On Mon, Jun 18, 2007 at 01:20:55AM -0700, Andrew Morton wrote: > On Mon, 18 Jun 2007 10:12:04 +0200 Ingo Molnar <[EMAIL PROTECTED]> wrote: > > > ----------------------------------------------------> > > Subject: [patch] x86: fix spin-loop starvation bug > > From: Ingo Molnar <[EMAIL PROTECTED]> > > > > Miklos Szeredi reported very long pauses (several seconds, sometimes > > more) on his T60 (with a Core2Duo) which he managed to track down to > > wait_task_inactive()'s open-coded busy-loop. He observed that an > > interrupt on one core tries to acquire the runqueue-lock but does not > > succeed in doing so for a very long time - while wait_task_inactive() on > > the other core loops waiting for the first core to deschedule a task > > (which it wont do while spinning in an interrupt handler). > > > > The problem is: both the spin_lock() code and the wait_task_inactive() > > loop uses cpu_relax()/rep_nop(), so in theory the CPU should have > > guaranteed MESI-fairness to the two cores - but that didnt happen: one > > of the cores was able to monopolize the cacheline that holds the > > runqueue lock, for extended periods of time. > > > > This patch changes the spin-loop to assert an atomic op after every REP > > NOP instance - this will cause the CPU to express its "MESI interest" in > > that cacheline after every REP NOP. > > Kiran, if you're still able to reproduce that zone->lru_lock starvation > problem, > this would be a good one to try...
We tried this approach a week back (speak of co-incidences), and it did not help the problem. I'd changed calls to the zone->lru_lock spin_lock to do spin_trylock in a while loop with cpu_relax instead. It did not help, This was on top of 2.6.17 kernels. But the good news is 2.6.21, as is does not have the starvation issue -- that is, zone->lru_lock does not seem to get contended that much under the same workload. However, this was not on the same hardware I reported zone->lru_lock contention on (8 socket dual core opteron). I don't have access to it anymore :( Thanks, Kiran - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/