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(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 // In the latest stable kernel 4.14.3 as of 2017-12-04, unix domain
666 // socket does not support MSG_ERRQUEUE. So recvmsg(MSG_ERRQUEUE)
667 // will read application data from unix doamin socket as error
668 // message, which breaks the message flow in application. Feel free
669 // to remove the next code block if MSG_ERRQUEUE is added for unix
670 // domain socket in the future.
671 if (callback != nullptr) {
673 if (localAddr_.getFamily() == AF_UNIX) {
674 LOG(ERROR) << "Failed to set ErrMessageCallback=" << callback
675 << " for Unix Doamin Socket where MSG_ERRQUEUE is unsupported,"
681 // Short circuit if callback is the same as the existing errMessageCallback_.
682 if (callback == errMessageCallback_) {
686 if (!msgErrQueueSupported) {
687 // Per-socket error message queue is not supported on this platform.
688 return invalidState(callback);
691 DestructorGuard dg(this);
692 eventBase_->dcheckIsInEventBaseThread();
694 if (callback == nullptr) {
695 // We should be able to reset the callback regardless of the
696 // socket state. It's important to have a reliable callback
697 // cancellation mechanism.
698 errMessageCallback_ = callback;
702 switch ((StateEnum)state_) {
703 case StateEnum::CONNECTING:
704 case StateEnum::FAST_OPEN:
705 case StateEnum::ESTABLISHED: {
706 errMessageCallback_ = callback;
709 case StateEnum::CLOSED:
710 case StateEnum::ERROR:
711 // We should never reach here. SHUT_READ should always be set
712 // if we are in STATE_CLOSED or STATE_ERROR.
714 return invalidState(callback);
715 case StateEnum::UNINIT:
716 // We do not allow setReadCallback() to be called before we start
718 return invalidState(callback);
721 // We don't put a default case in the switch statement, so that the compiler
722 // will warn us to update the switch statement if a new state is added.
723 return invalidState(callback);
726 AsyncSocket::ErrMessageCallback* AsyncSocket::getErrMessageCallback() const {
727 return errMessageCallback_;
730 void AsyncSocket::setSendMsgParamCB(SendMsgParamsCallback* callback) {
731 sendMsgParamCallback_ = callback;
734 AsyncSocket::SendMsgParamsCallback* AsyncSocket::getSendMsgParamsCB() const {
735 return sendMsgParamCallback_;
738 void AsyncSocket::setReadCB(ReadCallback *callback) {
739 VLOG(6) << "AsyncSocket::setReadCallback() this=" << this << ", fd=" << fd_
740 << ", callback=" << callback << ", state=" << state_;
742 // Short circuit if callback is the same as the existing readCallback_.
744 // Note that this is needed for proper functioning during some cleanup cases.
745 // During cleanup we allow setReadCallback(nullptr) to be called even if the
746 // read callback is already unset and we have been detached from an event
747 // base. This check prevents us from asserting
748 // eventBase_->isInEventBaseThread() when eventBase_ is nullptr.
749 if (callback == readCallback_) {
753 /* We are removing a read callback */
754 if (callback == nullptr &&
755 immediateReadHandler_.isLoopCallbackScheduled()) {
756 immediateReadHandler_.cancelLoopCallback();
759 if (shutdownFlags_ & SHUT_READ) {
760 // Reads have already been shut down on this socket.
762 // Allow setReadCallback(nullptr) to be called in this case, but don't
763 // allow a new callback to be set.
765 // For example, setReadCallback(nullptr) can happen after an error if we
766 // invoke some other error callback before invoking readError(). The other
767 // error callback that is invoked first may go ahead and clear the read
768 // callback before we get a chance to invoke readError().
769 if (callback != nullptr) {
770 return invalidState(callback);
772 assert((eventFlags_ & EventHandler::READ) == 0);
773 readCallback_ = nullptr;
777 DestructorGuard dg(this);
778 eventBase_->dcheckIsInEventBaseThread();
780 switch ((StateEnum)state_) {
781 case StateEnum::CONNECTING:
782 case StateEnum::FAST_OPEN:
783 // For convenience, we allow the read callback to be set while we are
784 // still connecting. We just store the callback for now. Once the
785 // connection completes we'll register for read events.
786 readCallback_ = callback;
788 case StateEnum::ESTABLISHED:
790 readCallback_ = callback;
791 uint16_t oldFlags = eventFlags_;
793 eventFlags_ |= EventHandler::READ;
795 eventFlags_ &= ~EventHandler::READ;
798 // Update our registration if our flags have changed
799 if (eventFlags_ != oldFlags) {
800 // We intentionally ignore the return value here.
801 // updateEventRegistration() will move us into the error state if it
802 // fails, and we don't need to do anything else here afterwards.
803 (void)updateEventRegistration();
807 checkForImmediateRead();
811 case StateEnum::CLOSED:
812 case StateEnum::ERROR:
813 // We should never reach here. SHUT_READ should always be set
814 // if we are in STATE_CLOSED or STATE_ERROR.
816 return invalidState(callback);
817 case StateEnum::UNINIT:
818 // We do not allow setReadCallback() to be called before we start
820 return invalidState(callback);
823 // We don't put a default case in the switch statement, so that the compiler
824 // will warn us to update the switch statement if a new state is added.
825 return invalidState(callback);
828 AsyncSocket::ReadCallback* AsyncSocket::getReadCallback() const {
829 return readCallback_;
832 bool AsyncSocket::setZeroCopy(bool enable) {
833 if (msgErrQueueSupported) {
834 zeroCopyVal_ = enable;
840 int val = enable ? 1 : 0;
841 int ret = setsockopt(fd_, SOL_SOCKET, SO_ZEROCOPY, &val, sizeof(val));
843 // if enable == false, set zeroCopyEnabled_ = false regardless
844 // if SO_ZEROCOPY is set or not
846 zeroCopyEnabled_ = enable;
850 /* if the setsockopt failed, try to see if the socket inherited the flag
851 * since we cannot set SO_ZEROCOPY on a socket s = accept
855 socklen_t optlen = sizeof(val);
856 ret = getsockopt(fd_, SOL_SOCKET, SO_ZEROCOPY, &val, &optlen);
859 enable = val ? true : false;
864 zeroCopyEnabled_ = enable;
873 bool AsyncSocket::isZeroCopyRequest(WriteFlags flags) {
874 return (zeroCopyEnabled_ && isSet(flags, WriteFlags::WRITE_MSG_ZEROCOPY));
877 void AsyncSocket::adjustZeroCopyFlags(folly::WriteFlags& flags) {
878 if (!zeroCopyEnabled_) {
879 flags = unSet(flags, folly::WriteFlags::WRITE_MSG_ZEROCOPY);
883 void AsyncSocket::addZeroCopyBuf(std::unique_ptr<folly::IOBuf>&& buf) {
884 uint32_t id = getNextZeroCopyBufId();
885 folly::IOBuf* ptr = buf.get();
887 idZeroCopyBufPtrMap_[id] = ptr;
888 auto& p = idZeroCopyBufInfoMap_[ptr];
890 CHECK(p.buf_.get() == nullptr);
891 p.buf_ = std::move(buf);
894 void AsyncSocket::addZeroCopyBuf(folly::IOBuf* ptr) {
895 uint32_t id = getNextZeroCopyBufId();
896 idZeroCopyBufPtrMap_[id] = ptr;
898 idZeroCopyBufInfoMap_[ptr].count_++;
901 void AsyncSocket::releaseZeroCopyBuf(uint32_t id) {
902 auto iter = idZeroCopyBufPtrMap_.find(id);
903 CHECK(iter != idZeroCopyBufPtrMap_.end());
904 auto ptr = iter->second;
905 auto iter1 = idZeroCopyBufInfoMap_.find(ptr);
906 CHECK(iter1 != idZeroCopyBufInfoMap_.end());
907 if (0 == --iter1->second.count_) {
908 idZeroCopyBufInfoMap_.erase(iter1);
911 idZeroCopyBufPtrMap_.erase(iter);
914 void AsyncSocket::setZeroCopyBuf(std::unique_ptr<folly::IOBuf>&& buf) {
915 folly::IOBuf* ptr = buf.get();
916 auto& p = idZeroCopyBufInfoMap_[ptr];
917 CHECK(p.buf_.get() == nullptr);
919 p.buf_ = std::move(buf);
922 bool AsyncSocket::containsZeroCopyBuf(folly::IOBuf* ptr) {
923 return (idZeroCopyBufInfoMap_.find(ptr) != idZeroCopyBufInfoMap_.end());
926 bool AsyncSocket::isZeroCopyMsg(const cmsghdr& cmsg) const {
927 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
928 if (zeroCopyEnabled_ &&
929 ((cmsg.cmsg_level == SOL_IP && cmsg.cmsg_type == IP_RECVERR) ||
930 (cmsg.cmsg_level == SOL_IPV6 && cmsg.cmsg_type == IPV6_RECVERR))) {
931 const struct sock_extended_err* serr =
932 reinterpret_cast<const struct sock_extended_err*>(CMSG_DATA(&cmsg));
934 (serr->ee_errno == 0) && (serr->ee_origin == SO_EE_ORIGIN_ZEROCOPY));
940 void AsyncSocket::processZeroCopyMsg(const cmsghdr& cmsg) {
941 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
942 const struct sock_extended_err* serr =
943 reinterpret_cast<const struct sock_extended_err*>(CMSG_DATA(&cmsg));
944 uint32_t hi = serr->ee_data;
945 uint32_t lo = serr->ee_info;
946 // disable zero copy if the buffer was actually copied
947 if ((serr->ee_code & SO_EE_CODE_ZEROCOPY_COPIED) && zeroCopyEnabled_) {
948 VLOG(2) << "AsyncSocket::processZeroCopyMsg(): setting "
949 << "zeroCopyEnabled_ = false due to SO_EE_CODE_ZEROCOPY_COPIED "
951 zeroCopyEnabled_ = false;
954 for (uint32_t i = lo; i <= hi; i++) {
955 releaseZeroCopyBuf(i);
960 void AsyncSocket::write(WriteCallback* callback,
961 const void* buf, size_t bytes, WriteFlags flags) {
963 op.iov_base = const_cast<void*>(buf);
965 writeImpl(callback, &op, 1, unique_ptr<IOBuf>(), flags);
968 void AsyncSocket::writev(WriteCallback* callback,
972 writeImpl(callback, vec, count, unique_ptr<IOBuf>(), flags);
975 void AsyncSocket::writeChain(WriteCallback* callback, unique_ptr<IOBuf>&& buf,
977 adjustZeroCopyFlags(flags);
979 constexpr size_t kSmallSizeMax = 64;
980 size_t count = buf->countChainElements();
981 if (count <= kSmallSizeMax) {
982 // suppress "warning: variable length array 'vec' is used [-Wvla]"
984 FOLLY_GCC_DISABLE_WARNING("-Wvla")
985 iovec vec[BOOST_PP_IF(FOLLY_HAVE_VLA, count, kSmallSizeMax)];
988 writeChainImpl(callback, vec, count, std::move(buf), flags);
990 iovec* vec = new iovec[count];
991 writeChainImpl(callback, vec, count, std::move(buf), flags);
996 void AsyncSocket::writeChainImpl(WriteCallback* callback, iovec* vec,
997 size_t count, unique_ptr<IOBuf>&& buf, WriteFlags flags) {
998 size_t veclen = buf->fillIov(vec, count);
999 writeImpl(callback, vec, veclen, std::move(buf), flags);
1002 void AsyncSocket::writeImpl(WriteCallback* callback, const iovec* vec,
1003 size_t count, unique_ptr<IOBuf>&& buf,
1005 VLOG(6) << "AsyncSocket::writev() this=" << this << ", fd=" << fd_
1006 << ", callback=" << callback << ", count=" << count
1007 << ", state=" << state_;
1008 DestructorGuard dg(this);
1009 unique_ptr<IOBuf>ioBuf(std::move(buf));
1010 eventBase_->dcheckIsInEventBaseThread();
1012 if (shutdownFlags_ & (SHUT_WRITE | SHUT_WRITE_PENDING)) {
1013 // No new writes may be performed after the write side of the socket has
1016 // We could just call callback->writeError() here to fail just this write.
1017 // However, fail hard and use invalidState() to fail all outstanding
1018 // callbacks and move the socket into the error state. There's most likely
1019 // a bug in the caller's code, so we abort everything rather than trying to
1020 // proceed as best we can.
1021 return invalidState(callback);
1024 uint32_t countWritten = 0;
1025 uint32_t partialWritten = 0;
1026 ssize_t bytesWritten = 0;
1027 bool mustRegister = false;
1028 if ((state_ == StateEnum::ESTABLISHED || state_ == StateEnum::FAST_OPEN) &&
1030 if (writeReqHead_ == nullptr) {
1031 // If we are established and there are no other writes pending,
1032 // we can attempt to perform the write immediately.
1033 assert(writeReqTail_ == nullptr);
1034 assert((eventFlags_ & EventHandler::WRITE) == 0);
1036 auto writeResult = performWrite(
1037 vec, uint32_t(count), flags, &countWritten, &partialWritten);
1038 bytesWritten = writeResult.writeReturn;
1039 if (bytesWritten < 0) {
1040 auto errnoCopy = errno;
1041 if (writeResult.exception) {
1042 return failWrite(__func__, callback, 0, *writeResult.exception);
1044 AsyncSocketException ex(
1045 AsyncSocketException::INTERNAL_ERROR,
1046 withAddr("writev failed"),
1048 return failWrite(__func__, callback, 0, ex);
1049 } else if (countWritten == count) {
1050 // done, add the whole buffer
1051 if (countWritten && isZeroCopyRequest(flags)) {
1052 addZeroCopyBuf(std::move(ioBuf));
1054 // We successfully wrote everything.
1055 // Invoke the callback and return.
1057 callback->writeSuccess();
1060 } else { // continue writing the next writeReq
1062 if (bytesWritten && isZeroCopyRequest(flags)) {
1063 addZeroCopyBuf(ioBuf.get());
1065 if (bufferCallback_) {
1066 bufferCallback_->onEgressBuffered();
1069 if (!connecting()) {
1070 // Writes might put the socket back into connecting state
1071 // if TFO is enabled, and using TFO fails.
1072 // This means that write timeouts would not be active, however
1073 // connect timeouts would affect this stage.
1074 mustRegister = true;
1077 } else if (!connecting()) {
1078 // Invalid state for writing
1079 return invalidState(callback);
1082 // Create a new WriteRequest to add to the queue
1085 req = BytesWriteRequest::newRequest(
1089 uint32_t(count - countWritten),
1091 uint32_t(bytesWritten),
1094 } catch (const std::exception& ex) {
1095 // we mainly expect to catch std::bad_alloc here
1096 AsyncSocketException tex(AsyncSocketException::INTERNAL_ERROR,
1097 withAddr(string("failed to append new WriteRequest: ") + ex.what()));
1098 return failWrite(__func__, callback, size_t(bytesWritten), tex);
1101 if (writeReqTail_ == nullptr) {
1102 assert(writeReqHead_ == nullptr);
1103 writeReqHead_ = writeReqTail_ = req;
1105 writeReqTail_->append(req);
1106 writeReqTail_ = req;
1109 // Register for write events if are established and not currently
1110 // waiting on write events
1112 assert(state_ == StateEnum::ESTABLISHED);
1113 assert((eventFlags_ & EventHandler::WRITE) == 0);
1114 if (!updateEventRegistration(EventHandler::WRITE, 0)) {
1115 assert(state_ == StateEnum::ERROR);
1118 if (sendTimeout_ > 0) {
1119 // Schedule a timeout to fire if the write takes too long.
1120 if (!writeTimeout_.scheduleTimeout(sendTimeout_)) {
1121 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
1122 withAddr("failed to schedule send timeout"));
1123 return failWrite(__func__, ex);
1129 void AsyncSocket::writeRequest(WriteRequest* req) {
1130 if (writeReqTail_ == nullptr) {
1131 assert(writeReqHead_ == nullptr);
1132 writeReqHead_ = writeReqTail_ = req;
1135 writeReqTail_->append(req);
1136 writeReqTail_ = req;
1140 void AsyncSocket::close() {
1141 VLOG(5) << "AsyncSocket::close(): this=" << this << ", fd_=" << fd_
1142 << ", state=" << state_ << ", shutdownFlags="
1143 << std::hex << (int) shutdownFlags_;
1145 // close() is only different from closeNow() when there are pending writes
1146 // that need to drain before we can close. In all other cases, just call
1149 // Note that writeReqHead_ can be non-nullptr even in STATE_CLOSED or
1150 // STATE_ERROR if close() is invoked while a previous closeNow() or failure
1151 // is still running. (e.g., If there are multiple pending writes, and we
1152 // call writeError() on the first one, it may call close(). In this case we
1153 // will already be in STATE_CLOSED or STATE_ERROR, but the remaining pending
1154 // writes will still be in the queue.)
1156 // We only need to drain pending writes if we are still in STATE_CONNECTING
1157 // or STATE_ESTABLISHED
1158 if ((writeReqHead_ == nullptr) ||
1159 !(state_ == StateEnum::CONNECTING ||
1160 state_ == StateEnum::ESTABLISHED)) {
1165 // Declare a DestructorGuard to ensure that the AsyncSocket cannot be
1166 // destroyed until close() returns.
1167 DestructorGuard dg(this);
1168 eventBase_->dcheckIsInEventBaseThread();
1170 // Since there are write requests pending, we have to set the
1171 // SHUT_WRITE_PENDING flag, and wait to perform the real close until the
1172 // connect finishes and we finish writing these requests.
1174 // Set SHUT_READ to indicate that reads are shut down, and set the
1175 // SHUT_WRITE_PENDING flag to mark that we want to shutdown once the
1176 // pending writes complete.
1177 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE_PENDING);
1179 // If a read callback is set, invoke readEOF() immediately to inform it that
1180 // the socket has been closed and no more data can be read.
1181 if (readCallback_) {
1182 // Disable reads if they are enabled
1183 if (!updateEventRegistration(0, EventHandler::READ)) {
1184 // We're now in the error state; callbacks have been cleaned up
1185 assert(state_ == StateEnum::ERROR);
1186 assert(readCallback_ == nullptr);
1188 ReadCallback* callback = readCallback_;
1189 readCallback_ = nullptr;
1190 callback->readEOF();
1195 void AsyncSocket::closeNow() {
1196 VLOG(5) << "AsyncSocket::closeNow(): this=" << this << ", fd_=" << fd_
1197 << ", state=" << state_ << ", shutdownFlags="
1198 << std::hex << (int) shutdownFlags_;
1199 DestructorGuard dg(this);
1201 eventBase_->dcheckIsInEventBaseThread();
1205 case StateEnum::ESTABLISHED:
1206 case StateEnum::CONNECTING:
1207 case StateEnum::FAST_OPEN: {
1208 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE);
1209 state_ = StateEnum::CLOSED;
1211 // If the write timeout was set, cancel it.
1212 writeTimeout_.cancelTimeout();
1214 // If we are registered for I/O events, unregister.
1215 if (eventFlags_ != EventHandler::NONE) {
1216 eventFlags_ = EventHandler::NONE;
1217 if (!updateEventRegistration()) {
1218 // We will have been moved into the error state.
1219 assert(state_ == StateEnum::ERROR);
1224 if (immediateReadHandler_.isLoopCallbackScheduled()) {
1225 immediateReadHandler_.cancelLoopCallback();
1229 ioHandler_.changeHandlerFD(-1);
1233 invokeConnectErr(socketClosedLocallyEx);
1235 failAllWrites(socketClosedLocallyEx);
1237 if (readCallback_) {
1238 ReadCallback* callback = readCallback_;
1239 readCallback_ = nullptr;
1240 callback->readEOF();
1244 case StateEnum::CLOSED:
1245 // Do nothing. It's possible that we are being called recursively
1246 // from inside a callback that we invoked inside another call to close()
1247 // that is still running.
1249 case StateEnum::ERROR:
1250 // Do nothing. The error handling code has performed (or is performing)
1253 case StateEnum::UNINIT:
1254 assert(eventFlags_ == EventHandler::NONE);
1255 assert(connectCallback_ == nullptr);
1256 assert(readCallback_ == nullptr);
1257 assert(writeReqHead_ == nullptr);
1258 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE);
1259 state_ = StateEnum::CLOSED;
1263 LOG(DFATAL) << "AsyncSocket::closeNow() (this=" << this << ", fd=" << fd_
1264 << ") called in unknown state " << state_;
1267 void AsyncSocket::closeWithReset() {
1268 // Enable SO_LINGER, with the linger timeout set to 0.
1269 // This will trigger a TCP reset when we close the socket.
1271 struct linger optLinger = {1, 0};
1272 if (setSockOpt(SOL_SOCKET, SO_LINGER, &optLinger) != 0) {
1273 VLOG(2) << "AsyncSocket::closeWithReset(): error setting SO_LINGER "
1274 << "on " << fd_ << ": errno=" << errno;
1278 // Then let closeNow() take care of the rest
1282 void AsyncSocket::shutdownWrite() {
1283 VLOG(5) << "AsyncSocket::shutdownWrite(): this=" << this << ", fd=" << fd_
1284 << ", state=" << state_ << ", shutdownFlags="
1285 << std::hex << (int) shutdownFlags_;
1287 // If there are no pending writes, shutdownWrite() is identical to
1288 // shutdownWriteNow().
1289 if (writeReqHead_ == nullptr) {
1294 eventBase_->dcheckIsInEventBaseThread();
1296 // There are pending writes. Set SHUT_WRITE_PENDING so that the actual
1297 // shutdown will be performed once all writes complete.
1298 shutdownFlags_ |= SHUT_WRITE_PENDING;
1301 void AsyncSocket::shutdownWriteNow() {
1302 VLOG(5) << "AsyncSocket::shutdownWriteNow(): this=" << this
1303 << ", fd=" << fd_ << ", state=" << state_
1304 << ", shutdownFlags=" << std::hex << (int) shutdownFlags_;
1306 if (shutdownFlags_ & SHUT_WRITE) {
1307 // Writes are already shutdown; nothing else to do.
1311 // If SHUT_READ is already set, just call closeNow() to completely
1312 // close the socket. This can happen if close() was called with writes
1313 // pending, and then shutdownWriteNow() is called before all pending writes
1315 if (shutdownFlags_ & SHUT_READ) {
1320 DestructorGuard dg(this);
1322 eventBase_->dcheckIsInEventBaseThread();
1325 switch (static_cast<StateEnum>(state_)) {
1326 case StateEnum::ESTABLISHED:
1328 shutdownFlags_ |= SHUT_WRITE;
1330 // If the write timeout was set, cancel it.
1331 writeTimeout_.cancelTimeout();
1333 // If we are registered for write events, unregister.
1334 if (!updateEventRegistration(0, EventHandler::WRITE)) {
1335 // We will have been moved into the error state.
1336 assert(state_ == StateEnum::ERROR);
1340 // Shutdown writes on the file descriptor
1341 shutdown(fd_, SHUT_WR);
1343 // Immediately fail all write requests
1344 failAllWrites(socketShutdownForWritesEx);
1347 case StateEnum::CONNECTING:
1349 // Set the SHUT_WRITE_PENDING flag.
1350 // When the connection completes, it will check this flag,
1351 // shutdown the write half of the socket, and then set SHUT_WRITE.
1352 shutdownFlags_ |= SHUT_WRITE_PENDING;
1354 // Immediately fail all write requests
1355 failAllWrites(socketShutdownForWritesEx);
1358 case StateEnum::UNINIT:
1359 // Callers normally shouldn't call shutdownWriteNow() before the socket
1360 // even starts connecting. Nonetheless, go ahead and set
1361 // SHUT_WRITE_PENDING. Once the socket eventually connects it will
1362 // immediately shut down the write side of the socket.
1363 shutdownFlags_ |= SHUT_WRITE_PENDING;
1365 case StateEnum::FAST_OPEN:
1366 // In fast open state we haven't call connected yet, and if we shutdown
1367 // the writes, we will never try to call connect, so shut everything down
1368 shutdownFlags_ |= SHUT_WRITE;
1369 // Immediately fail all write requests
1370 failAllWrites(socketShutdownForWritesEx);
1372 case StateEnum::CLOSED:
1373 case StateEnum::ERROR:
1374 // We should never get here. SHUT_WRITE should always be set
1375 // in STATE_CLOSED and STATE_ERROR.
1376 VLOG(4) << "AsyncSocket::shutdownWriteNow() (this=" << this
1377 << ", fd=" << fd_ << ") in unexpected state " << state_
1378 << " with SHUT_WRITE not set ("
1379 << std::hex << (int) shutdownFlags_ << ")";
1384 LOG(DFATAL) << "AsyncSocket::shutdownWriteNow() (this=" << this << ", fd="
1385 << fd_ << ") called in unknown state " << state_;
1388 bool AsyncSocket::readable() const {
1392 struct pollfd fds[1];
1394 fds[0].events = POLLIN;
1396 int rc = poll(fds, 1, 0);
1400 bool AsyncSocket::writable() const {
1404 struct pollfd fds[1];
1406 fds[0].events = POLLOUT;
1408 int rc = poll(fds, 1, 0);
1412 bool AsyncSocket::isPending() const {
1413 return ioHandler_.isPending();
1416 bool AsyncSocket::hangup() const {
1418 // sanity check, no one should ask for hangup if we are not connected.
1422 #ifdef POLLRDHUP // Linux-only
1423 struct pollfd fds[1];
1425 fds[0].events = POLLRDHUP|POLLHUP;
1428 return (fds[0].revents & (POLLRDHUP|POLLHUP)) != 0;
1434 bool AsyncSocket::good() const {
1436 (state_ == StateEnum::CONNECTING || state_ == StateEnum::FAST_OPEN ||
1437 state_ == StateEnum::ESTABLISHED) &&
1438 (shutdownFlags_ == 0) && (eventBase_ != nullptr));
1441 bool AsyncSocket::error() const {
1442 return (state_ == StateEnum::ERROR);
1445 void AsyncSocket::attachEventBase(EventBase* eventBase) {
1446 VLOG(5) << "AsyncSocket::attachEventBase(this=" << this << ", fd=" << fd_
1447 << ", old evb=" << eventBase_ << ", new evb=" << eventBase
1448 << ", state=" << state_ << ", events="
1449 << std::hex << eventFlags_ << ")";
1450 assert(eventBase_ == nullptr);
1451 eventBase->dcheckIsInEventBaseThread();
1453 eventBase_ = eventBase;
1454 ioHandler_.attachEventBase(eventBase);
1456 updateEventRegistration();
1458 writeTimeout_.attachEventBase(eventBase);
1460 evbChangeCb_->evbAttached(this);
1464 void AsyncSocket::detachEventBase() {
1465 VLOG(5) << "AsyncSocket::detachEventBase(this=" << this << ", fd=" << fd_
1466 << ", old evb=" << eventBase_ << ", state=" << state_
1467 << ", events=" << std::hex << eventFlags_ << ")";
1468 assert(eventBase_ != nullptr);
1469 eventBase_->dcheckIsInEventBaseThread();
1471 eventBase_ = nullptr;
1473 ioHandler_.unregisterHandler();
1475 ioHandler_.detachEventBase();
1476 writeTimeout_.detachEventBase();
1478 evbChangeCb_->evbDetached(this);
1482 bool AsyncSocket::isDetachable() const {
1483 DCHECK(eventBase_ != nullptr);
1484 eventBase_->dcheckIsInEventBaseThread();
1486 return !writeTimeout_.isScheduled();
1489 void AsyncSocket::cacheAddresses() {
1492 cacheLocalAddress();
1494 } catch (const std::system_error& e) {
1495 if (e.code() != std::error_code(ENOTCONN, std::system_category())) {
1496 VLOG(1) << "Error caching addresses: " << e.code().value() << ", "
1497 << e.code().message();
1503 void AsyncSocket::cacheLocalAddress() const {
1504 if (!localAddr_.isInitialized()) {
1505 localAddr_.setFromLocalAddress(fd_);
1509 void AsyncSocket::cachePeerAddress() const {
1510 if (!addr_.isInitialized()) {
1511 addr_.setFromPeerAddress(fd_);
1515 bool AsyncSocket::isZeroCopyWriteInProgress() const noexcept {
1516 eventBase_->dcheckIsInEventBaseThread();
1517 return (!idZeroCopyBufPtrMap_.empty());
1520 void AsyncSocket::getLocalAddress(folly::SocketAddress* address) const {
1521 cacheLocalAddress();
1522 *address = localAddr_;
1525 void AsyncSocket::getPeerAddress(folly::SocketAddress* address) const {
1530 bool AsyncSocket::getTFOSucceded() const {
1531 return detail::tfo_succeeded(fd_);
1534 int AsyncSocket::setNoDelay(bool noDelay) {
1536 VLOG(4) << "AsyncSocket::setNoDelay() called on non-open socket "
1537 << this << "(state=" << state_ << ")";
1542 int value = noDelay ? 1 : 0;
1543 if (setsockopt(fd_, IPPROTO_TCP, TCP_NODELAY, &value, sizeof(value)) != 0) {
1544 int errnoCopy = errno;
1545 VLOG(2) << "failed to update TCP_NODELAY option on AsyncSocket "
1546 << this << " (fd=" << fd_ << ", state=" << state_ << "): "
1547 << strerror(errnoCopy);
1554 int AsyncSocket::setCongestionFlavor(const std::string &cname) {
1556 #ifndef TCP_CONGESTION
1557 #define TCP_CONGESTION 13
1561 VLOG(4) << "AsyncSocket::setCongestionFlavor() called on non-open "
1562 << "socket " << this << "(state=" << state_ << ")";
1572 socklen_t(cname.length() + 1)) != 0) {
1573 int errnoCopy = errno;
1574 VLOG(2) << "failed to update TCP_CONGESTION option on AsyncSocket "
1575 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1576 << strerror(errnoCopy);
1583 int AsyncSocket::setQuickAck(bool quickack) {
1586 VLOG(4) << "AsyncSocket::setQuickAck() called on non-open socket "
1587 << this << "(state=" << state_ << ")";
1592 #ifdef TCP_QUICKACK // Linux-only
1593 int value = quickack ? 1 : 0;
1594 if (setsockopt(fd_, IPPROTO_TCP, TCP_QUICKACK, &value, sizeof(value)) != 0) {
1595 int errnoCopy = errno;
1596 VLOG(2) << "failed to update TCP_QUICKACK option on AsyncSocket"
1597 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1598 << strerror(errnoCopy);
1608 int AsyncSocket::setSendBufSize(size_t bufsize) {
1610 VLOG(4) << "AsyncSocket::setSendBufSize() called on non-open socket "
1611 << this << "(state=" << state_ << ")";
1615 if (setsockopt(fd_, SOL_SOCKET, SO_SNDBUF, &bufsize, sizeof(bufsize)) !=0) {
1616 int errnoCopy = errno;
1617 VLOG(2) << "failed to update SO_SNDBUF option on AsyncSocket"
1618 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1619 << strerror(errnoCopy);
1626 int AsyncSocket::setRecvBufSize(size_t bufsize) {
1628 VLOG(4) << "AsyncSocket::setRecvBufSize() called on non-open socket "
1629 << this << "(state=" << state_ << ")";
1633 if (setsockopt(fd_, SOL_SOCKET, SO_RCVBUF, &bufsize, sizeof(bufsize)) !=0) {
1634 int errnoCopy = errno;
1635 VLOG(2) << "failed to update SO_RCVBUF option on AsyncSocket"
1636 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1637 << strerror(errnoCopy);
1644 int AsyncSocket::setTCPProfile(int profd) {
1646 VLOG(4) << "AsyncSocket::setTCPProfile() called on non-open socket "
1647 << this << "(state=" << state_ << ")";
1651 if (setsockopt(fd_, SOL_SOCKET, SO_SET_NAMESPACE, &profd, sizeof(int)) !=0) {
1652 int errnoCopy = errno;
1653 VLOG(2) << "failed to set socket namespace option on AsyncSocket"
1654 << this << "(fd=" << fd_ << ", state=" << state_ << "): "
1655 << strerror(errnoCopy);
1662 void AsyncSocket::ioReady(uint16_t events) noexcept {
1663 VLOG(7) << "AsyncSocket::ioRead() this=" << this << ", fd=" << fd_
1664 << ", events=" << std::hex << events << ", state=" << state_;
1665 DestructorGuard dg(this);
1666 assert(events & EventHandler::READ_WRITE);
1667 eventBase_->dcheckIsInEventBaseThread();
1669 uint16_t relevantEvents = uint16_t(events & EventHandler::READ_WRITE);
1670 EventBase* originalEventBase = eventBase_;
1671 // If we got there it means that either EventHandler::READ or
1672 // EventHandler::WRITE is set. Any of these flags can
1673 // indicate that there are messages available in the socket
1674 // error message queue.
1675 handleErrMessages();
1677 // Return now if handleErrMessages() detached us from our EventBase
1678 if (eventBase_ != originalEventBase) {
1682 if (relevantEvents == EventHandler::READ) {
1684 } else if (relevantEvents == EventHandler::WRITE) {
1686 } else if (relevantEvents == EventHandler::READ_WRITE) {
1687 // If both read and write events are ready, process writes first.
1690 // Return now if handleWrite() detached us from our EventBase
1691 if (eventBase_ != originalEventBase) {
1695 // Only call handleRead() if a read callback is still installed.
1696 // (It's possible that the read callback was uninstalled during
1698 if (readCallback_) {
1702 VLOG(4) << "AsyncSocket::ioRead() called with unexpected events "
1703 << std::hex << events << "(this=" << this << ")";
1708 AsyncSocket::ReadResult
1709 AsyncSocket::performRead(void** buf, size_t* buflen, size_t* /* offset */) {
1710 VLOG(5) << "AsyncSocket::performRead() this=" << this << ", buf=" << *buf
1711 << ", buflen=" << *buflen;
1713 if (preReceivedData_ && !preReceivedData_->empty()) {
1714 VLOG(5) << "AsyncSocket::performRead() this=" << this
1715 << ", reading pre-received data";
1717 io::Cursor cursor(preReceivedData_.get());
1718 auto len = cursor.pullAtMost(*buf, *buflen);
1721 queue.append(std::move(preReceivedData_));
1722 queue.trimStart(len);
1723 preReceivedData_ = queue.move();
1725 appBytesReceived_ += len;
1726 return ReadResult(len);
1729 ssize_t bytes = recv(fd_, *buf, *buflen, MSG_DONTWAIT);
1731 if (errno == EAGAIN || errno == EWOULDBLOCK) {
1732 // No more data to read right now.
1733 return ReadResult(READ_BLOCKING);
1735 return ReadResult(READ_ERROR);
1738 appBytesReceived_ += bytes;
1739 return ReadResult(bytes);
1743 void AsyncSocket::prepareReadBuffer(void** buf, size_t* buflen) {
1744 // no matter what, buffer should be preapared for non-ssl socket
1745 CHECK(readCallback_);
1746 readCallback_->getReadBuffer(buf, buflen);
1749 void AsyncSocket::handleErrMessages() noexcept {
1750 // This method has non-empty implementation only for platforms
1751 // supporting per-socket error queues.
1752 VLOG(5) << "AsyncSocket::handleErrMessages() this=" << this << ", fd=" << fd_
1753 << ", state=" << state_;
1754 if (errMessageCallback_ == nullptr && idZeroCopyBufPtrMap_.empty()) {
1755 VLOG(7) << "AsyncSocket::handleErrMessages(): "
1756 << "no callback installed - exiting.";
1760 #ifdef FOLLY_HAVE_MSG_ERRQUEUE
1766 entry.iov_base = &data;
1767 entry.iov_len = sizeof(data);
1768 msg.msg_iov = &entry;
1770 msg.msg_name = nullptr;
1771 msg.msg_namelen = 0;
1772 msg.msg_control = ctrl;
1773 msg.msg_controllen = sizeof(ctrl);
1778 ret = recvmsg(fd_, &msg, MSG_ERRQUEUE);
1779 VLOG(5) << "AsyncSocket::handleErrMessages(): recvmsg returned " << ret;
1782 if (errno != EAGAIN) {
1783 auto errnoCopy = errno;
1784 LOG(ERROR) << "::recvmsg exited with code " << ret
1785 << ", errno: " << errnoCopy;
1786 AsyncSocketException ex(
1787 AsyncSocketException::INTERNAL_ERROR,
1788 withAddr("recvmsg() failed"),
1790 failErrMessageRead(__func__, ex);
1795 for (struct cmsghdr* cmsg = CMSG_FIRSTHDR(&msg);
1796 cmsg != nullptr && cmsg->cmsg_len != 0;
1797 cmsg = CMSG_NXTHDR(&msg, cmsg)) {
1798 if (isZeroCopyMsg(*cmsg)) {
1799 processZeroCopyMsg(*cmsg);
1801 if (errMessageCallback_) {
1802 errMessageCallback_->errMessage(*cmsg);
1807 #endif // FOLLY_HAVE_MSG_ERRQUEUE
1810 bool AsyncSocket::processZeroCopyWriteInProgress() noexcept {
1811 eventBase_->dcheckIsInEventBaseThread();
1812 if (idZeroCopyBufPtrMap_.empty()) {
1816 handleErrMessages();
1818 return idZeroCopyBufPtrMap_.empty();
1821 void AsyncSocket::handleRead() noexcept {
1822 VLOG(5) << "AsyncSocket::handleRead() this=" << this << ", fd=" << fd_
1823 << ", state=" << state_;
1824 assert(state_ == StateEnum::ESTABLISHED);
1825 assert((shutdownFlags_ & SHUT_READ) == 0);
1826 assert(readCallback_ != nullptr);
1827 assert(eventFlags_ & EventHandler::READ);
1830 // - a read attempt would block
1831 // - readCallback_ is uninstalled
1832 // - the number of loop iterations exceeds the optional maximum
1833 // - this AsyncSocket is moved to another EventBase
1835 // When we invoke readDataAvailable() it may uninstall the readCallback_,
1836 // which is why need to check for it here.
1838 // The last bullet point is slightly subtle. readDataAvailable() may also
1839 // detach this socket from this EventBase. However, before
1840 // readDataAvailable() returns another thread may pick it up, attach it to
1841 // a different EventBase, and install another readCallback_. We need to
1842 // exit immediately after readDataAvailable() returns if the eventBase_ has
1843 // changed. (The caller must perform some sort of locking to transfer the
1844 // AsyncSocket between threads properly. This will be sufficient to ensure
1845 // that this thread sees the updated eventBase_ variable after
1846 // readDataAvailable() returns.)
1847 uint16_t numReads = 0;
1848 EventBase* originalEventBase = eventBase_;
1849 while (readCallback_ && eventBase_ == originalEventBase) {
1850 // Get the buffer to read into.
1851 void* buf = nullptr;
1852 size_t buflen = 0, offset = 0;
1854 prepareReadBuffer(&buf, &buflen);
1855 VLOG(5) << "prepareReadBuffer() buf=" << buf << ", buflen=" << buflen;
1856 } catch (const AsyncSocketException& ex) {
1857 return failRead(__func__, ex);
1858 } catch (const std::exception& ex) {
1859 AsyncSocketException tex(AsyncSocketException::BAD_ARGS,
1860 string("ReadCallback::getReadBuffer() "
1861 "threw exception: ") +
1863 return failRead(__func__, tex);
1865 AsyncSocketException ex(AsyncSocketException::BAD_ARGS,
1866 "ReadCallback::getReadBuffer() threw "
1867 "non-exception type");
1868 return failRead(__func__, ex);
1870 if (!isBufferMovable_ && (buf == nullptr || buflen == 0)) {
1871 AsyncSocketException ex(AsyncSocketException::BAD_ARGS,
1872 "ReadCallback::getReadBuffer() returned "
1874 return failRead(__func__, ex);
1878 auto readResult = performRead(&buf, &buflen, &offset);
1879 auto bytesRead = readResult.readReturn;
1880 VLOG(4) << "this=" << this << ", AsyncSocket::handleRead() got "
1881 << bytesRead << " bytes";
1882 if (bytesRead > 0) {
1883 if (!isBufferMovable_) {
1884 readCallback_->readDataAvailable(size_t(bytesRead));
1886 CHECK(kOpenSslModeMoveBufferOwnership);
1887 VLOG(5) << "this=" << this << ", AsyncSocket::handleRead() got "
1888 << "buf=" << buf << ", " << bytesRead << "/" << buflen
1889 << ", offset=" << offset;
1890 auto readBuf = folly::IOBuf::takeOwnership(buf, buflen);
1891 readBuf->trimStart(offset);
1892 readBuf->trimEnd(buflen - offset - bytesRead);
1893 readCallback_->readBufferAvailable(std::move(readBuf));
1896 // Fall through and continue around the loop if the read
1897 // completely filled the available buffer.
1898 // Note that readCallback_ may have been uninstalled or changed inside
1899 // readDataAvailable().
1900 if (size_t(bytesRead) < buflen) {
1903 } else if (bytesRead == READ_BLOCKING) {
1904 // No more data to read right now.
1906 } else if (bytesRead == READ_ERROR) {
1907 readErr_ = READ_ERROR;
1908 if (readResult.exception) {
1909 return failRead(__func__, *readResult.exception);
1911 auto errnoCopy = errno;
1912 AsyncSocketException ex(
1913 AsyncSocketException::INTERNAL_ERROR,
1914 withAddr("recv() failed"),
1916 return failRead(__func__, ex);
1918 assert(bytesRead == READ_EOF);
1919 readErr_ = READ_EOF;
1921 shutdownFlags_ |= SHUT_READ;
1922 if (!updateEventRegistration(0, EventHandler::READ)) {
1923 // we've already been moved into STATE_ERROR
1924 assert(state_ == StateEnum::ERROR);
1925 assert(readCallback_ == nullptr);
1929 ReadCallback* callback = readCallback_;
1930 readCallback_ = nullptr;
1931 callback->readEOF();
1934 if (maxReadsPerEvent_ && (++numReads >= maxReadsPerEvent_)) {
1935 if (readCallback_ != nullptr) {
1936 // We might still have data in the socket.
1937 // (e.g. see comment in AsyncSSLSocket::checkForImmediateRead)
1938 scheduleImmediateRead();
1946 * This function attempts to write as much data as possible, until no more data
1949 * - If it sends all available data, it unregisters for write events, and stops
1950 * the writeTimeout_.
1952 * - If not all of the data can be sent immediately, it reschedules
1953 * writeTimeout_ (if a non-zero timeout is set), and ensures the handler is
1954 * registered for write events.
1956 void AsyncSocket::handleWrite() noexcept {
1957 VLOG(5) << "AsyncSocket::handleWrite() this=" << this << ", fd=" << fd_
1958 << ", state=" << state_;
1959 DestructorGuard dg(this);
1961 if (state_ == StateEnum::CONNECTING) {
1967 assert(state_ == StateEnum::ESTABLISHED);
1968 assert((shutdownFlags_ & SHUT_WRITE) == 0);
1969 assert(writeReqHead_ != nullptr);
1971 // Loop until we run out of write requests,
1972 // or until this socket is moved to another EventBase.
1973 // (See the comment in handleRead() explaining how this can happen.)
1974 EventBase* originalEventBase = eventBase_;
1975 while (writeReqHead_ != nullptr && eventBase_ == originalEventBase) {
1976 auto writeResult = writeReqHead_->performWrite();
1977 if (writeResult.writeReturn < 0) {
1978 if (writeResult.exception) {
1979 return failWrite(__func__, *writeResult.exception);
1981 auto errnoCopy = errno;
1982 AsyncSocketException ex(
1983 AsyncSocketException::INTERNAL_ERROR,
1984 withAddr("writev() failed"),
1986 return failWrite(__func__, ex);
1987 } else if (writeReqHead_->isComplete()) {
1988 // We finished this request
1989 WriteRequest* req = writeReqHead_;
1990 writeReqHead_ = req->getNext();
1992 if (writeReqHead_ == nullptr) {
1993 writeReqTail_ = nullptr;
1994 // This is the last write request.
1995 // Unregister for write events and cancel the send timer
1996 // before we invoke the callback. We have to update the state properly
1997 // before calling the callback, since it may want to detach us from
1999 if (eventFlags_ & EventHandler::WRITE) {
2000 if (!updateEventRegistration(0, EventHandler::WRITE)) {
2001 assert(state_ == StateEnum::ERROR);
2004 // Stop the send timeout
2005 writeTimeout_.cancelTimeout();
2007 assert(!writeTimeout_.isScheduled());
2009 // If SHUT_WRITE_PENDING is set, we should shutdown the socket after
2010 // we finish sending the last write request.
2012 // We have to do this before invoking writeSuccess(), since
2013 // writeSuccess() may detach us from our EventBase.
2014 if (shutdownFlags_ & SHUT_WRITE_PENDING) {
2015 assert(connectCallback_ == nullptr);
2016 shutdownFlags_ |= SHUT_WRITE;
2018 if (shutdownFlags_ & SHUT_READ) {
2019 // Reads have already been shutdown. Fully close the socket and
2020 // move to STATE_CLOSED.
2022 // Note: This code currently moves us to STATE_CLOSED even if
2023 // close() hasn't ever been called. This can occur if we have
2024 // received EOF from the peer and shutdownWrite() has been called
2025 // locally. Should we bother staying in STATE_ESTABLISHED in this
2026 // case, until close() is actually called? I can't think of a
2027 // reason why we would need to do so. No other operations besides
2028 // calling close() or destroying the socket can be performed at
2030 assert(readCallback_ == nullptr);
2031 state_ = StateEnum::CLOSED;
2033 ioHandler_.changeHandlerFD(-1);
2037 // Reads are still enabled, so we are only doing a half-shutdown
2038 shutdown(fd_, SHUT_WR);
2043 // Invoke the callback
2044 WriteCallback* callback = req->getCallback();
2047 callback->writeSuccess();
2049 // We'll continue around the loop, trying to write another request
2052 if (bufferCallback_) {
2053 bufferCallback_->onEgressBuffered();
2055 writeReqHead_->consume();
2056 // Stop after a partial write; it's highly likely that a subsequent write
2057 // attempt will just return EAGAIN.
2059 // Ensure that we are registered for write events.
2060 if ((eventFlags_ & EventHandler::WRITE) == 0) {
2061 if (!updateEventRegistration(EventHandler::WRITE, 0)) {
2062 assert(state_ == StateEnum::ERROR);
2067 // Reschedule the send timeout, since we have made some write progress.
2068 if (sendTimeout_ > 0) {
2069 if (!writeTimeout_.scheduleTimeout(sendTimeout_)) {
2070 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
2071 withAddr("failed to reschedule write timeout"));
2072 return failWrite(__func__, ex);
2078 if (!writeReqHead_ && bufferCallback_) {
2079 bufferCallback_->onEgressBufferCleared();
2083 void AsyncSocket::checkForImmediateRead() noexcept {
2084 // We currently don't attempt to perform optimistic reads in AsyncSocket.
2085 // (However, note that some subclasses do override this method.)
2087 // Simply calling handleRead() here would be bad, as this would call
2088 // readCallback_->getReadBuffer(), forcing the callback to allocate a read
2089 // buffer even though no data may be available. This would waste lots of
2090 // memory, since the buffer will sit around unused until the socket actually
2091 // becomes readable.
2093 // Checking if the socket is readable now also seems like it would probably
2094 // be a pessimism. In most cases it probably wouldn't be readable, and we
2095 // would just waste an extra system call. Even if it is readable, waiting to
2096 // find out from libevent on the next event loop doesn't seem that bad.
2098 // The exception to this is if we have pre-received data. In that case there
2099 // is definitely data available immediately.
2100 if (preReceivedData_ && !preReceivedData_->empty()) {
2105 void AsyncSocket::handleInitialReadWrite() noexcept {
2106 // Our callers should already be holding a DestructorGuard, but grab
2107 // one here just to make sure, in case one of our calling code paths ever
2109 DestructorGuard dg(this);
2110 // If we have a readCallback_, make sure we enable read events. We
2111 // may already be registered for reads if connectSuccess() set
2112 // the read calback.
2113 if (readCallback_ && !(eventFlags_ & EventHandler::READ)) {
2114 assert(state_ == StateEnum::ESTABLISHED);
2115 assert((shutdownFlags_ & SHUT_READ) == 0);
2116 if (!updateEventRegistration(EventHandler::READ, 0)) {
2117 assert(state_ == StateEnum::ERROR);
2120 checkForImmediateRead();
2121 } else if (readCallback_ == nullptr) {
2122 // Unregister for read events.
2123 updateEventRegistration(0, EventHandler::READ);
2126 // If we have write requests pending, try to send them immediately.
2127 // Since we just finished accepting, there is a very good chance that we can
2128 // write without blocking.
2130 // However, we only process them if EventHandler::WRITE is not already set,
2131 // which means that we're already blocked on a write attempt. (This can
2132 // happen if connectSuccess() called write() before returning.)
2133 if (writeReqHead_ && !(eventFlags_ & EventHandler::WRITE)) {
2134 // Call handleWrite() to perform write processing.
2136 } else if (writeReqHead_ == nullptr) {
2137 // Unregister for write event.
2138 updateEventRegistration(0, EventHandler::WRITE);
2142 void AsyncSocket::handleConnect() noexcept {
2143 VLOG(5) << "AsyncSocket::handleConnect() this=" << this << ", fd=" << fd_
2144 << ", state=" << state_;
2145 assert(state_ == StateEnum::CONNECTING);
2146 // SHUT_WRITE can never be set while we are still connecting;
2147 // SHUT_WRITE_PENDING may be set, be we only set SHUT_WRITE once the connect
2149 assert((shutdownFlags_ & SHUT_WRITE) == 0);
2151 // In case we had a connect timeout, cancel the timeout
2152 writeTimeout_.cancelTimeout();
2153 // We don't use a persistent registration when waiting on a connect event,
2154 // so we have been automatically unregistered now. Update eventFlags_ to
2156 assert(eventFlags_ == EventHandler::WRITE);
2157 eventFlags_ = EventHandler::NONE;
2159 // Call getsockopt() to check if the connect succeeded
2161 socklen_t len = sizeof(error);
2162 int rv = getsockopt(fd_, SOL_SOCKET, SO_ERROR, &error, &len);
2164 auto errnoCopy = errno;
2165 AsyncSocketException ex(
2166 AsyncSocketException::INTERNAL_ERROR,
2167 withAddr("error calling getsockopt() after connect"),
2169 VLOG(4) << "AsyncSocket::handleConnect(this=" << this << ", fd="
2170 << fd_ << " host=" << addr_.describe()
2171 << ") exception:" << ex.what();
2172 return failConnect(__func__, ex);
2176 AsyncSocketException ex(AsyncSocketException::NOT_OPEN,
2177 "connect failed", error);
2178 VLOG(1) << "AsyncSocket::handleConnect(this=" << this << ", fd="
2179 << fd_ << " host=" << addr_.describe()
2180 << ") exception: " << ex.what();
2181 return failConnect(__func__, ex);
2184 // Move into STATE_ESTABLISHED
2185 state_ = StateEnum::ESTABLISHED;
2187 // If SHUT_WRITE_PENDING is set and we don't have any write requests to
2188 // perform, immediately shutdown the write half of the socket.
2189 if ((shutdownFlags_ & SHUT_WRITE_PENDING) && writeReqHead_ == nullptr) {
2190 // SHUT_READ shouldn't be set. If close() is called on the socket while we
2191 // are still connecting we just abort the connect rather than waiting for
2193 assert((shutdownFlags_ & SHUT_READ) == 0);
2194 shutdown(fd_, SHUT_WR);
2195 shutdownFlags_ |= SHUT_WRITE;
2198 VLOG(7) << "AsyncSocket " << this << ": fd " << fd_
2199 << "successfully connected; state=" << state_;
2201 // Remember the EventBase we are attached to, before we start invoking any
2202 // callbacks (since the callbacks may call detachEventBase()).
2203 EventBase* originalEventBase = eventBase_;
2205 invokeConnectSuccess();
2206 // Note that the connect callback may have changed our state.
2207 // (set or unset the read callback, called write(), closed the socket, etc.)
2208 // The following code needs to handle these situations correctly.
2210 // If the socket has been closed, readCallback_ and writeReqHead_ will
2211 // always be nullptr, so that will prevent us from trying to read or write.
2213 // The main thing to check for is if eventBase_ is still originalEventBase.
2214 // If not, we have been detached from this event base, so we shouldn't
2215 // perform any more operations.
2216 if (eventBase_ != originalEventBase) {
2220 handleInitialReadWrite();
2223 void AsyncSocket::timeoutExpired() noexcept {
2224 VLOG(7) << "AsyncSocket " << this << ", fd " << fd_ << ": timeout expired: "
2225 << "state=" << state_ << ", events=" << std::hex << eventFlags_;
2226 DestructorGuard dg(this);
2227 eventBase_->dcheckIsInEventBaseThread();
2229 if (state_ == StateEnum::CONNECTING) {
2230 // connect() timed out
2231 // Unregister for I/O events.
2232 if (connectCallback_) {
2233 AsyncSocketException ex(
2234 AsyncSocketException::TIMED_OUT,
2236 "connect timed out after {}ms", connectTimeout_.count()));
2237 failConnect(__func__, ex);
2239 // we faced a connect error without a connect callback, which could
2240 // happen due to TFO.
2241 AsyncSocketException ex(
2242 AsyncSocketException::TIMED_OUT, "write timed out during connection");
2243 failWrite(__func__, ex);
2246 // a normal write operation timed out
2247 AsyncSocketException ex(
2248 AsyncSocketException::TIMED_OUT,
2249 folly::sformat("write timed out after {}ms", sendTimeout_));
2250 failWrite(__func__, ex);
2254 ssize_t AsyncSocket::tfoSendMsg(int fd, struct msghdr* msg, int msg_flags) {
2255 return detail::tfo_sendmsg(fd, msg, msg_flags);
2258 AsyncSocket::WriteResult
2259 AsyncSocket::sendSocketMessage(int fd, struct msghdr* msg, int msg_flags) {
2260 ssize_t totalWritten = 0;
2261 if (state_ == StateEnum::FAST_OPEN) {
2262 sockaddr_storage addr;
2263 auto len = addr_.getAddress(&addr);
2264 msg->msg_name = &addr;
2265 msg->msg_namelen = len;
2266 totalWritten = tfoSendMsg(fd_, msg, msg_flags);
2267 if (totalWritten >= 0) {
2268 tfoFinished_ = true;
2269 state_ = StateEnum::ESTABLISHED;
2270 // We schedule this asynchrously so that we don't end up
2271 // invoking initial read or write while a write is in progress.
2272 scheduleInitialReadWrite();
2273 } else if (errno == EINPROGRESS) {
2274 VLOG(4) << "TFO falling back to connecting";
2275 // A normal sendmsg doesn't return EINPROGRESS, however
2276 // TFO might fallback to connecting if there is no
2278 state_ = StateEnum::CONNECTING;
2280 scheduleConnectTimeout();
2281 registerForConnectEvents();
2282 } catch (const AsyncSocketException& ex) {
2284 WRITE_ERROR, std::make_unique<AsyncSocketException>(ex));
2286 // Let's fake it that no bytes were written and return an errno.
2289 } else if (errno == EOPNOTSUPP) {
2290 // Try falling back to connecting.
2291 VLOG(4) << "TFO not supported";
2292 state_ = StateEnum::CONNECTING;
2294 int ret = socketConnect((const sockaddr*)&addr, len);
2296 // connect succeeded immediately
2297 // Treat this like no data was written.
2298 state_ = StateEnum::ESTABLISHED;
2299 scheduleInitialReadWrite();
2301 // If there was no exception during connections,
2302 // we would return that no bytes were written.
2305 } catch (const AsyncSocketException& ex) {
2307 WRITE_ERROR, std::make_unique<AsyncSocketException>(ex));
2309 } else if (errno == EAGAIN) {
2310 // Normally sendmsg would indicate that the write would block.
2311 // However in the fast open case, it would indicate that sendmsg
2312 // fell back to a connect. This is a return code from connect()
2313 // instead, and is an error condition indicating no fds available.
2316 std::make_unique<AsyncSocketException>(
2317 AsyncSocketException::UNKNOWN, "No more free local ports"));
2320 totalWritten = ::sendmsg(fd, msg, msg_flags);
2322 return WriteResult(totalWritten);
2325 AsyncSocket::WriteResult AsyncSocket::performWrite(
2329 uint32_t* countWritten,
2330 uint32_t* partialWritten) {
2331 // We use sendmsg() instead of writev() so that we can pass in MSG_NOSIGNAL
2332 // We correctly handle EPIPE errors, so we never want to receive SIGPIPE
2333 // (since it may terminate the program if the main program doesn't explicitly
2336 msg.msg_name = nullptr;
2337 msg.msg_namelen = 0;
2338 msg.msg_iov = const_cast<iovec *>(vec);
2339 msg.msg_iovlen = std::min<size_t>(count, kIovMax);
2341 msg.msg_controllen = sendMsgParamCallback_->getAncillaryDataSize(flags);
2342 CHECK_GE(AsyncSocket::SendMsgParamsCallback::maxAncillaryDataSize,
2343 msg.msg_controllen);
2345 if (msg.msg_controllen != 0) {
2346 msg.msg_control = reinterpret_cast<char*>(alloca(msg.msg_controllen));
2347 sendMsgParamCallback_->getAncillaryData(flags, msg.msg_control);
2349 msg.msg_control = nullptr;
2351 int msg_flags = sendMsgParamCallback_->getFlags(flags, zeroCopyEnabled_);
2353 auto writeResult = sendSocketMessage(fd_, &msg, msg_flags);
2354 auto totalWritten = writeResult.writeReturn;
2355 if (totalWritten < 0) {
2356 bool tryAgain = (errno == EAGAIN);
2358 // Apple has a bug where doing a second write on a socket which we
2359 // have opened with TFO causes an ENOTCONN to be thrown. However the
2360 // socket is really connected, so treat ENOTCONN as a EAGAIN until
2361 // this bug is fixed.
2362 tryAgain |= (errno == ENOTCONN);
2365 // workaround for running with zerocopy enabled but without a proper
2366 // memlock value - see ulimit -l
2367 if (zeroCopyEnabled_ && (errno == ENOBUFS)) {
2369 zeroCopyEnabled_ = false;
2372 if (!writeResult.exception && tryAgain) {
2373 // TCP buffer is full; we can't write any more data right now.
2375 *partialWritten = 0;
2376 return WriteResult(0);
2380 *partialWritten = 0;
2384 appBytesWritten_ += totalWritten;
2386 uint32_t bytesWritten;
2388 for (bytesWritten = uint32_t(totalWritten), n = 0; n < count; ++n) {
2389 const iovec* v = vec + n;
2390 if (v->iov_len > bytesWritten) {
2391 // Partial write finished in the middle of this iovec
2393 *partialWritten = bytesWritten;
2394 return WriteResult(totalWritten);
2397 bytesWritten -= uint32_t(v->iov_len);
2400 assert(bytesWritten == 0);
2402 *partialWritten = 0;
2403 return WriteResult(totalWritten);
2407 * Re-register the EventHandler after eventFlags_ has changed.
2409 * If an error occurs, fail() is called to move the socket into the error state
2410 * and call all currently installed callbacks. After an error, the
2411 * AsyncSocket is completely unregistered.
2413 * @return Returns true on success, or false on error.
2415 bool AsyncSocket::updateEventRegistration() {
2416 VLOG(5) << "AsyncSocket::updateEventRegistration(this=" << this
2417 << ", fd=" << fd_ << ", evb=" << eventBase_ << ", state=" << state_
2418 << ", events=" << std::hex << eventFlags_;
2419 eventBase_->dcheckIsInEventBaseThread();
2420 if (eventFlags_ == EventHandler::NONE) {
2421 ioHandler_.unregisterHandler();
2425 // Always register for persistent events, so we don't have to re-register
2426 // after being called back.
2427 if (!ioHandler_.registerHandler(
2428 uint16_t(eventFlags_ | EventHandler::PERSIST))) {
2429 eventFlags_ = EventHandler::NONE; // we're not registered after error
2430 AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR,
2431 withAddr("failed to update AsyncSocket event registration"));
2432 fail("updateEventRegistration", ex);
2439 bool AsyncSocket::updateEventRegistration(uint16_t enable,
2441 uint16_t oldFlags = eventFlags_;
2442 eventFlags_ |= enable;
2443 eventFlags_ &= ~disable;
2444 if (eventFlags_ == oldFlags) {
2447 return updateEventRegistration();
2451 void AsyncSocket::startFail() {
2452 // startFail() should only be called once
2453 assert(state_ != StateEnum::ERROR);
2454 assert(getDestructorGuardCount() > 0);
2455 state_ = StateEnum::ERROR;
2456 // Ensure that SHUT_READ and SHUT_WRITE are set,
2457 // so all future attempts to read or write will be rejected
2458 shutdownFlags_ |= (SHUT_READ | SHUT_WRITE);
2460 if (eventFlags_ != EventHandler::NONE) {
2461 eventFlags_ = EventHandler::NONE;
2462 ioHandler_.unregisterHandler();
2464 writeTimeout_.cancelTimeout();
2467 ioHandler_.changeHandlerFD(-1);
2472 void AsyncSocket::invokeAllErrors(const AsyncSocketException& ex) {
2473 invokeConnectErr(ex);
2476 if (readCallback_) {
2477 ReadCallback* callback = readCallback_;
2478 readCallback_ = nullptr;
2479 callback->readErr(ex);
2483 void AsyncSocket::finishFail() {
2484 assert(state_ == StateEnum::ERROR);
2485 assert(getDestructorGuardCount() > 0);
2487 AsyncSocketException ex(
2488 AsyncSocketException::INTERNAL_ERROR,
2489 withAddr("socket closing after error"));
2490 invokeAllErrors(ex);
2493 void AsyncSocket::finishFail(const AsyncSocketException& ex) {
2494 assert(state_ == StateEnum::ERROR);
2495 assert(getDestructorGuardCount() > 0);
2496 invokeAllErrors(ex);
2499 void AsyncSocket::fail(const char* fn, const AsyncSocketException& ex) {
2500 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2501 << state_ << " host=" << addr_.describe()
2502 << "): failed in " << fn << "(): "
2508 void AsyncSocket::failConnect(const char* fn, const AsyncSocketException& ex) {
2509 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2510 << state_ << " host=" << addr_.describe()
2511 << "): failed while connecting in " << fn << "(): "
2515 invokeConnectErr(ex);
2519 void AsyncSocket::failRead(const char* fn, const AsyncSocketException& ex) {
2520 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2521 << state_ << " host=" << addr_.describe()
2522 << "): failed while reading in " << fn << "(): "
2526 if (readCallback_ != nullptr) {
2527 ReadCallback* callback = readCallback_;
2528 readCallback_ = nullptr;
2529 callback->readErr(ex);
2535 void AsyncSocket::failErrMessageRead(const char* fn,
2536 const AsyncSocketException& ex) {
2537 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2538 << state_ << " host=" << addr_.describe()
2539 << "): failed while reading message in " << fn << "(): "
2543 if (errMessageCallback_ != nullptr) {
2544 ErrMessageCallback* callback = errMessageCallback_;
2545 errMessageCallback_ = nullptr;
2546 callback->errMessageError(ex);
2552 void AsyncSocket::failWrite(const char* fn, const AsyncSocketException& ex) {
2553 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2554 << state_ << " host=" << addr_.describe()
2555 << "): failed while writing in " << fn << "(): "
2559 // Only invoke the first write callback, since the error occurred while
2560 // writing this request. Let any other pending write callbacks be invoked in
2562 if (writeReqHead_ != nullptr) {
2563 WriteRequest* req = writeReqHead_;
2564 writeReqHead_ = req->getNext();
2565 WriteCallback* callback = req->getCallback();
2566 uint32_t bytesWritten = req->getTotalBytesWritten();
2569 callback->writeErr(bytesWritten, ex);
2576 void AsyncSocket::failWrite(const char* fn, WriteCallback* callback,
2577 size_t bytesWritten,
2578 const AsyncSocketException& ex) {
2579 // This version of failWrite() is used when the failure occurs before
2580 // we've added the callback to writeReqHead_.
2581 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state="
2582 << state_ << " host=" << addr_.describe()
2583 <<"): failed while writing in " << fn << "(): "
2587 if (callback != nullptr) {
2588 callback->writeErr(bytesWritten, ex);
2594 void AsyncSocket::failAllWrites(const AsyncSocketException& ex) {
2595 // Invoke writeError() on all write callbacks.
2596 // This is used when writes are forcibly shutdown with write requests
2597 // pending, or when an error occurs with writes pending.
2598 while (writeReqHead_ != nullptr) {
2599 WriteRequest* req = writeReqHead_;
2600 writeReqHead_ = req->getNext();
2601 WriteCallback* callback = req->getCallback();
2603 callback->writeErr(req->getTotalBytesWritten(), ex);
2609 void AsyncSocket::invalidState(ConnectCallback* callback) {
2610 VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_
2611 << "): connect() called in invalid state " << state_;
2614 * The invalidState() methods don't use the normal failure mechanisms,
2615 * since we don't know what state we are in. We don't want to call
2616 * startFail()/finishFail() recursively if we are already in the middle of
2620 AsyncSocketException ex(AsyncSocketException::ALREADY_OPEN,
2621 "connect() called with socket in invalid state");
2622 connectEndTime_ = std::chrono::steady_clock::now();
2623 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2625 callback->connectErr(ex);
2628 // We can't use failConnect() here since connectCallback_
2629 // may already be set to another callback. Invoke this ConnectCallback
2630 // here; any other connectCallback_ will be invoked in finishFail()
2633 callback->connectErr(ex);
2639 void AsyncSocket::invalidState(ErrMessageCallback* callback) {
2640 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_
2641 << "): setErrMessageCB(" << callback
2642 << ") called in invalid state " << state_;
2644 AsyncSocketException ex(
2645 AsyncSocketException::NOT_OPEN,
2646 msgErrQueueSupported
2647 ? "setErrMessageCB() called with socket in invalid state"
2648 : "This platform does not support socket error message notifications");
2649 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2651 callback->errMessageError(ex);
2656 callback->errMessageError(ex);
2662 void AsyncSocket::invokeConnectErr(const AsyncSocketException& ex) {
2663 connectEndTime_ = std::chrono::steady_clock::now();
2664 if (connectCallback_) {
2665 ConnectCallback* callback = connectCallback_;
2666 connectCallback_ = nullptr;
2667 callback->connectErr(ex);
2671 void AsyncSocket::invokeConnectSuccess() {
2672 connectEndTime_ = std::chrono::steady_clock::now();
2673 if (connectCallback_) {
2674 ConnectCallback* callback = connectCallback_;
2675 connectCallback_ = nullptr;
2676 callback->connectSuccess();
2680 void AsyncSocket::invalidState(ReadCallback* callback) {
2681 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_
2682 << "): setReadCallback(" << callback
2683 << ") called in invalid state " << state_;
2685 AsyncSocketException ex(AsyncSocketException::NOT_OPEN,
2686 "setReadCallback() called with socket in "
2688 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2690 callback->readErr(ex);
2695 callback->readErr(ex);
2701 void AsyncSocket::invalidState(WriteCallback* callback) {
2702 VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_
2703 << "): write() called in invalid state " << state_;
2705 AsyncSocketException ex(AsyncSocketException::NOT_OPEN,
2706 withAddr("write() called with socket in invalid state"));
2707 if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) {
2709 callback->writeErr(0, ex);
2714 callback->writeErr(0, ex);
2720 void AsyncSocket::doClose() {
2724 if (const auto shutdownSocketSet = wShutdownSocketSet_.lock()) {
2725 shutdownSocketSet->close(fd_);
2732 std::ostream& operator << (std::ostream& os,
2733 const AsyncSocket::StateEnum& state) {
2734 os << static_cast<int>(state);
2738 std::string AsyncSocket::withAddr(const std::string& s) {
2739 // Don't use addr_ directly because it may not be initialized
2740 // e.g. if constructed from fd
2741 folly::SocketAddress peer, local;
2743 getPeerAddress(&peer);
2744 getLocalAddress(&local);
2745 } catch (const std::exception&) {
2750 return s + " (peer=" + peer.describe() + ", local=" + local.describe() + ")";
2753 void AsyncSocket::setBufferCallback(BufferCallback* cb) {
2754 bufferCallback_ = cb;
2757 } // namespace folly