On Thu, 2008-07-31 at 16:10 +0200, John Kacur wrote:
> On Thu, Jul 31, 2008 at 4:01 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Thu, 2008-07-31 at 15:49 +0200, John Kacur wrote:
> >> Signed-off-by: John Kacur <jkacur@xxxxxxxxx>
> >> Index: linux-2.6.26-rt1/net/core/sock.c
> >> ===================================================================
> >> --- linux-2.6.26-rt1.orig/net/core/sock.c
> >> +++ linux-2.6.26-rt1/net/core/sock.c
> >> @@ -1986,11 +1986,12 @@ static __init int net_inuse_init(void)
> >>
> >> core_initcall(net_inuse_init);
> >> #else
> >> -static DEFINE_PER_CPU(struct prot_inuse, prot_inuse);
> >> +static DEFINE_PER_CPU_LOCKED(struct prot_inuse, prot_inuse);
> >>
> >> void sock_prot_inuse_add(struct net *net, struct proto *prot, int val)
> >> {
> >> - __get_cpu_var(prot_inuse).val[prot->inuse_idx] += val;
> >> + int cpu = 0;
> >> + __get_cpu_var_locked(prot_inuse, cpu).val[prot->inuse_idx] += val;
> >> }
> >> EXPORT_SYMBOL_GPL(sock_prot_inuse_add);
> >>
> >> @@ -2000,7 +2001,7 @@ int sock_prot_inuse_get(struct net *net,
> >> int res = 0;
> >>
> >> for_each_possible_cpu(cpu)
> >> - res += per_cpu(prot_inuse, cpu).val[idx];
> >> + res += per_cpu_var_locked(prot_inuse, cpu).val[idx];
> >>
> >> return res >= 0 ? res : 0;
> >> }
> >
> > This doesn't look good. You declare it as a PER_CPU_LOCKED, but then
> > never use the extra lock to synchronize data.
> >
> > Given that sock_proc_inuse_get() is a racy read anyway, the 'right' fix
> > would be to do something like:
> >
> > diff --git a/net/core/sock.c b/net/core/sock.c
> > index 91f8bbc..5a8ace4 100644
> > --- a/net/core/sock.c
> > +++ b/net/core/sock.c
> > @@ -1941,8 +1941,9 @@ static DECLARE_BITMAP(proto_inuse_idx, PROTO_INUSE_NR);
> > #ifdef CONFIG_NET_NS
> > void sock_prot_inuse_add(struct net *net, struct proto *prot, int val)
> > {
> > - int cpu = smp_processor_id();
> > + int cpu = get_cpu();
> > per_cpu_ptr(net->core.inuse, cpu)->val[prot->inuse_idx] += val;
> > + put_cpu();
> > }
> > EXPORT_SYMBOL_GPL(sock_prot_inuse_add);
> >
> > @@ -1988,7 +1989,9 @@ static DEFINE_PER_CPU(struct prot_inuse, prot_inuse);
> >
> > void sock_prot_inuse_add(struct net *net, struct proto *prot, int val)
> > {
> > - __get_cpu_var(prot_inuse).val[prot->inuse_idx] += val;
> > + int cpu = get_cpu();
> > + per_cpu(prot_inuse, cpu).val[prot->inuse_idx] += val;
> > + put_cpu();
> > }
> > EXPORT_SYMBOL_GPL(sock_prot_inuse_add);
> >
> > This disables preemption, but only for a very short time - so it doesn't
> > hurt the preempt-latency.
> >
> > The alternative is to take a lock, do the inc, and drop the lock again,
> > which is much more expensive.
> >
> >
>
> Cool, thanks for the quick feedback. What kind of criteria are used to
> decide between disabling preemption for a short time, or using the
> more expensive lock?
Basically total cost of the operation.. in this case the cost of taking
the lock utterly dwarfs the cost of the operation.
And since its Real-Time we're talking about, its the WCET of the
operation that counts.
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
