Re: [RFC PATCH] tasklet: ignore disabled tasklet in tasklet_action

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 6 Nov 2012 09:22:16 +0800 Xiaotian Feng <xtfeng@gmail.com> wrote:

> On Tue, Nov 6, 2012 at 6:52 AM, Andrew Morton <akpm@linux-foundation.org> wrote:
> > On Fri,  2 Nov 2012 10:48:54 +0800
> > Xiaotian Feng <xtfeng@gmail.com> wrote:
> >
> >> We met a ksoftirqd 100% issue, the perf top shows kernel is busy
> >> with tasklet_action(), but no actual action is shown. From dumped
> >> kernel, there's only one disabled tasklet on the tasklet_vec.
> >>
> >> tasklet_action might be handled after tasklet is disabled, this will
> >> make disabled tasklet stayed on tasklet_vec. tasklet_action will not
> >> handle disabled tasklet, but place it on the tail of tasklet_vec,
> >> still raise softirq for this tasklet. Things will become worse if
> >> device driver uses tasklet_disable on its device remove/close code.
> >> The disabled tasklet will stay on the vec, frequently __raise_softirq_off()
> >> and make ksoftirqd wakeup even if no tasklets need to be handled.
> >>
> >> This patch introduced a new TASKLET_STATE_HI bit to indicate HI_SOFTIRQ,
> >> in tasklet_action(), simply ignore the disabled tasklet and don't raise
> >> the softirq nr. In my previous patch, I remove tasklet_hi_enable() since
> >> it is the same as tasklet_enable(). So only tasklet_enable() needs to be
> >> modified, if tasklet state is changed from disable to enable, use
> >> __tasklet_schedule() to put it on the right vec.
> >
> > gee, I haven't looked at the tasklet code in 100 years.  I think I'll
> > send this in Thomas's direction ;)
> >
> > The race description seems real and the patch looks sane to me.  Are
> > you sure we can get away with never clearing TASKLET_STATE_HI?  For
> > example, what would happen if code does a tasklet_hi_schedule(t) and
> > later does a tasklet_schedule(t)?
> 
> hmm, that will be a nightmare...
> tasklet_schedule(t)/tasklet_hi_schedule(t) doesn't use list_head, they
> simply
> make t->next = NULL, then put t on the tail of
> tasklet_vec/tasklet_hi_vec. If the code does a tasklet_hi_schedule()
> and then a tasklet_schedule, the tasklet will stay on tasklet_vec and
> tasklet_hi_vec .... tasklet_hi_action will handle it first and clear
> the TASKLET_SCHED_SCHED bit, later, in tasklet_action, it will be
> handled again and hit a BUG_ON ...

Well, actually I meant if the caller reuses the tassklet_struct after
its softirq has been run.

> But if code does a tasklet_hi_schedule(), then tasklet_kil and later
> does a tasklet_schedule(), we do need clear the TASKLET_STATE_HI.

That as well ;)

> Also
> we need to remove the tasklet_hi_enable() as it is the same as
> tasklet_enable() and there's
> only one user..
> 
> I'll send you V2 patch soon, thanks.

Sounds good.