Re: About ioremap() and L2-Cache

Ingo Molnar (mingo@pc7537.hil.siemens.at)
Wed, 23 Jul 1997 12:57:58 +0200 (MET DST)


On Wed, 23 Jul 1997, Systemkennung Linux wrote:

> > about the PWT bit. if it's needed, __ioremap(flags) can be called with any
> > caching flag. (which will be written into the page table entry). Is there
> > any RL use for the PWT bit?
>
> It might be interesting for clearing pages etc. to avoid unecessary
> dirtying the caches where the probability of the data being used again
> is relativly small. A bit a different cache trick speeds up the
> performance of clear_page() on MIPS by a factor of two.

yep, although the pentium does not write-allocate cachelines so there
it's not a problem.

The PPro does, there it makes sense to set up a 'page clearing' window
somewhere in the virtual memory space, and map the to-be-cleared page to
that address. (this mapping will be fast with invlpg).

[unlinke the MIPS case here we have to be careful with 4MB pages stuff on
pentiums and higher, clear_page() usually points to a bigpage-mapped
memory address, which we cannot and should not remap, but rather we should
use a page clearing window]

If MTRR's are used, then no need to set up such a window, as MTRR's are
working on the physical memory level.

-- mingo