[PATCH 3/4] [ARM] mm: add memory type for inner-writeback

Gary King GKing at nvidia.com
Tue Aug 3 11:24:46 EDT 2010


Russell,

 > No.  This is not "free for use".  Mapping 5 is unused because it's not
 > architecturally defined - CPU implementations may not implement it.
 > This is what the ARM ARM says:

According to the ARMv7-A ARM, n = 6 is implementation defined; n = 5 is
available:

"The meaning of the field with n = 6 is IMPLEMENTATION DEFINED and might
   differ from the meaning given here. This is because the meaning of the
   attribute combination {TEX[0] = 1, C = 1, B = 0} is IMPLEMENTATION
   DEFINED."

- Gary


On 08/03/2010 12:41 AM, Russell King - ARM Linux wrote:
>
> On Mon, Aug 02, 2010 at 07:42:48PM -0700, Gary King wrote:
> > This change uses the currently-unused mapping 5 (TEX[0]=1, C=0, B=1)
> > in the tex remapping tables as an inner-writeback-write-allocate, outer
> > non-cacheable memory type, so that this mapping will be available to
> > clients which will benefit from the reduced L2 maintenance.
>
> No.  This is not "free for use".  Mapping 5 is unused because it's not
> architecturally defined - CPU implementations may not implement it.
> This is what the ARM ARM says:
>
>   For seven of the eight possible combinations of the TEX[0], C and B 
> bits,
>   a field in the PRRR defines the corresponding memory region as being
>   Normal, Device or Strongly-ordered memory a field in the NMRR defines
>   the Inner cache attributes that apply if the PRRR field identifies the
>   region as Normal memory a second field in the NMRR defines the Outer
>   cache attributes that apply if the PRRR field identifies the region as
>   Normal memory.
>
>   The meaning of the eighth combination for the TEX[0], C and B bits is
>   IMPLEMENTATION DEFINED
>
> So we can't be sure that the PRRR and NMRR bits which correspond with
> mapping 5 even exist.
>



More information about the linux-arm-kernel mailing list