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