1 #include "cache.h"
2 #include "run-command.h"
3 #include "exec-cmd.h"
4 #include "sigchain.h"
5 #include "argv-array.h"
6 #include "thread-utils.h"
7 #include "strbuf.h"
8 #include "string-list.h"
9 #include "quote.h"
10 
child_process_init(struct child_process * child)11 void child_process_init(struct child_process *child)
12 {
13 	memset(child, 0, sizeof(*child));
14 	argv_array_init(&child->args);
15 	argv_array_init(&child->env_array);
16 }
17 
child_process_clear(struct child_process * child)18 void child_process_clear(struct child_process *child)
19 {
20 	argv_array_clear(&child->args);
21 	argv_array_clear(&child->env_array);
22 }
23 
24 struct child_to_clean {
25 	pid_t pid;
26 	struct child_process *process;
27 	struct child_to_clean *next;
28 };
29 static struct child_to_clean *children_to_clean;
30 static int installed_child_cleanup_handler;
31 
cleanup_children(int sig,int in_signal)32 static void cleanup_children(int sig, int in_signal)
33 {
34 	struct child_to_clean *children_to_wait_for = NULL;
35 
36 	while (children_to_clean) {
37 		struct child_to_clean *p = children_to_clean;
38 		children_to_clean = p->next;
39 
40 		if (p->process && !in_signal) {
41 			struct child_process *process = p->process;
42 			if (process->clean_on_exit_handler) {
43 				trace_printf(
44 					"trace: run_command: running exit handler for pid %"
45 					PRIuMAX, (uintmax_t)p->pid
46 				);
47 				process->clean_on_exit_handler(process);
48 			}
49 		}
50 
51 		kill(p->pid, sig);
52 
53 		if (p->process && p->process->wait_after_clean) {
54 			p->next = children_to_wait_for;
55 			children_to_wait_for = p;
56 		} else {
57 			if (!in_signal)
58 				free(p);
59 		}
60 	}
61 
62 	while (children_to_wait_for) {
63 		struct child_to_clean *p = children_to_wait_for;
64 		children_to_wait_for = p->next;
65 
66 		while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
67 			; /* spin waiting for process exit or error */
68 
69 		if (!in_signal)
70 			free(p);
71 	}
72 }
73 
cleanup_children_on_signal(int sig)74 static void cleanup_children_on_signal(int sig)
75 {
76 	cleanup_children(sig, 1);
77 	sigchain_pop(sig);
78 	raise(sig);
79 }
80 
cleanup_children_on_exit(void)81 static void cleanup_children_on_exit(void)
82 {
83 	cleanup_children(SIGTERM, 0);
84 }
85 
mark_child_for_cleanup(pid_t pid,struct child_process * process)86 static void mark_child_for_cleanup(pid_t pid, struct child_process *process)
87 {
88 	struct child_to_clean *p = xmalloc(sizeof(*p));
89 	p->pid = pid;
90 	p->process = process;
91 	p->next = children_to_clean;
92 	children_to_clean = p;
93 
94 	if (!installed_child_cleanup_handler) {
95 		atexit(cleanup_children_on_exit);
96 		sigchain_push_common(cleanup_children_on_signal);
97 		installed_child_cleanup_handler = 1;
98 	}
99 }
100 
clear_child_for_cleanup(pid_t pid)101 static void clear_child_for_cleanup(pid_t pid)
102 {
103 	struct child_to_clean **pp;
104 
105 	for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
106 		struct child_to_clean *clean_me = *pp;
107 
108 		if (clean_me->pid == pid) {
109 			*pp = clean_me->next;
110 			free(clean_me);
111 			return;
112 		}
113 	}
114 }
115 
close_pair(int fd[2])116 static inline void close_pair(int fd[2])
117 {
118 	close(fd[0]);
119 	close(fd[1]);
120 }
121 
is_executable(const char * name)122 int is_executable(const char *name)
123 {
124 	struct stat st;
125 
126 	if (stat(name, &st) || /* stat, not lstat */
127 	    !S_ISREG(st.st_mode))
128 		return 0;
129 
130 #if defined(GIT_WINDOWS_NATIVE)
131 	/*
132 	 * On Windows there is no executable bit. The file extension
133 	 * indicates whether it can be run as an executable, and Git
134 	 * has special-handling to detect scripts and launch them
135 	 * through the indicated script interpreter. We test for the
136 	 * file extension first because virus scanners may make
137 	 * it quite expensive to open many files.
138 	 */
139 	if (ends_with(name, ".exe"))
140 		return S_IXUSR;
141 
142 {
143 	/*
144 	 * Now that we know it does not have an executable extension,
145 	 * peek into the file instead.
146 	 */
147 	char buf[3] = { 0 };
148 	int n;
149 	int fd = open(name, O_RDONLY);
150 	st.st_mode &= ~S_IXUSR;
151 	if (fd >= 0) {
152 		n = read(fd, buf, 2);
153 		if (n == 2)
154 			/* look for a she-bang */
155 			if (!strcmp(buf, "#!"))
156 				st.st_mode |= S_IXUSR;
157 		close(fd);
158 	}
159 }
160 #endif
161 	return st.st_mode & S_IXUSR;
162 }
163 
164 /*
165  * Search $PATH for a command.  This emulates the path search that
166  * execvp would perform, without actually executing the command so it
167  * can be used before fork() to prepare to run a command using
168  * execve() or after execvp() to diagnose why it failed.
169  *
170  * The caller should ensure that file contains no directory
171  * separators.
172  *
173  * Returns the path to the command, as found in $PATH or NULL if the
174  * command could not be found.  The caller inherits ownership of the memory
175  * used to store the resultant path.
176  *
177  * This should not be used on Windows, where the $PATH search rules
178  * are more complicated (e.g., a search for "foo" should find
179  * "foo.exe").
180  */
locate_in_PATH(const char * file)181 static char *locate_in_PATH(const char *file)
182 {
183 	const char *p = getenv("PATH");
184 	struct strbuf buf = STRBUF_INIT;
185 
186 	if (!p || !*p)
187 		return NULL;
188 
189 	while (1) {
190 		const char *end = strchrnul(p, ':');
191 
192 		strbuf_reset(&buf);
193 
194 		/* POSIX specifies an empty entry as the current directory. */
195 		if (end != p) {
196 			strbuf_add(&buf, p, end - p);
197 			strbuf_addch(&buf, '/');
198 		}
199 		strbuf_addstr(&buf, file);
200 
201 		if (is_executable(buf.buf))
202 			return strbuf_detach(&buf, NULL);
203 
204 		if (!*end)
205 			break;
206 		p = end + 1;
207 	}
208 
209 	strbuf_release(&buf);
210 	return NULL;
211 }
212 
exists_in_PATH(const char * file)213 static int exists_in_PATH(const char *file)
214 {
215 	char *r = locate_in_PATH(file);
216 	int found = r != NULL;
217 	free(r);
218 	return found;
219 }
220 
sane_execvp(const char * file,char * const argv[])221 int sane_execvp(const char *file, char * const argv[])
222 {
223 #ifndef GIT_WINDOWS_NATIVE
224 	/*
225 	 * execvp() doesn't return, so we all we can do is tell trace2
226 	 * what we are about to do and let it leave a hint in the log
227 	 * (unless of course the execvp() fails).
228 	 *
229 	 * we skip this for Windows because the compat layer already
230 	 * has to emulate the execvp() call anyway.
231 	 */
232 	int exec_id = trace2_exec(file, (const char **)argv);
233 #endif
234 
235 	if (!execvp(file, argv))
236 		return 0; /* cannot happen ;-) */
237 
238 #ifndef GIT_WINDOWS_NATIVE
239 	{
240 		int ec = errno;
241 		trace2_exec_result(exec_id, ec);
242 		errno = ec;
243 	}
244 #endif
245 
246 	/*
247 	 * When a command can't be found because one of the directories
248 	 * listed in $PATH is unsearchable, execvp reports EACCES, but
249 	 * careful usability testing (read: analysis of occasional bug
250 	 * reports) reveals that "No such file or directory" is more
251 	 * intuitive.
252 	 *
253 	 * We avoid commands with "/", because execvp will not do $PATH
254 	 * lookups in that case.
255 	 *
256 	 * The reassignment of EACCES to errno looks like a no-op below,
257 	 * but we need to protect against exists_in_PATH overwriting errno.
258 	 */
259 	if (errno == EACCES && !strchr(file, '/'))
260 		errno = exists_in_PATH(file) ? EACCES : ENOENT;
261 	else if (errno == ENOTDIR && !strchr(file, '/'))
262 		errno = ENOENT;
263 	return -1;
264 }
265 
prepare_shell_cmd(struct argv_array * out,const char ** argv)266 static const char **prepare_shell_cmd(struct argv_array *out, const char **argv)
267 {
268 	if (!argv[0])
269 		BUG("shell command is empty");
270 
271 	if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
272 #ifndef GIT_WINDOWS_NATIVE
273 		argv_array_push(out, SHELL_PATH);
274 #else
275 		argv_array_push(out, "sh");
276 #endif
277 		argv_array_push(out, "-c");
278 
279 		/*
280 		 * If we have no extra arguments, we do not even need to
281 		 * bother with the "$@" magic.
282 		 */
283 		if (!argv[1])
284 			argv_array_push(out, argv[0]);
285 		else
286 			argv_array_pushf(out, "%s \"$@\"", argv[0]);
287 	}
288 
289 	argv_array_pushv(out, argv);
290 	return out->argv;
291 }
292 
293 #ifndef GIT_WINDOWS_NATIVE
294 static int child_notifier = -1;
295 
296 enum child_errcode {
297 	CHILD_ERR_CHDIR,
298 	CHILD_ERR_DUP2,
299 	CHILD_ERR_CLOSE,
300 	CHILD_ERR_SIGPROCMASK,
301 	CHILD_ERR_ENOENT,
302 	CHILD_ERR_SILENT,
303 	CHILD_ERR_ERRNO
304 };
305 
306 struct child_err {
307 	enum child_errcode err;
308 	int syserr; /* errno */
309 };
310 
child_die(enum child_errcode err)311 static void child_die(enum child_errcode err)
312 {
313 	struct child_err buf;
314 
315 	buf.err = err;
316 	buf.syserr = errno;
317 
318 	/* write(2) on buf smaller than PIPE_BUF (min 512) is atomic: */
319 	xwrite(child_notifier, &buf, sizeof(buf));
320 	_exit(1);
321 }
322 
child_dup2(int fd,int to)323 static void child_dup2(int fd, int to)
324 {
325 	if (dup2(fd, to) < 0)
326 		child_die(CHILD_ERR_DUP2);
327 }
328 
child_close(int fd)329 static void child_close(int fd)
330 {
331 	if (close(fd))
332 		child_die(CHILD_ERR_CLOSE);
333 }
334 
child_close_pair(int fd[2])335 static void child_close_pair(int fd[2])
336 {
337 	child_close(fd[0]);
338 	child_close(fd[1]);
339 }
340 
341 /*
342  * parent will make it look like the child spewed a fatal error and died
343  * this is needed to prevent changes to t0061.
344  */
fake_fatal(const char * err,va_list params)345 static void fake_fatal(const char *err, va_list params)
346 {
347 	vreportf("fatal: ", err, params);
348 }
349 
child_error_fn(const char * err,va_list params)350 static void child_error_fn(const char *err, va_list params)
351 {
352 	const char msg[] = "error() should not be called in child\n";
353 	xwrite(2, msg, sizeof(msg) - 1);
354 }
355 
child_warn_fn(const char * err,va_list params)356 static void child_warn_fn(const char *err, va_list params)
357 {
358 	const char msg[] = "warn() should not be called in child\n";
359 	xwrite(2, msg, sizeof(msg) - 1);
360 }
361 
child_die_fn(const char * err,va_list params)362 static void NORETURN child_die_fn(const char *err, va_list params)
363 {
364 	const char msg[] = "die() should not be called in child\n";
365 	xwrite(2, msg, sizeof(msg) - 1);
366 	_exit(2);
367 }
368 
369 /* this runs in the parent process */
child_err_spew(struct child_process * cmd,struct child_err * cerr)370 static void child_err_spew(struct child_process *cmd, struct child_err *cerr)
371 {
372 	static void (*old_errfn)(const char *err, va_list params);
373 
374 	old_errfn = get_error_routine();
375 	set_error_routine(fake_fatal);
376 	errno = cerr->syserr;
377 
378 	switch (cerr->err) {
379 	case CHILD_ERR_CHDIR:
380 		error_errno("exec '%s': cd to '%s' failed",
381 			    cmd->argv[0], cmd->dir);
382 		break;
383 	case CHILD_ERR_DUP2:
384 		error_errno("dup2() in child failed");
385 		break;
386 	case CHILD_ERR_CLOSE:
387 		error_errno("close() in child failed");
388 		break;
389 	case CHILD_ERR_SIGPROCMASK:
390 		error_errno("sigprocmask failed restoring signals");
391 		break;
392 	case CHILD_ERR_ENOENT:
393 		error_errno("cannot run %s", cmd->argv[0]);
394 		break;
395 	case CHILD_ERR_SILENT:
396 		break;
397 	case CHILD_ERR_ERRNO:
398 		error_errno("cannot exec '%s'", cmd->argv[0]);
399 		break;
400 	}
401 	set_error_routine(old_errfn);
402 }
403 
prepare_cmd(struct argv_array * out,const struct child_process * cmd)404 static int prepare_cmd(struct argv_array *out, const struct child_process *cmd)
405 {
406 	if (!cmd->argv[0])
407 		BUG("command is empty");
408 
409 	/*
410 	 * Add SHELL_PATH so in the event exec fails with ENOEXEC we can
411 	 * attempt to interpret the command with 'sh'.
412 	 */
413 	argv_array_push(out, SHELL_PATH);
414 
415 	if (cmd->git_cmd) {
416 		prepare_git_cmd(out, cmd->argv);
417 	} else if (cmd->use_shell) {
418 		prepare_shell_cmd(out, cmd->argv);
419 	} else {
420 		argv_array_pushv(out, cmd->argv);
421 	}
422 
423 	/*
424 	 * If there are no '/' characters in the command then perform a path
425 	 * lookup and use the resolved path as the command to exec.  If there
426 	 * are '/' characters, we have exec attempt to invoke the command
427 	 * directly.
428 	 */
429 	if (!strchr(out->argv[1], '/')) {
430 		char *program = locate_in_PATH(out->argv[1]);
431 		if (program) {
432 			free((char *)out->argv[1]);
433 			out->argv[1] = program;
434 		} else {
435 			argv_array_clear(out);
436 			errno = ENOENT;
437 			return -1;
438 		}
439 	}
440 
441 	return 0;
442 }
443 
prep_childenv(const char * const * deltaenv)444 static char **prep_childenv(const char *const *deltaenv)
445 {
446 	extern char **environ;
447 	char **childenv;
448 	struct string_list env = STRING_LIST_INIT_DUP;
449 	struct strbuf key = STRBUF_INIT;
450 	const char *const *p;
451 	int i;
452 
453 	/* Construct a sorted string list consisting of the current environ */
454 	for (p = (const char *const *) environ; p && *p; p++) {
455 		const char *equals = strchr(*p, '=');
456 
457 		if (equals) {
458 			strbuf_reset(&key);
459 			strbuf_add(&key, *p, equals - *p);
460 			string_list_append(&env, key.buf)->util = (void *) *p;
461 		} else {
462 			string_list_append(&env, *p)->util = (void *) *p;
463 		}
464 	}
465 	string_list_sort(&env);
466 
467 	/* Merge in 'deltaenv' with the current environ */
468 	for (p = deltaenv; p && *p; p++) {
469 		const char *equals = strchr(*p, '=');
470 
471 		if (equals) {
472 			/* ('key=value'), insert or replace entry */
473 			strbuf_reset(&key);
474 			strbuf_add(&key, *p, equals - *p);
475 			string_list_insert(&env, key.buf)->util = (void *) *p;
476 		} else {
477 			/* otherwise ('key') remove existing entry */
478 			string_list_remove(&env, *p, 0);
479 		}
480 	}
481 
482 	/* Create an array of 'char *' to be used as the childenv */
483 	ALLOC_ARRAY(childenv, env.nr + 1);
484 	for (i = 0; i < env.nr; i++)
485 		childenv[i] = env.items[i].util;
486 	childenv[env.nr] = NULL;
487 
488 	string_list_clear(&env, 0);
489 	strbuf_release(&key);
490 	return childenv;
491 }
492 
493 struct atfork_state {
494 #ifndef NO_PTHREADS
495 	int cs;
496 #endif
497 	sigset_t old;
498 };
499 
500 #define CHECK_BUG(err, msg) \
501 	do { \
502 		int e = (err); \
503 		if (e) \
504 			BUG("%s: %s", msg, strerror(e)); \
505 	} while(0)
506 
atfork_prepare(struct atfork_state * as)507 static void atfork_prepare(struct atfork_state *as)
508 {
509 	sigset_t all;
510 
511 	if (sigfillset(&all))
512 		die_errno("sigfillset");
513 #ifdef NO_PTHREADS
514 	if (sigprocmask(SIG_SETMASK, &all, &as->old))
515 		die_errno("sigprocmask");
516 #else
517 	CHECK_BUG(pthread_sigmask(SIG_SETMASK, &all, &as->old),
518 		"blocking all signals");
519 	CHECK_BUG(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &as->cs),
520 		"disabling cancellation");
521 #endif
522 }
523 
atfork_parent(struct atfork_state * as)524 static void atfork_parent(struct atfork_state *as)
525 {
526 #ifdef NO_PTHREADS
527 	if (sigprocmask(SIG_SETMASK, &as->old, NULL))
528 		die_errno("sigprocmask");
529 #else
530 	CHECK_BUG(pthread_setcancelstate(as->cs, NULL),
531 		"re-enabling cancellation");
532 	CHECK_BUG(pthread_sigmask(SIG_SETMASK, &as->old, NULL),
533 		"restoring signal mask");
534 #endif
535 }
536 #endif /* GIT_WINDOWS_NATIVE */
537 
set_cloexec(int fd)538 static inline void set_cloexec(int fd)
539 {
540 	int flags = fcntl(fd, F_GETFD);
541 	if (flags >= 0)
542 		fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
543 }
544 
wait_or_whine(pid_t pid,const char * argv0,int in_signal)545 static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
546 {
547 	int status, code = -1;
548 	pid_t waiting;
549 	int failed_errno = 0;
550 
551 	while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
552 		;	/* nothing */
553 	if (in_signal)
554 		return 0;
555 
556 	if (waiting < 0) {
557 		failed_errno = errno;
558 		error_errno("waitpid for %s failed", argv0);
559 	} else if (waiting != pid) {
560 		error("waitpid is confused (%s)", argv0);
561 	} else if (WIFSIGNALED(status)) {
562 		code = WTERMSIG(status);
563 		if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
564 			error("%s died of signal %d", argv0, code);
565 		/*
566 		 * This return value is chosen so that code & 0xff
567 		 * mimics the exit code that a POSIX shell would report for
568 		 * a program that died from this signal.
569 		 */
570 		code += 128;
571 	} else if (WIFEXITED(status)) {
572 		code = WEXITSTATUS(status);
573 	} else {
574 		error("waitpid is confused (%s)", argv0);
575 	}
576 
577 	clear_child_for_cleanup(pid);
578 
579 	errno = failed_errno;
580 	return code;
581 }
582 
trace_add_env(struct strbuf * dst,const char * const * deltaenv)583 static void trace_add_env(struct strbuf *dst, const char *const *deltaenv)
584 {
585 	struct string_list envs = STRING_LIST_INIT_DUP;
586 	const char *const *e;
587 	int i;
588 	int printed_unset = 0;
589 
590 	/* Last one wins, see run-command.c:prep_childenv() for context */
591 	for (e = deltaenv; e && *e; e++) {
592 		struct strbuf key = STRBUF_INIT;
593 		char *equals = strchr(*e, '=');
594 
595 		if (equals) {
596 			strbuf_add(&key, *e, equals - *e);
597 			string_list_insert(&envs, key.buf)->util = equals + 1;
598 		} else {
599 			string_list_insert(&envs, *e)->util = NULL;
600 		}
601 		strbuf_release(&key);
602 	}
603 
604 	/* "unset X Y...;" */
605 	for (i = 0; i < envs.nr; i++) {
606 		const char *var = envs.items[i].string;
607 		const char *val = envs.items[i].util;
608 
609 		if (val || !getenv(var))
610 			continue;
611 
612 		if (!printed_unset) {
613 			strbuf_addstr(dst, " unset");
614 			printed_unset = 1;
615 		}
616 		strbuf_addf(dst, " %s", var);
617 	}
618 	if (printed_unset)
619 		strbuf_addch(dst, ';');
620 
621 	/* ... followed by "A=B C=D ..." */
622 	for (i = 0; i < envs.nr; i++) {
623 		const char *var = envs.items[i].string;
624 		const char *val = envs.items[i].util;
625 		const char *oldval;
626 
627 		if (!val)
628 			continue;
629 
630 		oldval = getenv(var);
631 		if (oldval && !strcmp(val, oldval))
632 			continue;
633 
634 		strbuf_addf(dst, " %s=", var);
635 		sq_quote_buf_pretty(dst, val);
636 	}
637 	string_list_clear(&envs, 0);
638 }
639 
trace_run_command(const struct child_process * cp)640 static void trace_run_command(const struct child_process *cp)
641 {
642 	struct strbuf buf = STRBUF_INIT;
643 
644 	if (!trace_want(&trace_default_key))
645 		return;
646 
647 	strbuf_addstr(&buf, "trace: run_command:");
648 	if (cp->dir) {
649 		strbuf_addstr(&buf, " cd ");
650 		sq_quote_buf_pretty(&buf, cp->dir);
651 		strbuf_addch(&buf, ';');
652 	}
653 	/*
654 	 * The caller is responsible for initializing cp->env from
655 	 * cp->env_array if needed. We only check one place.
656 	 */
657 	if (cp->env)
658 		trace_add_env(&buf, cp->env);
659 	if (cp->git_cmd)
660 		strbuf_addstr(&buf, " git");
661 	sq_quote_argv_pretty(&buf, cp->argv);
662 
663 	trace_printf("%s", buf.buf);
664 	strbuf_release(&buf);
665 }
666 
start_command(struct child_process * cmd)667 int start_command(struct child_process *cmd)
668 {
669 	int need_in, need_out, need_err;
670 	int fdin[2], fdout[2], fderr[2];
671 	int failed_errno;
672 	char *str;
673 
674 	if (!cmd->argv)
675 		cmd->argv = cmd->args.argv;
676 	if (!cmd->env)
677 		cmd->env = cmd->env_array.argv;
678 
679 	/*
680 	 * In case of errors we must keep the promise to close FDs
681 	 * that have been passed in via ->in and ->out.
682 	 */
683 
684 	need_in = !cmd->no_stdin && cmd->in < 0;
685 	if (need_in) {
686 		if (pipe(fdin) < 0) {
687 			failed_errno = errno;
688 			if (cmd->out > 0)
689 				close(cmd->out);
690 			str = "standard input";
691 			goto fail_pipe;
692 		}
693 		cmd->in = fdin[1];
694 	}
695 
696 	need_out = !cmd->no_stdout
697 		&& !cmd->stdout_to_stderr
698 		&& cmd->out < 0;
699 	if (need_out) {
700 		if (pipe(fdout) < 0) {
701 			failed_errno = errno;
702 			if (need_in)
703 				close_pair(fdin);
704 			else if (cmd->in)
705 				close(cmd->in);
706 			str = "standard output";
707 			goto fail_pipe;
708 		}
709 		cmd->out = fdout[0];
710 	}
711 
712 	need_err = !cmd->no_stderr && cmd->err < 0;
713 	if (need_err) {
714 		if (pipe(fderr) < 0) {
715 			failed_errno = errno;
716 			if (need_in)
717 				close_pair(fdin);
718 			else if (cmd->in)
719 				close(cmd->in);
720 			if (need_out)
721 				close_pair(fdout);
722 			else if (cmd->out)
723 				close(cmd->out);
724 			str = "standard error";
725 fail_pipe:
726 			error("cannot create %s pipe for %s: %s",
727 				str, cmd->argv[0], strerror(failed_errno));
728 			child_process_clear(cmd);
729 			errno = failed_errno;
730 			return -1;
731 		}
732 		cmd->err = fderr[0];
733 	}
734 
735 	trace2_child_start(cmd);
736 	trace_run_command(cmd);
737 
738 	fflush(NULL);
739 
740 #ifndef GIT_WINDOWS_NATIVE
741 {
742 	int notify_pipe[2];
743 	int null_fd = -1;
744 	char **childenv;
745 	struct argv_array argv = ARGV_ARRAY_INIT;
746 	struct child_err cerr;
747 	struct atfork_state as;
748 
749 	if (prepare_cmd(&argv, cmd) < 0) {
750 		failed_errno = errno;
751 		cmd->pid = -1;
752 		if (!cmd->silent_exec_failure)
753 			error_errno("cannot run %s", cmd->argv[0]);
754 		goto end_of_spawn;
755 	}
756 
757 	if (pipe(notify_pipe))
758 		notify_pipe[0] = notify_pipe[1] = -1;
759 
760 	if (cmd->no_stdin || cmd->no_stdout || cmd->no_stderr) {
761 		null_fd = open("/dev/null", O_RDWR | O_CLOEXEC);
762 		if (null_fd < 0)
763 			die_errno(_("open /dev/null failed"));
764 		set_cloexec(null_fd);
765 	}
766 
767 	childenv = prep_childenv(cmd->env);
768 	atfork_prepare(&as);
769 
770 	/*
771 	 * NOTE: In order to prevent deadlocking when using threads special
772 	 * care should be taken with the function calls made in between the
773 	 * fork() and exec() calls.  No calls should be made to functions which
774 	 * require acquiring a lock (e.g. malloc) as the lock could have been
775 	 * held by another thread at the time of forking, causing the lock to
776 	 * never be released in the child process.  This means only
777 	 * Async-Signal-Safe functions are permitted in the child.
778 	 */
779 	cmd->pid = fork();
780 	failed_errno = errno;
781 	if (!cmd->pid) {
782 		int sig;
783 		/*
784 		 * Ensure the default die/error/warn routines do not get
785 		 * called, they can take stdio locks and malloc.
786 		 */
787 		set_die_routine(child_die_fn);
788 		set_error_routine(child_error_fn);
789 		set_warn_routine(child_warn_fn);
790 
791 		close(notify_pipe[0]);
792 		set_cloexec(notify_pipe[1]);
793 		child_notifier = notify_pipe[1];
794 
795 		if (cmd->no_stdin)
796 			child_dup2(null_fd, 0);
797 		else if (need_in) {
798 			child_dup2(fdin[0], 0);
799 			child_close_pair(fdin);
800 		} else if (cmd->in) {
801 			child_dup2(cmd->in, 0);
802 			child_close(cmd->in);
803 		}
804 
805 		if (cmd->no_stderr)
806 			child_dup2(null_fd, 2);
807 		else if (need_err) {
808 			child_dup2(fderr[1], 2);
809 			child_close_pair(fderr);
810 		} else if (cmd->err > 1) {
811 			child_dup2(cmd->err, 2);
812 			child_close(cmd->err);
813 		}
814 
815 		if (cmd->no_stdout)
816 			child_dup2(null_fd, 1);
817 		else if (cmd->stdout_to_stderr)
818 			child_dup2(2, 1);
819 		else if (need_out) {
820 			child_dup2(fdout[1], 1);
821 			child_close_pair(fdout);
822 		} else if (cmd->out > 1) {
823 			child_dup2(cmd->out, 1);
824 			child_close(cmd->out);
825 		}
826 
827 		if (cmd->dir && chdir(cmd->dir))
828 			child_die(CHILD_ERR_CHDIR);
829 
830 		/*
831 		 * restore default signal handlers here, in case
832 		 * we catch a signal right before execve below
833 		 */
834 		for (sig = 1; sig < NSIG; sig++) {
835 			/* ignored signals get reset to SIG_DFL on execve */
836 			if (signal(sig, SIG_DFL) == SIG_IGN)
837 				signal(sig, SIG_IGN);
838 		}
839 
840 		if (sigprocmask(SIG_SETMASK, &as.old, NULL) != 0)
841 			child_die(CHILD_ERR_SIGPROCMASK);
842 
843 		/*
844 		 * Attempt to exec using the command and arguments starting at
845 		 * argv.argv[1].  argv.argv[0] contains SHELL_PATH which will
846 		 * be used in the event exec failed with ENOEXEC at which point
847 		 * we will try to interpret the command using 'sh'.
848 		 */
849 		execve(argv.argv[1], (char *const *) argv.argv + 1,
850 		       (char *const *) childenv);
851 		if (errno == ENOEXEC)
852 			execve(argv.argv[0], (char *const *) argv.argv,
853 			       (char *const *) childenv);
854 
855 		if (errno == ENOENT) {
856 			if (cmd->silent_exec_failure)
857 				child_die(CHILD_ERR_SILENT);
858 			child_die(CHILD_ERR_ENOENT);
859 		} else {
860 			child_die(CHILD_ERR_ERRNO);
861 		}
862 	}
863 	atfork_parent(&as);
864 	if (cmd->pid < 0)
865 		error_errno("cannot fork() for %s", cmd->argv[0]);
866 	else if (cmd->clean_on_exit)
867 		mark_child_for_cleanup(cmd->pid, cmd);
868 
869 	/*
870 	 * Wait for child's exec. If the exec succeeds (or if fork()
871 	 * failed), EOF is seen immediately by the parent. Otherwise, the
872 	 * child process sends a child_err struct.
873 	 * Note that use of this infrastructure is completely advisory,
874 	 * therefore, we keep error checks minimal.
875 	 */
876 	close(notify_pipe[1]);
877 	if (xread(notify_pipe[0], &cerr, sizeof(cerr)) == sizeof(cerr)) {
878 		/*
879 		 * At this point we know that fork() succeeded, but exec()
880 		 * failed. Errors have been reported to our stderr.
881 		 */
882 		wait_or_whine(cmd->pid, cmd->argv[0], 0);
883 		child_err_spew(cmd, &cerr);
884 		failed_errno = errno;
885 		cmd->pid = -1;
886 	}
887 	close(notify_pipe[0]);
888 
889 	if (null_fd >= 0)
890 		close(null_fd);
891 	argv_array_clear(&argv);
892 	free(childenv);
893 }
894 end_of_spawn:
895 
896 #else
897 {
898 	int fhin = 0, fhout = 1, fherr = 2;
899 	const char **sargv = cmd->argv;
900 	struct argv_array nargv = ARGV_ARRAY_INIT;
901 
902 	if (cmd->no_stdin)
903 		fhin = open("/dev/null", O_RDWR);
904 	else if (need_in)
905 		fhin = dup(fdin[0]);
906 	else if (cmd->in)
907 		fhin = dup(cmd->in);
908 
909 	if (cmd->no_stderr)
910 		fherr = open("/dev/null", O_RDWR);
911 	else if (need_err)
912 		fherr = dup(fderr[1]);
913 	else if (cmd->err > 2)
914 		fherr = dup(cmd->err);
915 
916 	if (cmd->no_stdout)
917 		fhout = open("/dev/null", O_RDWR);
918 	else if (cmd->stdout_to_stderr)
919 		fhout = dup(fherr);
920 	else if (need_out)
921 		fhout = dup(fdout[1]);
922 	else if (cmd->out > 1)
923 		fhout = dup(cmd->out);
924 
925 	if (cmd->git_cmd)
926 		cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
927 	else if (cmd->use_shell)
928 		cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
929 
930 	cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
931 			cmd->dir, fhin, fhout, fherr);
932 	failed_errno = errno;
933 	if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
934 		error_errno("cannot spawn %s", cmd->argv[0]);
935 	if (cmd->clean_on_exit && cmd->pid >= 0)
936 		mark_child_for_cleanup(cmd->pid, cmd);
937 
938 	argv_array_clear(&nargv);
939 	cmd->argv = sargv;
940 	if (fhin != 0)
941 		close(fhin);
942 	if (fhout != 1)
943 		close(fhout);
944 	if (fherr != 2)
945 		close(fherr);
946 }
947 #endif
948 
949 	if (cmd->pid < 0) {
950 		trace2_child_exit(cmd, -1);
951 
952 		if (need_in)
953 			close_pair(fdin);
954 		else if (cmd->in)
955 			close(cmd->in);
956 		if (need_out)
957 			close_pair(fdout);
958 		else if (cmd->out)
959 			close(cmd->out);
960 		if (need_err)
961 			close_pair(fderr);
962 		else if (cmd->err)
963 			close(cmd->err);
964 		child_process_clear(cmd);
965 		errno = failed_errno;
966 		return -1;
967 	}
968 
969 	if (need_in)
970 		close(fdin[0]);
971 	else if (cmd->in)
972 		close(cmd->in);
973 
974 	if (need_out)
975 		close(fdout[1]);
976 	else if (cmd->out)
977 		close(cmd->out);
978 
979 	if (need_err)
980 		close(fderr[1]);
981 	else if (cmd->err)
982 		close(cmd->err);
983 
984 	return 0;
985 }
986 
finish_command(struct child_process * cmd)987 int finish_command(struct child_process *cmd)
988 {
989 	int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
990 	trace2_child_exit(cmd, ret);
991 	child_process_clear(cmd);
992 	return ret;
993 }
994 
finish_command_in_signal(struct child_process * cmd)995 int finish_command_in_signal(struct child_process *cmd)
996 {
997 	int ret = wait_or_whine(cmd->pid, cmd->argv[0], 1);
998 	trace2_child_exit(cmd, ret);
999 	return ret;
1000 }
1001 
1002 
run_command(struct child_process * cmd)1003 int run_command(struct child_process *cmd)
1004 {
1005 	int code;
1006 
1007 	if (cmd->out < 0 || cmd->err < 0)
1008 		BUG("run_command with a pipe can cause deadlock");
1009 
1010 	code = start_command(cmd);
1011 	if (code)
1012 		return code;
1013 	return finish_command(cmd);
1014 }
1015 
run_command_v_opt(const char ** argv,int opt)1016 int run_command_v_opt(const char **argv, int opt)
1017 {
1018 	return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
1019 }
1020 
run_command_v_opt_tr2(const char ** argv,int opt,const char * tr2_class)1021 int run_command_v_opt_tr2(const char **argv, int opt, const char *tr2_class)
1022 {
1023 	return run_command_v_opt_cd_env_tr2(argv, opt, NULL, NULL, tr2_class);
1024 }
1025 
run_command_v_opt_cd_env(const char ** argv,int opt,const char * dir,const char * const * env)1026 int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
1027 {
1028 	return run_command_v_opt_cd_env_tr2(argv, opt, dir, env, NULL);
1029 }
1030 
run_command_v_opt_cd_env_tr2(const char ** argv,int opt,const char * dir,const char * const * env,const char * tr2_class)1031 int run_command_v_opt_cd_env_tr2(const char **argv, int opt, const char *dir,
1032 				 const char *const *env, const char *tr2_class)
1033 {
1034 	struct child_process cmd = CHILD_PROCESS_INIT;
1035 	cmd.argv = argv;
1036 	cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
1037 	cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
1038 	cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
1039 	cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
1040 	cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
1041 	cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
1042 	cmd.dir = dir;
1043 	cmd.env = env;
1044 	cmd.trace2_child_class = tr2_class;
1045 	return run_command(&cmd);
1046 }
1047 
1048 #ifndef NO_PTHREADS
1049 static pthread_t main_thread;
1050 static int main_thread_set;
1051 static pthread_key_t async_key;
1052 static pthread_key_t async_die_counter;
1053 
run_thread(void * data)1054 static void *run_thread(void *data)
1055 {
1056 	struct async *async = data;
1057 	intptr_t ret;
1058 
1059 	if (async->isolate_sigpipe) {
1060 		sigset_t mask;
1061 		sigemptyset(&mask);
1062 		sigaddset(&mask, SIGPIPE);
1063 		if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
1064 			ret = error("unable to block SIGPIPE in async thread");
1065 			return (void *)ret;
1066 		}
1067 	}
1068 
1069 	pthread_setspecific(async_key, async);
1070 	ret = async->proc(async->proc_in, async->proc_out, async->data);
1071 	return (void *)ret;
1072 }
1073 
die_async(const char * err,va_list params)1074 static NORETURN void die_async(const char *err, va_list params)
1075 {
1076 	vreportf("fatal: ", err, params);
1077 
1078 	if (in_async()) {
1079 		struct async *async = pthread_getspecific(async_key);
1080 		if (async->proc_in >= 0)
1081 			close(async->proc_in);
1082 		if (async->proc_out >= 0)
1083 			close(async->proc_out);
1084 		pthread_exit((void *)128);
1085 	}
1086 
1087 	exit(128);
1088 }
1089 
async_die_is_recursing(void)1090 static int async_die_is_recursing(void)
1091 {
1092 	void *ret = pthread_getspecific(async_die_counter);
1093 	pthread_setspecific(async_die_counter, (void *)1);
1094 	return ret != NULL;
1095 }
1096 
in_async(void)1097 int in_async(void)
1098 {
1099 	if (!main_thread_set)
1100 		return 0; /* no asyncs started yet */
1101 	return !pthread_equal(main_thread, pthread_self());
1102 }
1103 
async_exit(int code)1104 static void NORETURN async_exit(int code)
1105 {
1106 	pthread_exit((void *)(intptr_t)code);
1107 }
1108 
1109 #else
1110 
1111 static struct {
1112 	void (**handlers)(void);
1113 	size_t nr;
1114 	size_t alloc;
1115 } git_atexit_hdlrs;
1116 
1117 static int git_atexit_installed;
1118 
git_atexit_dispatch(void)1119 static void git_atexit_dispatch(void)
1120 {
1121 	size_t i;
1122 
1123 	for (i=git_atexit_hdlrs.nr ; i ; i--)
1124 		git_atexit_hdlrs.handlers[i-1]();
1125 }
1126 
git_atexit_clear(void)1127 static void git_atexit_clear(void)
1128 {
1129 	free(git_atexit_hdlrs.handlers);
1130 	memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
1131 	git_atexit_installed = 0;
1132 }
1133 
1134 #undef atexit
git_atexit(void (* handler)(void))1135 int git_atexit(void (*handler)(void))
1136 {
1137 	ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
1138 	git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
1139 	if (!git_atexit_installed) {
1140 		if (atexit(&git_atexit_dispatch))
1141 			return -1;
1142 		git_atexit_installed = 1;
1143 	}
1144 	return 0;
1145 }
1146 #define atexit git_atexit
1147 
1148 static int process_is_async;
in_async(void)1149 int in_async(void)
1150 {
1151 	return process_is_async;
1152 }
1153 
async_exit(int code)1154 static void NORETURN async_exit(int code)
1155 {
1156 	exit(code);
1157 }
1158 
1159 #endif
1160 
check_pipe(int err)1161 void check_pipe(int err)
1162 {
1163 	if (err == EPIPE) {
1164 		if (in_async())
1165 			async_exit(141);
1166 
1167 		signal(SIGPIPE, SIG_DFL);
1168 		raise(SIGPIPE);
1169 		/* Should never happen, but just in case... */
1170 		exit(141);
1171 	}
1172 }
1173 
start_async(struct async * async)1174 int start_async(struct async *async)
1175 {
1176 	int need_in, need_out;
1177 	int fdin[2], fdout[2];
1178 	int proc_in, proc_out;
1179 
1180 	need_in = async->in < 0;
1181 	if (need_in) {
1182 		if (pipe(fdin) < 0) {
1183 			if (async->out > 0)
1184 				close(async->out);
1185 			return error_errno("cannot create pipe");
1186 		}
1187 		async->in = fdin[1];
1188 	}
1189 
1190 	need_out = async->out < 0;
1191 	if (need_out) {
1192 		if (pipe(fdout) < 0) {
1193 			if (need_in)
1194 				close_pair(fdin);
1195 			else if (async->in)
1196 				close(async->in);
1197 			return error_errno("cannot create pipe");
1198 		}
1199 		async->out = fdout[0];
1200 	}
1201 
1202 	if (need_in)
1203 		proc_in = fdin[0];
1204 	else if (async->in)
1205 		proc_in = async->in;
1206 	else
1207 		proc_in = -1;
1208 
1209 	if (need_out)
1210 		proc_out = fdout[1];
1211 	else if (async->out)
1212 		proc_out = async->out;
1213 	else
1214 		proc_out = -1;
1215 
1216 #ifdef NO_PTHREADS
1217 	/* Flush stdio before fork() to avoid cloning buffers */
1218 	fflush(NULL);
1219 
1220 	async->pid = fork();
1221 	if (async->pid < 0) {
1222 		error_errno("fork (async) failed");
1223 		goto error;
1224 	}
1225 	if (!async->pid) {
1226 		if (need_in)
1227 			close(fdin[1]);
1228 		if (need_out)
1229 			close(fdout[0]);
1230 		git_atexit_clear();
1231 		process_is_async = 1;
1232 		exit(!!async->proc(proc_in, proc_out, async->data));
1233 	}
1234 
1235 	mark_child_for_cleanup(async->pid, NULL);
1236 
1237 	if (need_in)
1238 		close(fdin[0]);
1239 	else if (async->in)
1240 		close(async->in);
1241 
1242 	if (need_out)
1243 		close(fdout[1]);
1244 	else if (async->out)
1245 		close(async->out);
1246 #else
1247 	if (!main_thread_set) {
1248 		/*
1249 		 * We assume that the first time that start_async is called
1250 		 * it is from the main thread.
1251 		 */
1252 		main_thread_set = 1;
1253 		main_thread = pthread_self();
1254 		pthread_key_create(&async_key, NULL);
1255 		pthread_key_create(&async_die_counter, NULL);
1256 		set_die_routine(die_async);
1257 		set_die_is_recursing_routine(async_die_is_recursing);
1258 	}
1259 
1260 	if (proc_in >= 0)
1261 		set_cloexec(proc_in);
1262 	if (proc_out >= 0)
1263 		set_cloexec(proc_out);
1264 	async->proc_in = proc_in;
1265 	async->proc_out = proc_out;
1266 	{
1267 		int err = pthread_create(&async->tid, NULL, run_thread, async);
1268 		if (err) {
1269 			error(_("cannot create async thread: %s"), strerror(err));
1270 			goto error;
1271 		}
1272 	}
1273 #endif
1274 	return 0;
1275 
1276 error:
1277 	if (need_in)
1278 		close_pair(fdin);
1279 	else if (async->in)
1280 		close(async->in);
1281 
1282 	if (need_out)
1283 		close_pair(fdout);
1284 	else if (async->out)
1285 		close(async->out);
1286 	return -1;
1287 }
1288 
finish_async(struct async * async)1289 int finish_async(struct async *async)
1290 {
1291 #ifdef NO_PTHREADS
1292 	return wait_or_whine(async->pid, "child process", 0);
1293 #else
1294 	void *ret = (void *)(intptr_t)(-1);
1295 
1296 	if (pthread_join(async->tid, &ret))
1297 		error("pthread_join failed");
1298 	return (int)(intptr_t)ret;
1299 #endif
1300 }
1301 
async_with_fork(void)1302 int async_with_fork(void)
1303 {
1304 #ifdef NO_PTHREADS
1305 	return 1;
1306 #else
1307 	return 0;
1308 #endif
1309 }
1310 
find_hook(const char * name)1311 const char *find_hook(const char *name)
1312 {
1313 	static struct strbuf path = STRBUF_INIT;
1314 
1315 	strbuf_reset(&path);
1316 	strbuf_git_path(&path, "hooks/%s", name);
1317 	if (access(path.buf, X_OK) < 0) {
1318 		int err = errno;
1319 
1320 #ifdef STRIP_EXTENSION
1321 		strbuf_addstr(&path, STRIP_EXTENSION);
1322 		if (access(path.buf, X_OK) >= 0)
1323 			return path.buf;
1324 		if (errno == EACCES)
1325 			err = errno;
1326 #endif
1327 
1328 		if (err == EACCES && advice_ignored_hook) {
1329 			static struct string_list advise_given = STRING_LIST_INIT_DUP;
1330 
1331 			if (!string_list_lookup(&advise_given, name)) {
1332 				string_list_insert(&advise_given, name);
1333 				advise(_("The '%s' hook was ignored because "
1334 					 "it's not set as executable.\n"
1335 					 "You can disable this warning with "
1336 					 "`git config advice.ignoredHook false`."),
1337 				       path.buf);
1338 			}
1339 		}
1340 		return NULL;
1341 	}
1342 	return path.buf;
1343 }
1344 
run_hook_ve(const char * const * env,const char * name,va_list args)1345 int run_hook_ve(const char *const *env, const char *name, va_list args)
1346 {
1347 	struct child_process hook = CHILD_PROCESS_INIT;
1348 	const char *p;
1349 
1350 	p = find_hook(name);
1351 	if (!p)
1352 		return 0;
1353 
1354 	argv_array_push(&hook.args, p);
1355 	while ((p = va_arg(args, const char *)))
1356 		argv_array_push(&hook.args, p);
1357 	hook.env = env;
1358 	hook.no_stdin = 1;
1359 	hook.stdout_to_stderr = 1;
1360 	hook.trace2_hook_name = name;
1361 
1362 	return run_command(&hook);
1363 }
1364 
run_hook_le(const char * const * env,const char * name,...)1365 int run_hook_le(const char *const *env, const char *name, ...)
1366 {
1367 	va_list args;
1368 	int ret;
1369 
1370 	va_start(args, name);
1371 	ret = run_hook_ve(env, name, args);
1372 	va_end(args);
1373 
1374 	return ret;
1375 }
1376 
1377 struct io_pump {
1378 	/* initialized by caller */
1379 	int fd;
1380 	int type; /* POLLOUT or POLLIN */
1381 	union {
1382 		struct {
1383 			const char *buf;
1384 			size_t len;
1385 		} out;
1386 		struct {
1387 			struct strbuf *buf;
1388 			size_t hint;
1389 		} in;
1390 	} u;
1391 
1392 	/* returned by pump_io */
1393 	int error; /* 0 for success, otherwise errno */
1394 
1395 	/* internal use */
1396 	struct pollfd *pfd;
1397 };
1398 
pump_io_round(struct io_pump * slots,int nr,struct pollfd * pfd)1399 static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
1400 {
1401 	int pollsize = 0;
1402 	int i;
1403 
1404 	for (i = 0; i < nr; i++) {
1405 		struct io_pump *io = &slots[i];
1406 		if (io->fd < 0)
1407 			continue;
1408 		pfd[pollsize].fd = io->fd;
1409 		pfd[pollsize].events = io->type;
1410 		io->pfd = &pfd[pollsize++];
1411 	}
1412 
1413 	if (!pollsize)
1414 		return 0;
1415 
1416 	if (poll(pfd, pollsize, -1) < 0) {
1417 		if (errno == EINTR)
1418 			return 1;
1419 		die_errno("poll failed");
1420 	}
1421 
1422 	for (i = 0; i < nr; i++) {
1423 		struct io_pump *io = &slots[i];
1424 
1425 		if (io->fd < 0)
1426 			continue;
1427 
1428 		if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
1429 			continue;
1430 
1431 		if (io->type == POLLOUT) {
1432 			ssize_t len = xwrite(io->fd,
1433 					     io->u.out.buf, io->u.out.len);
1434 			if (len < 0) {
1435 				io->error = errno;
1436 				close(io->fd);
1437 				io->fd = -1;
1438 			} else {
1439 				io->u.out.buf += len;
1440 				io->u.out.len -= len;
1441 				if (!io->u.out.len) {
1442 					close(io->fd);
1443 					io->fd = -1;
1444 				}
1445 			}
1446 		}
1447 
1448 		if (io->type == POLLIN) {
1449 			ssize_t len = strbuf_read_once(io->u.in.buf,
1450 						       io->fd, io->u.in.hint);
1451 			if (len < 0)
1452 				io->error = errno;
1453 			if (len <= 0) {
1454 				close(io->fd);
1455 				io->fd = -1;
1456 			}
1457 		}
1458 	}
1459 
1460 	return 1;
1461 }
1462 
pump_io(struct io_pump * slots,int nr)1463 static int pump_io(struct io_pump *slots, int nr)
1464 {
1465 	struct pollfd *pfd;
1466 	int i;
1467 
1468 	for (i = 0; i < nr; i++)
1469 		slots[i].error = 0;
1470 
1471 	ALLOC_ARRAY(pfd, nr);
1472 	while (pump_io_round(slots, nr, pfd))
1473 		; /* nothing */
1474 	free(pfd);
1475 
1476 	/* There may be multiple errno values, so just pick the first. */
1477 	for (i = 0; i < nr; i++) {
1478 		if (slots[i].error) {
1479 			errno = slots[i].error;
1480 			return -1;
1481 		}
1482 	}
1483 	return 0;
1484 }
1485 
1486 
pipe_command(struct child_process * cmd,const char * in,size_t in_len,struct strbuf * out,size_t out_hint,struct strbuf * err,size_t err_hint)1487 int pipe_command(struct child_process *cmd,
1488 		 const char *in, size_t in_len,
1489 		 struct strbuf *out, size_t out_hint,
1490 		 struct strbuf *err, size_t err_hint)
1491 {
1492 	struct io_pump io[3];
1493 	int nr = 0;
1494 
1495 	if (in)
1496 		cmd->in = -1;
1497 	if (out)
1498 		cmd->out = -1;
1499 	if (err)
1500 		cmd->err = -1;
1501 
1502 	if (start_command(cmd) < 0)
1503 		return -1;
1504 
1505 	if (in) {
1506 		io[nr].fd = cmd->in;
1507 		io[nr].type = POLLOUT;
1508 		io[nr].u.out.buf = in;
1509 		io[nr].u.out.len = in_len;
1510 		nr++;
1511 	}
1512 	if (out) {
1513 		io[nr].fd = cmd->out;
1514 		io[nr].type = POLLIN;
1515 		io[nr].u.in.buf = out;
1516 		io[nr].u.in.hint = out_hint;
1517 		nr++;
1518 	}
1519 	if (err) {
1520 		io[nr].fd = cmd->err;
1521 		io[nr].type = POLLIN;
1522 		io[nr].u.in.buf = err;
1523 		io[nr].u.in.hint = err_hint;
1524 		nr++;
1525 	}
1526 
1527 	if (pump_io(io, nr) < 0) {
1528 		finish_command(cmd); /* throw away exit code */
1529 		return -1;
1530 	}
1531 
1532 	return finish_command(cmd);
1533 }
1534 
1535 enum child_state {
1536 	GIT_CP_FREE,
1537 	GIT_CP_WORKING,
1538 	GIT_CP_WAIT_CLEANUP,
1539 };
1540 
1541 struct parallel_processes {
1542 	void *data;
1543 
1544 	int max_processes;
1545 	int nr_processes;
1546 
1547 	get_next_task_fn get_next_task;
1548 	start_failure_fn start_failure;
1549 	task_finished_fn task_finished;
1550 
1551 	struct {
1552 		enum child_state state;
1553 		struct child_process process;
1554 		struct strbuf err;
1555 		void *data;
1556 	} *children;
1557 	/*
1558 	 * The struct pollfd is logically part of *children,
1559 	 * but the system call expects it as its own array.
1560 	 */
1561 	struct pollfd *pfd;
1562 
1563 	unsigned shutdown : 1;
1564 
1565 	int output_owner;
1566 	struct strbuf buffered_output; /* of finished children */
1567 };
1568 
default_start_failure(struct strbuf * out,void * pp_cb,void * pp_task_cb)1569 static int default_start_failure(struct strbuf *out,
1570 				 void *pp_cb,
1571 				 void *pp_task_cb)
1572 {
1573 	return 0;
1574 }
1575 
default_task_finished(int result,struct strbuf * out,void * pp_cb,void * pp_task_cb)1576 static int default_task_finished(int result,
1577 				 struct strbuf *out,
1578 				 void *pp_cb,
1579 				 void *pp_task_cb)
1580 {
1581 	return 0;
1582 }
1583 
kill_children(struct parallel_processes * pp,int signo)1584 static void kill_children(struct parallel_processes *pp, int signo)
1585 {
1586 	int i, n = pp->max_processes;
1587 
1588 	for (i = 0; i < n; i++)
1589 		if (pp->children[i].state == GIT_CP_WORKING)
1590 			kill(pp->children[i].process.pid, signo);
1591 }
1592 
1593 static struct parallel_processes *pp_for_signal;
1594 
handle_children_on_signal(int signo)1595 static void handle_children_on_signal(int signo)
1596 {
1597 	kill_children(pp_for_signal, signo);
1598 	sigchain_pop(signo);
1599 	raise(signo);
1600 }
1601 
pp_init(struct parallel_processes * pp,int n,get_next_task_fn get_next_task,start_failure_fn start_failure,task_finished_fn task_finished,void * data)1602 static void pp_init(struct parallel_processes *pp,
1603 		    int n,
1604 		    get_next_task_fn get_next_task,
1605 		    start_failure_fn start_failure,
1606 		    task_finished_fn task_finished,
1607 		    void *data)
1608 {
1609 	int i;
1610 
1611 	if (n < 1)
1612 		n = online_cpus();
1613 
1614 	pp->max_processes = n;
1615 
1616 	trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
1617 
1618 	pp->data = data;
1619 	if (!get_next_task)
1620 		BUG("you need to specify a get_next_task function");
1621 	pp->get_next_task = get_next_task;
1622 
1623 	pp->start_failure = start_failure ? start_failure : default_start_failure;
1624 	pp->task_finished = task_finished ? task_finished : default_task_finished;
1625 
1626 	pp->nr_processes = 0;
1627 	pp->output_owner = 0;
1628 	pp->shutdown = 0;
1629 	pp->children = xcalloc(n, sizeof(*pp->children));
1630 	pp->pfd = xcalloc(n, sizeof(*pp->pfd));
1631 	strbuf_init(&pp->buffered_output, 0);
1632 
1633 	for (i = 0; i < n; i++) {
1634 		strbuf_init(&pp->children[i].err, 0);
1635 		child_process_init(&pp->children[i].process);
1636 		pp->pfd[i].events = POLLIN | POLLHUP;
1637 		pp->pfd[i].fd = -1;
1638 	}
1639 
1640 	pp_for_signal = pp;
1641 	sigchain_push_common(handle_children_on_signal);
1642 }
1643 
pp_cleanup(struct parallel_processes * pp)1644 static void pp_cleanup(struct parallel_processes *pp)
1645 {
1646 	int i;
1647 
1648 	trace_printf("run_processes_parallel: done");
1649 	for (i = 0; i < pp->max_processes; i++) {
1650 		strbuf_release(&pp->children[i].err);
1651 		child_process_clear(&pp->children[i].process);
1652 	}
1653 
1654 	free(pp->children);
1655 	free(pp->pfd);
1656 
1657 	/*
1658 	 * When get_next_task added messages to the buffer in its last
1659 	 * iteration, the buffered output is non empty.
1660 	 */
1661 	strbuf_write(&pp->buffered_output, stderr);
1662 	strbuf_release(&pp->buffered_output);
1663 
1664 	sigchain_pop_common();
1665 }
1666 
1667 /* returns
1668  *  0 if a new task was started.
1669  *  1 if no new jobs was started (get_next_task ran out of work, non critical
1670  *    problem with starting a new command)
1671  * <0 no new job was started, user wishes to shutdown early. Use negative code
1672  *    to signal the children.
1673  */
pp_start_one(struct parallel_processes * pp)1674 static int pp_start_one(struct parallel_processes *pp)
1675 {
1676 	int i, code;
1677 
1678 	for (i = 0; i < pp->max_processes; i++)
1679 		if (pp->children[i].state == GIT_CP_FREE)
1680 			break;
1681 	if (i == pp->max_processes)
1682 		BUG("bookkeeping is hard");
1683 
1684 	code = pp->get_next_task(&pp->children[i].process,
1685 				 &pp->children[i].err,
1686 				 pp->data,
1687 				 &pp->children[i].data);
1688 	if (!code) {
1689 		strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1690 		strbuf_reset(&pp->children[i].err);
1691 		return 1;
1692 	}
1693 	pp->children[i].process.err = -1;
1694 	pp->children[i].process.stdout_to_stderr = 1;
1695 	pp->children[i].process.no_stdin = 1;
1696 
1697 	if (start_command(&pp->children[i].process)) {
1698 		code = pp->start_failure(&pp->children[i].err,
1699 					 pp->data,
1700 					 pp->children[i].data);
1701 		strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1702 		strbuf_reset(&pp->children[i].err);
1703 		if (code)
1704 			pp->shutdown = 1;
1705 		return code;
1706 	}
1707 
1708 	pp->nr_processes++;
1709 	pp->children[i].state = GIT_CP_WORKING;
1710 	pp->pfd[i].fd = pp->children[i].process.err;
1711 	return 0;
1712 }
1713 
pp_buffer_stderr(struct parallel_processes * pp,int output_timeout)1714 static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
1715 {
1716 	int i;
1717 
1718 	while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
1719 		if (errno == EINTR)
1720 			continue;
1721 		pp_cleanup(pp);
1722 		die_errno("poll");
1723 	}
1724 
1725 	/* Buffer output from all pipes. */
1726 	for (i = 0; i < pp->max_processes; i++) {
1727 		if (pp->children[i].state == GIT_CP_WORKING &&
1728 		    pp->pfd[i].revents & (POLLIN | POLLHUP)) {
1729 			int n = strbuf_read_once(&pp->children[i].err,
1730 						 pp->children[i].process.err, 0);
1731 			if (n == 0) {
1732 				close(pp->children[i].process.err);
1733 				pp->children[i].state = GIT_CP_WAIT_CLEANUP;
1734 			} else if (n < 0)
1735 				if (errno != EAGAIN)
1736 					die_errno("read");
1737 		}
1738 	}
1739 }
1740 
pp_output(struct parallel_processes * pp)1741 static void pp_output(struct parallel_processes *pp)
1742 {
1743 	int i = pp->output_owner;
1744 	if (pp->children[i].state == GIT_CP_WORKING &&
1745 	    pp->children[i].err.len) {
1746 		strbuf_write(&pp->children[i].err, stderr);
1747 		strbuf_reset(&pp->children[i].err);
1748 	}
1749 }
1750 
pp_collect_finished(struct parallel_processes * pp)1751 static int pp_collect_finished(struct parallel_processes *pp)
1752 {
1753 	int i, code;
1754 	int n = pp->max_processes;
1755 	int result = 0;
1756 
1757 	while (pp->nr_processes > 0) {
1758 		for (i = 0; i < pp->max_processes; i++)
1759 			if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
1760 				break;
1761 		if (i == pp->max_processes)
1762 			break;
1763 
1764 		code = finish_command(&pp->children[i].process);
1765 
1766 		code = pp->task_finished(code,
1767 					 &pp->children[i].err, pp->data,
1768 					 pp->children[i].data);
1769 
1770 		if (code)
1771 			result = code;
1772 		if (code < 0)
1773 			break;
1774 
1775 		pp->nr_processes--;
1776 		pp->children[i].state = GIT_CP_FREE;
1777 		pp->pfd[i].fd = -1;
1778 		child_process_init(&pp->children[i].process);
1779 
1780 		if (i != pp->output_owner) {
1781 			strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
1782 			strbuf_reset(&pp->children[i].err);
1783 		} else {
1784 			strbuf_write(&pp->children[i].err, stderr);
1785 			strbuf_reset(&pp->children[i].err);
1786 
1787 			/* Output all other finished child processes */
1788 			strbuf_write(&pp->buffered_output, stderr);
1789 			strbuf_reset(&pp->buffered_output);
1790 
1791 			/*
1792 			 * Pick next process to output live.
1793 			 * NEEDSWORK:
1794 			 * For now we pick it randomly by doing a round
1795 			 * robin. Later we may want to pick the one with
1796 			 * the most output or the longest or shortest
1797 			 * running process time.
1798 			 */
1799 			for (i = 0; i < n; i++)
1800 				if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
1801 					break;
1802 			pp->output_owner = (pp->output_owner + i) % n;
1803 		}
1804 	}
1805 	return result;
1806 }
1807 
run_processes_parallel(int n,get_next_task_fn get_next_task,start_failure_fn start_failure,task_finished_fn task_finished,void * pp_cb)1808 int run_processes_parallel(int n,
1809 			   get_next_task_fn get_next_task,
1810 			   start_failure_fn start_failure,
1811 			   task_finished_fn task_finished,
1812 			   void *pp_cb)
1813 {
1814 	int i, code;
1815 	int output_timeout = 100;
1816 	int spawn_cap = 4;
1817 	struct parallel_processes pp;
1818 
1819 	pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
1820 	while (1) {
1821 		for (i = 0;
1822 		    i < spawn_cap && !pp.shutdown &&
1823 		    pp.nr_processes < pp.max_processes;
1824 		    i++) {
1825 			code = pp_start_one(&pp);
1826 			if (!code)
1827 				continue;
1828 			if (code < 0) {
1829 				pp.shutdown = 1;
1830 				kill_children(&pp, -code);
1831 			}
1832 			break;
1833 		}
1834 		if (!pp.nr_processes)
1835 			break;
1836 		pp_buffer_stderr(&pp, output_timeout);
1837 		pp_output(&pp);
1838 		code = pp_collect_finished(&pp);
1839 		if (code) {
1840 			pp.shutdown = 1;
1841 			if (code < 0)
1842 				kill_children(&pp, -code);
1843 		}
1844 	}
1845 
1846 	pp_cleanup(&pp);
1847 	return 0;
1848 }
1849 
run_processes_parallel_tr2(int n,get_next_task_fn get_next_task,start_failure_fn start_failure,task_finished_fn task_finished,void * pp_cb,const char * tr2_category,const char * tr2_label)1850 int run_processes_parallel_tr2(int n, get_next_task_fn get_next_task,
1851 			       start_failure_fn start_failure,
1852 			       task_finished_fn task_finished, void *pp_cb,
1853 			       const char *tr2_category, const char *tr2_label)
1854 {
1855 	int result;
1856 
1857 	trace2_region_enter_printf(tr2_category, tr2_label, NULL, "max:%d",
1858 				   ((n < 1) ? online_cpus() : n));
1859 
1860 	result = run_processes_parallel(n, get_next_task, start_failure,
1861 					task_finished, pp_cb);
1862 
1863 	trace2_region_leave(tr2_category, tr2_label, NULL);
1864 
1865 	return result;
1866 }
1867