Re: [PATCH] do not p2m_invalidate_root when iommu_use_hap_pt

[Julien Grall <julien@xxxxxxx>]

Hi Stefano,

On 04/08/2021 18:41, Stefano Stabellini wrote:
> On Wed, 4 Aug 2021, Julien Grall wrote:
> > Hi Stefano,
> >
> > On 04/08/2021 01:08, Stefano Stabellini wrote:
> > > > > diff --git a/xen/arch/arm/arm64/vsysreg.c b/xen/arch/arm/arm64/vsysreg.c
> > > > > index caf17174b8..125a9281fc 100644
> > > > > --- a/xen/arch/arm/arm64/vsysreg.c
> > > > > +++ b/xen/arch/arm/arm64/vsysreg.c
> > > > > @@ -105,6 +105,13 @@ void do_sysreg(struct cpu_user_regs *regs,
> > > > >         case HSR_SYSREG_DCISW:
> > > > >         case HSR_SYSREG_DCCSW:
> > > > >         case HSR_SYSREG_DCCISW:
> > > > > +        if ( iommu_use_hap_pt(current->domain) )
> > > > > +        {
> > > > > +            gdprintk(XENLOG_ERR,
> > > > > +                     "d%u uses set/way cache flushes with the IOMMU on.
> > > > > It
> > > > > cannot work. Replace set/way instructions with dc [ci]vac and retry.
> > > > > Injecting exception into the guest now.\n",
> > > > This line would be far too long to print on the serial. I think you want
> > > > to
> > > > add a few newline here.
> > > Fair enough but I'll try to keep most info on the same line because
> > > otherwise with a dom0less boot it can get confusing. I suggest:
> > I am not quite too sure to understand why it would get confusing with
> > dom0less. Can you give an example?
> I was doing tests with the error messages before implementing the undef
> exception injection. This is the output of a regular domU (not
> dom0less): https://pastebin.com/Wytg660j
>
> The entire message in this test should be:
>
> (XEN) d1v0 uses set/way cache flushes with the IOMMU on. It cannot work.
> (XEN) Replace them with dc [ci]vac and retry.
>
> But actually the first line gets eaten, so we only see:
>
> (XEN) Replace them with dc [ci]vac and retry.
>
> several times at the bottom of the logs.
That's most likely because you are hitting the ratelimit. However, 
writing a 80+ line is almost the wrong way to go. This making a lot more 
difficult to serial the console output...
Looking at the output, I think you can shorten to:

"The cache should be flushed by VA rather than by set/way."

At least this would make clear that using them even without the IOMMU is 
a bad idea for performance reason.
Cheers,

--
Julien Grall