On 27/11/14 21:37, Thomas Gleixner wrote:
On Thu, 27 Nov 2014, Daniel Thompson wrote:
Currently gic_raise_softirq() unconditionally takes and releases a lock whose only purpose is to synchronize with the b.L switcher.
Remove this lock if the b.L switcher is not compiled in.
I think the patches are in the wrong order. We optimize for the sane use case first, i.e BL=n. So you want to make the locking of irq_controller_lock in gic_raise_softirq() conditional in the first place, which should have been done when this was introduced.
Once you have isolated that you can apply your split lock patch for the BL=y nonsense.
Adding more locks first and then optimizing them out does not make any sense.
You original described the use of irq_controller_lock for its current dual purpose to be an abuse of the lock. Does it really make more sense to optimize before we correct the abuse?
How about just squashing them together? It reduces the combined diffstat by ~10%...
Daniel.