Wednesday, May 10, 2017

Doing some further Spice testing.

Spice warnings on cast alignment

Rebased my cast-warnings branch once more. Did some additional cleanup.

Tested spicy with the rebased branch. Crashes with assert in spice-channel.c:1362. I had removed this assert on the osx branch, but that removal somehow got lost in the transfer to the cast-warnings branch.

Fixed various cases that were mistakenly classified as aligned.

Built x11spice

X11 spice was added on gitlab. Also made a copy gitlab.

Screen capture glitches on M2000

There are some interesting glitches on both VNC and spice remote displays when running them on a M2000 card:

I never saw any of these on my other PCs. I suspect the common choke point between these two is the capture API, since I see this both with:

• VNC with a vino-server on Turbo and a Apple Screen Sharing client
• Spice with a x11spice on Turbo and a spicy client running over X11

This is with driver version 375.39:

ddd@turbo[master] x11spice> cat /sys/bus/pci/drivers/nvidia/module/version 
375.39

Filed Red Hat Bugzilla 1449747.

Testing 3D on Turbo

Installed Negativo repositories at Frediano's suggestion:

dnf config-manager --add-repo=http://negativo17.org/repos/fedora-multimedia.repo

Then installed Blender with Cuda support. Curious to see how well this works:

dnf install blender-cuda

Virgl 3D guest does not boot with M2000

Confirmed the issue with 3D accelerated virgl guest, filed Red Hat Bugzilla 1449774.

The host dmesg shows little of interest:

[164685.196979] kvm_get_msr_common: 218 callbacks suppressed
[164685.196981] kvm [20500]: vcpu0, guest rIP: 0xffffffff83060b52 unhandled rdmsr: 0x34
[164910.402610] kvm [20500]: vcpu3, guest rIP: 0xffffffff81060b52 unhandled rdmsr: 0x34
[165258.185043] kvm: zapping shadow pages for mmio generation wraparound
[165258.190447] kvm: zapping shadow pages for mmio generation wraparound
[165264.536855] kvm [19688]: vcpu2, guest rIP: 0xffffffff8a060b52 unhandled rdmsr: 0x34
[165273.834050] kvm: zapping shadow pages for mmio generation wraparound
[165273.840259] kvm: zapping shadow pages for mmio generation wraparound
[165280.171865] kvm [19788]: vcpu0, guest rIP: 0xffffffff8f060b52 unhandled rdmsr: 0x34
[165290.428404] kvm: zapping shadow pages for mmio generation wraparound
[165290.434195] kvm: zapping shadow pages for mmio generation wraparound
[165337.405823] kvm [19892]: vcpu2, guest rIP: 0xffffffffa1060b52 unhandled rdmsr: 0x34

So nothing interesting.

The guest boot log is not very informative either, but one aspect that may require investigation:

[    2.730251] [drm] virtio vbuffers: 288 bufs, 192B each, 54kB total.
[    2.730889] [TTM] Zone  kernel: Available graphics memory: 1023428 kiB

Having 54K of buffers for a video driver, even a virtual one, seems quite low. Muse is not in a good enough shape to test with another NVIDIA card, but I need to check if that's normal.

Controlling core dumps

Following a comment by Christophe Fergeau, I looked into coredumpctl. It's installed by default in Fedora 25 and 26. However, abrt gets in the way of it working the way you want (for a developer). So disable that:

systemctl disable --now abrt-ccpp

Now I have neatly sorted core dumps for processes that crash. Cool.