On Wed, Nov 03, 2021 at 10:54:09PM +0100, Christophe JAILLET wrote: > Le 03/11/2021 à 22:41, Matthew Wilcox a écrit : > > On Wed, Nov 03, 2021 at 10:35:34PM +0100, Christophe JAILLET wrote: > > > Use 'percpu_ref_tryget_live_rcu()' instead of 'percpu_ref_tryget_live()' to > > > save a few cycles when it is known that the rcu lock is already > > > taken/released. > > > > If this is really important, we can add an __xa_load() which doesn't > > take the RCU read lock. > > There are a few: > rcu_read_lock(); > mem = xa_load(...); > rcu_read_unlock(); > patterns here and there. If that's all they are, then the rcu_read_lock() and unlock can be deleted. if they're actually rcu_read_lock() mem = xa_load(...); try_get_ref(mem); rcu_read_unlock(); then of course they can't be. > I don't have any numbers of if saving some rcu_read_lock/rcu_read_unlock > would be useful in these cases. > > The only numbers I have are in [1]. > > [1]: https://lore.kernel.org/linux-kernel/cover.1634822969.git.asml.silence@xxxxxxxxx/ It may not be worth hyperoptimising the slow path like this patch ...
