Hi
Am 19.10.25 um 19:10 schrieb Dmitry Osipenko:
On 10/17/25 16:58, Dmitry Osipenko wrote:
Hi,
On 10/17/25 10:38, Thomas Zimmermann wrote:
...
There's little difference between the current event handling and the one
where no vblanks have been set up. I suspect that the vblank timer
callback interferes with the locking in atomic_flush. That would also
explain why the fps drop at no clear pattern.
Could you please test the attached patch? It enables/disables the vblank
timer depending on the buffer resources; as you suggested before. Does
this make a difference?
The attached patch doesn't work, please see the trace below.
@Vivek Please clarify whether you only see frames drop with your
multi-gpu guest-blob setup or with a usual virgl too. I haven't noticed
problem with frames pacing for virgl and nctx modes yesterday, will
check again.
On a second look, I now see that this RFC (not the attached) patch
doesn't work properly with host blobs.
I'm getting 100-150fps with this patch applied instead of expected
60fps. Without this RFC patch I'm getting constant 60fps with native
context displaying host blobs.
Not sure why guest blob would behave differently from the host blob.
Suspect something if off with the prime sharing that Vivek uses in the
vfio testing setup. I'd suggest to disable vblank timer only for guest
blobs if no quick solution will be found.
After reading your reply and Vivek's new results, I'm confused now. Does
it work or is there another patch needed?
Best regards
Thomas
--
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)
