Hi Keir,
We have been seeing some hangs in call function and
tlb shootdown. In our case the machine was an Intel Tylersburg.
One failure mode is this early return in smp_call_function_interrupt:
if ( !cpu_isset(smp_processor_id(), call_data->selected) )
return;
Without a mb(), the target processor takes this path as it doesn't yet
see the update
to call_data->selected from the initiating processor. Then the initiator
waits for ever
to get an acknowledgment from this processor.
On the initiating side, I added the mb()'s to the send_IPI functions,
x2apic already
having one.
On the target side, I added the mb()'s in the individual target functions.
Regards,
Dave
diff -r 4129f0f2f2ba xen/arch/x86/smp.c
--- a/xen/arch/x86/smp.c Fri Oct 17 14:15:37 2008 +0100
+++ b/xen/arch/x86/smp.c Mon Oct 20 10:53:24 2008 -0400
@@ -93,6 +93,8 @@ void send_IPI_mask_flat(cpumask_t cpumas
/* An IPI with no target generates a send accept error from P5/P6 APICs. */
WARN_ON(mask == 0);
+ mb();
+
local_irq_save(flags);
/*
@@ -123,6 +125,8 @@ void send_IPI_mask_phys(cpumask_t mask,
{
unsigned long cfg, flags;
unsigned int query_cpu;
+
+ mb();
local_irq_save(flags);
@@ -160,6 +164,7 @@ static unsigned int flush_flags;
fastcall void smp_invalidate_interrupt(void)
{
+ mb();
ack_APIC_irq();
perfc_incr(ipis);
irq_enter();
@@ -337,17 +342,22 @@ void smp_send_stop(void)
fastcall void smp_event_check_interrupt(struct cpu_user_regs *regs)
{
+ mb();
ack_APIC_irq();
perfc_incr(ipis);
}
fastcall void smp_call_function_interrupt(struct cpu_user_regs *regs)
{
- void (*func)(void *info) = call_data->func;
- void *info = call_data->info;
-
+ void (*func)(void *info);
+ void *info;
+
+ mb();
ack_APIC_irq();
perfc_incr(ipis);
+
+ func = call_data->func;
+ info = call_data->info;
if ( !cpu_isset(smp_processor_id(), call_data->selected) )
return;
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home