[PATCH] tracing/osnoise: Add option to align tlat threads

[Tomas Glozar]

Add an option called TIMERLAT_ALIGN to osnoise/options, together with a
corresponding setting osnoise/timerlat_align_us.

This option sets the alignment of wakeup times between different
timerlat threads, similarly to cyclictest's -A/--aligned option. If
TIMERLAT_ALIGN is set, the first thread that reaches the first cycle
records its first wake-up time. Each following thread sets its first
wake-up time to a fixed offset from the recorded time, and incremenets
it by the same offset.

Example:

osnoise/timerlat_period is set to 1000, osnoise/timerlat_align_us is
set to 50. There are four threads, on CPUs 1 to 4.

- CPU 4 enters first cycle first. The current time is 20000us, so
the wake-up of the first cycle is set to 21000us. This time is recorded.
- CPU 2 enter first cycle next. It reads the recorded time, increments
it to 21020us, and uses this value as its own wake-up time for the first
cycle.
- CPU 3 enters first cycle next. It reads the recorded time, increments
it to 21040 us, and uses the value as its own wake-up time.
- CPU 1 proceeds analogically.

In each next cycle, the wake-up time (called "absolute period" in
timerlat code) is incremented by the (relative) period of 1000us. Thus,
the wake-ups in the following cycles (provided the times are reached and
not in the past) will be as follows:

CPU 1		CPU 2		CPU 3	 	CPU 4
21080us		21020us		21040us		21000us
22080us		22020us		22040us		22000us
...		...		...		...

Even if any cycle is skipped due to e.g. the first cycle calculation
happening later, the alignment stays in place.

Signed-off-by: Tomas Glozar <tglozar@xxxxxxxxxx>
---

I tested this option using the following command:

$ bpftrace -e 'tracepoint:osnoise:timerlat_sample /!@time[cpu]/ {
      if (!@begin) { @begin = nsecs; }
      @time[cpu] = ((nsecs - @begin) / 1000) % 1000;
  }
 END { clear(@begin); }' -c 'rtla timerlat hist -d 1s -c 1-10'

This captures the alignment of first timerlat sample (which is +-
equivalent to the wake-up time).

With timerlat_align_us = 20:

@time[1]: 2
@time[2]: 18
@time[3]: 38
@time[4]: 57
@time[5]: 83
@time[6]: 103
@time[7]: 123
@time[8]: 143
@time[9]: 162
@time[10]: 182

With timerlat_align_us = 0

@time[1]: 1
@time[5]: 4
@time[7]: 4
@time[6]: 4
@time[8]: 4
@time[9]: 4
@time[10]: 4
@time[4]: 5
@time[3]: 5
@time[2]: 5

Only thing I am not too sure about is the absense of barriers. I feel
like I only touch that one atomic variable concurrently, so it should
be fine (unlike e.g. a mutex protecting another variable, where you need
acquire-release semantics) with relaxed variants of atomic functions;
but I don't have any other experience with barriers so far.

 kernel/trace/trace_osnoise.c | 34 +++++++++++++++++++++++++++++++++-
 1 file changed, 33 insertions(+), 1 deletion(-)

diff --git a/kernel/trace/trace_osnoise.c b/kernel/trace/trace_osnoise.c
index dee610e465b9..df1d4529d226 100644
--- a/kernel/trace/trace_osnoise.c
+++ b/kernel/trace/trace_osnoise.c
@@ -58,6 +58,7 @@ enum osnoise_options_index {
 	OSN_PANIC_ON_STOP,
 	OSN_PREEMPT_DISABLE,
 	OSN_IRQ_DISABLE,
+	OSN_TIMERLAT_ALIGN,
 	OSN_MAX
 };
 
@@ -66,7 +67,8 @@ static const char * const osnoise_options_str[OSN_MAX] = {
 							"OSNOISE_WORKLOAD",
 							"PANIC_ON_STOP",
 							"OSNOISE_PREEMPT_DISABLE",
-							"OSNOISE_IRQ_DISABLE" };
+							"OSNOISE_IRQ_DISABLE",
+							"TIMERLAT_ALIGN" };
 
 #define OSN_DEFAULT_OPTIONS		0x2
 static unsigned long osnoise_options	= OSN_DEFAULT_OPTIONS;
@@ -326,6 +328,7 @@ static struct osnoise_data {
 	u64	stop_tracing_total;	/* stop trace in the final operation (report/thread) */
 #ifdef CONFIG_TIMERLAT_TRACER
 	u64	timerlat_period;	/* timerlat period */
+	u64	timerlat_align_us;	/* timerlat alignment */
 	u64	print_stack;		/* print IRQ stack if total > */
 	int	timerlat_tracer;	/* timerlat tracer */
 #endif
@@ -338,6 +341,7 @@ static struct osnoise_data {
 #ifdef CONFIG_TIMERLAT_TRACER
 	.print_stack			= 0,
 	.timerlat_period		= DEFAULT_TIMERLAT_PERIOD,
+	.timerlat_align_us		= 0,
 	.timerlat_tracer		= 0,
 #endif
 };
@@ -1820,6 +1824,7 @@ static int wait_next_period(struct timerlat_variables *tlat)
 {
 	ktime_t next_abs_period, now;
 	u64 rel_period = osnoise_data.timerlat_period * 1000;
+	static atomic64_t align_next;
 
 	now = hrtimer_cb_get_time(&tlat->timer);
 	next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
@@ -1829,6 +1834,17 @@ static int wait_next_period(struct timerlat_variables *tlat)
 	 */
 	tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
 
+	if (test_bit(OSN_TIMERLAT_ALIGN, &osnoise_options) && !tlat->count
+	    && atomic64_cmpxchg_relaxed(&align_next, 0, tlat->abs_period)) {
+		/*
+		 * Align thread in first cycle on each CPU to the set alignment.
+		 */
+		tlat->abs_period = atomic64_fetch_add_relaxed(osnoise_data.timerlat_align_us * 1000,
+			&align_next);
+		tlat->abs_period += osnoise_data.timerlat_align_us * 1000;
+		next_abs_period = ns_to_ktime(tlat->abs_period);
+	}
+
 	/*
 	 * If the new abs_period is in the past, skip the activation.
 	 */
@@ -2650,6 +2666,17 @@ static struct trace_min_max_param timerlat_period = {
 	.min	= &timerlat_min_period,
 };
 
+/*
+ * osnoise/timerlat_align_us: align the first wakeup of all timerlat
+ * threads to a common boundary (in us). 0 means disabled.
+ */
+static struct trace_min_max_param timerlat_align_us = {
+	.lock	= &interface_lock,
+	.val	= &osnoise_data.timerlat_align_us,
+	.max	= NULL,
+	.min	= NULL,
+};
+
 static const struct file_operations timerlat_fd_fops = {
 	.open		= timerlat_fd_open,
 	.read		= timerlat_fd_read,
@@ -2746,6 +2773,11 @@ static int init_timerlat_tracefs(struct dentry *top_dir)
 	if (!tmp)
 		return -ENOMEM;
 
+	tmp = tracefs_create_file("timerlat_align_us", TRACE_MODE_WRITE, top_dir,
+				  &timerlat_align_us, &trace_min_max_fops);
+	if (!tmp)
+		return -ENOMEM;
+
 	retval = osnoise_create_cpu_timerlat_fd(top_dir);
 	if (retval)
 		return retval;
-- 
2.53.0