Am 24.09.21 um 07:50 schrieb Yu, Lang:
[AMD Official Use Only]
[SNIP]
Hi Christian,
Thanks for your explanation and advice. I got your point.
Actually, these BOs are allocated and pinned during a kfd process lifecycle.
I will try to add a flag into struct kgd_mem to indicate which BO is
pined and should be unpinned. Which will make
amdgpu_bo_pin/amdgpu_bo_unpin calls balanced. Thanks!
That isn't to much better. The real solution would be to unpin them
when the kfd process is destroyed.
Yes, will unpin them when the kfd process is destroyed.
But we should indicate which BO is pinned and should be unpinned. Right?
Well not with a flag or something like that.
The knowledge which BO is pinned and needs to be unpinned should come from
the control logic and not be papered over by some general handling.
That's the background why we have removed the general handling for this from
TTM in the first place.
In other words when you need to pin a BO because it is kmapped it should be
unpinned when it is kunmapped and if you don't kunmap at all then there is
something wrong with the code structure from a higher level point of view.
Yes, this function "amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel" did a kmap,
but without a kunmap when the kfd process is destroyed. The flag actually indicates kmap/kunmap.
Well that's the wrong approach then. I mean you need to have the BO
reference and the mapped pointer somewhere, don't you?
How do you clean those up?
Regards,
Christian.
