Messages in this thread | | | Date | Mon, 21 Sep 2015 13:39:48 -0700 | Subject | Re: [PATCH 4/5] locking/rwsem: Use acquire/release semantics | From | Linus Torvalds <> |
| |
On Mon, Sep 21, 2015 at 1:17 PM, Davidlohr Bueso <dave@stgolabs.net> wrote: > @@ -114,7 +114,7 @@ static inline void __downgrade_write(struct rw_semaphore *sem) > { > long tmp; > > - tmp = atomic_long_add_return(-RWSEM_WAITING_BIAS, > + tmp = atomic_long_add_return_acquire(-RWSEM_WAITING_BIAS, > (atomic_long_t *)&sem->count); > if (tmp < 0) > rwsem_downgrade_wake(sem);
Careful. I'm pretty sure this is wrong.
When we downgrade exclusive ownership to non-exclusive, that should be a *release* operation. Anything we did inside the write-locked region had damn better _stay_ inside the write-locked region, we can not allow it to escape down into the read-locked side. So it needs to be at least a release.
In contrast, anything that we do in the read-locked part is fine to be re-ordered into the write-locked exclusive part, so it does *not* need acquire ordering (the original write locking obviously did use acquire, and acts as a barrier for everything that comes in the locked region).
I tried to look through everything, and I think this is the only thing you got wrong, but I'd like somebody to double-checks. Getting the acquire/release semantics wrong will cause some really really subtle and hard-as-hell-to-find bugs. So let's be careful out there, ok?
Linus
| |