drm/i915: Force CPU synchronisation even if userspace requests ASYNC
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 21 Jul 2017 14:50:37 +0000 (15:50 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 27 Jul 2017 20:07:24 +0000 (22:07 +0200)
The goal here was to minimise doing any thing or any check inside the
kernel that was not strictly required. For a userspace that assumes
complete control over the cache domains, the kernel is usually using
outdated information and may trigger clflushes where none were
required.

However, swapping is a situation where userspace has no knowledge of the
domain transfer, and will leave the object in the CPU cache. The kernel
must flush this out to the backing storage prior to use with the GPU. As
we use an asynchronous task tracked by an implicit fence for this, we
also need to cancel the ASYNC flag on the object so that the object will
wait for the clflush to complete before being executed. This also absolves
userspace of the responsibility imposed by commit 77ae9957897d ("drm/i915:
Enable userspace to opt-out of implicit fencing") that its needed to ensure
that the object was out of the CPU cache prior to use on the GPU.

Fixes: 77ae9957897d ("drm/i915: Enable userspace to opt-out of implicit fencing")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101571
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20170721145037.25105-5-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
(cherry picked from commit 0f46daa1a273779a0b73d768a788ca3f04238f9c)
Cc: stable@vger.kernel.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem_clflush.c
drivers/gpu/drm/i915/i915_gem_clflush.h
drivers/gpu/drm/i915/i915_gem_execbuffer.c

index 152f16c11878ef937e48b9139c9b8ab844322df3..348b29a845c961c73b1605b23212692b438f0f19 100644 (file)
@@ -114,7 +114,7 @@ i915_clflush_notify(struct i915_sw_fence *fence,
        return NOTIFY_DONE;
 }
 
-void i915_gem_clflush_object(struct drm_i915_gem_object *obj,
+bool i915_gem_clflush_object(struct drm_i915_gem_object *obj,
                             unsigned int flags)
 {
        struct clflush *clflush;
@@ -128,7 +128,7 @@ void i915_gem_clflush_object(struct drm_i915_gem_object *obj,
         */
        if (!i915_gem_object_has_struct_page(obj)) {
                obj->cache_dirty = false;
-               return;
+               return false;
        }
 
        /* If the GPU is snooping the contents of the CPU cache,
@@ -140,7 +140,7 @@ void i915_gem_clflush_object(struct drm_i915_gem_object *obj,
         * tracking.
         */
        if (!(flags & I915_CLFLUSH_FORCE) && obj->cache_coherent)
-               return;
+               return false;
 
        trace_i915_gem_object_clflush(obj);
 
@@ -179,4 +179,5 @@ void i915_gem_clflush_object(struct drm_i915_gem_object *obj,
        }
 
        obj->cache_dirty = false;
+       return true;
 }
index 2455a7820937b0f461839de713693a9010950997..f390247561b37691d615fe0182d47cf40bcc4277 100644 (file)
@@ -28,7 +28,7 @@
 struct drm_i915_private;
 struct drm_i915_gem_object;
 
-void i915_gem_clflush_object(struct drm_i915_gem_object *obj,
+bool i915_gem_clflush_object(struct drm_i915_gem_object *obj,
                             unsigned int flags);
 #define I915_CLFLUSH_FORCE BIT(0)
 #define I915_CLFLUSH_SYNC BIT(1)
index d70ac429106f3f0a97e75f38b6cc9a1e557398e9..e9503f6d110016b86311003349ff5b8ef70613d5 100644 (file)
@@ -1825,7 +1825,7 @@ static int eb_move_to_gpu(struct i915_execbuffer *eb)
        int err;
 
        for (i = 0; i < count; i++) {
-               const struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
+               struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
                struct i915_vma *vma = exec_to_vma(entry);
                struct drm_i915_gem_object *obj = vma->obj;
 
@@ -1841,12 +1841,14 @@ static int eb_move_to_gpu(struct i915_execbuffer *eb)
                        eb->request->capture_list = capture;
                }
 
+               if (unlikely(obj->cache_dirty && !obj->cache_coherent)) {
+                       if (i915_gem_clflush_object(obj, 0))
+                               entry->flags &= ~EXEC_OBJECT_ASYNC;
+               }
+
                if (entry->flags & EXEC_OBJECT_ASYNC)
                        goto skip_flushes;
 
-               if (unlikely(obj->cache_dirty && !obj->cache_coherent))
-                       i915_gem_clflush_object(obj, 0);
-
                err = i915_gem_request_await_object
                        (eb->request, obj, entry->flags & EXEC_OBJECT_WRITE);
                if (err)