Skip to content

Use the flat VRAM identity mapping for migration and ditch 2M PTE support

Thomas Hellström requested to merge thomash/kernel:identity into xe

Using the identity mapping helps us avoid emitting VRAM PTEs and for VRAM only operations we can also blit preemptible and in larger chunks, since we no longer need to make page-table-updates + blits an atomic operation.

This makes the 2M mapping mode unneeded so remove it and its selftest.

v2:

  • Add a commit to remove support for 64K PTEs except for mapping the page-table bo itself. v3:
  • Always use the 8MiB blit limit (It seems like the limit otherwise used would cause lockups and other strange behaviour on DG2).
  • There was a bug in how the blit start address was calculated from the resource cursor. Fix.

Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

Edited by Thomas Hellström

Merge request reports