On Thu, May 12, 2016 at 02:19:07PM +0200, Michal Hocko wrote: > On Thu 12-05-16 14:12:04, Peter Zijlstra wrote: > > On Wed, May 11, 2016 at 08:03:46PM +0200, Michal Hocko wrote: > > > I still cannot say I would understand why the pending > > > RWSEM_WAITING_BIAS matters but I would probably need to look at the code > > > again with a clean head, __rwsem_wake is quite tricky... > > > > Ah, you're asking why an unconditional __rwsem_wake(ANY) isn't enough? > > > > Because; if at that point there's nobody waiting, we're left with an > > empty list and WAITER_BIAS set. This in turn will make all fast paths > > fail. > > > > Look at rwsem_down_read_failed() for instance; if we enter that we'll > > unconditionally queue ourself, with nobody left to come wake us. > > This is still not clear to me because rwsem_down_read_failed will call > __rwsem_do_wake if the count is RWSEM_WAITING_BIAS so we shouldn't go to > sleep and get the lock. So you are right that we would force everybody > to the slow path which is not great but shouldn't cause incorrect > behavior. I guess I must be missing something obvious here...
Ah me too; I missed the obvious: we do the __rwsem_do_wake() after we add ourselves to the list, which means we'll also wake ourselves. I'll have more thinking..