2 * Copyright 2014 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.
17 #include "folly/Subprocess.h"
20 #include <sys/prctl.h>
32 #include <system_error>
34 #include <boost/container/flat_set.hpp>
35 #include <boost/range/adaptors.hpp>
37 #include <glog/logging.h>
39 #include "folly/Conv.h"
40 #include "folly/Exception.h"
41 #include "folly/FileUtil.h"
42 #include "folly/ScopeGuard.h"
43 #include "folly/String.h"
44 #include "folly/io/Cursor.h"
46 extern char** environ;
48 constexpr int kExecFailure = 127;
49 constexpr int kChildFailure = 126;
53 ProcessReturnCode::State ProcessReturnCode::state() const {
54 if (rawStatus_ == RV_NOT_STARTED) return NOT_STARTED;
55 if (rawStatus_ == RV_RUNNING) return RUNNING;
56 if (WIFEXITED(rawStatus_)) return EXITED;
57 if (WIFSIGNALED(rawStatus_)) return KILLED;
58 throw std::runtime_error(to<std::string>(
59 "Invalid ProcessReturnCode: ", rawStatus_));
62 void ProcessReturnCode::enforce(State expected) const {
65 throw std::logic_error(to<std::string>(
66 "Bad use of ProcessReturnCode; state is ", s, " expected ", expected
71 int ProcessReturnCode::exitStatus() const {
73 return WEXITSTATUS(rawStatus_);
76 int ProcessReturnCode::killSignal() const {
78 return WTERMSIG(rawStatus_);
81 bool ProcessReturnCode::coreDumped() const {
83 return WCOREDUMP(rawStatus_);
86 std::string ProcessReturnCode::str() const {
93 return to<std::string>("exited with status ", exitStatus());
95 return to<std::string>("killed by signal ", killSignal(),
96 (coreDumped() ? " (core dumped)" : ""));
98 CHECK(false); // unreached
101 CalledProcessError::CalledProcessError(ProcessReturnCode rc)
103 what_(returnCode_.str()) {
106 SubprocessSpawnError::SubprocessSpawnError(const char* executable,
109 : errnoValue_(errnoValue),
110 what_(to<std::string>(errCode == kExecFailure ?
111 "failed to execute " :
112 "error preparing to execute ",
113 executable, ": ", errnoStr(errnoValue))) {
118 // Copy pointers to the given strings in a format suitable for posix_spawn
119 std::unique_ptr<const char*[]> cloneStrings(const std::vector<std::string>& s) {
120 std::unique_ptr<const char*[]> d(new const char*[s.size() + 1]);
121 for (int i = 0; i < s.size(); i++) {
124 d[s.size()] = nullptr;
128 // Check a wait() status, throw on non-successful
129 void checkStatus(ProcessReturnCode returnCode) {
130 if (returnCode.state() != ProcessReturnCode::EXITED ||
131 returnCode.exitStatus() != 0) {
132 throw CalledProcessError(returnCode);
138 Subprocess::Options& Subprocess::Options::fd(int fd, int action) {
139 if (action == Subprocess::PIPE) {
141 action = Subprocess::PIPE_IN;
142 } else if (fd == 1 || fd == 2) {
143 action = Subprocess::PIPE_OUT;
145 throw std::invalid_argument(
146 to<std::string>("Only fds 0, 1, 2 are valid for action=PIPE: ", fd));
149 fdActions_[fd] = action;
153 Subprocess::Subprocess(
154 const std::vector<std::string>& argv,
155 const Options& options,
156 const char* executable,
157 const std::vector<std::string>* env)
159 returnCode_(RV_NOT_STARTED) {
161 throw std::invalid_argument("argv must not be empty");
163 if (!executable) executable = argv[0].c_str();
164 spawn(cloneStrings(argv), executable, options, env);
167 Subprocess::Subprocess(
168 const std::string& cmd,
169 const Options& options,
170 const std::vector<std::string>* env)
172 returnCode_(RV_NOT_STARTED) {
173 if (options.usePath_) {
174 throw std::invalid_argument("usePath() not allowed when running in shell");
176 const char* shell = getenv("SHELL");
181 std::unique_ptr<const char*[]> argv(new const char*[4]);
184 argv[2] = cmd.c_str();
186 spawn(std::move(argv), shell, options, env);
189 Subprocess::~Subprocess() {
190 CHECK_NE(returnCode_.state(), ProcessReturnCode::RUNNING)
191 << "Subprocess destroyed without reaping child";
196 void closeChecked(int fd) {
197 checkUnixError(::close(fd), "close");
200 struct ChildErrorInfo {
205 FOLLY_NORETURN void childError(int errFd, int errCode, int errnoValue);
206 void childError(int errFd, int errCode, int errnoValue) {
207 ChildErrorInfo info = {errCode, errnoValue};
208 // Write the error information over the pipe to our parent process.
209 // We can't really do anything else if this write call fails.
210 writeNoInt(errFd, &info, sizeof(info));
217 void Subprocess::closeAll() {
218 for (auto& p : pipes_) {
219 closeChecked(p.parentFd);
224 void Subprocess::setAllNonBlocking() {
225 for (auto& p : pipes_) {
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 of the pipes_
248 auto pipesGuard = makeGuard([&] {
249 for (auto& p : this->pipes_) {
250 CHECK_ERR(::close(p.parentFd));
254 // Create a pipe to use to receive error information from the child,
255 // in case it fails before calling exec()
257 int r = ::pipe(errFds);
258 checkUnixError(r, "pipe");
260 CHECK_ERR(::close(errFds[0]));
261 if (errFds[1] >= 0) {
262 CHECK_ERR(::close(errFds[1]));
265 // Ask the child to close the read end of the error pipe.
266 options.fdActions_[errFds[0]] = CLOSE;
267 // Set the close-on-exec flag on the write side of the pipe.
268 // This way the pipe will be closed automatically in the child if execve()
269 // succeeds. If the exec fails the child can write error information to the
271 r = fcntl(errFds[1], F_SETFD, FD_CLOEXEC);
272 checkUnixError(r, "set FD_CLOEXEC");
274 // Perform the actual work of setting up pipes then forking and
275 // executing the child.
276 spawnInternal(std::move(argv), executable, options, env, errFds[1]);
278 // After spawnInternal() returns the child is alive. We have to be very
279 // careful about throwing after this point. We are inside the constructor,
280 // so if we throw the Subprocess object will have never existed, and the
281 // destructor will never be called.
283 // We should only throw if we got an error via the errFd, and we know the
284 // child has exited and can be immediately waited for. In all other cases,
285 // we have no way of cleaning up the child.
287 // Close writable side of the errFd pipe in the parent process
288 CHECK_ERR(::close(errFds[1]));
291 // Read from the errFd pipe, to tell if the child ran into any errors before
293 readChildErrorPipe(errFds[0], executable);
295 // We have fully succeeded now, so release the guard on pipes_
296 pipesGuard.dismiss();
299 void Subprocess::spawnInternal(
300 std::unique_ptr<const char*[]> argv,
301 const char* executable,
303 const std::vector<std::string>* env,
305 // Parent work, pre-fork: create pipes
306 std::vector<int> childFds;
307 // Close all of the childFds as we leave this scope
309 // These are only pipes, closing them shouldn't fail
310 for (int cfd : childFds) {
311 CHECK_ERR(::close(cfd));
316 for (auto& p : options.fdActions_) {
317 if (p.second == PIPE_IN || p.second == PIPE_OUT) {
320 checkUnixError(r, "pipe");
322 pinfo.direction = p.second;
324 if (p.second == PIPE_IN) {
325 // Child gets reading end
326 pinfo.parentFd = fds[1];
329 pinfo.parentFd = fds[0];
332 p.second = cfd; // ensure it gets dup2()ed
333 pinfo.childFd = p.first;
334 childFds.push_back(cfd);
335 pipes_.push_back(pinfo);
339 // This should already be sorted, as options.fdActions_ is
340 DCHECK(std::is_sorted(pipes_.begin(), pipes_.end()));
342 // Note that the const casts below are legit, per
343 // http://pubs.opengroup.org/onlinepubs/009695399/functions/exec.html
345 char** argVec = const_cast<char**>(argv.get());
347 // Set up environment
348 std::unique_ptr<const char*[]> envHolder;
351 envHolder = cloneStrings(*env);
352 envVec = const_cast<char**>(envHolder.get());
357 // Block all signals around vfork; see http://ewontfix.com/7/.
359 // As the child may run in the same address space as the parent until
360 // the actual execve() system call, any (custom) signal handlers that
361 // the parent has might alter parent's memory if invoked in the child,
362 // with undefined results. So we block all signals in the parent before
363 // vfork(), which will cause them to be blocked in the child as well (we
364 // rely on the fact that Linux, just like all sane implementations, only
365 // clones the calling thread). Then, in the child, we reset all signals
366 // to their default dispositions (while still blocked), and unblock them
367 // (so the exec()ed process inherits the parent's signal mask)
369 // The parent also unblocks all signals as soon as vfork() returns.
371 r = sigfillset(&allBlocked);
372 checkUnixError(r, "sigfillset");
375 r = pthread_sigmask(SIG_SETMASK, &allBlocked, &oldSignals);
376 checkPosixError(r, "pthread_sigmask");
378 // Restore signal mask
379 r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
380 CHECK_EQ(r, 0) << "pthread_sigmask: " << errnoStr(r); // shouldn't fail
385 int errnoValue = prepareChild(options, &oldSignals);
386 if (errnoValue != 0) {
387 childError(errFd, kChildFailure, errnoValue);
390 errnoValue = runChild(executable, argVec, envVec, options);
391 // If we get here, exec() failed.
392 childError(errFd, kExecFailure, errnoValue);
394 // In parent. Make sure vfork() succeeded.
395 checkUnixError(pid, errno, "vfork");
397 // Child is alive. We have to be very careful about throwing after this
398 // point. We are inside the constructor, so if we throw the Subprocess
399 // object will have never existed, and the destructor will never be called.
401 // We should only throw if we got an error via the errFd, and we know the
402 // child has exited and can be immediately waited for. In all other cases,
403 // we have no way of cleaning up the child.
405 returnCode_ = ProcessReturnCode(RV_RUNNING);
408 int Subprocess::prepareChild(const Options& options,
409 const sigset_t* sigmask) const {
410 // While all signals are blocked, we must reset their
411 // dispositions to default.
412 for (int sig = 1; sig < NSIG; ++sig) {
413 ::signal(sig, SIG_DFL);
415 // Unblock signals; restore signal mask.
416 int r = pthread_sigmask(SIG_SETMASK, sigmask, nullptr);
418 return r; // pthread_sigmask() returns an errno value
421 // Change the working directory, if one is given
422 if (!options.childDir_.empty()) {
423 r = ::chdir(options.childDir_.c_str());
429 // Close parent's ends of all pipes
430 for (auto& p : pipes_) {
431 r = ::close(p.parentFd);
437 // Close all fds that we're supposed to close.
438 // Note that we're ignoring errors here, in case some of these
439 // fds were set to close on exec.
440 for (auto& p : options.fdActions_) {
441 if (p.second == CLOSE) {
444 r = ::dup2(p.second, p.first);
451 // If requested, close all other file descriptors. Don't close
452 // any fds in options.fdActions_, and don't touch stdin, stdout, stderr.
454 if (options.closeOtherFds_) {
455 for (int fd = getdtablesize() - 1; fd >= 3; --fd) {
456 if (options.fdActions_.count(fd) == 0) {
463 // Opt to receive signal on parent death, if requested
464 if (options.parentDeathSignal_ != 0) {
465 r = prctl(PR_SET_PDEATHSIG, options.parentDeathSignal_, 0, 0, 0);
475 int Subprocess::runChild(const char* executable,
476 char** argv, char** env,
477 const Options& options) const {
478 // Now, finally, exec.
480 if (options.usePath_) {
481 ::execvp(executable, argv);
483 ::execve(executable, argv, env);
488 void Subprocess::readChildErrorPipe(int pfd, const char* executable) {
490 auto rc = readNoInt(pfd, &info, sizeof(info));
492 // No data means the child executed successfully, and the pipe
493 // was closed due to the close-on-exec flag being set.
495 } else if (rc != sizeof(ChildErrorInfo)) {
496 // An error occurred trying to read from the pipe, or we got a partial read.
497 // Neither of these cases should really occur in practice.
499 // We can't get any error data from the child in this case, and we don't
500 // know if it is successfully running or not. All we can do is to return
501 // normally, as if the child executed successfully. If something bad
502 // happened the caller should at least get a non-normal exit status from
504 LOG(ERROR) << "unexpected error trying to read from child error pipe " <<
505 "rc=" << rc << ", errno=" << errno;
509 // We got error data from the child. The child should exit immediately in
510 // this case, so wait on it to clean up.
513 // Throw to signal the error
514 throw SubprocessSpawnError(executable, info.errCode, info.errnoValue);
517 ProcessReturnCode Subprocess::poll() {
518 returnCode_.enforce(ProcessReturnCode::RUNNING);
521 pid_t found = ::waitpid(pid_, &status, WNOHANG);
522 checkUnixError(found, "waitpid");
524 returnCode_ = ProcessReturnCode(status);
530 bool Subprocess::pollChecked() {
531 if (poll().state() == ProcessReturnCode::RUNNING) {
534 checkStatus(returnCode_);
538 ProcessReturnCode Subprocess::wait() {
539 returnCode_.enforce(ProcessReturnCode::RUNNING);
544 found = ::waitpid(pid_, &status, 0);
545 } while (found == -1 && errno == EINTR);
546 checkUnixError(found, "waitpid");
547 DCHECK_EQ(found, pid_);
548 returnCode_ = ProcessReturnCode(status);
553 void Subprocess::waitChecked() {
555 checkStatus(returnCode_);
558 void Subprocess::sendSignal(int signal) {
559 returnCode_.enforce(ProcessReturnCode::RUNNING);
560 int r = ::kill(pid_, signal);
561 checkUnixError(r, "kill");
564 pid_t Subprocess::pid() const {
570 std::pair<const uint8_t*, size_t> queueFront(const IOBufQueue& queue) {
571 auto* p = queue.front();
572 if (!p) return std::make_pair(nullptr, 0);
573 return io::Cursor(p).peek();
577 bool handleWrite(int fd, IOBufQueue& queue) {
579 auto p = queueFront(queue);
584 ssize_t n = writeNoInt(fd, p.first, p.second);
585 if (n == -1 && errno == EAGAIN) {
588 checkUnixError(n, "write");
594 bool handleRead(int fd, IOBufQueue& queue) {
596 auto p = queue.preallocate(100, 65000);
597 ssize_t n = readNoInt(fd, p.first, p.second);
598 if (n == -1 && errno == EAGAIN) {
601 checkUnixError(n, "read");
605 queue.postallocate(n);
609 bool discardRead(int fd) {
610 static const size_t bufSize = 65000;
611 // Thread unsafe, but it doesn't matter.
612 static std::unique_ptr<char[]> buf(new char[bufSize]);
615 ssize_t n = readNoInt(fd, buf.get(), bufSize);
616 if (n == -1 && errno == EAGAIN) {
619 checkUnixError(n, "read");
628 std::pair<std::string, std::string> Subprocess::communicate(
630 IOBufQueue inputQueue;
631 inputQueue.wrapBuffer(input.data(), input.size());
633 auto outQueues = communicateIOBuf(std::move(inputQueue));
634 auto outBufs = std::make_pair(outQueues.first.move(),
635 outQueues.second.move());
636 std::pair<std::string, std::string> out;
638 outBufs.first->coalesce();
639 out.first.assign(reinterpret_cast<const char*>(outBufs.first->data()),
640 outBufs.first->length());
642 if (outBufs.second) {
643 outBufs.second->coalesce();
644 out.second.assign(reinterpret_cast<const char*>(outBufs.second->data()),
645 outBufs.second->length());
650 std::pair<IOBufQueue, IOBufQueue> Subprocess::communicateIOBuf(
652 // If the user supplied a non-empty input buffer, make sure
653 // that stdin is a pipe so we can write the data.
654 if (!input.empty()) {
655 // findByChildFd() will throw std::invalid_argument if no pipe for
656 // STDIN_FILENO exists
657 findByChildFd(STDIN_FILENO);
660 std::pair<IOBufQueue, IOBufQueue> out;
662 auto readCallback = [&] (int pfd, int cfd) -> bool {
663 if (cfd == STDOUT_FILENO) {
664 return handleRead(pfd, out.first);
665 } else if (cfd == STDERR_FILENO) {
666 return handleRead(pfd, out.second);
668 // Don't close the file descriptor, the child might not like SIGPIPE,
669 // just read and throw the data away.
670 return discardRead(pfd);
674 auto writeCallback = [&] (int pfd, int cfd) -> bool {
675 if (cfd == STDIN_FILENO) {
676 return handleWrite(pfd, input);
678 // If we don't want to write to this fd, just close it.
683 communicate(std::move(readCallback), std::move(writeCallback));
688 void Subprocess::communicate(FdCallback readCallback,
689 FdCallback writeCallback) {
690 returnCode_.enforce(ProcessReturnCode::RUNNING);
693 std::vector<pollfd> fds;
694 fds.reserve(pipes_.size());
695 std::vector<int> toClose;
696 toClose.reserve(pipes_.size());
698 while (!pipes_.empty()) {
702 for (auto& p : pipes_) {
705 // Yes, backwards, PIPE_IN / PIPE_OUT are defined from the
706 // child's point of view.
708 // Still keeping fd in watched set so we get notified of POLLHUP /
711 } else if (p.direction == PIPE_IN) {
712 pfd.events = POLLOUT;
721 r = ::poll(fds.data(), fds.size(), -1);
722 } while (r == -1 && errno == EINTR);
723 checkUnixError(r, "poll");
725 for (int i = 0; i < pipes_.size(); ++i) {
727 DCHECK_EQ(fds[i].fd, p.parentFd);
728 short events = fds[i].revents;
731 if (events & POLLOUT) {
732 DCHECK(!(events & POLLIN));
733 if (writeCallback(p.parentFd, p.childFd)) {
734 toClose.push_back(i);
739 // Call read callback on POLLHUP, to give it a chance to read (and act
741 if (events & (POLLIN | POLLHUP)) {
742 DCHECK(!(events & POLLOUT));
743 if (readCallback(p.parentFd, p.childFd)) {
744 toClose.push_back(i);
749 if ((events & (POLLHUP | POLLERR)) && !closed) {
750 toClose.push_back(i);
755 // Close the fds in reverse order so the indexes hold after erase()
756 for (int idx : boost::adaptors::reverse(toClose)) {
757 auto pos = pipes_.begin() + idx;
758 closeChecked(pos->parentFd);
764 void Subprocess::enableNotifications(int childFd, bool enabled) {
765 pipes_[findByChildFd(childFd)].enabled = enabled;
768 bool Subprocess::notificationsEnabled(int childFd) const {
769 return pipes_[findByChildFd(childFd)].enabled;
772 int Subprocess::findByChildFd(int childFd) const {
773 auto pos = std::lower_bound(
774 pipes_.begin(), pipes_.end(), childFd,
775 [] (const PipeInfo& info, int fd) { return info.childFd < fd; });
776 if (pos == pipes_.end() || pos->childFd != childFd) {
777 throw std::invalid_argument(folly::to<std::string>(
778 "child fd not found ", childFd));
780 return pos - pipes_.begin();
783 void Subprocess::closeParentFd(int childFd) {
784 int idx = findByChildFd(childFd);
785 closeChecked(pipes_[idx].parentFd);
786 pipes_.erase(pipes_.begin() + idx);
794 // We like EPIPE, thanks.
795 ::signal(SIGPIPE, SIG_IGN);
799 Initializer initializer;