Re: [Qemu-devel] [PATCH v2 1/3] q35: set split kernel irqchip as default

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 26 Apr 2019 13:27:44 -0600
Alex Williamson <alex.williamson@redhat.com> wrote:

> On Thu, 20 Dec 2018 13:40:35 +0800
> Peter Xu <peterx@redhat.com> wrote:
> 
> > Starting from QEMU 4.0, let's specify "split" as the default value for
> > kernel-irqchip.
> > 
> > So for QEMU>=4.0 we'll have: allowed=Y,required=N,split=Y
> >    for QEMU<=3.1 we'll have: allowed=Y,required=N,split=N
> >    (omitting all the "kernel_irqchip_" prefix)
> > 
> > Note that this will let the default q35 machine type to depend on
> > Linux version 4.4 or newer because that's where split irqchip is
> > introduced in kernel.  But it's fine since we're boosting supported
> > Linux version for QEMU 4.0 to around Linux 4.5.  For more information
> > please refer to the discussion on AMD's RDTSCP:
> > 
> >   https://lore.kernel.org/lkml/20181210181328.GA762@zn.tnic/  
> 
> It looks like this broke INTx for vfio-pci, see:
> 
> https://bugs.launchpad.net/qemu/+bug/1826422
> 
> In my testing it looks like KVM advertises supporting the KVM_IRQFD
> resample feature, but vfio never gets the unmask notification, so the
> device remains with DisINTx set and no further interrupts are
> generated.  Do we expect KVM's IRQFD with resampler to work in the
> split IRQ mode?  We can certainly hope that "high performance" devices
> use MSI or MSI/X, but this would be quite a performance regression with
> split mode if our userspace bypass for INTx goes away.  Thanks,

arch/x86/kvm/lapic.c:kvm_ioapic_send_eoi() dumps to userspace before
kvm_ioapic_update_eoi() can handle the irq_ack_notifier_list via
kvm_notify_acked_gsi(), so it looks like KVM really ought to return an
error when trying to register a resample IRQFD when irqchip_split(),
but do we have better options?  EOI handling in QEMU is pretty much
non-existent and even if it was in place, it's a big performance
regression for vfio INTx handling.  Is a split irqchip that retains
performant resampling (EOI) support possible?  Thanks,

Alex