Re: [patch] reduce IPI noise due to /dev/cdrom open/close

[Andrew Morton <akpm@osdl.org>]

On Tue, 04 Jul 2006 09:49:14 +0200
Jes Sorensen <jes@sgi.com> wrote:

> Keith Owens wrote:
> >> static void invalidate_bh_lrus(void)
> >> {
> >> -	on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
> >> +	/*
> >> +	 * Need to hand down a copy of the mask or we wouldn't be run
> >> +	 * anywhere due to the original mask being cleared
> >> +	 */
> >> +	cpumask_t mask = lru_in_use;
> >> +	cpus_clear(lru_in_use);
> >> +	schedule_on_each_cpu_mask(invalidate_bh_lru, NULL, mask);
> >> }
> > 
> > Racy?  Start with an empty lru_in_use.
> > 
> > Cpu A                         Cpu B
> > invalidate_bh_lrus()
> > mask = lru_in_use;
> > preempted
> >                               block I/O
> > 			      bh_lru_install()
> > 			      cpu_set(cpu, lru_in_use);
> > resume
> > cpus_clear(lru_in_use);
> > schedule_on_each_cpu_mask() - does not send IPI to cpu B
> 
> I guess the real question is whether the device is still valid for new
> IO by the time we hit the invalidate function. If not, and that was my,
> possibly flawed, assumption, then it's not an issue. Whatever bh's are
> left in the lrus from other devices will be handled on the next hit.
> 

The problem is that we can actually lose bits in invalidate_bh_lrus().

CPU0				CPU1

mask = lru_in_use;
				cpu_set(lru_in_use, 1);
cpus_clear(lru_in_use);

Now we have a stray bh which nobody knows about.