On 9/23/19 6:44 AM, marcandre.lur...@redhat.com wrote:
From: Marc-André Lureau <marcandre.lur...@redhat.com>
This series adds support for running virtio GPUs in seperate
processes, thanks to vhost-user backend.
The QEMU support landed for 4.1. There are several benefits of running
the GPU/virgl in an external process, since Mesa is rather heavy on
the qemu main loop, and may block for a while, or crash.
The external GPU process is started with one end of a socket pair, the
other end is given to a QEMU chardev attached to a device. The
external process is also added to the cgroup to limit resources usage.
vhost-user is a generic mechanism that allows implementing virtio device
dataplane handling in a separate userspace process. vhost-user-gpu here
notably moves the virgl 3d handling out of the main qemu process. The
external process will be /usr/libexec/vhost-user-gpu, which comes from
qemu.git contrib/vhost-user-gpu code, first released in qemu-4.1.
Part of this series deals with discovering the location on disk of the
vhost-user-gpu binary, and what capabilities it provides. This uses a
similar mechanism to firmware.json, described in qemu
docs/interop/vhost-user.json
https://github.com/qemu/qemu/blob/master/docs/interop/vhost-user.json
qemu 4.1 ships a 50-qemu-gpu.json to match. I believe virtio-fs
will use a similar mechanism when it lands in upstream qemu, as
virtiofsd is a separate process that communicates with qemu over
vhost-user.
For a bit more background on vhost-user-gpu process handling and
the json interop motivation, here's some of the qemu discussion:
https://lists.nongnu.org/archive/html/qemu-devel/2018-08/msg02610.html
https://lists.nongnu.org/archive/html/qemu-devel/2018-09/msg00807.html
For this series, the XML to enable this is:
<video model='virtio'>
<driver name='vhostuser'/>
<acceleration accel3d='yes' rendernode='/path/to/rendernode'/>
</video>
rendernode is optional
qemu_vhost_user.c handles vhost-user.json
qemu_vhost_user_gpu.c handles the process management for
vhost-user-gpu
I dropped patch #2 since there's discussion on it. Patch 10 needed a
tests/Makefile.am adjustment to fix RPM build. Pushed now with those changes
Thanks,
Cole
--
libvir-list mailing list
libvir-list@redhat.com
https://www.redhat.com/mailman/listinfo/libvir-list
