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("waitpid for %s failed: %s", argv0, strerror(errno));
 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("cannot fork() for %s: %s", cmd->argv[0],
 424                         strerror(errno));
 425         else if (cmd->clean_on_exit)
 426                 mark_child_for_cleanup(cmd->pid);
 427
 428         /*
 429          * Wait for child's execvp. If the execvp succeeds (or if fork()
 430          * failed), EOF is seen immediately by the parent. Otherwise, the
 431          * child process sends a single byte.
 432          * Note that use of this infrastructure is completely advisory,
 433          * therefore, we keep error checks minimal.
 434          */
 435         close(notify_pipe[1]);
 436         if (read(notify_pipe[0], &notify_pipe[1], 1) == 1) {
 437                 /*
 438                  * At this point we know that fork() succeeded, but execvp()
 439                  * failed. Errors have been reported to our stderr.
 440                  */
 441                 wait_or_whine(cmd->pid, cmd->argv[0], 0);
 442                 failed_errno = errno;
 443                 cmd->pid = -1;
 444         }
 445         close(notify_pipe[0]);
 446 }
 447 #else
 448 {
 449         int fhin = 0, fhout = 1, fherr = 2;
 450         const char **sargv = cmd->argv;
 451         struct argv_array nargv = ARGV_ARRAY_INIT;
 452
 453         if (cmd->no_stdin)
 454                 fhin = open("/dev/null", O_RDWR);
 455         else if (need_in)
 456                 fhin = dup(fdin[0]);
 457         else if (cmd->in)
 458                 fhin = dup(cmd->in);
 459
 460         if (cmd->no_stderr)
 461                 fherr = open("/dev/null", O_RDWR);
 462         else if (need_err)
 463                 fherr = dup(fderr[1]);
 464         else if (cmd->err > 2)
 465                 fherr = dup(cmd->err);
 466
 467         if (cmd->no_stdout)
 468                 fhout = open("/dev/null", O_RDWR);
 469         else if (cmd->stdout_to_stderr)
 470                 fhout = dup(fherr);
 471         else if (need_out)
 472                 fhout = dup(fdout[1]);
 473         else if (cmd->out > 1)
 474                 fhout = dup(cmd->out);
 475
 476         if (cmd->git_cmd)
 477                 cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
 478         else if (cmd->use_shell)
 479                 cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
 480
 481         cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
 482                         cmd->dir, fhin, fhout, fherr);
 483         failed_errno = errno;
 484         if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
 485                 error("cannot spawn %s: %s", cmd->argv[0], strerror(errno));
 486         if (cmd->clean_on_exit && cmd->pid >= 0)
 487                 mark_child_for_cleanup(cmd->pid);
 488
 489         argv_array_clear(&nargv);
 490         cmd->argv = sargv;
 491         if (fhin != 0)
 492                 close(fhin);
 493         if (fhout != 1)
 494                 close(fhout);
 495         if (fherr != 2)
 496                 close(fherr);
 497 }
 498 #endif
 499
 500         if (cmd->pid < 0) {
 501                 if (need_in)
 502                         close_pair(fdin);
 503                 else if (cmd->in)
 504                         close(cmd->in);
 505                 if (need_out)
 506                         close_pair(fdout);
 507                 else if (cmd->out)
 508                         close(cmd->out);
 509                 if (need_err)
 510                         close_pair(fderr);
 511                 else if (cmd->err)
 512                         close(cmd->err);
 513                 child_process_clear(cmd);
 514                 errno = failed_errno;
 515                 return -1;
 516         }
 517
 518         if (need_in)
 519                 close(fdin[0]);
 520         else if (cmd->in)
 521                 close(cmd->in);
 522
 523         if (need_out)
 524                 close(fdout[1]);
 525         else if (cmd->out)
 526                 close(cmd->out);
 527
 528         if (need_err)
 529                 close(fderr[1]);
 530         else if (cmd->err)
 531                 close(cmd->err);
 532
 533         return 0;
 534 }
 535
 536 int finish_command(struct child_process *cmd)
 537 {
 538         int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
 539         child_process_clear(cmd);
 540         return ret;
 541 }
 542
 543 int finish_command_in_signal(struct child_process *cmd)
 544 {
 545         return wait_or_whine(cmd->pid, cmd->argv[0], 1);
 546 }
 547
 548
 549 int run_command(struct child_process *cmd)
 550 {
 551         int code;
 552
 553         if (cmd->out < 0 || cmd->err < 0)
 554                 die("BUG: run_command with a pipe can cause deadlock");
 555
 556         code = start_command(cmd);
 557         if (code)
 558                 return code;
 559         return finish_command(cmd);
 560 }
 561
 562 int run_command_v_opt(const char **argv, int opt)
 563 {
 564         return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
 565 }
 566
 567 int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
 568 {
 569         struct child_process cmd = CHILD_PROCESS_INIT;
 570         cmd.argv = argv;
 571         cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
 572         cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
 573         cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
 574         cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
 575         cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
 576         cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
 577         cmd.dir = dir;
 578         cmd.env = env;
 579         return run_command(&cmd);
 580 }
 581
 582 #ifndef NO_PTHREADS
 583 static pthread_t main_thread;
 584 static int main_thread_set;
 585 static pthread_key_t async_key;
 586 static pthread_key_t async_die_counter;
 587
 588 static void *run_thread(void *data)
 589 {
 590         struct async *async = data;
 591         intptr_t ret;
 592
 593         if (async->isolate_sigpipe) {
 594                 sigset_t mask;
 595                 sigemptyset(&mask);
 596                 sigaddset(&mask, SIGPIPE);
 597                 if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
 598                         ret = error("unable to block SIGPIPE in async thread");
 599                         return (void *)ret;
 600                 }
 601         }
 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
 608 static 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
 624 static 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
 631 int 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 void NORETURN async_exit(int code)
 639 {
 640         pthread_exit((void *)(intptr_t)code);
 641 }
 642
 643 #else
 644
 645 static struct {
 646         void (**handlers)(void);
 647         size_t nr;
 648         size_t alloc;
 649 } git_atexit_hdlrs;
 650
 651 static int git_atexit_installed;
 652
 653 static void git_atexit_dispatch(void)
 654 {
 655         size_t i;
 656
 657         for (i=git_atexit_hdlrs.nr ; i ; i--)
 658                 git_atexit_hdlrs.handlers[i-1]();
 659 }
 660
 661 static void git_atexit_clear(void)
 662 {
 663         free(git_atexit_hdlrs.handlers);
 664         memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
 665         git_atexit_installed = 0;
 666 }
 667
 668 #undef atexit
 669 int git_atexit(void (*handler)(void))
 670 {
 671         ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
 672         git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
 673         if (!git_atexit_installed) {
 674                 if (atexit(&git_atexit_dispatch))
 675                         return -1;
 676                 git_atexit_installed = 1;
 677         }
 678         return 0;
 679 }
 680 #define atexit git_atexit
 681
 682 static int process_is_async;
 683 int in_async(void)
 684 {
 685         return process_is_async;
 686 }
 687
 688 void NORETURN async_exit(int code)
 689 {
 690         exit(code);
 691 }
 692
 693 #endif
 694
 695 int start_async(struct async *async)
 696 {
 697         int need_in, need_out;
 698         int fdin[2], fdout[2];
 699         int proc_in, proc_out;
 700
 701         need_in = async->in < 0;
 702         if (need_in) {
 703                 if (pipe(fdin) < 0) {
 704                         if (async->out > 0)
 705                                 close(async->out);
 706                         return error("cannot create pipe: %s", strerror(errno));
 707                 }
 708                 async->in = fdin[1];
 709         }
 710
 711         need_out = async->out < 0;
 712         if (need_out) {
 713                 if (pipe(fdout) < 0) {
 714                         if (need_in)
 715                                 close_pair(fdin);
 716                         else if (async->in)
 717                                 close(async->in);
 718                         return error("cannot create pipe: %s", strerror(errno));
 719                 }
 720                 async->out = fdout[0];
 721         }
 722
 723         if (need_in)
 724                 proc_in = fdin[0];
 725         else if (async->in)
 726                 proc_in = async->in;
 727         else
 728                 proc_in = -1;
 729
 730         if (need_out)
 731                 proc_out = fdout[1];
 732         else if (async->out)
 733                 proc_out = async->out;
 734         else
 735                 proc_out = -1;
 736
 737 #ifdef NO_PTHREADS
 738         /* Flush stdio before fork() to avoid cloning buffers */
 739         fflush(NULL);
 740
 741         async->pid = fork();
 742         if (async->pid < 0) {
 743                 error("fork (async) failed: %s", strerror(errno));
 744                 goto error;
 745         }
 746         if (!async->pid) {
 747                 if (need_in)
 748                         close(fdin[1]);
 749                 if (need_out)
 750                         close(fdout[0]);
 751                 git_atexit_clear();
 752                 process_is_async = 1;
 753                 exit(!!async->proc(proc_in, proc_out, async->data));
 754         }
 755
 756         mark_child_for_cleanup(async->pid);
 757
 758         if (need_in)
 759                 close(fdin[0]);
 760         else if (async->in)
 761                 close(async->in);
 762
 763         if (need_out)
 764                 close(fdout[1]);
 765         else if (async->out)
 766                 close(async->out);
 767 #else
 768         if (!main_thread_set) {
 769                 /*
 770                  * We assume that the first time that start_async is called
 771                  * it is from the main thread.
 772                  */
 773                 main_thread_set = 1;
 774                 main_thread = pthread_self();
 775                 pthread_key_create(&async_key, NULL);
 776                 pthread_key_create(&async_die_counter, NULL);
 777                 set_die_routine(die_async);
 778                 set_die_is_recursing_routine(async_die_is_recursing);
 779         }
 780
 781         if (proc_in >= 0)
 782                 set_cloexec(proc_in);
 783         if (proc_out >= 0)
 784                 set_cloexec(proc_out);
 785         async->proc_in = proc_in;
 786         async->proc_out = proc_out;
 787         {
 788                 int err = pthread_create(&async->tid, NULL, run_thread, async);
 789                 if (err) {
 790                         error("cannot create thread: %s", strerror(err));
 791                         goto error;
 792                 }
 793         }
 794 #endif
 795         return 0;
 796
 797 error:
 798         if (need_in)
 799                 close_pair(fdin);
 800         else if (async->in)
 801                 close(async->in);
 802
 803         if (need_out)
 804                 close_pair(fdout);
 805         else if (async->out)
 806                 close(async->out);
 807         return -1;
 808 }
 809
 810 int finish_async(struct async *async)
 811 {
 812 #ifdef NO_PTHREADS
 813         return wait_or_whine(async->pid, "child process", 0);
 814 #else
 815         void *ret = (void *)(intptr_t)(-1);
 816
 817         if (pthread_join(async->tid, &ret))
 818                 error("pthread_join failed");
 819         return (int)(intptr_t)ret;
 820 #endif
 821 }
 822
 823 const char *find_hook(const char *name)
 824 {
 825         static struct strbuf path = STRBUF_INIT;
 826
 827         strbuf_reset(&path);
 828         strbuf_git_path(&path, "hooks/%s", name);
 829         if (access(path.buf, X_OK) < 0)
 830                 return NULL;
 831         return path.buf;
 832 }
 833
 834 int run_hook_ve(const char *const *env, const char *name, va_list args)
 835 {
 836         struct child_process hook = CHILD_PROCESS_INIT;
 837         const char *p;
 838
 839         p = find_hook(name);
 840         if (!p)
 841                 return 0;
 842
 843         argv_array_push(&hook.args, p);
 844         while ((p = va_arg(args, const char *)))
 845                 argv_array_push(&hook.args, p);
 846         hook.env = env;
 847         hook.no_stdin = 1;
 848         hook.stdout_to_stderr = 1;
 849
 850         return run_command(&hook);
 851 }
 852
 853 int run_hook_le(const char *const *env, const char *name, ...)
 854 {
 855         va_list args;
 856         int ret;
 857
 858         va_start(args, name);
 859         ret = run_hook_ve(env, name, args);
 860         va_end(args);
 861
 862         return ret;
 863 }
 864
 865 int capture_command(struct child_process *cmd, struct strbuf *buf, size_t hint)
 866 {
 867         cmd->out = -1;
 868         if (start_command(cmd) < 0)
 869                 return -1;
 870
 871         if (strbuf_read(buf, cmd->out, hint) < 0) {
 872                 close(cmd->out);
 873                 finish_command(cmd); /* throw away exit code */
 874                 return -1;
 875         }
 876
 877         close(cmd->out);
 878         return finish_command(cmd);
 879 }
 880
 881 enum child_state {
 882         GIT_CP_FREE,
 883         GIT_CP_WORKING,
 884         GIT_CP_WAIT_CLEANUP,
 885 };
 886
 887 struct parallel_processes {
 888         void *data;
 889
 890         int max_processes;
 891         int nr_processes;
 892
 893         get_next_task_fn get_next_task;
 894         start_failure_fn start_failure;
 895         task_finished_fn task_finished;
 896
 897         struct {
 898                 enum child_state state;
 899                 struct child_process process;
 900                 struct strbuf err;
 901                 void *data;
 902         } *children;
 903         /*
 904          * The struct pollfd is logically part of *children,
 905          * but the system call expects it as its own array.
 906          */
 907         struct pollfd *pfd;
 908
 909         unsigned shutdown : 1;
 910
 911         int output_owner;
 912         struct strbuf buffered_output; /* of finished children */
 913 };
 914
 915 static int default_start_failure(struct strbuf *out,
 916                                  void *pp_cb,
 917                                  void *pp_task_cb)
 918 {
 919         return 0;
 920 }
 921
 922 static int default_task_finished(int result,
 923                                  struct strbuf *out,
 924                                  void *pp_cb,
 925                                  void *pp_task_cb)
 926 {
 927         return 0;
 928 }
 929
 930 static void kill_children(struct parallel_processes *pp, int signo)
 931 {
 932         int i, n = pp->max_processes;
 933
 934         for (i = 0; i < n; i++)
 935                 if (pp->children[i].state == GIT_CP_WORKING)
 936                         kill(pp->children[i].process.pid, signo);
 937 }
 938
 939 static struct parallel_processes *pp_for_signal;
 940
 941 static void handle_children_on_signal(int signo)
 942 {
 943         kill_children(pp_for_signal, signo);
 944         sigchain_pop(signo);
 945         raise(signo);
 946 }
 947
 948 static void pp_init(struct parallel_processes *pp,
 949                     int n,
 950                     get_next_task_fn get_next_task,
 951                     start_failure_fn start_failure,
 952                     task_finished_fn task_finished,
 953                     void *data)
 954 {
 955         int i;
 956
 957         if (n < 1)
 958                 n = online_cpus();
 959
 960         pp->max_processes = n;
 961
 962         trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
 963
 964         pp->data = data;
 965         if (!get_next_task)
 966                 die("BUG: you need to specify a get_next_task function");
 967         pp->get_next_task = get_next_task;
 968
 969         pp->start_failure = start_failure ? start_failure : default_start_failure;
 970         pp->task_finished = task_finished ? task_finished : default_task_finished;
 971
 972         pp->nr_processes = 0;
 973         pp->output_owner = 0;
 974         pp->shutdown = 0;
 975         pp->children = xcalloc(n, sizeof(*pp->children));
 976         pp->pfd = xcalloc(n, sizeof(*pp->pfd));
 977         strbuf_init(&pp->buffered_output, 0);
 978
 979         for (i = 0; i < n; i++) {
 980                 strbuf_init(&pp->children[i].err, 0);
 981                 child_process_init(&pp->children[i].process);
 982                 pp->pfd[i].events = POLLIN | POLLHUP;
 983                 pp->pfd[i].fd = -1;
 984         }
 985
 986         pp_for_signal = pp;
 987         sigchain_push_common(handle_children_on_signal);
 988 }
 989
 990 static void pp_cleanup(struct parallel_processes *pp)
 991 {
 992         int i;
 993
 994         trace_printf("run_processes_parallel: done");
 995         for (i = 0; i < pp->max_processes; i++) {
 996                 strbuf_release(&pp->children[i].err);
 997                 child_process_clear(&pp->children[i].process);
 998         }
 999
1000         free(pp->children);
1001         free(pp->pfd);
1002
1003         /*
1004          * When get_next_task added messages to the buffer in its last
1005          * iteration, the buffered output is non empty.
1006          */
1007         strbuf_write(&pp->buffered_output, stderr);
1008         strbuf_release(&pp->buffered_output);
1009
1010         sigchain_pop_common();
1011 }
1012
1013 /* returns
1014  *  0 if a new task was started.
1015  *  1 if no new jobs was started (get_next_task ran out of work, non critical
1016  *    problem with starting a new command)
1017  * <0 no new job was started, user wishes to shutdown early. Use negative code
1018  *    to signal the children.
1019  */
1020 static int pp_start_one(struct parallel_processes *pp)
1021 {
1022         int i, code;
1023
1024         for (i = 0; i < pp->max_processes; i++)
1025                 if (pp->children[i].state == GIT_CP_FREE)
1026                         break;
1027         if (i == pp->max_processes)
1028                 die("BUG: bookkeeping is hard");
1029
1030         code = pp->get_next_task(&pp->children[i].process,
1031                                  &pp->children[i].err,
1032                                  pp->data,
1033                                  &pp->children[i].data);
1034         if (!code) {
1035                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1036                 strbuf_reset(&pp->children[i].err);
1037                 return 1;
1038         }
1039         pp->children[i].process.err = -1;
1040         pp->children[i].process.stdout_to_stderr = 1;
1041         pp->children[i].process.no_stdin = 1;
1042
1043         if (start_command(&pp->children[i].process)) {
1044                 code = pp->start_failure(&pp->children[i].err,
1045                                          pp->data,
1046                                          &pp->children[i].data);
1047                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1048                 strbuf_reset(&pp->children[i].err);
1049                 if (code)
1050                         pp->shutdown = 1;
1051                 return code;
1052         }
1053
1054         pp->nr_processes++;
1055         pp->children[i].state = GIT_CP_WORKING;
1056         pp->pfd[i].fd = pp->children[i].process.err;
1057         return 0;
1058 }
1059
1060 static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1061 {
1062         int i;
1063
1064         while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1065                 if (errno == EINTR)
1066                         continue;
1067                 pp_cleanup(pp);
1068                 die_errno("poll");
1069         }
1070
1071         /* Buffer output from all pipes. */
1072         for (i = 0; i < pp->max_processes; i++) {
1073                 if (pp->children[i].state == GIT_CP_WORKING &&
1074                     pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1075                         int n = strbuf_read_once(&pp->children[i].err,
1076                                                  pp->children[i].process.err, 0);
1077                         if (n == 0) {
1078                                 close(pp->children[i].process.err);
1079                                 pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1080                         } else if (n < 0)
1081                                 if (errno != EAGAIN)
1082                                         die_errno("read");
1083                 }
1084         }
1085 }
1086
1087 static void pp_output(struct parallel_processes *pp)
1088 {
1089         int i = pp->output_owner;
1090         if (pp->children[i].state == GIT_CP_WORKING &&
1091             pp->children[i].err.len) {
1092                 strbuf_write(&pp->children[i].err, stderr);
1093                 strbuf_reset(&pp->children[i].err);
1094         }
1095 }
1096
1097 static int pp_collect_finished(struct parallel_processes *pp)
1098 {
1099         int i, code;
1100         int n = pp->max_processes;
1101         int result = 0;
1102
1103         while (pp->nr_processes > 0) {
1104                 for (i = 0; i < pp->max_processes; i++)
1105                         if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1106                                 break;
1107                 if (i == pp->max_processes)
1108                         break;
1109
1110                 code = finish_command(&pp->children[i].process);
1111
1112                 code = pp->task_finished(code,
1113                                          &pp->children[i].err, pp->data,
1114                                          &pp->children[i].data);
1115
1116                 if (code)
1117                         result = code;
1118                 if (code < 0)
1119                         break;
1120
1121                 pp->nr_processes--;
1122                 pp->children[i].state = GIT_CP_FREE;
1123                 pp->pfd[i].fd = -1;
1124                 child_process_init(&pp->children[i].process);
1125
1126                 if (i != pp->output_owner) {
1127                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1128                         strbuf_reset(&pp->children[i].err);
1129                 } else {
1130                         strbuf_write(&pp->children[i].err, stderr);
1131                         strbuf_reset(&pp->children[i].err);
1132
1133                         /* Output all other finished child processes */
1134                         strbuf_write(&pp->buffered_output, stderr);
1135                         strbuf_reset(&pp->buffered_output);
1136
1137                         /*
1138                          * Pick next process to output live.
1139                          * NEEDSWORK:
1140                          * For now we pick it randomly by doing a round
1141                          * robin. Later we may want to pick the one with
1142                          * the most output or the longest or shortest
1143                          * running process time.
1144                          */
1145                         for (i = 0; i < n; i++)
1146                                 if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1147                                         break;
1148                         pp->output_owner = (pp->output_owner + i) % n;
1149                 }
1150         }
1151         return result;
1152 }
1153
1154 int run_processes_parallel(int n,
1155                            get_next_task_fn get_next_task,
1156                            start_failure_fn start_failure,
1157                            task_finished_fn task_finished,
1158                            void *pp_cb)
1159 {
1160         int i, code;
1161         int output_timeout = 100;
1162         int spawn_cap = 4;
1163         struct parallel_processes pp;
1164
1165         pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1166         while (1) {
1167                 for (i = 0;
1168                     i < spawn_cap && !pp.shutdown &&
1169                     pp.nr_processes < pp.max_processes;
1170                     i++) {
1171                         code = pp_start_one(&pp);
1172                         if (!code)
1173                                 continue;
1174                         if (code < 0) {
1175                                 pp.shutdown = 1;
1176                                 kill_children(&pp, -code);
1177                         }
1178                         break;
1179                 }
1180                 if (!pp.nr_processes)
1181                         break;
1182                 pp_buffer_stderr(&pp, output_timeout);
1183                 pp_output(&pp);
1184                 code = pp_collect_finished(&pp);
1185                 if (code) {
1186                         pp.shutdown = 1;
1187                         if (code < 0)
1188                                 kill_children(&pp, -code);
1189                 }
1190         }
1191
1192         pp_cleanup(&pp);
1193         return 0;
1194 }