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");
186 auto argv = Subprocess::shellify(cmd);
187 spawn(cloneStrings(argv), argv[0].c_str(), options, env);
190 /* static */ std::vector<std::string> Subprocess::shellify(
191 const std::string& cmd) {
192 std::vector<std::string> argv;
194 const char* shell = getenv("SHELL");
199 argv.push_back(shell);
200 argv.push_back("-c");
206 Subprocess::~Subprocess() {
207 CHECK_NE(returnCode_.state(), ProcessReturnCode::RUNNING)
208 << "Subprocess destroyed without reaping child";
213 struct ChildErrorInfo {
218 [[noreturn]] void childError(int errFd, int errCode, int errnoValue) {
219 ChildErrorInfo info = {errCode, errnoValue};
220 // Write the error information over the pipe to our parent process.
221 // We can't really do anything else if this write call fails.
222 writeNoInt(errFd, &info, sizeof(info));
229 void Subprocess::setAllNonBlocking() {
230 for (auto& p : pipes_) {
231 int fd = p.pipe.fd();
232 int flags = ::fcntl(fd, F_GETFL);
233 checkUnixError(flags, "fcntl");
234 int r = ::fcntl(fd, F_SETFL, flags | O_NONBLOCK);
235 checkUnixError(r, "fcntl");
239 void Subprocess::spawn(
240 std::unique_ptr<const char*[]> argv,
241 const char* executable,
242 const Options& optionsIn,
243 const std::vector<std::string>* env) {
244 if (optionsIn.usePath_ && env) {
245 throw std::invalid_argument(
246 "usePath() not allowed when overriding environment");
249 // Make a copy, we'll mutate options
250 Options options(optionsIn);
252 // On error, close all pipes_ (ignoring errors, but that seems fine here).
253 auto pipesGuard = makeGuard([this] { pipes_.clear(); });
255 // Create a pipe to use to receive error information from the child,
256 // in case it fails before calling exec()
259 checkUnixError(::pipe2(errFds, O_CLOEXEC), "pipe2");
261 checkUnixError(::pipe(errFds), "pipe");
264 CHECK_ERR(::close(errFds[0]));
265 if (errFds[1] >= 0) {
266 CHECK_ERR(::close(errFds[1]));
270 #if !FOLLY_HAVE_PIPE2
271 // Ask the child to close the read end of the error pipe.
272 checkUnixError(fcntl(errFds[0], F_SETFD, FD_CLOEXEC), "set FD_CLOEXEC");
273 // Set the close-on-exec flag on the write side of the pipe.
274 // This way the pipe will be closed automatically in the child if execve()
275 // succeeds. If the exec fails the child can write error information to the
277 checkUnixError(fcntl(errFds[1], F_SETFD, FD_CLOEXEC), "set FD_CLOEXEC");
280 // Perform the actual work of setting up pipes then forking and
281 // executing the child.
282 spawnInternal(std::move(argv), executable, options, env, errFds[1]);
284 // After spawnInternal() returns the child is alive. We have to be very
285 // careful about throwing after this point. We are inside the constructor,
286 // so if we throw the Subprocess object will have never existed, and the
287 // destructor will never be called.
289 // We should only throw if we got an error via the errFd, and we know the
290 // child has exited and can be immediately waited for. In all other cases,
291 // we have no way of cleaning up the child.
293 // Close writable side of the errFd pipe in the parent process
294 CHECK_ERR(::close(errFds[1]));
297 // Read from the errFd pipe, to tell if the child ran into any errors before
299 readChildErrorPipe(errFds[0], executable);
301 // We have fully succeeded now, so release the guard on pipes_
302 pipesGuard.dismiss();
305 void Subprocess::spawnInternal(
306 std::unique_ptr<const char*[]> argv,
307 const char* executable,
309 const std::vector<std::string>* env,
311 // Parent work, pre-fork: create pipes
312 std::vector<int> childFds;
313 // Close all of the childFds as we leave this scope
315 // These are only pipes, closing them shouldn't fail
316 for (int cfd : childFds) {
317 CHECK_ERR(::close(cfd));
322 for (auto& p : options.fdActions_) {
323 if (p.second == PIPE_IN || p.second == PIPE_OUT) {
325 // We're setting both ends of the pipe as close-on-exec. The child
326 // doesn't need to reset the flag on its end, as we always dup2() the fd,
327 // and dup2() fds don't share the close-on-exec flag.
329 // If possible, set close-on-exec atomically. Otherwise, a concurrent
330 // Subprocess invocation can fork() between "pipe" and "fnctl",
331 // causing FDs to leak.
332 r = ::pipe2(fds, O_CLOEXEC);
333 checkUnixError(r, "pipe2");
336 checkUnixError(r, "pipe");
337 r = fcntl(fds[0], F_SETFD, FD_CLOEXEC);
338 checkUnixError(r, "set FD_CLOEXEC");
339 r = fcntl(fds[1], F_SETFD, FD_CLOEXEC);
340 checkUnixError(r, "set FD_CLOEXEC");
342 pipes_.emplace_back();
343 Pipe& pipe = pipes_.back();
344 pipe.direction = p.second;
346 if (p.second == PIPE_IN) {
347 // Child gets reading end
348 pipe.pipe = folly::File(fds[1], /*owns_fd=*/ true);
351 pipe.pipe = folly::File(fds[0], /*owns_fd=*/ true);
354 p.second = cfd; // ensure it gets dup2()ed
355 pipe.childFd = p.first;
356 childFds.push_back(cfd);
360 // This should already be sorted, as options.fdActions_ is
361 DCHECK(std::is_sorted(pipes_.begin(), pipes_.end()));
363 // Note that the const casts below are legit, per
364 // http://pubs.opengroup.org/onlinepubs/009695399/functions/exec.html
366 char** argVec = const_cast<char**>(argv.get());
368 // Set up environment
369 std::unique_ptr<const char*[]> envHolder;
372 envHolder = cloneStrings(*env);
373 envVec = const_cast<char**>(envHolder.get());
378 // Block all signals around vfork; see http://ewontfix.com/7/.
380 // As the child may run in the same address space as the parent until
381 // the actual execve() system call, any (custom) signal handlers that
382 // the parent has might alter parent's memory if invoked in the child,
383 // with undefined results. So we block all signals in the parent before
384 // vfork(), which will cause them to be blocked in the child as well (we
385 // rely on the fact that Linux, just like all sane implementations, only
386 // clones the calling thread). Then, in the child, we reset all signals
387 // to their default dispositions (while still blocked), and unblock them
388 // (so the exec()ed process inherits the parent's signal mask)
390 // The parent also unblocks all signals as soon as vfork() returns.
392 r = sigfillset(&allBlocked);
393 checkUnixError(r, "sigfillset");
396 r = pthread_sigmask(SIG_SETMASK, &allBlocked, &oldSignals);
397 checkPosixError(r, "pthread_sigmask");
399 // Restore signal mask
400 r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
401 CHECK_EQ(r, 0) << "pthread_sigmask: " << errnoStr(r); // shouldn't fail
404 // Call c_str() here, as it's not necessarily safe after fork.
405 const char* childDir =
406 options.childDir_.empty() ? nullptr : options.childDir_.c_str();
409 int errnoValue = prepareChild(options, &oldSignals, childDir);
410 if (errnoValue != 0) {
411 childError(errFd, kChildFailure, errnoValue);
414 errnoValue = runChild(executable, argVec, envVec, options);
415 // If we get here, exec() failed.
416 childError(errFd, kExecFailure, errnoValue);
418 // In parent. Make sure vfork() succeeded.
419 checkUnixError(pid, errno, "vfork");
421 // Child is alive. We have to be very careful about throwing after this
422 // point. We are inside the constructor, so if we throw the Subprocess
423 // object will have never existed, and the destructor will never be called.
425 // We should only throw if we got an error via the errFd, and we know the
426 // child has exited and can be immediately waited for. In all other cases,
427 // we have no way of cleaning up the child.
429 returnCode_ = ProcessReturnCode(RV_RUNNING);
432 int Subprocess::prepareChild(const Options& options,
433 const sigset_t* sigmask,
434 const char* childDir) const {
435 // While all signals are blocked, we must reset their
436 // dispositions to default.
437 for (int sig = 1; sig < NSIG; ++sig) {
438 ::signal(sig, SIG_DFL);
442 // Unblock signals; restore signal mask.
443 int r = pthread_sigmask(SIG_SETMASK, sigmask, nullptr);
445 return r; // pthread_sigmask() returns an errno value
449 // Change the working directory, if one is given
451 if (::chdir(childDir) == -1) {
456 // We don't have to explicitly close the parent's end of all pipes,
457 // as they all have the FD_CLOEXEC flag set and will be closed at
460 // Close all fds that we're supposed to close.
461 for (auto& p : options.fdActions_) {
462 if (p.second == CLOSE) {
463 if (::close(p.first) == -1) {
466 } else if (p.second != p.first) {
467 if (::dup2(p.second, p.first) == -1) {
473 // If requested, close all other file descriptors. Don't close
474 // any fds in options.fdActions_, and don't touch stdin, stdout, stderr.
476 if (options.closeOtherFds_) {
477 for (int fd = getdtablesize() - 1; fd >= 3; --fd) {
478 if (options.fdActions_.count(fd) == 0) {
485 // Opt to receive signal on parent death, if requested
486 if (options.parentDeathSignal_ != 0) {
487 const auto parentDeathSignal =
488 static_cast<unsigned long>(options.parentDeathSignal_);
489 if (prctl(PR_SET_PDEATHSIG, parentDeathSignal, 0, 0, 0) == -1) {
495 if (options.processGroupLeader_) {
496 if (setpgrp() == -1) {
501 // The user callback comes last, so that the child is otherwise all set up.
502 if (options.dangerousPostForkPreExecCallback_) {
503 if (int error = (*options.dangerousPostForkPreExecCallback_)()) {
511 int Subprocess::runChild(const char* executable,
512 char** argv, char** env,
513 const Options& options) const {
514 // Now, finally, exec.
515 if (options.usePath_) {
516 ::execvp(executable, argv);
518 ::execve(executable, argv, env);
523 void Subprocess::readChildErrorPipe(int pfd, const char* executable) {
525 auto rc = readNoInt(pfd, &info, sizeof(info));
527 // No data means the child executed successfully, and the pipe
528 // was closed due to the close-on-exec flag being set.
530 } else if (rc != sizeof(ChildErrorInfo)) {
531 // An error occurred trying to read from the pipe, or we got a partial read.
532 // Neither of these cases should really occur in practice.
534 // We can't get any error data from the child in this case, and we don't
535 // know if it is successfully running or not. All we can do is to return
536 // normally, as if the child executed successfully. If something bad
537 // happened the caller should at least get a non-normal exit status from
539 LOG(ERROR) << "unexpected error trying to read from child error pipe " <<
540 "rc=" << rc << ", errno=" << errno;
544 // We got error data from the child. The child should exit immediately in
545 // this case, so wait on it to clean up.
548 // Throw to signal the error
549 throw SubprocessSpawnError(executable, info.errCode, info.errnoValue);
552 ProcessReturnCode Subprocess::poll() {
553 returnCode_.enforce(ProcessReturnCode::RUNNING);
556 pid_t found = ::waitpid(pid_, &status, WNOHANG);
557 // The spec guarantees that EINTR does not occur with WNOHANG, so the only
558 // two remaining errors are ECHILD (other code reaped the child?), or
559 // EINVAL (cosmic rays?), both of which merit an abort:
560 PCHECK(found != -1) << "waitpid(" << pid_ << ", &status, WNOHANG)";
562 // Though the child process had quit, this call does not close the pipes
563 // since its descendants may still be using them.
564 returnCode_ = ProcessReturnCode(status);
570 bool Subprocess::pollChecked() {
571 if (poll().state() == ProcessReturnCode::RUNNING) {
574 checkStatus(returnCode_);
578 ProcessReturnCode Subprocess::wait() {
579 returnCode_.enforce(ProcessReturnCode::RUNNING);
584 found = ::waitpid(pid_, &status, 0);
585 } while (found == -1 && errno == EINTR);
586 // The only two remaining errors are ECHILD (other code reaped the
587 // child?), or EINVAL (cosmic rays?), and both merit an abort:
588 PCHECK(found != -1) << "waitpid(" << pid_ << ", &status, WNOHANG)";
589 // Though the child process had quit, this call does not close the pipes
590 // since its descendants may still be using them.
591 DCHECK_EQ(found, pid_);
592 returnCode_ = ProcessReturnCode(status);
597 void Subprocess::waitChecked() {
599 checkStatus(returnCode_);
602 void Subprocess::sendSignal(int signal) {
603 returnCode_.enforce(ProcessReturnCode::RUNNING);
604 int r = ::kill(pid_, signal);
605 checkUnixError(r, "kill");
608 pid_t Subprocess::pid() const {
614 ByteRange queueFront(const IOBufQueue& queue) {
615 auto* p = queue.front();
619 return io::Cursor(p).peekBytes();
623 bool handleWrite(int fd, IOBufQueue& queue) {
625 auto b = queueFront(queue);
630 ssize_t n = writeNoInt(fd, b.data(), b.size());
631 if (n == -1 && errno == EAGAIN) {
634 checkUnixError(n, "write");
640 bool handleRead(int fd, IOBufQueue& queue) {
642 auto p = queue.preallocate(100, 65000);
643 ssize_t n = readNoInt(fd, p.first, p.second);
644 if (n == -1 && errno == EAGAIN) {
647 checkUnixError(n, "read");
651 queue.postallocate(n);
655 bool discardRead(int fd) {
656 static const size_t bufSize = 65000;
657 // Thread unsafe, but it doesn't matter.
658 static std::unique_ptr<char[]> buf(new char[bufSize]);
661 ssize_t n = readNoInt(fd, buf.get(), bufSize);
662 if (n == -1 && errno == EAGAIN) {
665 checkUnixError(n, "read");
674 std::pair<std::string, std::string> Subprocess::communicate(
676 IOBufQueue inputQueue;
677 inputQueue.wrapBuffer(input.data(), input.size());
679 auto outQueues = communicateIOBuf(std::move(inputQueue));
680 auto outBufs = std::make_pair(outQueues.first.move(),
681 outQueues.second.move());
682 std::pair<std::string, std::string> out;
684 outBufs.first->coalesce();
685 out.first.assign(reinterpret_cast<const char*>(outBufs.first->data()),
686 outBufs.first->length());
688 if (outBufs.second) {
689 outBufs.second->coalesce();
690 out.second.assign(reinterpret_cast<const char*>(outBufs.second->data()),
691 outBufs.second->length());
696 std::pair<IOBufQueue, IOBufQueue> Subprocess::communicateIOBuf(
698 // If the user supplied a non-empty input buffer, make sure
699 // that stdin is a pipe so we can write the data.
700 if (!input.empty()) {
701 // findByChildFd() will throw std::invalid_argument if no pipe for
702 // STDIN_FILENO exists
703 findByChildFd(STDIN_FILENO);
706 std::pair<IOBufQueue, IOBufQueue> out;
708 auto readCallback = [&] (int pfd, int cfd) -> bool {
709 if (cfd == STDOUT_FILENO) {
710 return handleRead(pfd, out.first);
711 } else if (cfd == STDERR_FILENO) {
712 return handleRead(pfd, out.second);
714 // Don't close the file descriptor, the child might not like SIGPIPE,
715 // just read and throw the data away.
716 return discardRead(pfd);
720 auto writeCallback = [&] (int pfd, int cfd) -> bool {
721 if (cfd == STDIN_FILENO) {
722 return handleWrite(pfd, input);
724 // If we don't want to write to this fd, just close it.
729 communicate(std::move(readCallback), std::move(writeCallback));
734 void Subprocess::communicate(FdCallback readCallback,
735 FdCallback writeCallback) {
736 // This serves to prevent wait() followed by communicate(), but if you
737 // legitimately need that, send a patch to delete this line.
738 returnCode_.enforce(ProcessReturnCode::RUNNING);
741 std::vector<pollfd> fds;
742 fds.reserve(pipes_.size());
743 std::vector<size_t> toClose; // indexes into pipes_
744 toClose.reserve(pipes_.size());
746 while (!pipes_.empty()) {
750 for (auto& p : pipes_) {
752 pfd.fd = p.pipe.fd();
753 // Yes, backwards, PIPE_IN / PIPE_OUT are defined from the
754 // child's point of view.
756 // Still keeping fd in watched set so we get notified of POLLHUP /
759 } else if (p.direction == PIPE_IN) {
760 pfd.events = POLLOUT;
769 r = ::poll(fds.data(), fds.size(), -1);
770 } while (r == -1 && errno == EINTR);
771 checkUnixError(r, "poll");
773 for (size_t i = 0; i < pipes_.size(); ++i) {
775 auto parentFd = p.pipe.fd();
776 DCHECK_EQ(fds[i].fd, parentFd);
777 short events = fds[i].revents;
780 if (events & POLLOUT) {
781 DCHECK(!(events & POLLIN));
782 if (writeCallback(parentFd, p.childFd)) {
783 toClose.push_back(i);
788 // Call read callback on POLLHUP, to give it a chance to read (and act
790 if (events & (POLLIN | POLLHUP)) {
791 DCHECK(!(events & POLLOUT));
792 if (readCallback(parentFd, p.childFd)) {
793 toClose.push_back(i);
798 if ((events & (POLLHUP | POLLERR)) && !closed) {
799 toClose.push_back(i);
804 // Close the fds in reverse order so the indexes hold after erase()
805 for (int idx : boost::adaptors::reverse(toClose)) {
806 auto pos = pipes_.begin() + idx;
807 pos->pipe.close(); // Throws on error
813 void Subprocess::enableNotifications(int childFd, bool enabled) {
814 pipes_[findByChildFd(childFd)].enabled = enabled;
817 bool Subprocess::notificationsEnabled(int childFd) const {
818 return pipes_[findByChildFd(childFd)].enabled;
821 size_t Subprocess::findByChildFd(int childFd) const {
822 auto pos = std::lower_bound(
823 pipes_.begin(), pipes_.end(), childFd,
824 [] (const Pipe& pipe, int fd) { return pipe.childFd < fd; });
825 if (pos == pipes_.end() || pos->childFd != childFd) {
826 throw std::invalid_argument(folly::to<std::string>(
827 "child fd not found ", childFd));
829 return pos - pipes_.begin();
832 void Subprocess::closeParentFd(int childFd) {
833 int idx = findByChildFd(childFd);
834 pipes_[idx].pipe.close(); // May throw
835 pipes_.erase(pipes_.begin() + idx);
838 std::vector<Subprocess::ChildPipe> Subprocess::takeOwnershipOfPipes() {
839 std::vector<Subprocess::ChildPipe> pipes;
840 for (auto& p : pipes_) {
841 pipes.emplace_back(p.childFd, std::move(p.pipe));
844 std::vector<Pipe>().swap(pipes_);
853 // We like EPIPE, thanks.
854 ::signal(SIGPIPE, SIG_IGN);
858 Initializer initializer;