CPU high io wait

[9000h]

as I cannot reopen the closed one I did create a new issue
there is no disk io btw.
the vaapidevice did create high CPU io wait (the old softhddevice did the same)

when I use xinelibout the CPU io wait is very low as expected
vdr -P"xineliboutput --local=sxfe --video=vaapi --audio=alsa --remote=none" -Psatip

[Space2Man]

Hi,

I checked with perf and saw the following:

    17.25%     0.05%  vdr              [kernel.vmlinux]             [k] entry_SYSCALL_64
            |          
             --17.20%--entry_SYSCALL_64
                       do_syscall_64
                       |          
                       |--13.63%--sys_ioctl
                       |          |          
                       |          |--11.94%--do_vfs_ioctl
                       |          |          |          
                       |          |           --11.53%--drm_ioctl
                       |          |                     |          
                       |          |                      --10.66%--drm_ioctl_kernel
                       |          |                                |          
                       |          |                                |--7.66%--i915_gem_execbuffer2
                       |          |                                |          |          
                       |          |                                |           --7.46%--i915_gem_do_execbuffer

Seems like it accesses the intel card via vfs --> therfore it's listed as I/O.

Can you do the same check with both vaapidevice and xineliboutput? I did the following to get the above graph:

# perf record -g -a sleep 10
[ perf record: Woken up 18 times to write data ]
[ perf record: Captured and wrote 4.592 MB perf.data (34714 samples) ]
# perf report

I am not sure if this is really an issue or if it is just shows that vaapidevice accesses the gfx card via vfs call.

[rofafor]

If you're comparing vaapidevice to another implementation like xineliboutput, please, make sure that both of them have enabled exactly the same VA-API filters. Otherwise, you'll end up comparing apples to oranges.

[9000h]

sure on vaapidevice the filters are off
also mpv did not show high cpu io wait with hwdec=vaapi vo=vaapi and playing back HD/UHD recordings

[jupe76]

Hi,
I've also been observing this weirdness for some time now.
Defintily the kernel is involved is this.
With softhddevice vaapi with the last lts kernel 4.9 I always had a very low load of 0.4 or so.
Only using a newer kernel from 4.10 or 4.11 on shifts the load to where it is now (2.x ...)

[9000h]

I can also confirm on older kernels 4.8 and also older ffmpeg,libav etc... the iowait is normal, but why is mpv and xineliboutput not also affected?

[Quantomax]

Kernel 4.14.11 and newer may have Spectre / Meltdown Fixes enabled - maybe this is the game changer?

[9000h]

I don't think so as mpv and xinelibout do not show a similar behavior

[9000h]

all the thread locking around the ffmpeg api calls are probably not needed anymore as the newer libs have now his own mechanism for locking