On Thu, Nov 03, 2011 at 03:23:39PM -0700, Ben Widawsky wrote: > On Thu, 3 Nov 2011 20:19:23 +0000 > Dave Airlie <airlied at gmail.com> wrote: > > > >> > > > >> > The solution here is to add a new flag to the call chain which gives the > > >> > routines the information they need to possibly defer actions which may > > >> > cause us to recurse. A macro has been defined to replace i915_gpu_idle > > >> > which defaults to the old behavior. > > >> > > > >> > Kudos to Chris for tracking this one down. > > >> > > >> So this fixes the non-VTd case, the VT-d case still hits a recursion > > >> here, for posterity its below. > > > > Okay I take that back, I got my EL6 kernel rock stable with the > > correct blend of backported bits. > > > > So ignore that backtrace, however I did get another IOMMU hang on my > > upstream kernel with gem_linear_blits, > > > > so this should be fine to merge but I'm guessing we have more > > debugging to do on the VT-d cases. > > > > Dave. > > Does it pass your original failing case? Hi Ben, Is v3 of this patch the right one to merge or do we still have some outstanding issues on this? Also, have you looked at the recent ilk dmar fallout, I think strict dmar iotlb flushing doesn't sit too well with our gpus. My snb dies almost immediately with that (even when I do not enable rc6 nor semaphores). Cheers, Daniel -- Daniel Vetter Mail: daniel at ffwll.ch Mobile: +41 (0)79 365 57 48
