Re: [xen-unstable-smoke test] 169781: regressions - FAIL

[Julien Grall <julien@xxxxxxx>]

On 29/04/2022 17:15, Stefano Stabellini wrote:
> > > Which seems to be the right sequence to me. There must be an early boot
> > > phase where interrupts are disabled on a CPU but memory allocations are
> > > possible. If this was x86 with the tlbflush limitation, I would suggest
> > > to have per-cpu memory mapping areas so that we don't have to do any
> > > global tlb flushes with interrupts disabled.
> > >
> > > On ARM, we don't have the tlbflush limitation so we could do that but we
> > > wouldn't have much to gain from it.
> > >
> > > Also, this seems to be a bit of a special case, because in general we
> > > can move drivers initializations later after local_irq_enable(). But
> > > this is the interrupt controller driver itself -- we cannot move it
> > > after local_irq_enable().
> > >
> > > So maybe an ad-hoc solution could be acceptable?
> > We don't need any ad-hoc solution here. We can register a CPU notifier that
> > will notify us when a CPU will be prepared. Something like below should work
> > (untested yet):
> The CPU notifier is a good idea. It looks like the patch below got
> corrupted somehow by the mailer. If you send it as a proper patch I am
> happy to have a look.
Doh. I will send a proper patch once I have done some testing.

Cheers,

--
Julien Grall