On Sat, Nov 23, 2002 at 04:05:58PM -0800, Jason R Thorpe wrote:
> On Sat, Nov 23, 2002 at 03:56:49PM -0800, Chuck Silvers wrote:
> 
>  > are we sure it isn't the icache that's the problem?
>  > the icache isn't coherent with stores from the CPU, and I can't think of
>  > anything in the NFS i/o path that would do any icache flushing.
>  > is there some mechanism for handling this, or is it just missing?
> 
> Well, pmap_enter()/pmap_remove() should be flushing the I$ for a page
> when a mapping is entered/changed/removed.

well, it's not doing that.  it used to in 1.6, but I took it out in
that big performance tweak.  if you put it back, does this problem
go away?  (note that I have better ideas on how to do icache flushing
in general, I'll get into that once we're sure this is the problem.)

-Chuck