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>
33 #include <system_error>
35 #include <boost/container/flat_set.hpp>
36 #include <boost/range/adaptors.hpp>
38 #include <glog/logging.h>
40 #include <folly/Conv.h>
41 #include <folly/Exception.h>
42 #include <folly/ScopeGuard.h>
43 #include <folly/String.h>
44 #include <folly/io/Cursor.h>
45 #include <folly/portability/Environment.h>
47 constexpr int kExecFailure = 127;
48 constexpr int kChildFailure = 126;
52 ProcessReturnCode::ProcessReturnCode(ProcessReturnCode&& p) noexcept
53 : rawStatus_(p.rawStatus_) {
54 p.rawStatus_ = ProcessReturnCode::RV_NOT_STARTED;
57 ProcessReturnCode& ProcessReturnCode::operator=(ProcessReturnCode&& p)
59 rawStatus_ = p.rawStatus_;
60 p.rawStatus_ = ProcessReturnCode::RV_NOT_STARTED;
64 ProcessReturnCode::State ProcessReturnCode::state() const {
65 if (rawStatus_ == RV_NOT_STARTED) return NOT_STARTED;
66 if (rawStatus_ == RV_RUNNING) return RUNNING;
67 if (WIFEXITED(rawStatus_)) return EXITED;
68 if (WIFSIGNALED(rawStatus_)) return KILLED;
69 throw std::runtime_error(to<std::string>(
70 "Invalid ProcessReturnCode: ", rawStatus_));
73 void ProcessReturnCode::enforce(State expected) const {
76 throw std::logic_error(to<std::string>(
77 "Bad use of ProcessReturnCode; state is ", s, " expected ", expected
82 int ProcessReturnCode::exitStatus() const {
84 return WEXITSTATUS(rawStatus_);
87 int ProcessReturnCode::killSignal() const {
89 return WTERMSIG(rawStatus_);
92 bool ProcessReturnCode::coreDumped() const {
94 return WCOREDUMP(rawStatus_);
97 std::string ProcessReturnCode::str() const {
100 return "not started";
104 return to<std::string>("exited with status ", exitStatus());
106 return to<std::string>("killed by signal ", killSignal(),
107 (coreDumped() ? " (core dumped)" : ""));
109 CHECK(false); // unreached
110 return ""; // silence GCC warning
113 CalledProcessError::CalledProcessError(ProcessReturnCode rc)
115 what_(returnCode_.str()) {
118 SubprocessSpawnError::SubprocessSpawnError(const char* executable,
121 : errnoValue_(errnoValue),
122 what_(to<std::string>(errCode == kExecFailure ?
123 "failed to execute " :
124 "error preparing to execute ",
125 executable, ": ", errnoStr(errnoValue))) {
130 // Copy pointers to the given strings in a format suitable for posix_spawn
131 std::unique_ptr<const char*[]> cloneStrings(const std::vector<std::string>& s) {
132 std::unique_ptr<const char*[]> d(new const char*[s.size() + 1]);
133 for (size_t i = 0; i < s.size(); i++) {
136 d[s.size()] = nullptr;
140 // Check a wait() status, throw on non-successful
141 void checkStatus(ProcessReturnCode returnCode) {
142 if (returnCode.state() != ProcessReturnCode::EXITED ||
143 returnCode.exitStatus() != 0) {
144 throw CalledProcessError(returnCode);
150 Subprocess::Options& Subprocess::Options::fd(int fd, int action) {
151 if (action == Subprocess::PIPE) {
153 action = Subprocess::PIPE_IN;
154 } else if (fd == 1 || fd == 2) {
155 action = Subprocess::PIPE_OUT;
157 throw std::invalid_argument(
158 to<std::string>("Only fds 0, 1, 2 are valid for action=PIPE: ", fd));
161 fdActions_[fd] = action;
165 Subprocess::Subprocess(
166 const std::vector<std::string>& argv,
167 const Options& options,
168 const char* executable,
169 const std::vector<std::string>* env)
171 returnCode_(RV_NOT_STARTED) {
173 throw std::invalid_argument("argv must not be empty");
175 if (!executable) executable = argv[0].c_str();
176 spawn(cloneStrings(argv), executable, options, env);
179 Subprocess::Subprocess(
180 const std::string& cmd,
181 const Options& options,
182 const std::vector<std::string>* env)
184 returnCode_(RV_NOT_STARTED) {
185 if (options.usePath_) {
186 throw std::invalid_argument("usePath() not allowed when running in shell");
188 const char* shell = getenv("SHELL");
193 std::unique_ptr<const char*[]> argv(new const char*[4]);
196 argv[2] = cmd.c_str();
198 spawn(std::move(argv), shell, options, env);
201 Subprocess::~Subprocess() {
202 CHECK_NE(returnCode_.state(), ProcessReturnCode::RUNNING)
203 << "Subprocess destroyed without reaping child";
208 struct ChildErrorInfo {
213 [[noreturn]] void childError(int errFd, int errCode, int errnoValue) {
214 ChildErrorInfo info = {errCode, errnoValue};
215 // Write the error information over the pipe to our parent process.
216 // We can't really do anything else if this write call fails.
217 writeNoInt(errFd, &info, sizeof(info));
224 void Subprocess::setAllNonBlocking() {
225 for (auto& p : pipes_) {
226 int fd = p.pipe.fd();
227 int flags = ::fcntl(fd, F_GETFL);
228 checkUnixError(flags, "fcntl");
229 int r = ::fcntl(fd, F_SETFL, flags | O_NONBLOCK);
230 checkUnixError(r, "fcntl");
234 void Subprocess::spawn(
235 std::unique_ptr<const char*[]> argv,
236 const char* executable,
237 const Options& optionsIn,
238 const std::vector<std::string>* env) {
239 if (optionsIn.usePath_ && env) {
240 throw std::invalid_argument(
241 "usePath() not allowed when overriding environment");
244 // Make a copy, we'll mutate options
245 Options options(optionsIn);
247 // On error, close all pipes_ (ignoring errors, but that seems fine here).
248 auto pipesGuard = makeGuard([this] { pipes_.clear(); });
250 // Create a pipe to use to receive error information from the child,
251 // in case it fails before calling exec()
254 checkUnixError(::pipe2(errFds, O_CLOEXEC), "pipe2");
256 checkUnixError(::pipe(errFds), "pipe");
259 CHECK_ERR(::close(errFds[0]));
260 if (errFds[1] >= 0) {
261 CHECK_ERR(::close(errFds[1]));
265 #if !FOLLY_HAVE_PIPE2
266 // Ask the child to close the read end of the error pipe.
267 checkUnixError(fcntl(errFds[0], F_SETFD, FD_CLOEXEC), "set FD_CLOEXEC");
268 // Set the close-on-exec flag on the write side of the pipe.
269 // This way the pipe will be closed automatically in the child if execve()
270 // succeeds. If the exec fails the child can write error information to the
272 checkUnixError(fcntl(errFds[1], F_SETFD, FD_CLOEXEC), "set FD_CLOEXEC");
275 // Perform the actual work of setting up pipes then forking and
276 // executing the child.
277 spawnInternal(std::move(argv), executable, options, env, errFds[1]);
279 // After spawnInternal() returns the child is alive. We have to be very
280 // careful about throwing after this point. We are inside the constructor,
281 // so if we throw the Subprocess object will have never existed, and the
282 // destructor will never be called.
284 // We should only throw if we got an error via the errFd, and we know the
285 // child has exited and can be immediately waited for. In all other cases,
286 // we have no way of cleaning up the child.
288 // Close writable side of the errFd pipe in the parent process
289 CHECK_ERR(::close(errFds[1]));
292 // Read from the errFd pipe, to tell if the child ran into any errors before
294 readChildErrorPipe(errFds[0], executable);
296 // We have fully succeeded now, so release the guard on pipes_
297 pipesGuard.dismiss();
300 void Subprocess::spawnInternal(
301 std::unique_ptr<const char*[]> argv,
302 const char* executable,
304 const std::vector<std::string>* env,
306 // Parent work, pre-fork: create pipes
307 std::vector<int> childFds;
308 // Close all of the childFds as we leave this scope
310 // These are only pipes, closing them shouldn't fail
311 for (int cfd : childFds) {
312 CHECK_ERR(::close(cfd));
317 for (auto& p : options.fdActions_) {
318 if (p.second == PIPE_IN || p.second == PIPE_OUT) {
320 // We're setting both ends of the pipe as close-on-exec. The child
321 // doesn't need to reset the flag on its end, as we always dup2() the fd,
322 // and dup2() fds don't share the close-on-exec flag.
324 // If possible, set close-on-exec atomically. Otherwise, a concurrent
325 // Subprocess invocation can fork() between "pipe" and "fnctl",
326 // causing FDs to leak.
327 r = ::pipe2(fds, O_CLOEXEC);
328 checkUnixError(r, "pipe2");
331 checkUnixError(r, "pipe");
332 r = fcntl(fds[0], F_SETFD, FD_CLOEXEC);
333 checkUnixError(r, "set FD_CLOEXEC");
334 r = fcntl(fds[1], F_SETFD, FD_CLOEXEC);
335 checkUnixError(r, "set FD_CLOEXEC");
337 pipes_.emplace_back();
338 Pipe& pipe = pipes_.back();
339 pipe.direction = p.second;
341 if (p.second == PIPE_IN) {
342 // Child gets reading end
343 pipe.pipe = folly::File(fds[1], /*owns_fd=*/ true);
346 pipe.pipe = folly::File(fds[0], /*owns_fd=*/ true);
349 p.second = cfd; // ensure it gets dup2()ed
350 pipe.childFd = p.first;
351 childFds.push_back(cfd);
355 // This should already be sorted, as options.fdActions_ is
356 DCHECK(std::is_sorted(pipes_.begin(), pipes_.end()));
358 // Note that the const casts below are legit, per
359 // http://pubs.opengroup.org/onlinepubs/009695399/functions/exec.html
361 char** argVec = const_cast<char**>(argv.get());
363 // Set up environment
364 std::unique_ptr<const char*[]> envHolder;
367 envHolder = cloneStrings(*env);
368 envVec = const_cast<char**>(envHolder.get());
373 // Block all signals around vfork; see http://ewontfix.com/7/.
375 // As the child may run in the same address space as the parent until
376 // the actual execve() system call, any (custom) signal handlers that
377 // the parent has might alter parent's memory if invoked in the child,
378 // with undefined results. So we block all signals in the parent before
379 // vfork(), which will cause them to be blocked in the child as well (we
380 // rely on the fact that Linux, just like all sane implementations, only
381 // clones the calling thread). Then, in the child, we reset all signals
382 // to their default dispositions (while still blocked), and unblock them
383 // (so the exec()ed process inherits the parent's signal mask)
385 // The parent also unblocks all signals as soon as vfork() returns.
387 r = sigfillset(&allBlocked);
388 checkUnixError(r, "sigfillset");
391 r = pthread_sigmask(SIG_SETMASK, &allBlocked, &oldSignals);
392 checkPosixError(r, "pthread_sigmask");
394 // Restore signal mask
395 r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
396 CHECK_EQ(r, 0) << "pthread_sigmask: " << errnoStr(r); // shouldn't fail
399 // Call c_str() here, as it's not necessarily safe after fork.
400 const char* childDir =
401 options.childDir_.empty() ? nullptr : options.childDir_.c_str();
404 int errnoValue = prepareChild(options, &oldSignals, childDir);
405 if (errnoValue != 0) {
406 childError(errFd, kChildFailure, errnoValue);
409 errnoValue = runChild(executable, argVec, envVec, options);
410 // If we get here, exec() failed.
411 childError(errFd, kExecFailure, errnoValue);
413 // In parent. Make sure vfork() succeeded.
414 checkUnixError(pid, errno, "vfork");
416 // Child is alive. We have to be very careful about throwing after this
417 // point. We are inside the constructor, so if we throw the Subprocess
418 // object will have never existed, and the destructor will never be called.
420 // We should only throw if we got an error via the errFd, and we know the
421 // child has exited and can be immediately waited for. In all other cases,
422 // we have no way of cleaning up the child.
424 returnCode_ = ProcessReturnCode(RV_RUNNING);
427 int Subprocess::prepareChild(const Options& options,
428 const sigset_t* sigmask,
429 const char* childDir) const {
430 // While all signals are blocked, we must reset their
431 // dispositions to default.
432 for (int sig = 1; sig < NSIG; ++sig) {
433 ::signal(sig, SIG_DFL);
437 // Unblock signals; restore signal mask.
438 int r = pthread_sigmask(SIG_SETMASK, sigmask, nullptr);
440 return r; // pthread_sigmask() returns an errno value
444 // Change the working directory, if one is given
446 if (::chdir(childDir) == -1) {
451 // We don't have to explicitly close the parent's end of all pipes,
452 // as they all have the FD_CLOEXEC flag set and will be closed at
455 // Close all fds that we're supposed to close.
456 for (auto& p : options.fdActions_) {
457 if (p.second == CLOSE) {
458 if (::close(p.first) == -1) {
461 } else if (p.second != p.first) {
462 if (::dup2(p.second, p.first) == -1) {
468 // If requested, close all other file descriptors. Don't close
469 // any fds in options.fdActions_, and don't touch stdin, stdout, stderr.
471 if (options.closeOtherFds_) {
472 for (int fd = getdtablesize() - 1; fd >= 3; --fd) {
473 if (options.fdActions_.count(fd) == 0) {
480 // Opt to receive signal on parent death, if requested
481 if (options.parentDeathSignal_ != 0) {
482 const auto parentDeathSignal =
483 static_cast<unsigned long>(options.parentDeathSignal_);
484 if (prctl(PR_SET_PDEATHSIG, parentDeathSignal, 0, 0, 0) == -1) {
490 if (options.processGroupLeader_) {
491 if (setpgrp() == -1) {
496 // The user callback comes last, so that the child is otherwise all set up.
497 if (options.dangerousPostForkPreExecCallback_) {
498 if (int error = (*options.dangerousPostForkPreExecCallback_)()) {
506 int Subprocess::runChild(const char* executable,
507 char** argv, char** env,
508 const Options& options) const {
509 // Now, finally, exec.
510 if (options.usePath_) {
511 ::execvp(executable, argv);
513 ::execve(executable, argv, env);
518 void Subprocess::readChildErrorPipe(int pfd, const char* executable) {
520 auto rc = readNoInt(pfd, &info, sizeof(info));
522 // No data means the child executed successfully, and the pipe
523 // was closed due to the close-on-exec flag being set.
525 } else if (rc != sizeof(ChildErrorInfo)) {
526 // An error occurred trying to read from the pipe, or we got a partial read.
527 // Neither of these cases should really occur in practice.
529 // We can't get any error data from the child in this case, and we don't
530 // know if it is successfully running or not. All we can do is to return
531 // normally, as if the child executed successfully. If something bad
532 // happened the caller should at least get a non-normal exit status from
534 LOG(ERROR) << "unexpected error trying to read from child error pipe " <<
535 "rc=" << rc << ", errno=" << errno;
539 // We got error data from the child. The child should exit immediately in
540 // this case, so wait on it to clean up.
543 // Throw to signal the error
544 throw SubprocessSpawnError(executable, info.errCode, info.errnoValue);
547 ProcessReturnCode Subprocess::poll() {
548 returnCode_.enforce(ProcessReturnCode::RUNNING);
551 pid_t found = ::waitpid(pid_, &status, WNOHANG);
552 // The spec guarantees that EINTR does not occur with WNOHANG, so the only
553 // two remaining errors are ECHILD (other code reaped the child?), or
554 // EINVAL (cosmic rays?), both of which merit an abort:
555 PCHECK(found != -1) << "waitpid(" << pid_ << ", &status, WNOHANG)";
557 // Though the child process had quit, this call does not close the pipes
558 // since its descendants may still be using them.
559 returnCode_ = ProcessReturnCode(status);
565 bool Subprocess::pollChecked() {
566 if (poll().state() == ProcessReturnCode::RUNNING) {
569 checkStatus(returnCode_);
573 ProcessReturnCode Subprocess::wait() {
574 returnCode_.enforce(ProcessReturnCode::RUNNING);
579 found = ::waitpid(pid_, &status, 0);
580 } while (found == -1 && errno == EINTR);
581 // The only two remaining errors are ECHILD (other code reaped the
582 // child?), or EINVAL (cosmic rays?), and both merit an abort:
583 PCHECK(found != -1) << "waitpid(" << pid_ << ", &status, WNOHANG)";
584 // Though the child process had quit, this call does not close the pipes
585 // since its descendants may still be using them.
586 DCHECK_EQ(found, pid_);
587 returnCode_ = ProcessReturnCode(status);
592 void Subprocess::waitChecked() {
594 checkStatus(returnCode_);
597 void Subprocess::sendSignal(int signal) {
598 returnCode_.enforce(ProcessReturnCode::RUNNING);
599 int r = ::kill(pid_, signal);
600 checkUnixError(r, "kill");
603 pid_t Subprocess::pid() const {
609 std::pair<const uint8_t*, size_t> queueFront(const IOBufQueue& queue) {
610 auto* p = queue.front();
611 if (!p) return std::make_pair(nullptr, 0);
612 return io::Cursor(p).peek();
616 bool handleWrite(int fd, IOBufQueue& queue) {
618 auto p = queueFront(queue);
623 ssize_t n = writeNoInt(fd, p.first, p.second);
624 if (n == -1 && errno == EAGAIN) {
627 checkUnixError(n, "write");
633 bool handleRead(int fd, IOBufQueue& queue) {
635 auto p = queue.preallocate(100, 65000);
636 ssize_t n = readNoInt(fd, p.first, p.second);
637 if (n == -1 && errno == EAGAIN) {
640 checkUnixError(n, "read");
644 queue.postallocate(n);
648 bool discardRead(int fd) {
649 static const size_t bufSize = 65000;
650 // Thread unsafe, but it doesn't matter.
651 static std::unique_ptr<char[]> buf(new char[bufSize]);
654 ssize_t n = readNoInt(fd, buf.get(), bufSize);
655 if (n == -1 && errno == EAGAIN) {
658 checkUnixError(n, "read");
667 std::pair<std::string, std::string> Subprocess::communicate(
669 IOBufQueue inputQueue;
670 inputQueue.wrapBuffer(input.data(), input.size());
672 auto outQueues = communicateIOBuf(std::move(inputQueue));
673 auto outBufs = std::make_pair(outQueues.first.move(),
674 outQueues.second.move());
675 std::pair<std::string, std::string> out;
677 outBufs.first->coalesce();
678 out.first.assign(reinterpret_cast<const char*>(outBufs.first->data()),
679 outBufs.first->length());
681 if (outBufs.second) {
682 outBufs.second->coalesce();
683 out.second.assign(reinterpret_cast<const char*>(outBufs.second->data()),
684 outBufs.second->length());
689 std::pair<IOBufQueue, IOBufQueue> Subprocess::communicateIOBuf(
691 // If the user supplied a non-empty input buffer, make sure
692 // that stdin is a pipe so we can write the data.
693 if (!input.empty()) {
694 // findByChildFd() will throw std::invalid_argument if no pipe for
695 // STDIN_FILENO exists
696 findByChildFd(STDIN_FILENO);
699 std::pair<IOBufQueue, IOBufQueue> out;
701 auto readCallback = [&] (int pfd, int cfd) -> bool {
702 if (cfd == STDOUT_FILENO) {
703 return handleRead(pfd, out.first);
704 } else if (cfd == STDERR_FILENO) {
705 return handleRead(pfd, out.second);
707 // Don't close the file descriptor, the child might not like SIGPIPE,
708 // just read and throw the data away.
709 return discardRead(pfd);
713 auto writeCallback = [&] (int pfd, int cfd) -> bool {
714 if (cfd == STDIN_FILENO) {
715 return handleWrite(pfd, input);
717 // If we don't want to write to this fd, just close it.
722 communicate(std::move(readCallback), std::move(writeCallback));
727 void Subprocess::communicate(FdCallback readCallback,
728 FdCallback writeCallback) {
729 // This serves to prevent wait() followed by communicate(), but if you
730 // legitimately need that, send a patch to delete this line.
731 returnCode_.enforce(ProcessReturnCode::RUNNING);
734 std::vector<pollfd> fds;
735 fds.reserve(pipes_.size());
736 std::vector<size_t> toClose; // indexes into pipes_
737 toClose.reserve(pipes_.size());
739 while (!pipes_.empty()) {
743 for (auto& p : pipes_) {
745 pfd.fd = p.pipe.fd();
746 // Yes, backwards, PIPE_IN / PIPE_OUT are defined from the
747 // child's point of view.
749 // Still keeping fd in watched set so we get notified of POLLHUP /
752 } else if (p.direction == PIPE_IN) {
753 pfd.events = POLLOUT;
762 r = ::poll(fds.data(), fds.size(), -1);
763 } while (r == -1 && errno == EINTR);
764 checkUnixError(r, "poll");
766 for (size_t i = 0; i < pipes_.size(); ++i) {
768 auto parentFd = p.pipe.fd();
769 DCHECK_EQ(fds[i].fd, parentFd);
770 short events = fds[i].revents;
773 if (events & POLLOUT) {
774 DCHECK(!(events & POLLIN));
775 if (writeCallback(parentFd, p.childFd)) {
776 toClose.push_back(i);
781 // Call read callback on POLLHUP, to give it a chance to read (and act
783 if (events & (POLLIN | POLLHUP)) {
784 DCHECK(!(events & POLLOUT));
785 if (readCallback(parentFd, p.childFd)) {
786 toClose.push_back(i);
791 if ((events & (POLLHUP | POLLERR)) && !closed) {
792 toClose.push_back(i);
797 // Close the fds in reverse order so the indexes hold after erase()
798 for (int idx : boost::adaptors::reverse(toClose)) {
799 auto pos = pipes_.begin() + idx;
800 pos->pipe.close(); // Throws on error
806 void Subprocess::enableNotifications(int childFd, bool enabled) {
807 pipes_[findByChildFd(childFd)].enabled = enabled;
810 bool Subprocess::notificationsEnabled(int childFd) const {
811 return pipes_[findByChildFd(childFd)].enabled;
814 size_t Subprocess::findByChildFd(int childFd) const {
815 auto pos = std::lower_bound(
816 pipes_.begin(), pipes_.end(), childFd,
817 [] (const Pipe& pipe, int fd) { return pipe.childFd < fd; });
818 if (pos == pipes_.end() || pos->childFd != childFd) {
819 throw std::invalid_argument(folly::to<std::string>(
820 "child fd not found ", childFd));
822 return pos - pipes_.begin();
825 void Subprocess::closeParentFd(int childFd) {
826 int idx = findByChildFd(childFd);
827 pipes_[idx].pipe.close(); // May throw
828 pipes_.erase(pipes_.begin() + idx);
831 std::vector<Subprocess::ChildPipe> Subprocess::takeOwnershipOfPipes() {
832 std::vector<Subprocess::ChildPipe> pipes;
833 for (auto& p : pipes_) {
834 pipes.emplace_back(p.childFd, std::move(p.pipe));
845 // We like EPIPE, thanks.
846 ::signal(SIGPIPE, SIG_IGN);
850 Initializer initializer;