Add fence barriers for the repl backlog (important for AARCH64 and other weak memory models)

src/networking.cpp

@@ -1810,6 +1810,8 @@ int writeToClient(client *c, int handler_installed) {
        is a replica, so only attempt to do so if that's the case. */
     if (c->flags & CLIENT_SLAVE && !(c->flags & CLIENT_MONITOR) && c->replstate == SLAVE_STATE_ONLINE) {
         std::unique_lock<fastlock> repl_backlog_lock (g_pserver->repl_backlog_lock);
+        // Ensure all writes to the repl backlog are visible
+        std::atomic_thread_fence(std::memory_order_acquire);
 
         while (clientHasPendingReplies(c)) {
             long long repl_end_idx = getReplIndexFromOffset(c->repl_end_off);

src/replication.cpp

@@ -5599,6 +5599,9 @@ void flushReplBacklogToClients()
         serverAssert(g_pserver->master_repl_offset - g_pserver->repl_batch_offStart <= g_pserver->repl_backlog_size);
         serverAssert(g_pserver->repl_batch_idxStart != g_pserver->repl_backlog_idx);
 
+        // Repl backlog writes must become visible to all threads at this point
+        std::atomic_thread_fence(std::memory_order_release);
+
         listIter li;
         listNode *ln;
         listRewind(g_pserver->slaves, &li);