diff --git a/dlls/wined3d/adapter_gl.c b/dlls/wined3d/adapter_gl.c
index 80ba61414eb..b794415fb3d 100644
--- a/dlls/wined3d/adapter_gl.c
+++ b/dlls/wined3d/adapter_gl.c
@@ -58,6 +58,7 @@ static const struct wined3d_extension_map gl_extension_map[] =
     /* ARB */
     {"GL_ARB_base_instance",                ARB_BASE_INSTANCE             },
     {"GL_ARB_blend_func_extended",          ARB_BLEND_FUNC_EXTENDED       },
+    {"GL_ARB_buffer_storage",               ARB_BUFFER_STORAGE            },
     {"GL_ARB_clear_buffer_object",          ARB_CLEAR_BUFFER_OBJECT       },
     {"GL_ARB_clear_texture",                ARB_CLEAR_TEXTURE             },
     {"GL_ARB_clip_control",                 ARB_CLIP_CONTROL              },
@@ -103,6 +104,7 @@ static const struct wined3d_extension_map gl_extension_map[] =
     {"GL_ARB_point_parameters",             ARB_POINT_PARAMETERS          },
     {"GL_ARB_point_sprite",                 ARB_POINT_SPRITE              },
     {"GL_ARB_provoking_vertex",             ARB_PROVOKING_VERTEX          },
+    {"GL_ARB_query_buffer_object",          ARB_QUERY_BUFFER_OBJECT       },
     {"GL_ARB_sample_shading",               ARB_SAMPLE_SHADING            },
     {"GL_ARB_sampler_objects",              ARB_SAMPLER_OBJECTS           },
     {"GL_ARB_seamless_cube_map",            ARB_SEAMLESS_CUBE_MAP         },
@@ -2061,6 +2063,8 @@ static void load_gl_funcs(struct wined3d_gl_info *gl_info)
     /* GL_ARB_blend_func_extended */
     USE_GL_FUNC(glBindFragDataLocationIndexed)
     USE_GL_FUNC(glGetFragDataIndex)
+    /* GL_ARB_buffer_storage */
+    USE_GL_FUNC(glBufferStorage)
     /* GL_ARB_clear_buffer_object */
     USE_GL_FUNC(glClearBufferData)
     USE_GL_FUNC(glClearBufferSubData)
@@ -3301,7 +3305,9 @@ static BOOL wined3d_adapter_init_gl_caps(struct wined3d_adapter *adapter,
         {ARB_TEXTURE_STORAGE_MULTISAMPLE,  MAKEDWORD_VERSION(4, 2)},
         {ARB_TEXTURE_VIEW,                 MAKEDWORD_VERSION(4, 3)},
 
+        {ARB_BUFFER_STORAGE,               MAKEDWORD_VERSION(4, 4)},
         {ARB_CLEAR_TEXTURE,                MAKEDWORD_VERSION(4, 4)},
+        {ARB_QUERY_BUFFER_OBJECT,          MAKEDWORD_VERSION(4, 4)},
 
         {ARB_CLIP_CONTROL,                 MAKEDWORD_VERSION(4, 5)},
         {ARB_CULL_DISTANCE,                MAKEDWORD_VERSION(4, 5)},
diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
index ff4d562f106..2aae559a6cb 100644
--- a/dlls/wined3d/cs.c
+++ b/dlls/wined3d/cs.c
@@ -1986,7 +1986,10 @@ static void wined3d_cs_exec_query_issue(struct wined3d_cs *cs, const void *data)
 
     if (poll && list_empty(&query->poll_list_entry))
     {
-        list_add_tail(&cs->query_poll_list, &query->poll_list_entry);
+        if (query->buffer_object)
+            InterlockedIncrement(&query->counter_retrieved);
+        else
+            list_add_tail(&cs->query_poll_list, &query->poll_list_entry);
         return;
     }
 
diff --git a/dlls/wined3d/query.c b/dlls/wined3d/query.c
index 01e6bcb9905..26c32b1f13b 100644
--- a/dlls/wined3d/query.c
+++ b/dlls/wined3d/query.c
@@ -25,6 +25,88 @@
 
 WINE_DEFAULT_DEBUG_CHANNEL(d3d);
 
+static void wined3d_query_buffer_invalidate(struct wined3d_query *query)
+{
+    /* map[0] != map[1]: exact values do not have any significance. */
+    query->map_ptr[0] = 0;
+    query->map_ptr[1] = ~(UINT64)0;
+}
+
+static BOOL wined3d_query_buffer_is_valid(struct wined3d_query *query)
+{
+    return query->map_ptr[0] == query->map_ptr[1];
+}
+
+static void wined3d_query_create_buffer_object(struct wined3d_context *context, struct wined3d_query *query)
+{
+    const struct wined3d_gl_info *gl_info = context->gl_info;
+    const GLuint map_flags = GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT;
+    GLuint buffer_object;
+
+    GL_EXTCALL(glGenBuffers(1, &buffer_object));
+    GL_EXTCALL(glBindBuffer(GL_QUERY_BUFFER, buffer_object));
+    GL_EXTCALL(glBufferStorage(GL_QUERY_BUFFER, sizeof(query->map_ptr[0]) * 2, NULL, map_flags));
+    query->map_ptr = GL_EXTCALL(glMapBufferRange(GL_QUERY_BUFFER, 0, sizeof(query->map_ptr[0]) * 2, map_flags));
+    GL_EXTCALL(glBindBuffer(GL_QUERY_BUFFER, 0));
+    checkGLcall("query buffer object creation");
+
+    wined3d_query_buffer_invalidate(query);
+    query->buffer_object = buffer_object;
+}
+
+static void wined3d_query_destroy_buffer_object(struct wined3d_context *context, struct wined3d_query *query)
+{
+    const struct wined3d_gl_info *gl_info = context->gl_info;
+
+    GL_EXTCALL(glDeleteBuffers(1, &query->buffer_object));
+    checkGLcall("query buffer object destruction");
+
+    query->buffer_object = 0;
+    query->map_ptr = NULL;
+}
+
+/* From ARB_occlusion_query: "Querying the state for a given occlusion query
+ * forces that occlusion query to complete within a finite amount of time."
+ * In practice, that means drivers flush when retrieving
+ * GL_QUERY_RESULT_AVAILABLE, which can be undesirable when applications use a
+ * significant number of queries. Using a persistently mapped query buffer
+ * object allows us to avoid these implicit flushes. An additional benefit is
+ * that it allows us to poll the query status from the application-thread
+ * instead of from the csmt-thread. */
+static BOOL wined3d_query_buffer_queue_result(struct wined3d_context *context, struct wined3d_query *query, GLuint id)
+{
+    const struct wined3d_gl_info *gl_info = context->gl_info;
+
+    if (!gl_info->supported[ARB_QUERY_BUFFER_OBJECT] || !gl_info->supported[ARB_BUFFER_STORAGE])
+        return FALSE;
+    /* Don't use query buffers without CSMT, mainly for simplicity. */
+    if (!context->device->cs->thread)
+        return FALSE;
+
+    if (query->buffer_object)
+    {
+        /* If there's still a query result in-flight for the existing buffer
+         * object (i.e., the query was restarted before we received its
+         * result), we can't reuse the existing buffer object. */
+        if (wined3d_query_buffer_is_valid(query))
+            wined3d_query_buffer_invalidate(query);
+        else
+            wined3d_query_destroy_buffer_object(context, query);
+    }
+
+    if (!query->buffer_object)
+        wined3d_query_create_buffer_object(context, query);
+
+    GL_EXTCALL(glBindBuffer(GL_QUERY_BUFFER, query->buffer_object));
+    /* Read the same value twice. We know we have the result if map_ptr[0] == map_ptr[1]. */
+    GL_EXTCALL(glGetQueryObjectui64v(id, GL_QUERY_RESULT, (void *)0));
+    GL_EXTCALL(glGetQueryObjectui64v(id, GL_QUERY_RESULT, (void *)sizeof(query->map_ptr[0])));
+    GL_EXTCALL(glBindBuffer(GL_QUERY_BUFFER, 0));
+    checkGLcall("queue query result");
+
+    return TRUE;
+}
+
 static UINT64 get_query_result64(GLuint id, const struct wined3d_gl_info *gl_info)
 {
     if (gl_info->supported[ARB_TIMER_QUERY])
@@ -339,6 +421,14 @@ static void wined3d_query_destroy_object(void *object)
     if (!list_empty(&query->poll_list_entry))
         list_remove(&query->poll_list_entry);
 
+    if (query->buffer_object)
+    {
+        struct wined3d_context *context;
+        context = context_acquire(query->device, NULL, 0);
+        wined3d_query_destroy_buffer_object(context, query);
+        context_release(context);
+    }
+
     /* Queries are specific to the GL context that created them. Not
      * deleting the query will obviously leak it, but that's still better
      * than potentially deleting a different query with the same id in this
@@ -382,15 +472,20 @@ HRESULT CDECL wined3d_query_get_data(struct wined3d_query *query,
         return WINED3DERR_INVALIDCALL;
     }
 
-    if (!query->device->cs->thread)
+    if (query->device->cs->thread)
     {
-        if (!query->query_ops->query_poll(query, flags))
+        if (query->counter_main != query->counter_retrieved
+                || (query->buffer_object && !wined3d_query_buffer_is_valid(query)))
+        {
+            if (flags & WINED3DGETDATA_FLUSH && !query->device->cs->queries_flushed)
+                wined3d_cs_emit_flush(query->device->cs);
             return S_FALSE;
+        }
+        if (query->buffer_object)
+            query->data = query->map_ptr;
     }
-    else if (query->counter_main != query->counter_retrieved)
+    else if (!query->query_ops->query_poll(query, flags))
     {
-        if (flags & WINED3DGETDATA_FLUSH && !query->device->cs->queries_flushed)
-            wined3d_cs_emit_flush(query->device->cs);
         return S_FALSE;
     }
 
@@ -579,6 +674,7 @@ static BOOL wined3d_occlusion_query_ops_issue(struct wined3d_query *query, DWORD
                 gl_info = context->gl_info;
                 GL_EXTCALL(glEndQuery(GL_SAMPLES_PASSED));
                 checkGLcall("glEndQuery()");
+                wined3d_query_buffer_queue_result(context, query, oq->id);
 
                 context_release(context);
                 poll = TRUE;
diff --git a/dlls/wined3d/wined3d_gl.h b/dlls/wined3d/wined3d_gl.h
index 06aa21c33fa..2a985da9faf 100644
--- a/dlls/wined3d/wined3d_gl.h
+++ b/dlls/wined3d/wined3d_gl.h
@@ -44,6 +44,7 @@ enum wined3d_gl_extension
     /* ARB */
     ARB_BASE_INSTANCE,
     ARB_BLEND_FUNC_EXTENDED,
+    ARB_BUFFER_STORAGE,
     ARB_CLEAR_BUFFER_OBJECT,
     ARB_CLEAR_TEXTURE,
     ARB_CLIP_CONTROL,
@@ -89,6 +90,7 @@ enum wined3d_gl_extension
     ARB_POINT_PARAMETERS,
     ARB_POINT_SPRITE,
     ARB_PROVOKING_VERTEX,
+    ARB_QUERY_BUFFER_OBJECT,
     ARB_SAMPLE_SHADING,
     ARB_SAMPLER_OBJECTS,
     ARB_SEAMLESS_CUBE_MAP,
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 941bfa68fad..8cfa4c37ef8 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -1730,6 +1730,9 @@ struct wined3d_query
 
     LONG counter_main, counter_retrieved;
     struct list poll_list_entry;
+
+    GLuint buffer_object;
+    UINT64 *map_ptr;
 };
 
 struct wined3d_event_query