Re: Xen regression, Was: [PATCH] x86/irq: Probe for PIC presence before allocating descs for legacy IRQs

[Boris Ostrovsky]

On 04/13/2016 06:38 PM, Stefano Stabellini wrote:
> On Wed, 13 Apr 2016, Boris Ostrovsky wrote:
>
> > On 04/13/2016 01:36 PM, Boris Ostrovsky wrote:
> > > On 04/12/2016 09:27 PM, Boris Ostrovsky wrote:
> > > > On 04/12/2016 07:15 PM, Stefano Stabellini wrote:
> > > > > On Tue, 12 Apr 2016, Boris Ostrovsky wrote:
> > > > > > On 04/12/2016 05:56 PM, Stefano Stabellini wrote:
> > > > > > >     I am not sure, maybe you didn't have CONFIG_SPARSE_IRQ?
> > > > > > > But I am certain that 4.6-rc2, with the attached config, fails as
> > > > > > > Dom0
> > > > > > > on QEMU with the following sequence of calls:
> > > > > > I did have CONFIG_SPARSE_IRQ and I just rebuilt 4.5.0 with your config
> > > > > > (4.6-rc3 doesn't build for me for some reason) and that booted dom0 as
> > > > > > well.
> > > > > >
> > > > > > BTW, what do you mean by "dom0 on QEMU"?
> > > > >    I am running Xen and Linux inside a QEMU x86_64 emulated machine
> > > > > (nested
> > > > > virt).
> > > > This I, of course, never tried.
> > > >
> > > > But given that things work in a single-level virt, doesn't this imply that
> > > > perhaps there is something in the emulation that's not quite right?
> > > OK, so this *is* broken on single level virt as well. It's just that we
> > > always end up using AHCI so lack of irq 14 (and 15) does not affect the
> > > system. And I guess in QEMU case it's IDE only, right?
> > >
> > > You patch does fix this but I wonder if we could change something in
> > > probe_8259A() so that we can continue using nr_legacy_irqs(). Using
> > > nr_legacy_irqs() and NR_IRQS_LEGACY at the same time is inconsistent and may
> > > cause us headaches in the future.
> > I think we could use paravirt_has() feature that was added for similar reason
> > when we had a problem with RTC (commit
> > d8c98a1d1488747625ad6044d423406e17e99b7a). So we add paravirt_has(PIC) which
> > will only be set by dom0 and then probe_8259A() will not set legacy_pic to
> > null_legacy_pic when this flag is set.
> Maybe we could introduce a legacy_pic_xen in arch/x86/kernel/i8259.c or
> arch/x86/xen/enlighten.c? We could set legacy_pic = legacy_pic_xen from
> start_kernel, so that we can skip probe_8259A completely.

Something like

legacy_pic_xen = null_legacy_pic;
legacy_pic_xen.nr_legacy_irqs = NR_IRQS_LEGACY;
legacy_pic = legacy_pic_xen;

?

>   
> > Note that paravirt_has() is being removed by
> > http://lists.xenproject.org/archives/html/xen-devel/2016-04/msg01415.html so
> > presumably we'd use new struct x86_legacy_features instead (copying Luis so
> > that if this is acceptable he could add it to his next spin).
> I would prefer to come up with a fix that is backportable to 4.3, 4.4
> and 4.5.

I think once the series above makes it to mainline x86_legacy_features 
would be the way to go --- that's what it is created for. Given that we 
are almost at rc4 I assume that would be 4.7.

Using paravirt_has() now will create merge headaches so it's probably 
not a good idea. We can either use your original patch or do what you 
suggested with legacy_pic_xen (both of which are backportable) and then 
switch to x86_legacy_features once it shows up in the tree.

-boris