Closed nyanmisaka closed 1 year ago
@nyanmisaka since this i915 issue, can you help open bug in i915 project? thanks
@XinfengZhang , @Sherry-Lin
@haichund1 @FurongZhang Fixed by upstream i915 https://github.com/torvalds/linux/commit/3f882f2d4f689627c1566c2c92087bc3ff734953
Which component impacted?
Video Processing
Is it regression? Good in old configuration?
Yes, it's good in old version
What happened?
(vaapi dec + opencl filter + qsv enc):
(qsv dec + opencl filter + qsv enc):
Note that this issue can be more obvious if you create the opencl context with
CL_CONTEXT_INTEROP_USER_SYNC
set toCL_TRUE
.Also in the first vaapi-qsv hybrid ffmpeg cli this issue can be avoided by cap the input threads to 1, but it's not suitable for the second pure qsv ffmpeg cli since qsv requires 2+ threads.
What's the usage scenario when you are seeing the problem?
Transcode for media delivery
What impacted?
https://github.com/jellyfin/jellyfin QSV/VAAPI transcoding with OpenCL filtering on Intel Gen9+ iGPUs. older HD/UHD630 and new Xe graphics are both affected.
Debug Information
Do you want to contribute a patch to fix the issue?
I have bisected the issue to this kernel commit introduced in linux 5.18: https://github.com/torvalds/linux/commit/7e00897be8bf13ef9c68c95a8e386b714c29ad95 Removing
grab_vma()
andungrab_vma()
calls works fine again in this case.It seems to be a i915 kernel issue and it indeed affects the userspace programs such as iHD+NEO+ffmpeg use case but I don't really understand it's mechanism and how to fix it in i915, iHD or NEO, so I put it here to let you know.