On Wed, Apr 30 2008, Paul E. McKenney wrote:
quoted text
> On Wed, Apr 30, 2008 at 01:34:57PM +0200, Jens Axboe wrote:
> > On Wed, Apr 30 2008, Paul E. McKenney wrote:
> > > On Tue, Apr 29, 2008 at 06:59:36AM -0700, Paul E. McKenney wrote:
> > > > On Tue, Apr 29, 2008 at 09:26:21AM +0200, Jens Axboe wrote:
> > > > > This adds kernel/smp.c which contains helpers for IPI function calls. In
> > > > > addition to supporting the existing smp_call_function() in a more efficient
> > > > > manner, it also adds a more scalable variant called smp_call_function_single()
> > > > > for calling a given function on a single CPU only.
> > > > > 
> > > > > The core of this is based on the x86-64 patch from Nick Piggin, lots of
> > > > > changes since then. "Alan D. Brunelle" <Alan.Brunelle@hp.com> has
> > > > > contributed lots of fixes and suggestions as well.
> > > > 
> > > > Looks much better, but there still appears to be a potential deadlock
> > > > with a CPU spinning waiting (indirectly) for a grace period to complete.
> > > > Such spinning can prevent the grace period from ever completing.
> > > > 
> > > > See "!!!".
> > > 
> > > One additional question...  Why not handle memory allocation failure
> > > by pretending that the caller to smp_call_function() had specified
> > > "wait"?  The callee is in irq context, so cannot block, right?
> > 
> > (BTW a lot of thanks for your comments, I've read and understood most of
> > it, I'll reply in due time - perhaps not until next week, I'll be gone
> > from this afternoon and until monday).
> > 
> > We cannot always fallback to wait, unfortunately. If irqs are disabled,
> > you could deadlock between two CPUs each waiting for each others IPI
> > ack.
> 
> Good point!!!
> 
> > So the good question is how to handle the problem. The easiest would be
> > to return ENOMEM and get rid of the fallback, but the fallback deadlocks
> > are so far mostly in the theoretical realm since it PROBABLY would not
> > occur in practice. But still no good enough, so I'm still toying with
> > ideas on how to make it 100% bullet proof.
> 
> Here are some (probably totally broken) ideas:
> 
> 1.	Global lock so that only one smp_call_function() in the
> 	system proceeds.  Additional calls would be spinning with
> 	irqs -enabled- on the lock, avoiding deadlock.  Kind of
> 	defeats the purpose of your list, though...

That is what we used to do, that will obviously work. But defeats most
of the purpose, unfortunately :-)

quoted text
> 2.	Maintain a global mask of current targets of smp_call_function()
> 	CPUs.  A given CPU may proceed if it is not a current target
> 	and if none of its target CPUs are already in the mask.
> 	This mask would be manipulated under a global lock.
> 
> 3.	As in #2 above, but use per-CPU counters.  This allows the
> 	current CPU to proceed if it is not a target, but also allows
> 	concurrent smp_call_function()s to proceed even if their
> 	lists of target CPUs overlap.
> 
> 4.	#2 or #3, but where CPUs can proceed freely if their allocation
> 	succeeded.
> 
> 5.	If a given CPU is waiting for other CPUs to respond, it polls
> 	its own list (with irqs disabled), thus breaking the deadlock.
> 	This means that you cannot call smp_call_function() while holding
> 	a lock that might be acquired by the called function, but that
> 	is not a new prohibition -- the only safe way to hold such a
> 	lock is with irqs disabled, and you are not allowed to call
> 	the smp_call_function() with irqs disabled in the first place
> 	(right?).
> 
> #5 might actually work...

Yeah, #5 sounds quite promising. I'll see if I can work up a patch for
that, or if you feel so inclined, I'll definitely take patches :-)

The branch is 'generic-ipi' on git://git.kernel.dk/linux-2.6-block.git
The link is pretty slow, so it's best pull'ed off of Linus base. Or just
grab the patches from the gitweb interface:

http://git.kernel.dk/?p=linux-2.6-block.git;a=shortlog;h=refs/heads/generic-ipi

-- 
Jens Axboe

--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/