2 * Copyright 2012 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"
26 #include <system_error>
28 #include <boost/container/flat_set.hpp>
29 #include <boost/range/adaptors.hpp>
31 #include <glog/logging.h>
33 #include "folly/Conv.h"
34 #include "folly/ScopeGuard.h"
35 #include "folly/String.h"
36 #include "folly/experimental/io/Cursor.h"
38 extern char** environ;
42 ProcessReturnCode::State ProcessReturnCode::state() const {
43 if (rawStatus_ == RV_NOT_STARTED) return NOT_STARTED;
44 if (rawStatus_ == RV_RUNNING) return RUNNING;
45 if (WIFEXITED(rawStatus_)) return EXITED;
46 if (WIFSIGNALED(rawStatus_)) return KILLED;
47 throw std::runtime_error(to<std::string>(
48 "Invalid ProcessReturnCode: ", rawStatus_));
51 void ProcessReturnCode::enforce(State s) const {
53 throw std::logic_error(to<std::string>("Invalid state ", s));
57 int ProcessReturnCode::exitStatus() const {
59 return WEXITSTATUS(rawStatus_);
62 int ProcessReturnCode::killSignal() const {
64 return WTERMSIG(rawStatus_);
67 bool ProcessReturnCode::coreDumped() const {
69 return WCOREDUMP(rawStatus_);
72 std::string ProcessReturnCode::str() const {
79 return to<std::string>("exited with status ", exitStatus());
81 return to<std::string>("killed by signal ", killSignal(),
82 (coreDumped() ? " (core dumped)" : ""));
84 CHECK(false); // unreached
87 CalledProcessError::CalledProcessError(ProcessReturnCode rc)
89 what_(returnCode_.str()) {
94 // Copy pointers to the given strings in a format suitable for posix_spawn
95 std::unique_ptr<const char*[]> cloneStrings(const std::vector<std::string>& s) {
96 std::unique_ptr<const char*[]> d(new const char*[s.size() + 1]);
97 for (int i = 0; i < s.size(); i++) {
100 d[s.size()] = nullptr;
104 // Helper to throw std::system_error
105 void throwSystemError(int err, const char* msg) __attribute__((noreturn));
106 void throwSystemError(int err, const char* msg) {
107 throw std::system_error(err, std::system_category(), msg);
110 // Helper to throw std::system_error from errno
111 void throwSystemError(const char* msg) __attribute__((noreturn));
112 void throwSystemError(const char* msg) {
113 throwSystemError(errno, msg);
116 // Check a Posix return code (0 on success, error number on error), throw
118 void checkPosixError(int err, const char* msg) {
120 throwSystemError(err, msg);
124 // Check a traditional Uinx return code (-1 and sets errno on error), throw
126 void checkUnixError(ssize_t ret, const char* msg) {
128 throwSystemError(msg);
131 void checkUnixError(ssize_t ret, int savedErrno, const char* msg) {
133 throwSystemError(savedErrno, msg);
137 // Check a wait() status, throw on non-successful
138 void checkStatus(ProcessReturnCode returnCode) {
139 if (returnCode.state() != ProcessReturnCode::EXITED ||
140 returnCode.exitStatus() != 0) {
141 throw CalledProcessError(returnCode);
147 Subprocess::Options& Subprocess::Options::fd(int fd, int action) {
148 if (action == Subprocess::PIPE) {
150 action = Subprocess::PIPE_IN;
151 } else if (fd == 1 || fd == 2) {
152 action = Subprocess::PIPE_OUT;
154 throw std::invalid_argument(
155 to<std::string>("Only fds 0, 1, 2 are valid for action=PIPE: ", fd));
158 fdActions_[fd] = action;
162 Subprocess::Subprocess(
163 const std::vector<std::string>& argv,
164 const Options& options,
165 const char* executable,
166 const std::vector<std::string>* env)
168 returnCode_(RV_NOT_STARTED) {
170 throw std::invalid_argument("argv must not be empty");
172 if (!executable) executable = argv[0].c_str();
173 spawn(cloneStrings(argv), executable, options, env);
176 Subprocess::Subprocess(
177 const std::string& cmd,
178 const Options& options,
179 const std::vector<std::string>* env)
181 returnCode_(RV_NOT_STARTED) {
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 if (returnCode_.state() == ProcessReturnCode::RUNNING) {
200 LOG(ERROR) << "Subprocess destroyed without reaping; killing child.";
205 LOG(FATAL) << "Killing child failed, terminating: "
206 << exceptionStr(std::current_exception());
212 LOG(FATAL) << "close failed, terminating: "
213 << exceptionStr(std::current_exception());
218 void closeChecked(int fd) {
219 checkUnixError(::close(fd), "close");
223 void Subprocess::closeAll() {
224 for (auto& p : pipes_) {
225 closeChecked(p.parentFd);
230 void Subprocess::setAllNonBlocking() {
231 for (auto& p : pipes_) {
233 int flags = ::fcntl(fd, F_GETFL);
234 checkUnixError(flags, "fcntl");
235 int r = ::fcntl(fd, F_SETFL, flags | O_NONBLOCK);
236 checkUnixError(r, "fcntl");
240 void Subprocess::spawn(
241 std::unique_ptr<const char*[]> argv,
242 const char* executable,
243 const Options& optionsIn,
244 const std::vector<std::string>* env) {
245 if (optionsIn.usePath_ && env) {
246 throw std::invalid_argument(
247 "usePath() not allowed when overriding environment");
250 // Make a copy, we'll mutate options
251 Options options(optionsIn);
253 // Parent work, pre-fork: create pipes
254 std::vector<int> childFds;
255 for (auto& p : options.fdActions_) {
256 if (p.second == PIPE_IN || p.second == PIPE_OUT) {
259 checkUnixError(r, "pipe");
261 pinfo.direction = p.second;
263 if (p.second == PIPE_IN) {
264 // Child gets reading end
265 pinfo.parentFd = fds[1];
268 pinfo.parentFd = fds[0];
271 p.second = cfd; // ensure it gets dup2()ed
272 pinfo.childFd = p.first;
273 childFds.push_back(cfd);
274 pipes_.push_back(pinfo);
278 // This should already be sorted, as options.fdActions_ is
279 DCHECK(std::is_sorted(pipes_.begin(), pipes_.end()));
281 // Note that the const casts below are legit, per
282 // http://pubs.opengroup.org/onlinepubs/009695399/functions/exec.html
284 char** argVec = const_cast<char**>(argv.get());
286 // Set up environment
287 std::unique_ptr<const char*[]> envHolder;
290 envHolder = cloneStrings(*env);
291 envVec = const_cast<char**>(envHolder.get());
296 // Block all signals around vfork; see http://ewontfix.com/7/.
298 // As the child may run in the same address space as the parent until
299 // the actual execve() system call, any (custom) signal handlers that
300 // the parent has might alter parent's memory if invoked in the child,
301 // with undefined results. So we block all signals in the parent before
302 // vfork(), which will cause them to be blocked in the child as well (we
303 // rely on the fact that Linux, just like all sane implementations, only
304 // clones the calling thread). Then, in the child, we reset all signals
305 // to their default dispositions (while still blocked), and unblock them
306 // (so the exec()ed process inherits the parent's signal mask)
308 // The parent also unblocks all signals as soon as vfork() returns.
310 int r = ::sigfillset(&allBlocked);
311 checkUnixError(r, "sigfillset");
313 r = pthread_sigmask(SIG_SETMASK, &allBlocked, &oldSignals);
314 checkPosixError(r, "pthread_sigmask");
318 // While all signals are blocked, we must reset their
319 // dispositions to default.
320 for (int sig = 1; sig < NSIG; ++sig) {
321 ::signal(sig, SIG_DFL);
323 // Unblock signals; restore signal mask.
324 int r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
327 runChild(executable, argVec, envVec, options);
328 // This should never return, but there's nothing else we can do here.
331 // In parent. We want to restore the signal mask even if vfork fails,
332 // so we'll save errno here, restore the signal mask, and only then
334 int savedErrno = errno;
336 // Restore signal mask; do this even if vfork fails!
337 // We only check for errors from pthread_sigmask after we recorded state
338 // that the child is alive, so we know to reap it.
339 r = pthread_sigmask(SIG_SETMASK, &oldSignals, nullptr);
340 checkUnixError(pid, savedErrno, "vfork");
344 returnCode_ = ProcessReturnCode(RV_RUNNING);
346 // Parent work, post-fork: close child's ends of pipes
347 for (int f : childFds) {
351 checkPosixError(r, "pthread_sigmask");
356 // Checked version of close() to use in the child: abort() on error
357 void childClose(int fd) {
359 if (r == -1) abort();
362 // Checked version of dup2() to use in the child: abort() on error
363 void childDup2(int oldfd, int newfd) {
364 int r = ::dup2(oldfd, newfd);
365 if (r == -1) abort();
370 void Subprocess::runChild(const char* executable,
371 char** argv, char** env,
372 const Options& options) const {
373 // Close parent's ends of all pipes
374 for (auto& p : pipes_) {
375 childClose(p.parentFd);
378 // Close all fds that we're supposed to close.
379 // Note that we're ignoring errors here, in case some of these
380 // fds were set to close on exec.
381 for (auto& p : options.fdActions_) {
382 if (p.second == CLOSE) {
385 childDup2(p.second, p.first);
389 // If requested, close all other file descriptors. Don't close
390 // any fds in options.fdActions_, and don't touch stdin, stdout, stderr.
392 if (options.closeOtherFds_) {
393 for (int fd = getdtablesize() - 1; fd >= 3; --fd) {
394 if (options.fdActions_.count(fd) == 0) {
400 // Now, finally, exec.
402 if (options.usePath_) {
403 ::execvp(executable, argv);
405 ::execve(executable, argv, env);
408 // If we're here, something's wrong.
412 ProcessReturnCode Subprocess::poll() {
413 returnCode_.enforce(ProcessReturnCode::RUNNING);
416 pid_t found = ::waitpid(pid_, &status, WNOHANG);
417 checkUnixError(found, "waitpid");
419 returnCode_ = ProcessReturnCode(status);
425 bool Subprocess::pollChecked() {
426 if (poll().state() == ProcessReturnCode::RUNNING) {
429 checkStatus(returnCode_);
433 ProcessReturnCode Subprocess::wait() {
434 returnCode_.enforce(ProcessReturnCode::RUNNING);
439 found = ::waitpid(pid_, &status, 0);
440 } while (found == -1 && errno == EINTR);
441 checkUnixError(found, "waitpid");
442 DCHECK_EQ(found, pid_);
443 returnCode_ = ProcessReturnCode(status);
447 void Subprocess::waitChecked() {
449 checkStatus(returnCode_);
452 void Subprocess::sendSignal(int signal) {
453 returnCode_.enforce(ProcessReturnCode::RUNNING);
454 int r = ::kill(pid_, signal);
455 checkUnixError(r, "kill");
459 void setNonBlocking(int fd) {
460 int flags = ::fcntl(fd, F_GETFL);
461 checkUnixError(flags, "fcntl");
462 int r = ::fcntl(fd, F_SETFL, flags | O_NONBLOCK);
463 checkUnixError(r, "fcntl");
466 std::pair<const uint8_t*, size_t> queueFront(const IOBufQueue& queue) {
467 auto* p = queue.front();
468 if (!p) return std::make_pair(nullptr, 0);
469 return io::Cursor(p).peek();
473 bool handleWrite(int fd, IOBufQueue& queue) {
475 auto p = queueFront(queue);
482 n = ::write(fd, p.first, p.second);
483 } while (n == -1 && errno == EINTR);
484 if (n == -1 && errno == EAGAIN) {
487 checkUnixError(n, "write");
493 bool handleRead(int fd, IOBufQueue& queue) {
495 auto p = queue.preallocate(100, 65000);
498 n = ::read(fd, p.first, p.second);
499 } while (n == -1 && errno == EINTR);
500 if (n == -1 && errno == EAGAIN) {
503 checkUnixError(n, "read");
507 queue.postallocate(n);
511 bool discardRead(int fd) {
512 static const size_t bufSize = 65000;
513 // Thread unsafe, but it doesn't matter.
514 static std::unique_ptr<char[]> buf(new char[bufSize]);
519 n = ::read(fd, buf.get(), bufSize);
520 } while (n == -1 && errno == EINTR);
521 if (n == -1 && errno == EAGAIN) {
524 checkUnixError(n, "read");
533 std::pair<std::string, std::string> Subprocess::communicate(
536 IOBufQueue dataQueue;
537 dataQueue.wrapBuffer(data.data(), data.size());
539 auto outQueues = communicateIOBuf(flags, std::move(dataQueue));
540 auto outBufs = std::make_pair(outQueues.first.move(),
541 outQueues.second.move());
542 std::pair<std::string, std::string> out;
544 outBufs.first->coalesce();
545 out.first.assign(reinterpret_cast<const char*>(outBufs.first->data()),
546 outBufs.first->length());
548 if (outBufs.second) {
549 outBufs.second->coalesce();
550 out.second.assign(reinterpret_cast<const char*>(outBufs.second->data()),
551 outBufs.second->length());
556 std::pair<IOBufQueue, IOBufQueue> Subprocess::communicateIOBuf(
559 std::pair<IOBufQueue, IOBufQueue> out;
561 auto readCallback = [&, flags] (int pfd, int cfd) {
562 if (cfd == 1 && (flags & READ_STDOUT)) {
563 return handleRead(pfd, out.first);
564 } else if (cfd == 2 && (flags & READ_STDERR)) {
565 return handleRead(pfd, out.second);
567 // Don't close the file descriptor, the child might not like SIGPIPE,
568 // just read and throw the data away.
569 return discardRead(pfd);
573 auto writeCallback = [&, flags] (int pfd, int cfd) {
574 if (cfd == 0 && (flags & WRITE_STDIN)) {
575 return handleWrite(pfd, data);
577 // If we don't want to write to this fd, just close it.
582 communicate(std::move(readCallback), std::move(writeCallback));
587 void Subprocess::communicate(FdCallback readCallback,
588 FdCallback writeCallback) {
589 returnCode_.enforce(ProcessReturnCode::RUNNING);
592 std::vector<pollfd> fds;
593 fds.reserve(pipes_.size());
594 std::vector<int> toClose;
595 toClose.reserve(pipes_.size());
597 while (!pipes_.empty()) {
601 for (auto& p : pipes_) {
604 // Yes, backwards, PIPE_IN / PIPE_OUT are defined from the
605 // child's point of view.
606 pfd.events = (p.direction == PIPE_IN ? POLLOUT : POLLIN);
612 r = ::poll(fds.data(), fds.size(), -1);
613 } while (r == -1 && errno == EINTR);
614 checkUnixError(r, "poll");
616 for (int i = 0; i < pipes_.size(); ++i) {
618 DCHECK_EQ(fds[i].fd, p.parentFd);
619 short events = fds[i].revents;
622 if (events & POLLOUT) {
623 DCHECK(!(events & POLLIN));
624 if (writeCallback(p.parentFd, p.childFd)) {
625 toClose.push_back(i);
630 if (events & POLLIN) {
631 DCHECK(!(events & POLLOUT));
632 if (readCallback(p.parentFd, p.childFd)) {
633 toClose.push_back(i);
638 if ((events & (POLLHUP | POLLERR)) && !closed) {
639 toClose.push_back(i);
644 // Close the fds in reverse order so the indexes hold after erase()
645 for (int idx : boost::adaptors::reverse(toClose)) {
646 auto pos = pipes_.begin() + idx;
647 closeChecked(pos->parentFd);
653 int Subprocess::findByChildFd(int childFd) const {
654 auto pos = std::lower_bound(
655 pipes_.begin(), pipes_.end(), childFd,
656 [] (const PipeInfo& info, int fd) { return info.childFd < fd; });
657 if (pos == pipes_.end() || pos->childFd != childFd) {
658 throw std::invalid_argument(folly::to<std::string>(
659 "child fd not found ", childFd));
661 return pos - pipes_.begin();
664 void Subprocess::closeParentFd(int childFd) {
665 int idx = findByChildFd(childFd);
666 closeChecked(pipes_[idx].parentFd);
667 pipes_.erase(pipes_.begin() + idx);
675 // We like EPIPE, thanks.
676 ::signal(SIGPIPE, SIG_IGN);
680 Initializer initializer;