<qyliss>
don't remember if I said already, but the thing I'm working on at the moment is packaging up all the virtio-vhost-user research I did last year into a runnable demo I can post (and get paid for)
<cole-h>
!
<cole-h>
look forward to it
TheJollyRoger has quit [Remote host closed the connection]
TheJollyRoger has joined #spectrum
<MichaelRaskin>
Oh great
<qyliss>
I've been thinking a lot recently about how, even if I don't manage to do everything I want to with Spectrum, I'll have improved a lot of other software along the way
<qyliss>
and that's very reassuring
cole-h has quit [Quit: Goodbye]
cole-h has joined #spectrum
<MichaelRaskin>
Can virtio-vhost-user run the serving side inside another VM (and can that VM dynamically extend the set of client VMs served)?
<qyliss>
1. yes 2. yes, with VMM support
<qyliss>
but it is _very_ early days -- it can currently only do network traffic in one direction!
<qyliss>
so what's there now is really a proof of concept more than anything else
<qyliss>
they're working on getting it all into the appropriate specifications (virtio, etc.) before implementing it further
<MichaelRaskin>
Hmmm, but if I can dynamically add that thing in both direction, I get the desired network traffic in both directions, right?
<MichaelRaskin>
(I know it's not how it _should_ be used)
<qyliss>
uhh, pass :P
<MichaelRaskin>
Assymetric routing rules are not even _that_ weird a thing…
<MichaelRaskin>
Does VMM need to do anything after initial establishment of the connections?
<qyliss>
yes, it's a virtio device, so it needs to translate between virtio and vhost-user
<qyliss>
although now that I'm thinking about it, I think actually you only ever have one v-v-h device, and all v-v-h stuff goes through that one device?
<qyliss>
it which case you wouldn't need any VMM support at all to add connections at runtime
<MichaelRaskin>
Translation as in relaying all the traffic?
<qyliss>
yeah
<qyliss>
(think so)
<qyliss>
it's been a little while
<MichaelRaskin>
Suboptimal…
<qyliss>
pretty sure there was a TWiS explaining this
<MichaelRaskin>
Hm, I am not so sure…
<qyliss>
the conclusion I have been increasingly coming to is that a host memcpy is not going to be your bottleneck for anything
<MichaelRaskin>
If host memcpy is not a bottleneck I can as well run uncompressed VNC!
<qyliss>
^ that's what a VMM does on rx, but I'm having trouble remembering what exactly it's receiving in this context
<qyliss>
might just be control messages
<MichaelRaskin>
I also have some nagging doubts about hot-cache FS use
<qyliss>
FS shouldn't require a memcpy with DAX
<MichaelRaskin>
Even if FS driver is supposed to run un a VM?
<JJJollyjim>
what a coincidence, I too am a Trans virtio vhost user
<JJJollyjim>
:P
<qyliss>
:D
<qyliss>
MichaelRaskin: hmm no, in that case things are trickier
<qyliss>
oh also the link I sent is what the software in the VM does, not what the VMM does, oops
<qyliss>
fwiw, for networking v-v-h probably isn't what we want after all, because the thing it really helps with is talking to a userspace networking implementation running inside a vm
<MichaelRaskin>
If inside a dedicated VM, not sure it matters as much if user-space or kernel-space
<MichaelRaskin>
Yeah, a bit harder to stage attacks on VMM, I guess
<qyliss>
right, but if you just want to talk to a guest kernel, you don't need VVH
<qyliss>
you can just speak normal virtio to the guest kernel
<MichaelRaskin>
Right, I hoped v-v-h would allow just establishing a buffer once and free hypervisor from relaying
<qyliss>
it... might
<qyliss>
but if it does, a non-vvh vhost-user connector program could also do that
<MichaelRaskin>
Eventually… when spec-ced… ?
<qyliss>
no, it could do that today if somebody wrote it
<qyliss>
assuming that no copies are involved with vhost-user itself
<qyliss>
(if they are, then they'd also apply to virtio, because virtio-vhost-user is vhost-user over virtio)
<qyliss>
I'll probably be able to have this conversation better in a week or so when my head is less full of strace and more full of VMM things
<qyliss>
because I'm definitely getting things wrong right now
<MichaelRaskin>
heh
<qyliss>
currently I'm relying on things I still have in my head six months after I last looked into it, so it's all rather fuzzy