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(), setting the close-on-exec flag
256 // on both sides of the pipe.
257 // This way the pipe will be closed automatically in the child if execve()
258 // succeeds. If the exec fails the child can write error information to the
260 // Note that O_CLOEXEC must be set in a single call while we are creating
261 // the pipe instead of doing pipe()/fcntl separately, which might race if a
262 // another thread calls fork()/exec() concurrently and both sides of the pipe
263 // may be inherited by the corresponding child process without being closed.
265 int r = ::pipe2(errFds, O_CLOEXEC);
266 checkUnixError(r, "pipe2");
268 CHECK_ERR(::close(errFds[0]));
269 if (errFds[1] >= 0) {
270 CHECK_ERR(::close(errFds[1]));
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) {
319 // Set O_CLOEXEC on both ends of the pipe atomically while creating
320 // the pipe. The child will clear O_CLOEXEC on its side of the pipe
321 // before calling exec() so that stays open afterwards.
322 // This way even if a concurrently constructed Subprocess inherits
323 // both ends of this pipe, they will be automatically closed
324 // after the corresponding exec().
325 r = ::pipe2(fds, O_CLOEXEC);
326 checkUnixError(r, "pipe2");
328 pinfo.direction = p.second;
330 if (p.second == PIPE_IN) {
331 // Child gets reading end
332 pinfo.parentFd = fds[1];
335 pinfo.parentFd = fds[0];
338 p.second = cfd; // ensure it gets dup2()ed
339 pinfo.childFd = p.first;
340 childFds.push_back(cfd);
341 pipes_.push_back(pinfo);
345 // This should already be sorted, as options.fdActions_ is
346 DCHECK(std::is_sorted(pipes_.begin(), pipes_.end()));
348 // Note that the const casts below are legit, per
349 // http://pubs.opengroup.org/onlinepubs/009695399/functions/exec.html
351 char** argVec = const_cast<char**>(argv.get());
353 // Set up environment
354 std::unique_ptr<const char*[]> envHolder;
357 envHolder = cloneStrings(*env);
358 envVec = const_cast<char**>(envHolder.get());
363 // Block all signals around vfork; see http://ewontfix.com/7/.
365 // As the child may run in the same address space as the parent until
366 // the actual execve() system call, any (custom) signal handlers that
367 // the parent has might alter parent's memory if invoked in the child,
368 // with undefined results. So we block all signals in the parent before
369 // vfork(), which will cause them to be blocked in the child as well (we
370 // rely on the fact that Linux, just like all sane implementations, only
371 // clones the calling thread). Then, in the child, we reset all signals
372 // to their default dispositions (while still blocked), and unblock them
373 // (so the exec()ed process inherits the parent's signal mask)
375 // The parent also unblocks all signals as soon as vfork() returns.
377 r = sigfillset(&allBlocked);
378 checkUnixError(r, "sigfillset");
381 r = pthread_sigmask(SIG_SETMASK, &allBlocked, &oldSignals);
382 checkPosixError(r, "pthread_sigmask");
384 // Restore signal mask
385 r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
386 CHECK_EQ(r, 0) << "pthread_sigmask: " << errnoStr(r); // shouldn't fail
391 int errnoValue = prepareChild(options, &oldSignals);
392 if (errnoValue != 0) {
393 childError(errFd, kChildFailure, errnoValue);
396 errnoValue = runChild(executable, argVec, envVec, options);
397 // If we get here, exec() failed.
398 childError(errFd, kExecFailure, errnoValue);
400 // In parent. Make sure vfork() succeeded.
401 checkUnixError(pid, errno, "vfork");
403 // Child is alive. We have to be very careful about throwing after this
404 // point. We are inside the constructor, so if we throw the Subprocess
405 // object will have never existed, and the destructor will never be called.
407 // We should only throw if we got an error via the errFd, and we know the
408 // child has exited and can be immediately waited for. In all other cases,
409 // we have no way of cleaning up the child.
411 returnCode_ = ProcessReturnCode(RV_RUNNING);
414 int Subprocess::prepareChild(const Options& options,
415 const sigset_t* sigmask) const {
416 // While all signals are blocked, we must reset their
417 // dispositions to default.
418 for (int sig = 1; sig < NSIG; ++sig) {
419 ::signal(sig, SIG_DFL);
421 // Unblock signals; restore signal mask.
422 int r = pthread_sigmask(SIG_SETMASK, sigmask, nullptr);
424 return r; // pthread_sigmask() returns an errno value
427 // Change the working directory, if one is given
428 if (!options.childDir_.empty()) {
429 r = ::chdir(options.childDir_.c_str());
435 for (auto& p : pipes_) {
436 // Clear FD_CLOEXEC on the child side of the pipe so
437 // it stays open after exec() (so that the child could
439 // See spawnInternal() for why FD_CLOEXEC must be set
440 // by default on pipes.
441 r = fcntl(p.childFd, F_SETFD, 0);
447 // Close all fds that we're supposed to close.
448 // Note that we're ignoring errors here, in case some of these
449 // fds were set to close on exec.
450 for (auto& p : options.fdActions_) {
451 if (p.second == CLOSE) {
454 r = ::dup2(p.second, p.first);
461 // If requested, close all other file descriptors. Don't close
462 // any fds in options.fdActions_, and don't touch stdin, stdout, stderr.
464 if (options.closeOtherFds_) {
465 for (int fd = getdtablesize() - 1; fd >= 3; --fd) {
466 if (options.fdActions_.count(fd) == 0) {
473 // Opt to receive signal on parent death, if requested
474 if (options.parentDeathSignal_ != 0) {
475 r = prctl(PR_SET_PDEATHSIG, options.parentDeathSignal_, 0, 0, 0);
485 int Subprocess::runChild(const char* executable,
486 char** argv, char** env,
487 const Options& options) const {
488 // Now, finally, exec.
490 if (options.usePath_) {
491 ::execvp(executable, argv);
493 ::execve(executable, argv, env);
498 void Subprocess::readChildErrorPipe(int pfd, const char* executable) {
500 auto rc = readNoInt(pfd, &info, sizeof(info));
502 // No data means the child executed successfully, and the pipe
503 // was closed due to the close-on-exec flag being set.
505 } else if (rc != sizeof(ChildErrorInfo)) {
506 // An error occurred trying to read from the pipe, or we got a partial read.
507 // Neither of these cases should really occur in practice.
509 // We can't get any error data from the child in this case, and we don't
510 // know if it is successfully running or not. All we can do is to return
511 // normally, as if the child executed successfully. If something bad
512 // happened the caller should at least get a non-normal exit status from
514 LOG(ERROR) << "unexpected error trying to read from child error pipe " <<
515 "rc=" << rc << ", errno=" << errno;
519 // We got error data from the child. The child should exit immediately in
520 // this case, so wait on it to clean up.
523 // Throw to signal the error
524 throw SubprocessSpawnError(executable, info.errCode, info.errnoValue);
527 ProcessReturnCode Subprocess::poll() {
528 returnCode_.enforce(ProcessReturnCode::RUNNING);
531 pid_t found = ::waitpid(pid_, &status, WNOHANG);
532 checkUnixError(found, "waitpid");
534 returnCode_ = ProcessReturnCode(status);
540 bool Subprocess::pollChecked() {
541 if (poll().state() == ProcessReturnCode::RUNNING) {
544 checkStatus(returnCode_);
548 ProcessReturnCode Subprocess::wait() {
549 returnCode_.enforce(ProcessReturnCode::RUNNING);
554 found = ::waitpid(pid_, &status, 0);
555 } while (found == -1 && errno == EINTR);
556 checkUnixError(found, "waitpid");
557 DCHECK_EQ(found, pid_);
558 returnCode_ = ProcessReturnCode(status);
563 void Subprocess::waitChecked() {
565 checkStatus(returnCode_);
568 void Subprocess::sendSignal(int signal) {
569 returnCode_.enforce(ProcessReturnCode::RUNNING);
570 int r = ::kill(pid_, signal);
571 checkUnixError(r, "kill");
574 pid_t Subprocess::pid() const {
580 std::pair<const uint8_t*, size_t> queueFront(const IOBufQueue& queue) {
581 auto* p = queue.front();
582 if (!p) return std::make_pair(nullptr, 0);
583 return io::Cursor(p).peek();
587 bool handleWrite(int fd, IOBufQueue& queue) {
589 auto p = queueFront(queue);
594 ssize_t n = writeNoInt(fd, p.first, p.second);
595 if (n == -1 && errno == EAGAIN) {
598 checkUnixError(n, "write");
604 bool handleRead(int fd, IOBufQueue& queue) {
606 auto p = queue.preallocate(100, 65000);
607 ssize_t n = readNoInt(fd, p.first, p.second);
608 if (n == -1 && errno == EAGAIN) {
611 checkUnixError(n, "read");
615 queue.postallocate(n);
619 bool discardRead(int fd) {
620 static const size_t bufSize = 65000;
621 // Thread unsafe, but it doesn't matter.
622 static std::unique_ptr<char[]> buf(new char[bufSize]);
625 ssize_t n = readNoInt(fd, buf.get(), bufSize);
626 if (n == -1 && errno == EAGAIN) {
629 checkUnixError(n, "read");
638 std::pair<std::string, std::string> Subprocess::communicate(
640 IOBufQueue inputQueue;
641 inputQueue.wrapBuffer(input.data(), input.size());
643 auto outQueues = communicateIOBuf(std::move(inputQueue));
644 auto outBufs = std::make_pair(outQueues.first.move(),
645 outQueues.second.move());
646 std::pair<std::string, std::string> out;
648 outBufs.first->coalesce();
649 out.first.assign(reinterpret_cast<const char*>(outBufs.first->data()),
650 outBufs.first->length());
652 if (outBufs.second) {
653 outBufs.second->coalesce();
654 out.second.assign(reinterpret_cast<const char*>(outBufs.second->data()),
655 outBufs.second->length());
660 std::pair<IOBufQueue, IOBufQueue> Subprocess::communicateIOBuf(
662 // If the user supplied a non-empty input buffer, make sure
663 // that stdin is a pipe so we can write the data.
664 if (!input.empty()) {
665 // findByChildFd() will throw std::invalid_argument if no pipe for
666 // STDIN_FILENO exists
667 findByChildFd(STDIN_FILENO);
670 std::pair<IOBufQueue, IOBufQueue> out;
672 auto readCallback = [&] (int pfd, int cfd) -> bool {
673 if (cfd == STDOUT_FILENO) {
674 return handleRead(pfd, out.first);
675 } else if (cfd == STDERR_FILENO) {
676 return handleRead(pfd, out.second);
678 // Don't close the file descriptor, the child might not like SIGPIPE,
679 // just read and throw the data away.
680 return discardRead(pfd);
684 auto writeCallback = [&] (int pfd, int cfd) -> bool {
685 if (cfd == STDIN_FILENO) {
686 return handleWrite(pfd, input);
688 // If we don't want to write to this fd, just close it.
693 communicate(std::move(readCallback), std::move(writeCallback));
698 void Subprocess::communicate(FdCallback readCallback,
699 FdCallback writeCallback) {
700 returnCode_.enforce(ProcessReturnCode::RUNNING);
703 std::vector<pollfd> fds;
704 fds.reserve(pipes_.size());
705 std::vector<int> toClose;
706 toClose.reserve(pipes_.size());
708 while (!pipes_.empty()) {
712 for (auto& p : pipes_) {
715 // Yes, backwards, PIPE_IN / PIPE_OUT are defined from the
716 // child's point of view.
718 // Still keeping fd in watched set so we get notified of POLLHUP /
721 } else if (p.direction == PIPE_IN) {
722 pfd.events = POLLOUT;
731 r = ::poll(fds.data(), fds.size(), -1);
732 } while (r == -1 && errno == EINTR);
733 checkUnixError(r, "poll");
735 for (int i = 0; i < pipes_.size(); ++i) {
737 DCHECK_EQ(fds[i].fd, p.parentFd);
738 short events = fds[i].revents;
741 if (events & POLLOUT) {
742 DCHECK(!(events & POLLIN));
743 if (writeCallback(p.parentFd, p.childFd)) {
744 toClose.push_back(i);
749 // Call read callback on POLLHUP, to give it a chance to read (and act
751 if (events & (POLLIN | POLLHUP)) {
752 DCHECK(!(events & POLLOUT));
753 if (readCallback(p.parentFd, p.childFd)) {
754 toClose.push_back(i);
759 if ((events & (POLLHUP | POLLERR)) && !closed) {
760 toClose.push_back(i);
765 // Close the fds in reverse order so the indexes hold after erase()
766 for (int idx : boost::adaptors::reverse(toClose)) {
767 auto pos = pipes_.begin() + idx;
768 closeChecked(pos->parentFd);
774 void Subprocess::enableNotifications(int childFd, bool enabled) {
775 pipes_[findByChildFd(childFd)].enabled = enabled;
778 bool Subprocess::notificationsEnabled(int childFd) const {
779 return pipes_[findByChildFd(childFd)].enabled;
782 int Subprocess::findByChildFd(int childFd) const {
783 auto pos = std::lower_bound(
784 pipes_.begin(), pipes_.end(), childFd,
785 [] (const PipeInfo& info, int fd) { return info.childFd < fd; });
786 if (pos == pipes_.end() || pos->childFd != childFd) {
787 throw std::invalid_argument(folly::to<std::string>(
788 "child fd not found ", childFd));
790 return pos - pipes_.begin();
793 void Subprocess::closeParentFd(int childFd) {
794 int idx = findByChildFd(childFd);
795 closeChecked(pipes_[idx].parentFd);
796 pipes_.erase(pipes_.begin() + idx);
804 // We like EPIPE, thanks.
805 ::signal(SIGPIPE, SIG_IGN);
809 Initializer initializer;