Ignore stolen time in the softlockup watchdog
authorJeremy Fitzhardinge <jeremy@goop.org>
Tue, 8 May 2007 07:28:02 +0000 (00:28 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 8 May 2007 18:15:06 +0000 (11:15 -0700)
The softlockup watchdog is currently a nuisance in a virtual machine, since
the whole system could have the CPU stolen from it for a long period of
time.  While it would be unlikely for a guest domain to be denied timer
interrupts for over 10s, it could happen and any softlockup message would
be completely spurious.

Earlier I proposed that sched_clock() return time in unstolen nanoseconds,
which is how Xen and VMI currently implement it.  If the softlockup
watchdog uses sched_clock() to measure time, it would automatically ignore
stolen time, and therefore only report when the guest itself locked up.
When running native, sched_clock() returns real-time nanoseconds, so the
behaviour would be unchanged.

Note that sched_clock() used this way is inherently per-cpu, so this patch
makes sure that the per-processor watchdog thread initialized its own
timestamp.

Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: john stultz <johnstul@us.ibm.com>
Cc: Zachary Amsden <zach@vmware.com>
Cc: James Morris <jmorris@namei.org>
Cc: Dan Hecht <dhecht@vmware.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Cc: Chris Lalancette <clalance@redhat.com>
Cc: Rick Lindsley <ricklind@us.ibm.com>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kernel/softlockup.c

index 81d43caa2012feed6eff4a95f3cd6b1369483fe5..5ea631742dbcf246e7dbd772e33c798c74ee6625 100644 (file)
@@ -34,9 +34,19 @@ static struct notifier_block panic_block = {
        .notifier_call = softlock_panic,
 };
 
+/*
+ * Returns seconds, approximately.  We don't need nanosecond
+ * resolution, and we don't need to waste time with a big divide when
+ * 2^30ns == 1.074s.
+ */
+static unsigned long get_timestamp(void)
+{
+       return sched_clock() >> 30;  /* 2^30 ~= 10^9 */
+}
+
 void touch_softlockup_watchdog(void)
 {
-       __raw_get_cpu_var(touch_timestamp) = jiffies;
+       __raw_get_cpu_var(touch_timestamp) = get_timestamp();
 }
 EXPORT_SYMBOL(touch_softlockup_watchdog);
 
@@ -48,9 +58,17 @@ void softlockup_tick(void)
 {
        int this_cpu = smp_processor_id();
        unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
+       unsigned long print_timestamp;
+       unsigned long now;
 
-       /* prevent double reports: */
-       if (per_cpu(print_timestamp, this_cpu) == touch_timestamp ||
+       /* watchdog task hasn't updated timestamp yet */
+       if (touch_timestamp == 0)
+               return;
+
+       print_timestamp = per_cpu(print_timestamp, this_cpu);
+
+       /* report at most once a second */
+       if (print_timestamp < (touch_timestamp + 1) ||
                did_panic ||
                        !per_cpu(watchdog_task, this_cpu))
                return;
@@ -61,12 +79,14 @@ void softlockup_tick(void)
                return;
        }
 
+       now = get_timestamp();
+
        /* Wake up the high-prio watchdog task every second: */
-       if (time_after(jiffies, touch_timestamp + HZ))
+       if (now > (touch_timestamp + 1))
                wake_up_process(per_cpu(watchdog_task, this_cpu));
 
        /* Warn about unreasonable 10+ seconds delays: */
-       if (time_after(jiffies, touch_timestamp + 10*HZ)) {
+       if (now > (touch_timestamp + 10)) {
                per_cpu(print_timestamp, this_cpu) = touch_timestamp;
 
                spin_lock(&print_lock);
@@ -87,6 +107,9 @@ static int watchdog(void * __bind_cpu)
        sched_setscheduler(current, SCHED_FIFO, &param);
        current->flags |= PF_NOFREEZE;
 
+       /* initialize timestamp */
+       touch_softlockup_watchdog();
+
        /*
         * Run briefly once per second to reset the softlockup timestamp.
         * If this gets delayed for more than 10 seconds then the
@@ -118,7 +141,7 @@ cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
                        printk("watchdog for %i failed\n", hotcpu);
                        return NOTIFY_BAD;
                }
-               per_cpu(touch_timestamp, hotcpu) = jiffies;
+               per_cpu(touch_timestamp, hotcpu) = 0;
                per_cpu(watchdog_task, hotcpu) = p;
                kthread_bind(p, hotcpu);
                break;