From a4015bde81371a8b998361ccfb56d5fd714e3ec6 Mon Sep 17 00:00:00 2001
From: Determinant <ted.sybil@gmail.com>
Date: Tue, 26 Mar 2019 16:05:01 -0400
Subject: update atomic fences

---
 include/salticidae/conn.h  | 2 --
 include/salticidae/event.h | 3 ++-
 include/salticidae/queue.h | 6 +++++-
 3 files changed, 7 insertions(+), 4 deletions(-)

(limited to 'include')

diff --git a/include/salticidae/conn.h b/include/salticidae/conn.h
index e7d6295..cc1c5a6 100644
--- a/include/salticidae/conn.h
+++ b/include/salticidae/conn.h
@@ -188,7 +188,6 @@ class ConnPool {
 
         void feed(const conn_t &conn, int client_fd) {
             /* the caller should finalize all the preparation */
-            std::atomic_thread_fence(std::memory_order_release);
             tcall.async_call([this, conn, client_fd](ThreadCall::Handle &) {
                 if (conn->mode == Conn::ConnMode::DEAD)
                 {
@@ -381,7 +380,6 @@ class ConnPool {
             auto ret = *(static_cast<conn_t *>(disp_tcall->call(
                         [this, addr](ThreadCall::Handle &h) {
                 auto conn = _connect(addr);
-                std::atomic_thread_fence(std::memory_order_release);
                 h.set_result(std::move(conn));
             }).get()));
             return std::move(ret);
diff --git a/include/salticidae/event.h b/include/salticidae/event.h
index fafa884..b97ab9a 100644
--- a/include/salticidae/event.h
+++ b/include/salticidae/event.h
@@ -481,6 +481,7 @@ class ThreadCall {
         }
         template<typename T>
         void set_result(T &&data) {
+            std::atomic_thread_fence(std::memory_order_release);
             using _T = std::remove_reference_t<T>;
             result = Result(new _T(std::forward<T>(data)),
                             [](void *ptr) {delete static_cast<_T *>(ptr);});
@@ -591,7 +592,7 @@ class MPSCQueueEventDriven: public MPSCQueue<T> {
     template<typename U>
     bool try_enqueue(U &&e) {
         static const uint64_t dummy = 1;
-        if (!MPMCQueue<T>::try_enqueue(std::forward<U>(e)))
+        if (!MPSCQueue<T>::try_enqueue(std::forward<U>(e)))
             return false;
         // memory barrier here, so any load/store in enqueue must be finialized
         if (wait_sig.exchange(false, std::memory_order_acq_rel))
diff --git a/include/salticidae/queue.h b/include/salticidae/queue.h
index 0b493ea..1732238 100644
--- a/include/salticidae/queue.h
+++ b/include/salticidae/queue.h
@@ -104,7 +104,8 @@ class MPMCQueue {
     template<typename U>
     void _enqueue(Block *nblk, U &&e) {
         new (&(nblk->elem)) T(std::forward<U>(e));
-        nblk->next.store(nullptr, std::memory_order_release);
+        std::atomic_thread_fence(std::memory_order_release);
+        nblk->next.store(nullptr, std::memory_order_relaxed);
         auto prev = tail.exchange(nblk, std::memory_order_acq_rel);
         prev->next.store(nblk, std::memory_order_relaxed);
     }
@@ -158,6 +159,7 @@ class MPMCQueue {
                     blks.release_ref(h);
                     return false;
                 }
+                std::atomic_thread_fence(std::memory_order_acquire);
                 e = std::move(nh->elem);
                 auto hh = h;
                 if (head.compare_exchange_weak(hh, nh, std::memory_order_consume))
@@ -180,6 +182,7 @@ struct MPSCQueue: public MPMCQueue<T> {
         auto nh = h->next.load(std::memory_order_relaxed);
         if (nh == nullptr)
             return false;
+        std::atomic_thread_fence(std::memory_order_acquire);
         e = std::move(nh->elem);
         this->head.store(nh, std::memory_order_release);
         this->blks.push(h);
@@ -194,6 +197,7 @@ struct MPSCQueue: public MPMCQueue<T> {
         auto h = this->head.load(std::memory_order_acquire);
         nblk->next.store(h, std::memory_order_release);
         new (&(h->elem)) T(std::forward<U>(e));
+        std::atomic_thread_fence(std::memory_order_release);
         this->head.store(nblk, std::memory_order_release);
         return true;
     }
-- 
cgit v1.2.3