On 2018-03-05 12:14 PM, Christian König wrote: >> Also, we'd need to fix shadow page table handling with CPU updates, or >> get rid of shadow page tables. I'm not sure why they're needed, TBH. > > The idea behind shadow page tables is that you have the current state > of the pipeline if the GPU crashes. I see. So a GPU can be reset in the middle of a page table update without losing anything and without having to recreate the page table from scratch. > Since with CPU updates there is no pipeline they doesn't make sense at > all with them. That's true when all page table updates are done by the CPU. If we start mixing CPU and SDMA updates in the same VM, then CPU updates need to update the shadow page table as well. Regards,  Felix
