run-command.c

   1 #include "cache.h"
   2 #include "run-command.h"
   3 #include "exec-cmd.h"
   4 #include "sigchain.h"
   5 #include "strvec.h"
   6 #include "thread-utils.h"
   7 #include "strbuf.h"
   8 #include "string-list.h"
   9 #include "quote.h"
  10 #include "config.h"
  11 #include "packfile.h"
  12 #include "hook.h"
  13 #include "compat/nonblock.h"
  14
  15 void child_process_init(struct child_process *child)
  16 {
  17         struct child_process blank = CHILD_PROCESS_INIT;
  18         memcpy(child, &blank, sizeof(*child));
  19 }
  20
  21 void child_process_clear(struct child_process *child)
  22 {
  23         strvec_clear(&child->args);
  24         strvec_clear(&child->env);
  25 }
  26
  27 struct child_to_clean {
  28         pid_t pid;
  29         struct child_process *process;
  30         struct child_to_clean *next;
  31 };
  32 static struct child_to_clean *children_to_clean;
  33 static int installed_child_cleanup_handler;
  34
  35 static void cleanup_children(int sig, int in_signal)
  36 {
  37         struct child_to_clean *children_to_wait_for = NULL;
  38
  39         while (children_to_clean) {
  40                 struct child_to_clean *p = children_to_clean;
  41                 children_to_clean = p->next;
  42
  43                 if (p->process && !in_signal) {
  44                         struct child_process *process = p->process;
  45                         if (process->clean_on_exit_handler) {
  46                                 trace_printf(
  47                                         "trace: run_command: running exit handler for pid %"
  48                                         PRIuMAX, (uintmax_t)p->pid
  49                                 );
  50                                 process->clean_on_exit_handler(process);
  51                         }
  52                 }
  53
  54                 kill(p->pid, sig);
  55
  56                 if (p->process && p->process->wait_after_clean) {
  57                         p->next = children_to_wait_for;
  58                         children_to_wait_for = p;
  59                 } else {
  60                         if (!in_signal)
  61                                 free(p);
  62                 }
  63         }
  64
  65         while (children_to_wait_for) {
  66                 struct child_to_clean *p = children_to_wait_for;
  67                 children_to_wait_for = p->next;
  68
  69                 while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
  70                         ; /* spin waiting for process exit or error */
  71
  72                 if (!in_signal)
  73                         free(p);
  74         }
  75 }
  76
  77 static void cleanup_children_on_signal(int sig)
  78 {
  79         cleanup_children(sig, 1);
  80         sigchain_pop(sig);
  81         raise(sig);
  82 }
  83
  84 static void cleanup_children_on_exit(void)
  85 {
  86         cleanup_children(SIGTERM, 0);
  87 }
  88
  89 static void mark_child_for_cleanup(pid_t pid, struct child_process *process)
  90 {
  91         struct child_to_clean *p = xmalloc(sizeof(*p));
  92         p->pid = pid;
  93         p->process = process;
  94         p->next = children_to_clean;
  95         children_to_clean = p;
  96
  97         if (!installed_child_cleanup_handler) {
  98                 atexit(cleanup_children_on_exit);
  99                 sigchain_push_common(cleanup_children_on_signal);
 100                 installed_child_cleanup_handler = 1;
 101         }
 102 }
 103
 104 static void clear_child_for_cleanup(pid_t pid)
 105 {
 106         struct child_to_clean **pp;
 107
 108         for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
 109                 struct child_to_clean *clean_me = *pp;
 110
 111                 if (clean_me->pid == pid) {
 112                         *pp = clean_me->next;
 113                         free(clean_me);
 114                         return;
 115                 }
 116         }
 117 }
 118
 119 static inline void close_pair(int fd[2])
 120 {
 121         close(fd[0]);
 122         close(fd[1]);
 123 }
 124
 125 int is_executable(const char *name)
 126 {
 127         struct stat st;
 128
 129         if (stat(name, &st) || /* stat, not lstat */
 130             !S_ISREG(st.st_mode))
 131                 return 0;
 132
 133 #if defined(GIT_WINDOWS_NATIVE)
 134         /*
 135          * On Windows there is no executable bit. The file extension
 136          * indicates whether it can be run as an executable, and Git
 137          * has special-handling to detect scripts and launch them
 138          * through the indicated script interpreter. We test for the
 139          * file extension first because virus scanners may make
 140          * it quite expensive to open many files.
 141          */
 142         if (ends_with(name, ".exe"))
 143                 return S_IXUSR;
 144
 145 {
 146         /*
 147          * Now that we know it does not have an executable extension,
 148          * peek into the file instead.
 149          */
 150         char buf[3] = { 0 };
 151         int n;
 152         int fd = open(name, O_RDONLY);
 153         st.st_mode &= ~S_IXUSR;
 154         if (fd >= 0) {
 155                 n = read(fd, buf, 2);
 156                 if (n == 2)
 157                         /* look for a she-bang */
 158                         if (!strcmp(buf, "#!"))
 159                                 st.st_mode |= S_IXUSR;
 160                 close(fd);
 161         }
 162 }
 163 #endif
 164         return st.st_mode & S_IXUSR;
 165 }
 166
 167 /*
 168  * Search $PATH for a command.  This emulates the path search that
 169  * execvp would perform, without actually executing the command so it
 170  * can be used before fork() to prepare to run a command using
 171  * execve() or after execvp() to diagnose why it failed.
 172  *
 173  * The caller should ensure that file contains no directory
 174  * separators.
 175  *
 176  * Returns the path to the command, as found in $PATH or NULL if the
 177  * command could not be found.  The caller inherits ownership of the memory
 178  * used to store the resultant path.
 179  *
 180  * This should not be used on Windows, where the $PATH search rules
 181  * are more complicated (e.g., a search for "foo" should find
 182  * "foo.exe").
 183  */
 184 static char *locate_in_PATH(const char *file)
 185 {
 186         const char *p = getenv("PATH");
 187         struct strbuf buf = STRBUF_INIT;
 188
 189         if (!p || !*p)
 190                 return NULL;
 191
 192         while (1) {
 193                 const char *end = strchrnul(p, ':');
 194
 195                 strbuf_reset(&buf);
 196
 197                 /* POSIX specifies an empty entry as the current directory. */
 198                 if (end != p) {
 199                         strbuf_add(&buf, p, end - p);
 200                         strbuf_addch(&buf, '/');
 201                 }
 202                 strbuf_addstr(&buf, file);
 203
 204                 if (is_executable(buf.buf))
 205                         return strbuf_detach(&buf, NULL);
 206
 207                 if (!*end)
 208                         break;
 209                 p = end + 1;
 210         }
 211
 212         strbuf_release(&buf);
 213         return NULL;
 214 }
 215
 216 int exists_in_PATH(const char *command)
 217 {
 218         char *r = locate_in_PATH(command);
 219         int found = r != NULL;
 220         free(r);
 221         return found;
 222 }
 223
 224 int sane_execvp(const char *file, char * const argv[])
 225 {
 226 #ifndef GIT_WINDOWS_NATIVE
 227         /*
 228          * execvp() doesn't return, so we all we can do is tell trace2
 229          * what we are about to do and let it leave a hint in the log
 230          * (unless of course the execvp() fails).
 231          *
 232          * we skip this for Windows because the compat layer already
 233          * has to emulate the execvp() call anyway.
 234          */
 235         int exec_id = trace2_exec(file, (const char **)argv);
 236 #endif
 237
 238         if (!execvp(file, argv))
 239                 return 0; /* cannot happen ;-) */
 240
 241 #ifndef GIT_WINDOWS_NATIVE
 242         {
 243                 int ec = errno;
 244                 trace2_exec_result(exec_id, ec);
 245                 errno = ec;
 246         }
 247 #endif
 248
 249         /*
 250          * When a command can't be found because one of the directories
 251          * listed in $PATH is unsearchable, execvp reports EACCES, but
 252          * careful usability testing (read: analysis of occasional bug
 253          * reports) reveals that "No such file or directory" is more
 254          * intuitive.
 255          *
 256          * We avoid commands with "/", because execvp will not do $PATH
 257          * lookups in that case.
 258          *
 259          * The reassignment of EACCES to errno looks like a no-op below,
 260          * but we need to protect against exists_in_PATH overwriting errno.
 261          */
 262         if (errno == EACCES && !strchr(file, '/'))
 263                 errno = exists_in_PATH(file) ? EACCES : ENOENT;
 264         else if (errno == ENOTDIR && !strchr(file, '/'))
 265                 errno = ENOENT;
 266         return -1;
 267 }
 268
 269 static const char **prepare_shell_cmd(struct strvec *out, const char **argv)
 270 {
 271         if (!argv[0])
 272                 BUG("shell command is empty");
 273
 274         if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
 275 #ifndef GIT_WINDOWS_NATIVE
 276                 strvec_push(out, SHELL_PATH);
 277 #else
 278                 strvec_push(out, "sh");
 279 #endif
 280                 strvec_push(out, "-c");
 281
 282                 /*
 283                  * If we have no extra arguments, we do not even need to
 284                  * bother with the "$@" magic.
 285                  */
 286                 if (!argv[1])
 287                         strvec_push(out, argv[0]);
 288                 else
 289                         strvec_pushf(out, "%s \"$@\"", argv[0]);
 290         }
 291
 292         strvec_pushv(out, argv);
 293         return out->v;
 294 }
 295
 296 #ifndef GIT_WINDOWS_NATIVE
 297 static int child_notifier = -1;
 298
 299 enum child_errcode {
 300         CHILD_ERR_CHDIR,
 301         CHILD_ERR_DUP2,
 302         CHILD_ERR_CLOSE,
 303         CHILD_ERR_SIGPROCMASK,
 304         CHILD_ERR_ENOENT,
 305         CHILD_ERR_SILENT,
 306         CHILD_ERR_ERRNO
 307 };
 308
 309 struct child_err {
 310         enum child_errcode err;
 311         int syserr; /* errno */
 312 };
 313
 314 static void child_die(enum child_errcode err)
 315 {
 316         struct child_err buf;
 317
 318         buf.err = err;
 319         buf.syserr = errno;
 320
 321         /* write(2) on buf smaller than PIPE_BUF (min 512) is atomic: */
 322         xwrite(child_notifier, &buf, sizeof(buf));
 323         _exit(1);
 324 }
 325
 326 static void child_dup2(int fd, int to)
 327 {
 328         if (dup2(fd, to) < 0)
 329                 child_die(CHILD_ERR_DUP2);
 330 }
 331
 332 static void child_close(int fd)
 333 {
 334         if (close(fd))
 335                 child_die(CHILD_ERR_CLOSE);
 336 }
 337
 338 static void child_close_pair(int fd[2])
 339 {
 340         child_close(fd[0]);
 341         child_close(fd[1]);
 342 }
 343
 344 static void child_error_fn(const char *err, va_list params)
 345 {
 346         const char msg[] = "error() should not be called in child\n";
 347         xwrite(2, msg, sizeof(msg) - 1);
 348 }
 349
 350 static void child_warn_fn(const char *err, va_list params)
 351 {
 352         const char msg[] = "warn() should not be called in child\n";
 353         xwrite(2, msg, sizeof(msg) - 1);
 354 }
 355
 356 static void NORETURN child_die_fn(const char *err, va_list params)
 357 {
 358         const char msg[] = "die() should not be called in child\n";
 359         xwrite(2, msg, sizeof(msg) - 1);
 360         _exit(2);
 361 }
 362
 363 /* this runs in the parent process */
 364 static void child_err_spew(struct child_process *cmd, struct child_err *cerr)
 365 {
 366         static void (*old_errfn)(const char *err, va_list params);
 367         report_fn die_message_routine = get_die_message_routine();
 368
 369         old_errfn = get_error_routine();
 370         set_error_routine(die_message_routine);
 371         errno = cerr->syserr;
 372
 373         switch (cerr->err) {
 374         case CHILD_ERR_CHDIR:
 375                 error_errno("exec '%s': cd to '%s' failed",
 376                             cmd->args.v[0], cmd->dir);
 377                 break;
 378         case CHILD_ERR_DUP2:
 379                 error_errno("dup2() in child failed");
 380                 break;
 381         case CHILD_ERR_CLOSE:
 382                 error_errno("close() in child failed");
 383                 break;
 384         case CHILD_ERR_SIGPROCMASK:
 385                 error_errno("sigprocmask failed restoring signals");
 386                 break;
 387         case CHILD_ERR_ENOENT:
 388                 error_errno("cannot run %s", cmd->args.v[0]);
 389                 break;
 390         case CHILD_ERR_SILENT:
 391                 break;
 392         case CHILD_ERR_ERRNO:
 393                 error_errno("cannot exec '%s'", cmd->args.v[0]);
 394                 break;
 395         }
 396         set_error_routine(old_errfn);
 397 }
 398
 399 static int prepare_cmd(struct strvec *out, const struct child_process *cmd)
 400 {
 401         if (!cmd->args.v[0])
 402                 BUG("command is empty");
 403
 404         /*
 405          * Add SHELL_PATH so in the event exec fails with ENOEXEC we can
 406          * attempt to interpret the command with 'sh'.
 407          */
 408         strvec_push(out, SHELL_PATH);
 409
 410         if (cmd->git_cmd) {
 411                 prepare_git_cmd(out, cmd->args.v);
 412         } else if (cmd->use_shell) {
 413                 prepare_shell_cmd(out, cmd->args.v);
 414         } else {
 415                 strvec_pushv(out, cmd->args.v);
 416         }
 417
 418         /*
 419          * If there are no dir separator characters in the command then perform
 420          * a path lookup and use the resolved path as the command to exec. If
 421          * there are dir separator characters, we have exec attempt to invoke
 422          * the command directly.
 423          */
 424         if (!has_dir_sep(out->v[1])) {
 425                 char *program = locate_in_PATH(out->v[1]);
 426                 if (program) {
 427                         free((char *)out->v[1]);
 428                         out->v[1] = program;
 429                 } else {
 430                         strvec_clear(out);
 431                         errno = ENOENT;
 432                         return -1;
 433                 }
 434         }
 435
 436         return 0;
 437 }
 438
 439 static char **prep_childenv(const char *const *deltaenv)
 440 {
 441         extern char **environ;
 442         char **childenv;
 443         struct string_list env = STRING_LIST_INIT_DUP;
 444         struct strbuf key = STRBUF_INIT;
 445         const char *const *p;
 446         int i;
 447
 448         /* Construct a sorted string list consisting of the current environ */
 449         for (p = (const char *const *) environ; p && *p; p++) {
 450                 const char *equals = strchr(*p, '=');
 451
 452                 if (equals) {
 453                         strbuf_reset(&key);
 454                         strbuf_add(&key, *p, equals - *p);
 455                         string_list_append(&env, key.buf)->util = (void *) *p;
 456                 } else {
 457                         string_list_append(&env, *p)->util = (void *) *p;
 458                 }
 459         }
 460         string_list_sort(&env);
 461
 462         /* Merge in 'deltaenv' with the current environ */
 463         for (p = deltaenv; p && *p; p++) {
 464                 const char *equals = strchr(*p, '=');
 465
 466                 if (equals) {
 467                         /* ('key=value'), insert or replace entry */
 468                         strbuf_reset(&key);
 469                         strbuf_add(&key, *p, equals - *p);
 470                         string_list_insert(&env, key.buf)->util = (void *) *p;
 471                 } else {
 472                         /* otherwise ('key') remove existing entry */
 473                         string_list_remove(&env, *p, 0);
 474                 }
 475         }
 476
 477         /* Create an array of 'char *' to be used as the childenv */
 478         ALLOC_ARRAY(childenv, env.nr + 1);
 479         for (i = 0; i < env.nr; i++)
 480                 childenv[i] = env.items[i].util;
 481         childenv[env.nr] = NULL;
 482
 483         string_list_clear(&env, 0);
 484         strbuf_release(&key);
 485         return childenv;
 486 }
 487
 488 struct atfork_state {
 489 #ifndef NO_PTHREADS
 490         int cs;
 491 #endif
 492         sigset_t old;
 493 };
 494
 495 #define CHECK_BUG(err, msg) \
 496         do { \
 497                 int e = (err); \
 498                 if (e) \
 499                         BUG("%s: %s", msg, strerror(e)); \
 500         } while(0)
 501
 502 static void atfork_prepare(struct atfork_state *as)
 503 {
 504         sigset_t all;
 505
 506         if (sigfillset(&all))
 507                 die_errno("sigfillset");
 508 #ifdef NO_PTHREADS
 509         if (sigprocmask(SIG_SETMASK, &all, &as->old))
 510                 die_errno("sigprocmask");
 511 #else
 512         CHECK_BUG(pthread_sigmask(SIG_SETMASK, &all, &as->old),
 513                 "blocking all signals");
 514         CHECK_BUG(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &as->cs),
 515                 "disabling cancellation");
 516 #endif
 517 }
 518
 519 static void atfork_parent(struct atfork_state *as)
 520 {
 521 #ifdef NO_PTHREADS
 522         if (sigprocmask(SIG_SETMASK, &as->old, NULL))
 523                 die_errno("sigprocmask");
 524 #else
 525         CHECK_BUG(pthread_setcancelstate(as->cs, NULL),
 526                 "re-enabling cancellation");
 527         CHECK_BUG(pthread_sigmask(SIG_SETMASK, &as->old, NULL),
 528                 "restoring signal mask");
 529 #endif
 530 }
 531 #endif /* GIT_WINDOWS_NATIVE */
 532
 533 static inline void set_cloexec(int fd)
 534 {
 535         int flags = fcntl(fd, F_GETFD);
 536         if (flags >= 0)
 537                 fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
 538 }
 539
 540 static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
 541 {
 542         int status, code = -1;
 543         pid_t waiting;
 544         int failed_errno = 0;
 545
 546         while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
 547                 ;       /* nothing */
 548
 549         if (waiting < 0) {
 550                 failed_errno = errno;
 551                 if (!in_signal)
 552                         error_errno("waitpid for %s failed", argv0);
 553         } else if (waiting != pid) {
 554                 if (!in_signal)
 555                         error("waitpid is confused (%s)", argv0);
 556         } else if (WIFSIGNALED(status)) {
 557                 code = WTERMSIG(status);
 558                 if (!in_signal && code != SIGINT && code != SIGQUIT && code != SIGPIPE)
 559                         error("%s died of signal %d", argv0, code);
 560                 /*
 561                  * This return value is chosen so that code & 0xff
 562                  * mimics the exit code that a POSIX shell would report for
 563                  * a program that died from this signal.
 564                  */
 565                 code += 128;
 566         } else if (WIFEXITED(status)) {
 567                 code = WEXITSTATUS(status);
 568         } else {
 569                 if (!in_signal)
 570                         error("waitpid is confused (%s)", argv0);
 571         }
 572
 573         if (!in_signal)
 574                 clear_child_for_cleanup(pid);
 575
 576         errno = failed_errno;
 577         return code;
 578 }
 579
 580 static void trace_add_env(struct strbuf *dst, const char *const *deltaenv)
 581 {
 582         struct string_list envs = STRING_LIST_INIT_DUP;
 583         const char *const *e;
 584         int i;
 585         int printed_unset = 0;
 586
 587         /* Last one wins, see run-command.c:prep_childenv() for context */
 588         for (e = deltaenv; e && *e; e++) {
 589                 struct strbuf key = STRBUF_INIT;
 590                 char *equals = strchr(*e, '=');
 591
 592                 if (equals) {
 593                         strbuf_add(&key, *e, equals - *e);
 594                         string_list_insert(&envs, key.buf)->util = equals + 1;
 595                 } else {
 596                         string_list_insert(&envs, *e)->util = NULL;
 597                 }
 598                 strbuf_release(&key);
 599         }
 600
 601         /* "unset X Y...;" */
 602         for (i = 0; i < envs.nr; i++) {
 603                 const char *var = envs.items[i].string;
 604                 const char *val = envs.items[i].util;
 605
 606                 if (val || !getenv(var))
 607                         continue;
 608
 609                 if (!printed_unset) {
 610                         strbuf_addstr(dst, " unset");
 611                         printed_unset = 1;
 612                 }
 613                 strbuf_addf(dst, " %s", var);
 614         }
 615         if (printed_unset)
 616                 strbuf_addch(dst, ';');
 617
 618         /* ... followed by "A=B C=D ..." */
 619         for (i = 0; i < envs.nr; i++) {
 620                 const char *var = envs.items[i].string;
 621                 const char *val = envs.items[i].util;
 622                 const char *oldval;
 623
 624                 if (!val)
 625                         continue;
 626
 627                 oldval = getenv(var);
 628                 if (oldval && !strcmp(val, oldval))
 629                         continue;
 630
 631                 strbuf_addf(dst, " %s=", var);
 632                 sq_quote_buf_pretty(dst, val);
 633         }
 634         string_list_clear(&envs, 0);
 635 }
 636
 637 static void trace_run_command(const struct child_process *cp)
 638 {
 639         struct strbuf buf = STRBUF_INIT;
 640
 641         if (!trace_want(&trace_default_key))
 642                 return;
 643
 644         strbuf_addstr(&buf, "trace: run_command:");
 645         if (cp->dir) {
 646                 strbuf_addstr(&buf, " cd ");
 647                 sq_quote_buf_pretty(&buf, cp->dir);
 648                 strbuf_addch(&buf, ';');
 649         }
 650         trace_add_env(&buf, cp->env.v);
 651         if (cp->git_cmd)
 652                 strbuf_addstr(&buf, " git");
 653         sq_quote_argv_pretty(&buf, cp->args.v);
 654
 655         trace_printf("%s", buf.buf);
 656         strbuf_release(&buf);
 657 }
 658
 659 int start_command(struct child_process *cmd)
 660 {
 661         int need_in, need_out, need_err;
 662         int fdin[2], fdout[2], fderr[2];
 663         int failed_errno;
 664         char *str;
 665
 666         /*
 667          * In case of errors we must keep the promise to close FDs
 668          * that have been passed in via ->in and ->out.
 669          */
 670
 671         need_in = !cmd->no_stdin && cmd->in < 0;
 672         if (need_in) {
 673                 if (pipe(fdin) < 0) {
 674                         failed_errno = errno;
 675                         if (cmd->out > 0)
 676                                 close(cmd->out);
 677                         str = "standard input";
 678                         goto fail_pipe;
 679                 }
 680                 cmd->in = fdin[1];
 681         }
 682
 683         need_out = !cmd->no_stdout
 684                 && !cmd->stdout_to_stderr
 685                 && cmd->out < 0;
 686         if (need_out) {
 687                 if (pipe(fdout) < 0) {
 688                         failed_errno = errno;
 689                         if (need_in)
 690                                 close_pair(fdin);
 691                         else if (cmd->in)
 692                                 close(cmd->in);
 693                         str = "standard output";
 694                         goto fail_pipe;
 695                 }
 696                 cmd->out = fdout[0];
 697         }
 698
 699         need_err = !cmd->no_stderr && cmd->err < 0;
 700         if (need_err) {
 701                 if (pipe(fderr) < 0) {
 702                         failed_errno = errno;
 703                         if (need_in)
 704                                 close_pair(fdin);
 705                         else if (cmd->in)
 706                                 close(cmd->in);
 707                         if (need_out)
 708                                 close_pair(fdout);
 709                         else if (cmd->out)
 710                                 close(cmd->out);
 711                         str = "standard error";
 712 fail_pipe:
 713                         error("cannot create %s pipe for %s: %s",
 714                                 str, cmd->args.v[0], strerror(failed_errno));
 715                         child_process_clear(cmd);
 716                         errno = failed_errno;
 717                         return -1;
 718                 }
 719                 cmd->err = fderr[0];
 720         }
 721
 722         trace2_child_start(cmd);
 723         trace_run_command(cmd);
 724
 725         fflush(NULL);
 726
 727         if (cmd->close_object_store)
 728                 close_object_store(the_repository->objects);
 729
 730 #ifndef GIT_WINDOWS_NATIVE
 731 {
 732         int notify_pipe[2];
 733         int null_fd = -1;
 734         char **childenv;
 735         struct strvec argv = STRVEC_INIT;
 736         struct child_err cerr;
 737         struct atfork_state as;
 738
 739         if (prepare_cmd(&argv, cmd) < 0) {
 740                 failed_errno = errno;
 741                 cmd->pid = -1;
 742                 if (!cmd->silent_exec_failure)
 743                         error_errno("cannot run %s", cmd->args.v[0]);
 744                 goto end_of_spawn;
 745         }
 746
 747         if (pipe(notify_pipe))
 748                 notify_pipe[0] = notify_pipe[1] = -1;
 749
 750         if (cmd->no_stdin || cmd->no_stdout || cmd->no_stderr) {
 751                 null_fd = xopen("/dev/null", O_RDWR | O_CLOEXEC);
 752                 set_cloexec(null_fd);
 753         }
 754
 755         childenv = prep_childenv(cmd->env.v);
 756         atfork_prepare(&as);
 757
 758         /*
 759          * NOTE: In order to prevent deadlocking when using threads special
 760          * care should be taken with the function calls made in between the
 761          * fork() and exec() calls.  No calls should be made to functions which
 762          * require acquiring a lock (e.g. malloc) as the lock could have been
 763          * held by another thread at the time of forking, causing the lock to
 764          * never be released in the child process.  This means only
 765          * Async-Signal-Safe functions are permitted in the child.
 766          */
 767         cmd->pid = fork();
 768         failed_errno = errno;
 769         if (!cmd->pid) {
 770                 int sig;
 771                 /*
 772                  * Ensure the default die/error/warn routines do not get
 773                  * called, they can take stdio locks and malloc.
 774                  */
 775                 set_die_routine(child_die_fn);
 776                 set_error_routine(child_error_fn);
 777                 set_warn_routine(child_warn_fn);
 778
 779                 close(notify_pipe[0]);
 780                 set_cloexec(notify_pipe[1]);
 781                 child_notifier = notify_pipe[1];
 782
 783                 if (cmd->no_stdin)
 784                         child_dup2(null_fd, 0);
 785                 else if (need_in) {
 786                         child_dup2(fdin[0], 0);
 787                         child_close_pair(fdin);
 788                 } else if (cmd->in) {
 789                         child_dup2(cmd->in, 0);
 790                         child_close(cmd->in);
 791                 }
 792
 793                 if (cmd->no_stderr)
 794                         child_dup2(null_fd, 2);
 795                 else if (need_err) {
 796                         child_dup2(fderr[1], 2);
 797                         child_close_pair(fderr);
 798                 } else if (cmd->err > 1) {
 799                         child_dup2(cmd->err, 2);
 800                         child_close(cmd->err);
 801                 }
 802
 803                 if (cmd->no_stdout)
 804                         child_dup2(null_fd, 1);
 805                 else if (cmd->stdout_to_stderr)
 806                         child_dup2(2, 1);
 807                 else if (need_out) {
 808                         child_dup2(fdout[1], 1);
 809                         child_close_pair(fdout);
 810                 } else if (cmd->out > 1) {
 811                         child_dup2(cmd->out, 1);
 812                         child_close(cmd->out);
 813                 }
 814
 815                 if (cmd->dir && chdir(cmd->dir))
 816                         child_die(CHILD_ERR_CHDIR);
 817
 818                 /*
 819                  * restore default signal handlers here, in case
 820                  * we catch a signal right before execve below
 821                  */
 822                 for (sig = 1; sig < NSIG; sig++) {
 823                         /* ignored signals get reset to SIG_DFL on execve */
 824                         if (signal(sig, SIG_DFL) == SIG_IGN)
 825                                 signal(sig, SIG_IGN);
 826                 }
 827
 828                 if (sigprocmask(SIG_SETMASK, &as.old, NULL) != 0)
 829                         child_die(CHILD_ERR_SIGPROCMASK);
 830
 831                 /*
 832                  * Attempt to exec using the command and arguments starting at
 833                  * argv.argv[1].  argv.argv[0] contains SHELL_PATH which will
 834                  * be used in the event exec failed with ENOEXEC at which point
 835                  * we will try to interpret the command using 'sh'.
 836                  */
 837                 execve(argv.v[1], (char *const *) argv.v + 1,
 838                        (char *const *) childenv);
 839                 if (errno == ENOEXEC)
 840                         execve(argv.v[0], (char *const *) argv.v,
 841                                (char *const *) childenv);
 842
 843                 if (errno == ENOENT) {
 844                         if (cmd->silent_exec_failure)
 845                                 child_die(CHILD_ERR_SILENT);
 846                         child_die(CHILD_ERR_ENOENT);
 847                 } else {
 848                         child_die(CHILD_ERR_ERRNO);
 849                 }
 850         }
 851         atfork_parent(&as);
 852         if (cmd->pid < 0)
 853                 error_errno("cannot fork() for %s", cmd->args.v[0]);
 854         else if (cmd->clean_on_exit)
 855                 mark_child_for_cleanup(cmd->pid, cmd);
 856
 857         /*
 858          * Wait for child's exec. If the exec succeeds (or if fork()
 859          * failed), EOF is seen immediately by the parent. Otherwise, the
 860          * child process sends a child_err struct.
 861          * Note that use of this infrastructure is completely advisory,
 862          * therefore, we keep error checks minimal.
 863          */
 864         close(notify_pipe[1]);
 865         if (xread(notify_pipe[0], &cerr, sizeof(cerr)) == sizeof(cerr)) {
 866                 /*
 867                  * At this point we know that fork() succeeded, but exec()
 868                  * failed. Errors have been reported to our stderr.
 869                  */
 870                 wait_or_whine(cmd->pid, cmd->args.v[0], 0);
 871                 child_err_spew(cmd, &cerr);
 872                 failed_errno = errno;
 873                 cmd->pid = -1;
 874         }
 875         close(notify_pipe[0]);
 876
 877         if (null_fd >= 0)
 878                 close(null_fd);
 879         strvec_clear(&argv);
 880         free(childenv);
 881 }
 882 end_of_spawn:
 883
 884 #else
 885 {
 886         int fhin = 0, fhout = 1, fherr = 2;
 887         const char **sargv = cmd->args.v;
 888         struct strvec nargv = STRVEC_INIT;
 889
 890         if (cmd->no_stdin)
 891                 fhin = open("/dev/null", O_RDWR);
 892         else if (need_in)
 893                 fhin = dup(fdin[0]);
 894         else if (cmd->in)
 895                 fhin = dup(cmd->in);
 896
 897         if (cmd->no_stderr)
 898                 fherr = open("/dev/null", O_RDWR);
 899         else if (need_err)
 900                 fherr = dup(fderr[1]);
 901         else if (cmd->err > 2)
 902                 fherr = dup(cmd->err);
 903
 904         if (cmd->no_stdout)
 905                 fhout = open("/dev/null", O_RDWR);
 906         else if (cmd->stdout_to_stderr)
 907                 fhout = dup(fherr);
 908         else if (need_out)
 909                 fhout = dup(fdout[1]);
 910         else if (cmd->out > 1)
 911                 fhout = dup(cmd->out);
 912
 913         if (cmd->git_cmd)
 914                 cmd->args.v = prepare_git_cmd(&nargv, sargv);
 915         else if (cmd->use_shell)
 916                 cmd->args.v = prepare_shell_cmd(&nargv, sargv);
 917
 918         cmd->pid = mingw_spawnvpe(cmd->args.v[0], cmd->args.v,
 919                                   (char**) cmd->env.v,
 920                                   cmd->dir, fhin, fhout, fherr);
 921         failed_errno = errno;
 922         if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
 923                 error_errno("cannot spawn %s", cmd->args.v[0]);
 924         if (cmd->clean_on_exit && cmd->pid >= 0)
 925                 mark_child_for_cleanup(cmd->pid, cmd);
 926
 927         strvec_clear(&nargv);
 928         cmd->args.v = sargv;
 929         if (fhin != 0)
 930                 close(fhin);
 931         if (fhout != 1)
 932                 close(fhout);
 933         if (fherr != 2)
 934                 close(fherr);
 935 }
 936 #endif
 937
 938         if (cmd->pid < 0) {
 939                 trace2_child_exit(cmd, -1);
 940
 941                 if (need_in)
 942                         close_pair(fdin);
 943                 else if (cmd->in)
 944                         close(cmd->in);
 945                 if (need_out)
 946                         close_pair(fdout);
 947                 else if (cmd->out)
 948                         close(cmd->out);
 949                 if (need_err)
 950                         close_pair(fderr);
 951                 else if (cmd->err)
 952                         close(cmd->err);
 953                 child_process_clear(cmd);
 954                 errno = failed_errno;
 955                 return -1;
 956         }
 957
 958         if (need_in)
 959                 close(fdin[0]);
 960         else if (cmd->in)
 961                 close(cmd->in);
 962
 963         if (need_out)
 964                 close(fdout[1]);
 965         else if (cmd->out)
 966                 close(cmd->out);
 967
 968         if (need_err)
 969                 close(fderr[1]);
 970         else if (cmd->err)
 971                 close(cmd->err);
 972
 973         return 0;
 974 }
 975
 976 int finish_command(struct child_process *cmd)
 977 {
 978         int ret = wait_or_whine(cmd->pid, cmd->args.v[0], 0);
 979         trace2_child_exit(cmd, ret);
 980         child_process_clear(cmd);
 981         invalidate_lstat_cache();
 982         return ret;
 983 }
 984
 985 int finish_command_in_signal(struct child_process *cmd)
 986 {
 987         int ret = wait_or_whine(cmd->pid, cmd->args.v[0], 1);
 988         if (ret != -1)
 989                 trace2_child_exit(cmd, ret);
 990         return ret;
 991 }
 992
 993
 994 int run_command(struct child_process *cmd)
 995 {
 996         int code;
 997
 998         if (cmd->out < 0 || cmd->err < 0)
 999                 BUG("run_command with a pipe can cause deadlock");
1000
1001         code = start_command(cmd);
1002         if (code)
1003                 return code;
1004         return finish_command(cmd);
1005 }
1006
1007 int run_command_v_opt(const char **argv, int opt)
1008 {
1009         return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
1010 }
1011
1012 int run_command_v_opt_tr2(const char **argv, int opt, const char *tr2_class)
1013 {
1014         return run_command_v_opt_cd_env_tr2(argv, opt, NULL, NULL, tr2_class);
1015 }
1016
1017 int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
1018 {
1019         return run_command_v_opt_cd_env_tr2(argv, opt, dir, env, NULL);
1020 }
1021
1022 int run_command_v_opt_cd_env_tr2(const char **argv, int opt, const char *dir,
1023                                  const char *const *env, const char *tr2_class)
1024 {
1025         struct child_process cmd = CHILD_PROCESS_INIT;
1026         strvec_pushv(&cmd.args, argv);
1027         cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
1028         cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
1029         cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
1030         cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
1031         cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
1032         cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
1033         cmd.wait_after_clean = opt & RUN_WAIT_AFTER_CLEAN ? 1 : 0;
1034         cmd.close_object_store = opt & RUN_CLOSE_OBJECT_STORE ? 1 : 0;
1035         cmd.dir = dir;
1036         if (env)
1037                 strvec_pushv(&cmd.env, (const char **)env);
1038         cmd.trace2_child_class = tr2_class;
1039         return run_command(&cmd);
1040 }
1041
1042 #ifndef NO_PTHREADS
1043 static pthread_t main_thread;
1044 static int main_thread_set;
1045 static pthread_key_t async_key;
1046 static pthread_key_t async_die_counter;
1047
1048 static void *run_thread(void *data)
1049 {
1050         struct async *async = data;
1051         intptr_t ret;
1052
1053         if (async->isolate_sigpipe) {
1054                 sigset_t mask;
1055                 sigemptyset(&mask);
1056                 sigaddset(&mask, SIGPIPE);
1057                 if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
1058                         ret = error("unable to block SIGPIPE in async thread");
1059                         return (void *)ret;
1060                 }
1061         }
1062
1063         pthread_setspecific(async_key, async);
1064         ret = async->proc(async->proc_in, async->proc_out, async->data);
1065         return (void *)ret;
1066 }
1067
1068 static NORETURN void die_async(const char *err, va_list params)
1069 {
1070         report_fn die_message_fn = get_die_message_routine();
1071
1072         die_message_fn(err, params);
1073
1074         if (in_async()) {
1075                 struct async *async = pthread_getspecific(async_key);
1076                 if (async->proc_in >= 0)
1077                         close(async->proc_in);
1078                 if (async->proc_out >= 0)
1079                         close(async->proc_out);
1080                 pthread_exit((void *)128);
1081         }
1082
1083         exit(128);
1084 }
1085
1086 static int async_die_is_recursing(void)
1087 {
1088         void *ret = pthread_getspecific(async_die_counter);
1089         pthread_setspecific(async_die_counter, &async_die_counter); /* set to any non-NULL valid pointer */
1090         return ret != NULL;
1091 }
1092
1093 int in_async(void)
1094 {
1095         if (!main_thread_set)
1096                 return 0; /* no asyncs started yet */
1097         return !pthread_equal(main_thread, pthread_self());
1098 }
1099
1100 static void NORETURN async_exit(int code)
1101 {
1102         pthread_exit((void *)(intptr_t)code);
1103 }
1104
1105 #else
1106
1107 static struct {
1108         void (**handlers)(void);
1109         size_t nr;
1110         size_t alloc;
1111 } git_atexit_hdlrs;
1112
1113 static int git_atexit_installed;
1114
1115 static void git_atexit_dispatch(void)
1116 {
1117         size_t i;
1118
1119         for (i=git_atexit_hdlrs.nr ; i ; i--)
1120                 git_atexit_hdlrs.handlers[i-1]();
1121 }
1122
1123 static void git_atexit_clear(void)
1124 {
1125         free(git_atexit_hdlrs.handlers);
1126         memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
1127         git_atexit_installed = 0;
1128 }
1129
1130 #undef atexit
1131 int git_atexit(void (*handler)(void))
1132 {
1133         ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
1134         git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
1135         if (!git_atexit_installed) {
1136                 if (atexit(&git_atexit_dispatch))
1137                         return -1;
1138                 git_atexit_installed = 1;
1139         }
1140         return 0;
1141 }
1142 #define atexit git_atexit
1143
1144 static int process_is_async;
1145 int in_async(void)
1146 {
1147         return process_is_async;
1148 }
1149
1150 static void NORETURN async_exit(int code)
1151 {
1152         exit(code);
1153 }
1154
1155 #endif
1156
1157 void check_pipe(int err)
1158 {
1159         if (err == EPIPE) {
1160                 if (in_async())
1161                         async_exit(141);
1162
1163                 signal(SIGPIPE, SIG_DFL);
1164                 raise(SIGPIPE);
1165                 /* Should never happen, but just in case... */
1166                 exit(141);
1167         }
1168 }
1169
1170 int start_async(struct async *async)
1171 {
1172         int need_in, need_out;
1173         int fdin[2], fdout[2];
1174         int proc_in, proc_out;
1175
1176         need_in = async->in < 0;
1177         if (need_in) {
1178                 if (pipe(fdin) < 0) {
1179                         if (async->out > 0)
1180                                 close(async->out);
1181                         return error_errno("cannot create pipe");
1182                 }
1183                 async->in = fdin[1];
1184         }
1185
1186         need_out = async->out < 0;
1187         if (need_out) {
1188                 if (pipe(fdout) < 0) {
1189                         if (need_in)
1190                                 close_pair(fdin);
1191                         else if (async->in)
1192                                 close(async->in);
1193                         return error_errno("cannot create pipe");
1194                 }
1195                 async->out = fdout[0];
1196         }
1197
1198         if (need_in)
1199                 proc_in = fdin[0];
1200         else if (async->in)
1201                 proc_in = async->in;
1202         else
1203                 proc_in = -1;
1204
1205         if (need_out)
1206                 proc_out = fdout[1];
1207         else if (async->out)
1208                 proc_out = async->out;
1209         else
1210                 proc_out = -1;
1211
1212 #ifdef NO_PTHREADS
1213         /* Flush stdio before fork() to avoid cloning buffers */
1214         fflush(NULL);
1215
1216         async->pid = fork();
1217         if (async->pid < 0) {
1218                 error_errno("fork (async) failed");
1219                 goto error;
1220         }
1221         if (!async->pid) {
1222                 if (need_in)
1223                         close(fdin[1]);
1224                 if (need_out)
1225                         close(fdout[0]);
1226                 git_atexit_clear();
1227                 process_is_async = 1;
1228                 exit(!!async->proc(proc_in, proc_out, async->data));
1229         }
1230
1231         mark_child_for_cleanup(async->pid, NULL);
1232
1233         if (need_in)
1234                 close(fdin[0]);
1235         else if (async->in)
1236                 close(async->in);
1237
1238         if (need_out)
1239                 close(fdout[1]);
1240         else if (async->out)
1241                 close(async->out);
1242 #else
1243         if (!main_thread_set) {
1244                 /*
1245                  * We assume that the first time that start_async is called
1246                  * it is from the main thread.
1247                  */
1248                 main_thread_set = 1;
1249                 main_thread = pthread_self();
1250                 pthread_key_create(&async_key, NULL);
1251                 pthread_key_create(&async_die_counter, NULL);
1252                 set_die_routine(die_async);
1253                 set_die_is_recursing_routine(async_die_is_recursing);
1254         }
1255
1256         if (proc_in >= 0)
1257                 set_cloexec(proc_in);
1258         if (proc_out >= 0)
1259                 set_cloexec(proc_out);
1260         async->proc_in = proc_in;
1261         async->proc_out = proc_out;
1262         {
1263                 int err = pthread_create(&async->tid, NULL, run_thread, async);
1264                 if (err) {
1265                         error(_("cannot create async thread: %s"), strerror(err));
1266                         goto error;
1267                 }
1268         }
1269 #endif
1270         return 0;
1271
1272 error:
1273         if (need_in)
1274                 close_pair(fdin);
1275         else if (async->in)
1276                 close(async->in);
1277
1278         if (need_out)
1279                 close_pair(fdout);
1280         else if (async->out)
1281                 close(async->out);
1282         return -1;
1283 }
1284
1285 int finish_async(struct async *async)
1286 {
1287 #ifdef NO_PTHREADS
1288         int ret = wait_or_whine(async->pid, "child process", 0);
1289
1290         invalidate_lstat_cache();
1291
1292         return ret;
1293 #else
1294         void *ret = (void *)(intptr_t)(-1);
1295
1296         if (pthread_join(async->tid, &ret))
1297                 error("pthread_join failed");
1298         invalidate_lstat_cache();
1299         return (int)(intptr_t)ret;
1300
1301 #endif
1302 }
1303
1304 int async_with_fork(void)
1305 {
1306 #ifdef NO_PTHREADS
1307         return 1;
1308 #else
1309         return 0;
1310 #endif
1311 }
1312
1313 struct io_pump {
1314         /* initialized by caller */
1315         int fd;
1316         int type; /* POLLOUT or POLLIN */
1317         union {
1318                 struct {
1319                         const char *buf;
1320                         size_t len;
1321                 } out;
1322                 struct {
1323                         struct strbuf *buf;
1324                         size_t hint;
1325                 } in;
1326         } u;
1327
1328         /* returned by pump_io */
1329         int error; /* 0 for success, otherwise errno */
1330
1331         /* internal use */
1332         struct pollfd *pfd;
1333 };
1334
1335 static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
1336 {
1337         int pollsize = 0;
1338         int i;
1339
1340         for (i = 0; i < nr; i++) {
1341                 struct io_pump *io = &slots[i];
1342                 if (io->fd < 0)
1343                         continue;
1344                 pfd[pollsize].fd = io->fd;
1345                 pfd[pollsize].events = io->type;
1346                 io->pfd = &pfd[pollsize++];
1347         }
1348
1349         if (!pollsize)
1350                 return 0;
1351
1352         if (poll(pfd, pollsize, -1) < 0) {
1353                 if (errno == EINTR)
1354                         return 1;
1355                 die_errno("poll failed");
1356         }
1357
1358         for (i = 0; i < nr; i++) {
1359                 struct io_pump *io = &slots[i];
1360
1361                 if (io->fd < 0)
1362                         continue;
1363
1364                 if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
1365                         continue;
1366
1367                 if (io->type == POLLOUT) {
1368                         ssize_t len;
1369
1370                         /*
1371                          * Don't use xwrite() here. It loops forever on EAGAIN,
1372                          * and we're in our own poll() loop here.
1373                          *
1374                          * Note that we lose xwrite()'s handling of MAX_IO_SIZE
1375                          * and EINTR, so we have to implement those ourselves.
1376                          */
1377                         len = write(io->fd, io->u.out.buf,
1378                                     io->u.out.len <= MAX_IO_SIZE ?
1379                                     io->u.out.len : MAX_IO_SIZE);
1380                         if (len < 0) {
1381                                 if (errno != EINTR && errno != EAGAIN &&
1382                                     errno != ENOSPC) {
1383                                         io->error = errno;
1384                                         close(io->fd);
1385                                         io->fd = -1;
1386                                 }
1387                         } else {
1388                                 io->u.out.buf += len;
1389                                 io->u.out.len -= len;
1390                                 if (!io->u.out.len) {
1391                                         close(io->fd);
1392                                         io->fd = -1;
1393                                 }
1394                         }
1395                 }
1396
1397                 if (io->type == POLLIN) {
1398                         ssize_t len = strbuf_read_once(io->u.in.buf,
1399                                                        io->fd, io->u.in.hint);
1400                         if (len < 0)
1401                                 io->error = errno;
1402                         if (len <= 0) {
1403                                 close(io->fd);
1404                                 io->fd = -1;
1405                         }
1406                 }
1407         }
1408
1409         return 1;
1410 }
1411
1412 static int pump_io(struct io_pump *slots, int nr)
1413 {
1414         struct pollfd *pfd;
1415         int i;
1416
1417         for (i = 0; i < nr; i++)
1418                 slots[i].error = 0;
1419
1420         ALLOC_ARRAY(pfd, nr);
1421         while (pump_io_round(slots, nr, pfd))
1422                 ; /* nothing */
1423         free(pfd);
1424
1425         /* There may be multiple errno values, so just pick the first. */
1426         for (i = 0; i < nr; i++) {
1427                 if (slots[i].error) {
1428                         errno = slots[i].error;
1429                         return -1;
1430                 }
1431         }
1432         return 0;
1433 }
1434
1435
1436 int pipe_command(struct child_process *cmd,
1437                  const char *in, size_t in_len,
1438                  struct strbuf *out, size_t out_hint,
1439                  struct strbuf *err, size_t err_hint)
1440 {
1441         struct io_pump io[3];
1442         int nr = 0;
1443
1444         if (in)
1445                 cmd->in = -1;
1446         if (out)
1447                 cmd->out = -1;
1448         if (err)
1449                 cmd->err = -1;
1450
1451         if (start_command(cmd) < 0)
1452                 return -1;
1453
1454         if (in) {
1455                 if (enable_pipe_nonblock(cmd->in) < 0) {
1456                         error_errno("unable to make pipe non-blocking");
1457                         close(cmd->in);
1458                         if (out)
1459                                 close(cmd->out);
1460                         if (err)
1461                                 close(cmd->err);
1462                         return -1;
1463                 }
1464                 io[nr].fd = cmd->in;
1465                 io[nr].type = POLLOUT;
1466                 io[nr].u.out.buf = in;
1467                 io[nr].u.out.len = in_len;
1468                 nr++;
1469         }
1470         if (out) {
1471                 io[nr].fd = cmd->out;
1472                 io[nr].type = POLLIN;
1473                 io[nr].u.in.buf = out;
1474                 io[nr].u.in.hint = out_hint;
1475                 nr++;
1476         }
1477         if (err) {
1478                 io[nr].fd = cmd->err;
1479                 io[nr].type = POLLIN;
1480                 io[nr].u.in.buf = err;
1481                 io[nr].u.in.hint = err_hint;
1482                 nr++;
1483         }
1484
1485         if (pump_io(io, nr) < 0) {
1486                 finish_command(cmd); /* throw away exit code */
1487                 return -1;
1488         }
1489
1490         return finish_command(cmd);
1491 }
1492
1493 enum child_state {
1494         GIT_CP_FREE,
1495         GIT_CP_WORKING,
1496         GIT_CP_WAIT_CLEANUP,
1497 };
1498
1499 int run_processes_parallel_ungroup;
1500 struct parallel_processes {
1501         void *data;
1502
1503         int max_processes;
1504         int nr_processes;
1505
1506         get_next_task_fn get_next_task;
1507         start_failure_fn start_failure;
1508         task_finished_fn task_finished;
1509
1510         struct {
1511                 enum child_state state;
1512                 struct child_process process;
1513                 struct strbuf err;
1514                 void *data;
1515         } *children;
1516         /*
1517          * The struct pollfd is logically part of *children,
1518          * but the system call expects it as its own array.
1519          */
1520         struct pollfd *pfd;
1521
1522         unsigned shutdown : 1;
1523         unsigned ungroup : 1;
1524
1525         int output_owner;
1526         struct strbuf buffered_output; /* of finished children */
1527 };
1528
1529 static int default_start_failure(struct strbuf *out,
1530                                  void *pp_cb,
1531                                  void *pp_task_cb)
1532 {
1533         return 0;
1534 }
1535
1536 static int default_task_finished(int result,
1537                                  struct strbuf *out,
1538                                  void *pp_cb,
1539                                  void *pp_task_cb)
1540 {
1541         return 0;
1542 }
1543
1544 static void kill_children(struct parallel_processes *pp, int signo)
1545 {
1546         int i, n = pp->max_processes;
1547
1548         for (i = 0; i < n; i++)
1549                 if (pp->children[i].state == GIT_CP_WORKING)
1550                         kill(pp->children[i].process.pid, signo);
1551 }
1552
1553 static struct parallel_processes *pp_for_signal;
1554
1555 static void handle_children_on_signal(int signo)
1556 {
1557         kill_children(pp_for_signal, signo);
1558         sigchain_pop(signo);
1559         raise(signo);
1560 }
1561
1562 static void pp_init(struct parallel_processes *pp,
1563                     int n,
1564                     get_next_task_fn get_next_task,
1565                     start_failure_fn start_failure,
1566                     task_finished_fn task_finished,
1567                     void *data, int ungroup)
1568 {
1569         int i;
1570
1571         if (n < 1)
1572                 n = online_cpus();
1573
1574         pp->max_processes = n;
1575
1576         trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
1577
1578         pp->data = data;
1579         if (!get_next_task)
1580                 BUG("you need to specify a get_next_task function");
1581         pp->get_next_task = get_next_task;
1582
1583         pp->start_failure = start_failure ? start_failure : default_start_failure;
1584         pp->task_finished = task_finished ? task_finished : default_task_finished;
1585
1586         pp->nr_processes = 0;
1587         pp->output_owner = 0;
1588         pp->shutdown = 0;
1589         pp->ungroup = ungroup;
1590         CALLOC_ARRAY(pp->children, n);
1591         if (pp->ungroup)
1592                 pp->pfd = NULL;
1593         else
1594                 CALLOC_ARRAY(pp->pfd, n);
1595         strbuf_init(&pp->buffered_output, 0);
1596
1597         for (i = 0; i < n; i++) {
1598                 strbuf_init(&pp->children[i].err, 0);
1599                 child_process_init(&pp->children[i].process);
1600                 if (pp->pfd) {
1601                         pp->pfd[i].events = POLLIN | POLLHUP;
1602                         pp->pfd[i].fd = -1;
1603                 }
1604         }
1605
1606         pp_for_signal = pp;
1607         sigchain_push_common(handle_children_on_signal);
1608 }
1609
1610 static void pp_cleanup(struct parallel_processes *pp)
1611 {
1612         int i;
1613
1614         trace_printf("run_processes_parallel: done");
1615         for (i = 0; i < pp->max_processes; i++) {
1616                 strbuf_release(&pp->children[i].err);
1617                 child_process_clear(&pp->children[i].process);
1618         }
1619
1620         free(pp->children);
1621         free(pp->pfd);
1622
1623         /*
1624          * When get_next_task added messages to the buffer in its last
1625          * iteration, the buffered output is non empty.
1626          */
1627         strbuf_write(&pp->buffered_output, stderr);
1628         strbuf_release(&pp->buffered_output);
1629
1630         sigchain_pop_common();
1631 }
1632
1633 /* returns
1634  *  0 if a new task was started.
1635  *  1 if no new jobs was started (get_next_task ran out of work, non critical
1636  *    problem with starting a new command)
1637  * <0 no new job was started, user wishes to shutdown early. Use negative code
1638  *    to signal the children.
1639  */
1640 static int pp_start_one(struct parallel_processes *pp)
1641 {
1642         int i, code;
1643
1644         for (i = 0; i < pp->max_processes; i++)
1645                 if (pp->children[i].state == GIT_CP_FREE)
1646                         break;
1647         if (i == pp->max_processes)
1648                 BUG("bookkeeping is hard");
1649
1650         code = pp->get_next_task(&pp->children[i].process,
1651                                  pp->ungroup ? NULL : &pp->children[i].err,
1652                                  pp->data,
1653                                  &pp->children[i].data);
1654         if (!code) {
1655                 if (!pp->ungroup) {
1656                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1657                         strbuf_reset(&pp->children[i].err);
1658                 }
1659                 return 1;
1660         }
1661         if (!pp->ungroup) {
1662                 pp->children[i].process.err = -1;
1663                 pp->children[i].process.stdout_to_stderr = 1;
1664         }
1665         pp->children[i].process.no_stdin = 1;
1666
1667         if (start_command(&pp->children[i].process)) {
1668                 code = pp->start_failure(pp->ungroup ? NULL :
1669                                          &pp->children[i].err,
1670                                          pp->data,
1671                                          pp->children[i].data);
1672                 if (!pp->ungroup) {
1673                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1674                         strbuf_reset(&pp->children[i].err);
1675                 }
1676                 if (code)
1677                         pp->shutdown = 1;
1678                 return code;
1679         }
1680
1681         pp->nr_processes++;
1682         pp->children[i].state = GIT_CP_WORKING;
1683         if (pp->pfd)
1684                 pp->pfd[i].fd = pp->children[i].process.err;
1685         return 0;
1686 }
1687
1688 static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1689 {
1690         int i;
1691
1692         while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1693                 if (errno == EINTR)
1694                         continue;
1695                 pp_cleanup(pp);
1696                 die_errno("poll");
1697         }
1698
1699         /* Buffer output from all pipes. */
1700         for (i = 0; i < pp->max_processes; i++) {
1701                 if (pp->children[i].state == GIT_CP_WORKING &&
1702                     pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1703                         int n = strbuf_read_once(&pp->children[i].err,
1704                                                  pp->children[i].process.err, 0);
1705                         if (n == 0) {
1706                                 close(pp->children[i].process.err);
1707                                 pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1708                         } else if (n < 0)
1709                                 if (errno != EAGAIN)
1710                                         die_errno("read");
1711                 }
1712         }
1713 }
1714
1715 static void pp_output(struct parallel_processes *pp)
1716 {
1717         int i = pp->output_owner;
1718
1719         if (pp->children[i].state == GIT_CP_WORKING &&
1720             pp->children[i].err.len) {
1721                 strbuf_write(&pp->children[i].err, stderr);
1722                 strbuf_reset(&pp->children[i].err);
1723         }
1724 }
1725
1726 static int pp_collect_finished(struct parallel_processes *pp)
1727 {
1728         int i, code;
1729         int n = pp->max_processes;
1730         int result = 0;
1731
1732         while (pp->nr_processes > 0) {
1733                 for (i = 0; i < pp->max_processes; i++)
1734                         if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1735                                 break;
1736                 if (i == pp->max_processes)
1737                         break;
1738
1739                 code = finish_command(&pp->children[i].process);
1740
1741                 code = pp->task_finished(code, pp->ungroup ? NULL :
1742                                          &pp->children[i].err, pp->data,
1743                                          pp->children[i].data);
1744
1745                 if (code)
1746                         result = code;
1747                 if (code < 0)
1748                         break;
1749
1750                 pp->nr_processes--;
1751                 pp->children[i].state = GIT_CP_FREE;
1752                 if (pp->pfd)
1753                         pp->pfd[i].fd = -1;
1754                 child_process_init(&pp->children[i].process);
1755
1756                 if (pp->ungroup) {
1757                         ; /* no strbuf_*() work to do here */
1758                 } else if (i != pp->output_owner) {
1759                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1760                         strbuf_reset(&pp->children[i].err);
1761                 } else {
1762                         strbuf_write(&pp->children[i].err, stderr);
1763                         strbuf_reset(&pp->children[i].err);
1764
1765                         /* Output all other finished child processes */
1766                         strbuf_write(&pp->buffered_output, stderr);
1767                         strbuf_reset(&pp->buffered_output);
1768
1769                         /*
1770                          * Pick next process to output live.
1771                          * NEEDSWORK:
1772                          * For now we pick it randomly by doing a round
1773                          * robin. Later we may want to pick the one with
1774                          * the most output or the longest or shortest
1775                          * running process time.
1776                          */
1777                         for (i = 0; i < n; i++)
1778                                 if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1779                                         break;
1780                         pp->output_owner = (pp->output_owner + i) % n;
1781                 }
1782         }
1783         return result;
1784 }
1785
1786 int run_processes_parallel(int n,
1787                            get_next_task_fn get_next_task,
1788                            start_failure_fn start_failure,
1789                            task_finished_fn task_finished,
1790                            void *pp_cb)
1791 {
1792         int i, code;
1793         int output_timeout = 100;
1794         int spawn_cap = 4;
1795         int ungroup = run_processes_parallel_ungroup;
1796         struct parallel_processes pp;
1797
1798         /* unset for the next API user */
1799         run_processes_parallel_ungroup = 0;
1800
1801         pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb,
1802                 ungroup);
1803         while (1) {
1804                 for (i = 0;
1805                     i < spawn_cap && !pp.shutdown &&
1806                     pp.nr_processes < pp.max_processes;
1807                     i++) {
1808                         code = pp_start_one(&pp);
1809                         if (!code)
1810                                 continue;
1811                         if (code < 0) {
1812                                 pp.shutdown = 1;
1813                                 kill_children(&pp, -code);
1814                         }
1815                         break;
1816                 }
1817                 if (!pp.nr_processes)
1818                         break;
1819                 if (ungroup) {
1820                         int i;
1821
1822                         for (i = 0; i < pp.max_processes; i++)
1823                                 pp.children[i].state = GIT_CP_WAIT_CLEANUP;
1824                 } else {
1825                         pp_buffer_stderr(&pp, output_timeout);
1826                         pp_output(&pp);
1827                 }
1828                 code = pp_collect_finished(&pp);
1829                 if (code) {
1830                         pp.shutdown = 1;
1831                         if (code < 0)
1832                                 kill_children(&pp, -code);
1833                 }
1834         }
1835
1836         pp_cleanup(&pp);
1837         return 0;
1838 }
1839
1840 int run_processes_parallel_tr2(int n, get_next_task_fn get_next_task,
1841                                start_failure_fn start_failure,
1842                                task_finished_fn task_finished, void *pp_cb,
1843                                const char *tr2_category, const char *tr2_label)
1844 {
1845         int result;
1846
1847         trace2_region_enter_printf(tr2_category, tr2_label, NULL, "max:%d",
1848                                    ((n < 1) ? online_cpus() : n));
1849
1850         result = run_processes_parallel(n, get_next_task, start_failure,
1851                                         task_finished, pp_cb);
1852
1853         trace2_region_leave(tr2_category, tr2_label, NULL);
1854
1855         return result;
1856 }
1857
1858 int run_auto_maintenance(int quiet)
1859 {
1860         int enabled;
1861         struct child_process maint = CHILD_PROCESS_INIT;
1862
1863         if (!git_config_get_bool("maintenance.auto", &enabled) &&
1864             !enabled)
1865                 return 0;
1866
1867         maint.git_cmd = 1;
1868         maint.close_object_store = 1;
1869         strvec_pushl(&maint.args, "maintenance", "run", "--auto", NULL);
1870         strvec_push(&maint.args, quiet ? "--quiet" : "--no-quiet");
1871
1872         return run_command(&maint);
1873 }
1874
1875 void prepare_other_repo_env(struct strvec *env, const char *new_git_dir)
1876 {
1877         const char * const *var;
1878
1879         for (var = local_repo_env; *var; var++) {
1880                 if (strcmp(*var, CONFIG_DATA_ENVIRONMENT) &&
1881                     strcmp(*var, CONFIG_COUNT_ENVIRONMENT))
1882                         strvec_push(env, *var);
1883         }
1884         strvec_pushf(env, "%s=%s", GIT_DIR_ENVIRONMENT, new_git_dir);
1885 }
1886
1887 enum start_bg_result start_bg_command(struct child_process *cmd,
1888                                       start_bg_wait_cb *wait_cb,
1889                                       void *cb_data,
1890                                       unsigned int timeout_sec)
1891 {
1892         enum start_bg_result sbgr = SBGR_ERROR;
1893         int ret;
1894         int wait_status;
1895         pid_t pid_seen;
1896         time_t time_limit;
1897
1898         /*
1899          * We do not allow clean-on-exit because the child process
1900          * should persist in the background and possibly/probably
1901          * after this process exits.  So we don't want to kill the
1902          * child during our atexit routine.
1903          */
1904         if (cmd->clean_on_exit)
1905                 BUG("start_bg_command() does not allow non-zero clean_on_exit");
1906
1907         if (!cmd->trace2_child_class)
1908                 cmd->trace2_child_class = "background";
1909
1910         ret = start_command(cmd);
1911         if (ret) {
1912                 /*
1913                  * We assume that if `start_command()` fails, we
1914                  * either get a complete `trace2_child_start() /
1915                  * trace2_child_exit()` pair or it fails before the
1916                  * `trace2_child_start()` is emitted, so we do not
1917                  * need to worry about it here.
1918                  *
1919                  * We also assume that `start_command()` does not add
1920                  * us to the cleanup list.  And that it calls
1921                  * calls `child_process_clear()`.
1922                  */
1923                 sbgr = SBGR_ERROR;
1924                 goto done;
1925         }
1926
1927         time(&time_limit);
1928         time_limit += timeout_sec;
1929
1930 wait:
1931         pid_seen = waitpid(cmd->pid, &wait_status, WNOHANG);
1932
1933         if (!pid_seen) {
1934                 /*
1935                  * The child is currently running.  Ask the callback
1936                  * if the child is ready to do work or whether we
1937                  * should keep waiting for it to boot up.
1938                  */
1939                 ret = (*wait_cb)(cmd, cb_data);
1940                 if (!ret) {
1941                         /*
1942                          * The child is running and "ready".
1943                          */
1944                         trace2_child_ready(cmd, "ready");
1945                         sbgr = SBGR_READY;
1946                         goto done;
1947                 } else if (ret > 0) {
1948                         /*
1949                          * The callback said to give it more time to boot up
1950                          * (subject to our timeout limit).
1951                          */
1952                         time_t now;
1953
1954                         time(&now);
1955                         if (now < time_limit)
1956                                 goto wait;
1957
1958                         /*
1959                          * Our timeout has expired.  We don't try to
1960                          * kill the child, but rather let it continue
1961                          * (hopefully) trying to startup.
1962                          */
1963                         trace2_child_ready(cmd, "timeout");
1964                         sbgr = SBGR_TIMEOUT;
1965                         goto done;
1966                 } else {
1967                         /*
1968                          * The cb gave up on this child.  It is still running,
1969                          * but our cb got an error trying to probe it.
1970                          */
1971                         trace2_child_ready(cmd, "error");
1972                         sbgr = SBGR_CB_ERROR;
1973                         goto done;
1974                 }
1975         }
1976
1977         else if (pid_seen == cmd->pid) {
1978                 int child_code = -1;
1979
1980                 /*
1981                  * The child started, but exited or was terminated
1982                  * before becoming "ready".
1983                  *
1984                  * We try to match the behavior of `wait_or_whine()`
1985                  * WRT the handling of WIFSIGNALED() and WIFEXITED()
1986                  * and convert the child's status to a return code for
1987                  * tracing purposes and emit the `trace2_child_exit()`
1988                  * event.
1989                  *
1990                  * We do not want the wait_or_whine() error message
1991                  * because we will be called by client-side library
1992                  * routines.
1993                  */
1994                 if (WIFEXITED(wait_status))
1995                         child_code = WEXITSTATUS(wait_status);
1996                 else if (WIFSIGNALED(wait_status))
1997                         child_code = WTERMSIG(wait_status) + 128;
1998                 trace2_child_exit(cmd, child_code);
1999
2000                 sbgr = SBGR_DIED;
2001                 goto done;
2002         }
2003
2004         else if (pid_seen < 0 && errno == EINTR)
2005                 goto wait;
2006
2007         trace2_child_exit(cmd, -1);
2008         sbgr = SBGR_ERROR;
2009
2010 done:
2011         child_process_clear(cmd);
2012         invalidate_lstat_cache();
2013         return sbgr;
2014 }