ntdll: Cache the esync struct itself instead of a pointer to it.

Now that the esync object has a constant size, we can perform this
optimization.

dlls/ntdll/esync.c

@@ -156,11 +156,11 @@ static void *get_shm( unsigned int idx )
 /* We'd like lookup to be fast. To that end, we use a static list indexed by handle.
  * This is copied and adapted from the fd cache code. */
 
-#define ESYNC_LIST_BLOCK_SIZE  (65536 / sizeof(struct esync *))
-#define ESYNC_LIST_ENTRIES     128
+#define ESYNC_LIST_BLOCK_SIZE  (65536 / sizeof(struct esync))
+#define ESYNC_LIST_ENTRIES     256
 
-static struct esync * *esync_list[ESYNC_LIST_ENTRIES];
-static struct esync * esync_list_initial_block[ESYNC_LIST_BLOCK_SIZE];
+static struct esync *esync_list[ESYNC_LIST_ENTRIES];
+static struct esync esync_list_initial_block[ESYNC_LIST_BLOCK_SIZE];
 
 static inline UINT_PTR handle_to_index( HANDLE handle, UINT_PTR *entry )
 {
@@ -169,7 +169,7 @@ static inline UINT_PTR handle_to_index( HANDLE handle, UINT_PTR *entry )
     return idx % ESYNC_LIST_BLOCK_SIZE;
 }
 
-static BOOL add_to_list( HANDLE handle, struct esync *obj )
+static struct esync *add_to_list( HANDLE handle, enum esync_type type, int fd, void *shm )
 {
     UINT_PTR entry, idx = handle_to_index( handle, &entry );
 
@@ -191,18 +191,22 @@ static BOOL add_to_list( HANDLE handle, struct esync *obj )
         }
     }
 
-    obj = interlocked_xchg_ptr((void **)&esync_list[entry][idx], obj);
-    assert(!obj);
-    return TRUE;
+    if (!interlocked_cmpxchg((int *)&esync_list[entry][idx].type, type, 0))
+    {
+        esync_list[entry][idx].fd = fd;
+        esync_list[entry][idx].shm = shm;
+    }
+    return &esync_list[entry][idx];
 }
 
 static struct esync *get_cached_object( HANDLE handle )
 {
     UINT_PTR entry, idx = handle_to_index( handle, &entry );
 
     if (entry >= ESYNC_LIST_ENTRIES || !esync_list[entry]) return NULL;
+    if (!esync_list[entry][idx].type) return NULL;
 
-    return esync_list[entry][idx];
+    return &esync_list[entry][idx];
 }
 
 /* Gets an object. This is either a proper esync object (i.e. an event,
@@ -214,19 +218,18 @@ static struct esync *get_cached_object( HANDLE handle )
  * since we might be passed a duplicated or inherited handle. */
 static NTSTATUS get_object( HANDLE handle, struct esync **obj )
 {
+    NTSTATUS ret = STATUS_SUCCESS;
+    enum esync_type type = 0;
+    unsigned int shm_idx = 0;
     obj_handle_t fd_handle;
-    enum esync_type type;
-    unsigned int shm_idx;
-    struct esync *esync;
     sigset_t sigset;
-    NTSTATUS ret;
-    int fd;
+    int fd = -1;
 
     if ((*obj = get_cached_object( handle ))) return STATUS_SUCCESS;
 
     /* We need to try grabbing it from the server. */
     server_enter_uninterrupted_section( &fd_cache_section, &sigset );
-    if (!(esync = get_cached_object( handle )))
+    if (!(*obj = get_cached_object( handle )))
     {
         SERVER_START_REQ( get_esync_fd )
         {
@@ -243,10 +246,9 @@ static NTSTATUS get_object( HANDLE handle, struct esync **obj )
     }
     server_leave_uninterrupted_section( &fd_cache_section, &sigset );
 
-    if (esync)
+    if (*obj)
     {
         /* We managed to grab it while in the CS; return it. */
-        *obj = esync;
         return STATUS_SUCCESS;
     }
 
@@ -259,31 +261,21 @@ static NTSTATUS get_object( HANDLE handle, struct esync **obj )
 
     TRACE("Got fd %d for handle %p.\n", fd, handle);
 
-    esync = RtlAllocateHeap( GetProcessHeap(), 0, sizeof(*esync) );
-    esync->fd = fd;
-    esync->type = type;
-
-    esync->shm = shm_idx ? get_shm( shm_idx ) : 0;
-
-    add_to_list( handle, esync );
-
-    *obj = esync;
+    *obj = add_to_list( handle, type, fd, shm_idx ? get_shm( shm_idx ) : 0 );
     return ret;
 }
 
 NTSTATUS esync_close( HANDLE handle )
 {
     UINT_PTR entry, idx = handle_to_index( handle, &entry );
-    struct esync *obj;
 
     TRACE("%p.\n", handle);
 
     if (entry < ESYNC_LIST_ENTRIES && esync_list[entry])
     {
-        if ((obj = interlocked_xchg_ptr( (void **)&esync_list[entry][idx], 0 )))
+        if (interlocked_xchg((int *)&esync_list[entry][idx].type, 0))
         {
-            close( obj->fd );
-            RtlFreeHeap( GetProcessHeap(), 0, obj );
+            close( esync_list[entry][idx].fd );
             return STATUS_SUCCESS;
         }
     }
@@ -329,14 +321,7 @@ static NTSTATUS create_esync( enum esync_type type, HANDLE *handle,
 
     if (!ret || ret == STATUS_OBJECT_NAME_EXISTS)
     {
-        struct esync *obj = RtlAllocateHeap( GetProcessHeap(), 0, sizeof(*obj) );
-        if (!obj) return STATUS_NO_MEMORY;
-
-        obj->type = type;
-        obj->fd = fd;
-        obj->shm = shm_idx ? get_shm( shm_idx ) : 0;
-
-        add_to_list( *handle, obj );
+        add_to_list( *handle, type, fd, shm_idx ? get_shm( shm_idx ) : 0 );
 
         TRACE("-> handle %p, fd %d, shm index %d.\n", *handle, fd, shm_idx);
     }
@@ -377,14 +362,7 @@ static NTSTATUS open_esync( enum esync_type type, HANDLE *handle,
 
     if (!ret)
     {
-        struct esync *obj = RtlAllocateHeap( GetProcessHeap(), 0, sizeof(*obj) );
-        if (!obj) return STATUS_NO_MEMORY;
-
-        obj->type = type;
-        obj->fd = fd;
-        obj->shm = shm_idx ? get_shm( shm_idx ) : 0;
-
-        add_to_list( *handle, obj );
+        add_to_list( *handle, type, fd, shm_idx ? get_shm( shm_idx ) : 0 );
 
         TRACE("-> handle %p, fd %d.\n", *handle, fd);
     }