Merge branch 'jk/rev-parse-cleanup'
[git] / run-command.c
1 #include "cache.h"
2 #include "run-command.h"
3 #include "exec_cmd.h"
4 #include "sigchain.h"
5 #include "argv-array.h"
6 #include "thread-utils.h"
7 #include "strbuf.h"
8
9 void child_process_init(struct child_process *child)
10 {
11         memset(child, 0, sizeof(*child));
12         argv_array_init(&child->args);
13         argv_array_init(&child->env_array);
14 }
15
16 void child_process_clear(struct child_process *child)
17 {
18         argv_array_clear(&child->args);
19         argv_array_clear(&child->env_array);
20 }
21
22 struct child_to_clean {
23         pid_t pid;
24         struct child_process *process;
25         struct child_to_clean *next;
26 };
27 static struct child_to_clean *children_to_clean;
28 static int installed_child_cleanup_handler;
29
30 static void cleanup_children(int sig, int in_signal)
31 {
32         struct child_to_clean *children_to_wait_for = NULL;
33
34         while (children_to_clean) {
35                 struct child_to_clean *p = children_to_clean;
36                 children_to_clean = p->next;
37
38                 if (p->process && !in_signal) {
39                         struct child_process *process = p->process;
40                         if (process->clean_on_exit_handler) {
41                                 trace_printf(
42                                         "trace: run_command: running exit handler for pid %"
43                                         PRIuMAX, (uintmax_t)p->pid
44                                 );
45                                 process->clean_on_exit_handler(process);
46                         }
47                 }
48
49                 kill(p->pid, sig);
50
51                 if (p->process->wait_after_clean) {
52                         p->next = children_to_wait_for;
53                         children_to_wait_for = p;
54                 } else {
55                         if (!in_signal)
56                                 free(p);
57                 }
58         }
59
60         while (children_to_wait_for) {
61                 struct child_to_clean *p = children_to_wait_for;
62                 children_to_wait_for = p->next;
63
64                 while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
65                         ; /* spin waiting for process exit or error */
66
67                 if (!in_signal)
68                         free(p);
69         }
70 }
71
72 static void cleanup_children_on_signal(int sig)
73 {
74         cleanup_children(sig, 1);
75         sigchain_pop(sig);
76         raise(sig);
77 }
78
79 static void cleanup_children_on_exit(void)
80 {
81         cleanup_children(SIGTERM, 0);
82 }
83
84 static void mark_child_for_cleanup(pid_t pid, struct child_process *process)
85 {
86         struct child_to_clean *p = xmalloc(sizeof(*p));
87         p->pid = pid;
88         p->process = process;
89         p->next = children_to_clean;
90         children_to_clean = p;
91
92         if (!installed_child_cleanup_handler) {
93                 atexit(cleanup_children_on_exit);
94                 sigchain_push_common(cleanup_children_on_signal);
95                 installed_child_cleanup_handler = 1;
96         }
97 }
98
99 static void clear_child_for_cleanup(pid_t pid)
100 {
101         struct child_to_clean **pp;
102
103         for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
104                 struct child_to_clean *clean_me = *pp;
105
106                 if (clean_me->pid == pid) {
107                         *pp = clean_me->next;
108                         free(clean_me);
109                         return;
110                 }
111         }
112 }
113
114 static inline void close_pair(int fd[2])
115 {
116         close(fd[0]);
117         close(fd[1]);
118 }
119
120 #ifndef GIT_WINDOWS_NATIVE
121 static inline void dup_devnull(int to)
122 {
123         int fd = open("/dev/null", O_RDWR);
124         if (fd < 0)
125                 die_errno(_("open /dev/null failed"));
126         if (dup2(fd, to) < 0)
127                 die_errno(_("dup2(%d,%d) failed"), fd, to);
128         close(fd);
129 }
130 #endif
131
132 static char *locate_in_PATH(const char *file)
133 {
134         const char *p = getenv("PATH");
135         struct strbuf buf = STRBUF_INIT;
136
137         if (!p || !*p)
138                 return NULL;
139
140         while (1) {
141                 const char *end = strchrnul(p, ':');
142
143                 strbuf_reset(&buf);
144
145                 /* POSIX specifies an empty entry as the current directory. */
146                 if (end != p) {
147                         strbuf_add(&buf, p, end - p);
148                         strbuf_addch(&buf, '/');
149                 }
150                 strbuf_addstr(&buf, file);
151
152                 if (!access(buf.buf, F_OK))
153                         return strbuf_detach(&buf, NULL);
154
155                 if (!*end)
156                         break;
157                 p = end + 1;
158         }
159
160         strbuf_release(&buf);
161         return NULL;
162 }
163
164 static int exists_in_PATH(const char *file)
165 {
166         char *r = locate_in_PATH(file);
167         free(r);
168         return r != NULL;
169 }
170
171 int sane_execvp(const char *file, char * const argv[])
172 {
173         if (!execvp(file, argv))
174                 return 0; /* cannot happen ;-) */
175
176         /*
177          * When a command can't be found because one of the directories
178          * listed in $PATH is unsearchable, execvp reports EACCES, but
179          * careful usability testing (read: analysis of occasional bug
180          * reports) reveals that "No such file or directory" is more
181          * intuitive.
182          *
183          * We avoid commands with "/", because execvp will not do $PATH
184          * lookups in that case.
185          *
186          * The reassignment of EACCES to errno looks like a no-op below,
187          * but we need to protect against exists_in_PATH overwriting errno.
188          */
189         if (errno == EACCES && !strchr(file, '/'))
190                 errno = exists_in_PATH(file) ? EACCES : ENOENT;
191         else if (errno == ENOTDIR && !strchr(file, '/'))
192                 errno = ENOENT;
193         return -1;
194 }
195
196 static const char **prepare_shell_cmd(struct argv_array *out, const char **argv)
197 {
198         if (!argv[0])
199                 die("BUG: shell command is empty");
200
201         if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
202 #ifndef GIT_WINDOWS_NATIVE
203                 argv_array_push(out, SHELL_PATH);
204 #else
205                 argv_array_push(out, "sh");
206 #endif
207                 argv_array_push(out, "-c");
208
209                 /*
210                  * If we have no extra arguments, we do not even need to
211                  * bother with the "$@" magic.
212                  */
213                 if (!argv[1])
214                         argv_array_push(out, argv[0]);
215                 else
216                         argv_array_pushf(out, "%s \"$@\"", argv[0]);
217         }
218
219         argv_array_pushv(out, argv);
220         return out->argv;
221 }
222
223 #ifndef GIT_WINDOWS_NATIVE
224 static int execv_shell_cmd(const char **argv)
225 {
226         struct argv_array nargv = ARGV_ARRAY_INIT;
227         prepare_shell_cmd(&nargv, argv);
228         trace_argv_printf(nargv.argv, "trace: exec:");
229         sane_execvp(nargv.argv[0], (char **)nargv.argv);
230         argv_array_clear(&nargv);
231         return -1;
232 }
233 #endif
234
235 #ifndef GIT_WINDOWS_NATIVE
236 static int child_notifier = -1;
237
238 static void notify_parent(void)
239 {
240         /*
241          * execvp failed.  If possible, we'd like to let start_command
242          * know, so failures like ENOENT can be handled right away; but
243          * otherwise, finish_command will still report the error.
244          */
245         xwrite(child_notifier, "", 1);
246 }
247 #endif
248
249 static inline void set_cloexec(int fd)
250 {
251         int flags = fcntl(fd, F_GETFD);
252         if (flags >= 0)
253                 fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
254 }
255
256 static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
257 {
258         int status, code = -1;
259         pid_t waiting;
260         int failed_errno = 0;
261
262         while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
263                 ;       /* nothing */
264         if (in_signal)
265                 return 0;
266
267         if (waiting < 0) {
268                 failed_errno = errno;
269                 error_errno("waitpid for %s failed", argv0);
270         } else if (waiting != pid) {
271                 error("waitpid is confused (%s)", argv0);
272         } else if (WIFSIGNALED(status)) {
273                 code = WTERMSIG(status);
274                 if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
275                         error("%s died of signal %d", argv0, code);
276                 /*
277                  * This return value is chosen so that code & 0xff
278                  * mimics the exit code that a POSIX shell would report for
279                  * a program that died from this signal.
280                  */
281                 code += 128;
282         } else if (WIFEXITED(status)) {
283                 code = WEXITSTATUS(status);
284                 /*
285                  * Convert special exit code when execvp failed.
286                  */
287                 if (code == 127) {
288                         code = -1;
289                         failed_errno = ENOENT;
290                 }
291         } else {
292                 error("waitpid is confused (%s)", argv0);
293         }
294
295         clear_child_for_cleanup(pid);
296
297         errno = failed_errno;
298         return code;
299 }
300
301 int start_command(struct child_process *cmd)
302 {
303         int need_in, need_out, need_err;
304         int fdin[2], fdout[2], fderr[2];
305         int failed_errno;
306         char *str;
307
308         if (!cmd->argv)
309                 cmd->argv = cmd->args.argv;
310         if (!cmd->env)
311                 cmd->env = cmd->env_array.argv;
312
313         /*
314          * In case of errors we must keep the promise to close FDs
315          * that have been passed in via ->in and ->out.
316          */
317
318         need_in = !cmd->no_stdin && cmd->in < 0;
319         if (need_in) {
320                 if (pipe(fdin) < 0) {
321                         failed_errno = errno;
322                         if (cmd->out > 0)
323                                 close(cmd->out);
324                         str = "standard input";
325                         goto fail_pipe;
326                 }
327                 cmd->in = fdin[1];
328         }
329
330         need_out = !cmd->no_stdout
331                 && !cmd->stdout_to_stderr
332                 && cmd->out < 0;
333         if (need_out) {
334                 if (pipe(fdout) < 0) {
335                         failed_errno = errno;
336                         if (need_in)
337                                 close_pair(fdin);
338                         else if (cmd->in)
339                                 close(cmd->in);
340                         str = "standard output";
341                         goto fail_pipe;
342                 }
343                 cmd->out = fdout[0];
344         }
345
346         need_err = !cmd->no_stderr && cmd->err < 0;
347         if (need_err) {
348                 if (pipe(fderr) < 0) {
349                         failed_errno = errno;
350                         if (need_in)
351                                 close_pair(fdin);
352                         else if (cmd->in)
353                                 close(cmd->in);
354                         if (need_out)
355                                 close_pair(fdout);
356                         else if (cmd->out)
357                                 close(cmd->out);
358                         str = "standard error";
359 fail_pipe:
360                         error("cannot create %s pipe for %s: %s",
361                                 str, cmd->argv[0], strerror(failed_errno));
362                         child_process_clear(cmd);
363                         errno = failed_errno;
364                         return -1;
365                 }
366                 cmd->err = fderr[0];
367         }
368
369         trace_argv_printf(cmd->argv, "trace: run_command:");
370         fflush(NULL);
371
372 #ifndef GIT_WINDOWS_NATIVE
373 {
374         int notify_pipe[2];
375         if (pipe(notify_pipe))
376                 notify_pipe[0] = notify_pipe[1] = -1;
377
378         cmd->pid = fork();
379         failed_errno = errno;
380         if (!cmd->pid) {
381                 /*
382                  * Redirect the channel to write syscall error messages to
383                  * before redirecting the process's stderr so that all die()
384                  * in subsequent call paths use the parent's stderr.
385                  */
386                 if (cmd->no_stderr || need_err) {
387                         int child_err = dup(2);
388                         set_cloexec(child_err);
389                         set_error_handle(fdopen(child_err, "w"));
390                 }
391
392                 close(notify_pipe[0]);
393                 set_cloexec(notify_pipe[1]);
394                 child_notifier = notify_pipe[1];
395                 atexit(notify_parent);
396
397                 if (cmd->no_stdin)
398                         dup_devnull(0);
399                 else if (need_in) {
400                         dup2(fdin[0], 0);
401                         close_pair(fdin);
402                 } else if (cmd->in) {
403                         dup2(cmd->in, 0);
404                         close(cmd->in);
405                 }
406
407                 if (cmd->no_stderr)
408                         dup_devnull(2);
409                 else if (need_err) {
410                         dup2(fderr[1], 2);
411                         close_pair(fderr);
412                 } else if (cmd->err > 1) {
413                         dup2(cmd->err, 2);
414                         close(cmd->err);
415                 }
416
417                 if (cmd->no_stdout)
418                         dup_devnull(1);
419                 else if (cmd->stdout_to_stderr)
420                         dup2(2, 1);
421                 else if (need_out) {
422                         dup2(fdout[1], 1);
423                         close_pair(fdout);
424                 } else if (cmd->out > 1) {
425                         dup2(cmd->out, 1);
426                         close(cmd->out);
427                 }
428
429                 if (cmd->dir && chdir(cmd->dir))
430                         die_errno("exec '%s': cd to '%s' failed", cmd->argv[0],
431                             cmd->dir);
432                 if (cmd->env) {
433                         for (; *cmd->env; cmd->env++) {
434                                 if (strchr(*cmd->env, '='))
435                                         putenv((char *)*cmd->env);
436                                 else
437                                         unsetenv(*cmd->env);
438                         }
439                 }
440                 if (cmd->git_cmd)
441                         execv_git_cmd(cmd->argv);
442                 else if (cmd->use_shell)
443                         execv_shell_cmd(cmd->argv);
444                 else
445                         sane_execvp(cmd->argv[0], (char *const*) cmd->argv);
446                 if (errno == ENOENT) {
447                         if (!cmd->silent_exec_failure)
448                                 error("cannot run %s: %s", cmd->argv[0],
449                                         strerror(ENOENT));
450                         exit(127);
451                 } else {
452                         die_errno("cannot exec '%s'", cmd->argv[0]);
453                 }
454         }
455         if (cmd->pid < 0)
456                 error_errno("cannot fork() for %s", cmd->argv[0]);
457         else if (cmd->clean_on_exit)
458                 mark_child_for_cleanup(cmd->pid, cmd);
459
460         /*
461          * Wait for child's execvp. If the execvp succeeds (or if fork()
462          * failed), EOF is seen immediately by the parent. Otherwise, the
463          * child process sends a single byte.
464          * Note that use of this infrastructure is completely advisory,
465          * therefore, we keep error checks minimal.
466          */
467         close(notify_pipe[1]);
468         if (read(notify_pipe[0], &notify_pipe[1], 1) == 1) {
469                 /*
470                  * At this point we know that fork() succeeded, but execvp()
471                  * failed. Errors have been reported to our stderr.
472                  */
473                 wait_or_whine(cmd->pid, cmd->argv[0], 0);
474                 failed_errno = errno;
475                 cmd->pid = -1;
476         }
477         close(notify_pipe[0]);
478 }
479 #else
480 {
481         int fhin = 0, fhout = 1, fherr = 2;
482         const char **sargv = cmd->argv;
483         struct argv_array nargv = ARGV_ARRAY_INIT;
484
485         if (cmd->no_stdin)
486                 fhin = open("/dev/null", O_RDWR);
487         else if (need_in)
488                 fhin = dup(fdin[0]);
489         else if (cmd->in)
490                 fhin = dup(cmd->in);
491
492         if (cmd->no_stderr)
493                 fherr = open("/dev/null", O_RDWR);
494         else if (need_err)
495                 fherr = dup(fderr[1]);
496         else if (cmd->err > 2)
497                 fherr = dup(cmd->err);
498
499         if (cmd->no_stdout)
500                 fhout = open("/dev/null", O_RDWR);
501         else if (cmd->stdout_to_stderr)
502                 fhout = dup(fherr);
503         else if (need_out)
504                 fhout = dup(fdout[1]);
505         else if (cmd->out > 1)
506                 fhout = dup(cmd->out);
507
508         if (cmd->git_cmd)
509                 cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
510         else if (cmd->use_shell)
511                 cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
512
513         cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
514                         cmd->dir, fhin, fhout, fherr);
515         failed_errno = errno;
516         if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
517                 error_errno("cannot spawn %s", cmd->argv[0]);
518         if (cmd->clean_on_exit && cmd->pid >= 0)
519                 mark_child_for_cleanup(cmd->pid, cmd);
520
521         argv_array_clear(&nargv);
522         cmd->argv = sargv;
523         if (fhin != 0)
524                 close(fhin);
525         if (fhout != 1)
526                 close(fhout);
527         if (fherr != 2)
528                 close(fherr);
529 }
530 #endif
531
532         if (cmd->pid < 0) {
533                 if (need_in)
534                         close_pair(fdin);
535                 else if (cmd->in)
536                         close(cmd->in);
537                 if (need_out)
538                         close_pair(fdout);
539                 else if (cmd->out)
540                         close(cmd->out);
541                 if (need_err)
542                         close_pair(fderr);
543                 else if (cmd->err)
544                         close(cmd->err);
545                 child_process_clear(cmd);
546                 errno = failed_errno;
547                 return -1;
548         }
549
550         if (need_in)
551                 close(fdin[0]);
552         else if (cmd->in)
553                 close(cmd->in);
554
555         if (need_out)
556                 close(fdout[1]);
557         else if (cmd->out)
558                 close(cmd->out);
559
560         if (need_err)
561                 close(fderr[1]);
562         else if (cmd->err)
563                 close(cmd->err);
564
565         return 0;
566 }
567
568 int finish_command(struct child_process *cmd)
569 {
570         int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
571         child_process_clear(cmd);
572         return ret;
573 }
574
575 int finish_command_in_signal(struct child_process *cmd)
576 {
577         return wait_or_whine(cmd->pid, cmd->argv[0], 1);
578 }
579
580
581 int run_command(struct child_process *cmd)
582 {
583         int code;
584
585         if (cmd->out < 0 || cmd->err < 0)
586                 die("BUG: run_command with a pipe can cause deadlock");
587
588         code = start_command(cmd);
589         if (code)
590                 return code;
591         return finish_command(cmd);
592 }
593
594 int run_command_v_opt(const char **argv, int opt)
595 {
596         return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
597 }
598
599 int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
600 {
601         struct child_process cmd = CHILD_PROCESS_INIT;
602         cmd.argv = argv;
603         cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
604         cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
605         cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
606         cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
607         cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
608         cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
609         cmd.dir = dir;
610         cmd.env = env;
611         return run_command(&cmd);
612 }
613
614 #ifndef NO_PTHREADS
615 static pthread_t main_thread;
616 static int main_thread_set;
617 static pthread_key_t async_key;
618 static pthread_key_t async_die_counter;
619
620 static void *run_thread(void *data)
621 {
622         struct async *async = data;
623         intptr_t ret;
624
625         if (async->isolate_sigpipe) {
626                 sigset_t mask;
627                 sigemptyset(&mask);
628                 sigaddset(&mask, SIGPIPE);
629                 if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
630                         ret = error("unable to block SIGPIPE in async thread");
631                         return (void *)ret;
632                 }
633         }
634
635         pthread_setspecific(async_key, async);
636         ret = async->proc(async->proc_in, async->proc_out, async->data);
637         return (void *)ret;
638 }
639
640 static NORETURN void die_async(const char *err, va_list params)
641 {
642         vreportf("fatal: ", err, params);
643
644         if (in_async()) {
645                 struct async *async = pthread_getspecific(async_key);
646                 if (async->proc_in >= 0)
647                         close(async->proc_in);
648                 if (async->proc_out >= 0)
649                         close(async->proc_out);
650                 pthread_exit((void *)128);
651         }
652
653         exit(128);
654 }
655
656 static int async_die_is_recursing(void)
657 {
658         void *ret = pthread_getspecific(async_die_counter);
659         pthread_setspecific(async_die_counter, (void *)1);
660         return ret != NULL;
661 }
662
663 int in_async(void)
664 {
665         if (!main_thread_set)
666                 return 0; /* no asyncs started yet */
667         return !pthread_equal(main_thread, pthread_self());
668 }
669
670 static void NORETURN async_exit(int code)
671 {
672         pthread_exit((void *)(intptr_t)code);
673 }
674
675 #else
676
677 static struct {
678         void (**handlers)(void);
679         size_t nr;
680         size_t alloc;
681 } git_atexit_hdlrs;
682
683 static int git_atexit_installed;
684
685 static void git_atexit_dispatch(void)
686 {
687         size_t i;
688
689         for (i=git_atexit_hdlrs.nr ; i ; i--)
690                 git_atexit_hdlrs.handlers[i-1]();
691 }
692
693 static void git_atexit_clear(void)
694 {
695         free(git_atexit_hdlrs.handlers);
696         memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
697         git_atexit_installed = 0;
698 }
699
700 #undef atexit
701 int git_atexit(void (*handler)(void))
702 {
703         ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
704         git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
705         if (!git_atexit_installed) {
706                 if (atexit(&git_atexit_dispatch))
707                         return -1;
708                 git_atexit_installed = 1;
709         }
710         return 0;
711 }
712 #define atexit git_atexit
713
714 static int process_is_async;
715 int in_async(void)
716 {
717         return process_is_async;
718 }
719
720 static void NORETURN async_exit(int code)
721 {
722         exit(code);
723 }
724
725 #endif
726
727 void check_pipe(int err)
728 {
729         if (err == EPIPE) {
730                 if (in_async())
731                         async_exit(141);
732
733                 signal(SIGPIPE, SIG_DFL);
734                 raise(SIGPIPE);
735                 /* Should never happen, but just in case... */
736                 exit(141);
737         }
738 }
739
740 int start_async(struct async *async)
741 {
742         int need_in, need_out;
743         int fdin[2], fdout[2];
744         int proc_in, proc_out;
745
746         need_in = async->in < 0;
747         if (need_in) {
748                 if (pipe(fdin) < 0) {
749                         if (async->out > 0)
750                                 close(async->out);
751                         return error_errno("cannot create pipe");
752                 }
753                 async->in = fdin[1];
754         }
755
756         need_out = async->out < 0;
757         if (need_out) {
758                 if (pipe(fdout) < 0) {
759                         if (need_in)
760                                 close_pair(fdin);
761                         else if (async->in)
762                                 close(async->in);
763                         return error_errno("cannot create pipe");
764                 }
765                 async->out = fdout[0];
766         }
767
768         if (need_in)
769                 proc_in = fdin[0];
770         else if (async->in)
771                 proc_in = async->in;
772         else
773                 proc_in = -1;
774
775         if (need_out)
776                 proc_out = fdout[1];
777         else if (async->out)
778                 proc_out = async->out;
779         else
780                 proc_out = -1;
781
782 #ifdef NO_PTHREADS
783         /* Flush stdio before fork() to avoid cloning buffers */
784         fflush(NULL);
785
786         async->pid = fork();
787         if (async->pid < 0) {
788                 error_errno("fork (async) failed");
789                 goto error;
790         }
791         if (!async->pid) {
792                 if (need_in)
793                         close(fdin[1]);
794                 if (need_out)
795                         close(fdout[0]);
796                 git_atexit_clear();
797                 process_is_async = 1;
798                 exit(!!async->proc(proc_in, proc_out, async->data));
799         }
800
801         mark_child_for_cleanup(async->pid, NULL);
802
803         if (need_in)
804                 close(fdin[0]);
805         else if (async->in)
806                 close(async->in);
807
808         if (need_out)
809                 close(fdout[1]);
810         else if (async->out)
811                 close(async->out);
812 #else
813         if (!main_thread_set) {
814                 /*
815                  * We assume that the first time that start_async is called
816                  * it is from the main thread.
817                  */
818                 main_thread_set = 1;
819                 main_thread = pthread_self();
820                 pthread_key_create(&async_key, NULL);
821                 pthread_key_create(&async_die_counter, NULL);
822                 set_die_routine(die_async);
823                 set_die_is_recursing_routine(async_die_is_recursing);
824         }
825
826         if (proc_in >= 0)
827                 set_cloexec(proc_in);
828         if (proc_out >= 0)
829                 set_cloexec(proc_out);
830         async->proc_in = proc_in;
831         async->proc_out = proc_out;
832         {
833                 int err = pthread_create(&async->tid, NULL, run_thread, async);
834                 if (err) {
835                         error_errno("cannot create thread");
836                         goto error;
837                 }
838         }
839 #endif
840         return 0;
841
842 error:
843         if (need_in)
844                 close_pair(fdin);
845         else if (async->in)
846                 close(async->in);
847
848         if (need_out)
849                 close_pair(fdout);
850         else if (async->out)
851                 close(async->out);
852         return -1;
853 }
854
855 int finish_async(struct async *async)
856 {
857 #ifdef NO_PTHREADS
858         return wait_or_whine(async->pid, "child process", 0);
859 #else
860         void *ret = (void *)(intptr_t)(-1);
861
862         if (pthread_join(async->tid, &ret))
863                 error("pthread_join failed");
864         return (int)(intptr_t)ret;
865 #endif
866 }
867
868 const char *find_hook(const char *name)
869 {
870         static struct strbuf path = STRBUF_INIT;
871
872         strbuf_reset(&path);
873         strbuf_git_path(&path, "hooks/%s", name);
874         if (access(path.buf, X_OK) < 0) {
875 #ifdef STRIP_EXTENSION
876                 strbuf_addstr(&path, STRIP_EXTENSION);
877                 if (access(path.buf, X_OK) >= 0)
878                         return path.buf;
879 #endif
880                 return NULL;
881         }
882         return path.buf;
883 }
884
885 int run_hook_ve(const char *const *env, const char *name, va_list args)
886 {
887         struct child_process hook = CHILD_PROCESS_INIT;
888         const char *p;
889
890         p = find_hook(name);
891         if (!p)
892                 return 0;
893
894         argv_array_push(&hook.args, p);
895         while ((p = va_arg(args, const char *)))
896                 argv_array_push(&hook.args, p);
897         hook.env = env;
898         hook.no_stdin = 1;
899         hook.stdout_to_stderr = 1;
900
901         return run_command(&hook);
902 }
903
904 int run_hook_le(const char *const *env, const char *name, ...)
905 {
906         va_list args;
907         int ret;
908
909         va_start(args, name);
910         ret = run_hook_ve(env, name, args);
911         va_end(args);
912
913         return ret;
914 }
915
916 struct io_pump {
917         /* initialized by caller */
918         int fd;
919         int type; /* POLLOUT or POLLIN */
920         union {
921                 struct {
922                         const char *buf;
923                         size_t len;
924                 } out;
925                 struct {
926                         struct strbuf *buf;
927                         size_t hint;
928                 } in;
929         } u;
930
931         /* returned by pump_io */
932         int error; /* 0 for success, otherwise errno */
933
934         /* internal use */
935         struct pollfd *pfd;
936 };
937
938 static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
939 {
940         int pollsize = 0;
941         int i;
942
943         for (i = 0; i < nr; i++) {
944                 struct io_pump *io = &slots[i];
945                 if (io->fd < 0)
946                         continue;
947                 pfd[pollsize].fd = io->fd;
948                 pfd[pollsize].events = io->type;
949                 io->pfd = &pfd[pollsize++];
950         }
951
952         if (!pollsize)
953                 return 0;
954
955         if (poll(pfd, pollsize, -1) < 0) {
956                 if (errno == EINTR)
957                         return 1;
958                 die_errno("poll failed");
959         }
960
961         for (i = 0; i < nr; i++) {
962                 struct io_pump *io = &slots[i];
963
964                 if (io->fd < 0)
965                         continue;
966
967                 if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
968                         continue;
969
970                 if (io->type == POLLOUT) {
971                         ssize_t len = xwrite(io->fd,
972                                              io->u.out.buf, io->u.out.len);
973                         if (len < 0) {
974                                 io->error = errno;
975                                 close(io->fd);
976                                 io->fd = -1;
977                         } else {
978                                 io->u.out.buf += len;
979                                 io->u.out.len -= len;
980                                 if (!io->u.out.len) {
981                                         close(io->fd);
982                                         io->fd = -1;
983                                 }
984                         }
985                 }
986
987                 if (io->type == POLLIN) {
988                         ssize_t len = strbuf_read_once(io->u.in.buf,
989                                                        io->fd, io->u.in.hint);
990                         if (len < 0)
991                                 io->error = errno;
992                         if (len <= 0) {
993                                 close(io->fd);
994                                 io->fd = -1;
995                         }
996                 }
997         }
998
999         return 1;
1000 }
1001
1002 static int pump_io(struct io_pump *slots, int nr)
1003 {
1004         struct pollfd *pfd;
1005         int i;
1006
1007         for (i = 0; i < nr; i++)
1008                 slots[i].error = 0;
1009
1010         ALLOC_ARRAY(pfd, nr);
1011         while (pump_io_round(slots, nr, pfd))
1012                 ; /* nothing */
1013         free(pfd);
1014
1015         /* There may be multiple errno values, so just pick the first. */
1016         for (i = 0; i < nr; i++) {
1017                 if (slots[i].error) {
1018                         errno = slots[i].error;
1019                         return -1;
1020                 }
1021         }
1022         return 0;
1023 }
1024
1025
1026 int pipe_command(struct child_process *cmd,
1027                  const char *in, size_t in_len,
1028                  struct strbuf *out, size_t out_hint,
1029                  struct strbuf *err, size_t err_hint)
1030 {
1031         struct io_pump io[3];
1032         int nr = 0;
1033
1034         if (in)
1035                 cmd->in = -1;
1036         if (out)
1037                 cmd->out = -1;
1038         if (err)
1039                 cmd->err = -1;
1040
1041         if (start_command(cmd) < 0)
1042                 return -1;
1043
1044         if (in) {
1045                 io[nr].fd = cmd->in;
1046                 io[nr].type = POLLOUT;
1047                 io[nr].u.out.buf = in;
1048                 io[nr].u.out.len = in_len;
1049                 nr++;
1050         }
1051         if (out) {
1052                 io[nr].fd = cmd->out;
1053                 io[nr].type = POLLIN;
1054                 io[nr].u.in.buf = out;
1055                 io[nr].u.in.hint = out_hint;
1056                 nr++;
1057         }
1058         if (err) {
1059                 io[nr].fd = cmd->err;
1060                 io[nr].type = POLLIN;
1061                 io[nr].u.in.buf = err;
1062                 io[nr].u.in.hint = err_hint;
1063                 nr++;
1064         }
1065
1066         if (pump_io(io, nr) < 0) {
1067                 finish_command(cmd); /* throw away exit code */
1068                 return -1;
1069         }
1070
1071         return finish_command(cmd);
1072 }
1073
1074 enum child_state {
1075         GIT_CP_FREE,
1076         GIT_CP_WORKING,
1077         GIT_CP_WAIT_CLEANUP,
1078 };
1079
1080 struct parallel_processes {
1081         void *data;
1082
1083         int max_processes;
1084         int nr_processes;
1085
1086         get_next_task_fn get_next_task;
1087         start_failure_fn start_failure;
1088         task_finished_fn task_finished;
1089
1090         struct {
1091                 enum child_state state;
1092                 struct child_process process;
1093                 struct strbuf err;
1094                 void *data;
1095         } *children;
1096         /*
1097          * The struct pollfd is logically part of *children,
1098          * but the system call expects it as its own array.
1099          */
1100         struct pollfd *pfd;
1101
1102         unsigned shutdown : 1;
1103
1104         int output_owner;
1105         struct strbuf buffered_output; /* of finished children */
1106 };
1107
1108 static int default_start_failure(struct strbuf *out,
1109                                  void *pp_cb,
1110                                  void *pp_task_cb)
1111 {
1112         return 0;
1113 }
1114
1115 static int default_task_finished(int result,
1116                                  struct strbuf *out,
1117                                  void *pp_cb,
1118                                  void *pp_task_cb)
1119 {
1120         return 0;
1121 }
1122
1123 static void kill_children(struct parallel_processes *pp, int signo)
1124 {
1125         int i, n = pp->max_processes;
1126
1127         for (i = 0; i < n; i++)
1128                 if (pp->children[i].state == GIT_CP_WORKING)
1129                         kill(pp->children[i].process.pid, signo);
1130 }
1131
1132 static struct parallel_processes *pp_for_signal;
1133
1134 static void handle_children_on_signal(int signo)
1135 {
1136         kill_children(pp_for_signal, signo);
1137         sigchain_pop(signo);
1138         raise(signo);
1139 }
1140
1141 static void pp_init(struct parallel_processes *pp,
1142                     int n,
1143                     get_next_task_fn get_next_task,
1144                     start_failure_fn start_failure,
1145                     task_finished_fn task_finished,
1146                     void *data)
1147 {
1148         int i;
1149
1150         if (n < 1)
1151                 n = online_cpus();
1152
1153         pp->max_processes = n;
1154
1155         trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
1156
1157         pp->data = data;
1158         if (!get_next_task)
1159                 die("BUG: you need to specify a get_next_task function");
1160         pp->get_next_task = get_next_task;
1161
1162         pp->start_failure = start_failure ? start_failure : default_start_failure;
1163         pp->task_finished = task_finished ? task_finished : default_task_finished;
1164
1165         pp->nr_processes = 0;
1166         pp->output_owner = 0;
1167         pp->shutdown = 0;
1168         pp->children = xcalloc(n, sizeof(*pp->children));
1169         pp->pfd = xcalloc(n, sizeof(*pp->pfd));
1170         strbuf_init(&pp->buffered_output, 0);
1171
1172         for (i = 0; i < n; i++) {
1173                 strbuf_init(&pp->children[i].err, 0);
1174                 child_process_init(&pp->children[i].process);
1175                 pp->pfd[i].events = POLLIN | POLLHUP;
1176                 pp->pfd[i].fd = -1;
1177         }
1178
1179         pp_for_signal = pp;
1180         sigchain_push_common(handle_children_on_signal);
1181 }
1182
1183 static void pp_cleanup(struct parallel_processes *pp)
1184 {
1185         int i;
1186
1187         trace_printf("run_processes_parallel: done");
1188         for (i = 0; i < pp->max_processes; i++) {
1189                 strbuf_release(&pp->children[i].err);
1190                 child_process_clear(&pp->children[i].process);
1191         }
1192
1193         free(pp->children);
1194         free(pp->pfd);
1195
1196         /*
1197          * When get_next_task added messages to the buffer in its last
1198          * iteration, the buffered output is non empty.
1199          */
1200         strbuf_write(&pp->buffered_output, stderr);
1201         strbuf_release(&pp->buffered_output);
1202
1203         sigchain_pop_common();
1204 }
1205
1206 /* returns
1207  *  0 if a new task was started.
1208  *  1 if no new jobs was started (get_next_task ran out of work, non critical
1209  *    problem with starting a new command)
1210  * <0 no new job was started, user wishes to shutdown early. Use negative code
1211  *    to signal the children.
1212  */
1213 static int pp_start_one(struct parallel_processes *pp)
1214 {
1215         int i, code;
1216
1217         for (i = 0; i < pp->max_processes; i++)
1218                 if (pp->children[i].state == GIT_CP_FREE)
1219                         break;
1220         if (i == pp->max_processes)
1221                 die("BUG: bookkeeping is hard");
1222
1223         code = pp->get_next_task(&pp->children[i].process,
1224                                  &pp->children[i].err,
1225                                  pp->data,
1226                                  &pp->children[i].data);
1227         if (!code) {
1228                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1229                 strbuf_reset(&pp->children[i].err);
1230                 return 1;
1231         }
1232         pp->children[i].process.err = -1;
1233         pp->children[i].process.stdout_to_stderr = 1;
1234         pp->children[i].process.no_stdin = 1;
1235
1236         if (start_command(&pp->children[i].process)) {
1237                 code = pp->start_failure(&pp->children[i].err,
1238                                          pp->data,
1239                                          &pp->children[i].data);
1240                 strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1241                 strbuf_reset(&pp->children[i].err);
1242                 if (code)
1243                         pp->shutdown = 1;
1244                 return code;
1245         }
1246
1247         pp->nr_processes++;
1248         pp->children[i].state = GIT_CP_WORKING;
1249         pp->pfd[i].fd = pp->children[i].process.err;
1250         return 0;
1251 }
1252
1253 static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1254 {
1255         int i;
1256
1257         while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1258                 if (errno == EINTR)
1259                         continue;
1260                 pp_cleanup(pp);
1261                 die_errno("poll");
1262         }
1263
1264         /* Buffer output from all pipes. */
1265         for (i = 0; i < pp->max_processes; i++) {
1266                 if (pp->children[i].state == GIT_CP_WORKING &&
1267                     pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1268                         int n = strbuf_read_once(&pp->children[i].err,
1269                                                  pp->children[i].process.err, 0);
1270                         if (n == 0) {
1271                                 close(pp->children[i].process.err);
1272                                 pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1273                         } else if (n < 0)
1274                                 if (errno != EAGAIN)
1275                                         die_errno("read");
1276                 }
1277         }
1278 }
1279
1280 static void pp_output(struct parallel_processes *pp)
1281 {
1282         int i = pp->output_owner;
1283         if (pp->children[i].state == GIT_CP_WORKING &&
1284             pp->children[i].err.len) {
1285                 strbuf_write(&pp->children[i].err, stderr);
1286                 strbuf_reset(&pp->children[i].err);
1287         }
1288 }
1289
1290 static int pp_collect_finished(struct parallel_processes *pp)
1291 {
1292         int i, code;
1293         int n = pp->max_processes;
1294         int result = 0;
1295
1296         while (pp->nr_processes > 0) {
1297                 for (i = 0; i < pp->max_processes; i++)
1298                         if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1299                                 break;
1300                 if (i == pp->max_processes)
1301                         break;
1302
1303                 code = finish_command(&pp->children[i].process);
1304
1305                 code = pp->task_finished(code,
1306                                          &pp->children[i].err, pp->data,
1307                                          &pp->children[i].data);
1308
1309                 if (code)
1310                         result = code;
1311                 if (code < 0)
1312                         break;
1313
1314                 pp->nr_processes--;
1315                 pp->children[i].state = GIT_CP_FREE;
1316                 pp->pfd[i].fd = -1;
1317                 child_process_init(&pp->children[i].process);
1318
1319                 if (i != pp->output_owner) {
1320                         strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1321                         strbuf_reset(&pp->children[i].err);
1322                 } else {
1323                         strbuf_write(&pp->children[i].err, stderr);
1324                         strbuf_reset(&pp->children[i].err);
1325
1326                         /* Output all other finished child processes */
1327                         strbuf_write(&pp->buffered_output, stderr);
1328                         strbuf_reset(&pp->buffered_output);
1329
1330                         /*
1331                          * Pick next process to output live.
1332                          * NEEDSWORK:
1333                          * For now we pick it randomly by doing a round
1334                          * robin. Later we may want to pick the one with
1335                          * the most output or the longest or shortest
1336                          * running process time.
1337                          */
1338                         for (i = 0; i < n; i++)
1339                                 if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1340                                         break;
1341                         pp->output_owner = (pp->output_owner + i) % n;
1342                 }
1343         }
1344         return result;
1345 }
1346
1347 int run_processes_parallel(int n,
1348                            get_next_task_fn get_next_task,
1349                            start_failure_fn start_failure,
1350                            task_finished_fn task_finished,
1351                            void *pp_cb)
1352 {
1353         int i, code;
1354         int output_timeout = 100;
1355         int spawn_cap = 4;
1356         struct parallel_processes pp;
1357
1358         pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1359         while (1) {
1360                 for (i = 0;
1361                     i < spawn_cap && !pp.shutdown &&
1362                     pp.nr_processes < pp.max_processes;
1363                     i++) {
1364                         code = pp_start_one(&pp);
1365                         if (!code)
1366                                 continue;
1367                         if (code < 0) {
1368                                 pp.shutdown = 1;
1369                                 kill_children(&pp, -code);
1370                         }
1371                         break;
1372                 }
1373                 if (!pp.nr_processes)
1374                         break;
1375                 pp_buffer_stderr(&pp, output_timeout);
1376                 pp_output(&pp);
1377                 code = pp_collect_finished(&pp);
1378                 if (code) {
1379                         pp.shutdown = 1;
1380                         if (code < 0)
1381                                 kill_children(&pp, -code);
1382                 }
1383         }
1384
1385         pp_cleanup(&pp);
1386         return 0;
1387 }