Re: [PATCH] x86: q35: require split irqchip for large CPU count

[David Woodhouse]

On Mon, 2022-03-14 at 13:21 +0000, Daniel P. Berrangé wrote:
> On Mon, Mar 14, 2022 at 12:59:38PM +0000, David Woodhouse wrote:
> > On Mon, 2022-03-14 at 11:35 +0100, Igor Mammedov wrote:
> > > On Fri, 11 Mar 2022 14:58:41 +0000
> > > David Woodhouse <
> > > dwmw2@infradead.org
> > > 
> > > > wrote:
> > > > On Fri, 2022-03-11 at 09:39 -0500, Igor Mammedov wrote:
> > > > > if VM is started with:
> > > > > 
> > > > >    -enable-kvm -smp 256
> > > > > 
> > > > > without specifying 'split' irqchip, VM might eventually boot
> > > > > but no more than 255 CPUs will be operational and following
> > > > > error messages in guest could be observed:
> > > > >    ...
> > > > >    smpboot: native_cpu_up: bad cpu 256
> > > > >    ...
> > > > > It's a regression introduced by [1], which removed dependency
> > > > > on intremap=on that were implicitly requiring 'split' irqchip
> > > > > and forgot to check for 'split' irqchip.
> > > > > Instead of letting VM boot a broken VM, error out and tell
> > > > > user how to fix CLI.  
> > > > 
> > > > Hm, wasn't that already fixed in the patches I posted in December?
> > > 
> > > It might be, could you point to the commit/series that fixed it.
> > 
> > https://lore.kernel.org/all/20211209220840.14889-1-dwmw2@infradead.org/
> > 
> > is the patch I was thinking of, but although that moves the check to a
> > more useful place and fixes the X2APIC check, it *doesn't* include the
> > fix you're making; it's still using kvm_irqchip_in_kernel().
> > 
> > I can change that and repost the series, which is still sitting (with
> > fixed Reviewed-By/Acked-By attributions that I screwed up last time) in
> > https://git.infradead.org/users/dwmw2/qemu.git
> > 
> > 
> > > Regardless of that, fixing it in recent kernels doesn't help
> > > as still supported kernels are still affected by it.
> > > 
> > > If there is a way to detect that fix, I can add to q35 a compat
> > > property and an extra logic to enable kernel-irqchip if fix is present.
> > > Otherwise the fix does not exist until minimum supported kernel
> > > version reaches version where it was fixed.
> > 
> > Hm, I'm not sure I follow here. Do you mean recent versions of *qemu*
> > when you say 'kernels'? 
> > 
> > I'm not even sure I agree with the observation that qemu should error
> > out here. The guest boots fine and the guest can even *use* all the
> > CPUs. IPIs etc. will all work fine. The only thing that doesn't work is
> > delivering *external* interrupts to CPUs above 254.
> > 
> > Ultimately, this is the *guest's* problem. Some operating systems can
> > cope; some can't.
> > 
> > The fact that *Linux* has a fundamental assumption that *all* CPUs can
> > receive all interrupts and that affinity can't be limited in hardware,
> > is a Linux problem. I tried to fix it once but it was distinctly non-
> > trivial and eventually I gave up and took a different approach.
> > https://lore.kernel.org/linux-iommu/87lfgj59mp.fsf@nanos.tec.linutronix.de/T/
> > 
> > 
> > But even if we 'fix' the check as you suggest to bail out and refuse to
> > boot a certain configuration because Linux guest wouldn't be able to
> > fully utilize it... Even if we boot with the split IRQ chip and the 15-
> > bit MSI enlightenment, we're still in the same position. Some guests
> > will be able to use it; some won't.
> > 
> > In fact, there are operating systems that don't even know about X2APIC.
> > 
> > Why should qemu refuse to even start up?
> 
> We've generally said QEMU should not reject / block startup of valid
> hardware configurations, based on existance of bugs in certain guest
> OS, if the config would be valid for other guest.

Right, so I think the patches I sent in December are sufficient to fix
the existing bugs there. I'll repost them shortly.

In particular, I just rechecked that we aren't advertising the
KVM_FEATURE_MSI_EXT_DEST_ID feature unless we have the *split* irq
chip, which is correct. So we'll expose vCPUs with higher APIC IDs to
guests without any way to target *external* interrupts at them, and
that *might* mean that some guests refuse to use those vCPUs at all,
but that's just fine, and we don't advertise anything that isn't
working correctly.

smime.p7s
Description: S/MIME cryptographic signature