Re: [PATCH v2 6/6] x86, pat: Update documentation for WT changes

[Andy Lutomirski <luto@amacapital.net>]

On Mon, Sep 15, 2014 at 2:19 PM, Toshi Kani <toshi.kani@hp.com> wrote:
> On Wed, 2014-09-10 at 15:34 -0600, Toshi Kani wrote:
>> On Wed, 2014-09-10 at 13:29 -0700, Andy Lutomirski wrote:
>> > On Wed, Sep 10, 2014 at 1:12 PM, Toshi Kani <toshi.kani@hp.com> wrote:
>> > > On Wed, 2014-09-10 at 11:30 -0700, Andy Lutomirski wrote:
>> > >> On Wed, Sep 10, 2014 at 9:51 AM, Toshi Kani <toshi.kani@hp.com> wrote:
>> > >> > +Drivers may map the entire NV-DIMM range with ioremap_cache and then change
>> > >> > +a specific range to wt with set_memory_wt.
>> > >>
>> > >> That's mighty specific :)
>> > >
>> > > How about below?
>> > >
>> > > Drivers may use set_memory_wt to set WT type for cached reserve ranges.
>> >
>> > Do they have to be cached?
>>
>> Yes, set_memory_xyz only supports WB->type->WB transition.
>>
>> > How about:
>> >
>> > Drivers may call set_memory_wt on ioremapped ranges.  In this case,
>> > there is no need to change the memory type back before calling
>> > iounmap.
>> >
>> > (Or only on cached ioremapped ranges if that is, in fact, the case.)
>>
>> Sounds good.  Yes, I will use cashed ioremapped ranges.
>
> Well, testing "no need to change the memory type back before calling
> iounmap" turns out to be a good test case.  I realized that
> set_memory_xyz only works properly for RAM.  There are two problems for
> using this interface for ioremapped ranges.
>
> 1) set_memory_xyz calls reserve_memtype() with __pa(addr).  However,
> __pa() translates the addr into a fake physical address when it is an
> ioremapped address.
>
> 2) reserve_memtype() does not work for set_memory_xyz.  For RAM, the WB
> state is managed untracked.  Hence, WB->new->WB is not considered as a
> conflict.  For ioremapped ranges, WB is tracked in the same way as other
> cache types.  Hence, WB->new is considered as a conflict.
>
> In my previous testing, 2) was undetected since 1) led using a fake
> physical address which was not tracked for WB.  This made ioremapped
> ranges worked just like RAM. :-(
>
> Anyway, 1) can be fixed by using slow_virt_to_phys() instead of __pa().
> set_memory_xyz is already slow, but this makes it even slower, though.
>
> For 2), WB has to be continuously tracked in order to detect aliasing,
> ex. ioremap_cache and ioremap to a same address.  So, I think
> reserve_memtype() needs the following changes:
>  - Add a new arg to see if an operation is to create a new mapping or to
> change cache attribute.
>  - Track overlapping maps so that cache type change to an overlapping
> range can be detected and failed.
>
> This level of changes requires a separate set of patches if we pursue to
> support ioremapped ranges.  So, I am considering to take one of the two
> options below.
>
> A) Drop the patch for set_memory_wt.
>
> B) Keep the patch for set_memory_wt, but document that it fails with
> -EINVAL and its use is for RAM only.
>

I vote A.  I see no great reason to add code that can't be used.  Once
someone needs this ability, they can add it :)

It's too bad that ioremap is called ioremap and not iomap.  Otherwise
the natural solution would be to add a different function call
ioremap_wt that's like set_memory_wt but for ioremap ranges.  Calling
it ioreremap_wt sounds kind of disgusting :)

--Andy

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>