Re: [PATCH] locking/refcount: add sparse annotations to dec-and-lock functions

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Jan 06, 2020 at 06:54:59PM +0100, Luc Van Oostenryck wrote:
> On Mon, Jan 06, 2020 at 04:41:19PM +0100, Peter Zijlstra wrote:

> > extern bool spin_trylock(spinlock_t *lock) __attribute__((context(lock, 0, spin_trylock(lock));
> 
> Well, allowing arbitrary conditions would be hard/impossible but you're
> only asking to have the *return value* as condition, right? That looks
> as reasonably feasible.

Just the return value would cover all the known cases yes. At the time
I might have been somewhat over ambitious..

> > Basically have sparse do a transform on its own expression tree and
> > inject the very same crud we now do manually. This avoids cluttering the
> > kernel tree with this nonsense.
> 
> So, a call of a function declared with __acquires() or releases() is
> interpreted by Sparse as if the call is immediately followed by an
> increase or a decrease of the context. It wouldn't be very hard to
> add a new attribute (something like __cond_context) and let Sparse do
> as if a call to a function with such attribute is directly followed
> by a test of its return value and a corresponding change in the context.
> It would boil down to:
> 
> 	extern bool spin_trylock(lock) __cond_context(lock);
> 
> 	if (spin_trylock(lock)) {
> 		/* do crap */
> 		spin_unlock();
> 	}
> 
> behaving like the following code currently would:
> 
> 	extern bool spin_trylock(lock);
> 
> 	if (spin_trylock(lock)) {
> 		__acquire(lock);
> 		/* do crap */
> 		spin_unlock();
> 	}
> 
> 
> Would something like this be satisfactory?

Very much so, Thanks!