X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=folly%2Fio%2Fasync%2FAsyncSocket.cpp;h=2ecfbb49811abc0d32adae475978df4a7aa7241e;hb=1672380910a8c21cd36095661eb1360f43c93332;hp=0fe8e88613c013b759771b476ed6b7db33344d3c;hpb=46f3788d8acb3d351698d9ffff551807c807caa7;p=folly.git diff --git a/folly/io/async/AsyncSocket.cpp b/folly/io/async/AsyncSocket.cpp index 0fe8e886..2ecfbb49 100644 --- a/folly/io/async/AsyncSocket.cpp +++ b/folly/io/async/AsyncSocket.cpp @@ -1,5 +1,5 @@ /* - * Copyright 2016 Facebook, Inc. + * Copyright 2017 Facebook, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -16,26 +16,37 @@ #include -#include -#include +#include +#include #include +#include #include +#include #include #include #include #include +#include #include #include -#include #include -#include +#include using std::string; using std::unique_ptr; +namespace fsp = folly::portability::sockets; + namespace folly { +static constexpr bool msgErrQueueSupported = +#ifdef MSG_ERRQUEUE + true; +#else + false; +#endif // MSG_ERRQUEUE + // static members initializers const AsyncSocket::OptionMap AsyncSocket::emptyOptionMap; @@ -91,10 +102,12 @@ class AsyncSocket::BytesWriteRequest : public AsyncSocket::WriteRequest { WriteResult performWrite() override { WriteFlags writeFlags = flags_; if (getNext() != nullptr) { - writeFlags = writeFlags | WriteFlags::CORK; + writeFlags |= WriteFlags::CORK; } - return socket_->performWrite( + auto writeResult = socket_->performWrite( getOps(), getOpCount(), writeFlags, &opsWritten_, &partialBytes_); + bytesWritten_ = writeResult.writeReturn > 0 ? writeResult.writeReturn : 0; + return writeResult; } bool isComplete() override { @@ -122,7 +135,8 @@ class AsyncSocket::BytesWriteRequest : public AsyncSocket::WriteRequest { currentOp->iov_len -= partialBytes_; // Increment the totalBytesWritten_ count by bytesWritten_; - totalBytesWritten_ += bytesWritten_; + assert(bytesWritten_ >= 0); + totalBytesWritten_ += uint32_t(bytesWritten_); } private: @@ -171,20 +185,47 @@ class AsyncSocket::BytesWriteRequest : public AsyncSocket::WriteRequest { struct iovec writeOps_[]; ///< write operation(s) list }; +int AsyncSocket::SendMsgParamsCallback::getDefaultFlags(folly::WriteFlags flags) + noexcept { + int msg_flags = MSG_DONTWAIT; + +#ifdef MSG_NOSIGNAL // Linux-only + msg_flags |= MSG_NOSIGNAL; +#ifdef MSG_MORE + if (isSet(flags, WriteFlags::CORK)) { + // MSG_MORE tells the kernel we have more data to send, so wait for us to + // give it the rest of the data rather than immediately sending a partial + // frame, even when TCP_NODELAY is enabled. + msg_flags |= MSG_MORE; + } +#endif // MSG_MORE +#endif // MSG_NOSIGNAL + if (isSet(flags, WriteFlags::EOR)) { + // marks that this is the last byte of a record (response) + msg_flags |= MSG_EOR; + } + + return msg_flags; +} + +namespace { +static AsyncSocket::SendMsgParamsCallback defaultSendMsgParamsCallback; +} + AsyncSocket::AsyncSocket() - : eventBase_(nullptr) - , writeTimeout_(this, nullptr) - , ioHandler_(this, nullptr) - , immediateReadHandler_(this) { + : eventBase_(nullptr), + writeTimeout_(this, nullptr), + ioHandler_(this, nullptr), + immediateReadHandler_(this) { VLOG(5) << "new AsyncSocket()"; init(); } AsyncSocket::AsyncSocket(EventBase* evb) - : eventBase_(evb) - , writeTimeout_(this, evb) - , ioHandler_(this, evb) - , immediateReadHandler_(this) { + : eventBase_(evb), + writeTimeout_(this, evb), + ioHandler_(this, evb), + immediateReadHandler_(this) { VLOG(5) << "new AsyncSocket(" << this << ", evb=" << evb << ")"; init(); } @@ -205,10 +246,10 @@ AsyncSocket::AsyncSocket(EventBase* evb, } AsyncSocket::AsyncSocket(EventBase* evb, int fd) - : eventBase_(evb) - , writeTimeout_(this, evb) - , ioHandler_(this, evb, fd) - , immediateReadHandler_(this) { + : eventBase_(evb), + writeTimeout_(this, evb), + ioHandler_(this, evb, fd), + immediateReadHandler_(this) { VLOG(5) << "new AsyncSocket(" << this << ", evb=" << evb << ", fd=" << fd << ")"; init(); @@ -217,6 +258,11 @@ AsyncSocket::AsyncSocket(EventBase* evb, int fd) state_ = StateEnum::ESTABLISHED; } +AsyncSocket::AsyncSocket(AsyncSocket::UniquePtr oldAsyncSocket) + : AsyncSocket(oldAsyncSocket->getEventBase(), oldAsyncSocket->detachFd()) { + preReceivedData_ = std::move(oldAsyncSocket->preReceivedData_); +} + // init() method, since constructor forwarding isn't supported in most // compilers yet. void AsyncSocket::init() { @@ -228,12 +274,14 @@ void AsyncSocket::init() { sendTimeout_ = 0; maxReadsPerEvent_ = 16; connectCallback_ = nullptr; + errMessageCallback_ = nullptr; readCallback_ = nullptr; writeReqHead_ = nullptr; writeReqTail_ = nullptr; shutdownSocketSet_ = nullptr; appBytesWritten_ = 0; appBytesReceived_ = 0; + sendMsgParamCallback_ = &defaultSendMsgParamsCallback; } AsyncSocket::~AsyncSocket() { @@ -335,7 +383,7 @@ void AsyncSocket::connect(ConnectCallback* callback, // constant (PF_xxx) rather than an address family (AF_xxx), but the // distinction is mainly just historical. In pretty much all // implementations the PF_foo and AF_foo constants are identical. - fd_ = socket(address.getFamily(), SOCK_STREAM, 0); + fd_ = fsp::socket(address.getFamily(), SOCK_STREAM, 0); if (fd_ < 0) { auto errnoCopy = errno; throw AsyncSocketException( @@ -393,7 +441,7 @@ void AsyncSocket::connect(ConnectCallback* callback, // bind the socket if (bindAddr != anyAddress()) { int one = 1; - if (::setsockopt(fd_, SOL_SOCKET, SO_REUSEADDR, &one, sizeof(one))) { + if (setsockopt(fd_, SOL_SOCKET, SO_REUSEADDR, &one, sizeof(one))) { auto errnoCopy = errno; doClose(); throw AsyncSocketException( @@ -404,7 +452,7 @@ void AsyncSocket::connect(ConnectCallback* callback, bindAddr.getAddress(&addrStorage); - if (::bind(fd_, saddr, bindAddr.getActualSize()) != 0) { + if (bind(fd_, saddr, bindAddr.getActualSize()) != 0) { auto errnoCopy = errno; doClose(); throw AsyncSocketException( @@ -416,7 +464,7 @@ void AsyncSocket::connect(ConnectCallback* callback, // Apply the additional options if any. for (const auto& opt: options) { - int rv = opt.first.apply(fd_, opt.second); + rv = opt.first.apply(fd_, opt.second); if (rv != 0) { auto errnoCopy = errno; throw AsyncSocketException( @@ -429,34 +477,12 @@ void AsyncSocket::connect(ConnectCallback* callback, // Perform the connect() address.getAddress(&addrStorage); - rv = ::connect(fd_, saddr, address.getActualSize()); - if (rv < 0) { - auto errnoCopy = errno; - if (errnoCopy == EINPROGRESS) { - // Connection in progress. - if (timeout > 0) { - // Start a timer in case the connection takes too long. - if (!writeTimeout_.scheduleTimeout(timeout)) { - throw AsyncSocketException(AsyncSocketException::INTERNAL_ERROR, - withAddr("failed to schedule AsyncSocket connect timeout")); - } - } - - // Register for write events, so we'll - // be notified when the connection finishes/fails. - // Note that we don't register for a persistent event here. - assert(eventFlags_ == EventHandler::NONE); - eventFlags_ = EventHandler::WRITE; - if (!ioHandler_.registerHandler(eventFlags_)) { - throw AsyncSocketException(AsyncSocketException::INTERNAL_ERROR, - withAddr("failed to register AsyncSocket connect handler")); - } + if (tfoEnabled_) { + state_ = StateEnum::FAST_OPEN; + tfoAttempted_ = true; + } else { + if (socketConnect(saddr, addr_.getActualSize()) < 0) { return; - } else { - throw AsyncSocketException( - AsyncSocketException::NOT_OPEN, - "connect failed (immediately)", - errnoCopy); } } @@ -479,12 +505,64 @@ void AsyncSocket::connect(ConnectCallback* callback, // The read callback may not have been set yet, and no writes may be pending // yet, so we don't have to register for any events at the moment. VLOG(8) << "AsyncSocket::connect succeeded immediately; this=" << this; + assert(errMessageCallback_ == nullptr); assert(readCallback_ == nullptr); assert(writeReqHead_ == nullptr); - state_ = StateEnum::ESTABLISHED; + if (state_ != StateEnum::FAST_OPEN) { + state_ = StateEnum::ESTABLISHED; + } invokeConnectSuccess(); } +int AsyncSocket::socketConnect(const struct sockaddr* saddr, socklen_t len) { +#if __linux__ + if (noTransparentTls_) { + // Ignore return value, errors are ok + setsockopt(fd_, SOL_SOCKET, SO_NO_TRANSPARENT_TLS, nullptr, 0); + } +#endif + int rv = fsp::connect(fd_, saddr, len); + if (rv < 0) { + auto errnoCopy = errno; + if (errnoCopy == EINPROGRESS) { + scheduleConnectTimeout(); + registerForConnectEvents(); + } else { + throw AsyncSocketException( + AsyncSocketException::NOT_OPEN, + "connect failed (immediately)", + errnoCopy); + } + } + return rv; +} + +void AsyncSocket::scheduleConnectTimeout() { + // Connection in progress. + auto timeout = connectTimeout_.count(); + if (timeout > 0) { + // Start a timer in case the connection takes too long. + if (!writeTimeout_.scheduleTimeout(uint32_t(timeout))) { + throw AsyncSocketException( + AsyncSocketException::INTERNAL_ERROR, + withAddr("failed to schedule AsyncSocket connect timeout")); + } + } +} + +void AsyncSocket::registerForConnectEvents() { + // Register for write events, so we'll + // be notified when the connection finishes/fails. + // Note that we don't register for a persistent event here. + assert(eventFlags_ == EventHandler::NONE); + eventFlags_ = EventHandler::WRITE; + if (!ioHandler_.registerHandler(eventFlags_)) { + throw AsyncSocketException( + AsyncSocketException::INTERNAL_ERROR, + withAddr("failed to register AsyncSocket connect handler")); + } +} + void AsyncSocket::connect(ConnectCallback* callback, const string& ip, uint16_t port, int timeout, @@ -502,7 +580,7 @@ void AsyncSocket::connect(ConnectCallback* callback, void AsyncSocket::cancelConnect() { connectCallback_ = nullptr; - if (state_ == StateEnum::CONNECTING) { + if (state_ == StateEnum::CONNECTING || state_ == StateEnum::FAST_OPEN) { closeNow(); } } @@ -514,7 +592,7 @@ void AsyncSocket::setSendTimeout(uint32_t milliseconds) { // If we are currently pending on write requests, immediately update // writeTimeout_ with the new value. if ((eventFlags_ & EventHandler::WRITE) && - (state_ != StateEnum::CONNECTING)) { + (state_ != StateEnum::CONNECTING && state_ != StateEnum::FAST_OPEN)) { assert(state_ == StateEnum::ESTABLISHED); assert((shutdownFlags_ & SHUT_WRITE) == 0); if (sendTimeout_ > 0) { @@ -529,6 +607,68 @@ void AsyncSocket::setSendTimeout(uint32_t milliseconds) { } } +void AsyncSocket::setErrMessageCB(ErrMessageCallback* callback) { + VLOG(6) << "AsyncSocket::setErrMessageCB() this=" << this + << ", fd=" << fd_ << ", callback=" << callback + << ", state=" << state_; + + // Short circuit if callback is the same as the existing errMessageCallback_. + if (callback == errMessageCallback_) { + return; + } + + if (!msgErrQueueSupported) { + // Per-socket error message queue is not supported on this platform. + return invalidState(callback); + } + + DestructorGuard dg(this); + assert(eventBase_->isInEventBaseThread()); + + if (callback == nullptr) { + // We should be able to reset the callback regardless of the + // socket state. It's important to have a reliable callback + // cancellation mechanism. + errMessageCallback_ = callback; + return; + } + + switch ((StateEnum)state_) { + case StateEnum::CONNECTING: + case StateEnum::FAST_OPEN: + case StateEnum::ESTABLISHED: { + errMessageCallback_ = callback; + return; + } + case StateEnum::CLOSED: + case StateEnum::ERROR: + // We should never reach here. SHUT_READ should always be set + // if we are in STATE_CLOSED or STATE_ERROR. + assert(false); + return invalidState(callback); + case StateEnum::UNINIT: + // We do not allow setReadCallback() to be called before we start + // connecting. + return invalidState(callback); + } + + // We don't put a default case in the switch statement, so that the compiler + // will warn us to update the switch statement if a new state is added. + return invalidState(callback); +} + +AsyncSocket::ErrMessageCallback* AsyncSocket::getErrMessageCallback() const { + return errMessageCallback_; +} + +void AsyncSocket::setSendMsgParamCB(SendMsgParamsCallback* callback) { + sendMsgParamCallback_ = callback; +} + +AsyncSocket::SendMsgParamsCallback* AsyncSocket::getSendMsgParamsCB() const { + return sendMsgParamCallback_; +} + void AsyncSocket::setReadCB(ReadCallback *callback) { VLOG(6) << "AsyncSocket::setReadCallback() this=" << this << ", fd=" << fd_ << ", callback=" << callback << ", state=" << state_; @@ -573,6 +713,7 @@ void AsyncSocket::setReadCB(ReadCallback *callback) { switch ((StateEnum)state_) { case StateEnum::CONNECTING: + case StateEnum::FAST_OPEN: // For convenience, we allow the read callback to be set while we are // still connecting. We just store the callback for now. Once the // connection completes we'll register for read events. @@ -642,7 +783,12 @@ void AsyncSocket::writeChain(WriteCallback* callback, unique_ptr&& buf, constexpr size_t kSmallSizeMax = 64; size_t count = buf->countChainElements(); if (count <= kSmallSizeMax) { + // suppress "warning: variable length array 'vec' is used [-Wvla]" + FOLLY_PUSH_WARNING + FOLLY_GCC_DISABLE_WARNING("-Wvla") iovec vec[BOOST_PP_IF(FOLLY_HAVE_VLA, count, kSmallSizeMax)]; + FOLLY_POP_WARNING + writeChainImpl(callback, vec, count, std::move(buf), flags); } else { iovec* vec = new iovec[count]; @@ -681,17 +827,18 @@ void AsyncSocket::writeImpl(WriteCallback* callback, const iovec* vec, uint32_t countWritten = 0; uint32_t partialWritten = 0; - int bytesWritten = 0; + ssize_t bytesWritten = 0; bool mustRegister = false; - if (state_ == StateEnum::ESTABLISHED && !connecting()) { + if ((state_ == StateEnum::ESTABLISHED || state_ == StateEnum::FAST_OPEN) && + !connecting()) { if (writeReqHead_ == nullptr) { // If we are established and there are no other writes pending, // we can attempt to perform the write immediately. assert(writeReqTail_ == nullptr); assert((eventFlags_ & EventHandler::WRITE) == 0); - auto writeResult = - performWrite(vec, count, flags, &countWritten, &partialWritten); + auto writeResult = performWrite( + vec, uint32_t(count), flags, &countWritten, &partialWritten); bytesWritten = writeResult.writeReturn; if (bytesWritten < 0) { auto errnoCopy = errno; @@ -715,7 +862,13 @@ void AsyncSocket::writeImpl(WriteCallback* callback, const iovec* vec, bufferCallback_->onEgressBuffered(); } } - mustRegister = true; + if (!connecting()) { + // Writes might put the socket back into connecting state + // if TFO is enabled, and using TFO fails. + // This means that write timeouts would not be active, however + // connect timeouts would affect this stage. + mustRegister = true; + } } } else if (!connecting()) { // Invalid state for writing @@ -725,14 +878,20 @@ void AsyncSocket::writeImpl(WriteCallback* callback, const iovec* vec, // Create a new WriteRequest to add to the queue WriteRequest* req; try { - req = BytesWriteRequest::newRequest(this, callback, vec + countWritten, - count - countWritten, partialWritten, - bytesWritten, std::move(ioBuf), flags); + req = BytesWriteRequest::newRequest( + this, + callback, + vec + countWritten, + uint32_t(count - countWritten), + partialWritten, + uint32_t(bytesWritten), + std::move(ioBuf), + flags); } catch (const std::exception& ex) { // we mainly expect to catch std::bad_alloc here AsyncSocketException tex(AsyncSocketException::INTERNAL_ERROR, withAddr(string("failed to append new WriteRequest: ") + ex.what())); - return failWrite(__func__, callback, bytesWritten, tex); + return failWrite(__func__, callback, size_t(bytesWritten), tex); } req->consume(); if (writeReqTail_ == nullptr) { @@ -839,7 +998,7 @@ void AsyncSocket::closeNow() { switch (state_) { case StateEnum::ESTABLISHED: case StateEnum::CONNECTING: - { + case StateEnum::FAST_OPEN: { shutdownFlags_ |= (SHUT_READ | SHUT_WRITE); state_ = StateEnum::CLOSED; @@ -971,7 +1130,7 @@ void AsyncSocket::shutdownWriteNow() { } // Shutdown writes on the file descriptor - ::shutdown(fd_, SHUT_WR); + shutdown(fd_, SHUT_WR); // Immediately fail all write requests failAllWrites(socketShutdownForWritesEx); @@ -995,6 +1154,13 @@ void AsyncSocket::shutdownWriteNow() { // immediately shut down the write side of the socket. shutdownFlags_ |= SHUT_WRITE_PENDING; return; + case StateEnum::FAST_OPEN: + // In fast open state we haven't call connected yet, and if we shutdown + // the writes, we will never try to call connect, so shut everything down + shutdownFlags_ |= SHUT_WRITE; + // Immediately fail all write requests + failAllWrites(socketShutdownForWritesEx); + return; case StateEnum::CLOSED: case StateEnum::ERROR: // We should never get here. SHUT_WRITE should always be set @@ -1023,6 +1189,18 @@ bool AsyncSocket::readable() const { return rc == 1; } +bool AsyncSocket::writable() const { + if (fd_ == -1) { + return false; + } + struct pollfd fds[1]; + fds[0].fd = fd_; + fds[0].events = POLLOUT; + fds[0].revents = 0; + int rc = poll(fds, 1, 0); + return rc == 1; +} + bool AsyncSocket::isPending() const { return ioHandler_.isPending(); } @@ -1046,9 +1224,10 @@ bool AsyncSocket::hangup() const { } bool AsyncSocket::good() const { - return ((state_ == StateEnum::CONNECTING || - state_ == StateEnum::ESTABLISHED) && - (shutdownFlags_ == 0) && (eventBase_ != nullptr)); + return ( + (state_ == StateEnum::CONNECTING || state_ == StateEnum::FAST_OPEN || + state_ == StateEnum::ESTABLISHED) && + (shutdownFlags_ == 0) && (eventBase_ != nullptr)); } bool AsyncSocket::error() const { @@ -1066,6 +1245,9 @@ void AsyncSocket::attachEventBase(EventBase* eventBase) { eventBase_ = eventBase; ioHandler_.attachEventBase(eventBase); writeTimeout_.attachEventBase(eventBase); + if (evbChangeCb_) { + evbChangeCb_->evbAttached(this); + } } void AsyncSocket::detachEventBase() { @@ -1078,6 +1260,9 @@ void AsyncSocket::detachEventBase() { eventBase_ = nullptr; ioHandler_.detachEventBase(); writeTimeout_.detachEventBase(); + if (evbChangeCb_) { + evbChangeCb_->evbDetached(this); + } } bool AsyncSocket::isDetachable() const { @@ -1101,6 +1286,10 @@ void AsyncSocket::getPeerAddress(folly::SocketAddress* address) const { *address = addr_; } +bool AsyncSocket::getTFOSucceded() const { + return detail::tfo_succeeded(fd_); +} + int AsyncSocket::setNoDelay(bool noDelay) { if (fd_ < 0) { VLOG(4) << "AsyncSocket::setNoDelay() called on non-open socket " @@ -1134,8 +1323,12 @@ int AsyncSocket::setCongestionFlavor(const std::string &cname) { } - if (setsockopt(fd_, IPPROTO_TCP, TCP_CONGESTION, cname.c_str(), - cname.length() + 1) != 0) { + if (setsockopt( + fd_, + IPPROTO_TCP, + TCP_CONGESTION, + cname.c_str(), + socklen_t(cname.length() + 1)) != 0) { int errnoCopy = errno; VLOG(2) << "failed to update TCP_CONGESTION option on AsyncSocket " << this << "(fd=" << fd_ << ", state=" << state_ << "): " @@ -1147,6 +1340,7 @@ int AsyncSocket::setCongestionFlavor(const std::string &cname) { } int AsyncSocket::setQuickAck(bool quickack) { + (void)quickack; if (fd_ < 0) { VLOG(4) << "AsyncSocket::setQuickAck() called on non-open socket " << this << "(state=" << state_ << ")"; @@ -1225,19 +1419,30 @@ int AsyncSocket::setTCPProfile(int profd) { } void AsyncSocket::ioReady(uint16_t events) noexcept { - VLOG(7) << "AsyncSocket::ioRead() this=" << this << ", fd" << fd_ + VLOG(7) << "AsyncSocket::ioRead() this=" << this << ", fd=" << fd_ << ", events=" << std::hex << events << ", state=" << state_; DestructorGuard dg(this); assert(events & EventHandler::READ_WRITE); assert(eventBase_->isInEventBaseThread()); - uint16_t relevantEvents = events & EventHandler::READ_WRITE; + uint16_t relevantEvents = uint16_t(events & EventHandler::READ_WRITE); + EventBase* originalEventBase = eventBase_; + // If we got there it means that either EventHandler::READ or + // EventHandler::WRITE is set. Any of these flags can + // indicate that there are messages available in the socket + // error message queue. + handleErrMessages(); + + // Return now if handleErrMessages() detached us from our EventBase + if (eventBase_ != originalEventBase) { + return; + } + if (relevantEvents == EventHandler::READ) { handleRead(); } else if (relevantEvents == EventHandler::WRITE) { handleWrite(); } else if (relevantEvents == EventHandler::READ_WRITE) { - EventBase* originalEventBase = eventBase_; // If both read and write events are ready, process writes first. handleWrite(); @@ -1264,12 +1469,23 @@ AsyncSocket::performRead(void** buf, size_t* buflen, size_t* /* offset */) { VLOG(5) << "AsyncSocket::performRead() this=" << this << ", buf=" << *buf << ", buflen=" << *buflen; - int recvFlags = 0; - if (peek_) { - recvFlags |= MSG_PEEK; + if (preReceivedData_ && !preReceivedData_->empty()) { + VLOG(5) << "AsyncSocket::performRead() this=" << this + << ", reading pre-received data"; + + io::Cursor cursor(preReceivedData_.get()); + auto len = cursor.pullAtMost(*buf, *buflen); + + IOBufQueue queue; + queue.append(std::move(preReceivedData_)); + queue.trimStart(len); + preReceivedData_ = queue.move(); + + appBytesReceived_ += len; + return ReadResult(len); } - ssize_t bytes = recv(fd_, *buf, *buflen, MSG_DONTWAIT | recvFlags); + ssize_t bytes = recv(fd_, *buf, *buflen, MSG_DONTWAIT); if (bytes < 0) { if (errno == EAGAIN || errno == EWOULDBLOCK) { // No more data to read right now. @@ -1283,12 +1499,69 @@ AsyncSocket::performRead(void** buf, size_t* buflen, size_t* /* offset */) { } } -void AsyncSocket::prepareReadBuffer(void** buf, size_t* buflen) noexcept { +void AsyncSocket::prepareReadBuffer(void** buf, size_t* buflen) { // no matter what, buffer should be preapared for non-ssl socket CHECK(readCallback_); readCallback_->getReadBuffer(buf, buflen); } +void AsyncSocket::handleErrMessages() noexcept { + // This method has non-empty implementation only for platforms + // supporting per-socket error queues. + VLOG(5) << "AsyncSocket::handleErrMessages() this=" << this << ", fd=" << fd_ + << ", state=" << state_; + if (errMessageCallback_ == nullptr) { + VLOG(7) << "AsyncSocket::handleErrMessages(): " + << "no callback installed - exiting."; + return; + } + +#ifdef MSG_ERRQUEUE + uint8_t ctrl[1024]; + unsigned char data; + struct msghdr msg; + iovec entry; + + entry.iov_base = &data; + entry.iov_len = sizeof(data); + msg.msg_iov = &entry; + msg.msg_iovlen = 1; + msg.msg_name = nullptr; + msg.msg_namelen = 0; + msg.msg_control = ctrl; + msg.msg_controllen = sizeof(ctrl); + msg.msg_flags = 0; + + int ret; + while (true) { + ret = recvmsg(fd_, &msg, MSG_ERRQUEUE); + VLOG(5) << "AsyncSocket::handleErrMessages(): recvmsg returned " << ret; + + if (ret < 0) { + if (errno != EAGAIN) { + auto errnoCopy = errno; + LOG(ERROR) << "::recvmsg exited with code " << ret + << ", errno: " << errnoCopy; + AsyncSocketException ex( + AsyncSocketException::INTERNAL_ERROR, + withAddr("recvmsg() failed"), + errnoCopy); + failErrMessageRead(__func__, ex); + } + return; + } + + for (struct cmsghdr* cmsg = CMSG_FIRSTHDR(&msg); + cmsg != nullptr && + cmsg->cmsg_len != 0 && + errMessageCallback_ != nullptr; + cmsg = CMSG_NXTHDR(&msg, cmsg)) { + errMessageCallback_->errMessage(*cmsg); + } + } +#endif //MSG_ERRQUEUE +} + void AsyncSocket::handleRead() noexcept { VLOG(5) << "AsyncSocket::handleRead() this=" << this << ", fd=" << fd_ << ", state=" << state_; @@ -1352,7 +1625,7 @@ void AsyncSocket::handleRead() noexcept { << bytesRead << " bytes"; if (bytesRead > 0) { if (!isBufferMovable_) { - readCallback_->readDataAvailable(bytesRead); + readCallback_->readDataAvailable(size_t(bytesRead)); } else { CHECK(kOpenSslModeMoveBufferOwnership); VLOG(5) << "this=" << this << ", AsyncSocket::handleRead() got " @@ -1506,7 +1779,7 @@ void AsyncSocket::handleWrite() noexcept { } } else { // Reads are still enabled, so we are only doing a half-shutdown - ::shutdown(fd_, SHUT_WR); + shutdown(fd_, SHUT_WR); } } } @@ -1565,6 +1838,12 @@ void AsyncSocket::checkForImmediateRead() noexcept { // be a pessimism. In most cases it probably wouldn't be readable, and we // would just waste an extra system call. Even if it is readable, waiting to // find out from libevent on the next event loop doesn't seem that bad. + // + // The exception to this is if we have pre-received data. In that case there + // is definitely data available immediately. + if (preReceivedData_ && !preReceivedData_->empty()) { + handleRead(); + } } void AsyncSocket::handleInitialReadWrite() noexcept { @@ -1572,7 +1851,6 @@ void AsyncSocket::handleInitialReadWrite() noexcept { // one here just to make sure, in case one of our calling code paths ever // changes. DestructorGuard dg(this); - // If we have a readCallback_, make sure we enable read events. We // may already be registered for reads if connectSuccess() set // the read calback. @@ -1657,7 +1935,7 @@ void AsyncSocket::handleConnect() noexcept { // are still connecting we just abort the connect rather than waiting for // it to complete. assert((shutdownFlags_ & SHUT_READ) == 0); - ::shutdown(fd_, SHUT_WR); + shutdown(fd_, SHUT_WR); shutdownFlags_ |= SHUT_WRITE; } @@ -1695,17 +1973,99 @@ void AsyncSocket::timeoutExpired() noexcept { if (state_ == StateEnum::CONNECTING) { // connect() timed out // Unregister for I/O events. - AsyncSocketException ex(AsyncSocketException::TIMED_OUT, - "connect timed out"); - failConnect(__func__, ex); + if (connectCallback_) { + AsyncSocketException ex( + AsyncSocketException::TIMED_OUT, + folly::sformat( + "connect timed out after {}ms", connectTimeout_.count())); + failConnect(__func__, ex); + } else { + // we faced a connect error without a connect callback, which could + // happen due to TFO. + AsyncSocketException ex( + AsyncSocketException::TIMED_OUT, "write timed out during connection"); + failWrite(__func__, ex); + } } else { // a normal write operation timed out - assert(state_ == StateEnum::ESTABLISHED); - AsyncSocketException ex(AsyncSocketException::TIMED_OUT, "write timed out"); + AsyncSocketException ex( + AsyncSocketException::TIMED_OUT, + folly::sformat("write timed out after {}ms", sendTimeout_)); failWrite(__func__, ex); } } +ssize_t AsyncSocket::tfoSendMsg(int fd, struct msghdr* msg, int msg_flags) { + return detail::tfo_sendmsg(fd, msg, msg_flags); +} + +AsyncSocket::WriteResult +AsyncSocket::sendSocketMessage(int fd, struct msghdr* msg, int msg_flags) { + ssize_t totalWritten = 0; + if (state_ == StateEnum::FAST_OPEN) { + sockaddr_storage addr; + auto len = addr_.getAddress(&addr); + msg->msg_name = &addr; + msg->msg_namelen = len; + totalWritten = tfoSendMsg(fd_, msg, msg_flags); + if (totalWritten >= 0) { + tfoFinished_ = true; + state_ = StateEnum::ESTABLISHED; + // We schedule this asynchrously so that we don't end up + // invoking initial read or write while a write is in progress. + scheduleInitialReadWrite(); + } else if (errno == EINPROGRESS) { + VLOG(4) << "TFO falling back to connecting"; + // A normal sendmsg doesn't return EINPROGRESS, however + // TFO might fallback to connecting if there is no + // cookie. + state_ = StateEnum::CONNECTING; + try { + scheduleConnectTimeout(); + registerForConnectEvents(); + } catch (const AsyncSocketException& ex) { + return WriteResult( + WRITE_ERROR, std::make_unique(ex)); + } + // Let's fake it that no bytes were written and return an errno. + errno = EAGAIN; + totalWritten = -1; + } else if (errno == EOPNOTSUPP) { + // Try falling back to connecting. + VLOG(4) << "TFO not supported"; + state_ = StateEnum::CONNECTING; + try { + int ret = socketConnect((const sockaddr*)&addr, len); + if (ret == 0) { + // connect succeeded immediately + // Treat this like no data was written. + state_ = StateEnum::ESTABLISHED; + scheduleInitialReadWrite(); + } + // If there was no exception during connections, + // we would return that no bytes were written. + errno = EAGAIN; + totalWritten = -1; + } catch (const AsyncSocketException& ex) { + return WriteResult( + WRITE_ERROR, std::make_unique(ex)); + } + } else if (errno == EAGAIN) { + // Normally sendmsg would indicate that the write would block. + // However in the fast open case, it would indicate that sendmsg + // fell back to a connect. This is a return code from connect() + // instead, and is an error condition indicating no fds available. + return WriteResult( + WRITE_ERROR, + std::make_unique( + AsyncSocketException::UNKNOWN, "No more free local ports")); + } + } else { + totalWritten = ::sendmsg(fd, msg, msg_flags); + } + return WriteResult(totalWritten); +} + AsyncSocket::WriteResult AsyncSocket::performWrite( const iovec* vec, uint32_t count, @@ -1721,28 +2081,31 @@ AsyncSocket::WriteResult AsyncSocket::performWrite( msg.msg_namelen = 0; msg.msg_iov = const_cast(vec); msg.msg_iovlen = std::min(count, kIovMax); - msg.msg_control = nullptr; - msg.msg_controllen = 0; msg.msg_flags = 0; + msg.msg_controllen = sendMsgParamCallback_->getAncillaryDataSize(flags); + CHECK_GE(AsyncSocket::SendMsgParamsCallback::maxAncillaryDataSize, + msg.msg_controllen); - int msg_flags = MSG_DONTWAIT; - -#ifdef MSG_NOSIGNAL // Linux-only - msg_flags |= MSG_NOSIGNAL; - if (isSet(flags, WriteFlags::CORK)) { - // MSG_MORE tells the kernel we have more data to send, so wait for us to - // give it the rest of the data rather than immediately sending a partial - // frame, even when TCP_NODELAY is enabled. - msg_flags |= MSG_MORE; - } -#endif - if (isSet(flags, WriteFlags::EOR)) { - // marks that this is the last byte of a record (response) - msg_flags |= MSG_EOR; + if (msg.msg_controllen != 0) { + msg.msg_control = reinterpret_cast(alloca(msg.msg_controllen)); + sendMsgParamCallback_->getAncillaryData(flags, msg.msg_control); + } else { + msg.msg_control = nullptr; } - ssize_t totalWritten = ::sendmsg(fd_, &msg, msg_flags); + int msg_flags = sendMsgParamCallback_->getFlags(flags); + + auto writeResult = sendSocketMessage(fd_, &msg, msg_flags); + auto totalWritten = writeResult.writeReturn; if (totalWritten < 0) { - if (errno == EAGAIN) { + bool tryAgain = (errno == EAGAIN); +#ifdef __APPLE__ + // Apple has a bug where doing a second write on a socket which we + // have opened with TFO causes an ENOTCONN to be thrown. However the + // socket is really connected, so treat ENOTCONN as a EAGAIN until + // this bug is fixed. + tryAgain |= (errno == ENOTCONN); +#endif + if (!writeResult.exception && tryAgain) { // TCP buffer is full; we can't write any more data right now. *countWritten = 0; *partialWritten = 0; @@ -1751,14 +2114,14 @@ AsyncSocket::WriteResult AsyncSocket::performWrite( // error *countWritten = 0; *partialWritten = 0; - return WriteResult(WRITE_ERROR); + return writeResult; } appBytesWritten_ += totalWritten; uint32_t bytesWritten; uint32_t n; - for (bytesWritten = totalWritten, n = 0; n < count; ++n) { + for (bytesWritten = uint32_t(totalWritten), n = 0; n < count; ++n) { const iovec* v = vec + n; if (v->iov_len > bytesWritten) { // Partial write finished in the middle of this iovec @@ -1767,7 +2130,7 @@ AsyncSocket::WriteResult AsyncSocket::performWrite( return WriteResult(totalWritten); } - bytesWritten -= v->iov_len; + bytesWritten -= uint32_t(v->iov_len); } assert(bytesWritten == 0); @@ -1797,7 +2160,8 @@ bool AsyncSocket::updateEventRegistration() { // Always register for persistent events, so we don't have to re-register // after being called back. - if (!ioHandler_.registerHandler(eventFlags_ | EventHandler::PERSIST)) { + if (!ioHandler_.registerHandler( + uint16_t(eventFlags_ | EventHandler::PERSIST))) { eventFlags_ = EventHandler::NONE; // we're not registered after error AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR, withAddr("failed to update AsyncSocket event registration")); @@ -1841,12 +2205,7 @@ void AsyncSocket::startFail() { } } -void AsyncSocket::finishFail() { - assert(state_ == StateEnum::ERROR); - assert(getDestructorGuardCount() > 0); - - AsyncSocketException ex(AsyncSocketException::INTERNAL_ERROR, - withAddr("socket closing after error")); +void AsyncSocket::invokeAllErrors(const AsyncSocketException& ex) { invokeConnectErr(ex); failAllWrites(ex); @@ -1857,6 +2216,22 @@ void AsyncSocket::finishFail() { } } +void AsyncSocket::finishFail() { + assert(state_ == StateEnum::ERROR); + assert(getDestructorGuardCount() > 0); + + AsyncSocketException ex( + AsyncSocketException::INTERNAL_ERROR, + withAddr("socket closing after error")); + invokeAllErrors(ex); +} + +void AsyncSocket::finishFail(const AsyncSocketException& ex) { + assert(state_ == StateEnum::ERROR); + assert(getDestructorGuardCount() > 0); + invokeAllErrors(ex); +} + void AsyncSocket::fail(const char* fn, const AsyncSocketException& ex) { VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state=" << state_ << " host=" << addr_.describe() @@ -1874,7 +2249,7 @@ void AsyncSocket::failConnect(const char* fn, const AsyncSocketException& ex) { startFail(); invokeConnectErr(ex); - finishFail(); + finishFail(ex); } void AsyncSocket::failRead(const char* fn, const AsyncSocketException& ex) { @@ -1893,6 +2268,23 @@ void AsyncSocket::failRead(const char* fn, const AsyncSocketException& ex) { finishFail(); } +void AsyncSocket::failErrMessageRead(const char* fn, + const AsyncSocketException& ex) { + VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state=" + << state_ << " host=" << addr_.describe() + << "): failed while reading message in " << fn << "(): " + << ex.what(); + startFail(); + + if (errMessageCallback_ != nullptr) { + ErrMessageCallback* callback = errMessageCallback_; + errMessageCallback_ = nullptr; + callback->errMessageError(ex); + } + + finishFail(); +} + void AsyncSocket::failWrite(const char* fn, const AsyncSocketException& ex) { VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ << ", state=" << state_ << " host=" << addr_.describe() @@ -1952,7 +2344,7 @@ void AsyncSocket::failAllWrites(const AsyncSocketException& ex) { void AsyncSocket::invalidState(ConnectCallback* callback) { VLOG(5) << "AsyncSocket(this=" << this << ", fd=" << fd_ - << "): connect() called in invalid state " << state_; + << "): connect() called in invalid state " << state_; /* * The invalidState() methods don't use the normal failure mechanisms, @@ -1980,6 +2372,29 @@ void AsyncSocket::invalidState(ConnectCallback* callback) { } } +void AsyncSocket::invalidState(ErrMessageCallback* callback) { + VLOG(4) << "AsyncSocket(this=" << this << ", fd=" << fd_ + << "): setErrMessageCB(" << callback + << ") called in invalid state " << state_; + + AsyncSocketException ex( + AsyncSocketException::NOT_OPEN, + msgErrQueueSupported + ? "setErrMessageCB() called with socket in invalid state" + : "This platform does not support socket error message notifications"); + if (state_ == StateEnum::CLOSED || state_ == StateEnum::ERROR) { + if (callback) { + callback->errMessageError(ex); + } + } else { + startFail(); + if (callback) { + callback->errMessageError(ex); + } + finishFail(); + } +} + void AsyncSocket::invokeConnectErr(const AsyncSocketException& ex) { connectEndTime_ = std::chrono::steady_clock::now(); if (connectCallback_) {