Messages in this thread | | | Date | Fri, 18 Jan 2019 21:06:57 +0100 | From | Peter Zijlstra <> | Subject | Re: Question about qspinlock nest |
| |
On Fri, Jan 18, 2019 at 09:50:12AM -0500, Waiman Long wrote: > On 01/18/2019 05:02 AM, Peter Zijlstra wrote: > > > >> e.g. We can't take an SError during the SError handler. > >> > >> But we can take this SError/NMI on another CPU while the first one is still > >> running the handler. > >> > >> These multiple NMIlike notifications mean having multiple locks/fixmap-slots, > >> one per notification. This is where the qspinlock node limit comes in, as we > >> could have more than 4 contexts. > > Right; so Waiman was going to do a patch that reverts to test-and-set or > > something along those lines once we hit the queue limit, which seems > > like a good way out. Actually hitting that nesting level should be > > exceedingly rare. > > Yes, I am working on a patch to support arbitrary levels of nesting. It > is easy for PV qspinlock as lock stealing is supported. > > For native qspinlock, we cannot do lock stealing without incurring a > certain amount of overhead in the regular slowpath code. It was up to > 10% in my own testing. So I am exploring an alternative that can do the > job without incurring any noticeable performance degradation in the > slowpath. I ran into a race condition which I am still trying to find > out where that comes from. Hopefully, I will have something to post next > week.
Where does the overhead come from? Surely that's not just checking that bound?
| |