Hi Siqueira,
as discussed on the call if you can wrap your head around how the
amdgpu_device_enforce_isolation() function works it should be trivial to write
a new function or extend the function to insert a CPU bubble whenever the
ownership of one of the compute rings change.
IIRC we already
Limiting the number of available VMIDs to enforce isolation causes some
issues with gang submit and applying certain HW workarounds which
require multiple VMIDs to work correctly.
So instead start to track all submissions to the relevant engines in a
per partition data structure and use the dma_fe
Limiting the number of available VMIDs to enforce isolation causes some
issues with gang submit and applying certain HW workarounds which
require multiple VMIDs to work correctly.
So instead start to track all submissions to the relevant engines in a
per partition data structure and use the dma_fe
