2 * Copyright 2016 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/SysUio.h>
28 #include <openssl/ssl.h>
30 constexpr bool kOpenSslModeMoveBufferOwnership =
31 #ifdef SSL_MODE_MOVE_BUFFER_OWNERSHIP
40 class AsyncSocketException;
45 * flags given by the application for write* calls
47 enum class WriteFlags : uint32_t {
50 * Whether to delay the output until a subsequent non-corked write.
51 * (Note: may not be supported in all subclasses or on all platforms.)
55 * for a socket that has ACK latency enabled, it will cause the kernel
56 * to fire a TCP ESTATS event when the last byte of the given write call
57 * will be acknowledged.
61 * this indicates that only the write side of socket should be shutdown
63 WRITE_SHUTDOWN = 0x04,
69 inline WriteFlags operator|(WriteFlags a, WriteFlags b) {
70 return static_cast<WriteFlags>(
71 static_cast<uint32_t>(a) | static_cast<uint32_t>(b));
75 * intersection operator
77 inline WriteFlags operator&(WriteFlags a, WriteFlags b) {
78 return static_cast<WriteFlags>(
79 static_cast<uint32_t>(a) & static_cast<uint32_t>(b));
85 inline WriteFlags operator~(WriteFlags a) {
86 return static_cast<WriteFlags>(~static_cast<uint32_t>(a));
92 inline WriteFlags unSet(WriteFlags a, WriteFlags b) {
99 inline bool isSet(WriteFlags a, WriteFlags b) {
105 * AsyncTransport defines an asynchronous API for streaming I/O.
107 * This class provides an API to for asynchronously waiting for data
108 * on a streaming transport, and for asynchronously sending data.
110 * The APIs for reading and writing are intentionally asymmetric. Waiting for
111 * data to read is a persistent API: a callback is installed, and is notified
112 * whenever new data is available. It continues to be notified of new events
113 * until it is uninstalled.
115 * AsyncTransport does not provide read timeout functionality, because it
116 * typically cannot determine when the timeout should be active. Generally, a
117 * timeout should only be enabled when processing is blocked waiting on data
118 * from the remote endpoint. For server-side applications, the timeout should
119 * not be active if the server is currently processing one or more outstanding
120 * requests on this transport. For client-side applications, the timeout
121 * should not be active if there are no requests pending on the transport.
122 * Additionally, if a client has multiple pending requests, it will ususally
123 * want a separate timeout for each request, rather than a single read timeout.
125 * The write API is fairly intuitive: a user can request to send a block of
126 * data, and a callback will be informed once the entire block has been
127 * transferred to the kernel, or on error. AsyncTransport does provide a send
128 * timeout, since most callers want to give up if the remote end stops
129 * responding and no further progress can be made sending the data.
131 class AsyncTransport : public DelayedDestruction, public AsyncSocketBase {
133 typedef std::unique_ptr<AsyncTransport, Destructor> UniquePtr;
136 * Close the transport.
138 * This gracefully closes the transport, waiting for all pending write
139 * requests to complete before actually closing the underlying transport.
141 * If a read callback is set, readEOF() will be called immediately. If there
142 * are outstanding write requests, the close will be delayed until all
143 * remaining writes have completed. No new writes may be started after
144 * close() has been called.
146 virtual void close() = 0;
149 * Close the transport immediately.
151 * This closes the transport immediately, dropping any outstanding data
152 * waiting to be written.
154 * If a read callback is set, readEOF() will be called immediately.
155 * If there are outstanding write requests, these requests will be aborted
156 * and writeError() will be invoked immediately on all outstanding write
159 virtual void closeNow() = 0;
162 * Reset the transport immediately.
164 * This closes the transport immediately, sending a reset to the remote peer
165 * if possible to indicate abnormal shutdown.
167 * Note that not all subclasses implement this reset functionality: some
168 * subclasses may treat reset() the same as closeNow(). Subclasses that use
169 * TCP transports should terminate the connection with a TCP reset.
171 virtual void closeWithReset() {
176 * Perform a half-shutdown of the write side of the transport.
178 * The caller should not make any more calls to write() or writev() after
179 * shutdownWrite() is called. Any future write attempts will fail
182 * Not all transport types support half-shutdown. If the underlying
183 * transport does not support half-shutdown, it will fully shutdown both the
184 * read and write sides of the transport. (Fully shutting down the socket is
185 * better than doing nothing at all, since the caller may rely on the
186 * shutdownWrite() call to notify the other end of the connection that no
187 * more data can be read.)
189 * If there is pending data still waiting to be written on the transport,
190 * the actual shutdown will be delayed until the pending data has been
193 * Note: There is no corresponding shutdownRead() equivalent. Simply
194 * uninstall the read callback if you wish to stop reading. (On TCP sockets
195 * at least, shutting down the read side of the socket is a no-op anyway.)
197 virtual void shutdownWrite() = 0;
200 * Perform a half-shutdown of the write side of the transport.
202 * shutdownWriteNow() is identical to shutdownWrite(), except that it
203 * immediately performs the shutdown, rather than waiting for pending writes
204 * to complete. Any pending write requests will be immediately failed when
205 * shutdownWriteNow() is called.
207 virtual void shutdownWriteNow() = 0;
210 * Determine if transport is open and ready to read or write.
212 * Note that this function returns false on EOF; you must also call error()
213 * to distinguish between an EOF and an error.
215 * @return true iff the transport is open and ready, false otherwise.
217 virtual bool good() const = 0;
220 * Determine if the transport is readable or not.
222 * @return true iff the transport is readable, false otherwise.
224 virtual bool readable() const = 0;
227 * Determine if the there is pending data on the transport.
229 * @return true iff the if the there is pending data, false otherwise.
231 virtual bool isPending() const {
236 * Determine if transport is connected to the endpoint
238 * @return false iff the transport is connected, otherwise true
240 virtual bool connecting() const = 0;
243 * Determine if an error has occurred with this transport.
245 * @return true iff an error has occurred (not EOF).
247 virtual bool error() const = 0;
250 * Attach the transport to a EventBase.
252 * This may only be called if the transport is not currently attached to a
253 * EventBase (by an earlier call to detachEventBase()).
255 * This method must be invoked in the EventBase's thread.
257 virtual void attachEventBase(EventBase* eventBase) = 0;
260 * Detach the transport from its EventBase.
262 * This may only be called when the transport is idle and has no reads or
263 * writes pending. Once detached, the transport may not be used again until
264 * it is re-attached to a EventBase by calling attachEventBase().
266 * This method must be called from the current EventBase's thread.
268 virtual void detachEventBase() = 0;
271 * Determine if the transport can be detached.
273 * This method must be called from the current EventBase's thread.
275 virtual bool isDetachable() const = 0;
278 * Set the send timeout.
280 * If write requests do not make any progress for more than the specified
281 * number of milliseconds, fail all pending writes and close the transport.
283 * If write requests are currently pending when setSendTimeout() is called,
284 * the timeout interval is immediately restarted using the new value.
286 * @param milliseconds The timeout duration, in milliseconds. If 0, no
287 * timeout will be used.
289 virtual void setSendTimeout(uint32_t milliseconds) = 0;
292 * Get the send timeout.
294 * @return Returns the current send timeout, in milliseconds. A return value
295 * of 0 indicates that no timeout is set.
297 virtual uint32_t getSendTimeout() const = 0;
300 * Get the address of the local endpoint of this transport.
302 * This function may throw AsyncSocketException on error.
304 * @param address The local address will be stored in the specified
307 virtual void getLocalAddress(SocketAddress* address) const = 0;
309 virtual void getAddress(SocketAddress* address) const {
310 getLocalAddress(address);
314 * Get the address of the remote endpoint to which this transport is
317 * This function may throw AsyncSocketException on error.
319 * @param address The remote endpoint's address will be stored in the
320 * specified SocketAddress.
322 virtual void getPeerAddress(SocketAddress* address) const = 0;
325 * Get the certificate used to authenticate the peer.
327 virtual ssl::X509UniquePtr getPeerCert() const { return nullptr; }
330 * The local certificate used for this connection. May be null
332 virtual const X509* getSelfCert() const {
337 * @return True iff end of record tracking is enabled
339 virtual bool isEorTrackingEnabled() const = 0;
341 virtual void setEorTracking(bool track) = 0;
343 virtual size_t getAppBytesWritten() const = 0;
344 virtual size_t getRawBytesWritten() const = 0;
345 virtual size_t getAppBytesReceived() const = 0;
346 virtual size_t getRawBytesReceived() const = 0;
348 class BufferCallback {
350 virtual ~BufferCallback() {}
351 virtual void onEgressBuffered() = 0;
352 virtual void onEgressBufferCleared() = 0;
356 * Callback class to signal when a transport that did not have replay
357 * protection gains replay protection. This is needed for 0-RTT security
360 class ReplaySafetyCallback {
362 virtual ~ReplaySafetyCallback() = default;
365 * Called when the transport becomes replay safe.
367 virtual void onReplaySafe() = 0;
371 * False if the transport does not have replay protection, but will in the
374 virtual bool isReplaySafe() const { return true; }
377 * Set the ReplaySafeCallback on this transport.
379 * This should only be called if isReplaySafe() returns false.
381 virtual void setReplaySafetyCallback(ReplaySafetyCallback* callback) {
383 CHECK(false) << "setReplaySafetyCallback() not supported";
388 virtual ~AsyncTransport() = default;
395 virtual ~ReadCallback() = default;
398 * When data becomes available, getReadBuffer() will be invoked to get the
399 * buffer into which data should be read.
401 * This method allows the ReadCallback to delay buffer allocation until
402 * data becomes available. This allows applications to manage large
403 * numbers of idle connections, without having to maintain a separate read
404 * buffer for each idle connection.
406 * It is possible that in some cases, getReadBuffer() may be called
407 * multiple times before readDataAvailable() is invoked. In this case, the
408 * data will be written to the buffer returned from the most recent call to
409 * readDataAvailable(). If the previous calls to readDataAvailable()
410 * returned different buffers, the ReadCallback is responsible for ensuring
411 * that they are not leaked.
413 * If getReadBuffer() throws an exception, returns a nullptr buffer, or
414 * returns a 0 length, the ReadCallback will be uninstalled and its
415 * readError() method will be invoked.
417 * getReadBuffer() is not allowed to change the transport state before it
418 * returns. (For example, it should never uninstall the read callback, or
419 * set a different read callback.)
421 * @param bufReturn getReadBuffer() should update *bufReturn to contain the
422 * address of the read buffer. This parameter will never
424 * @param lenReturn getReadBuffer() should update *lenReturn to contain the
425 * maximum number of bytes that may be written to the read
426 * buffer. This parameter will never be nullptr.
428 virtual void getReadBuffer(void** bufReturn, size_t* lenReturn) = 0;
431 * readDataAvailable() will be invoked when data has been successfully read
432 * into the buffer returned by the last call to getReadBuffer().
434 * The read callback remains installed after readDataAvailable() returns.
435 * It must be explicitly uninstalled to stop receiving read events.
436 * getReadBuffer() will be called at least once before each call to
437 * readDataAvailable(). getReadBuffer() will also be called before any
440 * @param len The number of bytes placed in the buffer.
443 virtual void readDataAvailable(size_t len) noexcept = 0;
446 * When data becomes available, isBufferMovable() will be invoked to figure
447 * out which API will be used, readBufferAvailable() or
448 * readDataAvailable(). If isBufferMovable() returns true, that means
449 * ReadCallback supports the IOBuf ownership transfer and
450 * readBufferAvailable() will be used. Otherwise, not.
452 * By default, isBufferMovable() always return false. If
453 * readBufferAvailable() is implemented and to be invoked, You should
454 * overwrite isBufferMovable() and return true in the inherited class.
456 * This method allows the AsyncSocket/AsyncSSLSocket do buffer allocation by
457 * itself until data becomes available. Compared with the pre/post buffer
458 * allocation in getReadBuffer()/readDataAvailabe(), readBufferAvailable()
459 * has two advantages. First, this can avoid memcpy. E.g., in
460 * AsyncSSLSocket, the decrypted data was copied from the openssl internal
461 * buffer to the readbuf buffer. With the buffer ownership transfer, the
462 * internal buffer can be directly "moved" to ReadCallback. Second, the
463 * memory allocation can be more precise. The reason is
464 * AsyncSocket/AsyncSSLSocket can allocate the memory of precise size
465 * because they have more context about the available data than
466 * ReadCallback. Think about the getReadBuffer() pre-allocate 4072 bytes
467 * buffer, but the available data is always 16KB (max OpenSSL record size).
470 virtual bool isBufferMovable() noexcept {
475 * Suggested buffer size, allocated for read operations,
476 * if callback is movable and supports folly::IOBuf
479 virtual size_t maxBufferSize() const {
480 return 64 * 1024; // 64K
484 * readBufferAvailable() will be invoked when data has been successfully
487 * Note that only either readBufferAvailable() or readDataAvailable() will
488 * be invoked according to the return value of isBufferMovable(). The timing
489 * and aftereffect of readBufferAvailable() are the same as
490 * readDataAvailable()
492 * @param readBuf The unique pointer of read buffer.
495 virtual void readBufferAvailable(std::unique_ptr<IOBuf> /*readBuf*/)
499 * readEOF() will be invoked when the transport is closed.
501 * The read callback will be automatically uninstalled immediately before
502 * readEOF() is invoked.
504 virtual void readEOF() noexcept = 0;
507 * readError() will be invoked if an error occurs reading from the
510 * The read callback will be automatically uninstalled immediately before
511 * readError() is invoked.
513 * @param ex An exception describing the error that occurred.
515 virtual void readErr(const AsyncSocketException& ex) noexcept = 0;
518 // Read methods that aren't part of AsyncTransport.
519 virtual void setReadCB(ReadCallback* callback) = 0;
520 virtual ReadCallback* getReadCallback() const = 0;
523 virtual ~AsyncReader() = default;
528 class WriteCallback {
530 virtual ~WriteCallback() = default;
533 * writeSuccess() will be invoked when all of the data has been
534 * successfully written.
536 * Note that this mainly signals that the buffer containing the data to
537 * write is no longer needed and may be freed or re-used. It does not
538 * guarantee that the data has been fully transmitted to the remote
539 * endpoint. For example, on socket-based transports, writeSuccess() only
540 * indicates that the data has been given to the kernel for eventual
543 virtual void writeSuccess() noexcept = 0;
546 * writeError() will be invoked if an error occurs writing the data.
548 * @param bytesWritten The number of bytes that were successfull
549 * @param ex An exception describing the error that occurred.
551 virtual void writeErr(size_t bytesWritten,
552 const AsyncSocketException& ex) noexcept = 0;
555 // Write methods that aren't part of AsyncTransport
556 virtual void write(WriteCallback* callback, const void* buf, size_t bytes,
557 WriteFlags flags = WriteFlags::NONE) = 0;
558 virtual void writev(WriteCallback* callback, const iovec* vec, size_t count,
559 WriteFlags flags = WriteFlags::NONE) = 0;
560 virtual void writeChain(WriteCallback* callback,
561 std::unique_ptr<IOBuf>&& buf,
562 WriteFlags flags = WriteFlags::NONE) = 0;
565 virtual ~AsyncWriter() = default;
568 // Transitional intermediate interface. This is deprecated.
569 // Wrapper around folly::AsyncTransport, that includes read/write callbacks
570 class AsyncTransportWrapper : virtual public AsyncTransport,
571 virtual public AsyncReader,
572 virtual public AsyncWriter {
574 using UniquePtr = std::unique_ptr<AsyncTransportWrapper, Destructor>;
576 // Alias for inherited members from AsyncReader and AsyncWriter
577 // to keep compatibility.
578 using ReadCallback = AsyncReader::ReadCallback;
579 using WriteCallback = AsyncWriter::WriteCallback;
580 virtual void setReadCB(ReadCallback* callback) override = 0;
581 virtual ReadCallback* getReadCallback() const override = 0;
582 virtual void write(WriteCallback* callback, const void* buf, size_t bytes,
583 WriteFlags flags = WriteFlags::NONE) override = 0;
584 virtual void writev(WriteCallback* callback, const iovec* vec, size_t count,
585 WriteFlags flags = WriteFlags::NONE) override = 0;
586 virtual void writeChain(WriteCallback* callback,
587 std::unique_ptr<IOBuf>&& buf,
588 WriteFlags flags = WriteFlags::NONE) override = 0;
590 * The transport wrapper may wrap another transport. This returns the
591 * transport that is wrapped. It returns nullptr if there is no wrapped
594 virtual const AsyncTransportWrapper* getWrappedTransport() const {
599 * In many cases when we need to set socket properties or otherwise access the
600 * underlying transport from a wrapped transport. This method allows access to
601 * the derived classes of the underlying transport.
604 const T* getUnderlyingTransport() const {
605 const AsyncTransportWrapper* current = this;
607 auto sock = dynamic_cast<const T*>(current);
611 current = current->getWrappedTransport();
617 T* getUnderlyingTransport() {
618 return const_cast<T*>(static_cast<const AsyncTransportWrapper*>(this)
619 ->getUnderlyingTransport<T>());
623 * Return the application protocol being used by the underlying transport
624 * protocol. This is useful for transports which are used to tunnel other
627 virtual std::string getApplicationProtocol() noexcept {
632 * Returns the name of the security protocol being used.
634 virtual std::string getSecurityProtocol() const { return ""; }