Messages in this thread | | | Date | Mon, 12 Aug 2013 09:44:36 -0700 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH 5/8] rcu: eliminate deadlock for rcu read site |
| |
On Mon, Aug 12, 2013 at 06:21:26PM +0200, Peter Zijlstra wrote: > On Mon, Aug 12, 2013 at 08:16:18AM -0700, Paul E. McKenney wrote: > > On Mon, Aug 12, 2013 at 03:55:44PM +0200, Peter Zijlstra wrote: > > > On Fri, Aug 09, 2013 at 05:31:27PM +0800, Lai Jiangshan wrote: > > > > On 08/09/2013 04:40 AM, Paul E. McKenney wrote: > > > > > One problem here -- it may take quite some time for a set_need_resched() > > > > > to take effect. This is especially a problem for RCU priority boosting, > > > > > but can also needlessly delay preemptible-RCU grace periods because > > > > > local_irq_restore() and friends don't check the TIF_NEED_RESCHED bit. > > > > > > > > > > > > The final effect of deboosting(rt_mutex_unlock()) is also accomplished > > > > via set_need_resched()/set_tsk_need_resched(). > > > > set_need_resched() is enough for RCU priority boosting issue here. > > > > > > But there's a huge difference between the boosting and deboosting side > > > of things. rcu_read_unlock_special() starts the boost, the deboosting > > > only matters if/when you reschedule. > > > > Or if there is a pre-existing runnable task whose priority is such that > > deboosting makes it the highest-priority task. > > Right, I got horribly lost in rt_mutex, but I suspect we deal with that > case the right way. -rt people would've noticed us screwing that up ;-) > > But there too, we're fully limited to how fast we can get a > reschedule(). Deboosting sooner than we can reschedule to run the other > task is effectively pointless. The converse is obviously not true; we > must not be able to reschedule sooner than we can deboost ;-)
In addition, the proposed change was to defer the deboost based on a set_need_resched(), which would provide additional opportunity for delay -- the running task would retain its high priority until the scheduler acted on the set_need_resched().
Thanx, Paul
| |