apply: move libified code from builtin/apply.c to apply.{c,h}
[gitweb.git] / run-command.c
index bbb9c777e583c345d25a6651f9ddf7725c10f6af..33bc63a1de8d41b52a0ec353cf80b0f2e65dc8fd 100644 (file)
@@ -1,6 +1,82 @@
 #include "cache.h"
 #include "run-command.h"
 #include "exec_cmd.h"
+#include "sigchain.h"
+#include "argv-array.h"
+#include "thread-utils.h"
+#include "strbuf.h"
+
+void child_process_init(struct child_process *child)
+{
+       memset(child, 0, sizeof(*child));
+       argv_array_init(&child->args);
+       argv_array_init(&child->env_array);
+}
+
+void child_process_clear(struct child_process *child)
+{
+       argv_array_clear(&child->args);
+       argv_array_clear(&child->env_array);
+}
+
+struct child_to_clean {
+       pid_t pid;
+       struct child_to_clean *next;
+};
+static struct child_to_clean *children_to_clean;
+static int installed_child_cleanup_handler;
+
+static void cleanup_children(int sig, int in_signal)
+{
+       while (children_to_clean) {
+               struct child_to_clean *p = children_to_clean;
+               children_to_clean = p->next;
+               kill(p->pid, sig);
+               if (!in_signal)
+                       free(p);
+       }
+}
+
+static void cleanup_children_on_signal(int sig)
+{
+       cleanup_children(sig, 1);
+       sigchain_pop(sig);
+       raise(sig);
+}
+
+static void cleanup_children_on_exit(void)
+{
+       cleanup_children(SIGTERM, 0);
+}
+
+static void mark_child_for_cleanup(pid_t pid)
+{
+       struct child_to_clean *p = xmalloc(sizeof(*p));
+       p->pid = pid;
+       p->next = children_to_clean;
+       children_to_clean = p;
+
+       if (!installed_child_cleanup_handler) {
+               atexit(cleanup_children_on_exit);
+               sigchain_push_common(cleanup_children_on_signal);
+               installed_child_cleanup_handler = 1;
+       }
+}
+
+static void clear_child_for_cleanup(pid_t pid)
+{
+       struct child_to_clean **pp;
+
+       for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
+               struct child_to_clean *clean_me = *pp;
+
+               if (clean_me->pid == pid) {
+                       *pp = clean_me->next;
+                       free(clean_me);
+                       return;
+               }
+       }
+}
 
 static inline void close_pair(int fd[2])
 {
@@ -8,17 +84,198 @@ static inline void close_pair(int fd[2])
        close(fd[1]);
 }
 
+#ifndef GIT_WINDOWS_NATIVE
 static inline void dup_devnull(int to)
 {
        int fd = open("/dev/null", O_RDWR);
-       dup2(fd, to);
+       if (fd < 0)
+               die_errno(_("open /dev/null failed"));
+       if (dup2(fd, to) < 0)
+               die_errno(_("dup2(%d,%d) failed"), fd, to);
        close(fd);
 }
+#endif
+
+static char *locate_in_PATH(const char *file)
+{
+       const char *p = getenv("PATH");
+       struct strbuf buf = STRBUF_INIT;
+
+       if (!p || !*p)
+               return NULL;
+
+       while (1) {
+               const char *end = strchrnul(p, ':');
+
+               strbuf_reset(&buf);
+
+               /* POSIX specifies an empty entry as the current directory. */
+               if (end != p) {
+                       strbuf_add(&buf, p, end - p);
+                       strbuf_addch(&buf, '/');
+               }
+               strbuf_addstr(&buf, file);
+
+               if (!access(buf.buf, F_OK))
+                       return strbuf_detach(&buf, NULL);
+
+               if (!*end)
+                       break;
+               p = end + 1;
+       }
+
+       strbuf_release(&buf);
+       return NULL;
+}
+
+static int exists_in_PATH(const char *file)
+{
+       char *r = locate_in_PATH(file);
+       free(r);
+       return r != NULL;
+}
+
+int sane_execvp(const char *file, char * const argv[])
+{
+       if (!execvp(file, argv))
+               return 0; /* cannot happen ;-) */
+
+       /*
+        * When a command can't be found because one of the directories
+        * listed in $PATH is unsearchable, execvp reports EACCES, but
+        * careful usability testing (read: analysis of occasional bug
+        * reports) reveals that "No such file or directory" is more
+        * intuitive.
+        *
+        * We avoid commands with "/", because execvp will not do $PATH
+        * lookups in that case.
+        *
+        * The reassignment of EACCES to errno looks like a no-op below,
+        * but we need to protect against exists_in_PATH overwriting errno.
+        */
+       if (errno == EACCES && !strchr(file, '/'))
+               errno = exists_in_PATH(file) ? EACCES : ENOENT;
+       else if (errno == ENOTDIR && !strchr(file, '/'))
+               errno = ENOENT;
+       return -1;
+}
+
+static const char **prepare_shell_cmd(struct argv_array *out, const char **argv)
+{
+       if (!argv[0])
+               die("BUG: shell command is empty");
+
+       if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
+#ifndef GIT_WINDOWS_NATIVE
+               argv_array_push(out, SHELL_PATH);
+#else
+               argv_array_push(out, "sh");
+#endif
+               argv_array_push(out, "-c");
+
+               /*
+                * If we have no extra arguments, we do not even need to
+                * bother with the "$@" magic.
+                */
+               if (!argv[1])
+                       argv_array_push(out, argv[0]);
+               else
+                       argv_array_pushf(out, "%s \"$@\"", argv[0]);
+       }
+
+       argv_array_pushv(out, argv);
+       return out->argv;
+}
+
+#ifndef GIT_WINDOWS_NATIVE
+static int execv_shell_cmd(const char **argv)
+{
+       struct argv_array nargv = ARGV_ARRAY_INIT;
+       prepare_shell_cmd(&nargv, argv);
+       trace_argv_printf(nargv.argv, "trace: exec:");
+       sane_execvp(nargv.argv[0], (char **)nargv.argv);
+       argv_array_clear(&nargv);
+       return -1;
+}
+#endif
+
+#ifndef GIT_WINDOWS_NATIVE
+static int child_notifier = -1;
+
+static void notify_parent(void)
+{
+       /*
+        * execvp failed.  If possible, we'd like to let start_command
+        * know, so failures like ENOENT can be handled right away; but
+        * otherwise, finish_command will still report the error.
+        */
+       xwrite(child_notifier, "", 1);
+}
+#endif
+
+static inline void set_cloexec(int fd)
+{
+       int flags = fcntl(fd, F_GETFD);
+       if (flags >= 0)
+               fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
+}
+
+static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
+{
+       int status, code = -1;
+       pid_t waiting;
+       int failed_errno = 0;
+
+       while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
+               ;       /* nothing */
+       if (in_signal)
+               return 0;
+
+       if (waiting < 0) {
+               failed_errno = errno;
+               error_errno("waitpid for %s failed", argv0);
+       } else if (waiting != pid) {
+               error("waitpid is confused (%s)", argv0);
+       } else if (WIFSIGNALED(status)) {
+               code = WTERMSIG(status);
+               if (code != SIGINT && code != SIGQUIT && code != SIGPIPE)
+                       error("%s died of signal %d", argv0, code);
+               /*
+                * This return value is chosen so that code & 0xff
+                * mimics the exit code that a POSIX shell would report for
+                * a program that died from this signal.
+                */
+               code += 128;
+       } else if (WIFEXITED(status)) {
+               code = WEXITSTATUS(status);
+               /*
+                * Convert special exit code when execvp failed.
+                */
+               if (code == 127) {
+                       code = -1;
+                       failed_errno = ENOENT;
+               }
+       } else {
+               error("waitpid is confused (%s)", argv0);
+       }
+
+       clear_child_for_cleanup(pid);
+
+       errno = failed_errno;
+       return code;
+}
 
 int start_command(struct child_process *cmd)
 {
        int need_in, need_out, need_err;
        int fdin[2], fdout[2], fderr[2];
+       int failed_errno;
+       char *str;
+
+       if (!cmd->argv)
+               cmd->argv = cmd->args.argv;
+       if (!cmd->env)
+               cmd->env = cmd->env_array.argv;
 
        /*
         * In case of errors we must keep the promise to close FDs
@@ -28,9 +285,11 @@ int start_command(struct child_process *cmd)
        need_in = !cmd->no_stdin && cmd->in < 0;
        if (need_in) {
                if (pipe(fdin) < 0) {
+                       failed_errno = errno;
                        if (cmd->out > 0)
                                close(cmd->out);
-                       return -ERR_RUN_COMMAND_PIPE;
+                       str = "standard input";
+                       goto fail_pipe;
                }
                cmd->in = fdin[1];
        }
@@ -40,11 +299,13 @@ int start_command(struct child_process *cmd)
                && cmd->out < 0;
        if (need_out) {
                if (pipe(fdout) < 0) {
+                       failed_errno = errno;
                        if (need_in)
                                close_pair(fdin);
                        else if (cmd->in)
                                close(cmd->in);
-                       return -ERR_RUN_COMMAND_PIPE;
+                       str = "standard output";
+                       goto fail_pipe;
                }
                cmd->out = fdout[0];
        }
@@ -52,6 +313,7 @@ int start_command(struct child_process *cmd)
        need_err = !cmd->no_stderr && cmd->err < 0;
        if (need_err) {
                if (pipe(fderr) < 0) {
+                       failed_errno = errno;
                        if (need_in)
                                close_pair(fdin);
                        else if (cmd->in)
@@ -60,17 +322,45 @@ int start_command(struct child_process *cmd)
                                close_pair(fdout);
                        else if (cmd->out)
                                close(cmd->out);
-                       return -ERR_RUN_COMMAND_PIPE;
+                       str = "standard error";
+fail_pipe:
+                       error("cannot create %s pipe for %s: %s",
+                               str, cmd->argv[0], strerror(failed_errno));
+                       child_process_clear(cmd);
+                       errno = failed_errno;
+                       return -1;
                }
                cmd->err = fderr[0];
        }
 
        trace_argv_printf(cmd->argv, "trace: run_command:");
-
-#ifndef __MINGW32__
        fflush(NULL);
+
+#ifndef GIT_WINDOWS_NATIVE
+{
+       int notify_pipe[2];
+       if (pipe(notify_pipe))
+               notify_pipe[0] = notify_pipe[1] = -1;
+
        cmd->pid = fork();
+       failed_errno = errno;
        if (!cmd->pid) {
+               /*
+                * Redirect the channel to write syscall error messages to
+                * before redirecting the process's stderr so that all die()
+                * in subsequent call paths use the parent's stderr.
+                */
+               if (cmd->no_stderr || need_err) {
+                       int child_err = dup(2);
+                       set_cloexec(child_err);
+                       set_error_handle(fdopen(child_err, "w"));
+               }
+
+               close(notify_pipe[0]);
+               set_cloexec(notify_pipe[1]);
+               child_notifier = notify_pipe[1];
+               atexit(notify_parent);
+
                if (cmd->no_stdin)
                        dup_devnull(0);
                else if (need_in) {
@@ -86,6 +376,9 @@ int start_command(struct child_process *cmd)
                else if (need_err) {
                        dup2(fderr[1], 2);
                        close_pair(fderr);
+               } else if (cmd->err > 1) {
+                       dup2(cmd->err, 2);
+                       close(cmd->err);
                }
 
                if (cmd->no_stdout)
@@ -101,87 +394,106 @@ int start_command(struct child_process *cmd)
                }
 
                if (cmd->dir && chdir(cmd->dir))
-                       die("exec %s: cd to %s failed (%s)", cmd->argv[0],
-                           cmd->dir, strerror(errno));
+                       die_errno("exec '%s': cd to '%s' failed", cmd->argv[0],
+                           cmd->dir);
                if (cmd->env) {
                        for (; *cmd->env; cmd->env++) {
                                if (strchr(*cmd->env, '='))
-                                       putenv((char*)*cmd->env);
+                                       putenv((char *)*cmd->env);
                                else
                                        unsetenv(*cmd->env);
                        }
                }
-               if (cmd->git_cmd) {
+               if (cmd->git_cmd)
                        execv_git_cmd(cmd->argv);
+               else if (cmd->use_shell)
+                       execv_shell_cmd(cmd->argv);
+               else
+                       sane_execvp(cmd->argv[0], (char *const*) cmd->argv);
+               if (errno == ENOENT) {
+                       if (!cmd->silent_exec_failure)
+                               error("cannot run %s: %s", cmd->argv[0],
+                                       strerror(ENOENT));
+                       exit(127);
                } else {
-                       execvp(cmd->argv[0], (char *const*) cmd->argv);
+                       die_errno("cannot exec '%s'", cmd->argv[0]);
                }
-               die("exec %s failed.", cmd->argv[0]);
-       }
-#else
-       int s0 = -1, s1 = -1, s2 = -1;  /* backups of stdin, stdout, stderr */
-       const char **sargv = cmd->argv;
-       char **env = environ;
-
-       if (cmd->no_stdin) {
-               s0 = dup(0);
-               dup_devnull(0);
-       } else if (need_in) {
-               s0 = dup(0);
-               dup2(fdin[0], 0);
-       } else if (cmd->in) {
-               s0 = dup(0);
-               dup2(cmd->in, 0);
-       }
-
-       if (cmd->no_stderr) {
-               s2 = dup(2);
-               dup_devnull(2);
-       } else if (need_err) {
-               s2 = dup(2);
-               dup2(fderr[1], 2);
        }
+       if (cmd->pid < 0)
+               error_errno("cannot fork() for %s", cmd->argv[0]);
+       else if (cmd->clean_on_exit)
+               mark_child_for_cleanup(cmd->pid);
 
-       if (cmd->no_stdout) {
-               s1 = dup(1);
-               dup_devnull(1);
-       } else if (cmd->stdout_to_stderr) {
-               s1 = dup(1);
-               dup2(2, 1);
-       } else if (need_out) {
-               s1 = dup(1);
-               dup2(fdout[1], 1);
-       } else if (cmd->out > 1) {
-               s1 = dup(1);
-               dup2(cmd->out, 1);
+       /*
+        * Wait for child's execvp. If the execvp succeeds (or if fork()
+        * failed), EOF is seen immediately by the parent. Otherwise, the
+        * child process sends a single byte.
+        * Note that use of this infrastructure is completely advisory,
+        * therefore, we keep error checks minimal.
+        */
+       close(notify_pipe[1]);
+       if (read(notify_pipe[0], &notify_pipe[1], 1) == 1) {
+               /*
+                * At this point we know that fork() succeeded, but execvp()
+                * failed. Errors have been reported to our stderr.
+                */
+               wait_or_whine(cmd->pid, cmd->argv[0], 0);
+               failed_errno = errno;
+               cmd->pid = -1;
        }
+       close(notify_pipe[0]);
+}
+#else
+{
+       int fhin = 0, fhout = 1, fherr = 2;
+       const char **sargv = cmd->argv;
+       struct argv_array nargv = ARGV_ARRAY_INIT;
 
-       if (cmd->dir)
-               die("chdir in start_command() not implemented");
-       if (cmd->env) {
-               env = copy_environ();
-               for (; *cmd->env; cmd->env++)
-                       env = env_setenv(env, *cmd->env);
-       }
+       if (cmd->no_stdin)
+               fhin = open("/dev/null", O_RDWR);
+       else if (need_in)
+               fhin = dup(fdin[0]);
+       else if (cmd->in)
+               fhin = dup(cmd->in);
 
-       if (cmd->git_cmd) {
-               cmd->argv = prepare_git_cmd(cmd->argv);
-       }
+       if (cmd->no_stderr)
+               fherr = open("/dev/null", O_RDWR);
+       else if (need_err)
+               fherr = dup(fderr[1]);
+       else if (cmd->err > 2)
+               fherr = dup(cmd->err);
 
-       cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, env);
+       if (cmd->no_stdout)
+               fhout = open("/dev/null", O_RDWR);
+       else if (cmd->stdout_to_stderr)
+               fhout = dup(fherr);
+       else if (need_out)
+               fhout = dup(fdout[1]);
+       else if (cmd->out > 1)
+               fhout = dup(cmd->out);
 
-       if (cmd->env)
-               free_environ(env);
        if (cmd->git_cmd)
-               free(cmd->argv);
+               cmd->argv = prepare_git_cmd(&nargv, cmd->argv);
+       else if (cmd->use_shell)
+               cmd->argv = prepare_shell_cmd(&nargv, cmd->argv);
 
+       cmd->pid = mingw_spawnvpe(cmd->argv[0], cmd->argv, (char**) cmd->env,
+                       cmd->dir, fhin, fhout, fherr);
+       failed_errno = errno;
+       if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
+               error_errno("cannot spawn %s", cmd->argv[0]);
+       if (cmd->clean_on_exit && cmd->pid >= 0)
+               mark_child_for_cleanup(cmd->pid);
+
+       argv_array_clear(&nargv);
        cmd->argv = sargv;
-       if (s0 >= 0)
-               dup2(s0, 0), close(s0);
-       if (s1 >= 0)
-               dup2(s1, 1), close(s1);
-       if (s2 >= 0)
-               dup2(s2, 2), close(s2);
+       if (fhin != 0)
+               close(fhin);
+       if (fhout != 1)
+               close(fhout);
+       if (fherr != 2)
+               close(fherr);
+}
 #endif
 
        if (cmd->pid < 0) {
@@ -195,7 +507,11 @@ int start_command(struct child_process *cmd)
                        close(cmd->out);
                if (need_err)
                        close_pair(fderr);
-               return -ERR_RUN_COMMAND_FORK;
+               else if (cmd->err)
+                       close(cmd->err);
+               child_process_clear(cmd);
+               errno = failed_errno;
+               return -1;
        }
 
        if (need_in)
@@ -210,141 +526,813 @@ int start_command(struct child_process *cmd)
 
        if (need_err)
                close(fderr[1]);
+       else if (cmd->err)
+               close(cmd->err);
 
        return 0;
 }
 
-static int wait_or_whine(pid_t pid)
+int finish_command(struct child_process *cmd)
 {
-       for (;;) {
-               int status, code;
-               pid_t waiting = waitpid(pid, &status, 0);
-
-               if (waiting < 0) {
-                       if (errno == EINTR)
-                               continue;
-                       error("waitpid failed (%s)", strerror(errno));
-                       return -ERR_RUN_COMMAND_WAITPID;
-               }
-               if (waiting != pid)
-                       return -ERR_RUN_COMMAND_WAITPID_WRONG_PID;
-               if (WIFSIGNALED(status))
-                       return -ERR_RUN_COMMAND_WAITPID_SIGNAL;
-
-               if (!WIFEXITED(status))
-                       return -ERR_RUN_COMMAND_WAITPID_NOEXIT;
-               code = WEXITSTATUS(status);
-               if (code)
-                       return -code;
-               return 0;
-       }
+       int ret = wait_or_whine(cmd->pid, cmd->argv[0], 0);
+       child_process_clear(cmd);
+       return ret;
 }
 
-int finish_command(struct child_process *cmd)
+int finish_command_in_signal(struct child_process *cmd)
 {
-       return wait_or_whine(cmd->pid);
+       return wait_or_whine(cmd->pid, cmd->argv[0], 1);
 }
 
+
 int run_command(struct child_process *cmd)
 {
-       int code = start_command(cmd);
+       int code;
+
+       if (cmd->out < 0 || cmd->err < 0)
+               die("BUG: run_command with a pipe can cause deadlock");
+
+       code = start_command(cmd);
        if (code)
                return code;
        return finish_command(cmd);
 }
 
-static void prepare_run_command_v_opt(struct child_process *cmd,
-                                     const char **argv,
-                                     int opt)
+int run_command_v_opt(const char **argv, int opt)
 {
-       memset(cmd, 0, sizeof(*cmd));
-       cmd->argv = argv;
-       cmd->no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
-       cmd->git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
-       cmd->stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
+       return run_command_v_opt_cd_env(argv, opt, NULL, NULL);
 }
 
-int run_command_v_opt(const char **argv, int opt)
+int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
 {
-       struct child_process cmd;
-       prepare_run_command_v_opt(&cmd, argv, opt);
+       struct child_process cmd = CHILD_PROCESS_INIT;
+       cmd.argv = argv;
+       cmd.no_stdin = opt & RUN_COMMAND_NO_STDIN ? 1 : 0;
+       cmd.git_cmd = opt & RUN_GIT_CMD ? 1 : 0;
+       cmd.stdout_to_stderr = opt & RUN_COMMAND_STDOUT_TO_STDERR ? 1 : 0;
+       cmd.silent_exec_failure = opt & RUN_SILENT_EXEC_FAILURE ? 1 : 0;
+       cmd.use_shell = opt & RUN_USING_SHELL ? 1 : 0;
+       cmd.clean_on_exit = opt & RUN_CLEAN_ON_EXIT ? 1 : 0;
+       cmd.dir = dir;
+       cmd.env = env;
        return run_command(&cmd);
 }
 
-int run_command_v_opt_cd(const char **argv, int opt, const char *dir)
+#ifndef NO_PTHREADS
+static pthread_t main_thread;
+static int main_thread_set;
+static pthread_key_t async_key;
+static pthread_key_t async_die_counter;
+
+static void *run_thread(void *data)
 {
-       struct child_process cmd;
-       prepare_run_command_v_opt(&cmd, argv, opt);
-       cmd.dir = dir;
-       return run_command(&cmd);
+       struct async *async = data;
+       intptr_t ret;
+
+       if (async->isolate_sigpipe) {
+               sigset_t mask;
+               sigemptyset(&mask);
+               sigaddset(&mask, SIGPIPE);
+               if (pthread_sigmask(SIG_BLOCK, &mask, NULL) < 0) {
+                       ret = error("unable to block SIGPIPE in async thread");
+                       return (void *)ret;
+               }
+       }
+
+       pthread_setspecific(async_key, async);
+       ret = async->proc(async->proc_in, async->proc_out, async->data);
+       return (void *)ret;
 }
 
-int run_command_v_opt_cd_env(const char **argv, int opt, const char *dir, const char *const *env)
+static NORETURN void die_async(const char *err, va_list params)
 {
-       struct child_process cmd;
-       prepare_run_command_v_opt(&cmd, argv, opt);
-       cmd.dir = dir;
-       cmd.env = env;
-       return run_command(&cmd);
+       vreportf("fatal: ", err, params);
+
+       if (in_async()) {
+               struct async *async = pthread_getspecific(async_key);
+               if (async->proc_in >= 0)
+                       close(async->proc_in);
+               if (async->proc_out >= 0)
+                       close(async->proc_out);
+               pthread_exit((void *)128);
+       }
+
+       exit(128);
 }
 
-#ifdef __MINGW32__
-static __stdcall unsigned run_thread(void *data)
+static int async_die_is_recursing(void)
 {
-       struct async *async = data;
-       return async->proc(async->fd_for_proc, async->data);
+       void *ret = pthread_getspecific(async_die_counter);
+       pthread_setspecific(async_die_counter, (void *)1);
+       return ret != NULL;
+}
+
+int in_async(void)
+{
+       if (!main_thread_set)
+               return 0; /* no asyncs started yet */
+       return !pthread_equal(main_thread, pthread_self());
+}
+
+void NORETURN async_exit(int code)
+{
+       pthread_exit((void *)(intptr_t)code);
+}
+
+#else
+
+static struct {
+       void (**handlers)(void);
+       size_t nr;
+       size_t alloc;
+} git_atexit_hdlrs;
+
+static int git_atexit_installed;
+
+static void git_atexit_dispatch(void)
+{
+       size_t i;
+
+       for (i=git_atexit_hdlrs.nr ; i ; i--)
+               git_atexit_hdlrs.handlers[i-1]();
+}
+
+static void git_atexit_clear(void)
+{
+       free(git_atexit_hdlrs.handlers);
+       memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
+       git_atexit_installed = 0;
 }
+
+#undef atexit
+int git_atexit(void (*handler)(void))
+{
+       ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
+       git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
+       if (!git_atexit_installed) {
+               if (atexit(&git_atexit_dispatch))
+                       return -1;
+               git_atexit_installed = 1;
+       }
+       return 0;
+}
+#define atexit git_atexit
+
+static int process_is_async;
+int in_async(void)
+{
+       return process_is_async;
+}
+
+void NORETURN async_exit(int code)
+{
+       exit(code);
+}
+
 #endif
 
 int start_async(struct async *async)
 {
-       int pipe_out[2];
+       int need_in, need_out;
+       int fdin[2], fdout[2];
+       int proc_in, proc_out;
 
-       if (pipe(pipe_out) < 0)
-               return error("cannot create pipe: %s", strerror(errno));
-       async->out = pipe_out[0];
+       need_in = async->in < 0;
+       if (need_in) {
+               if (pipe(fdin) < 0) {
+                       if (async->out > 0)
+                               close(async->out);
+                       return error_errno("cannot create pipe");
+               }
+               async->in = fdin[1];
+       }
+
+       need_out = async->out < 0;
+       if (need_out) {
+               if (pipe(fdout) < 0) {
+                       if (need_in)
+                               close_pair(fdin);
+                       else if (async->in)
+                               close(async->in);
+                       return error_errno("cannot create pipe");
+               }
+               async->out = fdout[0];
+       }
 
-#ifndef __MINGW32__
+       if (need_in)
+               proc_in = fdin[0];
+       else if (async->in)
+               proc_in = async->in;
+       else
+               proc_in = -1;
+
+       if (need_out)
+               proc_out = fdout[1];
+       else if (async->out)
+               proc_out = async->out;
+       else
+               proc_out = -1;
+
+#ifdef NO_PTHREADS
        /* Flush stdio before fork() to avoid cloning buffers */
        fflush(NULL);
 
        async->pid = fork();
        if (async->pid < 0) {
-               error("fork (async) failed: %s", strerror(errno));
-               close_pair(pipe_out);
-               return -1;
+               error_errno("fork (async) failed");
+               goto error;
        }
        if (!async->pid) {
-               close(pipe_out[0]);
-               exit(!!async->proc(pipe_out[1], async->data));
+               if (need_in)
+                       close(fdin[1]);
+               if (need_out)
+                       close(fdout[0]);
+               git_atexit_clear();
+               process_is_async = 1;
+               exit(!!async->proc(proc_in, proc_out, async->data));
        }
-       close(pipe_out[1]);
+
+       mark_child_for_cleanup(async->pid);
+
+       if (need_in)
+               close(fdin[0]);
+       else if (async->in)
+               close(async->in);
+
+       if (need_out)
+               close(fdout[1]);
+       else if (async->out)
+               close(async->out);
 #else
-       async->fd_for_proc = pipe_out[1];
-       async->tid = (HANDLE) _beginthreadex(NULL, 0, run_thread, async, 0, NULL);
-       if (!async->tid) {
-               error("cannot create thread: %s", strerror(errno));
-               close_pair(pipe_out);
-               return -1;
+       if (!main_thread_set) {
+               /*
+                * We assume that the first time that start_async is called
+                * it is from the main thread.
+                */
+               main_thread_set = 1;
+               main_thread = pthread_self();
+               pthread_key_create(&async_key, NULL);
+               pthread_key_create(&async_die_counter, NULL);
+               set_die_routine(die_async);
+               set_die_is_recursing_routine(async_die_is_recursing);
+       }
+
+       if (proc_in >= 0)
+               set_cloexec(proc_in);
+       if (proc_out >= 0)
+               set_cloexec(proc_out);
+       async->proc_in = proc_in;
+       async->proc_out = proc_out;
+       {
+               int err = pthread_create(&async->tid, NULL, run_thread, async);
+               if (err) {
+                       error_errno("cannot create thread");
+                       goto error;
+               }
        }
 #endif
        return 0;
+
+error:
+       if (need_in)
+               close_pair(fdin);
+       else if (async->in)
+               close(async->in);
+
+       if (need_out)
+               close_pair(fdout);
+       else if (async->out)
+               close(async->out);
+       return -1;
 }
 
 int finish_async(struct async *async)
 {
-#ifndef __MINGW32__
-       int ret = 0;
-
-       if (wait_or_whine(async->pid))
-               ret = error("waitpid (async) failed");
+#ifdef NO_PTHREADS
+       return wait_or_whine(async->pid, "child process", 0);
 #else
-       DWORD ret = 0;
-       if (WaitForSingleObject(async->tid, INFINITE) != WAIT_OBJECT_0)
-               ret = error("waiting for thread failed: %lu", GetLastError());
-       else if (!GetExitCodeThread(async->tid, &ret))
-               ret = error("cannot get thread exit code: %lu", GetLastError());
-       CloseHandle(async->tid);
+       void *ret = (void *)(intptr_t)(-1);
+
+       if (pthread_join(async->tid, &ret))
+               error("pthread_join failed");
+       return (int)(intptr_t)ret;
 #endif
+}
+
+const char *find_hook(const char *name)
+{
+       static struct strbuf path = STRBUF_INIT;
+
+       strbuf_reset(&path);
+       if (git_hooks_path)
+               strbuf_addf(&path, "%s/%s", git_hooks_path, name);
+       else
+               strbuf_git_path(&path, "hooks/%s", name);
+       if (access(path.buf, X_OK) < 0)
+               return NULL;
+       return path.buf;
+}
+
+int run_hook_ve(const char *const *env, const char *name, va_list args)
+{
+       struct child_process hook = CHILD_PROCESS_INIT;
+       const char *p;
+
+       p = find_hook(name);
+       if (!p)
+               return 0;
+
+       argv_array_push(&hook.args, p);
+       while ((p = va_arg(args, const char *)))
+               argv_array_push(&hook.args, p);
+       hook.env = env;
+       hook.no_stdin = 1;
+       hook.stdout_to_stderr = 1;
+
+       return run_command(&hook);
+}
+
+int run_hook_le(const char *const *env, const char *name, ...)
+{
+       va_list args;
+       int ret;
+
+       va_start(args, name);
+       ret = run_hook_ve(env, name, args);
+       va_end(args);
+
        return ret;
 }
+
+struct io_pump {
+       /* initialized by caller */
+       int fd;
+       int type; /* POLLOUT or POLLIN */
+       union {
+               struct {
+                       const char *buf;
+                       size_t len;
+               } out;
+               struct {
+                       struct strbuf *buf;
+                       size_t hint;
+               } in;
+       } u;
+
+       /* returned by pump_io */
+       int error; /* 0 for success, otherwise errno */
+
+       /* internal use */
+       struct pollfd *pfd;
+};
+
+static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
+{
+       int pollsize = 0;
+       int i;
+
+       for (i = 0; i < nr; i++) {
+               struct io_pump *io = &slots[i];
+               if (io->fd < 0)
+                       continue;
+               pfd[pollsize].fd = io->fd;
+               pfd[pollsize].events = io->type;
+               io->pfd = &pfd[pollsize++];
+       }
+
+       if (!pollsize)
+               return 0;
+
+       if (poll(pfd, pollsize, -1) < 0) {
+               if (errno == EINTR)
+                       return 1;
+               die_errno("poll failed");
+       }
+
+       for (i = 0; i < nr; i++) {
+               struct io_pump *io = &slots[i];
+
+               if (io->fd < 0)
+                       continue;
+
+               if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
+                       continue;
+
+               if (io->type == POLLOUT) {
+                       ssize_t len = xwrite(io->fd,
+                                            io->u.out.buf, io->u.out.len);
+                       if (len < 0) {
+                               io->error = errno;
+                               close(io->fd);
+                               io->fd = -1;
+                       } else {
+                               io->u.out.buf += len;
+                               io->u.out.len -= len;
+                               if (!io->u.out.len) {
+                                       close(io->fd);
+                                       io->fd = -1;
+                               }
+                       }
+               }
+
+               if (io->type == POLLIN) {
+                       ssize_t len = strbuf_read_once(io->u.in.buf,
+                                                      io->fd, io->u.in.hint);
+                       if (len < 0)
+                               io->error = errno;
+                       if (len <= 0) {
+                               close(io->fd);
+                               io->fd = -1;
+                       }
+               }
+       }
+
+       return 1;
+}
+
+static int pump_io(struct io_pump *slots, int nr)
+{
+       struct pollfd *pfd;
+       int i;
+
+       for (i = 0; i < nr; i++)
+               slots[i].error = 0;
+
+       ALLOC_ARRAY(pfd, nr);
+       while (pump_io_round(slots, nr, pfd))
+               ; /* nothing */
+       free(pfd);
+
+       /* There may be multiple errno values, so just pick the first. */
+       for (i = 0; i < nr; i++) {
+               if (slots[i].error) {
+                       errno = slots[i].error;
+                       return -1;
+               }
+       }
+       return 0;
+}
+
+
+int 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)
+{
+       struct io_pump io[3];
+       int nr = 0;
+
+       if (in)
+               cmd->in = -1;
+       if (out)
+               cmd->out = -1;
+       if (err)
+               cmd->err = -1;
+
+       if (start_command(cmd) < 0)
+               return -1;
+
+       if (in) {
+               io[nr].fd = cmd->in;
+               io[nr].type = POLLOUT;
+               io[nr].u.out.buf = in;
+               io[nr].u.out.len = in_len;
+               nr++;
+       }
+       if (out) {
+               io[nr].fd = cmd->out;
+               io[nr].type = POLLIN;
+               io[nr].u.in.buf = out;
+               io[nr].u.in.hint = out_hint;
+               nr++;
+       }
+       if (err) {
+               io[nr].fd = cmd->err;
+               io[nr].type = POLLIN;
+               io[nr].u.in.buf = err;
+               io[nr].u.in.hint = err_hint;
+               nr++;
+       }
+
+       if (pump_io(io, nr) < 0) {
+               finish_command(cmd); /* throw away exit code */
+               return -1;
+       }
+
+       return finish_command(cmd);
+}
+
+enum child_state {
+       GIT_CP_FREE,
+       GIT_CP_WORKING,
+       GIT_CP_WAIT_CLEANUP,
+};
+
+struct parallel_processes {
+       void *data;
+
+       int max_processes;
+       int nr_processes;
+
+       get_next_task_fn get_next_task;
+       start_failure_fn start_failure;
+       task_finished_fn task_finished;
+
+       struct {
+               enum child_state state;
+               struct child_process process;
+               struct strbuf err;
+               void *data;
+       } *children;
+       /*
+        * The struct pollfd is logically part of *children,
+        * but the system call expects it as its own array.
+        */
+       struct pollfd *pfd;
+
+       unsigned shutdown : 1;
+
+       int output_owner;
+       struct strbuf buffered_output; /* of finished children */
+};
+
+static int default_start_failure(struct strbuf *out,
+                                void *pp_cb,
+                                void *pp_task_cb)
+{
+       return 0;
+}
+
+static int default_task_finished(int result,
+                                struct strbuf *out,
+                                void *pp_cb,
+                                void *pp_task_cb)
+{
+       return 0;
+}
+
+static void kill_children(struct parallel_processes *pp, int signo)
+{
+       int i, n = pp->max_processes;
+
+       for (i = 0; i < n; i++)
+               if (pp->children[i].state == GIT_CP_WORKING)
+                       kill(pp->children[i].process.pid, signo);
+}
+
+static struct parallel_processes *pp_for_signal;
+
+static void handle_children_on_signal(int signo)
+{
+       kill_children(pp_for_signal, signo);
+       sigchain_pop(signo);
+       raise(signo);
+}
+
+static void 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)
+{
+       int i;
+
+       if (n < 1)
+               n = online_cpus();
+
+       pp->max_processes = n;
+
+       trace_printf("run_processes_parallel: preparing to run up to %d tasks", n);
+
+       pp->data = data;
+       if (!get_next_task)
+               die("BUG: you need to specify a get_next_task function");
+       pp->get_next_task = get_next_task;
+
+       pp->start_failure = start_failure ? start_failure : default_start_failure;
+       pp->task_finished = task_finished ? task_finished : default_task_finished;
+
+       pp->nr_processes = 0;
+       pp->output_owner = 0;
+       pp->shutdown = 0;
+       pp->children = xcalloc(n, sizeof(*pp->children));
+       pp->pfd = xcalloc(n, sizeof(*pp->pfd));
+       strbuf_init(&pp->buffered_output, 0);
+
+       for (i = 0; i < n; i++) {
+               strbuf_init(&pp->children[i].err, 0);
+               child_process_init(&pp->children[i].process);
+               pp->pfd[i].events = POLLIN | POLLHUP;
+               pp->pfd[i].fd = -1;
+       }
+
+       pp_for_signal = pp;
+       sigchain_push_common(handle_children_on_signal);
+}
+
+static void pp_cleanup(struct parallel_processes *pp)
+{
+       int i;
+
+       trace_printf("run_processes_parallel: done");
+       for (i = 0; i < pp->max_processes; i++) {
+               strbuf_release(&pp->children[i].err);
+               child_process_clear(&pp->children[i].process);
+       }
+
+       free(pp->children);
+       free(pp->pfd);
+
+       /*
+        * When get_next_task added messages to the buffer in its last
+        * iteration, the buffered output is non empty.
+        */
+       strbuf_write(&pp->buffered_output, stderr);
+       strbuf_release(&pp->buffered_output);
+
+       sigchain_pop_common();
+}
+
+/* returns
+ *  0 if a new task was started.
+ *  1 if no new jobs was started (get_next_task ran out of work, non critical
+ *    problem with starting a new command)
+ * <0 no new job was started, user wishes to shutdown early. Use negative code
+ *    to signal the children.
+ */
+static int pp_start_one(struct parallel_processes *pp)
+{
+       int i, code;
+
+       for (i = 0; i < pp->max_processes; i++)
+               if (pp->children[i].state == GIT_CP_FREE)
+                       break;
+       if (i == pp->max_processes)
+               die("BUG: bookkeeping is hard");
+
+       code = pp->get_next_task(&pp->children[i].process,
+                                &pp->children[i].err,
+                                pp->data,
+                                &pp->children[i].data);
+       if (!code) {
+               strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
+               strbuf_reset(&pp->children[i].err);
+               return 1;
+       }
+       pp->children[i].process.err = -1;
+       pp->children[i].process.stdout_to_stderr = 1;
+       pp->children[i].process.no_stdin = 1;
+
+       if (start_command(&pp->children[i].process)) {
+               code = pp->start_failure(&pp->children[i].err,
+                                        pp->data,
+                                        &pp->children[i].data);
+               strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
+               strbuf_reset(&pp->children[i].err);
+               if (code)
+                       pp->shutdown = 1;
+               return code;
+       }
+
+       pp->nr_processes++;
+       pp->children[i].state = GIT_CP_WORKING;
+       pp->pfd[i].fd = pp->children[i].process.err;
+       return 0;
+}
+
+static void pp_buffer_stderr(struct parallel_processes *pp, int output_timeout)
+{
+       int i;
+
+       while ((i = poll(pp->pfd, pp->max_processes, output_timeout)) < 0) {
+               if (errno == EINTR)
+                       continue;
+               pp_cleanup(pp);
+               die_errno("poll");
+       }
+
+       /* Buffer output from all pipes. */
+       for (i = 0; i < pp->max_processes; i++) {
+               if (pp->children[i].state == GIT_CP_WORKING &&
+                   pp->pfd[i].revents & (POLLIN | POLLHUP)) {
+                       int n = strbuf_read_once(&pp->children[i].err,
+                                                pp->children[i].process.err, 0);
+                       if (n == 0) {
+                               close(pp->children[i].process.err);
+                               pp->children[i].state = GIT_CP_WAIT_CLEANUP;
+                       } else if (n < 0)
+                               if (errno != EAGAIN)
+                                       die_errno("read");
+               }
+       }
+}
+
+static void pp_output(struct parallel_processes *pp)
+{
+       int i = pp->output_owner;
+       if (pp->children[i].state == GIT_CP_WORKING &&
+           pp->children[i].err.len) {
+               strbuf_write(&pp->children[i].err, stderr);
+               strbuf_reset(&pp->children[i].err);
+       }
+}
+
+static int pp_collect_finished(struct parallel_processes *pp)
+{
+       int i, code;
+       int n = pp->max_processes;
+       int result = 0;
+
+       while (pp->nr_processes > 0) {
+               for (i = 0; i < pp->max_processes; i++)
+                       if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
+                               break;
+               if (i == pp->max_processes)
+                       break;
+
+               code = finish_command(&pp->children[i].process);
+
+               code = pp->task_finished(code,
+                                        &pp->children[i].err, pp->data,
+                                        &pp->children[i].data);
+
+               if (code)
+                       result = code;
+               if (code < 0)
+                       break;
+
+               pp->nr_processes--;
+               pp->children[i].state = GIT_CP_FREE;
+               pp->pfd[i].fd = -1;
+               child_process_init(&pp->children[i].process);
+
+               if (i != pp->output_owner) {
+                       strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
+                       strbuf_reset(&pp->children[i].err);
+               } else {
+                       strbuf_write(&pp->children[i].err, stderr);
+                       strbuf_reset(&pp->children[i].err);
+
+                       /* Output all other finished child processes */
+                       strbuf_write(&pp->buffered_output, stderr);
+                       strbuf_reset(&pp->buffered_output);
+
+                       /*
+                        * Pick next process to output live.
+                        * NEEDSWORK:
+                        * For now we pick it randomly by doing a round
+                        * robin. Later we may want to pick the one with
+                        * the most output or the longest or shortest
+                        * running process time.
+                        */
+                       for (i = 0; i < n; i++)
+                               if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
+                                       break;
+                       pp->output_owner = (pp->output_owner + i) % n;
+               }
+       }
+       return result;
+}
+
+int 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)
+{
+       int i, code;
+       int output_timeout = 100;
+       int spawn_cap = 4;
+       struct parallel_processes pp;
+
+       pp_init(&pp, n, get_next_task, start_failure, task_finished, pp_cb);
+       while (1) {
+               for (i = 0;
+                   i < spawn_cap && !pp.shutdown &&
+                   pp.nr_processes < pp.max_processes;
+                   i++) {
+                       code = pp_start_one(&pp);
+                       if (!code)
+                               continue;
+                       if (code < 0) {
+                               pp.shutdown = 1;
+                               kill_children(&pp, -code);
+                       }
+                       break;
+               }
+               if (!pp.nr_processes)
+                       break;
+               pp_buffer_stderr(&pp, output_timeout);
+               pp_output(&pp);
+               code = pp_collect_finished(&pp);
+               if (code) {
+                       pp.shutdown = 1;
+                       if (code < 0)
+                               kill_children(&pp, -code);
+               }
+       }
+
+       pp_cleanup(&pp);
+       return 0;
+}