2 * Copyright 2017 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/io/async/AsyncSocket.h>
19 #include <folly/ExceptionWrapper.h>
20 #include <folly/Format.h>
21 #include <folly/Portability.h>
22 #include <folly/SocketAddress.h>
23 #include <folly/io/Cursor.h>
24 #include <folly/io/IOBuf.h>
25 #include <folly/io/IOBufQueue.h>
26 #include <folly/portability/Fcntl.h>
27 #include <folly/portability/Sockets.h>
28 #include <folly/portability/SysUio.h>
29 #include <folly/portability/Unistd.h>
31 #include <boost/preprocessor/control/if.hpp>
34 #include <sys/types.h>
38 using std::unique_ptr;
40 namespace fsp = folly::portability::sockets;
44 static constexpr bool msgErrQueueSupported =
45 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
49 #endif // FOLLY_HAVE_MSG_ERRQUEUE
51 // static members initializers
52 const AsyncSocket::OptionMap AsyncSocket::emptyOptionMap;
54 const AsyncSocketException socketClosedLocallyEx(
55 AsyncSocketException::END_OF_FILE, "socket closed locally");
56 const AsyncSocketException socketShutdownForWritesEx(
57 AsyncSocketException::END_OF_FILE, "socket shutdown for writes");
59 // TODO: It might help performance to provide a version of BytesWriteRequest that
60 // users could derive from, so we can avoid the extra allocation for each call
61 // to write()/writev(). We could templatize TFramedAsyncChannel just like the
62 // protocols are currently templatized for transports.
64 // We would need the version for external users where they provide the iovec
65 // storage space, and only our internal version would allocate it at the end of
68 /* The default WriteRequest implementation, used for write(), writev() and
71 * A new BytesWriteRequest operation is allocated on the heap for all write
72 * operations that cannot be completed immediately.
74 class AsyncSocket::BytesWriteRequest : public AsyncSocket::WriteRequest {
76 static BytesWriteRequest* newRequest(AsyncSocket* socket,
77 WriteCallback* callback,
80 uint32_t partialWritten,
81 uint32_t bytesWritten,
82 unique_ptr<IOBuf>&& ioBuf,
85 // Since we put a variable size iovec array at the end
86 // of each BytesWriteRequest, we have to manually allocate the memory.
87 void* buf = malloc(sizeof(BytesWriteRequest) +
88 (opCount * sizeof(struct iovec)));
90 throw std::bad_alloc();
93 return new(buf) BytesWriteRequest(socket, callback, ops, opCount,
94 partialWritten, bytesWritten,
95 std::move(ioBuf), flags);
98 void destroy() override {
99 this->~BytesWriteRequest();
103 WriteResult performWrite() override {
104 WriteFlags writeFlags = flags_;
105 if (getNext() != nullptr) {
106 writeFlags |= WriteFlags::CORK;
109 socket_->adjustZeroCopyFlags(getOps(), getOpCount(), writeFlags);
111 auto writeResult = socket_->performWrite(
112 getOps(), getOpCount(), writeFlags, &opsWritten_, &partialBytes_);
113 bytesWritten_ = writeResult.writeReturn > 0 ? writeResult.writeReturn : 0;
115 if (socket_->isZeroCopyRequest(writeFlags)) {
117 socket_->addZeroCopyBuf(std::move(ioBuf_));
119 socket_->addZeroCopyBuf(ioBuf_.get());
122 // this happens if at least one of the prev requests were sent
123 // with zero copy but not the last one
124 if (isComplete() && socket_->getZeroCopy() &&
125 socket_->containsZeroCopyBuf(ioBuf_.get())) {
126 socket_->setZeroCopyBuf(std::move(ioBuf_));
133 bool isComplete() override {
134 return opsWritten_ == getOpCount();
137 void consume() override {
138 // Advance opIndex_ forward by opsWritten_
139 opIndex_ += opsWritten_;
140 assert(opIndex_ < opCount_);
142 if (!socket_->isZeroCopyRequest(flags_)) {
143 // If we've finished writing any IOBufs, release them
145 for (uint32_t i = opsWritten_; i != 0; --i) {
147 ioBuf_ = ioBuf_->pop();
152 // Move partialBytes_ forward into the current iovec buffer
153 struct iovec* currentOp = writeOps_ + opIndex_;
154 assert((partialBytes_ < currentOp->iov_len) || (currentOp->iov_len == 0));
155 currentOp->iov_base =
156 reinterpret_cast<uint8_t*>(currentOp->iov_base) + partialBytes_;
157 currentOp->iov_len -= partialBytes_;
159 // Increment the totalBytesWritten_ count by bytesWritten_;
160 assert(bytesWritten_ >= 0);
161 totalBytesWritten_ += uint32_t(bytesWritten_);
165 BytesWriteRequest(AsyncSocket* socket,
166 WriteCallback* callback,
167 const struct iovec* ops,
169 uint32_t partialBytes,
170 uint32_t bytesWritten,
171 unique_ptr<IOBuf>&& ioBuf,
173 : AsyncSocket::WriteRequest(socket, callback)
177 , ioBuf_(std::move(ioBuf))
179 , partialBytes_(partialBytes)
180 , bytesWritten_(bytesWritten) {
181 memcpy(writeOps_, ops, sizeof(*ops) * opCount_);
184 // private destructor, to ensure callers use destroy()
185 ~BytesWriteRequest() override = default;
187 const struct iovec* getOps() const {
188 assert(opCount_ > opIndex_);
189 return writeOps_ + opIndex_;
192 uint32_t getOpCount() const {
193 assert(opCount_ > opIndex_);
194 return opCount_ - opIndex_;
197 uint32_t opCount_; ///< number of entries in writeOps_
198 uint32_t opIndex_; ///< current index into writeOps_
199 WriteFlags flags_; ///< set for WriteFlags
200 unique_ptr<IOBuf> ioBuf_; ///< underlying IOBuf, or nullptr if N/A
202 // for consume(), how much we wrote on the last write
203 uint32_t opsWritten_; ///< complete ops written
204 uint32_t partialBytes_; ///< partial bytes of incomplete op written
205 ssize_t bytesWritten_; ///< bytes written altogether
207 struct iovec writeOps_[]; ///< write operation(s) list
210 int AsyncSocket::SendMsgParamsCallback::getDefaultFlags(
211 folly::WriteFlags flags,
212 bool zeroCopyEnabled) noexcept {
213 int msg_flags = MSG_DONTWAIT;
215 #ifdef MSG_NOSIGNAL // Linux-only
216 msg_flags |= MSG_NOSIGNAL;
218 if (isSet(flags, WriteFlags::CORK)) {
219 // MSG_MORE tells the kernel we have more data to send, so wait for us to
220 // give it the rest of the data rather than immediately sending a partial
221 // frame, even when TCP_NODELAY is enabled.
222 msg_flags |= MSG_MORE;
225 #endif // MSG_NOSIGNAL
226 if (isSet(flags, WriteFlags::EOR)) {
227 // marks that this is the last byte of a record (response)
228 msg_flags |= MSG_EOR;
231 if (zeroCopyEnabled && isSet(flags, WriteFlags::WRITE_MSG_ZEROCOPY)) {
232 msg_flags |= MSG_ZEROCOPY;
239 static AsyncSocket::SendMsgParamsCallback defaultSendMsgParamsCallback;
242 AsyncSocket::AsyncSocket()
243 : eventBase_(nullptr),
244 writeTimeout_(this, nullptr),
245 ioHandler_(this, nullptr),
246 immediateReadHandler_(this) {
247 VLOG(5) << "new AsyncSocket()";
251 AsyncSocket::AsyncSocket(EventBase* evb)
253 writeTimeout_(this, evb),
254 ioHandler_(this, evb),
255 immediateReadHandler_(this) {
256 VLOG(5) << "new AsyncSocket(" << this << ", evb=" << evb << ")";
260 AsyncSocket::AsyncSocket(EventBase* evb,
261 const folly::SocketAddress& address,
262 uint32_t connectTimeout)
264 connect(nullptr, address, connectTimeout);
267 AsyncSocket::AsyncSocket(EventBase* evb,
268 const std::string& ip,
270 uint32_t connectTimeout)
272 connect(nullptr, ip, port, connectTimeout);
275 AsyncSocket::AsyncSocket(EventBase* evb, int fd, uint32_t zeroCopyBufId)
276 : zeroCopyBufId_(zeroCopyBufId),
278 writeTimeout_(this, evb),
279 ioHandler_(this, evb, fd),
280 immediateReadHandler_(this) {
281 VLOG(5) << "new AsyncSocket(" << this << ", evb=" << evb << ", fd=" << fd
282 << ", zeroCopyBufId=" << zeroCopyBufId << ")";
286 state_ = StateEnum::ESTABLISHED;
289 AsyncSocket::AsyncSocket(AsyncSocket::UniquePtr oldAsyncSocket)
291 oldAsyncSocket->getEventBase(),
292 oldAsyncSocket->detachFd(),
293 oldAsyncSocket->getZeroCopyBufId()) {
294 preReceivedData_ = std::move(oldAsyncSocket->preReceivedData_);
297 // init() method, since constructor forwarding isn't supported in most
299 void AsyncSocket::init() {
301 eventBase_->dcheckIsInEventBaseThread();
304 state_ = StateEnum::UNINIT;
305 eventFlags_ = EventHandler::NONE;
308 maxReadsPerEvent_ = 16;
309 connectCallback_ = nullptr;
310 errMessageCallback_ = nullptr;
311 readCallback_ = nullptr;
312 writeReqHead_ = nullptr;
313 writeReqTail_ = nullptr;
314 wShutdownSocketSet_.reset();
315 appBytesWritten_ = 0;
316 appBytesReceived_ = 0;
317 sendMsgParamCallback_ = &defaultSendMsgParamsCallback;
320 AsyncSocket::~AsyncSocket() {
321 VLOG(7) << "actual destruction of AsyncSocket(this=" << this
322 << ", evb=" << eventBase_ << ", fd=" << fd_
323 << ", state=" << state_ << ")";
326 void AsyncSocket::destroy() {
327 VLOG(5) << "AsyncSocket::destroy(this=" << this << ", evb=" << eventBase_
328 << ", fd=" << fd_ << ", state=" << state_;
329 // When destroy is called, close the socket immediately
332 // Then call DelayedDestruction::destroy() to take care of
333 // whether or not we need immediate or delayed destruction
334 DelayedDestruction::destroy();
337 int AsyncSocket::detachFd() {
338 VLOG(6) << "AsyncSocket::detachFd(this=" << this << ", fd=" << fd_
339 << ", evb=" << eventBase_ << ", state=" << state_
340 << ", events=" << std::hex << eventFlags_ << ")";
341 // Extract the fd, and set fd_ to -1 first, so closeNow() won't
342 // actually close the descriptor.
343 if (const auto socketSet = wShutdownSocketSet_.lock()) {
344 socketSet->remove(fd_);
348 // Call closeNow() to invoke all pending callbacks with an error.
350 // Update the EventHandler to stop using this fd.
351 // This can only be done after closeNow() unregisters the handler.
352 ioHandler_.changeHandlerFD(-1);
356 const folly::SocketAddress& AsyncSocket::anyAddress() {
357 static const folly::SocketAddress anyAddress =
358 folly::SocketAddress("0.0.0.0", 0);
362 void AsyncSocket::setShutdownSocketSet(
363 const std::weak_ptr<ShutdownSocketSet>& wNewSS) {
364 const auto newSS = wNewSS.lock();
365 const auto shutdownSocketSet = wShutdownSocketSet_.lock();
367 if (newSS == shutdownSocketSet) {
371 if (shutdownSocketSet && fd_ != -1) {
372 shutdownSocketSet->remove(fd_);
375 if (newSS && fd_ != -1) {
379 wShutdownSocketSet_ = wNewSS;
382 void AsyncSocket::setCloseOnExec() {
383 int rv = fcntl(fd_, F_SETFD, FD_CLOEXEC);
385 auto errnoCopy = errno;
386 throw AsyncSocketException(
387 AsyncSocketException::INTERNAL_ERROR,
388 withAddr("failed to set close-on-exec flag"),
393 void AsyncSocket::connect(ConnectCallback* callback,
394 const folly::SocketAddress& address,
396 const OptionMap &options,
397 const folly::SocketAddress& bindAddr) noexcept {
398 DestructorGuard dg(this);
399 eventBase_->dcheckIsInEventBaseThread();
403 // Make sure we're in the uninitialized state
404 if (state_ != StateEnum::UNINIT) {
405 return invalidState(callback);
408 connectTimeout_ = std::chrono::milliseconds(timeout);
409 connectStartTime_ = std::chrono::steady_clock::now();
410 // Make connect end time at least >= connectStartTime.
411 connectEndTime_ = connectStartTime_;
414 state_ = StateEnum::CONNECTING;
415 connectCallback_ = callback;
417 sockaddr_storage addrStorage;
418 sockaddr* saddr = reinterpret_cast<sockaddr*>(&addrStorage);
422 // Technically the first parameter should actually be a protocol family
423 // constant (PF_xxx) rather than an address family (AF_xxx), but the
424 // distinction is mainly just historical. In pretty much all
425 // implementations the PF_foo and AF_foo constants are identical.
426 fd_ = fsp::socket(address.getFamily(), SOCK_STREAM, 0);
428 auto errnoCopy = errno;
429 throw AsyncSocketException(
430 AsyncSocketException::INTERNAL_ERROR,
431 withAddr("failed to create socket"),
434 if (const auto shutdownSocketSet = wShutdownSocketSet_.lock()) {
435 shutdownSocketSet->add(fd_);
437 ioHandler_.changeHandlerFD(fd_);
441 // Put the socket in non-blocking mode
442 int flags = fcntl(fd_, F_GETFL, 0);
444 auto errnoCopy = errno;
445 throw AsyncSocketException(
446 AsyncSocketException::INTERNAL_ERROR,
447 withAddr("failed to get socket flags"),
450 int rv = fcntl(fd_, F_SETFL, flags | O_NONBLOCK);
452 auto errnoCopy = errno;
453 throw AsyncSocketException(
454 AsyncSocketException::INTERNAL_ERROR,
455 withAddr("failed to put socket in non-blocking mode"),
459 #if !defined(MSG_NOSIGNAL) && defined(F_SETNOSIGPIPE)
460 // iOS and OS X don't support MSG_NOSIGNAL; set F_SETNOSIGPIPE instead
461 rv = fcntl(fd_, F_SETNOSIGPIPE, 1);
463 auto errnoCopy = errno;
464 throw AsyncSocketException(
465 AsyncSocketException::INTERNAL_ERROR,
466 "failed to enable F_SETNOSIGPIPE on socket",
471 // By default, turn on TCP_NODELAY
472 // If setNoDelay() fails, we continue anyway; this isn't a fatal error.
473 // setNoDelay() will log an error message if it fails.
474 // Also set the cached zeroCopyVal_ since it cannot be set earlier if the fd
476 if (address.getFamily() != AF_UNIX) {
477 (void)setNoDelay(true);
478 setZeroCopy(zeroCopyVal_);
481 VLOG(5) << "AsyncSocket::connect(this=" << this << ", evb=" << eventBase_
482 << ", fd=" << fd_ << ", host=" << address.describe().c_str();
485 if (bindAddr != anyAddress()) {
487 if (setsockopt(fd_, SOL_SOCKET, SO_REUSEADDR, &one, sizeof(one))) {
488 auto errnoCopy = errno;
490 throw AsyncSocketException(
491 AsyncSocketException::NOT_OPEN,
492 "failed to setsockopt prior to bind on " + bindAddr.describe(),
496 bindAddr.getAddress(&addrStorage);
498 if (bind(fd_, saddr, bindAddr.getActualSize()) != 0) {
499 auto errnoCopy = errno;
501 throw AsyncSocketException(
502 AsyncSocketException::NOT_OPEN,
503 "failed to bind to async socket: " + bindAddr.describe(),
508 // Apply the additional options if any.
509 for (const auto& opt: options) {
510 rv = opt.first.apply(fd_, opt.second);
512 auto errnoCopy = errno;
513 throw AsyncSocketException(
514 AsyncSocketException::INTERNAL_ERROR,
515 withAddr("failed to set socket option"),
520 // Perform the connect()
521 address.getAddress(&addrStorage);
524 state_ = StateEnum::FAST_OPEN;
525 tfoAttempted_ = true;
527 if (socketConnect(saddr, addr_.getActualSize()) < 0) {
532 // If we're still here the connect() succeeded immediately.
533 // Fall through to call the callback outside of this try...catch block
534 } catch (const AsyncSocketException& ex) {
535 return failConnect(__func__, ex);
536 } catch (const std::exception& ex) {
537 // shouldn't happen, but handle it just in case
538 VLOG(4) << "AsyncSocket::connect(this=" << this << ", fd=" << fd_
539 << "): unexpected " << typeid(ex).name() << " exception: "
541 AsyncSocketException tex(AsyncSocketException::INTERNAL_ERROR,
542 withAddr(string("unexpected exception: ") +
544 return failConnect(__func__, tex);
547 // The connection succeeded immediately
548 // The read callback may not have been set yet, and no writes may be pending
549 // yet, so we don't have to register for any events at the moment.
550 VLOG(8) << "AsyncSocket::connect succeeded immediately; this=" << this;
551 assert(errMessageCallback_ == nullptr);
552 assert(readCallback_ == nullptr);
553 assert(writeReqHead_ == nullptr);
554 if (state_ != StateEnum::FAST_OPEN) {
555 state_ = StateEnum::ESTABLISHED;
557 invokeConnectSuccess();
560 int AsyncSocket::socketConnect(const struct sockaddr* saddr, socklen_t len) {
562 if (noTransparentTls_) {
563 // Ignore return value, errors are ok
564 setsockopt(fd_, SOL_SOCKET, SO_NO_TRANSPARENT_TLS, nullptr, 0);
567 VLOG(4) << "Disabling TSOCKS for fd " << fd_;
568 // Ignore return value, errors are ok
569 setsockopt(fd_, SOL_SOCKET, SO_NO_TSOCKS, nullptr, 0);
572 int rv = fsp::connect(fd_, saddr, len);
574 auto errnoCopy = errno;
575 if (errnoCopy == EINPROGRESS) {
576 scheduleConnectTimeout();
577 registerForConnectEvents();
579 throw AsyncSocketException(
580 AsyncSocketException::NOT_OPEN,
581 "connect failed (immediately)",
588 void AsyncSocket::scheduleConnectTimeout() {
589 // Connection in progress.
590 auto timeout = connectTimeout_.count();
592 // Start a timer in case the connection takes too long.
593 if (!writeTimeout_.scheduleTimeout(uint32_t(timeout))) {
594 throw AsyncSocketException(
595 AsyncSocketException::INTERNAL_ERROR,
596 withAddr("failed to schedule AsyncSocket connect timeout"));
601 void AsyncSocket::registerForConnectEvents() {
602 // Register for write events, so we'll
603 // be notified when the connection finishes/fails.
604 // Note that we don't register for a persistent event here.
605 assert(eventFlags_ == EventHandler::NONE);
606 eventFlags_ = EventHandler::WRITE;
607 if (!ioHandler_.registerHandler(eventFlags_)) {
608 throw AsyncSocketException(
609 AsyncSocketException::INTERNAL_ERROR,
610 withAddr("failed to register AsyncSocket connect handler"));
614 void AsyncSocket::connect(ConnectCallback* callback,
615 const string& ip, uint16_t port,
617 const OptionMap &options) noexcept {
618 DestructorGuard dg(this);
620 connectCallback_ = callback;
621 connect(callback, folly::SocketAddress(ip, port), timeout, options);
622 } catch (const std::exception& ex) {
623 AsyncSocketException tex(AsyncSocketException::INTERNAL_ERROR,
625 return failConnect(__func__, tex);
629 void AsyncSocket::cancelConnect() {
630 connectCallback_ = nullptr;
631 if (state_ == StateEnum::CONNECTING || state_ == StateEnum::FAST_OPEN) {
636 void AsyncSocket::setSendTimeout(uint32_t milliseconds) {
637 sendTimeout_ = milliseconds;
639 eventBase_->dcheckIsInEventBaseThread();
642 // If we are currently pending on write requests, immediately update
643 // writeTimeout_ with the new value.
644 if ((eventFlags_ & EventHandler::WRITE) &&
645 (state_ != StateEnum::CONNECTING && state_ != StateEnum::FAST_OPEN)) {
646 assert(state_ == StateEnum::ESTABLISHED);
647 assert((shutdownFlags_ & SHUT_WRITE) == 0);
648 if (sendTimeout_ > 0) {
649 if (!writeTimeout_.scheduleTimeout(sendTimeout_)) {
650 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
651 withAddr("failed to reschedule send timeout in setSendTimeout"));
652 return failWrite(__func__, ex);
655 writeTimeout_.cancelTimeout();
660 void AsyncSocket::setErrMessageCB(ErrMessageCallback* callback) {
661 VLOG(6) << "AsyncSocket::setErrMessageCB() this=" << this
662 << ", fd=" << fd_ << ", callback=" << callback
663 << ", state=" << state_;
665 // Short circuit if callback is the same as the existing errMessageCallback_.
666 if (callback == errMessageCallback_) {
670 if (!msgErrQueueSupported) {
671 // Per-socket error message queue is not supported on this platform.
672 return invalidState(callback);
675 DestructorGuard dg(this);
676 eventBase_->dcheckIsInEventBaseThread();
678 if (callback == nullptr) {
679 // We should be able to reset the callback regardless of the
680 // socket state. It's important to have a reliable callback
681 // cancellation mechanism.
682 errMessageCallback_ = callback;
686 switch ((StateEnum)state_) {
687 case StateEnum::CONNECTING:
688 case StateEnum::FAST_OPEN:
689 case StateEnum::ESTABLISHED: {
690 errMessageCallback_ = callback;
693 case StateEnum::CLOSED:
694 case StateEnum::ERROR:
695 // We should never reach here. SHUT_READ should always be set
696 // if we are in STATE_CLOSED or STATE_ERROR.
698 return invalidState(callback);
699 case StateEnum::UNINIT:
700 // We do not allow setReadCallback() to be called before we start
702 return invalidState(callback);
705 // We don't put a default case in the switch statement, so that the compiler
706 // will warn us to update the switch statement if a new state is added.
707 return invalidState(callback);
710 AsyncSocket::ErrMessageCallback* AsyncSocket::getErrMessageCallback() const {
711 return errMessageCallback_;
714 void AsyncSocket::setSendMsgParamCB(SendMsgParamsCallback* callback) {
715 sendMsgParamCallback_ = callback;
718 AsyncSocket::SendMsgParamsCallback* AsyncSocket::getSendMsgParamsCB() const {
719 return sendMsgParamCallback_;
722 void AsyncSocket::setReadCB(ReadCallback *callback) {
723 VLOG(6) << "AsyncSocket::setReadCallback() this=" << this << ", fd=" << fd_
724 << ", callback=" << callback << ", state=" << state_;
726 // Short circuit if callback is the same as the existing readCallback_.
728 // Note that this is needed for proper functioning during some cleanup cases.
729 // During cleanup we allow setReadCallback(nullptr) to be called even if the
730 // read callback is already unset and we have been detached from an event
731 // base. This check prevents us from asserting
732 // eventBase_->isInEventBaseThread() when eventBase_ is nullptr.
733 if (callback == readCallback_) {
737 /* We are removing a read callback */
738 if (callback == nullptr &&
739 immediateReadHandler_.isLoopCallbackScheduled()) {
740 immediateReadHandler_.cancelLoopCallback();
743 if (shutdownFlags_ & SHUT_READ) {
744 // Reads have already been shut down on this socket.
746 // Allow setReadCallback(nullptr) to be called in this case, but don't
747 // allow a new callback to be set.
749 // For example, setReadCallback(nullptr) can happen after an error if we
750 // invoke some other error callback before invoking readError(). The other
751 // error callback that is invoked first may go ahead and clear the read
752 // callback before we get a chance to invoke readError().
753 if (callback != nullptr) {
754 return invalidState(callback);
756 assert((eventFlags_ & EventHandler::READ) == 0);
757 readCallback_ = nullptr;
761 DestructorGuard dg(this);
762 eventBase_->dcheckIsInEventBaseThread();
764 switch ((StateEnum)state_) {
765 case StateEnum::CONNECTING:
766 case StateEnum::FAST_OPEN:
767 // For convenience, we allow the read callback to be set while we are
768 // still connecting. We just store the callback for now. Once the
769 // connection completes we'll register for read events.
770 readCallback_ = callback;
772 case StateEnum::ESTABLISHED:
774 readCallback_ = callback;
775 uint16_t oldFlags = eventFlags_;
777 eventFlags_ |= EventHandler::READ;
779 eventFlags_ &= ~EventHandler::READ;
782 // Update our registration if our flags have changed
783 if (eventFlags_ != oldFlags) {
784 // We intentionally ignore the return value here.
785 // updateEventRegistration() will move us into the error state if it
786 // fails, and we don't need to do anything else here afterwards.
787 (void)updateEventRegistration();
791 checkForImmediateRead();
795 case StateEnum::CLOSED:
796 case StateEnum::ERROR:
797 // We should never reach here. SHUT_READ should always be set
798 // if we are in STATE_CLOSED or STATE_ERROR.
800 return invalidState(callback);
801 case StateEnum::UNINIT:
802 // We do not allow setReadCallback() to be called before we start
804 return invalidState(callback);
807 // We don't put a default case in the switch statement, so that the compiler
808 // will warn us to update the switch statement if a new state is added.
809 return invalidState(callback);
812 AsyncSocket::ReadCallback* AsyncSocket::getReadCallback() const {
813 return readCallback_;
816 bool AsyncSocket::setZeroCopy(bool enable) {
817 if (msgErrQueueSupported) {
818 zeroCopyVal_ = enable;
824 int val = enable ? 1 : 0;
825 int ret = setsockopt(fd_, SOL_SOCKET, SO_ZEROCOPY, &val, sizeof(val));
827 // if enable == false, set zeroCopyEnabled_ = false regardless
828 // if SO_ZEROCOPY is set or not
830 zeroCopyEnabled_ = enable;
834 /* if the setsockopt failed, try to see if the socket inherited the flag
835 * since we cannot set SO_ZEROCOPY on a socket s = accept
839 socklen_t optlen = sizeof(val);
840 ret = getsockopt(fd_, SOL_SOCKET, SO_ZEROCOPY, &val, &optlen);
843 enable = val ? true : false;
848 zeroCopyEnabled_ = enable;
857 void AsyncSocket::setZeroCopyWriteChainThreshold(size_t threshold) {
858 zeroCopyWriteChainThreshold_ = threshold;
861 bool AsyncSocket::isZeroCopyRequest(WriteFlags flags) {
862 return (zeroCopyEnabled_ && isSet(flags, WriteFlags::WRITE_MSG_ZEROCOPY));
865 void AsyncSocket::adjustZeroCopyFlags(
867 folly::WriteFlags& flags) {
868 if (zeroCopyEnabled_ && zeroCopyWriteChainThreshold_ && buf &&
870 if (buf->computeChainDataLength() >= zeroCopyWriteChainThreshold_) {
871 flags |= folly::WriteFlags::WRITE_MSG_ZEROCOPY;
873 flags = unSet(flags, folly::WriteFlags::WRITE_MSG_ZEROCOPY);
878 void AsyncSocket::adjustZeroCopyFlags(
881 folly::WriteFlags& flags) {
882 if (zeroCopyEnabled_ && zeroCopyWriteChainThreshold_) {
883 count = std::min<uint32_t>(count, kIovMax);
885 for (uint32_t i = 0; i < count; ++i) {
886 const iovec* v = vec + i;
890 if (sum >= zeroCopyWriteChainThreshold_) {
891 flags |= folly::WriteFlags::WRITE_MSG_ZEROCOPY;
893 flags = unSet(flags, folly::WriteFlags::WRITE_MSG_ZEROCOPY);
898 void AsyncSocket::addZeroCopyBuf(std::unique_ptr<folly::IOBuf>&& buf) {
899 uint32_t id = getNextZeroCopyBufId();
900 folly::IOBuf* ptr = buf.get();
902 idZeroCopyBufPtrMap_[id] = ptr;
903 auto& p = idZeroCopyBufInfoMap_[ptr];
905 CHECK(p.buf_.get() == nullptr);
906 p.buf_ = std::move(buf);
909 void AsyncSocket::addZeroCopyBuf(folly::IOBuf* ptr) {
910 uint32_t id = getNextZeroCopyBufId();
911 idZeroCopyBufPtrMap_[id] = ptr;
913 idZeroCopyBufInfoMap_[ptr].count_++;
916 void AsyncSocket::releaseZeroCopyBuf(uint32_t id) {
917 auto iter = idZeroCopyBufPtrMap_.find(id);
918 CHECK(iter != idZeroCopyBufPtrMap_.end());
919 auto ptr = iter->second;
920 auto iter1 = idZeroCopyBufInfoMap_.find(ptr);
921 CHECK(iter1 != idZeroCopyBufInfoMap_.end());
922 if (0 == --iter1->second.count_) {
923 idZeroCopyBufInfoMap_.erase(iter1);
927 void AsyncSocket::setZeroCopyBuf(std::unique_ptr<folly::IOBuf>&& buf) {
928 folly::IOBuf* ptr = buf.get();
929 auto& p = idZeroCopyBufInfoMap_[ptr];
930 CHECK(p.buf_.get() == nullptr);
932 p.buf_ = std::move(buf);
935 bool AsyncSocket::containsZeroCopyBuf(folly::IOBuf* ptr) {
936 return (idZeroCopyBufInfoMap_.find(ptr) != idZeroCopyBufInfoMap_.end());
939 bool AsyncSocket::isZeroCopyMsg(const cmsghdr& cmsg) const {
940 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
941 if (zeroCopyEnabled_ &&
942 ((cmsg.cmsg_level == SOL_IP && cmsg.cmsg_type == IP_RECVERR) ||
943 (cmsg.cmsg_level == SOL_IPV6 && cmsg.cmsg_type == IPV6_RECVERR))) {
944 const struct sock_extended_err* serr =
945 reinterpret_cast<const struct sock_extended_err*>(CMSG_DATA(&cmsg));
947 (serr->ee_errno == 0) && (serr->ee_origin == SO_EE_ORIGIN_ZEROCOPY));
953 void AsyncSocket::processZeroCopyMsg(const cmsghdr& cmsg) {
954 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
955 const struct sock_extended_err* serr =
956 reinterpret_cast<const struct sock_extended_err*>(CMSG_DATA(&cmsg));
957 uint32_t hi = serr->ee_data;
958 uint32_t lo = serr->ee_info;
959 // disable zero copy if the buffer was actually copied
960 if ((serr->ee_code & SO_EE_CODE_ZEROCOPY_COPIED) && zeroCopyEnabled_) {
961 VLOG(2) << "AsyncSocket::processZeroCopyMsg(): setting "
962 << "zeroCopyEnabled_ = false due to SO_EE_CODE_ZEROCOPY_COPIED "
964 zeroCopyEnabled_ = false;
967 for (uint32_t i = lo; i <= hi; i++) {
968 releaseZeroCopyBuf(i);
973 void AsyncSocket::write(WriteCallback* callback,
974 const void* buf, size_t bytes, WriteFlags flags) {
976 op.iov_base = const_cast<void*>(buf);
978 writeImpl(callback, &op, 1, unique_ptr<IOBuf>(), flags);
981 void AsyncSocket::writev(WriteCallback* callback,
985 writeImpl(callback, vec, count, unique_ptr<IOBuf>(), flags);
988 void AsyncSocket::writeChain(WriteCallback* callback, unique_ptr<IOBuf>&& buf,
990 adjustZeroCopyFlags(buf.get(), flags);
992 constexpr size_t kSmallSizeMax = 64;
993 size_t count = buf->countChainElements();
994 if (count <= kSmallSizeMax) {
995 // suppress "warning: variable length array 'vec' is used [-Wvla]"
997 FOLLY_GCC_DISABLE_WARNING("-Wvla")
998 iovec vec[BOOST_PP_IF(FOLLY_HAVE_VLA, count, kSmallSizeMax)];
1001 writeChainImpl(callback, vec, count, std::move(buf), flags);
1003 iovec* vec = new iovec[count];
1004 writeChainImpl(callback, vec, count, std::move(buf), flags);
1009 void AsyncSocket::writeChainImpl(WriteCallback* callback, iovec* vec,
1010 size_t count, unique_ptr<IOBuf>&& buf, WriteFlags flags) {
1011 size_t veclen = buf->fillIov(vec, count);
1012 writeImpl(callback, vec, veclen, std::move(buf), flags);
1015 void AsyncSocket::writeImpl(WriteCallback* callback, const iovec* vec,
1016 size_t count, unique_ptr<IOBuf>&& buf,
1018 VLOG(6) << "AsyncSocket::writev() this=" << this << ", fd=" << fd_
1019 << ", callback=" << callback << ", count=" << count
1020 << ", state=" << state_;
1021 DestructorGuard dg(this);
1022 unique_ptr<IOBuf>ioBuf(std::move(buf));
1023 eventBase_->dcheckIsInEventBaseThread();
1025 if (shutdownFlags_ & (SHUT_WRITE | SHUT_WRITE_PENDING)) {
1026 // No new writes may be performed after the write side of the socket has
1029 // We could just call callback->writeError() here to fail just this write.
1030 // However, fail hard and use invalidState() to fail all outstanding
1031 // callbacks and move the socket into the error state. There's most likely
1032 // a bug in the caller's code, so we abort everything rather than trying to
1033 // proceed as best we can.
1034 return invalidState(callback);
1037 uint32_t countWritten = 0;
1038 uint32_t partialWritten = 0;
1039 ssize_t bytesWritten = 0;
1040 bool mustRegister = false;
1041 if ((state_ == StateEnum::ESTABLISHED || state_ == StateEnum::FAST_OPEN) &&
1043 if (writeReqHead_ == nullptr) {
1044 // If we are established and there are no other writes pending,
1045 // we can attempt to perform the write immediately.
1046 assert(writeReqTail_ == nullptr);
1047 assert((eventFlags_ & EventHandler::WRITE) == 0);
1049 auto writeResult = performWrite(
1050 vec, uint32_t(count), flags, &countWritten, &partialWritten);
1051 bytesWritten = writeResult.writeReturn;
1052 if (bytesWritten < 0) {
1053 auto errnoCopy = errno;
1054 if (writeResult.exception) {
1055 return failWrite(__func__, callback, 0, *writeResult.exception);
1057 AsyncSocketException ex(
1058 AsyncSocketException::INTERNAL_ERROR,
1059 withAddr("writev failed"),
1061 return failWrite(__func__, callback, 0, ex);
1062 } else if (countWritten == count) {
1063 // done, add the whole buffer
1064 if (isZeroCopyRequest(flags)) {
1065 addZeroCopyBuf(std::move(ioBuf));
1067 // We successfully wrote everything.
1068 // Invoke the callback and return.
1070 callback->writeSuccess();
1073 } else { // continue writing the next writeReq
1075 if (isZeroCopyRequest(flags)) {
1076 addZeroCopyBuf(ioBuf.get());
1078 if (bufferCallback_) {
1079 bufferCallback_->onEgressBuffered();
1082 if (!connecting()) {
1083 // Writes might put the socket back into connecting state
1084 // if TFO is enabled, and using TFO fails.
1085 // This means that write timeouts would not be active, however
1086 // connect timeouts would affect this stage.
1087 mustRegister = true;
1090 } else if (!connecting()) {
1091 // Invalid state for writing
1092 return invalidState(callback);
1095 // Create a new WriteRequest to add to the queue
1098 req = BytesWriteRequest::newRequest(
1102 uint32_t(count - countWritten),
1104 uint32_t(bytesWritten),
1107 } catch (const std::exception& ex) {
1108 // we mainly expect to catch std::bad_alloc here
1109 AsyncSocketException tex(AsyncSocketException::INTERNAL_ERROR,
1110 withAddr(string("failed to append new WriteRequest: ") + ex.what()));
1111 return failWrite(__func__, callback, size_t(bytesWritten), tex);
1114 if (writeReqTail_ == nullptr) {
1115 assert(writeReqHead_ == nullptr);
1116 writeReqHead_ = writeReqTail_ = req;
1118 writeReqTail_->append(req);
1119 writeReqTail_ = req;
1122 // Register for write events if are established and not currently
1123 // waiting on write events
1125 assert(state_ == StateEnum::ESTABLISHED);
1126 assert((eventFlags_ & EventHandler::WRITE) == 0);
1127 if (!updateEventRegistration(EventHandler::WRITE, 0)) {
1128 assert(state_ == StateEnum::ERROR);
1131 if (sendTimeout_ > 0) {
1132 // Schedule a timeout to fire if the write takes too long.
1133 if (!writeTimeout_.scheduleTimeout(sendTimeout_)) {
1134 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
1135 withAddr("failed to schedule send timeout"));
1136 return failWrite(__func__, ex);
1142 void AsyncSocket::writeRequest(WriteRequest* req) {
1143 if (writeReqTail_ == nullptr) {
1144 assert(writeReqHead_ == nullptr);
1145 writeReqHead_ = writeReqTail_ = req;
1148 writeReqTail_->append(req);
1149 writeReqTail_ = req;
1153 void AsyncSocket::close() {
1154 VLOG(5) << "AsyncSocket::close(): this=" << this << ", fd_=" << fd_
1155 << ", state=" << state_ << ", shutdownFlags="
1156 << std::hex << (int) shutdownFlags_;
1158 // close() is only different from closeNow() when there are pending writes
1159 // that need to drain before we can close. In all other cases, just call
1162 // Note that writeReqHead_ can be non-nullptr even in STATE_CLOSED or
1163 // STATE_ERROR if close() is invoked while a previous closeNow() or failure
1164 // is still running. (e.g., If there are multiple pending writes, and we
1165 // call writeError() on the first one, it may call close(). In this case we
1166 // will already be in STATE_CLOSED or STATE_ERROR, but the remaining pending
1167 // writes will still be in the queue.)
1169 // We only need to drain pending writes if we are still in STATE_CONNECTING
1170 // or STATE_ESTABLISHED
1171 if ((writeReqHead_ == nullptr) ||
1172 !(state_ == StateEnum::CONNECTING ||
1173 state_ == StateEnum::ESTABLISHED)) {
1178 // Declare a DestructorGuard to ensure that the AsyncSocket cannot be
1179 // destroyed until close() returns.
1180 DestructorGuard dg(this);
1181 eventBase_->dcheckIsInEventBaseThread();
1183 // Since there are write requests pending, we have to set the
1184 // SHUT_WRITE_PENDING flag, and wait to perform the real close until the
1185 // connect finishes and we finish writing these requests.
1187 // Set SHUT_READ to indicate that reads are shut down, and set the
1188 // SHUT_WRITE_PENDING flag to mark that we want to shutdown once the
1189 // pending writes complete.
1190 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE_PENDING);
1192 // If a read callback is set, invoke readEOF() immediately to inform it that
1193 // the socket has been closed and no more data can be read.
1194 if (readCallback_) {
1195 // Disable reads if they are enabled
1196 if (!updateEventRegistration(0, EventHandler::READ)) {
1197 // We're now in the error state; callbacks have been cleaned up
1198 assert(state_ == StateEnum::ERROR);
1199 assert(readCallback_ == nullptr);
1201 ReadCallback* callback = readCallback_;
1202 readCallback_ = nullptr;
1203 callback->readEOF();
1208 void AsyncSocket::closeNow() {
1209 VLOG(5) << "AsyncSocket::closeNow(): this=" << this << ", fd_=" << fd_
1210 << ", state=" << state_ << ", shutdownFlags="
1211 << std::hex << (int) shutdownFlags_;
1212 DestructorGuard dg(this);
1214 eventBase_->dcheckIsInEventBaseThread();
1218 case StateEnum::ESTABLISHED:
1219 case StateEnum::CONNECTING:
1220 case StateEnum::FAST_OPEN: {
1221 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE);
1222 state_ = StateEnum::CLOSED;
1224 // If the write timeout was set, cancel it.
1225 writeTimeout_.cancelTimeout();
1227 // If we are registered for I/O events, unregister.
1228 if (eventFlags_ != EventHandler::NONE) {
1229 eventFlags_ = EventHandler::NONE;
1230 if (!updateEventRegistration()) {
1231 // We will have been moved into the error state.
1232 assert(state_ == StateEnum::ERROR);
1237 if (immediateReadHandler_.isLoopCallbackScheduled()) {
1238 immediateReadHandler_.cancelLoopCallback();
1242 ioHandler_.changeHandlerFD(-1);
1246 invokeConnectErr(socketClosedLocallyEx);
1248 failAllWrites(socketClosedLocallyEx);
1250 if (readCallback_) {
1251 ReadCallback* callback = readCallback_;
1252 readCallback_ = nullptr;
1253 callback->readEOF();
1257 case StateEnum::CLOSED:
1258 // Do nothing. It's possible that we are being called recursively
1259 // from inside a callback that we invoked inside another call to close()
1260 // that is still running.
1262 case StateEnum::ERROR:
1263 // Do nothing. The error handling code has performed (or is performing)
1266 case StateEnum::UNINIT:
1267 assert(eventFlags_ == EventHandler::NONE);
1268 assert(connectCallback_ == nullptr);
1269 assert(readCallback_ == nullptr);
1270 assert(writeReqHead_ == nullptr);
1271 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE);
1272 state_ = StateEnum::CLOSED;
1276 LOG(DFATAL) << "AsyncSocket::closeNow() (this=" << this << ", fd=" << fd_
1277 << ") called in unknown state " << state_;
1280 void AsyncSocket::closeWithReset() {
1281 // Enable SO_LINGER, with the linger timeout set to 0.
1282 // This will trigger a TCP reset when we close the socket.
1284 struct linger optLinger = {1, 0};
1285 if (setSockOpt(SOL_SOCKET, SO_LINGER, &optLinger) != 0) {
1286 VLOG(2) << "AsyncSocket::closeWithReset(): error setting SO_LINGER "
1287 << "on " << fd_ << ": errno=" << errno;
1291 // Then let closeNow() take care of the rest
1295 void AsyncSocket::shutdownWrite() {
1296 VLOG(5) << "AsyncSocket::shutdownWrite(): this=" << this << ", fd=" << fd_
1297 << ", state=" << state_ << ", shutdownFlags="
1298 << std::hex << (int) shutdownFlags_;
1300 // If there are no pending writes, shutdownWrite() is identical to
1301 // shutdownWriteNow().
1302 if (writeReqHead_ == nullptr) {
1307 eventBase_->dcheckIsInEventBaseThread();
1309 // There are pending writes. Set SHUT_WRITE_PENDING so that the actual
1310 // shutdown will be performed once all writes complete.
1311 shutdownFlags_ |= SHUT_WRITE_PENDING;
1314 void AsyncSocket::shutdownWriteNow() {
1315 VLOG(5) << "AsyncSocket::shutdownWriteNow(): this=" << this
1316 << ", fd=" << fd_ << ", state=" << state_
1317 << ", shutdownFlags=" << std::hex << (int) shutdownFlags_;
1319 if (shutdownFlags_ & SHUT_WRITE) {
1320 // Writes are already shutdown; nothing else to do.
1324 // If SHUT_READ is already set, just call closeNow() to completely
1325 // close the socket. This can happen if close() was called with writes
1326 // pending, and then shutdownWriteNow() is called before all pending writes
1328 if (shutdownFlags_ & SHUT_READ) {
1333 DestructorGuard dg(this);
1335 eventBase_->dcheckIsInEventBaseThread();
1338 switch (static_cast<StateEnum>(state_)) {
1339 case StateEnum::ESTABLISHED:
1341 shutdownFlags_ |= SHUT_WRITE;
1343 // If the write timeout was set, cancel it.
1344 writeTimeout_.cancelTimeout();
1346 // If we are registered for write events, unregister.
1347 if (!updateEventRegistration(0, EventHandler::WRITE)) {
1348 // We will have been moved into the error state.
1349 assert(state_ == StateEnum::ERROR);
1353 // Shutdown writes on the file descriptor
1354 shutdown(fd_, SHUT_WR);
1356 // Immediately fail all write requests
1357 failAllWrites(socketShutdownForWritesEx);
1360 case StateEnum::CONNECTING:
1362 // Set the SHUT_WRITE_PENDING flag.
1363 // When the connection completes, it will check this flag,
1364 // shutdown the write half of the socket, and then set SHUT_WRITE.
1365 shutdownFlags_ |= SHUT_WRITE_PENDING;
1367 // Immediately fail all write requests
1368 failAllWrites(socketShutdownForWritesEx);
1371 case StateEnum::UNINIT:
1372 // Callers normally shouldn't call shutdownWriteNow() before the socket
1373 // even starts connecting. Nonetheless, go ahead and set
1374 // SHUT_WRITE_PENDING. Once the socket eventually connects it will
1375 // immediately shut down the write side of the socket.
1376 shutdownFlags_ |= SHUT_WRITE_PENDING;
1378 case StateEnum::FAST_OPEN:
1379 // In fast open state we haven't call connected yet, and if we shutdown
1380 // the writes, we will never try to call connect, so shut everything down
1381 shutdownFlags_ |= SHUT_WRITE;
1382 // Immediately fail all write requests
1383 failAllWrites(socketShutdownForWritesEx);
1385 case StateEnum::CLOSED:
1386 case StateEnum::ERROR:
1387 // We should never get here. SHUT_WRITE should always be set
1388 // in STATE_CLOSED and STATE_ERROR.
1389 VLOG(4) << "AsyncSocket::shutdownWriteNow() (this=" << this
1390 << ", fd=" << fd_ << ") in unexpected state " << state_
1391 << " with SHUT_WRITE not set ("
1392 << std::hex << (int) shutdownFlags_ << ")";
1397 LOG(DFATAL) << "AsyncSocket::shutdownWriteNow() (this=" << this << ", fd="
1398 << fd_ << ") called in unknown state " << state_;
1401 bool AsyncSocket::readable() const {
1405 struct pollfd fds[1];
1407 fds[0].events = POLLIN;
1409 int rc = poll(fds, 1, 0);
1413 bool AsyncSocket::writable() const {
1417 struct pollfd fds[1];
1419 fds[0].events = POLLOUT;
1421 int rc = poll(fds, 1, 0);
1425 bool AsyncSocket::isPending() const {
1426 return ioHandler_.isPending();
1429 bool AsyncSocket::hangup() const {
1431 // sanity check, no one should ask for hangup if we are not connected.
1435 #ifdef POLLRDHUP // Linux-only
1436 struct pollfd fds[1];
1438 fds[0].events = POLLRDHUP|POLLHUP;
1441 return (fds[0].revents & (POLLRDHUP|POLLHUP)) != 0;
1447 bool AsyncSocket::good() const {
1449 (state_ == StateEnum::CONNECTING || state_ == StateEnum::FAST_OPEN ||
1450 state_ == StateEnum::ESTABLISHED) &&
1451 (shutdownFlags_ == 0) && (eventBase_ != nullptr));
1454 bool AsyncSocket::error() const {
1455 return (state_ == StateEnum::ERROR);
1458 void AsyncSocket::attachEventBase(EventBase* eventBase) {
1459 VLOG(5) << "AsyncSocket::attachEventBase(this=" << this << ", fd=" << fd_
1460 << ", old evb=" << eventBase_ << ", new evb=" << eventBase
1461 << ", state=" << state_ << ", events="
1462 << std::hex << eventFlags_ << ")";
1463 assert(eventBase_ == nullptr);
1464 eventBase->dcheckIsInEventBaseThread();
1466 eventBase_ = eventBase;
1467 ioHandler_.attachEventBase(eventBase);
1468 writeTimeout_.attachEventBase(eventBase);
1470 evbChangeCb_->evbAttached(this);
1474 void AsyncSocket::detachEventBase() {
1475 VLOG(5) << "AsyncSocket::detachEventBase(this=" << this << ", fd=" << fd_
1476 << ", old evb=" << eventBase_ << ", state=" << state_
1477 << ", events=" << std::hex << eventFlags_ << ")";
1478 assert(eventBase_ != nullptr);
1479 eventBase_->dcheckIsInEventBaseThread();
1481 eventBase_ = nullptr;
1482 ioHandler_.detachEventBase();
1483 writeTimeout_.detachEventBase();
1485 evbChangeCb_->evbDetached(this);
1489 bool AsyncSocket::isDetachable() const {
1490 DCHECK(eventBase_ != nullptr);
1491 eventBase_->dcheckIsInEventBaseThread();
1493 return !ioHandler_.isHandlerRegistered() && !writeTimeout_.isScheduled();
1496 void AsyncSocket::cacheAddresses() {
1499 cacheLocalAddress();
1501 } catch (const std::system_error& e) {
1502 if (e.code() != std::error_code(ENOTCONN, std::system_category())) {
1503 VLOG(1) << "Error caching addresses: " << e.code().value() << ", "
1504 << e.code().message();
1510 void AsyncSocket::cacheLocalAddress() const {
1511 if (!localAddr_.isInitialized()) {
1512 localAddr_.setFromLocalAddress(fd_);
1516 void AsyncSocket::cachePeerAddress() const {
1517 if (!addr_.isInitialized()) {
1518 addr_.setFromPeerAddress(fd_);
1522 bool AsyncSocket::isZeroCopyWriteInProgress() const noexcept {
1523 eventBase_->dcheckIsInEventBaseThread();
1524 return (!idZeroCopyBufPtrMap_.empty());
1527 void AsyncSocket::getLocalAddress(folly::SocketAddress* address) const {
1528 cacheLocalAddress();
1529 *address = localAddr_;
1532 void AsyncSocket::getPeerAddress(folly::SocketAddress* address) const {
1537 bool AsyncSocket::getTFOSucceded() const {
1538 return detail::tfo_succeeded(fd_);
1541 int AsyncSocket::setNoDelay(bool noDelay) {
1543 VLOG(4) << "AsyncSocket::setNoDelay() called on non-open socket "
1544 << this << "(state=" << state_ << ")";
1549 int value = noDelay ? 1 : 0;
1550 if (setsockopt(fd_, IPPROTO_TCP, TCP_NODELAY, &value, sizeof(value)) != 0) {
1551 int errnoCopy = errno;
1552 VLOG(2) << "failed to update TCP_NODELAY option on AsyncSocket "
1553 << this << " (fd=" << fd_ << ", state=" << state_ << "): "
1554 << strerror(errnoCopy);
1561 int AsyncSocket::setCongestionFlavor(const std::string &cname) {
1563 #ifndef TCP_CONGESTION
1564 #define TCP_CONGESTION 13
1568 VLOG(4) << "AsyncSocket::setCongestionFlavor() called on non-open "
1569 << "socket " << this << "(state=" << state_ << ")";
1579 socklen_t(cname.length() + 1)) != 0) {
1580 int errnoCopy = errno;
1581 VLOG(2) << "failed to update TCP_CONGESTION option on AsyncSocket "
1582 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1583 << strerror(errnoCopy);
1590 int AsyncSocket::setQuickAck(bool quickack) {
1593 VLOG(4) << "AsyncSocket::setQuickAck() called on non-open socket "
1594 << this << "(state=" << state_ << ")";
1599 #ifdef TCP_QUICKACK // Linux-only
1600 int value = quickack ? 1 : 0;
1601 if (setsockopt(fd_, IPPROTO_TCP, TCP_QUICKACK, &value, sizeof(value)) != 0) {
1602 int errnoCopy = errno;
1603 VLOG(2) << "failed to update TCP_QUICKACK option on AsyncSocket"
1604 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1605 << strerror(errnoCopy);
1615 int AsyncSocket::setSendBufSize(size_t bufsize) {
1617 VLOG(4) << "AsyncSocket::setSendBufSize() called on non-open socket "
1618 << this << "(state=" << state_ << ")";
1622 if (setsockopt(fd_, SOL_SOCKET, SO_SNDBUF, &bufsize, sizeof(bufsize)) !=0) {
1623 int errnoCopy = errno;
1624 VLOG(2) << "failed to update SO_SNDBUF option on AsyncSocket"
1625 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1626 << strerror(errnoCopy);
1633 int AsyncSocket::setRecvBufSize(size_t bufsize) {
1635 VLOG(4) << "AsyncSocket::setRecvBufSize() called on non-open socket "
1636 << this << "(state=" << state_ << ")";
1640 if (setsockopt(fd_, SOL_SOCKET, SO_RCVBUF, &bufsize, sizeof(bufsize)) !=0) {
1641 int errnoCopy = errno;
1642 VLOG(2) << "failed to update SO_RCVBUF option on AsyncSocket"
1643 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1644 << strerror(errnoCopy);
1651 int AsyncSocket::setTCPProfile(int profd) {
1653 VLOG(4) << "AsyncSocket::setTCPProfile() called on non-open socket "
1654 << this << "(state=" << state_ << ")";
1658 if (setsockopt(fd_, SOL_SOCKET, SO_SET_NAMESPACE, &profd, sizeof(int)) !=0) {
1659 int errnoCopy = errno;
1660 VLOG(2) << "failed to set socket namespace option on AsyncSocket"
1661 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1662 << strerror(errnoCopy);
1669 void AsyncSocket::ioReady(uint16_t events) noexcept {
1670 VLOG(7) << "AsyncSocket::ioRead() this=" << this << ", fd=" << fd_
1671 << ", events=" << std::hex << events << ", state=" << state_;
1672 DestructorGuard dg(this);
1673 assert(events & EventHandler::READ_WRITE);
1674 eventBase_->dcheckIsInEventBaseThread();
1676 uint16_t relevantEvents = uint16_t(events & EventHandler::READ_WRITE);
1677 EventBase* originalEventBase = eventBase_;
1678 // If we got there it means that either EventHandler::READ or
1679 // EventHandler::WRITE is set. Any of these flags can
1680 // indicate that there are messages available in the socket
1681 // error message queue.
1682 handleErrMessages();
1684 // Return now if handleErrMessages() detached us from our EventBase
1685 if (eventBase_ != originalEventBase) {
1689 if (relevantEvents == EventHandler::READ) {
1691 } else if (relevantEvents == EventHandler::WRITE) {
1693 } else if (relevantEvents == EventHandler::READ_WRITE) {
1694 // If both read and write events are ready, process writes first.
1697 // Return now if handleWrite() detached us from our EventBase
1698 if (eventBase_ != originalEventBase) {
1702 // Only call handleRead() if a read callback is still installed.
1703 // (It's possible that the read callback was uninstalled during
1705 if (readCallback_) {
1709 VLOG(4) << "AsyncSocket::ioRead() called with unexpected events "
1710 << std::hex << events << "(this=" << this << ")";
1715 AsyncSocket::ReadResult
1716 AsyncSocket::performRead(void** buf, size_t* buflen, size_t* /* offset */) {
1717 VLOG(5) << "AsyncSocket::performRead() this=" << this << ", buf=" << *buf
1718 << ", buflen=" << *buflen;
1720 if (preReceivedData_ && !preReceivedData_->empty()) {
1721 VLOG(5) << "AsyncSocket::performRead() this=" << this
1722 << ", reading pre-received data";
1724 io::Cursor cursor(preReceivedData_.get());
1725 auto len = cursor.pullAtMost(*buf, *buflen);
1728 queue.append(std::move(preReceivedData_));
1729 queue.trimStart(len);
1730 preReceivedData_ = queue.move();
1732 appBytesReceived_ += len;
1733 return ReadResult(len);
1736 ssize_t bytes = recv(fd_, *buf, *buflen, MSG_DONTWAIT);
1738 if (errno == EAGAIN || errno == EWOULDBLOCK) {
1739 // No more data to read right now.
1740 return ReadResult(READ_BLOCKING);
1742 return ReadResult(READ_ERROR);
1745 appBytesReceived_ += bytes;
1746 return ReadResult(bytes);
1750 void AsyncSocket::prepareReadBuffer(void** buf, size_t* buflen) {
1751 // no matter what, buffer should be preapared for non-ssl socket
1752 CHECK(readCallback_);
1753 readCallback_->getReadBuffer(buf, buflen);
1756 void AsyncSocket::handleErrMessages() noexcept {
1757 // This method has non-empty implementation only for platforms
1758 // supporting per-socket error queues.
1759 VLOG(5) << "AsyncSocket::handleErrMessages() this=" << this << ", fd=" << fd_
1760 << ", state=" << state_;
1761 if (errMessageCallback_ == nullptr &&
1762 (!zeroCopyEnabled_ || idZeroCopyBufPtrMap_.empty())) {
1763 VLOG(7) << "AsyncSocket::handleErrMessages(): "
1764 << "no callback installed - exiting.";
1768 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
1774 entry.iov_base = &data;
1775 entry.iov_len = sizeof(data);
1776 msg.msg_iov = &entry;
1778 msg.msg_name = nullptr;
1779 msg.msg_namelen = 0;
1780 msg.msg_control = ctrl;
1781 msg.msg_controllen = sizeof(ctrl);
1786 ret = recvmsg(fd_, &msg, MSG_ERRQUEUE);
1787 VLOG(5) << "AsyncSocket::handleErrMessages(): recvmsg returned " << ret;
1790 if (errno != EAGAIN) {
1791 auto errnoCopy = errno;
1792 LOG(ERROR) << "::recvmsg exited with code " << ret
1793 << ", errno: " << errnoCopy;
1794 AsyncSocketException ex(
1795 AsyncSocketException::INTERNAL_ERROR,
1796 withAddr("recvmsg() failed"),
1798 failErrMessageRead(__func__, ex);
1803 for (struct cmsghdr* cmsg = CMSG_FIRSTHDR(&msg);
1804 cmsg != nullptr && cmsg->cmsg_len != 0;
1805 cmsg = CMSG_NXTHDR(&msg, cmsg)) {
1806 if (isZeroCopyMsg(*cmsg)) {
1807 processZeroCopyMsg(*cmsg);
1809 if (errMessageCallback_) {
1810 errMessageCallback_->errMessage(*cmsg);
1815 #endif // FOLLY_HAVE_MSG_ERRQUEUE
1818 void AsyncSocket::handleRead() noexcept {
1819 VLOG(5) << "AsyncSocket::handleRead() this=" << this << ", fd=" << fd_
1820 << ", state=" << state_;
1821 assert(state_ == StateEnum::ESTABLISHED);
1822 assert((shutdownFlags_ & SHUT_READ) == 0);
1823 assert(readCallback_ != nullptr);
1824 assert(eventFlags_ & EventHandler::READ);
1827 // - a read attempt would block
1828 // - readCallback_ is uninstalled
1829 // - the number of loop iterations exceeds the optional maximum
1830 // - this AsyncSocket is moved to another EventBase
1832 // When we invoke readDataAvailable() it may uninstall the readCallback_,
1833 // which is why need to check for it here.
1835 // The last bullet point is slightly subtle. readDataAvailable() may also
1836 // detach this socket from this EventBase. However, before
1837 // readDataAvailable() returns another thread may pick it up, attach it to
1838 // a different EventBase, and install another readCallback_. We need to
1839 // exit immediately after readDataAvailable() returns if the eventBase_ has
1840 // changed. (The caller must perform some sort of locking to transfer the
1841 // AsyncSocket between threads properly. This will be sufficient to ensure
1842 // that this thread sees the updated eventBase_ variable after
1843 // readDataAvailable() returns.)
1844 uint16_t numReads = 0;
1845 EventBase* originalEventBase = eventBase_;
1846 while (readCallback_ && eventBase_ == originalEventBase) {
1847 // Get the buffer to read into.
1848 void* buf = nullptr;
1849 size_t buflen = 0, offset = 0;
1851 prepareReadBuffer(&buf, &buflen);
1852 VLOG(5) << "prepareReadBuffer() buf=" << buf << ", buflen=" << buflen;
1853 } catch (const AsyncSocketException& ex) {
1854 return failRead(__func__, ex);
1855 } catch (const std::exception& ex) {
1856 AsyncSocketException tex(AsyncSocketException::BAD_ARGS,
1857 string("ReadCallback::getReadBuffer() "
1858 "threw exception: ") +
1860 return failRead(__func__, tex);
1862 AsyncSocketException ex(AsyncSocketException::BAD_ARGS,
1863 "ReadCallback::getReadBuffer() threw "
1864 "non-exception type");
1865 return failRead(__func__, ex);
1867 if (!isBufferMovable_ && (buf == nullptr || buflen == 0)) {
1868 AsyncSocketException ex(AsyncSocketException::BAD_ARGS,
1869 "ReadCallback::getReadBuffer() returned "
1871 return failRead(__func__, ex);
1875 auto readResult = performRead(&buf, &buflen, &offset);
1876 auto bytesRead = readResult.readReturn;
1877 VLOG(4) << "this=" << this << ", AsyncSocket::handleRead() got "
1878 << bytesRead << " bytes";
1879 if (bytesRead > 0) {
1880 if (!isBufferMovable_) {
1881 readCallback_->readDataAvailable(size_t(bytesRead));
1883 CHECK(kOpenSslModeMoveBufferOwnership);
1884 VLOG(5) << "this=" << this << ", AsyncSocket::handleRead() got "
1885 << "buf=" << buf << ", " << bytesRead << "/" << buflen
1886 << ", offset=" << offset;
1887 auto readBuf = folly::IOBuf::takeOwnership(buf, buflen);
1888 readBuf->trimStart(offset);
1889 readBuf->trimEnd(buflen - offset - bytesRead);
1890 readCallback_->readBufferAvailable(std::move(readBuf));
1893 // Fall through and continue around the loop if the read
1894 // completely filled the available buffer.
1895 // Note that readCallback_ may have been uninstalled or changed inside
1896 // readDataAvailable().
1897 if (size_t(bytesRead) < buflen) {
1900 } else if (bytesRead == READ_BLOCKING) {
1901 // No more data to read right now.
1903 } else if (bytesRead == READ_ERROR) {
1904 readErr_ = READ_ERROR;
1905 if (readResult.exception) {
1906 return failRead(__func__, *readResult.exception);
1908 auto errnoCopy = errno;
1909 AsyncSocketException ex(
1910 AsyncSocketException::INTERNAL_ERROR,
1911 withAddr("recv() failed"),
1913 return failRead(__func__, ex);
1915 assert(bytesRead == READ_EOF);
1916 readErr_ = READ_EOF;
1918 shutdownFlags_ |= SHUT_READ;
1919 if (!updateEventRegistration(0, EventHandler::READ)) {
1920 // we've already been moved into STATE_ERROR
1921 assert(state_ == StateEnum::ERROR);
1922 assert(readCallback_ == nullptr);
1926 ReadCallback* callback = readCallback_;
1927 readCallback_ = nullptr;
1928 callback->readEOF();
1931 if (maxReadsPerEvent_ && (++numReads >= maxReadsPerEvent_)) {
1932 if (readCallback_ != nullptr) {
1933 // We might still have data in the socket.
1934 // (e.g. see comment in AsyncSSLSocket::checkForImmediateRead)
1935 scheduleImmediateRead();
1943 * This function attempts to write as much data as possible, until no more data
1946 * - If it sends all available data, it unregisters for write events, and stops
1947 * the writeTimeout_.
1949 * - If not all of the data can be sent immediately, it reschedules
1950 * writeTimeout_ (if a non-zero timeout is set), and ensures the handler is
1951 * registered for write events.
1953 void AsyncSocket::handleWrite() noexcept {
1954 VLOG(5) << "AsyncSocket::handleWrite() this=" << this << ", fd=" << fd_
1955 << ", state=" << state_;
1956 DestructorGuard dg(this);
1958 if (state_ == StateEnum::CONNECTING) {
1964 assert(state_ == StateEnum::ESTABLISHED);
1965 assert((shutdownFlags_ & SHUT_WRITE) == 0);
1966 assert(writeReqHead_ != nullptr);
1968 // Loop until we run out of write requests,
1969 // or until this socket is moved to another EventBase.
1970 // (See the comment in handleRead() explaining how this can happen.)
1971 EventBase* originalEventBase = eventBase_;
1972 while (writeReqHead_ != nullptr && eventBase_ == originalEventBase) {
1973 auto writeResult = writeReqHead_->performWrite();
1974 if (writeResult.writeReturn < 0) {
1975 if (writeResult.exception) {
1976 return failWrite(__func__, *writeResult.exception);
1978 auto errnoCopy = errno;
1979 AsyncSocketException ex(
1980 AsyncSocketException::INTERNAL_ERROR,
1981 withAddr("writev() failed"),
1983 return failWrite(__func__, ex);
1984 } else if (writeReqHead_->isComplete()) {
1985 // We finished this request
1986 WriteRequest* req = writeReqHead_;
1987 writeReqHead_ = req->getNext();
1989 if (writeReqHead_ == nullptr) {
1990 writeReqTail_ = nullptr;
1991 // This is the last write request.
1992 // Unregister for write events and cancel the send timer
1993 // before we invoke the callback. We have to update the state properly
1994 // before calling the callback, since it may want to detach us from
1996 if (eventFlags_ & EventHandler::WRITE) {
1997 if (!updateEventRegistration(0, EventHandler::WRITE)) {
1998 assert(state_ == StateEnum::ERROR);
2001 // Stop the send timeout
2002 writeTimeout_.cancelTimeout();
2004 assert(!writeTimeout_.isScheduled());
2006 // If SHUT_WRITE_PENDING is set, we should shutdown the socket after
2007 // we finish sending the last write request.
2009 // We have to do this before invoking writeSuccess(), since
2010 // writeSuccess() may detach us from our EventBase.
2011 if (shutdownFlags_ & SHUT_WRITE_PENDING) {
2012 assert(connectCallback_ == nullptr);
2013 shutdownFlags_ |= SHUT_WRITE;
2015 if (shutdownFlags_ & SHUT_READ) {
2016 // Reads have already been shutdown. Fully close the socket and
2017 // move to STATE_CLOSED.
2019 // Note: This code currently moves us to STATE_CLOSED even if
2020 // close() hasn't ever been called. This can occur if we have
2021 // received EOF from the peer and shutdownWrite() has been called
2022 // locally. Should we bother staying in STATE_ESTABLISHED in this
2023 // case, until close() is actually called? I can't think of a
2024 // reason why we would need to do so. No other operations besides
2025 // calling close() or destroying the socket can be performed at
2027 assert(readCallback_ == nullptr);
2028 state_ = StateEnum::CLOSED;
2030 ioHandler_.changeHandlerFD(-1);
2034 // Reads are still enabled, so we are only doing a half-shutdown
2035 shutdown(fd_, SHUT_WR);
2040 // Invoke the callback
2041 WriteCallback* callback = req->getCallback();
2044 callback->writeSuccess();
2046 // We'll continue around the loop, trying to write another request
2049 if (bufferCallback_) {
2050 bufferCallback_->onEgressBuffered();
2052 writeReqHead_->consume();
2053 // Stop after a partial write; it's highly likely that a subsequent write
2054 // attempt will just return EAGAIN.
2056 // Ensure that we are registered for write events.
2057 if ((eventFlags_ & EventHandler::WRITE) == 0) {
2058 if (!updateEventRegistration(EventHandler::WRITE, 0)) {
2059 assert(state_ == StateEnum::ERROR);
2064 // Reschedule the send timeout, since we have made some write progress.
2065 if (sendTimeout_ > 0) {
2066 if (!writeTimeout_.scheduleTimeout(sendTimeout_)) {
2067 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
2068 withAddr("failed to reschedule write timeout"));
2069 return failWrite(__func__, ex);
2075 if (!writeReqHead_ && bufferCallback_) {
2076 bufferCallback_->onEgressBufferCleared();
2080 void AsyncSocket::checkForImmediateRead() noexcept {
2081 // We currently don't attempt to perform optimistic reads in AsyncSocket.
2082 // (However, note that some subclasses do override this method.)
2084 // Simply calling handleRead() here would be bad, as this would call
2085 // readCallback_->getReadBuffer(), forcing the callback to allocate a read
2086 // buffer even though no data may be available. This would waste lots of
2087 // memory, since the buffer will sit around unused until the socket actually
2088 // becomes readable.
2090 // Checking if the socket is readable now also seems like it would probably
2091 // be a pessimism. In most cases it probably wouldn't be readable, and we
2092 // would just waste an extra system call. Even if it is readable, waiting to
2093 // find out from libevent on the next event loop doesn't seem that bad.
2095 // The exception to this is if we have pre-received data. In that case there
2096 // is definitely data available immediately.
2097 if (preReceivedData_ && !preReceivedData_->empty()) {
2102 void AsyncSocket::handleInitialReadWrite() noexcept {
2103 // Our callers should already be holding a DestructorGuard, but grab
2104 // one here just to make sure, in case one of our calling code paths ever
2106 DestructorGuard dg(this);
2107 // If we have a readCallback_, make sure we enable read events. We
2108 // may already be registered for reads if connectSuccess() set
2109 // the read calback.
2110 if (readCallback_ && !(eventFlags_ & EventHandler::READ)) {
2111 assert(state_ == StateEnum::ESTABLISHED);
2112 assert((shutdownFlags_ & SHUT_READ) == 0);
2113 if (!updateEventRegistration(EventHandler::READ, 0)) {
2114 assert(state_ == StateEnum::ERROR);
2117 checkForImmediateRead();
2118 } else if (readCallback_ == nullptr) {
2119 // Unregister for read events.
2120 updateEventRegistration(0, EventHandler::READ);
2123 // If we have write requests pending, try to send them immediately.
2124 // Since we just finished accepting, there is a very good chance that we can
2125 // write without blocking.
2127 // However, we only process them if EventHandler::WRITE is not already set,
2128 // which means that we're already blocked on a write attempt. (This can
2129 // happen if connectSuccess() called write() before returning.)
2130 if (writeReqHead_ && !(eventFlags_ & EventHandler::WRITE)) {
2131 // Call handleWrite() to perform write processing.
2133 } else if (writeReqHead_ == nullptr) {
2134 // Unregister for write event.
2135 updateEventRegistration(0, EventHandler::WRITE);
2139 void AsyncSocket::handleConnect() noexcept {
2140 VLOG(5) << "AsyncSocket::handleConnect() this=" << this << ", fd=" << fd_
2141 << ", state=" << state_;
2142 assert(state_ == StateEnum::CONNECTING);
2143 // SHUT_WRITE can never be set while we are still connecting;
2144 // SHUT_WRITE_PENDING may be set, be we only set SHUT_WRITE once the connect
2146 assert((shutdownFlags_ & SHUT_WRITE) == 0);
2148 // In case we had a connect timeout, cancel the timeout
2149 writeTimeout_.cancelTimeout();
2150 // We don't use a persistent registration when waiting on a connect event,
2151 // so we have been automatically unregistered now. Update eventFlags_ to
2153 assert(eventFlags_ == EventHandler::WRITE);
2154 eventFlags_ = EventHandler::NONE;
2156 // Call getsockopt() to check if the connect succeeded
2158 socklen_t len = sizeof(error);
2159 int rv = getsockopt(fd_, SOL_SOCKET, SO_ERROR, &error, &len);
2161 auto errnoCopy = errno;
2162 AsyncSocketException ex(
2163 AsyncSocketException::INTERNAL_ERROR,
2164 withAddr("error calling getsockopt() after connect"),
2166 VLOG(4) << "AsyncSocket::handleConnect(this=" << this << ", fd="
2167 << fd_ << " host=" << addr_.describe()
2168 << ") exception:" << ex.what();
2169 return failConnect(__func__, ex);
2173 AsyncSocketException ex(AsyncSocketException::NOT_OPEN,
2174 "connect failed", error);
2175 VLOG(1) << "AsyncSocket::handleConnect(this=" << this << ", fd="
2176 << fd_ << " host=" << addr_.describe()
2177 << ") exception: " << ex.what();
2178 return failConnect(__func__, ex);
2181 // Move into STATE_ESTABLISHED
2182 state_ = StateEnum::ESTABLISHED;
2184 // If SHUT_WRITE_PENDING is set and we don't have any write requests to
2185 // perform, immediately shutdown the write half of the socket.
2186 if ((shutdownFlags_ & SHUT_WRITE_PENDING) && writeReqHead_ == nullptr) {
2187 // SHUT_READ shouldn't be set. If close() is called on the socket while we
2188 // are still connecting we just abort the connect rather than waiting for
2190 assert((shutdownFlags_ & SHUT_READ) == 0);
2191 shutdown(fd_, SHUT_WR);
2192 shutdownFlags_ |= SHUT_WRITE;
2195 VLOG(7) << "AsyncSocket " << this << ": fd " << fd_
2196 << "successfully connected; state=" << state_;
2198 // Remember the EventBase we are attached to, before we start invoking any
2199 // callbacks (since the callbacks may call detachEventBase()).
2200 EventBase* originalEventBase = eventBase_;
2202 invokeConnectSuccess();
2203 // Note that the connect callback may have changed our state.
2204 // (set or unset the read callback, called write(), closed the socket, etc.)
2205 // The following code needs to handle these situations correctly.
2207 // If the socket has been closed, readCallback_ and writeReqHead_ will
2208 // always be nullptr, so that will prevent us from trying to read or write.
2210 // The main thing to check for is if eventBase_ is still originalEventBase.
2211 // If not, we have been detached from this event base, so we shouldn't
2212 // perform any more operations.
2213 if (eventBase_ != originalEventBase) {
2217 handleInitialReadWrite();
2220 void AsyncSocket::timeoutExpired() noexcept {
2221 VLOG(7) << "AsyncSocket " << this << ", fd " << fd_ << ": timeout expired: "
2222 << "state=" << state_ << ", events=" << std::hex << eventFlags_;
2223 DestructorGuard dg(this);
2224 eventBase_->dcheckIsInEventBaseThread();
2226 if (state_ == StateEnum::CONNECTING) {
2227 // connect() timed out
2228 // Unregister for I/O events.
2229 if (connectCallback_) {
2230 AsyncSocketException ex(
2231 AsyncSocketException::TIMED_OUT,
2233 "connect timed out after {}ms", connectTimeout_.count()));
2234 failConnect(__func__, ex);
2236 // we faced a connect error without a connect callback, which could
2237 // happen due to TFO.
2238 AsyncSocketException ex(
2239 AsyncSocketException::TIMED_OUT, "write timed out during connection");
2240 failWrite(__func__, ex);
2243 // a normal write operation timed out
2244 AsyncSocketException ex(
2245 AsyncSocketException::TIMED_OUT,
2246 folly::sformat("write timed out after {}ms", sendTimeout_));
2247 failWrite(__func__, ex);
2251 ssize_t AsyncSocket::tfoSendMsg(int fd, struct msghdr* msg, int msg_flags) {
2252 return detail::tfo_sendmsg(fd, msg, msg_flags);
2255 AsyncSocket::WriteResult
2256 AsyncSocket::sendSocketMessage(int fd, struct msghdr* msg, int msg_flags) {
2257 ssize_t totalWritten = 0;
2258 if (state_ == StateEnum::FAST_OPEN) {
2259 sockaddr_storage addr;
2260 auto len = addr_.getAddress(&addr);
2261 msg->msg_name = &addr;
2262 msg->msg_namelen = len;
2263 totalWritten = tfoSendMsg(fd_, msg, msg_flags);
2264 if (totalWritten >= 0) {
2265 tfoFinished_ = true;
2266 state_ = StateEnum::ESTABLISHED;
2267 // We schedule this asynchrously so that we don't end up
2268 // invoking initial read or write while a write is in progress.
2269 scheduleInitialReadWrite();
2270 } else if (errno == EINPROGRESS) {
2271 VLOG(4) << "TFO falling back to connecting";
2272 // A normal sendmsg doesn't return EINPROGRESS, however
2273 // TFO might fallback to connecting if there is no
2275 state_ = StateEnum::CONNECTING;
2277 scheduleConnectTimeout();
2278 registerForConnectEvents();
2279 } catch (const AsyncSocketException& ex) {
2281 WRITE_ERROR, std::make_unique<AsyncSocketException>(ex));
2283 // Let's fake it that no bytes were written and return an errno.
2286 } else if (errno == EOPNOTSUPP) {
2287 // Try falling back to connecting.
2288 VLOG(4) << "TFO not supported";
2289 state_ = StateEnum::CONNECTING;
2291 int ret = socketConnect((const sockaddr*)&addr, len);
2293 // connect succeeded immediately
2294 // Treat this like no data was written.
2295 state_ = StateEnum::ESTABLISHED;
2296 scheduleInitialReadWrite();
2298 // If there was no exception during connections,
2299 // we would return that no bytes were written.
2302 } catch (const AsyncSocketException& ex) {
2304 WRITE_ERROR, std::make_unique<AsyncSocketException>(ex));
2306 } else if (errno == EAGAIN) {
2307 // Normally sendmsg would indicate that the write would block.
2308 // However in the fast open case, it would indicate that sendmsg
2309 // fell back to a connect. This is a return code from connect()
2310 // instead, and is an error condition indicating no fds available.
2313 std::make_unique<AsyncSocketException>(
2314 AsyncSocketException::UNKNOWN, "No more free local ports"));
2317 totalWritten = ::sendmsg(fd, msg, msg_flags);
2319 return WriteResult(totalWritten);
2322 AsyncSocket::WriteResult AsyncSocket::performWrite(
2326 uint32_t* countWritten,
2327 uint32_t* partialWritten) {
2328 // We use sendmsg() instead of writev() so that we can pass in MSG_NOSIGNAL
2329 // We correctly handle EPIPE errors, so we never want to receive SIGPIPE
2330 // (since it may terminate the program if the main program doesn't explicitly
2333 msg.msg_name = nullptr;
2334 msg.msg_namelen = 0;
2335 msg.msg_iov = const_cast<iovec *>(vec);
2336 msg.msg_iovlen = std::min<size_t>(count, kIovMax);
2338 msg.msg_controllen = sendMsgParamCallback_->getAncillaryDataSize(flags);
2339 CHECK_GE(AsyncSocket::SendMsgParamsCallback::maxAncillaryDataSize,
2340 msg.msg_controllen);
2342 if (msg.msg_controllen != 0) {
2343 msg.msg_control = reinterpret_cast<char*>(alloca(msg.msg_controllen));
2344 sendMsgParamCallback_->getAncillaryData(flags, msg.msg_control);
2346 msg.msg_control = nullptr;
2348 int msg_flags = sendMsgParamCallback_->getFlags(flags, zeroCopyEnabled_);
2350 auto writeResult = sendSocketMessage(fd_, &msg, msg_flags);
2351 auto totalWritten = writeResult.writeReturn;
2352 if (totalWritten < 0) {
2353 bool tryAgain = (errno == EAGAIN);
2355 // Apple has a bug where doing a second write on a socket which we
2356 // have opened with TFO causes an ENOTCONN to be thrown. However the
2357 // socket is really connected, so treat ENOTCONN as a EAGAIN until
2358 // this bug is fixed.
2359 tryAgain |= (errno == ENOTCONN);
2361 if (!writeResult.exception && tryAgain) {
2362 // TCP buffer is full; we can't write any more data right now.
2364 *partialWritten = 0;
2365 return WriteResult(0);
2369 *partialWritten = 0;
2373 appBytesWritten_ += totalWritten;
2375 uint32_t bytesWritten;
2377 for (bytesWritten = uint32_t(totalWritten), n = 0; n < count; ++n) {
2378 const iovec* v = vec + n;
2379 if (v->iov_len > bytesWritten) {
2380 // Partial write finished in the middle of this iovec
2382 *partialWritten = bytesWritten;
2383 return WriteResult(totalWritten);
2386 bytesWritten -= uint32_t(v->iov_len);
2389 assert(bytesWritten == 0);
2391 *partialWritten = 0;
2392 return WriteResult(totalWritten);
2396 * Re-register the EventHandler after eventFlags_ has changed.
2398 * If an error occurs, fail() is called to move the socket into the error state
2399 * and call all currently installed callbacks. After an error, the
2400 * AsyncSocket is completely unregistered.
2402 * @return Returns true on success, or false on error.
2404 bool AsyncSocket::updateEventRegistration() {
2405 VLOG(5) << "AsyncSocket::updateEventRegistration(this=" << this
2406 << ", fd=" << fd_ << ", evb=" << eventBase_ << ", state=" << state_
2407 << ", events=" << std::hex << eventFlags_;
2408 eventBase_->dcheckIsInEventBaseThread();
2409 if (eventFlags_ == EventHandler::NONE) {
2410 ioHandler_.unregisterHandler();
2414 // Always register for persistent events, so we don't have to re-register
2415 // after being called back.
2416 if (!ioHandler_.registerHandler(
2417 uint16_t(eventFlags_ | EventHandler::PERSIST))) {
2418 eventFlags_ = EventHandler::NONE; // we're not registered after error
2419 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
2420 withAddr("failed to update AsyncSocket event registration"));
2421 fail("updateEventRegistration", ex);
2428 bool AsyncSocket::updateEventRegistration(uint16_t enable,
2430 uint16_t oldFlags = eventFlags_;
2431 eventFlags_ |= enable;
2432 eventFlags_ &= ~disable;
2433 if (eventFlags_ == oldFlags) {
2436 return updateEventRegistration();
2440 void AsyncSocket::startFail() {
2441 // startFail() should only be called once
2442 assert(state_ != StateEnum::ERROR);
2443 assert(getDestructorGuardCount() > 0);
2444 state_ = StateEnum::ERROR;
2445 // Ensure that SHUT_READ and SHUT_WRITE are set,
2446 // so all future attempts to read or write will be rejected
2447 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE);
2449 if (eventFlags_ != EventHandler::NONE) {
2450 eventFlags_ = EventHandler::NONE;
2451 ioHandler_.unregisterHandler();
2453 writeTimeout_.cancelTimeout();
2456 ioHandler_.changeHandlerFD(-1);
2461 void AsyncSocket::invokeAllErrors(const AsyncSocketException& ex) {
2462 invokeConnectErr(ex);
2465 if (readCallback_) {
2466 ReadCallback* callback = readCallback_;
2467 readCallback_ = nullptr;
2468 callback->readErr(ex);
2472 void AsyncSocket::finishFail() {
2473 assert(state_ == StateEnum::ERROR);
2474 assert(getDestructorGuardCount() > 0);
2476 AsyncSocketException ex(
2477 AsyncSocketException::INTERNAL_ERROR,
2478 withAddr("socket closing after error"));
2479 invokeAllErrors(ex);
2482 void AsyncSocket::finishFail(const AsyncSocketException& ex) {
2483 assert(state_ == StateEnum::ERROR);
2484 assert(getDestructorGuardCount() > 0);
2485 invokeAllErrors(ex);
2488 void AsyncSocket::fail(const char* fn, const AsyncSocketException& ex) {
2489 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2490 << state_ << " host=" << addr_.describe()
2491 << "): failed in " << fn << "(): "
2497 void AsyncSocket::failConnect(const char* fn, const AsyncSocketException& ex) {
2498 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2499 << state_ << " host=" << addr_.describe()
2500 << "): failed while connecting in " << fn << "(): "
2504 invokeConnectErr(ex);
2508 void AsyncSocket::failRead(const char* fn, const AsyncSocketException& ex) {
2509 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2510 << state_ << " host=" << addr_.describe()
2511 << "): failed while reading in " << fn << "(): "
2515 if (readCallback_ != nullptr) {
2516 ReadCallback* callback = readCallback_;
2517 readCallback_ = nullptr;
2518 callback->readErr(ex);
2524 void AsyncSocket::failErrMessageRead(const char* fn,
2525 const AsyncSocketException& ex) {
2526 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2527 << state_ << " host=" << addr_.describe()
2528 << "): failed while reading message in " << fn << "(): "
2532 if (errMessageCallback_ != nullptr) {
2533 ErrMessageCallback* callback = errMessageCallback_;
2534 errMessageCallback_ = nullptr;
2535 callback->errMessageError(ex);
2541 void AsyncSocket::failWrite(const char* fn, const AsyncSocketException& ex) {
2542 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2543 << state_ << " host=" << addr_.describe()
2544 << "): failed while writing in " << fn << "(): "
2548 // Only invoke the first write callback, since the error occurred while
2549 // writing this request. Let any other pending write callbacks be invoked in
2551 if (writeReqHead_ != nullptr) {
2552 WriteRequest* req = writeReqHead_;
2553 writeReqHead_ = req->getNext();
2554 WriteCallback* callback = req->getCallback();
2555 uint32_t bytesWritten = req->getTotalBytesWritten();
2558 callback->writeErr(bytesWritten, ex);
2565 void AsyncSocket::failWrite(const char* fn, WriteCallback* callback,
2566 size_t bytesWritten,
2567 const AsyncSocketException& ex) {
2568 // This version of failWrite() is used when the failure occurs before
2569 // we've added the callback to writeReqHead_.
2570 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2571 << state_ << " host=" << addr_.describe()
2572 <<"): failed while writing in " << fn << "(): "
2576 if (callback != nullptr) {
2577 callback->writeErr(bytesWritten, ex);
2583 void AsyncSocket::failAllWrites(const AsyncSocketException& ex) {
2584 // Invoke writeError() on all write callbacks.
2585 // This is used when writes are forcibly shutdown with write requests
2586 // pending, or when an error occurs with writes pending.
2587 while (writeReqHead_ != nullptr) {
2588 WriteRequest* req = writeReqHead_;
2589 writeReqHead_ = req->getNext();
2590 WriteCallback* callback = req->getCallback();
2592 callback->writeErr(req->getTotalBytesWritten(), ex);
2598 void AsyncSocket::invalidState(ConnectCallback* callback) {
2599 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_
2600 << "): connect() called in invalid state " << state_;
2603 * The invalidState() methods don't use the normal failure mechanisms,
2604 * since we don't know what state we are in. We don't want to call
2605 * startFail()/finishFail() recursively if we are already in the middle of
2609 AsyncSocketException ex(AsyncSocketException::ALREADY_OPEN,
2610 "connect() called with socket in invalid state");
2611 connectEndTime_ = std::chrono::steady_clock::now();
2612 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2614 callback->connectErr(ex);
2617 // We can't use failConnect() here since connectCallback_
2618 // may already be set to another callback. Invoke this ConnectCallback
2619 // here; any other connectCallback_ will be invoked in finishFail()
2622 callback->connectErr(ex);
2628 void AsyncSocket::invalidState(ErrMessageCallback* callback) {
2629 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_
2630 << "): setErrMessageCB(" << callback
2631 << ") called in invalid state " << state_;
2633 AsyncSocketException ex(
2634 AsyncSocketException::NOT_OPEN,
2635 msgErrQueueSupported
2636 ? "setErrMessageCB() called with socket in invalid state"
2637 : "This platform does not support socket error message notifications");
2638 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2640 callback->errMessageError(ex);
2645 callback->errMessageError(ex);
2651 void AsyncSocket::invokeConnectErr(const AsyncSocketException& ex) {
2652 connectEndTime_ = std::chrono::steady_clock::now();
2653 if (connectCallback_) {
2654 ConnectCallback* callback = connectCallback_;
2655 connectCallback_ = nullptr;
2656 callback->connectErr(ex);
2660 void AsyncSocket::invokeConnectSuccess() {
2661 connectEndTime_ = std::chrono::steady_clock::now();
2662 if (connectCallback_) {
2663 ConnectCallback* callback = connectCallback_;
2664 connectCallback_ = nullptr;
2665 callback->connectSuccess();
2669 void AsyncSocket::invalidState(ReadCallback* callback) {
2670 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_
2671 << "): setReadCallback(" << callback
2672 << ") called in invalid state " << state_;
2674 AsyncSocketException ex(AsyncSocketException::NOT_OPEN,
2675 "setReadCallback() called with socket in "
2677 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2679 callback->readErr(ex);
2684 callback->readErr(ex);
2690 void AsyncSocket::invalidState(WriteCallback* callback) {
2691 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_
2692 << "): write() called in invalid state " << state_;
2694 AsyncSocketException ex(AsyncSocketException::NOT_OPEN,
2695 withAddr("write() called with socket in invalid state"));
2696 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2698 callback->writeErr(0, ex);
2703 callback->writeErr(0, ex);
2709 void AsyncSocket::doClose() {
2713 if (const auto shutdownSocketSet = wShutdownSocketSet_.lock()) {
2714 shutdownSocketSet->close(fd_);
2721 std::ostream& operator << (std::ostream& os,
2722 const AsyncSocket::StateEnum& state) {
2723 os << static_cast<int>(state);
2727 std::string AsyncSocket::withAddr(const std::string& s) {
2728 // Don't use addr_ directly because it may not be initialized
2729 // e.g. if constructed from fd
2730 folly::SocketAddress peer, local;
2732 getPeerAddress(&peer);
2733 getLocalAddress(&local);
2734 } catch (const std::exception&) {
2739 return s + " (peer=" + peer.describe() + ", local=" + local.describe() + ")";
2742 void AsyncSocket::setBufferCallback(BufferCallback* cb) {
2743 bufferCallback_ = cb;
2746 } // namespace folly