On 2023-04-25 15:59, Shashank Sharma wrote:
On 24/04/2023 21:56, Felix Kuehling wrote:
On 2023-04-22 2:39, Shashank Sharma wrote:
- KFD process level doorbells: doorbell pages which are allocated by
kernel but mapped and written by userspace processes, saved in
struct pdd->qpd->doorbells
size = kfd_doorbell_process_slice.
We realized that we only need 1-2 doorbells for KFD kernel level
stuff (so kept it one page), but need 2-page of doorbells for KFD
process, so they are sized accordingly.
We have also run kfd_test_suit and verified the changes for any
regression. Hope this helps in explaining the design.
Right, I missed that this was only for kernel doorbells. I wonder
whether KFD really needs its own page here. I think we only need a
doorbell for HWS. And when we use MES, I think even that isn't needed
because MES packet submissions go through amdgpu. So maybe KFD
doesn't need its own kernel-mode doorbell page any more on systems
with user graphics mode queues.
Yeah, for any IP with MES enabled, KFD doesn't need kernel level
doorbells. But I still allocated a page just to make sure we do not
break any non-MES platforms or use cases where MES is deliberately
disabled from kernel command line. Hope that works for you.
Even without MES, we still only need one doorbell for HWS. Allocating a
whole page for that is wasteful. Anyway, I'm OK with cleaning that up later.
Regards,
Felix
- Shashank
Regards,
Felix
- Shashank
