On 2014/09/08 23:35, Mark Kettenis wrote: > The more code & documentation I read, the more I'm convinced that > coordinating state changes between logical processors isn't necessary > and actually is responsible for the hangs people have been seeing. > > So here is a diff that does away with it all. I've tested it on a few > laptops here, but it could use testing on a somewhat wider range of > machines. I'm especially interested in seeing this tested on a dual > socket machine with apmd -A.
I'm running with this on my amd64 X220 with apm -C, this configuration used to hang every couple of days. It's a bit soon to say if it fixes things yet (IIRC some others hit the hangs more easily than me), but I haven't noticed any regressions. I've also run a cycle of lots of apm -L / apm -H in a loop while otherwise stressing the cpu, no problems seen there.