"Goel, Akash" <akash.goel@xxxxxxxxx> writes: >>> What happens when we GTT-mapped write a batchbuffer that had > previously been silently made RO by the kernel? Does the CPU take a > fault? We tested this particular case, doing the relocation inside > the Batch buffer, which is mapped to GTT as read-only, from the > exec-buffer path. On doing so, there were no faults observed on the > CPU side. > > Also we don't see any ring hangs, when forcing the GPU to do a write > access on the buffers marked as RO. Just the writes were getting > rejected. Interesting. Good to know. So far we've never done self-modifying batches. If we could do it, we had some interesting ideas for handling some classes of state updates, but I think at least at one point the kernel was banning self-modifying batches.
Attachment:
pgpXsvUM4zQVV.pgp
Description: PGP signature
_______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
