2 * Copyright 2016 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
21 #include <folly/Subprocess.h>
24 #include <sys/prctl.h>
30 #include <system_error>
32 #include <boost/container/flat_set.hpp>
33 #include <boost/range/adaptors.hpp>
35 #include <glog/logging.h>
37 #include <folly/Conv.h>
38 #include <folly/Exception.h>
39 #include <folly/ScopeGuard.h>
40 #include <folly/String.h>
41 #include <folly/io/Cursor.h>
42 #include <folly/portability/Environment.h>
43 #include <folly/portability/Sockets.h>
44 #include <folly/portability/Unistd.h>
46 constexpr int kExecFailure = 127;
47 constexpr int kChildFailure = 126;
51 ProcessReturnCode::ProcessReturnCode(ProcessReturnCode&& p) noexcept
52 : rawStatus_(p.rawStatus_) {
53 p.rawStatus_ = ProcessReturnCode::RV_NOT_STARTED;
56 ProcessReturnCode& ProcessReturnCode::operator=(ProcessReturnCode&& p)
58 rawStatus_ = p.rawStatus_;
59 p.rawStatus_ = ProcessReturnCode::RV_NOT_STARTED;
63 ProcessReturnCode::State ProcessReturnCode::state() const {
64 if (rawStatus_ == RV_NOT_STARTED) return NOT_STARTED;
65 if (rawStatus_ == RV_RUNNING) return RUNNING;
66 if (WIFEXITED(rawStatus_)) return EXITED;
67 if (WIFSIGNALED(rawStatus_)) return KILLED;
68 throw std::runtime_error(to<std::string>(
69 "Invalid ProcessReturnCode: ", rawStatus_));
72 void ProcessReturnCode::enforce(State expected) const {
75 throw std::logic_error(to<std::string>(
76 "Bad use of ProcessReturnCode; state is ", s, " expected ", expected
81 int ProcessReturnCode::exitStatus() const {
83 return WEXITSTATUS(rawStatus_);
86 int ProcessReturnCode::killSignal() const {
88 return WTERMSIG(rawStatus_);
91 bool ProcessReturnCode::coreDumped() const {
93 return WCOREDUMP(rawStatus_);
96 std::string ProcessReturnCode::str() const {
103 return to<std::string>("exited with status ", exitStatus());
105 return to<std::string>("killed by signal ", killSignal(),
106 (coreDumped() ? " (core dumped)" : ""));
108 CHECK(false); // unreached
109 return ""; // silence GCC warning
112 CalledProcessError::CalledProcessError(ProcessReturnCode rc)
114 what_(returnCode_.str()) {
117 SubprocessSpawnError::SubprocessSpawnError(const char* executable,
120 : errnoValue_(errnoValue),
121 what_(to<std::string>(errCode == kExecFailure ?
122 "failed to execute " :
123 "error preparing to execute ",
124 executable, ": ", errnoStr(errnoValue))) {
129 // Copy pointers to the given strings in a format suitable for posix_spawn
130 std::unique_ptr<const char*[]> cloneStrings(const std::vector<std::string>& s) {
131 std::unique_ptr<const char*[]> d(new const char*[s.size() + 1]);
132 for (size_t i = 0; i < s.size(); i++) {
135 d[s.size()] = nullptr;
139 // Check a wait() status, throw on non-successful
140 void checkStatus(ProcessReturnCode returnCode) {
141 if (returnCode.state() != ProcessReturnCode::EXITED ||
142 returnCode.exitStatus() != 0) {
143 throw CalledProcessError(returnCode);
149 Subprocess::Options& Subprocess::Options::fd(int fd, int action) {
150 if (action == Subprocess::PIPE) {
152 action = Subprocess::PIPE_IN;
153 } else if (fd == 1 || fd == 2) {
154 action = Subprocess::PIPE_OUT;
156 throw std::invalid_argument(
157 to<std::string>("Only fds 0, 1, 2 are valid for action=PIPE: ", fd));
160 fdActions_[fd] = action;
164 Subprocess::Subprocess() {}
166 Subprocess::Subprocess(
167 const std::vector<std::string>& argv,
168 const Options& options,
169 const char* executable,
170 const std::vector<std::string>* env) {
172 throw std::invalid_argument("argv must not be empty");
174 if (!executable) executable = argv[0].c_str();
175 spawn(cloneStrings(argv), executable, options, env);
178 Subprocess::Subprocess(
179 const std::string& cmd,
180 const Options& options,
181 const std::vector<std::string>* env) {
182 if (options.usePath_) {
183 throw std::invalid_argument("usePath() not allowed when running in shell");
185 const char* shell = getenv("SHELL");
190 std::unique_ptr<const char*[]> argv(new const char*[4]);
193 argv[2] = cmd.c_str();
195 spawn(std::move(argv), shell, options, env);
198 Subprocess::~Subprocess() {
199 CHECK_NE(returnCode_.state(), ProcessReturnCode::RUNNING)
200 << "Subprocess destroyed without reaping child";
205 struct ChildErrorInfo {
210 [[noreturn]] void childError(int errFd, int errCode, int errnoValue) {
211 ChildErrorInfo info = {errCode, errnoValue};
212 // Write the error information over the pipe to our parent process.
213 // We can't really do anything else if this write call fails.
214 writeNoInt(errFd, &info, sizeof(info));
221 void Subprocess::setAllNonBlocking() {
222 for (auto& p : pipes_) {
223 int fd = p.pipe.fd();
224 int flags = ::fcntl(fd, F_GETFL);
225 checkUnixError(flags, "fcntl");
226 int r = ::fcntl(fd, F_SETFL, flags | O_NONBLOCK);
227 checkUnixError(r, "fcntl");
231 void Subprocess::spawn(
232 std::unique_ptr<const char*[]> argv,
233 const char* executable,
234 const Options& optionsIn,
235 const std::vector<std::string>* env) {
236 if (optionsIn.usePath_ && env) {
237 throw std::invalid_argument(
238 "usePath() not allowed when overriding environment");
241 // Make a copy, we'll mutate options
242 Options options(optionsIn);
244 // On error, close all pipes_ (ignoring errors, but that seems fine here).
245 auto pipesGuard = makeGuard([this] { pipes_.clear(); });
247 // Create a pipe to use to receive error information from the child,
248 // in case it fails before calling exec()
251 checkUnixError(::pipe2(errFds, O_CLOEXEC), "pipe2");
253 checkUnixError(::pipe(errFds), "pipe");
256 CHECK_ERR(::close(errFds[0]));
257 if (errFds[1] >= 0) {
258 CHECK_ERR(::close(errFds[1]));
262 #if !FOLLY_HAVE_PIPE2
263 // Ask the child to close the read end of the error pipe.
264 checkUnixError(fcntl(errFds[0], F_SETFD, FD_CLOEXEC), "set FD_CLOEXEC");
265 // Set the close-on-exec flag on the write side of the pipe.
266 // This way the pipe will be closed automatically in the child if execve()
267 // succeeds. If the exec fails the child can write error information to the
269 checkUnixError(fcntl(errFds[1], F_SETFD, FD_CLOEXEC), "set FD_CLOEXEC");
272 // Perform the actual work of setting up pipes then forking and
273 // executing the child.
274 spawnInternal(std::move(argv), executable, options, env, errFds[1]);
276 // After spawnInternal() returns the child is alive. We have to be very
277 // careful about throwing after this point. We are inside the constructor,
278 // so if we throw the Subprocess object will have never existed, and the
279 // destructor will never be called.
281 // We should only throw if we got an error via the errFd, and we know the
282 // child has exited and can be immediately waited for. In all other cases,
283 // we have no way of cleaning up the child.
285 // Close writable side of the errFd pipe in the parent process
286 CHECK_ERR(::close(errFds[1]));
289 // Read from the errFd pipe, to tell if the child ran into any errors before
291 readChildErrorPipe(errFds[0], executable);
293 // We have fully succeeded now, so release the guard on pipes_
294 pipesGuard.dismiss();
297 void Subprocess::spawnInternal(
298 std::unique_ptr<const char*[]> argv,
299 const char* executable,
301 const std::vector<std::string>* env,
303 // Parent work, pre-fork: create pipes
304 std::vector<int> childFds;
305 // Close all of the childFds as we leave this scope
307 // These are only pipes, closing them shouldn't fail
308 for (int cfd : childFds) {
309 CHECK_ERR(::close(cfd));
314 for (auto& p : options.fdActions_) {
315 if (p.second == PIPE_IN || p.second == PIPE_OUT) {
317 // We're setting both ends of the pipe as close-on-exec. The child
318 // doesn't need to reset the flag on its end, as we always dup2() the fd,
319 // and dup2() fds don't share the close-on-exec flag.
321 // If possible, set close-on-exec atomically. Otherwise, a concurrent
322 // Subprocess invocation can fork() between "pipe" and "fnctl",
323 // causing FDs to leak.
324 r = ::pipe2(fds, O_CLOEXEC);
325 checkUnixError(r, "pipe2");
328 checkUnixError(r, "pipe");
329 r = fcntl(fds[0], F_SETFD, FD_CLOEXEC);
330 checkUnixError(r, "set FD_CLOEXEC");
331 r = fcntl(fds[1], F_SETFD, FD_CLOEXEC);
332 checkUnixError(r, "set FD_CLOEXEC");
334 pipes_.emplace_back();
335 Pipe& pipe = pipes_.back();
336 pipe.direction = p.second;
338 if (p.second == PIPE_IN) {
339 // Child gets reading end
340 pipe.pipe = folly::File(fds[1], /*owns_fd=*/ true);
343 pipe.pipe = folly::File(fds[0], /*owns_fd=*/ true);
346 p.second = cfd; // ensure it gets dup2()ed
347 pipe.childFd = p.first;
348 childFds.push_back(cfd);
352 // This should already be sorted, as options.fdActions_ is
353 DCHECK(std::is_sorted(pipes_.begin(), pipes_.end()));
355 // Note that the const casts below are legit, per
356 // http://pubs.opengroup.org/onlinepubs/009695399/functions/exec.html
358 char** argVec = const_cast<char**>(argv.get());
360 // Set up environment
361 std::unique_ptr<const char*[]> envHolder;
364 envHolder = cloneStrings(*env);
365 envVec = const_cast<char**>(envHolder.get());
370 // Block all signals around vfork; see http://ewontfix.com/7/.
372 // As the child may run in the same address space as the parent until
373 // the actual execve() system call, any (custom) signal handlers that
374 // the parent has might alter parent's memory if invoked in the child,
375 // with undefined results. So we block all signals in the parent before
376 // vfork(), which will cause them to be blocked in the child as well (we
377 // rely on the fact that Linux, just like all sane implementations, only
378 // clones the calling thread). Then, in the child, we reset all signals
379 // to their default dispositions (while still blocked), and unblock them
380 // (so the exec()ed process inherits the parent's signal mask)
382 // The parent also unblocks all signals as soon as vfork() returns.
384 r = sigfillset(&allBlocked);
385 checkUnixError(r, "sigfillset");
388 r = pthread_sigmask(SIG_SETMASK, &allBlocked, &oldSignals);
389 checkPosixError(r, "pthread_sigmask");
391 // Restore signal mask
392 r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
393 CHECK_EQ(r, 0) << "pthread_sigmask: " << errnoStr(r); // shouldn't fail
396 // Call c_str() here, as it's not necessarily safe after fork.
397 const char* childDir =
398 options.childDir_.empty() ? nullptr : options.childDir_.c_str();
401 int errnoValue = prepareChild(options, &oldSignals, childDir);
402 if (errnoValue != 0) {
403 childError(errFd, kChildFailure, errnoValue);
406 errnoValue = runChild(executable, argVec, envVec, options);
407 // If we get here, exec() failed.
408 childError(errFd, kExecFailure, errnoValue);
410 // In parent. Make sure vfork() succeeded.
411 checkUnixError(pid, errno, "vfork");
413 // Child is alive. We have to be very careful about throwing after this
414 // point. We are inside the constructor, so if we throw the Subprocess
415 // object will have never existed, and the destructor will never be called.
417 // We should only throw if we got an error via the errFd, and we know the
418 // child has exited and can be immediately waited for. In all other cases,
419 // we have no way of cleaning up the child.
421 returnCode_ = ProcessReturnCode(RV_RUNNING);
424 int Subprocess::prepareChild(const Options& options,
425 const sigset_t* sigmask,
426 const char* childDir) const {
427 // While all signals are blocked, we must reset their
428 // dispositions to default.
429 for (int sig = 1; sig < NSIG; ++sig) {
430 ::signal(sig, SIG_DFL);
434 // Unblock signals; restore signal mask.
435 int r = pthread_sigmask(SIG_SETMASK, sigmask, nullptr);
437 return r; // pthread_sigmask() returns an errno value
441 // Change the working directory, if one is given
443 if (::chdir(childDir) == -1) {
448 // We don't have to explicitly close the parent's end of all pipes,
449 // as they all have the FD_CLOEXEC flag set and will be closed at
452 // Close all fds that we're supposed to close.
453 for (auto& p : options.fdActions_) {
454 if (p.second == CLOSE) {
455 if (::close(p.first) == -1) {
458 } else if (p.second != p.first) {
459 if (::dup2(p.second, p.first) == -1) {
465 // If requested, close all other file descriptors. Don't close
466 // any fds in options.fdActions_, and don't touch stdin, stdout, stderr.
468 if (options.closeOtherFds_) {
469 for (int fd = getdtablesize() - 1; fd >= 3; --fd) {
470 if (options.fdActions_.count(fd) == 0) {
477 // Opt to receive signal on parent death, if requested
478 if (options.parentDeathSignal_ != 0) {
479 const auto parentDeathSignal =
480 static_cast<unsigned long>(options.parentDeathSignal_);
481 if (prctl(PR_SET_PDEATHSIG, parentDeathSignal, 0, 0, 0) == -1) {
487 if (options.processGroupLeader_) {
488 if (setpgrp() == -1) {
493 // The user callback comes last, so that the child is otherwise all set up.
494 if (options.dangerousPostForkPreExecCallback_) {
495 if (int error = (*options.dangerousPostForkPreExecCallback_)()) {
503 int Subprocess::runChild(const char* executable,
504 char** argv, char** env,
505 const Options& options) const {
506 // Now, finally, exec.
507 if (options.usePath_) {
508 ::execvp(executable, argv);
510 ::execve(executable, argv, env);
515 void Subprocess::readChildErrorPipe(int pfd, const char* executable) {
517 auto rc = readNoInt(pfd, &info, sizeof(info));
519 // No data means the child executed successfully, and the pipe
520 // was closed due to the close-on-exec flag being set.
522 } else if (rc != sizeof(ChildErrorInfo)) {
523 // An error occurred trying to read from the pipe, or we got a partial read.
524 // Neither of these cases should really occur in practice.
526 // We can't get any error data from the child in this case, and we don't
527 // know if it is successfully running or not. All we can do is to return
528 // normally, as if the child executed successfully. If something bad
529 // happened the caller should at least get a non-normal exit status from
531 LOG(ERROR) << "unexpected error trying to read from child error pipe " <<
532 "rc=" << rc << ", errno=" << errno;
536 // We got error data from the child. The child should exit immediately in
537 // this case, so wait on it to clean up.
540 // Throw to signal the error
541 throw SubprocessSpawnError(executable, info.errCode, info.errnoValue);
544 ProcessReturnCode Subprocess::poll() {
545 returnCode_.enforce(ProcessReturnCode::RUNNING);
548 pid_t found = ::waitpid(pid_, &status, WNOHANG);
549 // The spec guarantees that EINTR does not occur with WNOHANG, so the only
550 // two remaining errors are ECHILD (other code reaped the child?), or
551 // EINVAL (cosmic rays?), both of which merit an abort:
552 PCHECK(found != -1) << "waitpid(" << pid_ << ", &status, WNOHANG)";
554 // Though the child process had quit, this call does not close the pipes
555 // since its descendants may still be using them.
556 returnCode_ = ProcessReturnCode(status);
562 bool Subprocess::pollChecked() {
563 if (poll().state() == ProcessReturnCode::RUNNING) {
566 checkStatus(returnCode_);
570 ProcessReturnCode Subprocess::wait() {
571 returnCode_.enforce(ProcessReturnCode::RUNNING);
576 found = ::waitpid(pid_, &status, 0);
577 } while (found == -1 && errno == EINTR);
578 // The only two remaining errors are ECHILD (other code reaped the
579 // child?), or EINVAL (cosmic rays?), and both merit an abort:
580 PCHECK(found != -1) << "waitpid(" << pid_ << ", &status, WNOHANG)";
581 // Though the child process had quit, this call does not close the pipes
582 // since its descendants may still be using them.
583 DCHECK_EQ(found, pid_);
584 returnCode_ = ProcessReturnCode(status);
589 void Subprocess::waitChecked() {
591 checkStatus(returnCode_);
594 void Subprocess::sendSignal(int signal) {
595 returnCode_.enforce(ProcessReturnCode::RUNNING);
596 int r = ::kill(pid_, signal);
597 checkUnixError(r, "kill");
600 pid_t Subprocess::pid() const {
606 ByteRange queueFront(const IOBufQueue& queue) {
607 auto* p = queue.front();
611 return io::Cursor(p).peekBytes();
615 bool handleWrite(int fd, IOBufQueue& queue) {
617 auto b = queueFront(queue);
622 ssize_t n = writeNoInt(fd, b.data(), b.size());
623 if (n == -1 && errno == EAGAIN) {
626 checkUnixError(n, "write");
632 bool handleRead(int fd, IOBufQueue& queue) {
634 auto p = queue.preallocate(100, 65000);
635 ssize_t n = readNoInt(fd, p.first, p.second);
636 if (n == -1 && errno == EAGAIN) {
639 checkUnixError(n, "read");
643 queue.postallocate(n);
647 bool discardRead(int fd) {
648 static const size_t bufSize = 65000;
649 // Thread unsafe, but it doesn't matter.
650 static std::unique_ptr<char[]> buf(new char[bufSize]);
653 ssize_t n = readNoInt(fd, buf.get(), bufSize);
654 if (n == -1 && errno == EAGAIN) {
657 checkUnixError(n, "read");
666 std::pair<std::string, std::string> Subprocess::communicate(
668 IOBufQueue inputQueue;
669 inputQueue.wrapBuffer(input.data(), input.size());
671 auto outQueues = communicateIOBuf(std::move(inputQueue));
672 auto outBufs = std::make_pair(outQueues.first.move(),
673 outQueues.second.move());
674 std::pair<std::string, std::string> out;
676 outBufs.first->coalesce();
677 out.first.assign(reinterpret_cast<const char*>(outBufs.first->data()),
678 outBufs.first->length());
680 if (outBufs.second) {
681 outBufs.second->coalesce();
682 out.second.assign(reinterpret_cast<const char*>(outBufs.second->data()),
683 outBufs.second->length());
688 std::pair<IOBufQueue, IOBufQueue> Subprocess::communicateIOBuf(
690 // If the user supplied a non-empty input buffer, make sure
691 // that stdin is a pipe so we can write the data.
692 if (!input.empty()) {
693 // findByChildFd() will throw std::invalid_argument if no pipe for
694 // STDIN_FILENO exists
695 findByChildFd(STDIN_FILENO);
698 std::pair<IOBufQueue, IOBufQueue> out;
700 auto readCallback = [&] (int pfd, int cfd) -> bool {
701 if (cfd == STDOUT_FILENO) {
702 return handleRead(pfd, out.first);
703 } else if (cfd == STDERR_FILENO) {
704 return handleRead(pfd, out.second);
706 // Don't close the file descriptor, the child might not like SIGPIPE,
707 // just read and throw the data away.
708 return discardRead(pfd);
712 auto writeCallback = [&] (int pfd, int cfd) -> bool {
713 if (cfd == STDIN_FILENO) {
714 return handleWrite(pfd, input);
716 // If we don't want to write to this fd, just close it.
721 communicate(std::move(readCallback), std::move(writeCallback));
726 void Subprocess::communicate(FdCallback readCallback,
727 FdCallback writeCallback) {
728 // This serves to prevent wait() followed by communicate(), but if you
729 // legitimately need that, send a patch to delete this line.
730 returnCode_.enforce(ProcessReturnCode::RUNNING);
733 std::vector<pollfd> fds;
734 fds.reserve(pipes_.size());
735 std::vector<size_t> toClose; // indexes into pipes_
736 toClose.reserve(pipes_.size());
738 while (!pipes_.empty()) {
742 for (auto& p : pipes_) {
744 pfd.fd = p.pipe.fd();
745 // Yes, backwards, PIPE_IN / PIPE_OUT are defined from the
746 // child's point of view.
748 // Still keeping fd in watched set so we get notified of POLLHUP /
751 } else if (p.direction == PIPE_IN) {
752 pfd.events = POLLOUT;
761 r = ::poll(fds.data(), fds.size(), -1);
762 } while (r == -1 && errno == EINTR);
763 checkUnixError(r, "poll");
765 for (size_t i = 0; i < pipes_.size(); ++i) {
767 auto parentFd = p.pipe.fd();
768 DCHECK_EQ(fds[i].fd, parentFd);
769 short events = fds[i].revents;
772 if (events & POLLOUT) {
773 DCHECK(!(events & POLLIN));
774 if (writeCallback(parentFd, p.childFd)) {
775 toClose.push_back(i);
780 // Call read callback on POLLHUP, to give it a chance to read (and act
782 if (events & (POLLIN | POLLHUP)) {
783 DCHECK(!(events & POLLOUT));
784 if (readCallback(parentFd, p.childFd)) {
785 toClose.push_back(i);
790 if ((events & (POLLHUP | POLLERR)) && !closed) {
791 toClose.push_back(i);
796 // Close the fds in reverse order so the indexes hold after erase()
797 for (int idx : boost::adaptors::reverse(toClose)) {
798 auto pos = pipes_.begin() + idx;
799 pos->pipe.close(); // Throws on error
805 void Subprocess::enableNotifications(int childFd, bool enabled) {
806 pipes_[findByChildFd(childFd)].enabled = enabled;
809 bool Subprocess::notificationsEnabled(int childFd) const {
810 return pipes_[findByChildFd(childFd)].enabled;
813 size_t Subprocess::findByChildFd(int childFd) const {
814 auto pos = std::lower_bound(
815 pipes_.begin(), pipes_.end(), childFd,
816 [] (const Pipe& pipe, int fd) { return pipe.childFd < fd; });
817 if (pos == pipes_.end() || pos->childFd != childFd) {
818 throw std::invalid_argument(folly::to<std::string>(
819 "child fd not found ", childFd));
821 return pos - pipes_.begin();
824 void Subprocess::closeParentFd(int childFd) {
825 int idx = findByChildFd(childFd);
826 pipes_[idx].pipe.close(); // May throw
827 pipes_.erase(pipes_.begin() + idx);
830 std::vector<Subprocess::ChildPipe> Subprocess::takeOwnershipOfPipes() {
831 std::vector<Subprocess::ChildPipe> pipes;
832 for (auto& p : pipes_) {
833 pipes.emplace_back(p.childFd, std::move(p.pipe));
836 std::vector<Pipe>().swap(pipes_);
845 // We like EPIPE, thanks.
846 ::signal(SIGPIPE, SIG_IGN);
850 Initializer initializer;