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

[Julia Cartwright]

On Thu, Feb 15, 2018 at 03:07:07PM -0500, Steven Rostedt wrote:
> On Thu, 15 Feb 2018 18:20:41 +0100
> Sebastian Andrzej Siewior <bige...@linutronix.de> 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.
> 
> >     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.

Well, it can be, but the percpu access needs to go with it; so an
alternative solution would be the below.

I'm also wondering whether the t->next = NULL assignment could be lifted
out from irqs-disabled region.

   Julia

diff --git a/kernel/softirq.c b/kernel/softirq.c
index fa7ed89a9fcf..177de3640c78 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -461,12 +461,14 @@ static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 
 static void __tasklet_schedule_common(struct tasklet_struct *t,
-                                     struct tasklet_head *head,
+                                     struct tasklet_head __percpu *headp,
                                      unsigned int softirq_nr)
 {
+       struct tasklet_head *head;
        unsigned long flags;
 
        local_irq_save(flags);
+       head = this_cpu_ptr(headp);
        t->next = NULL;
        *head->tail = t;
        head->tail = &(t->next);
@@ -476,14 +478,14 @@ static void __tasklet_schedule_common(struct 
tasklet_struct *t,
 
 void __tasklet_schedule(struct tasklet_struct *t)
 {
-       __tasklet_schedule_common(t, this_cpu_ptr(&tasklet_vec),
+       __tasklet_schedule_common(t, &tasklet_vec,
                                  TASKLET_SOFTIRQ);
 }
 EXPORT_SYMBOL(__tasklet_schedule);
 
 void __tasklet_hi_schedule(struct tasklet_struct *t)
 {
-       __tasklet_schedule_common(t, this_cpu_ptr(&tasklet_hi_vec),
+       __tasklet_schedule_common(t, &tasklet_hi_vec,
                                  HI_SOFTIRQ);
 }
 EXPORT_SYMBOL(__tasklet_hi_schedule);