Re: xennet performance collapses with multiple vCPU

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Mon, Apr 13, 2020 at 07:28:32PM +0000, Andrew Doran wrote:
> Ok that's interesting.  I've been all over the IPI code a few times now, and
> tried different things to exclude problems with it, and have found no issue
> there.

I found the problem and commited a fix.

Basically, the scheduler now choose to runs the xennet attach handler
on a different CPU than the boot CPU. The kernel expects the interrupt handler
to be called from this CPU, but Xen still sends the event to CPU 0
(which is the default). The event on CPU 0 is then ignored (the kernel
doens't expect it), and is later processed on the right CPU when a clock
tick causes events processing here.

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--