On (Tue) Sep 29 2009 [22:41:45], Rusty Russell wrote:
> On Fri, 11 Sep 2009 11:43:06 pm Amit Shah wrote:
> > Expose multiple char devices ("ports") for simple communication
> > between the host userspace and guest.
>
> OK, I think other comments have died down, so it's time for a review. This
> was the latest patch I could find.
I just sent a new patch in this thread; can you please review that one?
> The obvious way to do this is to use multiple pairs of virtqueues. Is this
> silly for some reason? Yes, we're restricted to 32k ports.
There were a couple of things to note about that:
- Avi, in the past, mentioned separate ivq and ovqs for each port for
the kind of usage intended would be overkill.
- With the MSI rework, we can't init new i/ovqs for hotplugged ports as
we find them. The queues all have to be init'ed at probe() time. I had
some discussion with Michael some time back on virtualization@, but
thought the overhead of implementing that would be high.
> It means that if we see the MULTIQUEUE feature, we can look for the other
> queues. Our current MSI-X friendly API makes that a little painful, but
> if we resolve that our code should be sweetness, no?
If this is to solve the port numbering, etc., issues, I've thought of a
new way this can be done in the newest patch. What's more, Christian
tested it on s390 as well and it worked with a minor tweak. I tested the
new patch with combinations of qemu and kernel (old/new) and they worked
fine.
Thanks,
Amit
_______________________________________________
Virtualization mailing list
Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/virtualization
