On Fri, Feb 07, 2025 at 09:28AM +0100, Peter Zijlstra wrote:
> On Thu, Feb 06, 2025 at 07:09:56PM +0100, Marco Elver wrote:
> > Just like the pairing of attribute __acquires() with a matching
> > function-like macro __acquire(), the attribute __cond_acquires() should
> > have a matching function-like macro __cond_acquire().
> >
> > To be consistent, rename __cond_lock() to __cond_acquire().
>
> So I hate this __cond_lock() thing we have with a passion. I think it is
> one of the very worst annotations possible since it makes a trainwreck
> of the trylock code.
>
> It is a major reason why mutex is not annotated with this nonsense.
>
> Also, I think very dim of sparse in general -- I don't think I've ever
> managed to get a useful warning from between all the noise it generates.
Happy to reduce the use of __cond_lock(). :-)
Though one problem I found is it's still needed for those complex
statement-expression *_trylock that spinlock.h/rwlock.h has, where we
e.g. have (with my changes):
#define raw_spin_trylock_irqsave(lock, flags) \
__cond_acquire(lock, ({ \
local_irq_save(flags); \
_raw_spin_trylock(lock) ? \
1 : ({ local_irq_restore(flags); 0; }); \
}))
Because there's an inner condition using _raw_spin_trylock() and the
result of _raw_spin_trylock() is no longer directly used in a branch
that also does the unlock, Clang becomes unhappy and complains. I.e.
annotating _raw_spin_trylock with __cond_acquires(1, lock) doesn't work
for this case because it's in a complex statement-expression. The only
way to make it work was to wrap it into a function that has attribute
__cond_acquires(1, lock) which is what I made __cond_lock/acquire do.
For some of the trivial uses, like e.g.
#define raw_spin_trylock(lock) __cond_acquire(lock, _raw_spin_trylock(lock))
it's easy enough to remove the outer __cond_lock/acquire if e.g. the
_raw_spin_trylock has the attribute __cond_acquires. I kept these around
for Sparse compatibility, but if we want to get rid of Sparse
compatibility, some of those can be simplified.
