On Sat, 2009-11-14 at 20:08 +0100, Joakim Tjernlund wrote:

Dan Malek [off-list ref] wrote on 14/11/2009 19:08:43:

quoted

On Nov 14, 2009, at 2:42 AM, Joakim Tjernlund wrote:

quoted

..... Avoid this by always pinning
kernel instruction TLB space.

You may as well map the data space, too, since you have
reserved the entries.  Take advantage of that performance.
Also, some processor variants have very few TLB entries,
and may only reserve two entries (although the flag says
reserve 4).  Ensure there are sufficient resources to do
what you want.  This is the reason the option is configurable.

Scott had some concerns about pinning the data space too. That is
is why I left the data TLB pinning behind the the config option.

How to make better use of the remaining ITLB slots is tricky.
Somehow one would want to map at lest one to modules but I cannot see how.

No. If you use modules, you pay the price. Sane embedded solutions
running in "tight" environments don't use modules :-) No point pinning
TLB entries on the vmalloc space, really.

What -might- be more useful is to look at Grant work on re-doing the
early ioremap and providing a way to provide what the old
io_block_mapping() did, but with dynamically chosen virtual addresses,
to have a pinned entry covering most common IOs.

Cheers,
Ben.