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