Re: [PULL 12/17] virtio-gpu: support context init multiple timeline

["Alex Bennée" <alex.bennee@linaro.org>]

Yiwei Zhang <zzyiwei@gmail.com> writes:

> On Sun, Jun 8, 2025 at 1:24 AM Akihiko Odaki
> <odaki@rsg.ci.i.u-tokyo.ac.jp> wrote:
>>
>> On 2025/06/06 1:26, Alex Bennée wrote:
>> > From: Yiwei Zhang <zzyiwei@gmail.com>
>> >
>> > Venus and later native contexts have their own fence context along with
>> > multiple timelines within. Fences wtih VIRTIO_GPU_FLAG_INFO_RING_IDX in
>> > the flags must be dispatched to be created on the target context. Fence
>> > signaling also has to be handled on the specific timeline within that
>> > target context.
>> >
>> > Before this change, venus fencing is completely broken if the host
>> > driver doesn't support implicit fencing with external memory objects.
>> > Frames can go backwards along with random artifacts on screen if the
>> > host driver doesn't attach an implicit fence to the render target. The
>> > symptom could be hidden by certain guest wsi backend that waits on a
>> > venus native VkFence object for the actual payload with limited present
>> > modes or under special configs. e.g. x11 mailbox or xwayland.
>> >
>> > After this change, everything related to venus fencing starts making
>> > sense. Confirmed this via guest and host side perfetto tracing.
>> >
>> > Cc: qemu-stable@nongnu.org
>> > Fixes: 94d0ea1c1928 ("virtio-gpu: Support Venus context")
>>
>> I suggest moving this in the front of the patch series to ease backporting.
>>
>> I also wonder if "[PULL 11/17] ui/gtk-gl-area: Remove extra draw call in
>> refresh" requires Cc: qemu-stable@nongnu.org. Fixing -display gtk,gl=on
>> for Wayland sounds as important as this patch.
>>
>> Regards,
>> Akihiko Odaki
>
> Hi Alex,
>
> +1 for Akihiko's point. I'm also curious when will the venus fix land
> in-tree?

We have a problem that there are two contradictory bugs - one that shows
up in the x86/kvm case and one in the aarch64/tcg case. Both are caused
by the weird lifetime semantics of the virgl resource vs QEMU memory
region and we haven't found a solution that solves both yet.

I'm currently busy with other stuff and need to do a sweep of my other
maintainer trees for 10.1. Once I've done that I'll have another go at
coming up with a solution unless someone else beats me to it.

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro