Re: [v4.14-rt][report] arm: run: stress-ng --class os --all 0 -t 5m

[Steven Rostedt]

On Tue, 19 Dec 2017 16:33:10 +0100
Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> wrote:

> On 2017-12-19 10:28:39 [-0500], Steven Rostedt wrote:
> > On Tue, 19 Dec 2017 16:04:18 +0100
> > Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> wrote:
> >   
> > > > 
> > > > The above just seems wrong. local_irq_disable() should imply
> > > > local_bh_disable(), as it doesn't let softirqs run either.    
> > > 
> > > Where does local_irq_disable() imply this?  
> > 
> > If it doesn't explicitly do so, it probably should. How can we have a
> > softirq execute when irqs are disabled?  
> 
> There are not. With local_bh_disable() the softirq will run on
> local_bh_enable(). Without it (and with or without local_irq_disable())
> the softirq won't run but wakeup the ksoftirq thread. We can't do the
> wake while holding the hrtimer lock. This is not RT specific.
> 

Then there should be a comment there, as it is way too subtle. As
local_bh_disable() is usually used only to prevent softirq from running
on the current CPU during a critical period. Where, here we are using
it to avoid a wake up of ksoftirqd.

-- Steve