Re: [PATCH] Documentation/locking: Document the semantics of spin_is_locked()

[Will Deacon]

On Wed, Feb 28, 2018 at 12:24:03PM +0100, Andrea Parri wrote:
> On Wed, Feb 28, 2018 at 10:56:32AM +0000, Will Deacon wrote:
> > On Wed, Feb 28, 2018 at 11:39:32AM +0100, Andrea Parri wrote:
> > > There appeared to be a certain, recurrent uncertainty concerning the
> > > semantics of spin_is_locked(), likely a consequence of the fact that
> > > this semantics remains undocumented or that it has been historically
> > > linked to the (likewise unclear) semantics of spin_unlock_wait().
> > > 
> > > Document this semantics.
> > > 
> > > Signed-off-by: Andrea Parri <parri.andrea@xxxxxxxxx>
> > > Cc: Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>
> > > Cc: Will Deacon <will.deacon@xxxxxxx>
> > > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > > Cc: Boqun Feng <boqun.feng@xxxxxxxxx>
> > > Cc: Nicholas Piggin <npiggin@xxxxxxxxx>
> > > Cc: David Howells <dhowells@xxxxxxxxxx>
> > > Cc: Jade Alglave <j.alglave@xxxxxxxxx>
> > > Cc: Luc Maranget <luc.maranget@xxxxxxxx>
> > > Cc: "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx>
> > > Cc: Akira Yokosawa <akiyks@xxxxxxxxx>
> > > ---
> > >  include/linux/spinlock.h | 11 +++++++++++
> > >  1 file changed, 11 insertions(+)
> > > 
> > > diff --git a/include/linux/spinlock.h b/include/linux/spinlock.h
> > > index 4894d322d2584..2639fdc9a916c 100644
> > > --- a/include/linux/spinlock.h
> > > +++ b/include/linux/spinlock.h
> > > @@ -380,6 +380,17 @@ static __always_inline int spin_trylock_irq(spinlock_t *lock)
> > >  	raw_spin_trylock_irqsave(spinlock_check(lock), flags); \
> > >  })
> > >  
> > > +/**
> > > + * spin_is_locked() - Check whether a spinlock is locked.
> > > + * @lock: Pointer to the spinlock.
> > > + *
> > > + * This function is NOT required to provide any memory ordering
> > > + * guarantees; it could be used for debugging purposes or, when
> > > + * additional synchronization is needed, accompanied with other
> > > + * constructs (memory barriers) enforcing the synchronization.
> > > + *
> > > + * Return: 1, if @lock is (found to be) locked; 0, otherwise.
> > > + */
> > 
> > I also don't think this is quite right, since the spin_is_locked check
> > must be ordered after all prior lock acquisitions (to any lock) on the same
> > CPU. That's why we have an smp_mb() in there on arm64 (see 38b850a73034f).
> 
> So, arm64 (and powerpc) complies to the semantics I _have_ in mind ...

Sure, but they're offering more than that at present. If I can remove the
smp_mb() in our spin_is_locked implementation, I will, but we need to know
what that will break even if you consider that code to be broken because it
relies on something undocumented.

> > So this is a change in semantics and we need to audit the users before
> > proceeding. We should also keep spin_is_locked consistent with the versions
> > for mutex, rwsem, bit_spin.
> 
> Well, strictly speaking, it isn't (given that the current semantics is,
> as reported above, currently undocumented); for the same reason, cases
> relying on anything more than _nothing_ (if any) are already broken ...

I suppose it depends on whether you consider the code or the documentation
to be authoritative. I tend to err on the side of the former for the kernel.
To be clear: I'm perfectly ok relaxing the semantics, but only if there's
some evidence that you've looked at the callsites and determined that they
won't break.  That's why I think a better first step would be to convert a
bunch of them to using lockdep for the "assert that I hold this lock"
checks, so we can start to see where the interesting cases are.

Will