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.
21 #include <folly/io/IOBuf.h>
22 #include <folly/io/async/AsyncSocketBase.h>
23 #include <folly/io/async/DelayedDestruction.h>
24 #include <folly/io/async/EventBase.h>
25 #include <folly/io/async/ssl/OpenSSLPtrTypes.h>
26 #include <folly/portability/OpenSSL.h>
27 #include <folly/portability/SysUio.h>
29 constexpr bool kOpenSslModeMoveBufferOwnership =
30 #ifdef SSL_MODE_MOVE_BUFFER_OWNERSHIP
39 class AsyncSocketException;
44 * flags given by the application for write* calls
46 enum class WriteFlags : uint32_t {
49 * Whether to delay the output until a subsequent non-corked write.
50 * (Note: may not be supported in all subclasses or on all platforms.)
54 * for a socket that has ACK latency enabled, it will cause the kernel
55 * to fire a TCP ESTATS event when the last byte of the given write call
56 * will be acknowledged.
60 * this indicates that only the write side of socket should be shutdown
62 WRITE_SHUTDOWN = 0x04,
68 inline WriteFlags operator|(WriteFlags a, WriteFlags b) {
69 return static_cast<WriteFlags>(
70 static_cast<uint32_t>(a) | static_cast<uint32_t>(b));
74 * compound assignment union operator
76 inline WriteFlags& operator|=(WriteFlags& a, WriteFlags b) {
82 * intersection operator
84 inline WriteFlags operator&(WriteFlags a, WriteFlags b) {
85 return static_cast<WriteFlags>(
86 static_cast<uint32_t>(a) & static_cast<uint32_t>(b));
90 * compound assignment intersection operator
92 inline WriteFlags& operator&=(WriteFlags& a, WriteFlags b) {
100 inline WriteFlags operator~(WriteFlags a) {
101 return static_cast<WriteFlags>(~static_cast<uint32_t>(a));
107 inline WriteFlags unSet(WriteFlags a, WriteFlags b) {
114 inline bool isSet(WriteFlags a, WriteFlags b) {
120 * AsyncTransport defines an asynchronous API for streaming I/O.
122 * This class provides an API to for asynchronously waiting for data
123 * on a streaming transport, and for asynchronously sending data.
125 * The APIs for reading and writing are intentionally asymmetric. Waiting for
126 * data to read is a persistent API: a callback is installed, and is notified
127 * whenever new data is available. It continues to be notified of new events
128 * until it is uninstalled.
130 * AsyncTransport does not provide read timeout functionality, because it
131 * typically cannot determine when the timeout should be active. Generally, a
132 * timeout should only be enabled when processing is blocked waiting on data
133 * from the remote endpoint. For server-side applications, the timeout should
134 * not be active if the server is currently processing one or more outstanding
135 * requests on this transport. For client-side applications, the timeout
136 * should not be active if there are no requests pending on the transport.
137 * Additionally, if a client has multiple pending requests, it will ususally
138 * want a separate timeout for each request, rather than a single read timeout.
140 * The write API is fairly intuitive: a user can request to send a block of
141 * data, and a callback will be informed once the entire block has been
142 * transferred to the kernel, or on error. AsyncTransport does provide a send
143 * timeout, since most callers want to give up if the remote end stops
144 * responding and no further progress can be made sending the data.
146 class AsyncTransport : public DelayedDestruction, public AsyncSocketBase {
148 typedef std::unique_ptr<AsyncTransport, Destructor> UniquePtr;
151 * Close the transport.
153 * This gracefully closes the transport, waiting for all pending write
154 * requests to complete before actually closing the underlying transport.
156 * If a read callback is set, readEOF() will be called immediately. If there
157 * are outstanding write requests, the close will be delayed until all
158 * remaining writes have completed. No new writes may be started after
159 * close() has been called.
161 virtual void close() = 0;
164 * Close the transport immediately.
166 * This closes the transport immediately, dropping any outstanding data
167 * waiting to be written.
169 * If a read callback is set, readEOF() will be called immediately.
170 * If there are outstanding write requests, these requests will be aborted
171 * and writeError() will be invoked immediately on all outstanding write
174 virtual void closeNow() = 0;
177 * Reset the transport immediately.
179 * This closes the transport immediately, sending a reset to the remote peer
180 * if possible to indicate abnormal shutdown.
182 * Note that not all subclasses implement this reset functionality: some
183 * subclasses may treat reset() the same as closeNow(). Subclasses that use
184 * TCP transports should terminate the connection with a TCP reset.
186 virtual void closeWithReset() {
191 * Perform a half-shutdown of the write side of the transport.
193 * The caller should not make any more calls to write() or writev() after
194 * shutdownWrite() is called. Any future write attempts will fail
197 * Not all transport types support half-shutdown. If the underlying
198 * transport does not support half-shutdown, it will fully shutdown both the
199 * read and write sides of the transport. (Fully shutting down the socket is
200 * better than doing nothing at all, since the caller may rely on the
201 * shutdownWrite() call to notify the other end of the connection that no
202 * more data can be read.)
204 * If there is pending data still waiting to be written on the transport,
205 * the actual shutdown will be delayed until the pending data has been
208 * Note: There is no corresponding shutdownRead() equivalent. Simply
209 * uninstall the read callback if you wish to stop reading. (On TCP sockets
210 * at least, shutting down the read side of the socket is a no-op anyway.)
212 virtual void shutdownWrite() = 0;
215 * Perform a half-shutdown of the write side of the transport.
217 * shutdownWriteNow() is identical to shutdownWrite(), except that it
218 * immediately performs the shutdown, rather than waiting for pending writes
219 * to complete. Any pending write requests will be immediately failed when
220 * shutdownWriteNow() is called.
222 virtual void shutdownWriteNow() = 0;
225 * Determine if transport is open and ready to read or write.
227 * Note that this function returns false on EOF; you must also call error()
228 * to distinguish between an EOF and an error.
230 * @return true iff the transport is open and ready, false otherwise.
232 virtual bool good() const = 0;
235 * Determine if the transport is readable or not.
237 * @return true iff the transport is readable, false otherwise.
239 virtual bool readable() const = 0;
242 * Determine if the there is pending data on the transport.
244 * @return true iff the if the there is pending data, false otherwise.
246 virtual bool isPending() const {
251 * Determine if transport is connected to the endpoint
253 * @return false iff the transport is connected, otherwise true
255 virtual bool connecting() const = 0;
258 * Determine if an error has occurred with this transport.
260 * @return true iff an error has occurred (not EOF).
262 virtual bool error() const = 0;
265 * Attach the transport to a EventBase.
267 * This may only be called if the transport is not currently attached to a
268 * EventBase (by an earlier call to detachEventBase()).
270 * This method must be invoked in the EventBase's thread.
272 virtual void attachEventBase(EventBase* eventBase) = 0;
275 * Detach the transport from its EventBase.
277 * This may only be called when the transport is idle and has no reads or
278 * writes pending. Once detached, the transport may not be used again until
279 * it is re-attached to a EventBase by calling attachEventBase().
281 * This method must be called from the current EventBase's thread.
283 virtual void detachEventBase() = 0;
286 * Determine if the transport can be detached.
288 * This method must be called from the current EventBase's thread.
290 virtual bool isDetachable() const = 0;
293 * Set the send timeout.
295 * If write requests do not make any progress for more than the specified
296 * number of milliseconds, fail all pending writes and close the transport.
298 * If write requests are currently pending when setSendTimeout() is called,
299 * the timeout interval is immediately restarted using the new value.
301 * @param milliseconds The timeout duration, in milliseconds. If 0, no
302 * timeout will be used.
304 virtual void setSendTimeout(uint32_t milliseconds) = 0;
307 * Get the send timeout.
309 * @return Returns the current send timeout, in milliseconds. A return value
310 * of 0 indicates that no timeout is set.
312 virtual uint32_t getSendTimeout() const = 0;
315 * Get the address of the local endpoint of this transport.
317 * This function may throw AsyncSocketException on error.
319 * @param address The local address will be stored in the specified
322 virtual void getLocalAddress(SocketAddress* address) const = 0;
325 * Get the address of the remote endpoint to which this transport is
328 * This function may throw AsyncSocketException on error.
330 * @return Return the local address
332 SocketAddress getLocalAddress() const {
334 getLocalAddress(&addr);
338 virtual void getAddress(SocketAddress* address) const {
339 getLocalAddress(address);
343 * Get the address of the remote endpoint to which this transport is
346 * This function may throw AsyncSocketException on error.
348 * @param address The remote endpoint's address will be stored in the
349 * specified SocketAddress.
351 virtual void getPeerAddress(SocketAddress* address) const = 0;
354 * Get the address of the remote endpoint to which this transport is
357 * This function may throw AsyncSocketException on error.
359 * @return Return the remote endpoint's address
361 SocketAddress getPeerAddress() const {
363 getPeerAddress(&addr);
368 * Get the certificate used to authenticate the peer.
370 virtual ssl::X509UniquePtr getPeerCert() const { return nullptr; }
373 * The local certificate used for this connection. May be null
375 virtual const X509* getSelfCert() const {
380 * @return True iff end of record tracking is enabled
382 virtual bool isEorTrackingEnabled() const = 0;
384 virtual void setEorTracking(bool track) = 0;
386 virtual size_t getAppBytesWritten() const = 0;
387 virtual size_t getRawBytesWritten() const = 0;
388 virtual size_t getAppBytesReceived() const = 0;
389 virtual size_t getRawBytesReceived() const = 0;
391 class BufferCallback {
393 virtual ~BufferCallback() {}
394 virtual void onEgressBuffered() = 0;
395 virtual void onEgressBufferCleared() = 0;
399 * Callback class to signal when a transport that did not have replay
400 * protection gains replay protection. This is needed for 0-RTT security
403 class ReplaySafetyCallback {
405 virtual ~ReplaySafetyCallback() = default;
408 * Called when the transport becomes replay safe.
410 virtual void onReplaySafe() = 0;
414 * False if the transport does not have replay protection, but will in the
417 virtual bool isReplaySafe() const { return true; }
420 * Set the ReplaySafeCallback on this transport.
422 * This should only be called if isReplaySafe() returns false.
424 virtual void setReplaySafetyCallback(ReplaySafetyCallback* callback) {
426 CHECK(false) << "setReplaySafetyCallback() not supported";
431 virtual ~AsyncTransport() = default;
438 virtual ~ReadCallback() = default;
441 * When data becomes available, getReadBuffer() will be invoked to get the
442 * buffer into which data should be read.
444 * This method allows the ReadCallback to delay buffer allocation until
445 * data becomes available. This allows applications to manage large
446 * numbers of idle connections, without having to maintain a separate read
447 * buffer for each idle connection.
449 * It is possible that in some cases, getReadBuffer() may be called
450 * multiple times before readDataAvailable() is invoked. In this case, the
451 * data will be written to the buffer returned from the most recent call to
452 * readDataAvailable(). If the previous calls to readDataAvailable()
453 * returned different buffers, the ReadCallback is responsible for ensuring
454 * that they are not leaked.
456 * If getReadBuffer() throws an exception, returns a nullptr buffer, or
457 * returns a 0 length, the ReadCallback will be uninstalled and its
458 * readError() method will be invoked.
460 * getReadBuffer() is not allowed to change the transport state before it
461 * returns. (For example, it should never uninstall the read callback, or
462 * set a different read callback.)
464 * @param bufReturn getReadBuffer() should update *bufReturn to contain the
465 * address of the read buffer. This parameter will never
467 * @param lenReturn getReadBuffer() should update *lenReturn to contain the
468 * maximum number of bytes that may be written to the read
469 * buffer. This parameter will never be nullptr.
471 virtual void getReadBuffer(void** bufReturn, size_t* lenReturn) = 0;
474 * readDataAvailable() will be invoked when data has been successfully read
475 * into the buffer returned by the last call to getReadBuffer().
477 * The read callback remains installed after readDataAvailable() returns.
478 * It must be explicitly uninstalled to stop receiving read events.
479 * getReadBuffer() will be called at least once before each call to
480 * readDataAvailable(). getReadBuffer() will also be called before any
483 * @param len The number of bytes placed in the buffer.
486 virtual void readDataAvailable(size_t len) noexcept = 0;
489 * When data becomes available, isBufferMovable() will be invoked to figure
490 * out which API will be used, readBufferAvailable() or
491 * readDataAvailable(). If isBufferMovable() returns true, that means
492 * ReadCallback supports the IOBuf ownership transfer and
493 * readBufferAvailable() will be used. Otherwise, not.
495 * By default, isBufferMovable() always return false. If
496 * readBufferAvailable() is implemented and to be invoked, You should
497 * overwrite isBufferMovable() and return true in the inherited class.
499 * This method allows the AsyncSocket/AsyncSSLSocket do buffer allocation by
500 * itself until data becomes available. Compared with the pre/post buffer
501 * allocation in getReadBuffer()/readDataAvailabe(), readBufferAvailable()
502 * has two advantages. First, this can avoid memcpy. E.g., in
503 * AsyncSSLSocket, the decrypted data was copied from the openssl internal
504 * buffer to the readbuf buffer. With the buffer ownership transfer, the
505 * internal buffer can be directly "moved" to ReadCallback. Second, the
506 * memory allocation can be more precise. The reason is
507 * AsyncSocket/AsyncSSLSocket can allocate the memory of precise size
508 * because they have more context about the available data than
509 * ReadCallback. Think about the getReadBuffer() pre-allocate 4072 bytes
510 * buffer, but the available data is always 16KB (max OpenSSL record size).
513 virtual bool isBufferMovable() noexcept {
518 * Suggested buffer size, allocated for read operations,
519 * if callback is movable and supports folly::IOBuf
522 virtual size_t maxBufferSize() const {
523 return 64 * 1024; // 64K
527 * readBufferAvailable() will be invoked when data has been successfully
530 * Note that only either readBufferAvailable() or readDataAvailable() will
531 * be invoked according to the return value of isBufferMovable(). The timing
532 * and aftereffect of readBufferAvailable() are the same as
533 * readDataAvailable()
535 * @param readBuf The unique pointer of read buffer.
538 virtual void readBufferAvailable(std::unique_ptr<IOBuf> /*readBuf*/)
542 * readEOF() will be invoked when the transport is closed.
544 * The read callback will be automatically uninstalled immediately before
545 * readEOF() is invoked.
547 virtual void readEOF() noexcept = 0;
550 * readError() will be invoked if an error occurs reading from the
553 * The read callback will be automatically uninstalled immediately before
554 * readError() is invoked.
556 * @param ex An exception describing the error that occurred.
558 virtual void readErr(const AsyncSocketException& ex) noexcept = 0;
561 // Read methods that aren't part of AsyncTransport.
562 virtual void setReadCB(ReadCallback* callback) = 0;
563 virtual ReadCallback* getReadCallback() const = 0;
566 virtual ~AsyncReader() = default;
571 class WriteCallback {
573 virtual ~WriteCallback() = default;
576 * writeSuccess() will be invoked when all of the data has been
577 * successfully written.
579 * Note that this mainly signals that the buffer containing the data to
580 * write is no longer needed and may be freed or re-used. It does not
581 * guarantee that the data has been fully transmitted to the remote
582 * endpoint. For example, on socket-based transports, writeSuccess() only
583 * indicates that the data has been given to the kernel for eventual
586 virtual void writeSuccess() noexcept = 0;
589 * writeError() will be invoked if an error occurs writing the data.
591 * @param bytesWritten The number of bytes that were successfull
592 * @param ex An exception describing the error that occurred.
594 virtual void writeErr(size_t bytesWritten,
595 const AsyncSocketException& ex) noexcept = 0;
598 // Write methods that aren't part of AsyncTransport
599 virtual void write(WriteCallback* callback, const void* buf, size_t bytes,
600 WriteFlags flags = WriteFlags::NONE) = 0;
601 virtual void writev(WriteCallback* callback, const iovec* vec, size_t count,
602 WriteFlags flags = WriteFlags::NONE) = 0;
603 virtual void writeChain(WriteCallback* callback,
604 std::unique_ptr<IOBuf>&& buf,
605 WriteFlags flags = WriteFlags::NONE) = 0;
608 virtual ~AsyncWriter() = default;
611 // Transitional intermediate interface. This is deprecated.
612 // Wrapper around folly::AsyncTransport, that includes read/write callbacks
613 class AsyncTransportWrapper : virtual public AsyncTransport,
614 virtual public AsyncReader,
615 virtual public AsyncWriter {
617 using UniquePtr = std::unique_ptr<AsyncTransportWrapper, Destructor>;
619 // Alias for inherited members from AsyncReader and AsyncWriter
620 // to keep compatibility.
621 using ReadCallback = AsyncReader::ReadCallback;
622 using WriteCallback = AsyncWriter::WriteCallback;
623 virtual void setReadCB(ReadCallback* callback) override = 0;
624 virtual ReadCallback* getReadCallback() const override = 0;
625 virtual void write(WriteCallback* callback, const void* buf, size_t bytes,
626 WriteFlags flags = WriteFlags::NONE) override = 0;
627 virtual void writev(WriteCallback* callback, const iovec* vec, size_t count,
628 WriteFlags flags = WriteFlags::NONE) override = 0;
629 virtual void writeChain(WriteCallback* callback,
630 std::unique_ptr<IOBuf>&& buf,
631 WriteFlags flags = WriteFlags::NONE) override = 0;
633 * The transport wrapper may wrap another transport. This returns the
634 * transport that is wrapped. It returns nullptr if there is no wrapped
637 virtual const AsyncTransportWrapper* getWrappedTransport() const {
642 * In many cases when we need to set socket properties or otherwise access the
643 * underlying transport from a wrapped transport. This method allows access to
644 * the derived classes of the underlying transport.
647 const T* getUnderlyingTransport() const {
648 const AsyncTransportWrapper* current = this;
650 auto sock = dynamic_cast<const T*>(current);
654 current = current->getWrappedTransport();
660 T* getUnderlyingTransport() {
661 return const_cast<T*>(static_cast<const AsyncTransportWrapper*>(this)
662 ->getUnderlyingTransport<T>());
666 * Return the application protocol being used by the underlying transport
667 * protocol. This is useful for transports which are used to tunnel other
670 virtual std::string getApplicationProtocol() noexcept {
675 * Returns the name of the security protocol being used.
677 virtual std::string getSecurityProtocol() const { return ""; }