On Wed, 17 Nov 2021 at 10:07, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> On Wed, Nov 17, 2021 at 10:03 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> >
> > On Wed, Nov 17, 2021 at 9:36 AM Tony Lindgren <tony@xxxxxxxxxxx> wrote:
> > > * Ard Biesheuvel <ardb@xxxxxxxxxx> [211117 08:29]:
> > > >
> > > > Thanks Tony, that is very helpful. I have a Beaglebone white somewhere
> > > > so I'll try and reproduce it locally as well.
> > >
> > > I think with Beaglebone you may hit this only with suspend/resume if at
> > > all. On am335x cpuidle is not shutting down the CPU. And only some models
> > > will suspend to deeper idle states as it depends on the PMIC.
> > >
> > > If you have some test patch to try, just let me know.
> >
> > I looked at how the sleep code is called and found that cpu_suspend()/
> > __cpu_suspend() has interesting manipulation of the stack pointer to
> > call the platform specific function with a simple 1:1 page table,
> > I would expect the problem somewhere in there, haven't pinpointed
> > the exact line yet, but if any of that code tries to local the physical
> > address of the stack using virt_to_phys or its asm equivalent, this
> > fails for a vmap stack.
>
> and just after sending this I see
>
> void __cpu_suspend_save(u32 *ptr, u32 ptrsz, u32 sp, u32 *save_ptr)
> {
> *save_ptr = virt_to_phys(ptr);
>
> 'ptr' is a pointer to the stack here. It might not be the only place that
> needs fixing, but this clearly has to do a page table walk like
> vmalloc_to_page() does to get to the correct physical address.
>
I had just arrived at the same conclusion. I'll fix this up and drop
it in kernelci.
