Re: [linuxsh-dev] 2.5.8-pre1

[NIIBE Y. <gn...@m1...>]

Robert Love wrote:
 > So what will it take to get kernel preemption support added? :)
 > 
 > preempt-kernel has been in 2.5 since 2.5.4-pre.  The work required to
 > add it to the SH architecture is fairly minimal, with most of the work
 > being in entry.S.  If you look at MontaVista's (Jeremy Siegel's) SH
 > preempt support for 2.4, the work for 2.5 is about the same.  The main
 > difference in 2.5 is the use of the thread_info structure for
 > preempt_count and that the interrupt return path no longer calls
 > preempt_schedule, but schedule itself.
 > 
 > I'd be happy to help ...

Thanks to point it out.  I was thinking about it.  Yes, I think that
it's ready to merge the preemption work.  If you or Jeremy go ahead to
do so, it's welcome.

Well, I have a issue.

__copy_user_page or __clear_user_page is the most time consuming
routine in the kernel.  For SH-4, we could use FPU registers to
transfer 8-byte at a instruction to improve speed (in theory, I
haven't tested).  This would improve performance of usual case, but I
think that we need to inhibit preemption to protect FPU usage in the
kernel.  Here we would have large latency.

I'm afraid of this, so I haven't moved to page copying/clearing with
FPU registers...
--