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.
19 #include <folly/Optional.h>
20 #include <folly/SocketAddress.h>
21 #include <folly/detail/SocketFastOpen.h>
22 #include <folly/io/IOBuf.h>
23 #include <folly/io/ShutdownSocketSet.h>
24 #include <folly/io/async/AsyncSocketException.h>
25 #include <folly/io/async/AsyncTimeout.h>
26 #include <folly/io/async/AsyncTransport.h>
27 #include <folly/io/async/DelayedDestruction.h>
28 #include <folly/io/async/EventHandler.h>
29 #include <folly/portability/Sockets.h>
31 #include <sys/types.h>
40 * A class for performing asynchronous I/O on a socket.
42 * AsyncSocket allows users to asynchronously wait for data on a socket, and
43 * to asynchronously send data.
45 * The APIs for reading and writing are intentionally asymmetric. Waiting for
46 * data to read is a persistent API: a callback is installed, and is notified
47 * whenever new data is available. It continues to be notified of new events
48 * until it is uninstalled.
50 * AsyncSocket does not provide read timeout functionality, because it
51 * typically cannot determine when the timeout should be active. Generally, a
52 * timeout should only be enabled when processing is blocked waiting on data
53 * from the remote endpoint. For server sockets, the timeout should not be
54 * active if the server is currently processing one or more outstanding
55 * requests for this socket. For client sockets, the timeout should not be
56 * active if there are no requests pending on the socket. Additionally, if a
57 * client has multiple pending requests, it will ususally want a separate
58 * timeout for each request, rather than a single read timeout.
60 * The write API is fairly intuitive: a user can request to send a block of
61 * data, and a callback will be informed once the entire block has been
62 * transferred to the kernel, or on error. AsyncSocket does provide a send
63 * timeout, since most callers want to give up if the remote end stops
64 * responding and no further progress can be made sending the data.
67 #if defined __linux__ && !defined SO_NO_TRANSPARENT_TLS
68 #define SO_NO_TRANSPARENT_TLS 200
71 #if defined __linux__ && !defined SO_NO_TSOCKS
72 #define SO_NO_TSOCKS 201
76 // We do a dynamic_cast on this, in
77 // AsyncTransportWrapper::getUnderlyingTransport so be safe and
78 // force displacements for it. See:
79 // https://msdn.microsoft.com/en-us/library/7sf3txa8.aspx
80 #pragma vtordisp(push, 2)
82 class AsyncSocket : virtual public AsyncTransportWrapper {
84 typedef std::unique_ptr<AsyncSocket, Destructor> UniquePtr;
86 class ConnectCallback {
88 virtual ~ConnectCallback() = default;
91 * connectSuccess() will be invoked when the connection has been
92 * successfully established.
94 virtual void connectSuccess() noexcept = 0;
97 * connectErr() will be invoked if the connection attempt fails.
99 * @param ex An exception describing the error that occurred.
101 virtual void connectErr(const AsyncSocketException& ex)
105 class EvbChangeCallback {
107 virtual ~EvbChangeCallback() = default;
109 // Called when the socket has been attached to a new EVB
110 // and is called from within that EVB thread
111 virtual void evbAttached(AsyncSocket* socket) = 0;
113 // Called when the socket is detached from an EVB and
114 // is called from the EVB thread being detached
115 virtual void evbDetached(AsyncSocket* socket) = 0;
119 * This interface is implemented only for platforms supporting
120 * per-socket error queues.
122 class ErrMessageCallback {
124 virtual ~ErrMessageCallback() = default;
127 * errMessage() will be invoked when kernel puts a message to
128 * the error queue associated with the socket.
130 * @param cmsg Reference to cmsghdr structure describing
131 * a message read from error queue associated
135 errMessage(const cmsghdr& cmsg) noexcept = 0;
138 * errMessageError() will be invoked if an error occurs reading a message
139 * from the socket error stream.
141 * @param ex An exception describing the error that occurred.
143 virtual void errMessageError(const AsyncSocketException& ex) noexcept = 0;
146 class SendMsgParamsCallback {
148 virtual ~SendMsgParamsCallback() = default;
151 * getFlags() will be invoked to retrieve the desired flags to be passed
152 * to ::sendmsg() system call. This method was intentionally declared
153 * non-virtual, so there is no way to override it. Instead feel free to
154 * override getFlagsImpl(flags, defaultFlags) method instead, and enjoy
155 * the convenience of defaultFlags passed there.
157 * @param flags Write flags requested for the given write operation
159 int getFlags(folly::WriteFlags flags, bool zeroCopyEnabled) noexcept {
160 return getFlagsImpl(flags, getDefaultFlags(flags, zeroCopyEnabled));
164 * getAncillaryData() will be invoked to initialize ancillary data
165 * buffer referred by "msg_control" field of msghdr structure passed to
166 * ::sendmsg() system call. The function assumes that the size of buffer
167 * is not smaller than the value returned by getAncillaryDataSize() method
168 * for the same combination of flags.
170 * @param flags Write flags requested for the given write operation
171 * @param data Pointer to ancillary data buffer to initialize.
173 virtual void getAncillaryData(
174 folly::WriteFlags /*flags*/,
175 void* /*data*/) noexcept {}
178 * getAncillaryDataSize() will be invoked to retrieve the size of
179 * ancillary data buffer which should be passed to ::sendmsg() system call
181 * @param flags Write flags requested for the given write operation
183 virtual uint32_t getAncillaryDataSize(folly::WriteFlags /*flags*/)
188 static const size_t maxAncillaryDataSize{0x5000};
192 * getFlagsImpl() will be invoked by getFlags(folly::WriteFlags flags)
193 * method to retrieve the flags to be passed to ::sendmsg() system call.
194 * SendMsgParamsCallback::getFlags() is calling this method, and returns
195 * its results directly to the caller in AsyncSocket.
196 * Classes inheriting from SendMsgParamsCallback are welcome to override
197 * this method to force SendMsgParamsCallback to return its own set
200 * @param flags Write flags requested for the given write operation
201 * @param defaultflags A set of message flags returned by getDefaultFlags()
202 * method for the given "flags" mask.
204 virtual int getFlagsImpl(folly::WriteFlags /*flags*/, int defaultFlags) {
209 * getDefaultFlags() will be invoked by getFlags(folly::WriteFlags flags)
210 * to retrieve the default set of flags, and pass them to getFlagsImpl(...)
212 * @param flags Write flags requested for the given write operation
214 int getDefaultFlags(folly::WriteFlags flags, bool zeroCopyEnabled) noexcept;
217 explicit AsyncSocket();
219 * Create a new unconnected AsyncSocket.
221 * connect() must later be called on this socket to establish a connection.
223 explicit AsyncSocket(EventBase* evb);
225 void setShutdownSocketSet(const std::weak_ptr<ShutdownSocketSet>& wSS);
228 * Create a new AsyncSocket and begin the connection process.
230 * @param evb EventBase that will manage this socket.
231 * @param address The address to connect to.
232 * @param connectTimeout Optional timeout in milliseconds for the connection
235 AsyncSocket(EventBase* evb,
236 const folly::SocketAddress& address,
237 uint32_t connectTimeout = 0);
240 * Create a new AsyncSocket and begin the connection process.
242 * @param evb EventBase that will manage this socket.
243 * @param ip IP address to connect to (dotted-quad).
244 * @param port Destination port in host byte order.
245 * @param connectTimeout Optional timeout in milliseconds for the connection
248 AsyncSocket(EventBase* evb,
249 const std::string& ip,
251 uint32_t connectTimeout = 0);
254 * Create a AsyncSocket from an already connected socket file descriptor.
256 * Note that while AsyncSocket enables TCP_NODELAY for sockets it creates
257 * when connecting, it does not change the socket options when given an
258 * existing file descriptor. If callers want TCP_NODELAY enabled when using
259 * this version of the constructor, they need to explicitly call
260 * setNoDelay(true) after the constructor returns.
262 * @param evb EventBase that will manage this socket.
263 * @param fd File descriptor to take over (should be a connected socket).
264 * @param zeroCopyBufId Zerocopy buf id to start with.
266 AsyncSocket(EventBase* evb, int fd, uint32_t zeroCopyBufId = 0);
269 * Create an AsyncSocket from a different, already connected AsyncSocket.
271 * Similar to AsyncSocket(evb, fd) when fd was previously owned by an
274 explicit AsyncSocket(AsyncSocket::UniquePtr);
277 * Helper function to create a shared_ptr<AsyncSocket>.
279 * This passes in the correct destructor object, since AsyncSocket's
280 * destructor is protected and cannot be invoked directly.
282 static std::shared_ptr<AsyncSocket> newSocket(EventBase* evb) {
283 return std::shared_ptr<AsyncSocket>(new AsyncSocket(evb),
288 * Helper function to create a shared_ptr<AsyncSocket>.
290 static std::shared_ptr<AsyncSocket> newSocket(
292 const folly::SocketAddress& address,
293 uint32_t connectTimeout = 0) {
294 return std::shared_ptr<AsyncSocket>(
295 new AsyncSocket(evb, address, connectTimeout),
300 * Helper function to create a shared_ptr<AsyncSocket>.
302 static std::shared_ptr<AsyncSocket> newSocket(
304 const std::string& ip,
306 uint32_t connectTimeout = 0) {
307 return std::shared_ptr<AsyncSocket>(
308 new AsyncSocket(evb, ip, port, connectTimeout),
313 * Helper function to create a shared_ptr<AsyncSocket>.
315 static std::shared_ptr<AsyncSocket> newSocket(EventBase* evb, int fd) {
316 return std::shared_ptr<AsyncSocket>(new AsyncSocket(evb, fd),
321 * Destroy the socket.
323 * AsyncSocket::destroy() must be called to destroy the socket.
324 * The normal destructor is private, and should not be invoked directly.
325 * This prevents callers from deleting a AsyncSocket while it is invoking a
328 void destroy() override;
331 * Get the EventBase used by this socket.
333 EventBase* getEventBase() const override {
338 * Get the file descriptor used by the AsyncSocket.
340 virtual int getFd() const {
345 * Extract the file descriptor from the AsyncSocket.
347 * This will immediately cause any installed callbacks to be invoked with an
348 * error. The AsyncSocket may no longer be used after the file descriptor
349 * has been extracted.
351 * This method should be used with care as the resulting fd is not guaranteed
352 * to perfectly reflect the state of the AsyncSocket (security state,
353 * pre-received data, etc.).
355 * Returns the file descriptor. The caller assumes ownership of the
356 * descriptor, and it will not be closed when the AsyncSocket is destroyed.
358 virtual int detachFd();
361 * Uniquely identifies a handle to a socket option value. Each
362 * combination of level and option name corresponds to one socket
367 bool operator<(const OptionKey& other) const {
368 if (level == other.level) {
369 return optname < other.optname;
371 return level < other.level;
373 int apply(int fd, int val) const {
374 return setsockopt(fd, level, optname, &val, sizeof(val));
380 // Maps from a socket option key to its value
381 typedef std::map<OptionKey, int> OptionMap;
383 static const OptionMap emptyOptionMap;
384 static const folly::SocketAddress& anyAddress();
387 * Initiate a connection.
389 * @param callback The callback to inform when the connection attempt
391 * @param address The address to connect to.
392 * @param timeout A timeout value, in milliseconds. If the connection
393 * does not succeed within this period,
394 * callback->connectError() will be invoked.
396 virtual void connect(
397 ConnectCallback* callback,
398 const folly::SocketAddress& address,
400 const OptionMap& options = emptyOptionMap,
401 const folly::SocketAddress& bindAddr = anyAddress()) noexcept;
404 ConnectCallback* callback,
405 const std::string& ip,
408 const OptionMap& options = emptyOptionMap) noexcept;
411 * If a connect request is in-flight, cancels it and closes the socket
412 * immediately. Otherwise, this is a no-op.
414 * This does not invoke any connection related callbacks. Call this to
415 * prevent any connect callback while cleaning up, etc.
417 void cancelConnect();
420 * Set the send timeout.
422 * If write requests do not make any progress for more than the specified
423 * number of milliseconds, fail all pending writes and close the socket.
425 * If write requests are currently pending when setSendTimeout() is called,
426 * the timeout interval is immediately restarted using the new value.
428 * (See the comments for AsyncSocket for an explanation of why AsyncSocket
429 * provides setSendTimeout() but not setRecvTimeout().)
431 * @param milliseconds The timeout duration, in milliseconds. If 0, no
432 * timeout will be used.
434 void setSendTimeout(uint32_t milliseconds) override;
437 * Get the send timeout.
439 * @return Returns the current send timeout, in milliseconds. A return value
440 * of 0 indicates that no timeout is set.
442 uint32_t getSendTimeout() const override {
447 * Set the maximum number of reads to execute from the underlying
448 * socket each time the EventBase detects that new ingress data is
449 * available. The default is unlimited, but callers can use this method
450 * to limit the amount of data read from the socket per event loop
453 * @param maxReads Maximum number of reads per data-available event;
454 * a value of zero means unlimited.
456 void setMaxReadsPerEvent(uint16_t maxReads) {
457 maxReadsPerEvent_ = maxReads;
461 * Get the maximum number of reads this object will execute from
462 * the underlying socket each time the EventBase detects that new
463 * ingress data is available.
465 * @returns Maximum number of reads per data-available event; a value
466 * of zero means unlimited.
468 uint16_t getMaxReadsPerEvent() const {
469 return maxReadsPerEvent_;
473 * Set a pointer to ErrMessageCallback implementation which will be
474 * receiving notifications for messages posted to the error queue
475 * associated with the socket.
476 * ErrMessageCallback is implemented only for platforms with
477 * per-socket error message queus support (recvmsg() system call must
481 virtual void setErrMessageCB(ErrMessageCallback* callback);
484 * Get a pointer to ErrMessageCallback implementation currently
485 * registered with this socket.
488 virtual ErrMessageCallback* getErrMessageCallback() const;
491 * Set a pointer to SendMsgParamsCallback implementation which
492 * will be used to form ::sendmsg() system call parameters
495 virtual void setSendMsgParamCB(SendMsgParamsCallback* callback);
498 * Get a pointer to SendMsgParamsCallback implementation currently
499 * registered with this socket.
502 virtual SendMsgParamsCallback* getSendMsgParamsCB() const;
504 // Read and write methods
505 void setReadCB(ReadCallback* callback) override;
506 ReadCallback* getReadCallback() const override;
508 bool setZeroCopy(bool enable);
509 bool getZeroCopy() const {
510 return zeroCopyEnabled_;
513 uint32_t getZeroCopyBufId() const {
514 return zeroCopyBufId_;
517 void write(WriteCallback* callback, const void* buf, size_t bytes,
518 WriteFlags flags = WriteFlags::NONE) override;
519 void writev(WriteCallback* callback, const iovec* vec, size_t count,
520 WriteFlags flags = WriteFlags::NONE) override;
521 void writeChain(WriteCallback* callback,
522 std::unique_ptr<folly::IOBuf>&& buf,
523 WriteFlags flags = WriteFlags::NONE) override;
526 virtual void writeRequest(WriteRequest* req);
527 void writeRequestReady() {
531 // Methods inherited from AsyncTransport
532 void close() override;
533 void closeNow() override;
534 void closeWithReset() override;
535 void shutdownWrite() override;
536 void shutdownWriteNow() override;
538 bool readable() const override;
539 bool writable() const override;
540 bool isPending() const override;
541 virtual bool hangup() const;
542 bool good() const override;
543 bool error() const override;
544 void attachEventBase(EventBase* eventBase) override;
545 void detachEventBase() override;
546 bool isDetachable() const override;
548 void getLocalAddress(
549 folly::SocketAddress* address) const override;
551 folly::SocketAddress* address) const override;
553 bool isEorTrackingEnabled() const override {
557 void setEorTracking(bool track) override {
561 bool connecting() const override {
562 return (state_ == StateEnum::CONNECTING);
565 virtual bool isClosedByPeer() const {
566 return (state_ == StateEnum::CLOSED &&
567 (readErr_ == READ_EOF || readErr_ == READ_ERROR));
570 virtual bool isClosedBySelf() const {
571 return (state_ == StateEnum::CLOSED &&
572 (readErr_ != READ_EOF && readErr_ != READ_ERROR));
575 size_t getAppBytesWritten() const override {
576 return appBytesWritten_;
579 size_t getRawBytesWritten() const override {
580 return getAppBytesWritten();
583 size_t getAppBytesReceived() const override {
584 return appBytesReceived_;
587 size_t getRawBytesReceived() const override {
588 return getAppBytesReceived();
591 std::chrono::nanoseconds getConnectTime() const {
592 return connectEndTime_ - connectStartTime_;
595 std::chrono::milliseconds getConnectTimeout() const {
596 return connectTimeout_;
599 bool getTFOAttempted() const {
600 return tfoAttempted_;
604 * Returns whether or not the attempt to use TFO
605 * finished successfully. This does not necessarily
606 * mean TFO worked, just that trying to use TFO
609 bool getTFOFinished() const {
614 * Returns whether or not TFO attempt succeded on this
616 * For servers this is pretty straightforward API and can
617 * be invoked right after the connection is accepted. This API
618 * will perform one syscall.
619 * This API is a bit tricky to use for clients, since clients
620 * only know this for sure after the SYN-ACK is returned. So it's
621 * appropriate to call this only after the first application
622 * data is read from the socket when the caller knows that
623 * the SYN has been ACKed by the server.
625 bool getTFOSucceded() const;
627 // Methods controlling socket options
630 * Force writes to be transmitted immediately.
632 * This controls the TCP_NODELAY socket option. When enabled, TCP segments
633 * are sent as soon as possible, even if it is not a full frame of data.
634 * When disabled, the data may be buffered briefly to try and wait for a full
637 * By default, TCP_NODELAY is enabled for AsyncSocket objects.
639 * This method will fail if the socket is not currently open.
641 * @return Returns 0 if the TCP_NODELAY flag was successfully updated,
642 * or a non-zero errno value on error.
644 int setNoDelay(bool noDelay);
648 * Set the FD_CLOEXEC flag so that the socket will be closed if the program
649 * later forks and execs.
651 void setCloseOnExec();
654 * Set the Flavor of Congestion Control to be used for this Socket
655 * Please check '/lib/modules/<kernel>/kernel/net/ipv4' for tcp_*.ko
656 * first to make sure the module is available for plugging in
657 * Alternatively you can choose from net.ipv4.tcp_allowed_congestion_control
659 int setCongestionFlavor(const std::string &cname);
662 * Forces ACKs to be sent immediately
664 * @return Returns 0 if the TCP_QUICKACK flag was successfully updated,
665 * or a non-zero errno value on error.
667 int setQuickAck(bool quickack);
670 * Set the send bufsize
672 int setSendBufSize(size_t bufsize);
675 * Set the recv bufsize
677 int setRecvBufSize(size_t bufsize);
680 * Sets a specific tcp personality
681 * Available only on kernels 3.2 and greater
683 #define SO_SET_NAMESPACE 41
684 int setTCPProfile(int profd);
687 * Generic API for reading a socket option.
689 * @param level same as the "level" parameter in getsockopt().
690 * @param optname same as the "optname" parameter in getsockopt().
691 * @param optval pointer to the variable in which the option value should
693 * @param optlen value-result argument, initially containing the size of
694 * the buffer pointed to by optval, and modified on return
695 * to indicate the actual size of the value returned.
696 * @return same as the return value of getsockopt().
698 template <typename T>
699 int getSockOpt(int level, int optname, T* optval, socklen_t* optlen) {
700 return getsockopt(fd_, level, optname, (void*) optval, optlen);
704 * Generic API for setting a socket option.
706 * @param level same as the "level" parameter in getsockopt().
707 * @param optname same as the "optname" parameter in getsockopt().
708 * @param optval the option value to set.
709 * @return same as the return value of setsockopt().
711 template <typename T>
712 int setSockOpt(int level, int optname, const T *optval) {
713 return setsockopt(fd_, level, optname, optval, sizeof(T));
717 * Virtual method for reading a socket option returning integer
718 * value, which is the most typical case. Convenient for overriding
721 * @param level same as the "level" parameter in getsockopt().
722 * @param optname same as the "optname" parameter in getsockopt().
723 * @param optval same as "optval" parameter in getsockopt().
724 * @param optlen same as "optlen" parameter in getsockopt().
725 * @return same as the return value of getsockopt().
728 getSockOptVirtual(int level, int optname, void* optval, socklen_t* optlen) {
729 return getsockopt(fd_, level, optname, optval, optlen);
733 * Virtual method for setting a socket option accepting integer
734 * value, which is the most typical case. Convenient for overriding
737 * @param level same as the "level" parameter in setsockopt().
738 * @param optname same as the "optname" parameter in setsockopt().
739 * @param optval same as "optval" parameter in setsockopt().
740 * @param optlen same as "optlen" parameter in setsockopt().
741 * @return same as the return value of setsockopt().
743 virtual int setSockOptVirtual(
748 return setsockopt(fd_, level, optname, optval, optlen);
752 * Set pre-received data, to be returned to read callback before any data
755 virtual void setPreReceivedData(std::unique_ptr<IOBuf> data) {
756 if (preReceivedData_) {
757 preReceivedData_->prependChain(std::move(data));
759 preReceivedData_ = std::move(data);
764 * Enables TFO behavior on the AsyncSocket if FOLLY_ALLOW_TFO
768 // No-op if folly does not allow tfo
774 void disableTransparentTls() {
775 noTransparentTls_ = true;
778 void disableTSocks() {
782 enum class StateEnum : uint8_t {
791 void setBufferCallback(BufferCallback* cb);
793 // Callers should set this prior to connecting the socket for the safest
795 void setEvbChangedCallback(std::unique_ptr<EvbChangeCallback> cb) {
796 evbChangeCb_ = std::move(cb);
800 * Attempt to cache the current local and peer addresses (if not already
801 * cached) so that they are available from getPeerAddress() and
802 * getLocalAddress() even after the socket is closed.
804 void cacheAddresses();
807 * Returns true if there is any zero copy write in progress
808 * Needs to be called from within the socket's EVB thread
810 bool isZeroCopyWriteInProgress() const noexcept;
813 * writeReturn is the total number of bytes written, or WRITE_ERROR on error.
814 * If no data has been written, 0 is returned.
815 * exception is a more specific exception that cause a write error.
816 * Not all writes have exceptions associated with them thus writeReturn
817 * should be checked to determine whether the operation resulted in an error.
820 explicit WriteResult(ssize_t ret) : writeReturn(ret) {}
822 WriteResult(ssize_t ret, std::unique_ptr<const AsyncSocketException> e)
823 : writeReturn(ret), exception(std::move(e)) {}
826 std::unique_ptr<const AsyncSocketException> exception;
830 * readReturn is the number of bytes read, or READ_EOF on EOF, or
831 * READ_ERROR on error, or READ_BLOCKING if the operation will
833 * exception is a more specific exception that may have caused a read error.
834 * Not all read errors have exceptions associated with them thus readReturn
835 * should be checked to determine whether the operation resulted in an error.
838 explicit ReadResult(ssize_t ret) : readReturn(ret) {}
840 ReadResult(ssize_t ret, std::unique_ptr<const AsyncSocketException> e)
841 : readReturn(ret), exception(std::move(e)) {}
844 std::unique_ptr<const AsyncSocketException> exception;
848 * A WriteRequest object tracks information about a pending write operation.
852 WriteRequest(AsyncSocket* socket, WriteCallback* callback) :
853 socket_(socket), callback_(callback) {}
855 virtual void start() {}
857 virtual void destroy() = 0;
859 virtual WriteResult performWrite() = 0;
861 virtual void consume() = 0;
863 virtual bool isComplete() = 0;
865 WriteRequest* getNext() const {
869 WriteCallback* getCallback() const {
873 uint32_t getTotalBytesWritten() const {
874 return totalBytesWritten_;
877 void append(WriteRequest* next) {
878 assert(next_ == nullptr);
882 void fail(const char* fn, const AsyncSocketException& ex) {
883 socket_->failWrite(fn, ex);
886 void bytesWritten(size_t count) {
887 totalBytesWritten_ += uint32_t(count);
888 socket_->appBytesWritten_ += count;
892 // protected destructor, to ensure callers use destroy()
893 virtual ~WriteRequest() {}
895 AsyncSocket* socket_; ///< parent socket
896 WriteRequest* next_{nullptr}; ///< pointer to next WriteRequest
897 WriteCallback* callback_; ///< completion callback
898 uint32_t totalBytesWritten_{0}; ///< total bytes written
902 enum ReadResultEnum {
909 enum WriteResultEnum {
914 * Protected destructor.
916 * Users of AsyncSocket must never delete it directly. Instead, invoke
917 * destroy() instead. (See the documentation in DelayedDestruction.h for
920 ~AsyncSocket() override;
922 friend std::ostream& operator << (std::ostream& os, const StateEnum& state);
925 /// shutdownWrite() called, but we are still waiting on writes to drain
926 SHUT_WRITE_PENDING = 0x01,
927 /// writes have been completely shut down
930 * Reads have been shutdown.
932 * At the moment we don't distinguish between remote read shutdown
933 * (received EOF from the remote end) and local read shutdown. We can
934 * only receive EOF when a read callback is set, and we immediately inform
935 * it of the EOF. Therefore there doesn't seem to be any reason to have a
936 * separate state of "received EOF but the local side may still want to
939 * We also don't currently provide any API for only shutting down the read
940 * side of a socket. (This is a no-op as far as TCP is concerned, anyway.)
945 class BytesWriteRequest;
947 class WriteTimeout : public AsyncTimeout {
949 WriteTimeout(AsyncSocket* socket, EventBase* eventBase)
950 : AsyncTimeout(eventBase)
953 void timeoutExpired() noexcept override {
954 socket_->timeoutExpired();
958 AsyncSocket* socket_;
961 class IoHandler : public EventHandler {
963 IoHandler(AsyncSocket* socket, EventBase* eventBase)
964 : EventHandler(eventBase, -1)
966 IoHandler(AsyncSocket* socket, EventBase* eventBase, int fd)
967 : EventHandler(eventBase, fd)
970 void handlerReady(uint16_t events) noexcept override {
971 socket_->ioReady(events);
975 AsyncSocket* socket_;
980 class ImmediateReadCB : public folly::EventBase::LoopCallback {
982 explicit ImmediateReadCB(AsyncSocket* socket) : socket_(socket) {}
983 void runLoopCallback() noexcept override {
984 DestructorGuard dg(socket_);
985 socket_->checkForImmediateRead();
988 AsyncSocket* socket_;
992 * Schedule checkForImmediateRead to be executed in the next loop
995 void scheduleImmediateRead() noexcept {
997 eventBase_->runInLoop(&immediateReadHandler_);
1002 * Schedule handleInitalReadWrite to run in the next iteration.
1004 void scheduleInitialReadWrite() noexcept {
1006 DestructorGuard dg(this);
1007 eventBase_->runInLoop([this, dg] {
1009 handleInitialReadWrite();
1015 // event notification methods
1016 void ioReady(uint16_t events) noexcept;
1017 virtual void checkForImmediateRead() noexcept;
1018 virtual void handleInitialReadWrite() noexcept;
1019 virtual void prepareReadBuffer(void** buf, size_t* buflen);
1020 virtual void handleErrMessages() noexcept;
1021 virtual void handleRead() noexcept;
1022 virtual void handleWrite() noexcept;
1023 virtual void handleConnect() noexcept;
1024 void timeoutExpired() noexcept;
1027 * Attempt to read from the socket.
1029 * @param buf The buffer to read data into.
1030 * @param buflen The length of the buffer.
1032 * @return Returns a read result. See read result for details.
1034 virtual ReadResult performRead(void** buf, size_t* buflen, size_t* offset);
1037 * Populate an iovec array from an IOBuf and attempt to write it.
1039 * @param callback Write completion/error callback.
1040 * @param vec Target iovec array; caller retains ownership.
1041 * @param count Number of IOBufs to write, beginning at start of buf.
1042 * @param buf Chain of iovecs.
1043 * @param flags set of flags for the underlying write calls, like cork
1045 void writeChainImpl(WriteCallback* callback, iovec* vec,
1046 size_t count, std::unique_ptr<folly::IOBuf>&& buf,
1050 * Write as much data as possible to the socket without blocking,
1051 * and queue up any leftover data to send when the socket can
1052 * handle writes again.
1054 * @param callback The callback to invoke when the write is completed.
1055 * @param vec Array of buffers to write; this method will make a
1056 * copy of the vector (but not the buffers themselves)
1057 * if the write has to be completed asynchronously.
1058 * @param count Number of elements in vec.
1059 * @param buf The IOBuf that manages the buffers referenced by
1060 * vec, or a pointer to nullptr if the buffers are not
1061 * associated with an IOBuf. Note that ownership of
1062 * the IOBuf is transferred here; upon completion of
1063 * the write, the AsyncSocket deletes the IOBuf.
1064 * @param flags Set of write flags.
1066 void writeImpl(WriteCallback* callback, const iovec* vec, size_t count,
1067 std::unique_ptr<folly::IOBuf>&& buf,
1068 WriteFlags flags = WriteFlags::NONE);
1071 * Attempt to write to the socket.
1073 * @param vec The iovec array pointing to the buffers to write.
1074 * @param count The length of the iovec array.
1075 * @param flags Set of write flags.
1076 * @param countWritten On return, the value pointed to by this parameter
1077 * will contain the number of iovec entries that were
1079 * @param partialWritten On return, the value pointed to by this parameter
1080 * will contain the number of bytes written in the
1081 * partially written iovec entry.
1083 * @return Returns a WriteResult. See WriteResult for more details.
1085 virtual WriteResult performWrite(
1089 uint32_t* countWritten,
1090 uint32_t* partialWritten);
1093 * Sends the message over the socket using sendmsg
1095 * @param msg Message to send
1096 * @param msg_flags Flags to pass to sendmsg
1098 AsyncSocket::WriteResult
1099 sendSocketMessage(int fd, struct msghdr* msg, int msg_flags);
1101 virtual ssize_t tfoSendMsg(int fd, struct msghdr* msg, int msg_flags);
1103 int socketConnect(const struct sockaddr* addr, socklen_t len);
1105 virtual void scheduleConnectTimeout();
1106 void registerForConnectEvents();
1108 bool updateEventRegistration();
1111 * Update event registration.
1113 * @param enable Flags of events to enable. Set it to 0 if no events
1114 * need to be enabled in this call.
1115 * @param disable Flags of events
1116 * to disable. Set it to 0 if no events need to be disabled in this
1119 * @return true iff the update is successful.
1121 bool updateEventRegistration(uint16_t enable, uint16_t disable);
1123 // Actually close the file descriptor and set it to -1 so we don't
1124 // accidentally close it again.
1127 // error handling methods
1130 void finishFail(const AsyncSocketException& ex);
1131 void invokeAllErrors(const AsyncSocketException& ex);
1132 void fail(const char* fn, const AsyncSocketException& ex);
1133 void failConnect(const char* fn, const AsyncSocketException& ex);
1134 void failRead(const char* fn, const AsyncSocketException& ex);
1135 void failErrMessageRead(const char* fn, const AsyncSocketException& ex);
1136 void failWrite(const char* fn, WriteCallback* callback, size_t bytesWritten,
1137 const AsyncSocketException& ex);
1138 void failWrite(const char* fn, const AsyncSocketException& ex);
1139 void failAllWrites(const AsyncSocketException& ex);
1140 virtual void invokeConnectErr(const AsyncSocketException& ex);
1141 virtual void invokeConnectSuccess();
1142 void invalidState(ConnectCallback* callback);
1143 void invalidState(ErrMessageCallback* callback);
1144 void invalidState(ReadCallback* callback);
1145 void invalidState(WriteCallback* callback);
1147 std::string withAddr(const std::string& s);
1149 void cacheLocalAddress() const;
1150 void cachePeerAddress() const;
1152 bool isZeroCopyRequest(WriteFlags flags);
1154 bool isZeroCopyMsg(const cmsghdr& cmsg) const;
1155 void processZeroCopyMsg(const cmsghdr& cmsg);
1157 uint32_t getNextZeroCopyBufId() {
1158 return zeroCopyBufId_++;
1160 void adjustZeroCopyFlags(folly::WriteFlags& flags);
1161 void addZeroCopyBuf(std::unique_ptr<folly::IOBuf>&& buf);
1162 void addZeroCopyBuf(folly::IOBuf* ptr);
1163 void setZeroCopyBuf(std::unique_ptr<folly::IOBuf>&& buf);
1164 bool containsZeroCopyBuf(folly::IOBuf* ptr);
1165 void releaseZeroCopyBuf(uint32_t id);
1167 // a folly::IOBuf can be used in multiple partial requests
1168 // there is a that maps a buffer id to a raw folly::IOBuf ptr
1169 // and another one that adds a ref count for a folly::IOBuf that is either
1170 // the original ptr or nullptr
1171 uint32_t zeroCopyBufId_{0};
1175 std::unique_ptr<folly::IOBuf> buf_;
1178 std::unordered_map<uint32_t, folly::IOBuf*> idZeroCopyBufPtrMap_;
1179 std::unordered_map<folly::IOBuf*, IOBufInfo> idZeroCopyBufInfoMap_;
1181 StateEnum state_; ///< StateEnum describing current state
1182 uint8_t shutdownFlags_; ///< Shutdown state (ShutdownFlags)
1183 uint16_t eventFlags_; ///< EventBase::HandlerFlags settings
1184 int fd_; ///< The socket file descriptor
1185 mutable folly::SocketAddress addr_; ///< The address we tried to connect to
1186 mutable folly::SocketAddress localAddr_;
1187 ///< The address we are connecting from
1188 uint32_t sendTimeout_; ///< The send timeout, in milliseconds
1189 uint16_t maxReadsPerEvent_; ///< Max reads per event loop iteration
1191 bool isBufferMovable_{false};
1193 int8_t readErr_{READ_NO_ERROR}; ///< The read error encountered, if any
1195 EventBase* eventBase_; ///< The EventBase
1196 WriteTimeout writeTimeout_; ///< A timeout for connect and write
1197 IoHandler ioHandler_; ///< A EventHandler to monitor the fd
1198 ImmediateReadCB immediateReadHandler_; ///< LoopCallback for checking read
1200 ConnectCallback* connectCallback_; ///< ConnectCallback
1201 ErrMessageCallback* errMessageCallback_; ///< TimestampCallback
1202 SendMsgParamsCallback* ///< Callback for retrieving
1203 sendMsgParamCallback_; ///< ::sendmsg() parameters
1204 ReadCallback* readCallback_; ///< ReadCallback
1205 WriteRequest* writeReqHead_; ///< Chain of WriteRequests
1206 WriteRequest* writeReqTail_; ///< End of WriteRequest chain
1207 std::weak_ptr<ShutdownSocketSet> wShutdownSocketSet_;
1208 size_t appBytesReceived_; ///< Num of bytes received from socket
1209 size_t appBytesWritten_; ///< Num of bytes written to socket
1211 // Pre-received data, to be returned to read callback before any data from the
1213 std::unique_ptr<IOBuf> preReceivedData_;
1215 std::chrono::steady_clock::time_point connectStartTime_;
1216 std::chrono::steady_clock::time_point connectEndTime_;
1218 std::chrono::milliseconds connectTimeout_{0};
1220 std::unique_ptr<EvbChangeCallback> evbChangeCb_{nullptr};
1222 BufferCallback* bufferCallback_{nullptr};
1223 bool tfoEnabled_{false};
1224 bool tfoAttempted_{false};
1225 bool tfoFinished_{false};
1226 bool noTransparentTls_{false};
1227 bool noTSocks_{false};
1228 // Whether to track EOR or not.
1229 bool trackEor_{false};
1230 bool zeroCopyEnabled_{false};
1231 bool zeroCopyVal_{false};
1234 #pragma vtordisp(pop)
1237 } // namespace folly