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