x86: Warn when NMI handlers take large amounts of time
authorDave Hansen <dave.hansen@linux.intel.com>
Fri, 21 Jun 2013 15:51:35 +0000 (08:51 -0700)
committerIngo Molnar <mingo@kernel.org>
Sun, 23 Jun 2013 09:52:56 +0000 (11:52 +0200)
I have a system which is causing all kinds of problems.  It has
8 NUMA nodes, and lots of cores that can fight over cachelines.
If things are not working _perfectly_, then NMIs can take longer
than expected.

If we get too many of them backed up to each other, we can
easily end up in a situation where we are doing nothing *but*
running NMIs.  The biggest problem, though, is that this happens
_silently_.  You might be lucky to get an hrtimer warning, but
most of the time system simply hangs.

This patch should at least give us some warning before we fall
off the cliff.  the warnings look like this:

nmi_handle: perf_event_nmi_handler() took: 26095071 ns

The message is triggered whenever we notice the longest NMI
we've seen to date.  You can always view and reset this value
via the debugfs interface if you like.

Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: paulus@samba.org
Cc: acme@ghostprotocols.net
Cc: Dave Hansen <dave@sr71.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/kernel/nmi.c

index 60308053fdb2aeb2aaaa35112096e3b23a8dc2f8..e9bae4c2f2ddabf3c9458338791cd039b160878f 100644 (file)
@@ -14,6 +14,7 @@
 #include <linux/kprobes.h>
 #include <linux/kdebug.h>
 #include <linux/nmi.h>
+#include <linux/debugfs.h>
 #include <linux/delay.h>
 #include <linux/hardirq.h>
 #include <linux/slab.h>
@@ -82,6 +83,15 @@ __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 
 #define nmi_to_desc(type) (&nmi_desc[type])
 
+static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
+static int __init nmi_warning_debugfs(void)
+{
+       debugfs_create_u64("nmi_longest_ns", 0644,
+                       arch_debugfs_dir, &nmi_longest_ns);
+       return 0;
+}
+fs_initcall(nmi_warning_debugfs);
+
 static int __kprobes nmi_handle(unsigned int type, struct pt_regs *regs, bool b2b)
 {
        struct nmi_desc *desc = nmi_to_desc(type);
@@ -96,8 +106,25 @@ static int __kprobes nmi_handle(unsigned int type, struct pt_regs *regs, bool b2
         * can be latched at any given time.  Walk the whole list
         * to handle those situations.
         */
-       list_for_each_entry_rcu(a, &desc->head, list)
+       list_for_each_entry_rcu(a, &desc->head, list) {
+               u64 before, delta, whole_msecs;
+               int decimal_msecs;
+
+               before = local_clock();
                handled += a->handler(type, regs);
+               delta = local_clock() - before;
+
+               if (delta < nmi_longest_ns)
+                       continue;
+
+               nmi_longest_ns = delta;
+               whole_msecs = do_div(delta, (1000 * 1000));
+               decimal_msecs = do_div(delta, 1000) % 1000;
+               printk_ratelimited(KERN_INFO
+                       "INFO: NMI handler (%ps) took too long to run: "
+                       "%lld.%03d msecs\n", a->handler, whole_msecs,
+                       decimal_msecs);
+       }
 
        rcu_read_unlock();