ntdll: Use an acquire/release pair on the IOSB status.

dlls/kernelbase/file.c

@@ -3156,7 +3156,9 @@ BOOL WINAPI DECLSPEC_HOTPATCH GetOverlappedResultEx( HANDLE file, OVERLAPPED *ov
 
     TRACE( "(%p %p %p %lu %d)\n", file, overlapped, result, timeout, alertable );
 
-    status = overlapped->Internal;
+    /* Paired with the write-release in set_async_iosb() in ntdll; see the
+     * latter for details. */
+    status = ReadAcquire( (LONG *)&overlapped->Internal );
     if (status == STATUS_PENDING)
     {
         if (!timeout)
@@ -3173,6 +3175,8 @@ BOOL WINAPI DECLSPEC_HOTPATCH GetOverlappedResultEx( HANDLE file, OVERLAPPED *ov
             return FALSE;
         }
 
+        /* We don't need to give this load acquire semantics; the wait above
+         * already guarantees that the IOSB and output buffer are filled. */
         status = overlapped->Internal;
         if (status == STATUS_PENDING) status = STATUS_SUCCESS;
     }

dlls/ntdll/unix/unix_private.h

@@ -367,6 +367,24 @@ static inline void set_async_iosb( client_ptr_t iosb, NTSTATUS status, ULONG_PTR
 {
     if (!iosb) return;
 
+    /* GetOverlappedResult() and WSAGetOverlappedResult() expect that if the
+     * status is written, that the information (and buffer, which was written
+     * earlier from the async callback) will be available. Hence we need to
+     * store the status last, with release semantics to ensure that those
+     * writes are visible. This release is paired with a read-acquire in
+     * GetOverlappedResult() and WSAGetOverlappedResult():
+     *
+     * CPU 0 (set_async_iosb)            CPU 1 (GetOverlappedResultEx)
+     * ===========================       ===========================
+     * write buffer
+     * write Information
+     * WriteRelease(Status) <--------.
+     *                               |
+     *                               |
+     *                (paired with)  `-> ReadAcquire(Status)
+     *                                   read Information
+     */
+
     if (in_wow64_call())
     {
         struct iosb32
@@ -374,18 +392,18 @@ static inline void set_async_iosb( client_ptr_t iosb, NTSTATUS status, ULONG_PTR
             NTSTATUS Status;
             ULONG    Information;
         } *io = wine_server_get_ptr( iosb );
-        io->Status = status;
         io->Information = info;
+        WriteRelease( &io->Status, status );
     }
     else
     {
         IO_STATUS_BLOCK *io = wine_server_get_ptr( iosb );
+        io->Information = info;
 #ifdef NONAMELESSUNION
-        io->u.Status = status;
+        WriteRelease( &io->u.Status, status );
 #else
-        io->Status = status;
+        WriteRelease( &io->Status, status );
 #endif
-        io->Information = info;
     }
 }
 

dlls/ws2_32/socket.c

@@ -3710,7 +3710,9 @@ BOOL WINAPI WSAGetOverlappedResult( SOCKET s, LPWSAOVERLAPPED lpOverlapped,
         return FALSE;
     }
 
-    status = lpOverlapped->Internal;
+    /* Paired with the write-release in set_async_iosb() in ntdll; see the
+     * latter for details. */
+    status = ReadAcquire( (LONG *)&lpOverlapped->Internal );
     if (status == STATUS_PENDING)
     {
         if (!fWait)
@@ -3722,6 +3724,8 @@ BOOL WINAPI WSAGetOverlappedResult( SOCKET s, LPWSAOVERLAPPED lpOverlapped,
         if (WaitForSingleObject( lpOverlapped->hEvent ? lpOverlapped->hEvent : SOCKET2HANDLE(s),
                                  INFINITE ) == WAIT_FAILED)
             return FALSE;
+        /* We don't need to give this load acquire semantics; the wait above
+         * already guarantees that the IOSB and output buffer are filled. */
         status = lpOverlapped->Internal;
     }