On Fri, 5 Aug 2016 16:35:55 +0200
Sebastian Andrzej Siewior [off-list ref] wrote:

* Steven Rostedt | 2016-08-04 13:16:45 [-0400]:

quoted

diff --git a/include/linux/ftrace_irq.h b/include/linux/ftrace_irq.h
index dca7bf8cffe2..4ec2c9b205f2 100644
--- a/include/linux/ftrace_irq.h
+++ b/include/linux/ftrace_irq.h

@@ -3,11 +3,34 @@  

…

quoted

+static inline void ftrace_nmi_enter(void)
+{
+#ifdef CONFIG_HWLAT_TRACER
+	if (trace_hwlat_callback_enabled)
+		trace_hwlat_callback(true);  

so we take a tracepoint while we enter an nmi

It's not technically a tracepoint. I'm not sure tracepoints
(jumplabels) may be located this early in the NMI handler. This is
before some of the magic of having NMIs dealing with page faults and
break points.

quoted

--- a/kernel/trace/trace_hwlat.c
+++ b/kernel/trace/trace_hwlat.c

@@ -64,6 +64,15 @@ static struct dentry *hwlat_sample_window;	/* sample window us */

/* Save the previous tracing_thresh value */
static unsigned long save_tracing_thresh;

+/* NMI timestamp counters */
+static u64 nmi_ts_start;
+static u64 nmi_total_ts;
+static int nmi_count;
+static int nmi_cpu;  

and this is always limited to one CPU at a time?

Yes. Hence the "nmi_cpu".

…

quoted

@@ -125,6 +138,19 @@ static void trace_hwlat_sample(struct hwlat_sample *sample)

#define init_time(a, b)	(a = b)
#define time_u64(a)	a

+void trace_hwlat_callback(bool enter)
+{
+	if (smp_processor_id() != nmi_cpu)
+		return;
+
+	if (enter)
+		nmi_ts_start = time_get();  

but more interestingly: trace_clock_local() -> sched_clock()
and of kernel/time/sched_clock.c we do raw_read_seqcount(&cd.seq) which
means we are busted if the NMI triggers during update_clock_read_data().

Hmm, interesting. Because this is true for general tracing from an NMI.

/me looks at code.

Ah, this is when we have GENERIC_SCHED_CLOCK, which would break tracing
if any arch that has this also has NMIs. Probably need to look at arm64.

For x86, it has its own NMI safe sched_clock. I could make this "NMI"
code depend on:

#ifndef CONFIG_GENERIC_SCHED_CLOCK


-- Steve

quoted

+	else {
+		nmi_total_ts = time_get() - nmi_ts_start;
+		nmi_count++;
+	}
+}  

Sebastian