On Wed, 30 Mar 2011 09:17:23 +0800 Shaohua Li <shaohua.li@xxxxxxxxx> wrote:
> On Wed, 2011-03-30 at 09:06 +0800, Andrew Morton wrote:
> > On Wed, 30 Mar 2011 09:01:22 +0800 Shaohua Li <shaohua.li@xxxxxxxxx> wrote:
> >
> > > +/*
> > > + * Make sure vm_committed_as in one cacheline and not cacheline shared with
> > > + * other variables. It can be updated by several CPUs frequently.
> > > + */
> > > +struct percpu_counter vm_committed_as ____cacheline_internodealigned_in_smp;
> >
> > The mystery deepens. The only cross-cpu writeable fields in there are
> > percpu_counter.lock and its companion percpu_counter.count. If CPUs
> > are contending for the lock then that itself is a problem - how does
> > adding some padding to the struct help anything?
> I had another patch trying to address the lock contention (for case
> OVERCOMMIT_GUESS), will send out soon. But thought better to have the
> correct alignment for OVERCOMMIT_NEVER case.
I still don't understand why adding
____cacheline_internodealigned_in_smp to vm_committed_as improves
anything.
Here it is:
struct percpu_counter {
spinlock_t lock;
s64 count;
#ifdef CONFIG_HOTPLUG_CPU
struct list_head list; /* All percpu_counters are on a list */
#endif
s32 __percpu *counters;
};
and your patch effectively converts this to
struct percpu_counter {
spinlock_t lock;
s64 count;
#ifdef CONFIG_HOTPLUG_CPU
struct list_head list; /* All percpu_counters are on a list */
#endif
s32 __percpu *counters;
+ char large_waste_of_space[lots];
};
how is it that this improves things?
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>