Add iouring folly support, refactor the async IO

Summary: Add iouring folly support, refactor the async IO

Reviewed By: kevin-vigor

Differential Revision: D17834511

fbshipit-source-id: e20c876a32730549f305334fd5eed02cccf23638

CMake/FindLibUring.cmake

+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+find_path(LIBURING_INCLUDE_DIR NAMES liburing.h)
+mark_as_advanced(LIBURING_INCLUDE_DIR)
+
+find_library(LIBURING_LIBRARY NAMES uring)
+mark_as_advanced(LIBURING_LIBRARY)
+
+include(FindPackageHandleStandardArgs)
+FIND_PACKAGE_HANDLE_STANDARD_ARGS(
+        LIBURING
+        REQUIRED_VARS LIBURING_LIBRARY LIBURING_INCLUDE_DIR)
+
+if(LIBURING_FOUND)
+  set(LIBURING_LIBRARIES ${LIBURING_LIBRARY})
+  set(LIBURING_INCLUDE_DIRS ${LIBURING_INCLUDE_DIR})
+endif()

CMake/folly-deps.cmake

@@ -113,6 +113,10 @@ find_package(LibAIO)
 list(APPEND FOLLY_LINK_LIBRARIES ${LIBAIO_LIBRARIES})
 list(APPEND FOLLY_INCLUDE_DIRECTORIES ${LIBAIO_INCLUDE_DIRS})
 
+find_package(LibUring)
+list(APPEND FOLLY_LINK_LIBRARIES ${LIBURING_LIBRARIES})
+list(APPEND FOLLY_INCLUDE_DIRECTORIES ${LIBURING_INCLUDE_DIRS})
+
 find_package(Libsodium)
 list(APPEND FOLLY_LINK_LIBRARIES ${LIBSODIUM_LIBRARIES})
 list(APPEND FOLLY_INCLUDE_DIRECTORIES ${LIBSODIUM_INCLUDE_DIRS})

CMakeLists.txt

@@ -203,6 +203,23 @@ if (NOT ${LIBAIO_FOUND})
     ${FOLLY_DIR}/experimental/io/AsyncIO.h
   )
 endif()
+if (NOT ${LIBURING_FOUND})
+  list(REMOVE_ITEM files
+    ${FOLLY_DIR}/experimental/io/IoUring.cpp
+  )
+  list(REMOVE_ITEM hfiles
+    ${FOLLY_DIR}/experimental/io/IoUring.h
+  )
+endif()
+if (NOT ${LIBAIO_FOUND} AND NOT ${LIBURING_FOUND})
+  list(REMOVE_ITEM files
+    ${FOLLY_DIR}/experimental/io/AsyncBase.cpp
+  )
+  list(REMOVE_ITEM hfiles
+    ${FOLLY_DIR}/experimental/io/AsyncBase.h
+  )
+endif()
+
 if (${LIBSODIUM_FOUND})
   string(FIND "${CMAKE_LIBRARY_ARCHITECTURE}" "x86_64" IS_X86_64_ARCH)
   if (${IS_X86_64_ARCH} STREQUAL "-1")

folly/experimental/io/AsyncBase.cpp

+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/experimental/io/AsyncBase.h>
+
+#include <sys/eventfd.h>
+#include <cerrno>
+#include <ostream>
+#include <stdexcept>
+#include <string>
+
+#include <boost/intrusive/parent_from_member.hpp>
+#include <glog/logging.h>
+
+#include <folly/Exception.h>
+#include <folly/Format.h>
+#include <folly/Likely.h>
+#include <folly/String.h>
+#include <folly/portability/Unistd.h>
+
+namespace folly {
+
+AsyncBaseOp::AsyncBaseOp(NotificationCallback cb)
+    : cb_(std::move(cb)), state_(State::UNINITIALIZED), result_(-EINVAL) {}
+
+void AsyncBaseOp::reset(NotificationCallback cb) {
+  CHECK_NE(state_, State::PENDING);
+  cb_ = std::move(cb);
+  state_ = State::UNINITIALIZED;
+  result_ = -EINVAL;
+}
+
+AsyncBaseOp::~AsyncBaseOp() {
+  CHECK_NE(state_, State::PENDING);
+}
+
+void AsyncBaseOp::start() {
+  DCHECK_EQ(state_, State::INITIALIZED);
+  state_ = State::PENDING;
+}
+
+void AsyncBaseOp::complete(ssize_t result) {
+  DCHECK_EQ(state_, State::PENDING);
+  state_ = State::COMPLETED;
+  result_ = result;
+  if (cb_) {
+    cb_(this);
+  }
+}
+
+void AsyncBaseOp::cancel() {
+  DCHECK_EQ(state_, State::PENDING);
+  state_ = State::CANCELED;
+}
+
+ssize_t AsyncBaseOp::result() const {
+  CHECK_EQ(state_, State::COMPLETED);
+  return result_;
+}
+
+void AsyncBaseOp::init() {
+  CHECK_EQ(state_, State::UNINITIALIZED);
+  state_ = State::INITIALIZED;
+}
+
+std::string AsyncBaseOp::fd2name(int fd) {
+  std::string path = folly::to<std::string>("/proc/self/fd/", fd);
+  char link[PATH_MAX];
+  const ssize_t length =
+      std::max<ssize_t>(readlink(path.c_str(), link, PATH_MAX), 0);
+  return path.assign(link, length);
+}
+
+AsyncBase::AsyncBase(size_t capacity, PollMode pollMode) : capacity_(capacity) {
+  CHECK_GT(capacity_, 0);
+  completed_.reserve(capacity_);
+  if (pollMode == POLLABLE) {
+    pollFd_ = eventfd(0, EFD_NONBLOCK);
+    checkUnixError(pollFd_, "AsyncBase: eventfd creation failed");
+  }
+}
+
+AsyncBase::~AsyncBase() {
+  CHECK_EQ(pending_, 0);
+  if (pollFd_ != -1) {
+    CHECK_ERR(close(pollFd_));
+  }
+}
+
+void AsyncBase::decrementPending() {
+  auto p =
+      pending_.fetch_add(static_cast<size_t>(-1), std::memory_order_acq_rel);
+  DCHECK_GE(p, 1);
+}
+
+void AsyncBase::submit(Op* op) {
+  CHECK_EQ(op->state(), Op::State::INITIALIZED);
+  initializeContext(); // on demand
+
+  // We can increment past capacity, but we'll clean up after ourselves.
+  auto p = pending_.fetch_add(1, std::memory_order_acq_rel);
+  if (p >= capacity_) {
+    decrementPending();
+    throw std::range_error("AsyncBase: too many pending requests");
+  }
+
+  int rc = submitOne(op);
+
+  if (rc < 0) {
+    decrementPending();
+    throwSystemErrorExplicit(-rc, "AsyncBase: io_submit failed");
+  }
+  submitted_++;
+  DCHECK_EQ(rc, 1);
+  op->start();
+}
+
+Range<AsyncBase::Op**> AsyncBase::wait(size_t minRequests) {
+  CHECK(isInit());
+  CHECK_EQ(pollFd_, -1) << "wait() only allowed on non-pollable object";
+  auto p = pending_.load(std::memory_order_acquire);
+  CHECK_LE(minRequests, p);
+  return doWait(WaitType::COMPLETE, minRequests, p, completed_);
+}
+
+Range<AsyncBase::Op**> AsyncBase::cancel() {
+  CHECK(isInit());
+  auto p = pending_.load(std::memory_order_acquire);
+  return doWait(WaitType::CANCEL, p, p, canceled_);
+}
+
+Range<AsyncBase::Op**> AsyncBase::pollCompleted() {
+  CHECK(isInit());
+  CHECK_NE(pollFd_, -1) << "pollCompleted() only allowed on pollable object";
+  uint64_t numEvents;
+  // This sets the eventFd counter to 0, see
+  // http://www.kernel.org/doc/man-pages/online/pages/man2/eventfd.2.html
+  ssize_t rc;
+  do {
+    rc = ::read(pollFd_, &numEvents, 8);
+  } while (rc == -1 && errno == EINTR);
+  if (UNLIKELY(rc == -1 && errno == EAGAIN)) {
+    return Range<Op**>(); // nothing completed
+  }
+  checkUnixError(rc, "AsyncBase: read from event fd failed");
+  DCHECK_EQ(rc, 8);
+
+  DCHECK_GT(numEvents, 0);
+  DCHECK_LE(numEvents, pending_);
+
+  // Don't reap more than numEvents, as we've just reset the counter to 0.
+  return doWait(WaitType::COMPLETE, numEvents, numEvents, completed_);
+}
+
+AsyncBaseQueue::AsyncBaseQueue(AsyncBase* asyncBase) : asyncBase_(asyncBase) {}
+
+AsyncBaseQueue::~AsyncBaseQueue() {
+  CHECK_EQ(asyncBase_->pending(), 0);
+}
+
+void AsyncBaseQueue::submit(AsyncBaseOp* op) {
+  submit([op]() { return op; });
+}
+
+void AsyncBaseQueue::submit(OpFactory op) {
+  queue_.push_back(op);
+  maybeDequeue();
+}
+
+void AsyncBaseQueue::onCompleted(AsyncBaseOp* /* op */) {
+  maybeDequeue();
+}
+
+void AsyncBaseQueue::maybeDequeue() {
+  while (!queue_.empty() && asyncBase_->pending() < asyncBase_->capacity()) {
+    auto& opFactory = queue_.front();
+    auto op = opFactory();
+    queue_.pop_front();
+
+    // Interpose our completion callback
+    auto& nextCb = op->notificationCallback();
+    op->setNotificationCallback([this, nextCb](AsyncBaseOp* op2) {
+      this->onCompleted(op2);
+      if (nextCb) {
+        nextCb(op2);
+      }
+    });
+
+    asyncBase_->submit(op);
+  }
+}
+
+// debugging helpers:
+
+namespace {
+
+#define X(c) \
+  case c:    \
+    return #c
+
+const char* asyncIoOpStateToString(AsyncBaseOp::State state) {
+  switch (state) {
+    X(AsyncBaseOp::State::UNINITIALIZED);
+    X(AsyncBaseOp::State::INITIALIZED);
+    X(AsyncBaseOp::State::PENDING);
+    X(AsyncBaseOp::State::COMPLETED);
+    X(AsyncBaseOp::State::CANCELED);
+  }
+  return "<INVALID AsyncBaseOp::State>";
+}
+#undef X
+} // namespace
+
+std::ostream& operator<<(std::ostream& os, const AsyncBaseOp& op) {
+  op.toStream(os);
+  return os;
+}
+
+std::ostream& operator<<(std::ostream& os, AsyncBaseOp::State state) {
+  return os << asyncIoOpStateToString(state);
+}
+
+} // namespace folly

folly/experimental/io/AsyncBase.h

+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <sys/types.h>
+
+#include <atomic>
+#include <cstdint>
+#include <deque>
+#include <functional>
+#include <iosfwd>
+#include <mutex>
+#include <utility>
+#include <vector>
+
+#include <folly/Portability.h>
+#include <folly/Range.h>
+#include <folly/portability/SysUio.h>
+
+namespace folly {
+class AsyncIOOp;
+class IoUringOp;
+/**
+ * An AsyncBaseOp represents a pending operation.  You may set a notification
+ * callback or you may use this class's methods directly.
+ *
+ * The op must remain allocated until it is completed or canceled.
+ */
+class AsyncBaseOp {
+  friend class AsyncBase;
+
+ public:
+  using NotificationCallback = std::function<void(AsyncBaseOp*)>;
+
+  explicit AsyncBaseOp(NotificationCallback cb = NotificationCallback());
+  AsyncBaseOp(const AsyncBaseOp&) = delete;
+  AsyncBaseOp& operator=(const AsyncBaseOp&) = delete;
+  virtual ~AsyncBaseOp();
+
+  enum class State {
+    UNINITIALIZED,
+    INITIALIZED,
+    PENDING,
+    COMPLETED,
+    CANCELED,
+  };
+
+  /**
+   * Initiate a read request.
+   */
+  virtual void pread(int fd, void* buf, size_t size, off_t start) = 0;
+  void pread(int fd, Range<unsigned char*> range, off_t start) {
+    pread(fd, range.begin(), range.size(), start);
+  }
+  virtual void preadv(int fd, const iovec* iov, int iovcnt, off_t start) = 0;
+
+  /**
+   * Initiate a write request.
+   */
+  virtual void pwrite(int fd, const void* buf, size_t size, off_t start) = 0;
+  void pwrite(int fd, Range<const unsigned char*> range, off_t start) {
+    pwrite(fd, range.begin(), range.size(), start);
+  }
+  virtual void pwritev(int fd, const iovec* iov, int iovcnt, off_t start) = 0;
+
+  // we support only these subclasses
+  virtual AsyncIOOp* getAsyncIOOp() = 0;
+  virtual IoUringOp* getIoUringOp() = 0;
+
+  // ostream output
+  virtual void toStream(std::ostream& os) const = 0;
+
+  /**
+   * Return the current operation state.
+   */
+  State state() const {
+    return state_;
+  }
+
+  /**
+   * user data get/set
+   */
+  void* getUserData() const {
+    return userData_;
+  }
+
+  void setUserData(void* userData) {
+    userData_ = userData;
+  }
+
+  /**
+   * Reset the operation for reuse.  It is an error to call reset() on
+   * an Op that is still pending.
+   */
+  virtual void reset(NotificationCallback cb = NotificationCallback()) = 0;
+
+  void setNotificationCallback(NotificationCallback cb) {
+    cb_ = std::move(cb);
+  }
+  const NotificationCallback& notificationCallback() const {
+    return cb_;
+  }
+
+  /**
+   * Retrieve the result of this operation.  Returns >=0 on success,
+   * -errno on failure (that is, using the Linux kernel error reporting
+   * conventions).  Use checkKernelError (folly/Exception.h) on the result to
+   * throw a std::system_error in case of error instead.
+   *
+   * It is an error to call this if the Op hasn't completed.
+   */
+  ssize_t result() const;
+
+  // debug helper
+  static std::string fd2name(int fd);
+
+ protected:
+  void init();
+  void start();
+  void complete(ssize_t result);
+  void cancel();
+
+  NotificationCallback cb_;
+  State state_;
+  ssize_t result_;
+  void* userData_{nullptr};
+};
+
+std::ostream& operator<<(std::ostream& stream, const AsyncBaseOp& op);
+std::ostream& operator<<(std::ostream& stream, AsyncBaseOp::State state);
+
+/**
+ * Generic C++ interface around Linux IO(io_submit, io_uring)
+ */
+class AsyncBase {
+ public:
+  using Op = AsyncBaseOp;
+
+  enum PollMode {
+    NOT_POLLABLE,
+    POLLABLE,
+  };
+
+  /**
+   * Create an AsyncBase context capable of holding at most 'capacity' pending
+   * requests at the same time.  As requests complete, others can be scheduled,
+   * as long as this limit is not exceeded.
+   *
+   * If pollMode is POLLABLE, pollFd() will return a file descriptor that
+   * can be passed to poll / epoll / select and will become readable when
+   * any IOs on this AsyncBase have completed.  If you do this, you must use
+   * pollCompleted() instead of wait() -- do not read from the pollFd()
+   * file descriptor directly.
+   *
+   * You may use the same AsyncBase object from multiple threads, as long as
+   * there is only one concurrent caller of wait() / pollCompleted() / cancel()
+   * (perhaps by always calling it from the same thread, or by providing
+   * appropriate mutual exclusion).  In this case, pending() returns a snapshot
+   * of the current number of pending requests.
+   */
+  explicit AsyncBase(size_t capacity, PollMode pollMode = NOT_POLLABLE);
+  AsyncBase(const AsyncBase&) = delete;
+  AsyncBase& operator=(const AsyncBase&) = delete;
+  virtual ~AsyncBase();
+
+  /**
+   * Wait for at least minRequests to complete.  Returns the requests that
+   * have completed; the returned range is valid until the next call to
+   * wait().  minRequests may be 0 to not block.
+   */
+  Range<Op**> wait(size_t minRequests);
+
+  /**
+   * Cancel all pending requests and return them; the returned range is
+   * valid until the next call to cancel().
+   */
+  Range<Op**> cancel();
+
+  /**
+   * Return the number of pending requests.
+   */
+  size_t pending() const {
+    return pending_;
+  }
+
+  /**
+   * Return the maximum number of requests that can be kept outstanding
+   * at any one time.
+   */
+  size_t capacity() const {
+    return capacity_;
+  }
+
+  /**
+   * Return the accumulative number of submitted I/O, since this object
+   * has been created.
+   */
+  size_t totalSubmits() const {
+    return submitted_;
+  }
+
+  /**
+   * If POLLABLE, return a file descriptor that can be passed to poll / epoll
+   * and will become readable when any async IO operations have completed.
+   * If NOT_POLLABLE, return -1.
+   */
+  int pollFd() const {
+    return pollFd_;
+  }
+
+  /**
+   * If POLLABLE, call instead of wait after the file descriptor returned
+   * by pollFd() became readable.  The returned range is valid until the next
+   * call to pollCompleted().
+   */
+  Range<Op**> pollCompleted();
+
+  /**
+   * Submit an op for execution.
+   */
+  void submit(Op* op);
+
+ protected:
+  void complete(Op* op, ssize_t result) {
+    op->complete(result);
+  }
+
+  void cancel(Op* op) {
+    op->cancel();
+  }
+
+  bool isInit() const {
+    return init_.load(std::memory_order_relaxed);
+  }
+
+  void decrementPending();
+  virtual void initializeContext() = 0;
+  virtual int submitOne(AsyncBase::Op* op) = 0;
+
+  enum class WaitType { COMPLETE, CANCEL };
+  virtual Range<AsyncBase::Op**> doWait(
+      WaitType type,
+      size_t minRequests,
+      size_t maxRequests,
+      std::vector<Op*>& result) = 0;
+
+  std::atomic<bool> init_{false};
+  std::mutex initMutex_;
+
+  std::atomic<size_t> pending_{0};
+  std::atomic<size_t> submitted_{0};
+  const size_t capacity_;
+  int pollFd_{-1};
+  std::vector<Op*> completed_;
+  std::vector<Op*> canceled_;
+};
+
+/**
+ * Wrapper around AsyncBase that allows you to schedule more requests than
+ * the AsyncBase's object capacity.  Other requests are queued and processed
+ * in a FIFO order.
+ */
+class AsyncBaseQueue {
+ public:
+  /**
+   * Create a queue, using the given AsyncBase object.
+   * The AsyncBase object may not be used by anything else until the
+   * queue is destroyed.
+   */
+  explicit AsyncBaseQueue(AsyncBase* asyncBase);
+  ~AsyncBaseQueue();
+
+  size_t queued() const {
+    return queue_.size();
+  }
+
+  /**
+   * Submit an op to the AsyncBase queue.  The op will be queued until
+   * the AsyncBase object has room.
+   */
+  void submit(AsyncBaseOp* op);
+
+  /**
+   * Submit a delayed op to the AsyncBase queue; this allows you to postpone
+   * creation of the Op (which may require allocating memory, etc) until
+   * the AsyncBase object has room.
+   */
+  using OpFactory = std::function<AsyncBaseOp*()>;
+  void submit(OpFactory op);
+
+ private:
+  void onCompleted(AsyncBaseOp* op);
+  void maybeDequeue();
+
+  AsyncBase* asyncBase_;
+
+  std::deque<OpFactory> queue_;
+};
+
+} // namespace folly

folly/experimental/io/AsyncIO.h

@@ -16,267 +16,80 @@
 
 #pragma once
 
-#include <sys/types.h>
-
-#include <atomic>
-#include <cstdint>
-#include <deque>
-#include <functional>
-#include <iosfwd>
-#include <mutex>
-#include <utility>
-#include <vector>
-
 #include <libaio.h>
 
-#include <folly/Portability.h>
-#include <folly/Range.h>
-#include <folly/portability/SysUio.h>
+#include <folly/experimental/io/AsyncBase.h>
 
 namespace folly {
 
-/**
- * An AsyncIOOp represents a pending operation.  You may set a notification
- * callback or you may use this class's methods directly.
- *
- * The op must remain allocated until it is completed or canceled.
- */
-class AsyncIOOp {
+class AsyncIOOp : public AsyncBaseOp {
   friend class AsyncIO;
   friend std::ostream& operator<<(std::ostream& stream, const AsyncIOOp& o);
 
  public:
-  typedef std::function<void(AsyncIOOp*)> NotificationCallback;
-
   explicit AsyncIOOp(NotificationCallback cb = NotificationCallback());
   AsyncIOOp(const AsyncIOOp&) = delete;
   AsyncIOOp& operator=(const AsyncIOOp&) = delete;
-  ~AsyncIOOp();
-
-  enum class State {
-    UNINITIALIZED,
-    INITIALIZED,
-    PENDING,
-    COMPLETED,
-    CANCELED,
-  };
+  ~AsyncIOOp() override;
 
   /**
    * Initiate a read request.
    */
-  void pread(int fd, void* buf, size_t size, off_t start);
-  void pread(int fd, Range<unsigned char*> range, off_t start);
-  void preadv(int fd, const iovec* iov, int iovcnt, off_t start);
+  void pread(int fd, void* buf, size_t size, off_t start) override;
+  void preadv(int fd, const iovec* iov, int iovcnt, off_t start) override;
 
   /**
    * Initiate a write request.
    */
-  void pwrite(int fd, const void* buf, size_t size, off_t start);
-  void pwrite(int fd, Range<const unsigned char*> range, off_t start);
-  void pwritev(int fd, const iovec* iov, int iovcnt, off_t start);
-
-  /**
-   * Return the current operation state.
-   */
-  State state() const {
-    return state_;
-  }
+  void pwrite(int fd, const void* buf, size_t size, off_t start) override;
+  void pwritev(int fd, const iovec* iov, int iovcnt, off_t start) override;
 
-  /**
-   * Reset the operation for reuse.  It is an error to call reset() on
-   * an Op that is still pending.
-   */
-  void reset(NotificationCallback cb = NotificationCallback());
+  void reset(NotificationCallback cb = NotificationCallback()) override;
 
-  void setNotificationCallback(NotificationCallback cb) {
-    cb_ = std::move(cb);
+  AsyncIOOp* getAsyncIOOp() override {
+    return this;
   }
-  const NotificationCallback& notificationCallback() const {
-    return cb_;
+
+  IoUringOp* getIoUringOp() override {
+    return nullptr;
   }
 
-  /**
-   * Retrieve the result of this operation.  Returns >=0 on success,
-   * -errno on failure (that is, using the Linux kernel error reporting
-   * conventions).  Use checkKernelError (folly/Exception.h) on the result to
-   * throw a std::system_error in case of error instead.
-   *
-   * It is an error to call this if the Op hasn't completed.
-   */
-  ssize_t result() const;
+  void toStream(std::ostream& os) const override;
 
  private:
-  void init();
-  void start();
-  void complete(ssize_t result);
-  void cancel();
-
-  NotificationCallback cb_;
   iocb iocb_;
-  State state_;
-  ssize_t result_;
 };
 
-std::ostream& operator<<(std::ostream& stream, const AsyncIOOp& o);
-std::ostream& operator<<(std::ostream& stream, AsyncIOOp::State state);
+std::ostream& operator<<(std::ostream& stream, const AsyncIOOp& op);
 
 /**
  * C++ interface around Linux Async IO.
  */
-class AsyncIO {
+class AsyncIO : public AsyncBase {
  public:
-  typedef AsyncIOOp Op;
-
-  enum PollMode {
-    NOT_POLLABLE,
-    POLLABLE,
-  };
+  using Op = AsyncIOOp;
 
   /**
-   * Create an AsyncIO context capable of holding at most 'capacity' pending
-   * requests at the same time.  As requests complete, others can be scheduled,
-   * as long as this limit is not exceeded.
-   *
    * Note: the maximum number of allowed concurrent requests is controlled
    * by the fs.aio-max-nr sysctl, the default value is usually 64K.
-   *
-   * If pollMode is POLLABLE, pollFd() will return a file descriptor that
-   * can be passed to poll / epoll / select and will become readable when
-   * any IOs on this AsyncIO have completed.  If you do this, you must use
-   * pollCompleted() instead of wait() -- do not read from the pollFd()
-   * file descriptor directly.
-   *
-   * You may use the same AsyncIO object from multiple threads, as long as
-   * there is only one concurrent caller of wait() / pollCompleted() / cancel()
-   * (perhaps by always calling it from the same thread, or by providing
-   * appropriate mutual exclusion).  In this case, pending() returns a snapshot
-   * of the current number of pending requests.
    */
   explicit AsyncIO(size_t capacity, PollMode pollMode = NOT_POLLABLE);
   AsyncIO(const AsyncIO&) = delete;
   AsyncIO& operator=(const AsyncIO&) = delete;
-  ~AsyncIO();
-
-  /**
-   * Wait for at least minRequests to complete.  Returns the requests that
-   * have completed; the returned range is valid until the next call to
-   * wait().  minRequests may be 0 to not block.
-   */
-  Range<Op**> wait(size_t minRequests);
-
-  /**
-   * Cancel all pending requests and return them; the returned range is
-   * valid until the next call to cancel().
-   */
-  Range<Op**> cancel();
-
-  /**
-   * Return the number of pending requests.
-   */
-  size_t pending() const {
-    return pending_;
-  }
-
-  /**
-   * Return the maximum number of requests that can be kept outstanding
-   * at any one time.
-   */
-  size_t capacity() const {
-    return capacity_;
-  }
-
-  /**
-   * Return the accumulative number of submitted I/O, since this object
-   * has been created.
-   */
-  size_t totalSubmits() const {
-    return submitted_;
-  }
-
-  /**
-   * If POLLABLE, return a file descriptor that can be passed to poll / epoll
-   * and will become readable when any async IO operations have completed.
-   * If NOT_POLLABLE, return -1.
-   */
-  int pollFd() const {
-    return pollFd_;
-  }
-
-  /**
-   * If POLLABLE, call instead of wait after the file descriptor returned
-   * by pollFd() became readable.  The returned range is valid until the next
-   * call to pollCompleted().
-   */
-  Range<Op**> pollCompleted();
-
-  /**
-   * Submit an op for execution.
-   */
-  void submit(Op* op);
+  ~AsyncIO() override;
 
  private:
-  void decrementPending();
-  void initializeContext();
+  void initializeContext() override;
+  int submitOne(AsyncBase::Op* op) override;
 
-  enum class WaitType { COMPLETE, CANCEL };
-  Range<AsyncIO::Op**> doWait(
+  Range<AsyncBase::Op**> doWait(
       WaitType type,
       size_t minRequests,
       size_t maxRequests,
-      std::vector<Op*>& result);
+      std::vector<AsyncBase::Op*>& result) override;
 
   io_context_t ctx_{nullptr};
-  std::atomic<bool> ctxSet_{false};
-  std::mutex initMutex_;
-
-  std::atomic<size_t> pending_{0};
-  std::atomic<size_t> submitted_{0};
-  const size_t capacity_;
-  int pollFd_{-1};
-  std::vector<Op*> completed_;
-  std::vector<Op*> canceled_;
-};
-
-/**
- * Wrapper around AsyncIO that allows you to schedule more requests than
- * the AsyncIO's object capacity.  Other requests are queued and processed
- * in a FIFO order.
- */
-class AsyncIOQueue {
- public:
-  /**
-   * Create a queue, using the given AsyncIO object.
-   * The AsyncIO object may not be used by anything else until the
-   * queue is destroyed.
-   */
-  explicit AsyncIOQueue(AsyncIO* asyncIO);
-  ~AsyncIOQueue();
-
-  size_t queued() const {
-    return queue_.size();
-  }
-
-  /**
-   * Submit an op to the AsyncIO queue.  The op will be queued until
-   * the AsyncIO object has room.
-   */
-  void submit(AsyncIOOp* op);
-
-  /**
-   * Submit a delayed op to the AsyncIO queue; this allows you to postpone
-   * creation of the Op (which may require allocating memory, etc) until
-   * the AsyncIO object has room.
-   */
-  typedef std::function<AsyncIOOp*()> OpFactory;
-  void submit(OpFactory op);
-
- private:
-  void onCompleted(AsyncIOOp* op);
-  void maybeDequeue();
-
-  AsyncIO* asyncIO_;
-
-  std::deque<OpFactory> queue_;
 };
 
+using AsyncIOQueue = AsyncBaseQueue;
 } // namespace folly

folly/experimental/io/IoUring.cpp

+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/experimental/io/IoUring.h>
+
+#include <sys/eventfd.h>
+#include <cerrno>
+#include <ostream>
+#include <stdexcept>
+#include <string>
+
+#include <boost/intrusive/parent_from_member.hpp>
+#include <glog/logging.h>
+
+#include <folly/Exception.h>
+#include <folly/Format.h>
+#include <folly/Likely.h>
+#include <folly/String.h>
+#include <folly/portability/Unistd.h>
+
+// helpers
+namespace {
+// http://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2
+uint32_t roundUpToNextPowerOfTwo(uint32_t num) {
+  if (num == 0) {
+    return 0;
+  }
+  num--;
+  num |= num >> 1;
+  num |= num >> 2;
+  num |= num >> 4;
+  num |= num >> 8;
+  num |= num >> 16;
+  return num + 1;
+}
+
+#define X(c) \
+  case c:    \
+    return #c
+
+const char* ioUringOpToString(unsigned char op) {
+  switch (op) {
+    X(IORING_OP_NOP);
+    X(IORING_OP_READV);
+    X(IORING_OP_WRITEV);
+    X(IORING_OP_FSYNC);
+    X(IORING_OP_READ_FIXED);
+    X(IORING_OP_WRITE_FIXED);
+    X(IORING_OP_POLL_ADD);
+    X(IORING_OP_POLL_REMOVE);
+    X(IORING_OP_SYNC_FILE_RANGE);
+    X(IORING_OP_SENDMSG);
+    X(IORING_OP_RECVMSG);
+    X(IORING_OP_TIMEOUT);
+  };
+  return "<INVALID op>";
+}
+
+#undef X
+
+void toStream(std::ostream& os, const struct io_uring_sqe& sqe) {
+  os << folly::format(
+      "user_data={}, opcode={}, ioprio={}, f={}, ",
+      sqe.user_data,
+      ioUringOpToString(sqe.opcode),
+      sqe.ioprio,
+      folly::AsyncBaseOp::fd2name(sqe.fd));
+
+  switch (sqe.opcode) {
+    case IORING_OP_READV:
+    case IORING_OP_WRITEV: {
+      auto offset = sqe.off;
+      auto* iovec = reinterpret_cast<struct iovec*>(sqe.addr);
+      os << "{";
+      for (unsigned int i = 0; i < sqe.len; i++) {
+        if (i) {
+          os << ",";
+        }
+        os << folly::format(
+            "buf={}, offset={}, nbytes={}",
+            iovec[i].iov_base,
+            offset,
+            iovec[i].iov_len);
+        // advance the offset
+        offset += iovec[i].iov_len;
+      }
+      os << "}";
+    } break;
+    default:
+      os << "[TODO: write debug string for " << ioUringOpToString(sqe.opcode)
+         << "] ";
+      break;
+  }
+}
+
+} // namespace
+
+namespace folly {
+
+IoUringOp::IoUringOp(NotificationCallback cb) : AsyncBaseOp(std::move(cb)) {}
+
+void IoUringOp::reset(NotificationCallback cb) {
+  CHECK_NE(state_, State::PENDING);
+  cb_ = std::move(cb);
+  state_ = State::UNINITIALIZED;
+  result_ = -EINVAL;
+}
+
+IoUringOp::~IoUringOp() {}
+
+void IoUringOp::pread(int fd, void* buf, size_t size, off_t start) {
+  init();
+  iov_[0].iov_base = buf;
+  iov_[0].iov_len = size;
+  io_uring_prep_readv(&sqe_, fd, iov_, 1, start);
+  io_uring_sqe_set_data(&sqe_, this);
+}
+
+void IoUringOp::preadv(int fd, const iovec* iov, int iovcnt, off_t start) {
+  init();
+  io_uring_prep_readv(&sqe_, fd, iov, iovcnt, start);
+  io_uring_sqe_set_data(&sqe_, this);
+}
+
+void IoUringOp::pwrite(int fd, const void* buf, size_t size, off_t start) {
+  init();
+  iov_[0].iov_base = const_cast<void*>(buf);
+  iov_[0].iov_len = size;
+  io_uring_prep_writev(&sqe_, fd, iov_, 1, start);
+  io_uring_sqe_set_data(&sqe_, this);
+}
+
+void IoUringOp::pwritev(int fd, const iovec* iov, int iovcnt, off_t start) {
+  init();
+  io_uring_prep_writev(&sqe_, fd, iov, iovcnt, start);
+  io_uring_sqe_set_data(&sqe_, this);
+}
+
+void IoUringOp::toStream(std::ostream& os) const {
+  os << "{" << state_ << ", ";
+
+  if (state_ != AsyncBaseOp::State::UNINITIALIZED) {
+    ::toStream(os, sqe_);
+  }
+
+  if (state_ == AsyncBaseOp::State::COMPLETED) {
+    os << "result=" << result_;
+    if (result_ < 0) {
+      os << " (" << errnoStr(-result_) << ')';
+    }
+    os << ", ";
+  }
+
+  os << "}";
+}
+
+std::ostream& operator<<(std::ostream& os, const IoUringOp& op) {
+  op.toStream(os);
+  return os;
+}
+
+IoUring::IoUring(size_t capacity, PollMode pollMode, size_t maxSubmit)
+    : AsyncBase(capacity, pollMode),
+      maxSubmit_((maxSubmit <= capacity) ? maxSubmit : capacity) {
+  ::memset(&ioRing_, 0, sizeof(ioRing_));
+  ::memset(&params_, 0, sizeof(params_));
+
+  params_.flags |= IORING_SETUP_CQSIZE;
+  params_.cq_entries = roundUpToNextPowerOfTwo(capacity_);
+}
+
+IoUring::~IoUring() {
+  CHECK_EQ(pending_, 0);
+  if (ioRing_.ring_fd > 0) {
+    ::io_uring_queue_exit(&ioRing_);
+    ioRing_.ring_fd = -1;
+  }
+}
+
+bool IoUring::isAvailable() {
+  IoUring ioUring(1);
+
+  try {
+    ioUring.initializeContext();
+  } catch (...) {
+    return false;
+  }
+
+  return true;
+}
+
+void IoUring::initializeContext() {
+  if (!init_.load(std::memory_order_acquire)) {
+    std::lock_guard<std::mutex> lock(initMutex_);
+    if (!init_.load(std::memory_order_relaxed)) {
+      int rc = ::io_uring_queue_init_params(
+          roundUpToNextPowerOfTwo(maxSubmit_), &ioRing_, &params_);
+      checkKernelError(rc, "IoUring: io_uring_queue_init_params failed");
+      DCHECK_GT(ioRing_.ring_fd, 0);
+      if (pollFd_ != -1) {
+        CHECK_ERR(io_uring_register_eventfd(&ioRing_, pollFd_));
+      }
+      init_.store(true, std::memory_order_release);
+    }
+  }
+}
+
+int IoUring::submitOne(AsyncBase::Op* op) {
+  // -1 return here will trigger throw if op isn't an IoUringOp
+  IoUringOp* iop = op->getIoUringOp();
+
+  if (!iop) {
+    return -1;
+  }
+
+  SharedMutex::WriteHolder lk(submitMutex_);
+  auto* sqe = io_uring_get_sqe(&ioRing_);
+  if (!sqe) {
+    return -1;
+  }
+
+  *sqe = iop->getSqe();
+
+  return io_uring_submit(&ioRing_);
+}
+
+Range<AsyncBase::Op**> IoUring::doWait(
+    WaitType type,
+    size_t minRequests,
+    size_t maxRequests,
+    std::vector<AsyncBase::Op*>& result) {
+  result.clear();
+
+  size_t count = 0;
+  while (count < maxRequests) {
+    struct io_uring_cqe* cqe = nullptr;
+    if (!io_uring_peek_cqe(&ioRing_, &cqe) && cqe) {
+      count++;
+      Op* op = reinterpret_cast<Op*>(io_uring_cqe_get_data(cqe));
+      CHECK(op);
+      auto res = cqe->res;
+      io_uring_cqe_seen(&ioRing_, cqe);
+      decrementPending();
+      switch (type) {
+        case WaitType::COMPLETE:
+          op->complete(res);
+          break;
+        case WaitType::CANCEL:
+          op->cancel();
+          break;
+      }
+      result.push_back(op);
+    } else {
+      if (count < minRequests) {
+        io_uring_enter(
+            ioRing_.ring_fd,
+            0,
+            minRequests - count,
+            IORING_ENTER_GETEVENTS,
+            nullptr);
+      } else {
+        break;
+      }
+    }
+  }
+
+  return range(result);
+}
+
+} // namespace folly

folly/experimental/io/IoUring.h

+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+extern "C" {
+#include <liburing.h>
+}
+
+#include <folly/SharedMutex.h>
+#include <folly/experimental/io/AsyncBase.h>
+
+namespace folly {
+
+/**
+ * An IoUringOp represents a pending operation.  You may set a notification
+ * callback or you may use this class's methods directly.
+ *
+ * The op must remain allocated until it is completed or canceled.
+ */
+class IoUringOp : public AsyncBaseOp {
+  friend class IoUring;
+  friend std::ostream& operator<<(std::ostream& stream, const IoUringOp& o);
+
+ public:
+  explicit IoUringOp(NotificationCallback cb = NotificationCallback());
+  IoUringOp(const IoUringOp&) = delete;
+  IoUringOp& operator=(const IoUringOp&) = delete;
+  ~IoUringOp() override;
+
+  /**
+   * Initiate a read request.
+   */
+  void pread(int fd, void* buf, size_t size, off_t start) override;
+  void preadv(int fd, const iovec* iov, int iovcnt, off_t start) override;
+
+  /**
+   * Initiate a write request.
+   */
+  void pwrite(int fd, const void* buf, size_t size, off_t start) override;
+  void pwritev(int fd, const iovec* iov, int iovcnt, off_t start) override;
+
+  void reset(NotificationCallback cb = NotificationCallback()) override;
+
+  AsyncIOOp* getAsyncIOOp() override {
+    return nullptr;
+  }
+
+  IoUringOp* getIoUringOp() override {
+    return this;
+  }
+
+  void toStream(std::ostream& os) const override;
+
+  const struct io_uring_sqe& getSqe() const {
+    return sqe_;
+  }
+
+ private:
+  struct io_uring_sqe sqe_;
+  struct iovec iov_[1];
+};
+
+std::ostream& operator<<(std::ostream& stream, const IoUringOp& op);
+
+/**
+ * C++ interface around Linux io_uring
+ */
+class IoUring : public AsyncBase {
+ public:
+  using Op = IoUringOp;
+
+  /**
+   * Note: the maximum number of allowed concurrent requests is controlled
+   * by the kernel IORING_MAX_ENTRIES and the memlock limit,
+   * The default IORING_MAX_ENTRIES value is usually 32K.
+   */
+  explicit IoUring(
+      size_t capacity,
+      PollMode pollMode = NOT_POLLABLE,
+      size_t maxSubmit = 1);
+  IoUring(const IoUring&) = delete;
+  IoUring& operator=(const IoUring&) = delete;
+  ~IoUring() override;
+
+  static bool isAvailable();
+
+ private:
+  void initializeContext() override;
+  int submitOne(AsyncBase::Op* op) override;
+
+  Range<AsyncBase::Op**> doWait(
+      WaitType type,
+      size_t minRequests,
+      size_t maxRequests,
+      std::vector<AsyncBase::Op*>& result) override;
+
+  size_t maxSubmit_;
+  struct io_uring_params params_;
+  struct io_uring ioRing_;
+  SharedMutex submitMutex_;
+};
+
+using IoUringQueue = AsyncBaseQueue;
+} // namespace folly

folly/experimental/io/test/AsyncBaseTestLib.cpp

+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/experimental/io/test/AsyncBaseTestLib.h>
+
+namespace folly {
+namespace test {
+void TestUtil::waitUntilReadable(int fd) {
+  pollfd pfd;
+  pfd.fd = fd;
+  pfd.events = POLLIN;
+
+  int r;
+  do {
+    r = poll(&pfd, 1, -1); // wait forever
+  } while (r == -1 && errno == EINTR);
+  PCHECK(r == 1);
+  CHECK_EQ(pfd.revents, POLLIN); // no errors etc
+}
+
+folly::Range<folly::AsyncBase::Op**> TestUtil::readerWait(
+    folly::AsyncBase* reader) {
+  int fd = reader->pollFd();
+  if (fd == -1) {
+    return reader->wait(1);
+  } else {
+    waitUntilReadable(fd);
+    return reader->pollCompleted();
+  }
+}
+
+TestUtil::ManagedBuffer TestUtil::allocateAligned(size_t size) {
+  void* buf;
+  int rc = posix_memalign(&buf, kAlign, size);
+  CHECK_EQ(rc, 0) << folly::errnoStr(rc);
+  return ManagedBuffer(reinterpret_cast<char*>(buf), free);
+}
+
+TemporaryFile::TemporaryFile(size_t size)
+    : path_(folly::fs::temp_directory_path() / folly::fs::unique_path()) {
+  CHECK_EQ(size % sizeof(uint32_t), 0);
+  size /= sizeof(uint32_t);
+  const uint32_t seed = 42;
+  std::mt19937 rnd(seed);
+
+  const size_t bufferSize = 1U << 16;
+  uint32_t buffer[bufferSize];
+
+  FILE* fp = ::fopen(path_.c_str(), "wb");
+  PCHECK(fp != nullptr);
+  while (size) {
+    size_t n = std::min(size, bufferSize);
+    for (size_t i = 0; i < n; ++i) {
+      buffer[i] = rnd();
+    }
+    size_t written = ::fwrite(buffer, sizeof(uint32_t), n, fp);
+    PCHECK(written == n);
+    size -= written;
+  }
+  PCHECK(::fclose(fp) == 0);
+}
+
+TemporaryFile::~TemporaryFile() {
+  try {
+    folly::fs::remove(path_);
+  } catch (const folly::fs::filesystem_error& e) {
+    LOG(ERROR) << "fs::remove: " << folly::exceptionStr(e);
+  }
+}
+
+TemporaryFile& TemporaryFile::getTempFile() {
+  static TemporaryFile sTempFile(6 << 20); // 6MiB
+  return sTempFile;
+}
+} // namespace test
+} // namespace folly

folly/experimental/io/test/IoUringTest.cpp

+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/experimental/io/IoUring.h>
+#include <folly/experimental/io/test/AsyncBaseTestLib.h>
+#include <folly/init/Init.h>
+
+using folly::IoUring;
+
+namespace folly {
+namespace test {
+INSTANTIATE_TYPED_TEST_CASE_P(AsyncTest, AsyncTest, IoUring);
+} // namespace test
+} // namespace folly
+
+int main(int argc, char** argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+  folly::init(&argc, &argv);
+
+  bool avail = IoUring::isAvailable();
+  if (!avail) {
+    LOG(INFO)
+        << "Not running tests since this kernel version does not support io_uring";
+    return 0;
+  }
+
+  return RUN_ALL_TESTS();
+}