The SIE instruction is interruptible, so instead of having a guest
exit on a host interrupt we basically return to guest mode.
We have some logic in the interrupt handler to check for
need_resched, machine checks or sigpending to exit SIE the hard
way, but RCU is currently not handled, leading to several second
delays on cpu bound guests.
Lets mark SIE (guest context) as quiescing state in the external
interrupt handler (hz tick, timers sigp and others) thus making
RCU working properly again.
Long term we might want to use proper state tracking (just like
the dynticks folks) and mark guest state similar to user space
as an extended grace period, but this is not ready yet.
Signed-off-by: Christian Borntraeger <borntraeger@xxxxxxxxxx>
Cc: Cornelia Huck <cornelia.huck@xxxxxxxxxx>
Cc: Dipankar Sarma <dipankar@xxxxxxxxxx>
Cc: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
Cc: Martin Schwidefsky <schwidefsky@xxxxxxxxxx>
Cc: Heiko Carstens <heiko.carstens@xxxxxxxxxx>
Cc: Gleb Natapov <gleb@xxxxxxxxxx>
Cc: Marcelo Tosatti <mtosatti@xxxxxxxxxx>
---
arch/s390/kernel/irq.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/arch/s390/kernel/irq.c b/arch/s390/kernel/irq.c
index 1630f43..d6ccb1d 100644
--- a/arch/s390/kernel/irq.c
+++ b/arch/s390/kernel/irq.c
@@ -244,6 +244,17 @@ void __irq_entry do_extint(struct pt_regs *regs, struct ext_code ext_code,
int index;
old_regs = set_irq_regs(regs);
+ /*
+ * The SIE instruction is interruptible, so instead of having a guest
+ * exit on a host interrupt we basically return to guest mode if there
+ * is no need_resched, machine check or signal pending. So we can
+ * stay in guest mode for several seconds or even minutes. This
+ * lets RCU wait for a grace period much too long. In case of PF_VCPU
+ * we know that we do not hold any rcu data, so lets claim that a
+ * context switch happened, which is a quiescing state.
+ */
+ if (current->flags & PF_VCPU)
+ rcu_sched_qs(smp_processor_id());
irq_enter();
if (S390_lowcore.int_clock >= S390_lowcore.clock_comparator) {
/* Serve timer interrupts first. */
--
1.8.1.4
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
