 100;
1151         int spawn_cap = 4;
1152         struct parallel_processes pp;
1153
1154         pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1155         while (1) {
1156                 for (i = 0;
1157                     i < spawn_cap && !pp.shutdown &&
1158                     pp.nr_processes < pp.max_processes;
1159                     i++) {
1160                         code = pp_start_one(&pp);
1161                         if (!code)
1162                                 continue;
1163                         if (code < 0) {
1164                                 pp.shutdown = 1;
1165                                 kill_children(&pp, -code);
1166                         }
1167                         break;
1168                 }
1169                 if (!pp.nr_processes)
1170                         break;
1171                 pp_buffer_stderr(&pp, output_timeout);
1172                 pp_output(&pp);
1173                 code = pp_collect_finished(&pp);
1174                 if (code) {
1175                         pp.shutdown = 1;
1176                         if (code < 0)
1177                                 kill_children(&pp, -code);
1178                 }
1179         }
1180
1181         pp_cleanup(&pp);
1182         return 0;
1183 }