Am 13.12.23 um 16:39 schrieb Felix Kuehling:
On 2023-12-13 9:20, Christian König wrote:
Am 12.12.23 um 00:32 schrieb Felix Kuehling:
On 2023-12-11 04:50, Christian König wrote:
Am 08.12.23 um 20:53 schrieb Alex Deucher:
[SNIP]
You also need a functionality which resets all cleared blocks to
uncleared after suspend/resume.
No idea how to do this, maybe Alex knows of hand.
Since the buffers are cleared on creation, is there actually
anything to do?
Well exactly that's the problem, the buffers are no longer always
cleared on creation with this patch.
Instead we clear on free, track which areas are cleared and clear
only the ones which aren't cleared yet on creation.
The code I added for clearing-on-free a long time ago, does not
clear to 0, but to a non-0 poison value. That was meant to make it
easier to catch applications incorrectly relying on 0-initialized
memory. Is that being changed? I didn't see it in this patch series.
Yeah, Arun stumbled over that as well. Any objections that we fill
with zeros instead or is that poison value something necessary for
debugging?
I don't think it's strictly necessary. But it may encourage sloppy
user mode programming to rely on 0-initialized memory that ends up
breaking in corner cases or on older kernels.
That said, I see that this patch series adds clearing of memory in the
VRAM manager, but it doesn't remove the clearing for
AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE in amdgpu_bo_release_notify and
amdgpu_move_blit. This will lead to duplicate work.
I'm also not sure how the clearing added in this patch series will
affect free-latency observed in user mode. Will this be synchronous
and cause the user mode thread to stall while the memory is being
cleared?
Yeah, that's not fully working at the moment. I already pointed out as
well that Arun should remove the clearing in the VRAM manager.
Regards,
Christian.
Regards,
Felix
Regards,
Christian.
Regards,
Felix
So some cases need special handling. E.g. when the engine is not
initialized yet or suspend/resume.
In theory after a suspend/resume cycle the VRAM is cleared to
zeros, but in practice that's not always true.
Christian.
Alex
