Problem with atomic accesses in pstore on some ARM CPUs

[Mark Rutland]

On Tue, Aug 16, 2016 at 08:02:53AM -0700, Guenter Roeck wrote:
> On Tue, Aug 16, 2016 at 6:21 AM, Will Deacon <will.deacon at arm.com> wrote:
> > On Tue, Aug 16, 2016 at 06:14:53AM -0700, Guenter Roeck wrote:
> >> On Tue, Aug 16, 2016 at 3:32 AM, Robin Murphy <robin.murphy at arm.com> wrote:
> >> > On 16/08/16 00:19, Guenter Roeck wrote:
> >> >> we are having a problem with atomic accesses in pstore on some ARM
> >> >> CPUs (specifically rk3288 and rk3399). With those chips, atomic
> >> >> accesses fail with both pgprot_noncached and pgprot_writecombine
> >> >> memory. Atomic accesses do work when selecting PAGE_KERNEL protection.
> >> >
> >> > What's the pstore backed by? I'm guessing it's not normal DRAM.
> >> >
> >>
> >> it is normal DRAM.
> >
> > In which case, why does it need to be mapped with weird attributes?
> > Is there an alias in the linear map you can use?
> >
> 
> I don't really _want_ to do anything besides using pstore as-is, or,
> in other words, to have the upstream kernel work with the affected
> systems.
> 
> The current pstore code runs the following code for memory with
> pfn_valid() = true.
> 
>         if (memtype)
>                 prot = pgprot_noncached(PAGE_KERNEL);
>         else
>                 prot = pgprot_writecombine(PAGE_KERNEL);
>         ...
>         vaddr = vmap(pages, page_count, VM_MAP, prot);
> 
> It then uses the memory pointed to by vaddr for atomic operations.

This means that the generic ramoops / pstore code is making non-portable
assumptions about memory types.

So _something_ has to happen to that code.

> In my case, both protection options don't work. Everything works fine
> (or at least doesn't create an exception) if I use
>         vaddr = vmap(pages, page_count, VM_MAP, PAGE_KERNEL);
> instead.

Architecturally, that will give you a memory type to which we can safely use
atomics on.

It would be nice to know why the ramoops/pstore code must use atomics, and
exactly what it's trying to achieve (i.e. is this just for serialisation, or an
attempt to ensure persistence).

Depending on that, it may be possible to fix things more generically by using
memremap by default, for example, and only allowing uncached mappings on those
architectures which support them.

Thanks,
Mark.