librbd: Modify locks and atomics in AioCompletion to improve performance

Signed-off-by: Adam Lyon-Jones <adamlyon@uk.ibm.com>

diff --git a/src/librbd/io/AioCompletion.cc b/src/librbd/io/AioCompletion.cc
index 5452edf5fd37d15e5a175cf2569a58b054583fd5..df628909bc0964162f068ca4ade0f4f07a10e64d 100644 (file)
--- a/src/librbd/io/AioCompletion.cc
+++ b/src/librbd/io/AioCompletion.cc
@@ -33,12 +33,7 @@ namespace io {
 
 int AioCompletion::wait_for_complete() {
   tracepoint(librbd, aio_wait_for_complete_enter, this);
-  {
-    std::unique_lock<std::mutex> locker(lock);
-    while (state != AIO_STATE_COMPLETE) {
-      cond.wait(locker);
-    }
-  }
+  completed.wait(false, std::memory_order_acquire);
   tracepoint(librbd, aio_wait_for_complete_exit, 0);
   return 0;
 }
@@ -237,7 +232,7 @@ void AioCompletion::complete_request(ssize_t r)
 
 bool AioCompletion::is_complete() {
   tracepoint(librbd, aio_is_complete_enter, this);
-  bool done = (this->state == AIO_STATE_COMPLETE);
+  bool done = completed.load(std::memory_order_acquire);
   tracepoint(librbd, aio_is_complete_exit, done);
   return done;
 }
@@ -262,17 +257,14 @@ void AioCompletion::complete_external_callback() {
 }
 
 void AioCompletion::mark_complete_and_notify() {
-  state = AIO_STATE_COMPLETE;
+  completed.store(true, std::memory_order_release);
 
   if (ictx != nullptr && event_notify && ictx->event_socket.is_valid()) {
     ictx->event_socket_completions.push(this);
     ictx->event_socket.notify();
   }
 
-  {
-    std::unique_lock<std::mutex> locker(lock);
-    cond.notify_all();
-  }
+  completed.notify_all();
 
   if (image_dispatcher_ctx != nullptr) {
     image_dispatcher_ctx->complete(rval);

diff --git a/src/librbd/io/AioCompletion.h b/src/librbd/io/AioCompletion.h
index d98055878cd9aed452a4d285cfbb2e08ad5ca940..c50b282b13089de3d693eb80fd24efdc085b5c71 100644 (file)
--- a/src/librbd/io/AioCompletion.h
+++ b/src/librbd/io/AioCompletion.h
@@ -38,20 +38,13 @@ namespace io {
  * context or via a thread pool context for cache read hits).
  */
 struct AioCompletion {
-  typedef enum {
-    AIO_STATE_PENDING = 0,
-    AIO_STATE_COMPLETE,
-  } aio_state_t;
-
   mutable std::mutex lock;
-  std::condition_variable cond;
 
   callback_t complete_cb = nullptr;
   void *complete_arg = nullptr;
   rbd_completion_t rbd_comp = nullptr;
 
-  /// note: only using atomic for built-in memory barrier
-  std::atomic<aio_state_t> state{AIO_STATE_PENDING};
+  std::atomic<bool> completed{false};
 
   std::atomic<ssize_t> rval{0};
   std::atomic<int> error_rval{0};