Re: [RFC tip/locking/lockdep v5 06/17] lockdep: Support deadlock detection for recursive read in check_noncircular()

[Peter Zijlstra]

On Thu, Feb 22, 2018 at 03:54:34PM +0100, Peter Zijlstra wrote:
> On Thu, Feb 22, 2018 at 03:08:53PM +0800, Boqun Feng wrote:
> 
> > +static inline int hlock_conflict(struct lock_list *entry, void *data)
> > +{
> > +	struct held_lock *hlock = (struct held_lock *)data;
> > +
> > +	return hlock_class(hlock) == entry->class &&
> > +	       (hlock->read != 2 || !entry->is_rr);
> > +}
> 
> Bah, brain hurts.
> 
> So before we add prev -> this, relation, we check if there's a this ->
> prev relation already in the graph -- if so that would be a problem.
> 
> The above function has @data == @prev (__bfs_forward starts at @this,
> looking for @prev), and the above patch augments the 'class_equal' test
> with @prev not having read==2 or @entry not having xr;
> 
> This is because.... (insert brain hurt)

(hlock->read != 2 || !entry->have_xr) := !(hlock->read == 2 && entry->have_xr)

hlock->read == 2 := prev->read == 2
entry->have_xr means the last fwd link has read==2.

Together this gives that:

  @prev (Rx) ---> X ---> @entry (xR)

does not form a cycle, because:

  @enrty (xR) -> @prev (Rx)

is not strong and can be ignored.

Did I get that right? If so, the Changelog needs serious help and code
does too.