Re: [RFC PATCH 2/3] lockdep: Merge hardirq_threaded and irq_config together

[Frederic Weisbecker]

On Mon, Mar 23, 2020 at 03:53:14PM +0100, Sebastian Andrzej Siewior wrote:
> On 2020-03-23 15:02:20 [+0100], Peter Zijlstra wrote:
> > On Mon, Mar 23, 2020 at 04:32:06AM +0100, Frederic Weisbecker wrote:
> > > These fields describe the same state: a code block running in hardirq
> > > that might be threaded under specific configurations.
> > > 
> > > Merge them together in the same field. Also rename the result as
> > > "hardirq_threadable" as we are talking about a possible state and not
> > > an actual one.
> > 
> > What isn't instantly obvious is that they cannot overlap. For instance
> > mainline with force threaded interrupt handlers on, can't that have the
> > irq_work nest inside a threaded handler ?
> 
> I remember we kept them due to the nesting. A threaded-interrupt can be
> interrupted by irq_work/hrtimer/posix-timer. 
> So in a threaded interrupt it is okay to use a sleeping lock. It is not
> okay with irq_work on-top - unless it is the non-atomic one.

Hmm, with the current layout which is:

    if (curr->hardirq_threaded || curr->irq_config)
        curr_inner = LD_WAIT_CONFIG;
    else
        curr_inner = LD_WAIT_SPIN;

you are not protected against that.

But anyway none of that should happen if hardirq_threaded and irq_config
are respectively only used when hardirq threads are disabled and irq_work
is always hardirq, right?

> 
> > I *think* it just about works out, but it definitely wants a little more
> > than this.
> 
> Sebastian