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 #include "string-list.h"
   9 #include "quote.h"
  10
  11 void child_process_init(struct child_process *child)
  12 {
  13         memset(child, 0, sizeof(*child));
  14         argv_array_init(&child->args);
  15         argv_array_init(&child->env_array);
  16 }
  17
  18 void child_process_clear(struct child_process *child)
  19 {
  20         argv_array_clear(&child->args);
  21         argv_array_clear(&child->env_array);
  22 }
  23
  24 struct child_to_clean {
  25         pid_t pid;
  26         struct child_process *process;
  27         struct child_to_clean *next;
  28 };
  29 static struct child_to_clean *children_to_clean;
  30 static int installed_child_cleanup_handler;
  31
  32 static void cleanup_children(int sig, int in_signal)
  33 {
  34         struct child_to_clean *children_to_wait_for = NULL;
  35
  36         while (children_to_clean) {
  37                 struct child_to_clean *p = children_to_clean;
  38                 children_to_clean = p->next;
  39
  40                 if (p->process && !in_signal) {
  41                         struct child_process *process = p->process;
  42                         if (process->clean_on_exit_handler) {
  43                                 trace_printf(
  44                                         "trace: run_command: running exit handler for pid %"
  45                                         PRIuMAX, (uintmax_t)p->pid
  46                                 );
  47                                 process->clean_on_exit_handler(process);
  48                         }
  49                 }
  50
  51                 kill(p->pid, sig);
  52
  53                 if (p->process && p->process->wait_after_clean) {
  54                         p->next = children_to_wait_for;
  55                         children_to_wait_for = p;
  56                 } else {
  57                         if (!in_signal)
  58                                 free(p);
  59                 }
  60         }
  61
  62         while (children_to_wait_for) {
  63                 struct child_to_clean *p = children_to_wait_for;
  64                 children_to_wait_for = p->next;
  65
  66                 while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
  67                         ; /* spin waiting for process exit or error */
  68
  69                 if (!in_signal)
  70                         free(p);
  71         }
  72 }
  73
  74 static void cleanup_children_on_signal(int sig)
  75 {
  76         cleanup_children(sig, 1);
  77         sigchain_pop(sig);
  78         raise(sig);
  79 }
  80
  81 static void cleanup_children_on_exit(void)
  82 {
  83         cleanup_children(SIGTERM, 0);
  84 }
  85
  86 static void mark_child_for_cleanup(pid_t pid, struct child_process *process)
  87 {
  88         struct child_to_clean *p = xmalloc(sizeof(*p));
  89         p->pid = pid;
  90         p->process = process;
  91         p->next = children_to_clean;
  92         children_to_clean = p;
  93
  94         if (!installed_child_cleanup_handler) {
  95                 atexit(cleanup_children_on_exit);
  96                 sigchain_push_common(cleanup_children_on_signal);
  97                 installed_child_cleanup_handler = 1;
  98         }
  99 }
 100
 101 static void clear_child_for_cleanup(pid_t pid)
 102 {
 103         struct child_to_clean **pp;
 104
 105         for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
 106                 struct child_to_clean *clean_me = *pp;
 107
 108                 if (clean_me->pid == pid) {
 109                         *pp = clean_me->next;
 110                         free(clean_me);
 111                         return;
 112                 }
 113         }
 114 }
 115
 116 static inline void close_pair(int fd[2])
 117 {
 118         close(fd[0]);
 119         close(fd[1]);
 120 }
 121
 122 int is_executable(const char *name)
 123 {
 124         struct stat st;
 125
 126         if (stat(name, &st) || /* stat, not lstat */
 127             !S_ISREG(st.st_mode))
 128                 return 0;
 129
 130 #if defined(GIT_WINDOWS_NATIVE)
 131         /*
 132          * On Windows there is no executable bit. The file extension
 133          * indicates whether it can be run as an executable, and Git
 134          * has special-handling to detect scripts and launch them
 135          * through the indicated script interpreter. We test for the
 136          * file extension first because virus scanners may make
 137          * it quite expensive to open many files.
 138          */
 139         if (ends_with(name, ".exe"))
 140                 return S_IXUSR;
 141
 142 {
 143         /*
 144          * Now that we know it does not have an executable extension,
 145          * peek into the file instead.
 146          */
 147         char buf[3] = { 0 };
 148         int n;
 149         int fd = open(name, O_RDONLY);
 150         st.st_mode &= ~S_IXUSR;
 151         if (fd >= 0) {
 152                 n = read(fd, buf, 2);
 153                 if (n == 2)
 154                         /* look for a she-bang */
 155                         if (!strcmp(buf, "#!"))
 156                                 st.st_mode |= S_IXUSR;
 157                 close(fd);
 158         }
 159 }
 160 #endif
 161         return st.st_mode & S_IXUSR;
 162 }
 163
 164 /*
 165  * Search $PATH for a command.  This emulates the path search that
 166  * execvp would perform, without actually executing the command so it
 167  * can be used before fork() to prepare to run a command using
 168  * execve() or after execvp() to diagnose why it failed.
 169  *
 170  * The caller should ensure that file contains no directory
 171  * separators.
 172  *
 173  * Returns the path to the command, as found in $PATH or NULL if the
 174  * command could not be found.  The caller inherits ownership of the memory
 175  * used to store the resultant path.
 176  *
 177  * This should not be used on Windows, where the $PATH search rules
 178  * are more complicated (e.g., a search for "foo" should find
 179  * "foo.exe").
 180  */
 181 static char *locate_in_PATH(const char *file)
 182 {
 183         const char *p = getenv("PATH");
 184         struct strbuf buf = STRBUF_INIT;
 185
 186         if (!p || !*p)
 187                 return NULL;
 188
 189         while (1) {
 190                 const char *end = strchrnul(p, ':');
 191
 192                 strbuf_reset(&buf);
 193
 194                 /* POSIX specifies an empty entry as the current directory. */
 195                 if (end != p) {
 196                         strbuf_add(&buf, p, end - p);
 197                         strbuf_addch(&buf, '/');
 198                 }
 199                 strbuf_addstr(&buf, file);
 200
 201                 if (is_executable(buf.buf))
 202                         return strbuf_detach(&buf, NULL);
 203
 204                 if (!*end)
 205                         break;
 206                 p = end + 1;
 207         }
 208
 209         strbuf_release(&buf);
 210         return NULL;
 211 }
 212
 213 static int exists_in_PATH(const char *file)
 214 {
 215         char *r = locate_in_PATH(file);
 216         free(r);
 217         return r != NULL;
 218 }
 219
 220 int sane_execvp(const char *file, char * const argv[])
 221 {
 222         if (!execvp(file, argv))
 223                 return 0; /* cannot happen ;-) */
 224
 225         /*
 226          * When a command can't be found because one of the directories
 227          * listed in $PATH is unsearchable, execvp reports EACCES, but
 228          * careful usability testing (read: analysis of occasional bug
 229          * reports) reveals that "No such file or directory" is more
 230          * intuitive.
 231          *
 232          * We avoid commands with "/", because execvp will not do $PATH
 233          * lookups in that case.
 234          *
 235          * The reassignment of EACCES to errno looks like a no-op below,
 236          * but we need to protect against exists_in_PATH overwriting errno.
 237          */
 238         if (errno == EACCES && !strchr(file, '/'))
 239                 errno = exists_in_PATH(file) ? EACCES : ENOENT;
 240         else if (errno == ENOTDIR && !strchr(file, '/'))
 241                 errno = ENOENT;
 242         return -1;
 243 }
 244
 245 static const char **prepare_shell_cmd(struct argv_array *out, const char **argv)
 246 {
 247         if (!argv[0])
 248                 die("BUG: shell command is empty");
 249
 250         if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
 251 #ifndef GIT_WINDOWS_NATIVE
 252                 argv_array_push(out, SHELL_PATH);
 253 #else
 254                 argv_array_push(out, "sh");
 255 #endif
 256                 argv_array_push(out, "-c");
 257
 258                 /*
 259                  * If we have no extra arguments, we do not even need to
 260                  * bother with the "$@" magic.
 261                  */
 262                 if (!argv[1])
 263                         argv_array_push(out, argv[0]);
 264                 else
 265                         argv_array_pushf(out, "%s \"$@\"", argv[0]);
 266         }
 267
 268         argv_array_pushv(out, argv);
 269         return out->argv;
 270 }
 271
 272 #ifndef GIT_WINDOWS_NATIVE
 273 static int child_notifier = -1;
 274
 275 enum child_errcode {
 276         CHILD_ERR_CHDIR,
 277         CHILD_ERR_DUP2,
 278         CHILD_ERR_CLOSE,
 279         CHILD_ERR_SIGPROCMASK,
 280         CHILD_ERR_ENOENT,
 281         CHILD_ERR_SILENT,
 282         CHILD_ERR_ERRNO
 283 };
 284
 285 struct child_err {
 286         enum child_errcode err;
 287         int syserr; /* errno */
 288 };
 289
 290 static void child_die(enum child_errcode err)
 291 {
 292         struct child_err buf;
 293
 294         buf.err = err;
 295         buf.syserr = errno;
 296
 297         /* write(2) on buf smaller than PIPE_BUF (min 512) is atomic: */
 298         xwrite(child_notifier, &buf, sizeof(buf));
 299         _exit(1);
 300 }
 301
 302 static void child_dup2(int fd, int to)
 303 {
 304         if (dup2(fd, to) < 0)
 305                 child_die(CHILD_ERR_DUP2);
 306 }
 307
 308 static void child_close(int fd)
 309 {
 310         if (close(fd))
 311                 child_die(CHILD_ERR_CLOSE);
 312 }
 313
 314 static void child_close_pair(int fd[2])
 315 {
 316         child_close(fd[0]);
 317         child_close(fd[1]);
 318 }
 319
 320 /*
 321  * parent will make it look like the child spewed a fatal error and died
 322  * this is needed to prevent changes to t0061.
 323  */
 324 static void fake_fatal(const char *err, va_list params)
 325 {
 326         vreportf("fatal: ", err, params);
 327 }
 328
 329 static void child_error_fn(const char *err, va_list params)
 330 {
 331         const char msg[] = "error() should not be called in child\n";
 332         xwrite(2, msg, sizeof(msg) - 1);
 333 }
 334
 335 static void child_warn_fn(const char *err, va_list params)
 336 {
 337         const char msg[] = "warn() should not be called in child\n";
 338         xwrite(2, msg, sizeof(msg) - 1);
 339 }
 340
 341 static void NORETURN child_die_fn(const char *err, va_list params)
 342 {
 343         const char msg[] = "die() should not be called in child\n";
 344         xwrite(2, msg, sizeof(msg) - 1);
 345         _exit(2);
 346 }
 347
 348 /* this runs in the parent process */
 349 static void child_err_spew(struct child_process *cmd, struct child_err *cerr)
 350 {
 351         static void (*old_errfn)(const char *err, va_list params);
 352
 353         old_errfn = get_error_routine();
 354         set_error_routine(fake_fatal);
 355         errno = cerr->syserr;
 356
 357         switch (cerr->err) {
 358         case CHILD_ERR_CHDIR:
 359                 error_errno("exec '%s': cd to '%s' failed",
 360                             cmd->argv[0], cmd->dir);
 361                 break;
 362         case CHILD_ERR_DUP2:
 363                 error_errno("dup2() in child failed");
 364                 break;
 365         case CHILD_ERR_CLOSE:
 366                 error_errno("close() in child failed");
 367                 break;
 368         case CHILD_ERR_SIGPROCMASK:
 369                 error_errno("sigprocmask failed restoring signals");
 370                 break;
 371         case CHILD_ERR_ENOENT:
 372                 error_errno("cannot run %s", cmd->argv[0]);
 373                 break;
 374         case CHILD_ERR_SILENT:
 375                 break;
 376         case CHILD_ERR_ERRNO:
 377                 error_errno("cannot exec '%s'", cmd->argv[0]);
 378                 break;
 379         }
 380         set_error_routine(old_errfn);
 381 }
 382
 383 static void prepare_cmd(struct argv_array *out, const struct child_process *cmd)
 384 {
 385         if (!cmd->argv[0])
 386                 die("BUG: command is empty");
 387
 388         /*
 389          * Add SHELL_PATH so in the event exec fails with ENOEXEC we can
 390          * attempt to interpret the command with 'sh'.
 391          */
 392         argv_array_push(out, SHELL_PATH);
 393
 394         if (cmd->git_cmd) {
 395                 argv_array_push(out, "git");
 396                 argv_array_pushv(out, cmd->argv);
 397         } else if (cmd->use_shell) {
 398                 prepare_shell_cmd(out, cmd->argv);
 399         } else {
 400                 argv_array_pushv(out, cmd->argv);
 401         }
 402
 403         /*
 404          * If there are no '/' characters in the command then perform a path
 405          * lookup and use the resolved path as the command to exec.  If there
 406          * are no '/' characters or if the command wasn't found in the path,
 407          * have exec attempt to invoke the command directly.
 408          */
 409         if (!strchr(out->argv[1], '/')) {
 410                 char *program = locate_in_PATH(out->argv[1]);
 411                 if (program) {
 412                         free((char *)out->argv[1]);
 413                         out->argv[1] = program;
 414                 }
 415         }
 416 }
 417
 418 static char **prep_childenv(const char *const *deltaenv)
 419 {
 420         extern char **environ;
 421         char **childenv;
 422         struct string_list env = STRING_LIST_INIT_DUP;
 423         struct strbuf key = STRBUF_INIT;
 424         const char *const *p;
 425         int i;
 426
 427         /* Construct a sorted string list consisting of the current environ */
 428         for (p = (const char *const *) environ; p && *p; p++) {
 429                 const char *equals = strchr(*p, '=');
 430
 431                 if (equals) {
 432                         strbuf_reset(&key);
 433                         strbuf_add(&key, *p, equals - *p);
 434                         string_list_append(&env, key.buf)->util = (void *) *p;
 435                 } else {
 436                         string_list_append(&env, *p)->util = (void *) *p;
 437                 }
 438         }
 439         string_list_sort(&env);
 440
 441         /* Merge in 'deltaenv' with the current environ */
 442         for (p = deltaenv; p && *p; p++) {
 443                 const char *equals = strchr(*p, '=');
 444
 445                 if (equals) {
 446                         /* ('key=value'), insert or replace entry */
 447                         strbuf_reset(&key);
 448                         strbuf_add(&key, *p, equals - *p);
 449                         string_list_insert(&env, key.buf)->util = (void *) *p;
 450                 } else {
 451                         /* otherwise ('key') remove existing entry */
 452                         string_list_remove(&env, *p, 0);
 453                 }
 454         }
 455
 456         /* Create an array of 'char *' to be used as the childenv */
 457         ALLOC_ARRAY(childenv, env.nr + 1);
 458         for (i = 0; i < env.nr; i++)
 459                 childenv[i] = env.items[i].util;
 460         childenv[env.nr] = NULL;
 461
 462         string_list_clear(&env, 0);
 463         strbuf_release(&key);
 464         return childenv;
 465 }
 466
 467 struct atfork_state {
 468 #ifndef NO_PTHREADS
 469         int cs;
 470 #endif
 471         sigset_t old;
 472 };
 473
 474 #ifndef NO_PTHREADS
 475 static void bug_die(int err, const char *msg)
 476 {
 477         if (err) {
 478                 errno = err;
 479                 die_errno("BUG: %s", msg);
 480         }
 481 }
 482 #endif
 483
 484 static void atfork_prepare(struct atfork_state *as)
 485 {
 486         sigset_t all;
 487
 488         if (sigfillset(&all))
 489                 die_errno("sigfillset");
 490 #ifdef NO_PTHREADS
 491         if (sigprocmask(SIG_SETMASK, &all, &as->old))
 492                 die_errno("sigprocmask");
 493 #else
 494         bug_die(pthread_sigmask(SIG_SETMASK, &all, &as->old),
 495                 "blocking all signals");
 496         bug_die(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &as->cs),
 497                 "disabling cancellation");
 498 #endif
 499 }
 500
 501 static void atfork_parent(struct atfork_state *as)
 502 {
 503 #ifdef NO_PTHREADS
 504         if (sigprocmask(SIG_SETMASK, &as->old, NULL))
 505                 die_errno("sigprocmask");
 506 #else
 507         bug_die(pthread_setcancelstate(as->cs, NULL),
 508                 "re-enabling cancellation");
 509         bug_die(pthread_sigmask(SIG_SETMASK, &as->old, NULL),
 510                 "restoring signal mask");
 511 #endif
 512 }
 513 #endif /* GIT_WINDOWS_NATIVE */
 514
 515 static inline void set_cloexec(int fd)
 516 {
 517         int flags = fcntl(fd, F_GETFD);
 518         if (flags >= 0)
 519                 fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
 520 }
 521
 522 static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
 523 {
 524         int status, code = -1;
 525         pid_t waiting;
 526         int failed_errno = 0;
 527
 528         while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
 529                 ;       /* nothing */
 530         if (in_signal)
 531                 return 0;
 532
 533         if (waiting < 0) {
 534                 failed_errno = errno;
 535                 error_errno("waitpid for %s failed", argv0);
 536         } else if (waiting != pid) {
 537                 error("waitpid is confused (%s)", argv0);
 538         } else if (WIFSIGNALED(status)) {
 539                 code = WTERMSIG(status);
 540                 if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
 541                         error("%s died of signal %d", argv0, code);
 542                 /*
 543                  * This return value is chosen so that code & 0xff
 544                  * mimics the exit code that a POSIX shell would report for
 545                  * a program that died from this signal.
 546                  */
 547                 code += 128;
 548         } else if (WIFEXITED(status)) {
 549                 code = WEXITSTATUS(status);
 550         } else {
 551                 error("waitpid is confused (%s)", argv0);
 552         }
 553
 554         clear_child_for_cleanup(pid);
 555
 556         errno = failed_errno;
 557         return code;
 558 }
 559
 560 static void trace_add_env(struct strbuf *dst, const char *const *deltaenv)
 561 {
 562         struct string_list envs = STRING_LIST_INIT_DUP;
 563         const char *const *e;
 564         int i;
 565         int printed_unset = 0;
 566
 567         /* Last one wins, see run-command.c:prep_childenv() for context */
 568         for (e = deltaenv; e && *e; e++) {
 569                 struct strbuf key = STRBUF_INIT;
 570                 char *equals = strchr(*e, '=');
 571
 572                 if (equals) {
 573                         strbuf_add(&key, *e, equals - *e);
 574                         string_list_insert(&envs, key.buf)->util = equals + 1;
 575                 } else {
 576                         string_list_insert(&envs, *e)->util = NULL;
 577                 }
 578                 strbuf_release(&key);
 579         }
 580
 581         /* "unset X Y...;" */
 582         for (i = 0; i < envs.nr; i++) {
 583                 const char *var = envs.items[i].string;
 584                 const char *val = envs.items[i].util;
 585
 586                 if (val || !getenv(var))
 587                         continue;
 588
 589                 if (!printed_unset) {
 590                         strbuf_addstr(dst, " unset");
 591                         printed_unset = 1;
 592                 }
 593                 strbuf_addf(dst, " %s", var);
 594         }
 595         if (printed_unset)
 596                 strbuf_addch(dst, ';');
 597
 598         /* ... followed by "A=B C=D ..." */
 599         for (i = 0; i < envs.nr; i++) {
 600                 const char *var = envs.items[i].string;
 601                 const char *val = envs.items[i].util;
 602                 const char *oldval;
 603
 604                 if (!val)
 605                         continue;
 606
 607                 oldval = getenv(var);
 608                 if (oldval && !strcmp(val, oldval))
 609                         continue;
 610
 611                 strbuf_addf(dst, " %s=", var);
 612                 sq_quote_buf_pretty(dst, val);
 613         }
 614         string_list_clear(&envs, 0);
 615 }
 616
 617 static void trace_run_command(const struct child_process *cp)
 618 {
 619         struct strbuf buf = STRBUF_INIT;
 620
 621         if (!trace_want(&trace_default_key))
 622                 return;
 623
 624         strbuf_addstr(&buf, "trace: run_command:");
 625         if (cp->dir) {
 626                 strbuf_addstr(&buf, " cd ");
 627                 sq_quote_buf_pretty(&buf, cp->dir);
 628                 strbuf_addch(&buf, ';');
 629         }
 630         /*
 631          * The caller is responsible for initializing cp->env from
 632          * cp->env_array if needed. We only check one place.
 633          */
 634         if (cp->env)
 635                 trace_add_env(&buf, cp->env);
 636         if (cp->git_cmd)
 637                 strbuf_addstr(&buf, " git");
 638         sq_quote_argv_pretty(&buf, cp->argv);
 639
 640         trace_printf("%s", buf.buf);
 641         strbuf_release(&buf);
 642 }
 643
 644 int start_command(struct child_process *cmd)
 645 {
 646         int need_in, need_out, need_err;
 647         int fdin[2], fdout[2], fderr[2];
 648         int failed_errno;
 649         char *str;
 650
 651         if (!cmd->argv)
 652                 cmd->argv = cmd->args.argv;
 653         if (!cmd->env)
 654                 cmd->env = cmd->env_array.argv;
 655
 656         /*
 657          * In case of errors we must keep the promise to close FDs
 658          * that have been passed in via ->in and ->out.
 659          */
 660
 661         need_in = !cmd->no_stdin && cmd->in < 0;
 662         if (need_in) {
 663                 if (pipe(fdin) < 0) {
 664                         failed_errno = errno;
 665                         if (cmd->out > 0)
 666                                 close(cmd->out);
 667                         str = "standard input";
 668                         goto fail_pipe;
 669                 }
 670                 cmd->in = fdin[1];
 671         }
 672
 673         need_out = !cmd->no_stdout
 674                 && !cmd->stdout_to_stderr
 675                 && cmd->out < 0;
 676         if (need_out) {
 677                 if (pipe(fdout) < 0) {
 678                         failed_errno = errno;
 679                         if (need_in)
 680                                 close_pair(fdin);
 681                         else if (cmd->in)
 682                                 close(cmd->in);
 683                         str = "standard output";
 684                         goto fail_pipe;
 685                 }
 686                 cmd->out = fdout[0];
 687         }
 688
 689         need_err = !cmd->no_stderr && cmd->err < 0;
 690         if (need_err) {
 691                 if (pipe(fderr) < 0) {
 692                         failed_errno = errno;
 693                         if (need_in)
 694                                 close_pair(fdin);
 695                         else if (cmd->in)
 696                                 close(cmd->in);
 697                         if (need_out)
 698                                 close_pair(fdout);
 699                         else if (cmd->out)
 700                                 close(cmd->out);
 701                         str = "standard error";
 702 fail_pipe:
 703                         error("cannot create %s pipe for %s: %s",
 704                                 str, cmd->argv[0], strerror(failed_errno));
 705                         child_process_clear(cmd);
 706                         errno = failed_errno;
 707                         return -1;
 708                 }
 709                 cmd->err = fderr[0];
 710         }
 711
 712         trace_run_command(cmd);
 713
 714         fflush(NULL);
 715
 716 #ifndef GIT_WINDOWS_NATIVE
 717 {
 718         int notify_pipe[2];
 719         int null_fd = -1;
 720         char **childenv;
 721         struct argv_array argv = ARGV_ARRAY_INIT;
 722         struct child_err cerr;
 723         struct atfork_state as;
 724
 725         if (pipe(notify_pipe))
 726                 notify_pipe[0] = notify_pipe[1] = -1;
 727
 728         if (cmd->no_stdin || cmd->no_stdout || cmd->no_stderr) {
 729                 null_fd = open("/dev/null", O_RDWR | O_CLOEXEC);
 730                 if (null_fd < 0)
 731                         die_errno(_("open /dev/null failed"));
 732                 set_cloexec(null_fd);
 733         }
 734
 735         prepare_cmd(&argv, cmd);
 736         childenv = prep_childenv(cmd->env);
 737         atfork_prepare(&as);
 738
 739         /*
 740          * NOTE: In order to prevent deadlocking when using threads special
 741          * care should be taken with the function calls made in between the
 742          * fork() and exec() calls.  No calls should be made to functions which
 743          * require acquiring a lock (e.g. malloc) as the lock could have been
 744          * held by another thread at the time of forking, causing the lock to
 745          * never be released in the child process.  This means only
 746          * Async-Signal-Safe functions are permitted in the child.
 747          */
 748         cmd->pid = fork();
 749         failed_errno = errno;
 750         if (!cmd->pid) {
 751                 int sig;
 752                 /*
 753                  * Ensure the default die/error/warn routines do not get
 754                  * called, they can take stdio locks and malloc.
 755                  */
 756                 set_die_routine(child_die_fn);
 757                 set_error_routine(child_error_fn);
 758                 set_warn_routine(child_warn_fn);
 759
 760                 close(notify_pipe[0]);
 761                 set_cloexec(notify_pipe[1]);
 762                 child_notifier = notify_pipe[1];
 763
 764                 if (cmd->no_stdin)
 765                         child_dup2(null_fd, 0);
 766                 else if (need_in) {
 767                         child_dup2(fdin[0], 0);
 768                         child_close_pair(fdin);
 769                 } else if (cmd->in) {
 770                         child_dup2(cmd->in, 0);
 771                         child_close(cmd->in);
 772                 }
 773
 774                 if (cmd->no_stderr)
 775                         child_dup2(null_fd, 2);
 776                 else if (need_err) {
 777                         child_dup2(fderr[1], 2);
 778                         child_close_pair(fderr);
 779                 } else if (cmd->err > 1) {
 780                         child_dup2(cmd->err, 2);
 781                         child_close(cmd->err);
 782                 }
 783
 784                 if (cmd->no_stdout)
 785                         child_dup2(null_fd, 1);
 786                 else if (cmd->stdout_to_stderr)
 787                         child_dup2(2, 1);
 788                 else if (need_out) {
 789                         child_dup2(fdout[1], 1);
 790                         child_close_pair(fdout);
 791                 } else if (cmd->out > 1) {
 792                         child_dup2(cmd->out, 1);
 793                         child_close(cmd->out);
 794                 }
 795
 796                 if (cmd->dir && chdir(cmd->dir))
 797                         child_die(CHILD_ERR_CHDIR);
 798
 799                 /*
 800                  * restore default signal handlers here, in case
 801                  * we catch a signal right before execve below
 802                  */
 803                 for (sig = 1; sig < NSIG; sig++) {
 804                         /* ignored signals get reset to SIG_DFL on execve */
 805                         if (signal(sig, SIG_DFL) == SIG_IGN)
 806                                 signal(sig, SIG_IGN);
 807                 }
 808
 809                 if (sigprocmask(SIG_SETMASK, &as.old, NULL) != 0)
 810                         child_die(CHILD_ERR_SIGPROCMASK);
 811
 812                 /*
 813                  * Attempt to exec using the command and arguments starting at
 814                  * argv.argv[1].  argv.argv[0] contains SHELL_PATH which will
 815                  * be used in the event exec failed with ENOEXEC at which point
 816                  * we will try to interpret the command using 'sh'.
 817                  */
 818                 execve(argv.argv[1], (char *const *) argv.argv + 1,
 819                        (char *const *) childenv);
 820                 if (errno == ENOEXEC)
 821                         execve(argv.argv[0], (char *const *) argv.argv,
 822                                (char *const *) childenv);
 823
 824                 if (errno == ENOENT) {
 825                         if (cmd->silent_exec_failure)
 826                                 child_die(CHILD_ERR_SILENT);
 827                         child_die(CHILD_ERR_ENOENT);
 828                 } else {
 829                         child_die(CHILD_ERR_ERRNO);
 830                 }
 831         }
 832         atfork_parent(&as);
 833         if (cmd->pid < 0)
 834                 error_errno("cannot fork() for %s", cmd->argv[0]);
 835         else if (cmd->clean_on_exit)
 836                 mark_child_for_cleanup(cmd->pid, cmd);
 837
 838         /*
 839          * Wait for child's exec. If the exec succeeds (or if fork()
 840          * failed), EOF is seen immediately by the parent. Otherwise, the
 841          * child process sends a child_err struct.
 842          * Note that use of this infrastructure is completely advisory,
 843          * therefore, we keep error checks minimal.
 844          */
 845         close(notify_pipe[1]);
 846         if (xread(notify_pipe[0], &cerr, sizeof(cerr)) == sizeof(cerr)) {
 847                 /*
 848                  * At this point we know that fork() succeeded, but exec()
 849                  * failed. Errors have been reported to our stderr.
 850                  */
 851                 wait_or_whine(cmd->pid, cmd->argv[0], 0);
 852                 child_err_spew(cmd, &cerr);
 853                 failed_errno = errno;
 854                 cmd->pid = -1;
 855         }
 856         close(notify_pipe[0]);
 857
 858         if (null_fd >= 0)
 859                 close(null_fd);
 860         argv_array_clear(&argv);
 861         free(childenv);
 862 }
 863 #else
 864 {
 865         int fhin = 0, fhout = 1, fherr = 2;
 866         const char **sargv = cmd->argv;
 867         struct argv_array nargv = ARGV_ARRAY_INIT;
 868
 869         if (cmd->no_stdin)
 870                 fhin = open("/dev/null", O_RDWR);
 871         else if (need_in)
 872                 fhin = dup(fdin[0]);
 873         else if (cmd->in)
 874                 fhin = dup(cmd->in);
 875
 876         if (cmd->no_stderr)
 877                 fherr = open("/dev/null", O_RDWR);
 878         else if (need_err)
 879                 fherr = dup(fderr[1]);
 880         else if (cmd->err > 2)
 881                 fherr = dup(cmd->err);
 882
 883         if (cmd->no_stdout)
 884                 fhout = open("/dev/null", O_RDWR);
 885         else if (cmd->stdout_to_stderr)
 886                 fhout = dup(fherr);
 887         else if (need_out)
 888                 fhout = dup(fdout[1]);
 889         else if (cmd->out > 1)
 890                 fhout = dup(cmd->out);
 891
 892         if (cmd->git_cmd)
 893                 cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
 894         else if (cmd->use_shell)
 895                 cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
 896
 897         cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
 898                         cmd->dir, fhin, fhout, fherr);
 899         failed_errno = errno;
 900         if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
 901                 error_errno("cannot spawn %s", cmd->argv[0]);
 902         if (cmd->clean_on_exit && cmd->pid >= 0)
 903                 mark_child_for_cleanup(cmd->pid, cmd);
 904
 905         argv_array_clear(&nargv);
 906         cmd->argv = sargv;
 907         if (fhin != 0)
 908                 close(fhin);
 909         if (fhout != 1)
 910                 close(fhout);
 911         if (fherr != 2)
 912                 close(fherr);
 913 }
 914 #endif
 915
 916         if (cmd->pid < 0) {
 917                 if (need_in)
 918                         close_pair(fdin);
 919                 else if (cmd->in)
 920                         close(cmd->in);
 921                 if (need_out)
 922                         close_pair(fdout);
 923                 else if (cmd->out)
 924                         close(cmd->out);
 925                 if (need_err)
 926                         close_pair(fderr);
 927                 else if (cmd->err)
 928                         close(cmd->err);
 929                 child_process_clear(cmd);
 930                 errno = failed_errno;
 931                 return -1;
 932         }
 933
 934         if (need_in)
 935                 close(fdin[0]);
 936         else if (cmd->in)
 937                 close(cmd->in);
 938
 939         if (need_out)
 940                 close(fdout[1]);
 941         else if (cmd->out)
 942                 close(cmd->out);
 943
 944         if (need_err)
 945                 close(fderr[1]);
 946         else if (cmd->err)
 947                 close(cmd->err);
 948
 949         return 0;
 950 }
 951
 952 int finish_command(struct child_process *cmd)
 953 {
 954         int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
 955         child_process_clear(cmd);
 956         return ret;
 957 }
 958
 959 int finish_command_in_signal(struct child_process *cmd)
 960 {
 961         return wait_or_whine(cmd->pid, cmd->argv[0], 1);
 962 }
 963
 964
 965 int run_command(struct child_process *cmd)
 966 {
 967         int code;
 968
 969         if (cmd->out < 0 || cmd->err < 0)
 970                 die("BUG: run_command with a pipe can cause deadlock");
 971
 972         code = start_command(cmd);
 973         if (code)
 974                 return code;
 975         return finish_command(cmd);
 976 }
 977
 978 int run_command_v_opt(const char **argv, int opt)
 979 {
 980         return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
 981 }
 982
 983 int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
 984 {
 985         struct child_process cmd = CHILD_PROCESS_INIT;
 986         cmd.argv = argv;
 987         cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
 988         cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
 989         cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
 990         cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
 991         cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
 992         cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
 993         cmd.dir = dir;
 994         cmd.env = env;
 995         return run_command(&cmd);
 996 }
 997
 998 #ifndef NO_PTHREADS
 999 static pthread_t main_thread;
1000 static int main_thread_set;
1001 static pthread_key_t async_key;
1002 static pthread_key_t async_die_counter;
1003
1004 static void *run_thread(void *data)
1005 {
1006         struct async *async = data;
1007         intptr_t ret;
1008
1009         if (async->isolate_sigpipe) {
1010                 sigset_t mask;
1011                 sigemptyset(&mask);
1012                 sigaddset(&mask, SIGPIPE);
1013                 if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
1014                         ret = error("unable to block SIGPIPE in async thread");
1015                         return (void *)ret;
1016                 }
1017         }
1018
1019         pthread_setspecific(async_key, async);
1020         ret = async->proc(async->proc_in, async->proc_out, async->data);
1021         return (void *)ret;
1022 }
1023
1024 static NORETURN void die_async(const char *err, va_list params)
1025 {
1026         vreportf("fatal: ", err, params);
1027
1028         if (in_async()) {
1029                 struct async *async = pthread_getspecific(async_key);
1030                 if (async->proc_in >= 0)
1031                         close(async->proc_in);
1032                 if (async->proc_out >= 0)
1033                         close(async->proc_out);
1034                 pthread_exit((void *)128);
1035         }
1036
1037         exit(128);
1038 }
1039
1040 static int async_die_is_recursing(void)
1041 {
1042         void *ret = pthread_getspecific(async_die_counter);
1043         pthread_setspecific(async_die_counter, (void *)1);
1044         return ret != NULL;
1045 }
1046
1047 int in_async(void)
1048 {
1049         if (!main_thread_set)
1050                 return 0; /* no asyncs started yet */
1051         return !pthread_equal(main_thread, pthread_self());
1052 }
1053
1054 static void NORETURN async_exit(int code)
1055 {
1056         pthread_exit((void *)(intptr_t)code);
1057 }
1058
1059 #else
1060
1061 static struct {
1062         void (**handlers)(void);
1063         size_t nr;
1064         size_t alloc;
1065 } git_atexit_hdlrs;
1066
1067 static int git_atexit_installed;
1068
1069 static void git_atexit_dispatch(void)
1070 {
1071         size_t i;
1072
1073         for (i=git_atexit_hdlrs.nr ; i ; i--)
1074                 git_atexit_hdlrs.handlers[i-1]();
1075 }
1076
1077 static void git_atexit_clear(void)
1078 {
1079         free(git_atexit_hdlrs.handlers);
1080         memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
1081         git_atexit_installed = 0;
1082 }
1083
1084 #undef atexit
1085 int git_atexit(void (*handler)(void))
1086 {
1087         ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
1088         git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
1089         if (!git_atexit_installed) {
1090                 if (atexit(&git_atexit_dispatch))
1091                         return -1;
1092                 git_atexit_installed = 1;
1093         }
1094         return 0;
1095 }
1096 #define atexit git_atexit
1097
1098 static int process_is_async;
1099 int in_async(void)
1100 {
1101         return process_is_async;
1102 }
1103
1104 static void NORETURN async_exit(int code)
1105 {
1106         exit(code);
1107 }
1108
1109 #endif
1110
1111 void check_pipe(int err)
1112 {
1113         if (err == EPIPE) {
1114                 if (in_async())
1115                         async_exit(141);
1116
1117                 signal(SIGPIPE, SIG_DFL);
1118                 raise(SIGPIPE);
1119                 /* Should never happen, but just in case... */
1120                 exit(141);
1121         }
1122 }
1123
1124 int start_async(struct async *async)
1125 {
1126         int need_in, need_out;
1127         int fdin[2], fdout[2];
1128         int proc_in, proc_out;
1129
1130         need_in = async->in < 0;
1131         if (need_in) {
1132                 if (pipe(fdin) < 0) {
1133                         if (async->out > 0)
1134                                 close(async->out);
1135                         return error_errno("cannot create pipe");
1136                 }
1137                 async->in = fdin[1];
1138         }
1139
1140         need_out = async->out < 0;
1141         if (need_out) {
1142                 if (pipe(fdout) < 0) {
1143                         if (need_in)
1144                                 close_pair(fdin);
1145                         else if (async->in)
1146                                 close(async->in);
1147                         return error_errno("cannot create pipe");
1148                 }
1149                 async->out = fdout[0];
1150         }
1151
1152         if (need_in)
1153                 proc_in = fdin[0];
1154         else if (async->in)
1155                 proc_in = async->in;
1156         else
1157                 proc_in = -1;
1158
1159         if (need_out)
1160                 proc_out = fdout[1];
1161         else if (async->out)
1162                 proc_out = async->out;
1163         else
1164                 proc_out = -1;
1165
1166 #ifdef NO_PTHREADS
1167         /* Flush stdio before fork() to avoid cloning buffers */
1168         fflush(NULL);
1169
1170         async->pid = fork();
1171         if (async->pid < 0) {
1172                 error_errno("fork (async) failed");
1173                 goto error;
1174         }
1175         if (!async->pid) {
1176                 if (need_in)
1177                         close(fdin[1]);
1178                 if (need_out)
1179                         close(fdout[0]);
1180                 git_atexit_clear();
1181                 process_is_async = 1;
1182                 exit(!!async->proc(proc_in, proc_out, async->data));
1183         }
1184
1185         mark_child_for_cleanup(async->pid, NULL);
1186
1187         if (need_in)
1188                 close(fdin[0]);
1189         else if (async->in)
1190                 close(async->in);
1191
1192         if (need_out)
1193                 close(fdout[1]);
1194         else if (async->out)
1195                 close(async->out);
1196 #else
1197         if (!main_thread_set) {
1198                 /*
1199                  * We assume that the first time that start_async is called
1200                  * it is from the main thread.
1201                  */
1202                 main_thread_set = 1;
1203                 main_thread = pthread_self();
1204                 pthread_key_create(&async_key, NULL);
1205                 pthread_key_create(&async_die_counter, NULL);
1206                 set_die_routine(die_async);
1207                 set_die_is_recursing_routine(async_die_is_recursing);
1208         }
1209
1210         if (proc_in >= 0)
1211                 set_cloexec(proc_in);
1212         if (proc_out >= 0)
1213                 set_cloexec(proc_out);
1214         async->proc_in = proc_in;
1215         async->proc_out = proc_out;
1216         {
1217                 int err = pthread_create(&async->tid, NULL, run_thread, async);
1218                 if (err) {
1219                         error_errno("cannot create thread");
1220                         goto error;
1221                 }
1222         }
1223 #endif
1224         return 0;
1225
1226 error:
1227         if (need_in)
1228                 close_pair(fdin);
1229         else if (async->in)
1230                 close(async->in);
1231
1232         if (need_out)
1233                 close_pair(fdout);
1234         else if (async->out)
1235                 close(async->out);
1236         return -1;
1237 }
1238
1239 int finish_async(struct async *async)
1240 {
1241 #ifdef NO_PTHREADS
1242         return wait_or_whine(async->pid, "child process", 0);
1243 #else
1244         void *ret = (void *)(intptr_t)(-1);
1245
1246         if (pthread_join(async->tid, &ret))
1247                 error("pthread_join failed");
1248         return (int)(intptr_t)ret;
1249 #endif
1250 }
1251
1252 const char *find_hook(const char *name)
1253 {
1254         static struct strbuf path = STRBUF_INIT;
1255
1256         strbuf_reset(&path);
1257         strbuf_git_path(&path, "hooks/%s", name);
1258         if (access(path.buf, X_OK) < 0) {
1259                 int err = errno;
1260
1261 #ifdef STRIP_EXTENSION
1262                 strbuf_addstr(&path, STRIP_EXTENSION);
1263                 if (access(path.buf, X_OK) >= 0)
1264                         return path.buf;
1265                 if (errno == EACCES)
1266                         err = errno;
1267 #endif
1268
1269                 if (err == EACCES && advice_ignored_hook) {
1270                         static struct string_list advise_given = STRING_LIST_INIT_DUP;
1271
1272                         if (!string_list_lookup(&advise_given, name)) {
1273                                 string_list_insert(&advise_given, name);
1274                                 advise(_("The '%s' hook was ignored because "
1275                                          "it's not set as executable.\n"
1276                                          "You can disable this warning with "
1277                                          "`git config advice.ignoredHook false`."),
1278                                        path.buf);
1279                         }
1280                 }
1281                 return NULL;
1282         }
1283         return path.buf;
1284 }
1285
1286 int run_hook_ve(const char *const *env, const char *name, va_list args)
1287 {
1288         struct child_process hook = CHILD_PROCESS_INIT;
1289         const char *p;
1290
1291         p = find_hook(name);
1292         if (!p)
1293                 return 0;
1294
1295         argv_array_push(&hook.args, p);
1296         while ((p = va_arg(args, const char *)))
1297                 argv_array_push(&hook.args, p);
1298         hook.env = env;
1299         hook.no_stdin = 1;
1300         hook.stdout_to_stderr = 1;
1301
1302         return run_command(&hook);
1303 }
1304
1305 int run_hook_le(const char *const *env, const char *name, ...)
1306 {
1307         va_list args;
1308         int ret;
1309
1310         va_start(args, name);
1311         ret = run_hook_ve(env, name, args);
1312         va_end(args);
1313
1314         return ret;
1315 }
1316
1317 struct io_pump {
1318         /* initialized by caller */
1319         int fd;
1320         int type; /* POLLOUT or POLLIN */
1321         union {
1322                 struct {
1323                         const char *buf;
1324                         size_t len;
1325                 } out;
1326                 struct {
1327                         struct strbuf *buf;
1328                         size_t hint;
1329                 } in;
1330         } u;
1331
1332         /* returned by pump_io */
1333         int error; /* 0 for success, otherwise errno */
1334
1335         /* internal use */
1336         struct pollfd *pfd;
1337 };
1338
1339 static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
1340 {
1341         int pollsize = 0;
1342         int i;
1343
1344         for (i = 0; i < nr; i++) {
1345                 struct io_pump *io = &slots[i];
1346                 if (io->fd < 0)
1347                         continue;
1348                 pfd[pollsize].fd = io->fd;
1349                 pfd[pollsize].events = io->type;
1350                 io->pfd = &pfd[pollsize++];
1351         }
1352
1353         if (!pollsize)
1354                 return 0;
1355
1356         if (poll(pfd, pollsize, -1) < 0) {
1357                 if (errno == EINTR)
1358                         return 1;
1359                 die_errno("poll failed");
1360         }
1361
1362         for (i = 0; i < nr; i++) {
1363                 struct io_pump *io = &slots[i];
1364
1365                 if (io->fd < 0)
1366                         continue;
1367
1368                 if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
1369                         continue;
1370
1371                 if (io->type == POLLOUT) {
1372                         ssize_t len = xwrite(io->fd,
1373                                              io->u.out.buf, io->u.out.len);
1374                         if (len < 0) {
1375                                 io->error = errno;
1376                                 close(io->fd);
1377                                 io->fd = -1;
1378                         } else {
1379                                 io->u.out.buf += len;
1380                                 io->u.out.len -= len;
1381                                 if (!io->u.out.len) {
1382                                         close(io->fd);
1383                                         io->fd = -1;
1384                                 }
1385                         }
1386                 }
1387
1388                 if (io->type == POLLIN) {
1389                         ssize_t len = strbuf_read_once(io->u.in.buf,
1390                                                        io->fd, io->u.in.hint);
1391                         if (len < 0)
1392                                 io->error = errno;
1393                         if (len <= 0) {
1394                                 close(io->fd);
1395                                 io->fd = -1;
1396                         }
1397                 }
1398         }
1399
1400         return 1;
1401 }
1402
1403 static int pump_io(struct io_pump *slots, int nr)
1404 {
1405         struct pollfd *pfd;
1406         int i;
1407
1408         for (i = 0; i < nr; i++)
1409                 slots[i].error = 0;
1410
1411         ALLOC_ARRAY(pfd, nr);
1412         while (pump_io_round(slots, nr, pfd))
1413                 ; /* nothing */
1414         free(pfd);
1415
1416         /* There may be multiple errno values, so just pick the first. */
1417         for (i = 0; i < nr; i++) {
1418                 if (slots[i].error) {
1419                         errno = slots[i].error;
1420                         return -1;
1421                 }
1422         }
1423         return 0;
1424 }
1425
1426
1427 int pipe_command(struct child_process *cmd,
1428                  const char *in, size_t in_len,
1429                  struct strbuf *out, size_t out_hint,
1430                  struct strbuf *err, size_t err_hint)
1431 {
1432         struct io_pump io[3];
1433         int nr = 0;
1434
1435         if (in)
1436                 cmd->in = -1;
1437         if (out)
1438                 cmd->out = -1;
1439         if (err)
1440                 cmd->err = -1;
1441
1442         if (start_command(cmd) < 0)
1443                 return -1;
1444
1445         if (in) {
1446                 io[nr].fd = cmd->in;
1447                 io[nr].type = POLLOUT;
1448                 io[nr].u.out.buf = in;
1449                 io[nr].u.out.len = in_len;
1450                 nr++;
1451         }
1452         if (out) {
1453                 io[nr].fd = cmd->out;
1454                 io[nr].type = POLLIN;
1455                 io[nr].u.in.buf = out;
1456                 io[nr].u.in.hint = out_hint;
1457                 nr++;
1458         }
1459         if (err) {
1460                 io[nr].fd = cmd->err;
1461                 io[nr].type = POLLIN;
1462                 io[nr].u.in.buf = err;
1463                 io[nr].u.in.hint = err_hint;
1464                 nr++;
1465         }
1466
1467         if (pump_io(io, nr) < 0) {
1468                 finish_command(cmd); /* throw away exit code */
1469                 return -1;
1470         }
1471
1472         return finish_command(cmd);
1473 }
1474
1475 enum child_state {
1476         GIT_CP_FREE,
1477         GIT_CP_WORKING,
1478         GIT_CP_WAIT_CLEANUP,
1479 };
1480
1481 struct parallel_processes {
1482         void *data;
1483
1484         int max_processes;
1485         int nr_processes;
1486
1487         get_next_task_fn get_next_task;
1488         start_failure_fn start_failure;
1489         task_finished_fn task_finished;
1490
1491         struct {
1492                 enum child_state state;
1493                 struct child_process process;
1494                 struct strbuf err;
1495                 void *data;
1496         } *children;
1497         /*
1498          * The struct pollfd is logically part of *children,
1499          * but the system call expects it as its own array.
1500          */
1501         struct pollfd *pfd;
1502
1503         unsigned shutdown : 1;
1504
1505         int output_owner;
1506         struct strbuf buffered_output; /* of finished children */
1507 };
1508
1509 static int default_start_failure(struct strbuf *out,
1510                                  void *pp_cb,
1511                                  void *pp_task_cb)
1512 {
1513         return 0;
1514 }
1515
1516 static int default_task_finished(int result,
1517                                  struct strbuf *out,
1518                                  void *pp_cb,
1519                                  void *pp_task_cb)
1520 {
1521         return 0;
1522 }
1523
1524 static void kill_children(struct parallel_processes *pp, int signo)
1525 {
1526         int i, n = pp->max_processes;
1527
1528         for (i = 0; i < n; i++)
1529                 if (pp->children[i].state == GIT_CP_WORKING)
1530                         kill(pp->children[i].process.pid, signo);
1531 }
1532
1533 static struct parallel_processes *pp_for_signal;
1534
1535 static void handle_children_on_signal(int signo)
1536 {
1537         kill_children(pp_for_signal, signo);
1538         sigchain_pop(signo);
1539         raise(signo);
1540 }
1541
1542 static void pp_init(struct parallel_processes *pp,
1543                     int n,
1544                     get_next_task_fn get_next_task,
1545                     start_failure_fn start_failure,
1546                     task_finished_fn task_finished,
1547                     void *data)
1548 {
1549         int i;
1550
1551         if (n < 1)
1552                 n = online_cpus();
1553
1554         pp->max_processes = n;
1555
1556         trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
1557
1558         pp->data = data;
1559         if (!get_next_task)
1560                 die("BUG: you need to specify a get_next_task function");
1561         pp->get_next_task = get_next_task;
1562
1563         pp->start_failure = start_failure ? start_failure : default_start_failure;
1564         pp->task_finished = task_finished ? task_finished : default_task_finished;
1565
1566         pp->nr_processes = 0;
1567         pp->output_owner = 0;
1568         pp->shutdown = 0;
1569         pp->children = xcalloc(n, sizeof(*pp->children));
1570         pp->pfd = xcalloc(n, sizeof(*pp->pfd));
1571         strbuf_init(&pp->buffered_output, 0);
1572
1573         for (i = 0; i < n; i++) {
1574                 strbuf_init(&pp->children[i].err, 0);
1575                 child_process_init(&pp->children[i].process);
1576                 pp->pfd[i].events = POLLIN | POLLHUP;
1577                 pp->pfd[i].fd = -1;
1578         }
1579
1580         pp_for_signal = pp;
1581         sigchain_push_common(handle_children_on_signal);
1582 }
1583
1584 static void pp_cleanup(struct parallel_processes *pp)
1585 {
1586         int i;
1587
1588         trace_printf("run_processes_parallel: done");
1589         for (i = 0; i < pp->max_processes; i++) {
1590                 strbuf_release(&pp->children[i].err);
1591                 child_process_clear(&pp->children[i].process);
1592         }
1593
1594         free(pp->children);
1595         free(pp->pfd);
1596
1597         /*
1598          * When get_next_task added messages to the buffer in its last
1599          * iteration, the buffered output is non empty.
1600          */
1601         strbuf_write(&pp->buffered_output, stderr);
1602         strbuf_release(&pp->buffered_output);
1603
1604         sigchain_pop_common();
1605 }
1606
1607 /* returns
1608  *  0 if a new task was started.
1609  *  1 if no new jobs was started (get_next_task ran out of work, non critical
1610  *    problem with starting a new command)
1611  * <0 no new job was started, user wishes to shutdown early. Use negative code
1612  *    to signal the children.
1613  */
1614 static int pp_start_one(struct parallel_processes *pp)
1615 {
1616         int i, code;
1617
1618         for (i = 0; i < pp->max_processes; i++)
1619                 if (pp->children[i].state == GIT_CP_FREE)
1620                         break;
1621         if (i == pp->max_processes)
1622                 die("BUG: bookkeeping is hard");
1623
1624         code = pp->get_next_task(&pp->children[i].process,
1625                                  &pp->children[i].err,
1626                                  pp->data,
1627                                  &pp->children[i].data);
1628         if (!code) {
1629                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1630                 strbuf_reset(&pp->children[i].err);
1631                 return 1;
1632         }
1633         pp->children[i].process.err = -1;
1634         pp->children[i].process.stdout_to_stderr = 1;
1635         pp->children[i].process.no_stdin = 1;
1636
1637         if (start_command(&pp->children[i].process)) {
1638                 code = pp->start_failure(&pp->children[i].err,
1639                                          pp->data,
1640                                          pp->children[i].data);
1641                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1642                 strbuf_reset(&pp->children[i].err);
1643                 if (code)
1644                         pp->shutdown = 1;
1645                 return code;
1646         }
1647
1648         pp->nr_processes++;
1649         pp->children[i].state = GIT_CP_WORKING;
1650         pp->pfd[i].fd = pp->children[i].process.err;
1651         return 0;
1652 }
1653
1654 static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1655 {
1656         int i;
1657
1658         while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1659                 if (errno == EINTR)
1660                         continue;
1661                 pp_cleanup(pp);
1662                 die_errno("poll");
1663         }
1664
1665         /* Buffer output from all pipes. */
1666         for (i = 0; i < pp->max_processes; i++) {
1667                 if (pp->children[i].state == GIT_CP_WORKING &&
1668                     pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1669                         int n = strbuf_read_once(&pp->children[i].err,
1670                                                  pp->children[i].process.err, 0);
1671                         if (n == 0) {
1672                                 close(pp->children[i].process.err);
1673                                 pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1674                         } else if (n < 0)
1675                                 if (errno != EAGAIN)
1676                                         die_errno("read");
1677                 }
1678         }
1679 }
1680
1681 static void pp_output(struct parallel_processes *pp)
1682 {
1683         int i = pp->output_owner;
1684         if (pp->children[i].state == GIT_CP_WORKING &&
1685             pp->children[i].err.len) {
1686                 strbuf_write(&pp->children[i].err, stderr);
1687                 strbuf_reset(&pp->children[i].err);
1688         }
1689 }
1690
1691 static int pp_collect_finished(struct parallel_processes *pp)
1692 {
1693         int i, code;
1694         int n = pp->max_processes;
1695         int result = 0;
1696
1697         while (pp->nr_processes > 0) {
1698                 for (i = 0; i < pp->max_processes; i++)
1699                         if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1700                                 break;
1701                 if (i == pp->max_processes)
1702                         break;
1703
1704                 code = finish_command(&pp->children[i].process);
1705
1706                 code = pp->task_finished(code,
1707                                          &pp->children[i].err, pp->data,
1708                                          pp->children[i].data);
1709
1710                 if (code)
1711                         result = code;
1712                 if (code < 0)
1713                         break;
1714
1715                 pp->nr_processes--;
1716                 pp->children[i].state = GIT_CP_FREE;
1717                 pp->pfd[i].fd = -1;
1718                 child_process_init(&pp->children[i].process);
1719
1720                 if (i != pp->output_owner) {
1721                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1722                         strbuf_reset(&pp->children[i].err);
1723                 } else {
1724                         strbuf_write(&pp->children[i].err, stderr);
1725                         strbuf_reset(&pp->children[i].err);
1726
1727                         /* Output all other finished child processes */
1728                         strbuf_write(&pp->buffered_output, stderr);
1729                         strbuf_reset(&pp->buffered_output);
1730
1731                         /*
1732                          * Pick next process to output live.
1733                          * NEEDSWORK:
1734                          * For now we pick it randomly by doing a round
1735                          * robin. Later we may want to pick the one with
1736                          * the most output or the longest or shortest
1737                          * running process time.
1738                          */
1739                         for (i = 0; i < n; i++)
1740                                 if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1741                                         break;
1742                         pp->output_owner = (pp->output_owner + i) % n;
1743                 }
1744         }
1745         return result;
1746 }
1747
1748 int run_processes_parallel(int n,
1749                            get_next_task_fn get_next_task,
1750                            start_failure_fn start_failure,
1751                            task_finished_fn task_finished,
1752                            void *pp_cb)
1753 {
1754         int i, code;
1755         int output_timeout = 100;
1756         int spawn_cap = 4;
1757         struct parallel_processes pp;
1758
1759         pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1760         while (1) {
1761                 for (i = 0;
1762                     i < spawn_cap && !pp.shutdown &&
1763                     pp.nr_processes < pp.max_processes;
1764                     i++) {
1765                         code = pp_start_one(&pp);
1766                         if (!code)
1767                                 continue;
1768                         if (code < 0) {
1769                                 pp.shutdown = 1;
1770                                 kill_children(&pp, -code);
1771                         }
1772                         break;
1773                 }
1774                 if (!pp.nr_processes)
1775                         break;
1776                 pp_buffer_stderr(&pp, output_timeout);
1777                 pp_output(&pp);
1778                 code = pp_collect_finished(&pp);
1779                 if (code) {
1780                         pp.shutdown = 1;
1781                         if (code < 0)
1782                                 kill_children(&pp, -code);
1783                 }
1784         }
1785
1786         pp_cleanup(&pp);
1787         return 0;
1788 }