On 830/845, the BLT unit invalidates the wrong PTE in its TLB after the GATT is updated. A simple solution is then to always invalidate the TLB of the BLT prior to each execbuffer.
This does appear to improve the stability slighty, but I am still seeing spurious GPU deaths under memory pressure. References: https://bugs.freedesktop.org/show_bug.cgi?id=26345 Signed-off-by: Chris Wilson <[email protected]> --- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c index 60fc132..b825c06 100644 --- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c @@ -883,6 +883,13 @@ i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring, int ret; memset(&cd, 0, sizeof(cd)); + + /* We need to invalidate the BLT's prefetched entries after + * updating the GATT (as the hardware invalidates the wrong PTEs). + */ + if (IS_I830(ring->dev) || IS_845G(ring->dev)) + cd.invalidate_domains = I915_GEM_DOMAIN_RENDER; + list_for_each_entry(obj, objects, exec_list) i915_gem_object_set_to_gpu_domain(obj, ring, &cd); -- 1.7.10 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
