On Mon, 20 Nov 2017 11:30:40 -0500 joe.ko...@concurrent-rt.com wrote: > Hi Steve, > A quick perusal of 4.11.12-rt16 shows that it has an > entirely new version of migrate_disable which to me appears > correct. > > In that new implementation, migrate_enable() recalculates > p->nr_cpus_allowed when it switches the task back to > using p->cpus_mask. This brings the two back into sync > if anything had happened to get them out of sync while > migration was disabled (as would happen on an affinity > change during that disable period). > > 4.9.47-rt37 has the old implementation and it appears to > have same bug as 4.4-rt though I have yet to test 4.9-rt. > > The fix in these older versions could take one of two > forms: either we recalculate p->nr_cpus_allowed when > migrate_enable goes back to using p->cpus_allowed, > as the 4.11-rt version does, or the one place where we > allow p->nr_cpus_allowed to diverge from p->cpus_allowed > be fixed. The patch I submitted earlier takes this second > approach. >
Ideally, I would like to stay close to what upstream -rt does. Would you be able to backport the 4.11-rt patch? I'm currently working on releasing 4.9-rt and 4.4-rt with the latest backports. I could easily add this one too. -- Steve
