Messages in this thread Patch in this message | | | Date | Sun, 4 Jan 2009 20:30:06 -0800 | From | "Paul E. McKenney" <> | Subject | [PATCH] Fix rcutree grace-period-latency bug on small systems |
| |
Kudos to Andi Kleen for finding a grace-period-latency problem! The problem was that the special-case code for small machines never updated the ->signaled field to indicate that grace-period initialization had completed, which prevented force_quiescent_state() from ever expediting grace periods. This problem resulted in grace periods extending for more than 20 seconds. Not subtle. I introduced this bug during my inspection process when I fixed a race between grace-period initialization and force_quiescent_state() execution.
The following patch properly updates the ->signaled field for the "small"-system case (no more than 32 CPUs for 32-bit kernels and no more than 64 CPUs for 64-bit kernels).
I believe that this patch should be included for 2.6.29.
Located-by: Andi Kleen <andi@firstfloor.org> Tested-by: Andi Kleen <andi@firstfloor.org> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> ---
rcutree.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/kernel/rcutree.c b/kernel/rcutree.c index a342b03..88d921c 100644 --- a/kernel/rcutree.c +++ b/kernel/rcutree.c @@ -572,6 +572,7 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags) /* Special-case the common single-level case. */ if (NUM_RCU_NODES == 1) { rnp->qsmask = rnp->qsmaskinit; + rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */ spin_unlock_irqrestore(&rnp->lock, flags); return; }
| |