On 7/15/20 1:51 PM, Chris Wilson wrote:
It is reasonably common for userspace (even modern drivers like iris) to reuse an active address for a new buffer. This would cause the application to stall under its mutex (originally struct_mutex) until the old batches were idle and it could synchronously remove the stale PTE. However, we can queue up a job that waits on the signal for the old nodes to complete and upon those signals, remove the old nodes replacing them with the new ones for the batch. This is still CPU driven, but in theory we can do the GTT patching from the GPU. The job itself has a completion signal allowing the execbuf to wait upon the rebinding, and also other observers to coordinate with the common VM activity. Letting userspace queue up more work, lets it do more stuff without blocking other clients. In turn, we take care not to let it too much concurrent work, creating a small number of queues for each context to limit the number of concurrent tasks. The implementation relies on only scheduling one unbind operation per vma as we use the unbound vma->node location to track the stale PTE. Closes: https://gitlab.freedesktop.org/drm/intel/issues/1402 Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> Cc: Matthew Auld <matthew.auld@xxxxxxxxx> Cc: Andi Shyti <andi.shyti@xxxxxxxxx>
Reviewed-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxx> _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
