On Sun, Aug 25, 2002 at 11:36:28AM +1200, Gregory McGarry wrote:

 > I'm sure you're just itching to hear the story behind this one.

Hehee :-)

 > Originally, I did the mips work which allowed an atomic sequence
 > to be restarted on any preemption or interrupt/exception.  Restarting
 > on an interrupt is easy on mips, since there is only two possible
 > entry points.  Seems like a simple generalisation, although I
 > wasn't ever able to justify it with a good practical reason.  The
 > closest I came was an example of profiling some short code sequence.
 > Would also be useful though for atomic sequences inside the kernel.
 > i386 was a little harder, but still possible.  m68k convinced me to
 > only restart on preemption.

I think for userland ras, we should stick with preemption.

However, on some platforms, there is a good reason to have kernel RAS
(it could be really useful on real-i386 and sparcv7/v8 on my "newlock"
branch).  I think we could think about extending the current ras API
or adding a kras API at some point in the future ... but it's not a biggie
right now.

-- 
        -- Jason R. Thorpe <thorpej@wasabisystems.com>