Use relaxed_atomic in IOThreadPoolExecutor

Summary:
[Folly] Use `relaxed_atomic` in `IOThreadPoolExecutor`.

For the `nextThread_` field, which may be mutated within `pickThread` in any number of threads concurrently. While `pickThread` is only ever called with a shared lock held, the shared lock does not actually help with concurrent mutations.

Since the mutations of `nextThread_` by `pickThread` is always a `std::atomic::fetch_add` with `std::memory_order_relaxed`, and since that case is a bit more nicely solved with `folly::relaxed_atomic`, switch to that.

Reviewed By: ot

Differential Revision: D6290451

fbshipit-source-id: e473fa6321e462c9b3df9d2bd9650967cca3444a

folly/executors/IOThreadPoolExecutor.cpp

@@ -159,7 +159,7 @@ IOThreadPoolExecutor::pickThread() {
     // the second case, `!me` so we'll crash anyway.
     return me;
   }
-  auto thread = ths[nextThread_.fetch_add(1, std::memory_order_relaxed) % n];
+  auto thread = ths[nextThread_++ % n];
   return std::static_pointer_cast<IOThread>(thread);
 }
 

folly/executors/IOThreadPoolExecutor.h

@@ -16,12 +16,11 @@
 
 #pragma once
 
-#include <atomic>
-
 #include <folly/Portability.h>
 #include <folly/executors/IOExecutor.h>
 #include <folly/executors/ThreadPoolExecutor.h>
 #include <folly/io/async/EventBaseManager.h>
+#include <folly/synchronization/RelaxedAtomic.h>
 
 namespace folly {
 
@@ -103,7 +102,7 @@ class IOThreadPoolExecutor : public ThreadPoolExecutor, public IOExecutor {
   size_t getPendingTaskCountImpl() const override final;
 
   const bool isWaitForAll_; // whether to wait till event base loop exits
-  std::atomic<size_t> nextThread_;
+  relaxed_atomic<size_t> nextThread_;
   folly::ThreadLocal<std::shared_ptr<IOThread>> thisThread_;
   folly::EventBaseManager* eventBaseManager_;
 };