[linuxsh-dev] PCMCIA patch & question

[Greg B. <gb...@po...>]

G'day

  I've uploaded a new version of my HD64465 PCMCIA patch to sourceforge
http://sourceforge.net/patch/download.php?id=101674

  Niibe-san, I want to check if you're happy about an
architectural aspect of my patch, which Peter Hung and
I have been discussing.

  I've used the MMU to set up mappings for PCMCIA IO-space.
My approach has been to use get_vm_area() to allocate a
region of kernel-only VM, then setup mappings using 
remap_page_range() (i.e. like ioremap() does on x86).
I've extended the PTE bits to encode the TC and SA
bits needed for PCMCIA mappings, and extended the
TLB miss handler to load the PTEA register by shifting
and masking bits from the PTE.  To supply TC&SA bits
for PCMCIA mappings, a special pgprot_t value is
constructed by the PCMCIA host bridge driver and passed
to remap_page_range().

  This means that PCMCIA IO-space mappings are loaded
into the TLB in exactly the same way as other mappings.
In particular, wired TLB entries are *not* used.  But
the downside is a small number of extra instructions in
the main path for all TLB misses (PCMCIA and non-PCMCIA
alike).

  My question is, is this approach acceptable?  Is the
performance hit in the TLB miss handler worth freeing
up the TLB slots when PCMCIA is not being heavily used?


Greg.
-- 
These are my opinions not PPIs.