Re: [RT PATCH] sched: rt: fix two possible deadlocks in push_irq_work_func

[Thomas Gleixner]

On Sat, 14 Nov 2015, yjin wrote:
> On 2015å11æ14æ 12:25, Steven Rostedt wrote:
> > On Sat, 14 Nov 2015 10:53:18 +0800
> > <yanjiang.jin@xxxxxxxxxxxxx> wrote:
> > 
> > > From: Yanjiang Jin <yanjiang.jin@xxxxxxxxxxxxx>
> > > 
> > > This can only happen in RT kernel due to run_timer_softirq() calls
> > > irq_work_tick() when CONFIG_PREEMPT_RT_FULL is enabled as below:
> > > 
> > > static void run_timer_softirq(struct softirq_action *h)
> > > {
> > > ........
> > > if defined(CONFIG_IRQ_WORK) && defined(CONFIG_PREEMPT_RT_FULL)
> > >          irq_work_tick();
> > > endif
> > > ........
> > > }
> > > 
> > > Use raw_spin_{un,}lock_irq{save,restore} in push_irq_work_func() to
> > > prevent following potentially deadlock scenario:
> > Ug. No, the real fix is that the irq work is to be run from hard
> > interrupt context.
> But if so, we shouldn't call irq_work_tick() in run_timer_softirq(), right?

The work is marked IRQ_WORK_HARD_IRQ so it should be run from hard irq
context. We only run the work, which is not marked IRQ_WORK_HARD_IRQ
from the softirq on RT.
 
Thanks,

	tglx