Messages in this thread | | | Date | Thu, 14 Jun 2012 17:02:11 +0200 (CEST) | From | Thomas Gleixner <> | Subject | Re: [RFC patch 2/5] smpboot: Provide infrastructure for percpu hotplug threads |
| |
On Thu, 14 Jun 2012, Paul E. McKenney wrote: > On Thu, Jun 14, 2012 at 03:32:19PM +0200, Thomas Gleixner wrote: > > On Thu, 14 Jun 2012, Paul E. McKenney wrote: > > > On Thu, Jun 14, 2012 at 01:20:39PM +0200, Thomas Gleixner wrote: > > > > I gave it a quick shot, but I was not able to reproduce the hang yet. > > > > > > Really? I have a strictly Western-Hemisphere bug? ;-) > > > > I guess I need to fire up rcu torture to make it surface. > > A simple offline was triggering it for me. Perhaps some of my debug > code was inappropriate, will retry. > > > > > But looking at the thread function made me look into rcu_yield() and I > > > > really wonder what kind of drug induced that particular piece of > > > > horror. > > > > > > When you are working on something like RCU priority boosting, no other > > > drug is in any way necessary. ;-) > > > > And how do we protect minors from that ? > > We rely on their own sense of self-preservation preventing them from > getting involved in such insanity. > > > > > I can't figure out why this yield business is necessary at all. The > > > > commit logs are as helpful as the missing code comments :) > > > > > > > > I suspect that it's some starvation issue. But if we need it, then > > > > can't we replace it with something sane like the (untested) patch > > > > below? > > > > > > Yep, starvation. I will take a look at your approach after I wake > > > up a bit more. > > > > Btw, if that simpler yield approach is working and I can't see why it > > shouldn't then you can get rid of the node task as well. The only > > purpose of it is to push up the priority of yielding tasks, right? > > It also boosts the priority of preempted RCU read-side critical sections.
Though the only way how this thread is invoked is via the timeout of that yield timer. So I really have a hard time for understanding that.
cpu_kthread() .... yield() timer fires -> mark cpu in mask and wakeup node kthread
node_kthread() do magic boost invocation push priority of cpu_kthread marked in mask
For the boost thread it looks like:
boost_kthread() ..... yield() timer fires -> wakeup node kthread
node_kthread() do magic boost invocation, but no prio adjustment of boost thread.
/me scratches head
| |