Re: [PATCH] drivers: dma-coherent: use ioremap_wc() for DMA_MEMORY_MAP

[Brian Starkey <brian.starkey@arm.com>]

On Fri, Dec 04, 2015 at 05:15:54PM +0000, Catalin Marinas wrote:
>On Fri, Dec 04, 2015 at 08:59:10AM -0800, Dan Williams wrote:
>> On Fri, Dec 4, 2015 at 2:50 AM, Catalin Marinas <catalin.marinas@arm.com> wrote:
>> > On Fri, Nov 20, 2015 at 02:20:26PM +0000, Brian Starkey wrote:
>> >> When the DMA_MEMORY_MAP flag is used, memory which can be accessed
>> >> directly should be returned, so use ioremap_wc() instead of ioremap().
>> >> Also, ensure that the correct memset operation is used in
>> >> dma_alloc_from_coherent() with respect to the region's flags.
>> >>
>> >> This fixes the below alignment fault on arm64, caused by invalid use
>> >> of memset() on Device memory.
>> >
>> > This is indeed affecting both arm32 and arm64 systems.
>> >
>> >> diff --git a/drivers/base/dma-coherent.c b/drivers/base/dma-coherent.c
>> >> index 55b8398..45358d0 100644
>> >> --- a/drivers/base/dma-coherent.c
>> >> +++ b/drivers/base/dma-coherent.c
>> >> @@ -31,7 +31,10 @@ static int dma_init_coherent_memory(phys_addr_t phys_addr, dma_addr_t device_add
>> >>       if (!size)
>> >>               goto out;
>> >>
>> >> -     mem_base = ioremap(phys_addr, size);
>> >> +     if (flags & DMA_MEMORY_MAP)
>> >> +             mem_base = ioremap_wc(phys_addr, size);
>> >> +     else
>> >> +             mem_base = ioremap(phys_addr, size);
>> >
>> > I wonder whether a memremap() approach for the DMA_MEMORY_MAP case would
>> > be better. This API was added recently by commit 92281dee825f ("arch:
>> > introduce memremap()"). It only supports write-back and write-through
>> > but we could add a MEMREMAP_WC flag for this case.
>>
>> I originally included both MEMREMAP_WC and MEMREAMP_UC as potential
>> flags to this api, but ultimately decided against it.  The memremap()
>> api is meant for memory that is known to have no i/o side effects.  As
>> far as I can see WC and UC usages are a muddy mix of "sometimes
>> there's I/O side effects, but it depends by arch and driver".  In
>> other words we can't drop the "__iomem" annotation from WC and UC
>> mappings by default.

The DMA_MEMORY_MAP flag is pretty much a statement of "no side-
effects", so as Catalin says it would fit OK here. That said, if it's
not possible to deprecate ioremap_wc() in the same way as
ioremap_cache() then I wonder if there's even much benefit in adding
it to memremap().

>
>In this context, the dma_declare_coherent_memory(DMA_MEMORY_MAP)
>implementation is aimed at normal RAM with no side effects as later
>returned by dma_alloc_coherent(). To me it looks like memremap is better
>suited here for the DMA_MEMORY_MAP case. As per the
>Documentation/DMA-API.txt:
>
>  DMA_MEMORY_MAP - request that the memory returned from
>  dma_alloc_coherent() be directly writable.
>
>which means no __iomem. Of course, we still need ioremap for
>DMA_MEMORY_IO which is supposed to be written with memcpy_toio() etc.
>
>Which memory type should be used is left to the driver and it should
>pass the corresponding DMA_MEMORY_* flag.
>

This can still be achieved without adding _WC to memremap(). I can
look at adding _WC to memremap() if that's deemed the preferred
approach, but if that isn't clear-cut, then it would be nice to get
this bug fixed now and worry about adding it to memremap() later.

-Brian

>-- Catalin
>