On 23.11.2016 15:25, Daniel Vetter wrote:
On Wed, Nov 23, 2016 at 03:03:36PM +0100, Peter Zijlstra wrote:
On Wed, Nov 23, 2016 at 12:25:22PM +0100, Nicolai Hähnle wrote:
@@ -473,7 +476,14 @@ void __sched ww_mutex_unlock(struct ww_mutex *lock) */ mutex_clear_owner(&lock->base); #endif
- __mutex_fastpath_unlock(&lock->base.count, __mutex_unlock_slowpath);
- /*
* A previously _not_ waiting task may acquire the lock via the fast* path during our unlock. In that case, already waiting tasks may have* to back off to avoid a deadlock. Wake up all waiters so that they* can check their acquire context stamp against the new owner.*/- __mutex_fastpath_unlock(&lock->base.count,
__mutex_unlock_slowpath_wakeall);}
So doing a wake-all has obvious issues with thundering herd etc.. Also, with the new mutex, you'd not be able to do hand-off, which would introduce starvation cases.
Ideally we'd iterate the blocked list and pick the waiter with the earliest stamp, or we'd maintain the list in stamp order instead of FIFO, for ww_mutex.
Not sure we'll win that much, at least I think we still need to wake up everyone with earlier stamp than the one of the task that just released the lock. Otherwise there's deadlocks. So just cuts the wakeups in half, on average.
What we could do is do a handoff-hint with the timestamp of earliest task we believe should get the lock. Everyone with a later timestamp that gets woken then knows that they definitely have a deadlock situation and need to back off (thread 2 in the example).
thread 1 would get woken, and would be able to take the lock, except when thread 0 successfully raced it and stole the lock. And anyone else racing in with later timestamps would also immediately back off, ensuring fairness.
I did consider maintaining stamp order in the waiter list and originally decided against it because I just wanted a simple and conservative fix to avoid adding new regressions.
Now that a different approach is needed for >= 4.9 anyway mostly due to the hand-off logic, I'm reconsidering this.
I do believe we can win a bit by keeping the wait list sorted, if we also make sure that waiters don't add themselves in the first place if they see that a deadlock situation cannot be avoided.
I will probably want to extend struct mutex_waiter with ww_mutex-specific fields to facilitate this (i.e. ctx pointer, perhaps stamp as well to reduce pointer-chasing). That should be fine since it lives on the stack.
In the meantime, I'd appreciate it if patch #1 could be accepted as-is for stable updates to <= 4.8. It fixes a real (if rare) bug, and the stampede inefficiency isn't a problem in practice at least for GPU applications.
Thanks, Nicolai
Without thinking it through in detail this is a PI issue, except that we replace boosting with wakeup&back-off. Could we perhaps steal something from rt mutexes to make it fair&efficient? -Daniel