On Wed, 2011-11-23 at 16:31 +0100, Daniel Vetter wrote: > > Things I've tried that don't work around the issue: > - Disable dmar for the igfx with intel_iommu=igfx_off > - Apply the ilk workaround (i.e. synchronous dmar tlb flushes + gpu > idling while flushing). Well, the ILK workaround was only ensuring that the gfx was idle while the *graphics* IOTLB was flushed. We are still flushing the IOTLB for *other* IOMMUs at arbitrary times. Might be interesting to go for *deferred* IOTLB flushes (i.e. the normal mode without the workaround or 'intel_iommu=strict' on the command line), *BUT* to make sure the batched flush (the flush_unmaps() call) can *only* happen when the GPU is quiescent. You could do that with a dirty hack for testing, and if *that* fixes the issue.... well, $DEITY knows, someone still needs a good kicking, but at least we'll have learned something :) -- dwmw2 -------------- next part -------------- A non-text attachment was scrubbed... Name: smime.p7s Type: application/x-pkcs7-signature Size: 5818 bytes Desc: not available URL: <http://lists.freedesktop.org/archives/intel-gfx/attachments/20111123/be93cddf/attachment-0001.bin>
