Re: [RFC PATCH 1/2] memremap: add arch specific hook for MEMREMAP_WB mappings

[Dan Williams]

On Tue, Feb 23, 2016 at 4:26 AM, Ard Biesheuvel
<ard.biesheuvel@xxxxxxxxxx> wrote:
> On 23 February 2016 at 13:03, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
>> On 23 February 2016 at 12:58, Russell King - ARM Linux
>> <linux@xxxxxxxxxxxxxxxx> wrote:
>>> On Mon, Feb 22, 2016 at 09:35:24PM +0100, Ard Biesheuvel wrote:
>>>> OK, thanks for the historical context.
>>>>
>>>> So what is your opinion on this series, i.e., to wire up memremap() to
>>>> remap arbitrary memory regions into the vmalloc area with MT_MEMORY_RW
>>>> attributes, and at the same time lift the restriction that the region
>>>> must be disjoint from memory covered by lowmem or kmap?
>>>
>>> The historical context is still present, because pxa2xx-flash has
>>> been converted to use memremap() from ioremap_cache() - possibly
>>> inappropriately.
>>>
>>> I've already described the semantics of ioremap_cache(), which are
>>> to always create a cacheable mapping irrespective of the system
>>> memory mapping type.  However, memremap() says that MEMREMAP_WB
>>> matches system RAM, which on ARM it doesn't right now.
>>>
>>
>> Indeed. Hence this series, to decouple memremap(MEMREMAP_WB) from
>> ioremap_cache() for ARM
>>
>>> Changing it to MT_MEMORY_RW would satisfy that comment against
>>> memremap(), but at the same time changes what happens with
>>> pxa2xx-flash - the memory region (which is not system RAM) then
>>> changes with the cache status of system RAM.
>>>
>>> So, I'm not that happy about the memremap() stuff right now, and
>>> I don't like the idea of making memremap() conform to its stated
>>> requirements without first preventing pxa2xx-flash being affected
>>> by such a change.
>>>
>>
>> Actually, my change fixes this issue, since it will cause memremap()
>> to always create MT_MEMORY_RW mappings, and not fallback to
>> ioremap_cache() for ranges that are not covered by lowmem.
>>
>>> Perhaps we need to reinstate the original ioremap_cached() API for
>>> pxa2xx-flash, and then switch memremap() to MT_MEMORY_RW - that
>>> would seem to result in the expected behaviour by all parties.
>>>
>>
>> I think we can simply revert the change to pxa2xx-flash if it is
>> deemed inappropriate.
>
> OK, I see what you mean. I find it unfortunate that ioremap_cache()
> instances are blindly being replaced with memremap(), and I wonder if
> this wasted test by and/or cc'ed to people who can actually test this
> driver. Dan?

I included that change in my original "convert ARM to memremap"
patchset [1].  I admit I didn't see the problem initially, but in
hindsight I should have told Brian to hold off until the whole
approach was sanity checked by ARM core maintainers.  Since then I've
been deferring the deprecation of ioremap_cache() until we could have
a conversation like this one.

> Anyway, I don't think it makes sense to stipulate at the generic level
> that ioremap_cache() and memremap(MEMREMAP_WB) shall be the same, and
> deprecating it is a bit premature since the cross-architecturally
> loosely defined semantics of ioremap_cache() can never be replaced 1:1
> with what memremap() promises.

Ok, my goal was to clean all the cases the were mishandling the
__iomem annotation where the *accesses* did not have I/O side effects.
What I overlooked was the difference between varying flavors of
writeback cacheable mappings.

> So what I suggest is that I revert the change to pxa2xx-flash as a new
> 1/3 in this series, and put these existing two on top to decouple
> memremap(MEMREMAP_WB) from ioremap_cache() entirely.

Should we formalize the pxa2xx-flash case with a new MEMREMAP_<type>?
Part of the original confusion is that we have ioremap_cache() with
varying semantics across architectures.

[1]: http://lists.infradead.org/pipermail/linux-arm-kernel/2015-July/360888.html