2013-03-22 19:55:53

by Andi Kleen

Subject: racy jump label users

Jason,

I noticed that a lot of the jump label users are racy,
because they implement something like this

static void sched_feat_disable(int i)
{
if (static_key_enabled(&sched_feat_keys[i]))
static_key_slow_dec(&sched_feat_keys[i]);
}

static void sched_feat_enable(int i)
{
if (!static_key_enabled(&sched_feat_keys[i]))
static_key_slow_inc(&sched_feat_keys[i]);
}

with no extra locking, controlled by sysfs. If two
CPUs do this in parallel the reference can be set multiple
times, which gives very unexpected semantics for a sysfs boolean.

Most likely you need a static_key_slow_setstate()
that does the check and set inside the jump label lock.

I understand that for inside kernel use reference
counts are the right semantics, but they are not so
good for sysfs interfaces.

-Andi

--
[email protected] -- Speaking for myself only

2013-04-18 18:26:09

by Jason Baron

Subject: Re: racy jump label users

Hi Andi,

Agreed. However, other users of 'static_key_enabled()', provide their own
locking. For example, in kernel/tracepoint.c, 'static_key_enabled()',
relies on
the tracepoints_mutex. Were there any other users that are problematic?

I agree a 'setstate' would be nice. Maybe something like:

static_key_slow_set_true();
static_key_slow_set_false();

Thanks,

-Jason

On 03/22/2013 03:55 PM, Andi Kleen wrote:
> Jason,
>
> I noticed that a lot of the jump label users are racy,
> because they implement something like this
>
> static void sched_feat_disable(int i)
> {
> if (static_key_enabled(&sched_feat_keys[i]))
> static_key_slow_dec(&sched_feat_keys[i]);
> }
>
> static void sched_feat_enable(int i)
> {
> if (!static_key_enabled(&sched_feat_keys[i]))
> static_key_slow_inc(&sched_feat_keys[i]);
> }
>
> with no extra locking, controlled by sysfs. If two
> CPUs do this in parallel the reference can be set multiple
> times, which gives very unexpected semantics for a sysfs boolean.
>
> Most likely you need a static_key_slow_setstate()
> that does the check and set inside the jump label lock.
>
> I understand that for inside kernel use reference
> counts are the right semantics, but they are not so
> good for sysfs interfaces.
>
> -Andi
>