Messages in this thread | | | Date | Mon, 8 Sep 2008 13:16:48 -0400 | From | Christoph Hellwig <> | Subject | adaptive mutexes, was Re: btrfs_tree_lock & trylock |
| |
On Mon, Sep 08, 2008 at 08:07:51AM -0700, Stephen Hemminger wrote: > The problem is that they are a nuisance. It is impossible to choose > the right trade off between spin an no-spin, also they optimize for > a case that doesn't occur often enough to be justified. > > People seem to repeatedly come up with adaptive mutex based on intuitive > hunch, and never do much analysis before or afterwards. > > You need some facts to come up with a useful model: > % of time lock is contended > average lock hold time > overhead of entry-exit for lock primitive (spin time) > overhead of the adaptive version either pure spin or pure mutex > > Also, adaptive locks have even worse unfairness than spin locks under > contention.
Well, the traditional wisdom in kernel land is that you want a spinlock if the contention phases are short. But we grow more an more places where we might do sleep under the lock. One optimization would be to spin, but only if the mutex holder is not sleeping. Or we make the spinning a completely different API, mutex_lock_adaptive()
| |