On 5/29/26 13:24, Christian König wrote:
Hi everyone,
this patch set aims to fix a couple of problems we still have with the VM, page
fault and HMM/SVM handling.
First I rework a bit what we have on headers/defines for the suballocator,
especially that the right GFP flags are now used for GPU reset and graphics
page faults is important to avoid hangs.
Then I rework the VM handling, especially splitting up the
amdgpu_vm_update_range() into an amdgpu_vm_map_range() and unmap_range(). The
later is only meant to be used for MMU notifier/HMM/SVM use, but I couldn't
come up with a better name.
And finally it fixes the problem with KFD SVM that we can't properly allocate
page tables under the SVM lock. The missing piece here is to remove the SVM
lock and rely on the eviction lock of the VM for that.
Timur & Natalie I've put you on the patch set because I know you have page
fault patches in the pipeline, it would be nice if you could apply this set to
amd-staging-drm-next and rebase your stuff on top of it.
Thanks! I just finished drafting up a patchset for explicit sync of VM
unmaps, I'll also rebase it onto this patchset.
I also sent the current state of the patchset to the ML earlier
today[1]. That patchset includes a non-trivial rework to add a "context"
struct to VM maps, in order to track changed/freed VM mappings on a
per-operation basis. Happy for feedback on that approach - if there's
something to rework I can roll that into the rebase right away.
Thanks,
Natalie
[1]
https://lore.kernel.org/amd-gfx/[email protected]/
Honglei and Philip this is basically the pre-requisite work I have been talking
about. Please integrate that into the new SVM patches and re-test.
I've only did some very brief testing, so something more extensional testing is
clearly a good idea.
Thanks,
Christian.
