Re: [PATCH 11/11] s390: convert to generic helpers for IPI function calls

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, Apr 23 2008, Heiko Carstens wrote:
> > > > Your generic patch changes semantics of smp_call_function(). At least the
> > > > s390 version of smp_call_function() mask used to not return unless all IPIs
> > > > where received on the target cpus.
> > > > Your patch changes that. It will already return as soon as all IPIs have
> > > > been sent. Unfortunately Martin's etr code in arch/s390/kernel/time.c
> > > 
> > > Yep, this is what I mentioned in the intro mail.
> > > 
> > > > relies on exactly the old semantics.
> > > 
> > > OK, I could not see anything which relied on that and I did think that
> > > it was a somewhat odd interface. By the time that smp_call_function()
> > > stops waiting for startups, we could easily be in the condition that the
> > > N-1 CPUs executed the function part a long time ago. I just didn't see
> > > much usefulness in that, you may as well just set wait == 1 in that
> > > case.
> > > 
> > > > Since this seems to be an s390 only issue, could you just drop the s390
> > > > conversion patch? We'll have to think about how to change the err code
> > > > and convert later I guess.
> > > 
> > > I'll drop the s390 bits. I can easily add something like
> > > smp_call_function_foo() that has this behaviour.
> > 
> > I added an smp_call_function_sync() for this purpose and kept the s390
> > conversion, I would very much appreciate if you could look it over. I
> > wont post the full patchset again today, but you can inspect it here:
> > 
> > http://git.kernel.dk/?p=linux-2.6-block.git;a=shortlog;h=refs/heads/generic-ipi
> > 
> > The interesting bits are patch #1 for the generic bits and patch #11 for
> > the s390 part. Thanks!
> 
> I don't think that works: the old code also relied on the fact that there
> could be only one cpu sending an smp_call_* IPI (serialized by call_lock).
> The current etr code exploits that to serialize all cpus:
> 
> Do smp_call_function() to force all other cpus into clock_sync_cpu_start()
> where they sort of busy wait(!). Then the master cpu is syncing the clock
> and when it finished the other cpus may continue again and finally leave
> the smp_call_function interrupt handler.
> 
> Your patch series however doesn't make sure anymore that there is only
> one cpu doing an smp_call_function*, so we may deadlock as soon as two
> cpus are trying to synchronize all cpus this way.
> After all I think the etr code should be converted to use stop_machine_run
> instead. However that doesn't allow for master/slave cpus.

You are right. I'll drop the code again, it would be nice to get rid of
this interface, I guess s390 can do that when they convert to the
generic helpers.

-- 
Jens Axboe