Re: [PATCH 2/2] arm64: mmu: use pagetable_alloc_nolock() while stop_machine()

[Yeoreum Yun <yeoreum.yun@arm.com>]

> On Thu 18-12-25 12:02:14, Ryan Roberts wrote:
> > On 18/12/2025 09:36, Yeoreum Yun wrote:
> > > Hi,
> > >> On Fri 12-12-25 16:18:32, Yeoreum Yun wrote:
> > >>> linear_map_split_to_ptes() and __kpti_install_ng_mappings()
> > >>> are called as callback of stop_machine().
> > >>> That means these functions context are preemption disabled.
> > >>>
> > >>> Unfortunately, under PREEMPT_RT, the pagetable_alloc() or
> > >>> __get_free_pages() couldn't be called in this context
> > >>> since spin lock that becomes sleepable on RT,
> > >>> potentially causing a sleep during page allocation.
> > >>>
> > >>> To address this, pagetable_alloc_nolock().
> > >>
> > >> As you cannot tolerate allocation failure and this is pretty much
> > >> permanent allocation (AFAIU) why don't you use a static allocation?
> > >
> > > Because of when bbl2_noabort is supported, that pages doesn't need to.
> > > If static alloc, that would be a waste in the system where bbl2_noabort
> > > is supported.
> > >
> > > When I tested, these extra pages are more than 40 in my FVP.
> > > So, it would be better dynamic allocation and I think since it's quite a
> > > early time, it's probably not failed that's why former code runs as it
> > > is.
> >
> > The required allocation size is also a function of the size of the installed RAM
> > so a static worst case allocation would consume all the RAM on small systems.
>
> Understood. But is it possible to pre-allocate early on so that the
> allocation itself doesn't have to happen from a constrained context.

That's the same suggestion from Ryan
(https://lore.kernel.org/all/100cc8da-b826-4fc2-a624-746bf6fb049d@arm.com/)

And here is the v2 which have accepted his suggestion :)
(But soon respin):
https://lore.kernel.org/all/20251217182007.2345700-1-yeoreum.yun@arm.com/

Thanks ;)

--
Sincerely,
Yeoreum Yun