Re: [patch V2 4/9] softirq: Make softirq control and processing RT aware

[Peter Zijlstra]

On Wed, Dec 09, 2020 at 01:36:54PM +0100, Thomas Gleixner wrote:
> On Wed, Dec 09 2020 at 11:11, Peter Zijlstra wrote:
> > On Fri, Dec 04, 2020 at 06:01:55PM +0100, Thomas Gleixner wrote:
> >> From: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> >> +	/* First entry of a task into a BH disabled section? */
> >> +	if (!current->softirq_disable_cnt) {
> >> +		if (preemptible()) {
> >> +			local_lock(&softirq_ctrl.lock);
> >
> > AFAICT this significantly changes the locking rules.
> >
> > Where previously we could do:
> >
> > 	spin_lock(&ponies)
> > 	spin_lock_bh(&foo);
> >
> > vs
> >
> > 	spin_lock_bh(&bar);
> > 	spin_lock(&ponies)
> >
> > provided the rest of the code observed: bar -> ponies -> foo
> > and never takes ponies from in-softirq.
> >
> > This is now a genuine deadlock on RT.
> 
> I know, but making this work is trying to square the circle.

:-)

> Any approach we tried before going this way had worse problems than
> this particular limitation.

OK, but that would've been very good Changelog material methinks.

Also, then we should probably make sure PREEMPT_RT=n builds start
suffering the same problem by adding the local_lock unconditionally,
otherwise this keeps being a PREEMPT_RT special and we'll keep having to
fix it up.