Add some details around non-LLC platforms and cflushing, when dealing
with the flush-on-acquire, which is potentially security sensitive.
Signed-off-by: Matthew Auld <matthew.auld@xxxxxxxxx>
Cc: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
Cc: Daniel Vetter <daniel@xxxxxxxx>
---
.../gpu/drm/i915/gem/i915_gem_execbuffer.c | 11 ++++++++
.../gpu/drm/i915/gem/i915_gem_object_types.h | 27 +++++++++++++++++++
2 files changed, 38 insertions(+)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
index 1231224728e4..9c323666bd7c 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
@@ -1933,6 +1933,17 @@ static int eb_move_to_gpu(struct i915_execbuffer *eb)
* !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)
* but gcc's optimiser doesn't handle that as well and emits
* two jumps instead of one. Maybe one day...
+ *
+ * FIXME: There is also sync flushing in set_pages(), which
+ * serves a different purpose(some of the time at least).
+ *
+ * We should consider:
+ *
+ * 1. Rip out the async flush code.
+ *
+ * 2. Or make the sync flushing use the async clflush path
+ * using mandatory fences underneath. Currently the below
+ * async flush happens after we bind the object.
*/
if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) {
if (i915_gem_clflush_object(obj, 0))
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h
index 7c3da4e3e737..da85169006d4 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h
+++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h
@@ -427,6 +427,33 @@ struct drm_i915_gem_object {
* can freely bypass the CPU cache when touching the pages with the GPU,
* where the kernel is completely unaware. On such platform we need
* apply the sledgehammer-on-acquire regardless of the @cache_coherent.
+ *
+ * Special care is taken on non-LLC platforms, to prevent potential
+ * information leak. The driver currently ensures:
+ *
+ * 1. All userspace objects, by default, have @cache_level set as
+ * I915_CACHE_NONE. The only exception is userptr objects, where we
+ * instead force I915_CACHE_LLC, but we also don't allow userspace to
+ * ever change the @cache_level for such objects. Another special case
+ * is dma-buf, which doesn't rely on @cache_dirty, but there we
+ * always do a forced flush when acquiring the pages, if there is a
+ * chance that the pages can be read directly from main memory with
+ * the GPU.
+ *
+ * 2. All I915_CACHE_NONE objects have @cache_dirty initially true.
+ *
+ * 3. All swapped-out objects(i.e shmem) have @cache_dirty set to
+ * true.
+ *
+ * 4. The @cache_dirty is never freely reset before the initial
+ * flush, even if userspace adjusts the @cache_level through the
+ * i915_gem_set_caching_ioctl.
+ *
+ * 5. All @cache_dirty objects(including swapped-in) are initially
+ * flushed with a synchronous call to drm_clflush_sg in
+ * __i915_gem_object_set_pages. The @cache_dirty can be freely reset
+ * at this point. All further asynchronous clfushes are never security
+ * critical, i.e userspace is free to race against itself.
*/
unsigned int cache_dirty:1;
--
2.26.3
