Re: [Lse-tech] Re: [Discontig-devel] Re: [PATCH] Nonlinear kernel virtual to physical mapping

[Matthew W. <wi...@fc...>]

On Tue, Apr 16, 2002 at 05:58:55PM +0200, Daniel Phillips wrote:
> On Friday 12 April 2002 14:40, Matthew Wilcox wrote:
> > FYI, parisc64 is likely to retain the ->virtual field in order to turn
> > flush_dcache_page into a no-op.  But nobody's written that code yet :-)
> 
> Hi Matthew,
> 
> The flush_dcache_page part is a little cryptic, could you please elaborate?

Certainly.  PA-RISC 2.0 has a 96-bit virtual address space.  The top
bits are determined by space registers, and the bottom bits by the
general register you're using.  While this is a virtually indexed cache,
it is possible to get read-write coherency by using an /equivalent alias/.
Aliases which are separated by a multiple of 4MB in this 96-bit virtual
address space are guaranteed to be coherent.  We already force
user-space shared mmaps to be equivalent aliases, but rely on the kernel
calling flush_dcache_page after it has written to a non-equivalent
alias.

Now, if the kernel only writes to an equivalent alias, the processor
guarantees coherency and flush_dcache_page can be a no-op.  So my
proposed scheme is to use kmap/kunmap to return an address for a page
which will be equivalent.  The plan is to take up a 256TB chunk of the
kernel's virtual address space to allow 256GB of memory to be mapped.
We'll use the ->virtual member to indicate what address to return.

-- 
It's always legal to use Linux (TM) systems
http://www.gnu.org/philosophy/why-free.html