Background:
In Hypervisor mode dGPU device is configured in pass through mode for
the Guest (QNX/Linux). GMMU programming is handled by the guest which
converts a mapped buffer's GVA into SGLes in IPA (Intermediate/Guest
Physical address) which is then translated into PA (Acutual Physical
address) and programs the GMMU PTEes with correct GVA to PA mapping.
Incase of the vgpu this work is delegated to the RM server which takes care
of the GMMU programming and IPA to PA conversion.
Problem:
The current GMMU mapping logic in the guest assumes that PA range is
continuous over a given IPA range. Hence, it doesn't account for holes being
present in the PA range. But this is not the case, a continous IPA range
can be mapped to dis-contiguous PA ranges. In this situation the mapping
logic sets up GMMU PTEes ignoring the holes in physical memory and
creates GVA => PA mapping which intrudes into the PA ranges which are
reserved. This results in memory being corrupted.
This change takes into account holes being present in a given PA range and
for a given IPA range it also identifies the discontiguous PA ranges and
sets up the PTE's appropriately.
Bug 200451447
Jira VQRM-5069
Change-Id: I354d984f6c44482e4576a173fce1e90ab52283ac
Signed-off-by: aalex <aalex@nvidia.com>
Signed-off-by: Antony Clince Alex <aalex@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1850972
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
