On Thu, 16 Aug 2001, Andrea Arcangeli wrote:
> On Thu, Aug 16, 2001 at 07:44:04PM +0100, Mark Hemment wrote:
> >   KM_BOUNCE_WRITE may waste virtual address-space, but it saves on
> > expensive shootdowns.
>
> I believe the slowdown is more a sign that kmap is not fast enoguh, not
> that you really need the BOUNCE_WRITE.

  I did knock up a patch where the shooter didn't need to busy-wait.  It
worked, but it assumed an IPI would be delivered as soon as a processor
re-enabled interrupts.
  After posting the patch, I was also made unsure that it would cover all
conditions after seeing a comment in arch/i386/kernel/smp.c;
	/*
	 * This was a BUG() but until someone can quote me the
	 * line from the intel manual that guarantees an IPI to
	 * multiple CPUs is retried _only_ on the erroring CPUs
	 * its staying as a return
	 */

  Can't remember now why I thought this might cause my patch problems
...but I sure it does.
  Any, I've attached my posting to lse and the patch I sent.  I think Andi
Kleen replied saying you had played with using the NMI for shootdowns?

  I do believe that improving the shootdown code won't remove the need for
BOUNCE_WRITE.  Only a local, temporary, mapping is need and the lighter
mapping is, the better.

Mark