Quoting Prabhakaran, Krishna (2026-11-12 07:29:14)
> > No bug opened for issue, no exact hardware config described, no dmesg
> > logs.
> >
> > Also, if you import a new dma-buf too frequently, probably shouldn't
> > do that.
> 
> This feedback was on v2, but it still applies to v4. The setup is an ARL-S 
> host running multiple VMs, with all VM's playing video. A persistent audio 
> glitch was heard during the playback. Here's a perf call tree captured from 
> the host QEMU process during video playback indicating that this was starving 
> the QEMU processing thread.
> 
> virtio_gpu_resource_flush                     45.45%
>   i915_gem_execbuffer2_ioctl
>     i915_gem_do_execbuffer
>       eb_validate_vmas
>         i915_vma_pin_ww
>           i915_vma_get_pages

This is already a point where you have fallen off the hot path and you
should really look into your application code or virtio gpu code to
optimize it to stay on the hot path (try_qad_pin succeeds).

You should not be importing a new DMA-BUF per frame, but should instead
keep recycling existing ones. If you keep using new buffers, you will be
finding more and more hot spots in the code as a fresh BO (DMA-BUF or
not) each frame is going to take a hit on the performance no matter
what. It's not how you're supposed to write applications.

>             i915_gem_object_get_pages_dmabuf   36.65%
>               wbinvd_on_all_cpus               35.95%
>                 on_each_cpu_cond_mask
>                   smp_call_function_many_cond  35.92%  (29.74% self)
>                     __wbinvd                    7.38%  (self)
> 
> This profiling data is what motivated the patch.

This profiling data is not that interesting in itself because it's
for a very suboptimal usage scenario of the uAPI.

How do the numbers look like after the patch applied? It's the
relative improvement of i915_gem_object_get_pages_dmabuf()
that is of interest between baseline and this patch applied.

Regards, Joonas

Reply via email to