]> git.apps.os.sepia.ceph.com Git - ceph.git/commitdiff
common: add an async SharedMutex
authorCasey Bodley <cbodley@redhat.com>
Fri, 30 Mar 2018 03:53:40 +0000 (23:53 -0400)
committerCasey Bodley <cbodley@redhat.com>
Mon, 25 Jun 2018 19:32:48 +0000 (15:32 -0400)
Signed-off-by: Casey Bodley <cbodley@redhat.com>
src/common/async/shared_mutex.h [new file with mode: 0644]
src/test/common/CMakeLists.txt
src/test/common/test_async_shared_mutex.cc [new file with mode: 0644]

diff --git a/src/common/async/shared_mutex.h b/src/common/async/shared_mutex.h
new file mode 100644 (file)
index 0000000..f71928f
--- /dev/null
@@ -0,0 +1,428 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2018 Red Hat
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation. See file COPYING.
+ *
+ */
+
+#ifndef CEPH_ASYNC_SHARED_MUTEX_H
+#define CEPH_ASYNC_SHARED_MUTEX_H
+
+#include <condition_variable>
+#include <mutex>
+#include <shared_mutex> // for std::shared_lock
+
+#include <boost/intrusive/list.hpp>
+
+#include "common/async/completion.h"
+
+namespace ceph::async {
+
+/**
+ * An asynchronous shared mutex for use with boost::asio.
+ *
+ * A shared mutex class with asynchronous lock operations that complete on a
+ * boost::asio executor. The class also has synchronous interfaces that meet
+ * most of the standard library's requirements for the SharedMutex concept,
+ * which makes it compatible with lock_guard, unique_lock, and shared_lock.
+ *
+ * All lock requests can fail with operation_aborted on cancel() or destruction.
+ * The non-error_code overloads of lock() and lock_shared() will throw this
+ * error as an exception of type boost::system::system_error.
+ *
+ * Exclusive locks are prioritized over shared locks. Locks of the same type
+ * are granted in fifo order. The implementation defines a limit on the number
+ * of shared locks to 65534 at a time.
+ *
+ * Example use:
+ *
+ *   boost::asio::io_context context;
+ *   SharedMutex mutex{context.get_executor()};
+ *
+ *   mutex.async_lock([&] (boost::system::error_code ec, auto lock) {
+ *       if (!ec) {
+ *         // mutate shared state ...
+ *       }
+ *     });
+ *   mutex.async_lock_shared([&] (boost::system::error_code ec, auto lock) {
+ *       if (!ec) {
+ *         // read shared state ...
+ *       }
+ *     });
+ *
+ *   context.run();
+ */
+template <typename Executor>
+class SharedMutex {
+ public:
+  SharedMutex(const Executor& ex1);
+
+  /// on destruction, all pending lock requests are canceled
+  ~SharedMutex();
+
+  using executor_type = Executor;
+  executor_type get_executor() const noexcept { return ex1; }
+
+  /// initiate an asynchronous request for an exclusive lock. when the lock is
+  /// granted, the completion handler is invoked with a successful error code
+  /// and a std::unique_lock that owns this mutex.
+  /// Signature = void(boost::system::error_code, std::unique_lock)
+  template <typename CompletionToken>
+  auto async_lock(CompletionToken&& token);
+
+  /// wait synchronously for an exclusive lock. if an error occurs before the
+  /// lock is granted, that error is thrown as an exception
+  void lock();
+
+  /// wait synchronously for an exclusive lock. if an error occurs before the
+  /// lock is granted, that error is assigned to 'ec'
+  void lock(boost::system::error_code& ec);
+
+  /// try to acquire an exclusive lock. if the lock is not immediately
+  /// available, returns false
+  bool try_lock();
+
+  /// releases an exclusive lock. not required to be called from the same thread
+  /// that initiated the lock
+  void unlock();
+
+  /// initiate an asynchronous request for a shared lock. when the lock is
+  /// granted, the completion handler is invoked with a successful error code
+  /// and a std::shared_lock that owns this mutex.
+  /// Signature = void(boost::system::error_code, std::shared_lock)
+  template <typename CompletionToken>
+  auto async_lock_shared(CompletionToken&& token);
+
+  /// wait synchronously for a shared lock. if an error occurs before the
+  /// lock is granted, that error is thrown as an exception
+  void lock_shared();
+
+  /// wait synchronously for a shared lock. if an error occurs before the lock
+  /// is granted, that error is assigned to 'ec'
+  void lock_shared(boost::system::error_code& ec);
+
+  /// try to acquire a shared lock. if the lock is not immediately available,
+  /// returns false
+  bool try_lock_shared();
+
+  /// releases a shared lock. not required to be called from the same thread
+  /// that initiated the lock
+  void unlock_shared();
+
+  /// cancel any pending requests for exclusive or shared locks with an
+  /// operation_aborted error
+  void cancel();
+
+ private:
+  Executor ex1; //< default callback executor
+
+  struct LockRequest : public boost::intrusive::list_base_hook<> {
+    virtual ~LockRequest() {}
+    virtual void complete(boost::system::error_code ec) = 0;
+    virtual void destroy() = 0;
+  };
+  using RequestList = boost::intrusive::list<LockRequest>;
+
+  RequestList shared_queue; //< requests waiting on a shared lock
+  RequestList exclusive_queue; //< requests waiting on an exclusive lock
+
+  /// lock state encodes the number of shared lockers, or 'max' for exclusive
+  using LockState = uint16_t;
+  static constexpr LockState Unlocked = 0;
+  static constexpr LockState Exclusive = std::numeric_limits<LockState>::max();
+  static constexpr LockState MaxShared = Exclusive - 1;
+  LockState state = Unlocked; //< current lock state
+
+  std::mutex mutex; //< protects lock state and wait queues
+
+  // sync requests live on the stack and wait on a condition variable
+  class SyncRequest;
+
+  // async requests use async::Completion to invoke a handler on its executor
+  template <template <typename Mutex> typename Lock>
+  class AsyncRequest;
+
+  using AsyncExclusiveRequest = AsyncRequest<std::unique_lock>;
+  using AsyncSharedRequest = AsyncRequest<std::shared_lock>;
+
+  void complete(RequestList&& requests, boost::system::error_code ec);
+};
+
+template <typename Executor>
+class SharedMutex<Executor>::SyncRequest : public LockRequest {
+  std::condition_variable cond;
+  std::optional<boost::system::error_code> ec;
+ public:
+  boost::system::error_code wait(std::unique_lock<std::mutex>& lock) {
+    // return the error code once its been set
+    cond.wait(lock, [this] { return ec; });
+    return *ec;
+  }
+  void complete(boost::system::error_code ec) override {
+    this->ec = ec;
+    cond.notify_one();
+  }
+  void destroy() override {
+    // nothing, SyncRequests live on the stack
+  }
+};
+
+template <typename Executor>
+template <template <typename Mutex> typename Lock>
+class SharedMutex<Executor>::AsyncRequest : public LockRequest {
+  SharedMutex& mutex; //< mutex argument for lock guard
+ public:
+  AsyncRequest(SharedMutex& mutex) : mutex(mutex) {}
+
+  using Signature = void(boost::system::error_code, Lock<SharedMutex>);
+  using LockCompletion = Completion<Signature, AsBase<AsyncRequest>>;
+
+  void complete(boost::system::error_code ec) override {
+    auto r = static_cast<LockCompletion*>(this);
+    // pass ownership of ourselves to post(). on error, pass an empty lock
+    post(std::unique_ptr<LockCompletion>{r}, ec,
+         ec ? Lock{mutex, std::defer_lock} : Lock{mutex, std::adopt_lock});
+  }
+  void destroy() override {
+    delete static_cast<LockCompletion*>(this);
+  }
+};
+
+
+template <typename Executor>
+inline SharedMutex<Executor>::SharedMutex(const Executor& ex1)
+  : ex1(ex1)
+{
+}
+
+template <typename Executor>
+inline SharedMutex<Executor>::~SharedMutex()
+{
+  try {
+    cancel();
+  } catch (const std::exception&) {
+    // swallow any exceptions, the destructor can't throw
+  }
+}
+
+template <typename Executor>
+template <typename CompletionToken>
+auto SharedMutex<Executor>::async_lock(CompletionToken&& token)
+{
+  using Signature = typename AsyncExclusiveRequest::Signature;
+  boost::asio::async_completion<CompletionToken, Signature> init(token);
+  auto& handler = init.completion_handler;
+  {
+    std::lock_guard lock{mutex};
+
+    if (state == Unlocked) {
+      state = Exclusive;
+
+      // post the completion
+      auto ex2 = boost::asio::get_associated_executor(handler, ex1);
+      auto alloc2 = boost::asio::get_associated_allocator(handler);
+      auto b = bind_handler(std::move(handler), boost::system::error_code{},
+                            std::unique_lock{*this, std::adopt_lock});
+      ex2.post(forward_handler(std::move(b)), alloc2);
+    } else {
+      // create a request and add it to the exclusive list
+      using LockCompletion = typename AsyncExclusiveRequest::LockCompletion;
+      auto request = LockCompletion::create(ex1, std::move(handler), *this);
+      exclusive_queue.push_back(*request.release());
+    }
+  }
+  return init.result.get();
+}
+
+template <typename Executor>
+inline void SharedMutex<Executor>::lock()
+{
+  boost::system::error_code ec;
+  lock(ec);
+  if (ec) {
+    throw boost::system::system_error(ec);
+  }
+}
+
+template <typename Executor>
+void SharedMutex<Executor>::lock(boost::system::error_code& ec)
+{
+  std::unique_lock lock{mutex};
+
+  if (state == Unlocked) {
+    state = Exclusive;
+    ec.clear();
+  } else {
+    SyncRequest request;
+    exclusive_queue.push_back(request);
+    ec = request.wait(lock);
+  }
+}
+
+template <typename Executor>
+inline bool SharedMutex<Executor>::try_lock()
+{
+  std::lock_guard lock{mutex};
+
+  if (state == Unlocked) {
+    state = Exclusive;
+    return true;
+  }
+  return false;
+}
+
+template <typename Executor>
+void SharedMutex<Executor>::unlock()
+{
+  RequestList granted;
+  {
+    std::lock_guard lock{mutex};
+    assert(state == Exclusive);
+
+    if (!exclusive_queue.empty()) {
+      // grant next exclusive lock
+      auto& request = exclusive_queue.front();
+      exclusive_queue.pop_front();
+      granted.push_back(request);
+    } else {
+      // grant shared locks, if any
+      state = shared_queue.size();
+      if (state > MaxShared) {
+        state = MaxShared;
+        auto end = std::next(shared_queue.begin(), MaxShared);
+        granted.splice(granted.end(), shared_queue,
+                       shared_queue.begin(), end, MaxShared);
+      } else {
+        granted.splice(granted.end(), shared_queue);
+      }
+    }
+  }
+  complete(std::move(granted), boost::system::error_code{});
+}
+
+template <typename Executor>
+template <typename CompletionToken>
+auto SharedMutex<Executor>::async_lock_shared(CompletionToken&& token)
+{
+  using Signature = typename AsyncSharedRequest::Signature;
+  boost::asio::async_completion<CompletionToken, Signature> init(token);
+  auto& handler = init.completion_handler;
+  {
+    std::lock_guard lock{mutex};
+
+    if (exclusive_queue.empty() && state < MaxShared) {
+      state++;
+
+      auto ex2 = boost::asio::get_associated_executor(handler, ex1);
+      auto alloc2 = boost::asio::get_associated_allocator(handler);
+      auto b = bind_handler(std::move(handler), boost::system::error_code{},
+                            std::shared_lock{*this, std::adopt_lock});
+      ex2.post(forward_handler(std::move(b)), alloc2);
+    } else {
+      using LockCompletion = typename AsyncSharedRequest::LockCompletion;
+      auto request = LockCompletion::create(ex1, std::move(handler), *this);
+      shared_queue.push_back(*request.release());
+    }
+  }
+  return init.result.get();
+}
+
+template <typename Executor>
+inline void SharedMutex<Executor>::lock_shared()
+{
+  boost::system::error_code ec;
+  lock_shared(ec);
+  if (ec) {
+    throw boost::system::system_error(ec);
+  }
+}
+
+template <typename Executor>
+void SharedMutex<Executor>::lock_shared(boost::system::error_code& ec)
+{
+  std::unique_lock lock{mutex};
+
+  if (exclusive_queue.empty() && state < MaxShared) {
+    state++;
+    ec.clear();
+  } else {
+    SyncRequest request;
+    shared_queue.push_back(request);
+    ec = request.wait(lock);
+  }
+}
+
+template <typename Executor>
+inline bool SharedMutex<Executor>::try_lock_shared()
+{
+  std::lock_guard lock{mutex};
+
+  if (exclusive_queue.empty() && state < MaxShared) {
+    state++;
+    return true;
+  }
+  return false;
+}
+
+template <typename Executor>
+inline void SharedMutex<Executor>::unlock_shared()
+{
+  std::lock_guard lock{mutex};
+  assert(state != Unlocked && state <= MaxShared);
+
+  if (state == 1 && !exclusive_queue.empty()) {
+    // grant next exclusive lock
+    state = Exclusive;
+    auto& request = exclusive_queue.front();
+    exclusive_queue.pop_front();
+    request.complete(boost::system::error_code{});
+  } else if (state == MaxShared && !shared_queue.empty() &&
+             exclusive_queue.empty()) {
+    // grant next shared lock
+    auto& request = shared_queue.front();
+    shared_queue.pop_front();
+    request.complete(boost::system::error_code{});
+  } else {
+    state--;
+  }
+}
+
+template <typename Executor>
+inline void SharedMutex<Executor>::cancel()
+{
+  RequestList canceled;
+  {
+    std::lock_guard lock{mutex};
+    canceled.splice(canceled.end(), shared_queue);
+    canceled.splice(canceled.end(), exclusive_queue);
+  }
+  complete(std::move(canceled), boost::asio::error::operation_aborted);
+}
+
+template <typename Executor>
+void SharedMutex<Executor>::complete(RequestList&& requests,
+                                     boost::system::error_code ec)
+{
+  while (!requests.empty()) {
+    auto& request = requests.front();
+    requests.pop_front();
+    try {
+      request.complete(ec);
+    } catch (...) {
+      // clean up any remaining completions and rethrow
+      requests.clear_and_dispose([] (LockRequest *r) { r->destroy(); });
+      throw;
+    }
+  }
+}
+
+} // namespace ceph::async
+
+#endif // CEPH_ASYNC_SHARED_MUTEX_H
index b76ea2b7ca7742be6bfa3e0ab319864f78e87c5c..0f9981c62049bb6de44fbb4a56ed13158cc7e869 100644 (file)
@@ -294,3 +294,7 @@ add_ceph_unittest(unittest_hobject)
 add_executable(unittest_async_completion test_async_completion.cc)
 add_ceph_unittest(unittest_async_completion)
 target_link_libraries(unittest_async_completion Boost::system)
+
+add_executable(unittest_async_shared_mutex test_async_shared_mutex.cc)
+add_ceph_unittest(unittest_async_shared_mutex)
+target_link_libraries(unittest_async_shared_mutex Boost::system)
diff --git a/src/test/common/test_async_shared_mutex.cc b/src/test/common/test_async_shared_mutex.cc
new file mode 100644 (file)
index 0000000..aed6e7b
--- /dev/null
@@ -0,0 +1,428 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2018 Red Hat
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation. See file COPYING.
+ *
+ */
+
+#include "common/async/shared_mutex.h"
+#include <optional>
+#include <gtest/gtest.h>
+
+namespace ceph::async {
+
+using executor_type = boost::asio::io_context::executor_type;
+using unique_lock = std::unique_lock<SharedMutex<executor_type>>;
+using shared_lock = std::shared_lock<SharedMutex<executor_type>>;
+
+// return a lambda that captures its error code and lock
+auto capture(std::optional<boost::system::error_code>& ec, unique_lock& lock)
+{
+  return [&] (boost::system::error_code e, unique_lock l) {
+    ec = e;
+    lock = std::move(l);
+  };
+}
+auto capture(std::optional<boost::system::error_code>& ec, shared_lock& lock)
+{
+  return [&] (boost::system::error_code e, shared_lock l) {
+    ec = e;
+    lock = std::move(l);
+  };
+}
+
+TEST(SharedMutex, async_exclusive)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::optional<boost::system::error_code> ec1, ec2, ec3;
+  unique_lock lock1, lock2, lock3;
+
+  // request three exclusive locks
+  mutex.async_lock(capture(ec1, lock1));
+  mutex.async_lock(capture(ec2, lock2));
+  mutex.async_lock(capture(ec3, lock3));
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped()); // second lock still pending
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  EXPECT_FALSE(ec2);
+
+  lock1.unlock();
+
+  EXPECT_FALSE(ec2);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped());
+
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::system::errc::success, *ec2);
+  ASSERT_TRUE(lock2);
+  EXPECT_FALSE(ec3);
+
+  lock2.unlock();
+
+  EXPECT_FALSE(ec3);
+
+  context.poll();
+  EXPECT_TRUE(context.stopped());
+
+  ASSERT_TRUE(ec3);
+  EXPECT_EQ(boost::system::errc::success, *ec3);
+  ASSERT_TRUE(lock3);
+}
+
+TEST(SharedMutex, async_shared)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::optional<boost::system::error_code> ec1, ec2;
+  shared_lock lock1, lock2;
+
+  // request two shared locks
+  mutex.async_lock_shared(capture(ec1, lock1));
+  mutex.async_lock_shared(capture(ec2, lock2));
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+
+  context.poll();
+  EXPECT_TRUE(context.stopped());
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::system::errc::success, *ec2);
+  ASSERT_TRUE(lock2);
+}
+
+TEST(SharedMutex, async_exclusive_while_shared)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::optional<boost::system::error_code> ec1, ec2;
+  shared_lock lock1;
+  unique_lock lock2;
+
+  // request a shared and exclusive lock
+  mutex.async_lock_shared(capture(ec1, lock1));
+  mutex.async_lock(capture(ec2, lock2));
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped()); // second lock still pending
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  EXPECT_FALSE(ec2);
+
+  lock1.unlock();
+
+  EXPECT_FALSE(ec2);
+
+  context.poll();
+  EXPECT_TRUE(context.stopped());
+
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::system::errc::success, *ec2);
+  ASSERT_TRUE(lock2);
+}
+
+TEST(SharedMutex, async_shared_while_exclusive)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::optional<boost::system::error_code> ec1, ec2;
+  unique_lock lock1;
+  shared_lock lock2;
+
+  // request an exclusive and shared lock
+  mutex.async_lock(capture(ec1, lock1));
+  mutex.async_lock_shared(capture(ec2, lock2));
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped()); // second lock still pending
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  EXPECT_FALSE(ec2);
+
+  lock1.unlock();
+
+  EXPECT_FALSE(ec2);
+
+  context.poll();
+  EXPECT_TRUE(context.stopped());
+
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::system::errc::success, *ec2);
+  ASSERT_TRUE(lock2);
+}
+
+TEST(SharedMutex, async_prioritize_exclusive)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::optional<boost::system::error_code> ec1, ec2, ec3;
+  shared_lock lock1, lock3;
+  unique_lock lock2;
+
+  // acquire a shared lock, then request an exclusive and another shared lock
+  mutex.async_lock_shared(capture(ec1, lock1));
+  mutex.async_lock(capture(ec2, lock2));
+  mutex.async_lock_shared(capture(ec3, lock3));
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped());
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  EXPECT_FALSE(ec2);
+  // exclusive waiter blocks the second shared lock
+  EXPECT_FALSE(ec3);
+
+  lock1.unlock();
+
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped());
+
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::system::errc::success, *ec2);
+  ASSERT_TRUE(lock2);
+  EXPECT_FALSE(ec3);
+}
+
+TEST(SharedMutex, async_cancel)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::optional<boost::system::error_code> ec1, ec2, ec3, ec4;
+  unique_lock lock1, lock2;
+  shared_lock lock3, lock4;
+
+  // request 2 exclusive and shared locks
+  mutex.async_lock(capture(ec1, lock1));
+  mutex.async_lock(capture(ec2, lock2));
+  mutex.async_lock_shared(capture(ec3, lock3));
+  mutex.async_lock_shared(capture(ec4, lock4));
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+  EXPECT_FALSE(ec4);
+
+  context.poll();
+  EXPECT_FALSE(context.stopped());
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+  EXPECT_FALSE(ec4);
+
+  mutex.cancel();
+
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+  EXPECT_FALSE(ec4);
+
+  context.poll();
+  EXPECT_TRUE(context.stopped());
+
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::asio::error::operation_aborted, *ec2);
+  EXPECT_FALSE(lock2);
+  ASSERT_TRUE(ec3);
+  EXPECT_EQ(boost::asio::error::operation_aborted, *ec3);
+  EXPECT_FALSE(lock3);
+  ASSERT_TRUE(ec4);
+  EXPECT_EQ(boost::asio::error::operation_aborted, *ec4);
+  EXPECT_FALSE(lock4);
+}
+
+TEST(SharedMutex, async_destruct)
+{
+  boost::asio::io_context context;
+
+  std::optional<boost::system::error_code> ec1, ec2, ec3, ec4;
+  unique_lock lock1, lock2;
+  shared_lock lock3, lock4;
+
+  {
+    SharedMutex mutex(context.get_executor());
+
+    // request 2 exclusive and shared locks
+    mutex.async_lock(capture(ec1, lock1));
+    mutex.async_lock(capture(ec2, lock2));
+    mutex.async_lock_shared(capture(ec3, lock3));
+    mutex.async_lock_shared(capture(ec4, lock4));
+  }
+
+  EXPECT_FALSE(ec1); // no callbacks until poll()
+  EXPECT_FALSE(ec2);
+  EXPECT_FALSE(ec3);
+  EXPECT_FALSE(ec4);
+
+  context.poll();
+  EXPECT_TRUE(context.stopped());
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::asio::error::operation_aborted, *ec2);
+  EXPECT_FALSE(lock2);
+  ASSERT_TRUE(ec3);
+  EXPECT_EQ(boost::asio::error::operation_aborted, *ec3);
+  EXPECT_FALSE(lock3);
+  ASSERT_TRUE(ec4);
+  EXPECT_EQ(boost::asio::error::operation_aborted, *ec4);
+  EXPECT_FALSE(lock4);
+}
+
+// return a capture() lambda that's bound to the given executor
+template <typename Executor, typename ...Args>
+auto capture_ex(const Executor& ex, Args&& ...args)
+{
+  return boost::asio::bind_executor(ex, capture(std::forward<Args>(args)...));
+}
+
+TEST(SharedMutex, cross_executor)
+{
+  boost::asio::io_context mutex_context;
+  SharedMutex mutex(mutex_context.get_executor());
+
+  boost::asio::io_context callback_context;
+  auto ex2 = callback_context.get_executor();
+
+  std::optional<boost::system::error_code> ec1, ec2;
+  unique_lock lock1, lock2;
+
+  // request two exclusive locks
+  mutex.async_lock(capture_ex(ex2, ec1, lock1));
+  mutex.async_lock(capture_ex(ex2, ec2, lock2));
+
+  EXPECT_FALSE(ec1);
+  EXPECT_FALSE(ec2);
+
+  mutex_context.poll();
+  EXPECT_FALSE(mutex_context.stopped()); // maintains work on both executors
+
+  EXPECT_FALSE(ec1); // no callbacks until poll() on callback_context
+  EXPECT_FALSE(ec2);
+
+  callback_context.poll();
+  EXPECT_FALSE(callback_context.stopped()); // second lock still pending
+
+  ASSERT_TRUE(ec1);
+  EXPECT_EQ(boost::system::errc::success, *ec1);
+  ASSERT_TRUE(lock1);
+  EXPECT_FALSE(ec2);
+
+  lock1.unlock();
+
+  mutex_context.poll();
+  EXPECT_TRUE(mutex_context.stopped());
+
+  EXPECT_FALSE(ec2);
+
+  callback_context.poll();
+  EXPECT_TRUE(callback_context.stopped());
+
+  ASSERT_TRUE(ec2);
+  EXPECT_EQ(boost::system::errc::success, *ec2);
+  ASSERT_TRUE(lock2);
+}
+
+TEST(SharedMutex, try_exclusive)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+  {
+    std::lock_guard lock{mutex};
+    ASSERT_FALSE(mutex.try_lock()); // fail during exclusive
+  }
+  {
+    std::shared_lock lock{mutex};
+    ASSERT_FALSE(mutex.try_lock()); // fail during shared
+  }
+  ASSERT_TRUE(mutex.try_lock());
+  mutex.unlock();
+}
+
+TEST(SharedMutex, try_shared)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+  {
+    std::lock_guard lock{mutex};
+    ASSERT_FALSE(mutex.try_lock_shared()); // fail during exclusive
+  }
+  {
+    std::shared_lock lock{mutex};
+    ASSERT_TRUE(mutex.try_lock_shared()); // succeed during shared
+    mutex.unlock_shared();
+  }
+  ASSERT_TRUE(mutex.try_lock_shared());
+  mutex.unlock_shared();
+}
+
+TEST(SharedMutex, cancel)
+{
+  boost::asio::io_context context;
+  SharedMutex mutex(context.get_executor());
+
+  std::lock_guard l{mutex}; // exclusive lock blocks others
+
+  // make synchronous lock calls in other threads
+  auto f1 = std::async(std::launch::async, [&] { mutex.lock(); });
+  auto f2 = std::async(std::launch::async, [&] { mutex.lock_shared(); });
+
+  // this will race with spawned threads. just keep canceling until the
+  // futures are ready
+  const auto t = std::chrono::milliseconds(1);
+  do { mutex.cancel(); } while (f1.wait_for(t) != std::future_status::ready);
+  do { mutex.cancel(); } while (f2.wait_for(t) != std::future_status::ready);
+
+  EXPECT_THROW(f1.get(), boost::system::system_error);
+  EXPECT_THROW(f2.get(), boost::system::system_error);
+}
+
+} // namespace ceph::async