On 09/10/19 14:12, Scott Wood wrote: > On Wed, 2019-10-09 at 09:27 +0200, Juri Lelli wrote: > > On 09/10/19 01:25, Scott Wood wrote: > > > On Tue, 2019-10-01 at 10:52 +0200, Juri Lelli wrote: > > > > On 30/09/19 11:24, Scott Wood wrote: > > > > > On Mon, 2019-09-30 at 09:12 +0200, Juri Lelli wrote: > > > > > > > > [...] > > > > > > > > > > Hummm, I was actually more worried about the fact that we call > > > > > > free_old_ > > > > > > cpuset_bw_dl() only if p->state != TASK_WAKING. > > > > > > > > > > Oh, right. :-P Not sure what I had in mind there; we want to call > > > > > it > > > > > regardless. > > > > > > > > > > I assume we need rq->lock in free_old_cpuset_bw_dl()? So something > > > > > like > > > > > > > > I think we can do with rcu_read_lock_sched() (see > > > > dl_task_can_attach()). > > > > > > RCU will keep dl_bw from being freed under us (we're implicitly in an > > > RCU > > > sched read section due to atomic context). It won't stop rq->rd from > > > changing, but that could have happened before we took rq->lock. If the > > > cpu > > > the task was running on was removed from the cpuset, and that raced with > > > the > > > task being moved to a different cpuset, couldn't we end up erroneously > > > subtracting from the cpu's new root domain (or failing to subtract at > > > all if > > > the old cpu's new cpuset happens to be the task's new cpuset)? I don't > > > see > > > anything that forces tasks off of the cpu when a cpu is removed from a > > > cpuset (though maybe I'm not looking in the right place), so the race > > > window > > > could be quite large. In any case, that's an existing problem that's > > > not > > > going to get solved in this patchset. > > > > OK. So, mainline has got cpuset_read_lock() which should be enough to > > guard against changes to rd(s). > > > > I agree that rq->lock is needed here. > > My point was that rq->lock isn't actually helping, because rq->rd could have > changed before rq->lock is acquired (and it's still the old rd that needs > the bandwidth subtraction). cpuset_mutex/cpuset_rwsem helps somewhat, > though there's still a problem due to the subtraction not happening until > the task is woken up (by which time cpuset_mutex may have been released and > further reconfiguration could have happened). This would be an issue even > without lazy migrate, since in that case ->set_cpus_allowed() can get > deferred, but this patch makes the window much bigger. > > The right solution is probably to explicitly track the rd for which a task > has a pending bandwidth subtraction (if any), and to continue doing it from > set_cpus_allowed() if the task is not migrate-disabled. In the meantime, I > think we should drop this patch from the patchset -- without it, lazy > migrate disable doesn't really make the race situation any worse than it > already was. I'm OK with dropping it for now (as we also have other possible issues as you point out below), but I really wonder what would be a solution here. Problem is that if a domain(s) reconfiguration happened while the task was migrate disabled, and we let the reconf destroy/rebuild domains, the old rd could be gone by the time the task gets migrate enabled again and the task could continue running, w/o its bandwidth been accounted for, in a new rd since the migrate enable instant, no? :-/ > BTW, what happens to the bw addition in dl_task_can_attach() if a subsequent > can_attach fails and the whole operation is cancelled? Oh, yeah, that doesn't look good. :-( Maybe we can use cancel_attach() to fix things up? Thanks, Juri
