Re: [PATCH 1/2] kernel/sofirq: consolidate common code in __tasklet_schedule() + _hi_

[Sebastian Andrzej Siewior]

On 2018-02-15 15:07:07 [-0500], Steven Rostedt wrote:
> On Thu, 15 Feb 2018 18:20:41 +0100
> Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> wrote:
> 
> > -void __tasklet_schedule(struct tasklet_struct *t)
> > +static void __tasklet_schedule_common(struct tasklet_struct *t,
> > +				      struct tasklet_head *head,
> > +				      unsigned int softirq_nr)
> >  {
> >  	unsigned long flags;
> >  
> >  	local_irq_save(flags);
> 
> If you look at the original patch, it did not move local_irq_save()
> into the common function.
correct butâ

> >  	t->next = NULL;
> > -	*__this_cpu_read(tasklet_vec.tail) = t;
> > -	__this_cpu_write(tasklet_vec.tail, &(t->next));
> > -	raise_softirq_irqoff(TASKLET_SOFTIRQ);
> > +	*head->tail = t;
> > +	head->tail = &(t->next);
> > +	raise_softirq_irqoff(softirq_nr);
> >  	local_irq_restore(flags);
> >  }
> > +
> > +void __tasklet_schedule(struct tasklet_struct *t)
> > +{
> > +	__tasklet_schedule_common(t, this_cpu_ptr(&tasklet_vec),
> 
> What can happen is, we reference (tasklet_vec) on one CPU, get
> preempted (running in ksoftirqd), scheduled on another CPU, then when
> inside the common code, we are executing on a different CPU than the
> tasklet is for. The rasise_softirq() is happening on the wrong CPU.
> 
> The local_irq_save() can't be moved to the common function. It must be
> done by each individual function.

That __tasklet_schedule_common() part is usually invoked from an
interrupt context which attempts to schedule the tasklet so it should
with invoked with interrupts off. However there was one warn_on()
because something early in the boot managed to invoke it without
interrupts disabled (which I missed). Okay, granted, fixed.

As for the second invocation (tasklet_action_common() part) is always
invoked in BH-disabled context (even if called from ksoftirqd) so you
are never preemptible() and can't switch CPUs.
So I am going to correct this patch as you suggested but I don't see the
reason to do the same in the second one.

> -- Steve
Sebastian