(no subject)

[Daniel Bristot de Oliveira]

Subject: [PATCH 0/2] osnoise improvements

These two patches are improvements for the osnoise/timerlat tracers,
mainly to support rtla. The first one is a dependency for the -C <cgroup>
command line option to be added to rtla. It is essential to allow
timerlat/osnoise to measure the latencies from the container/cgroup
point of view. The second improves rtla time for saving tracing after
a stop tracing condition on osnoise/hwnoise.

Daniel Bristot de Oliveira (2):
  tracing/osnoise: Switch from PF_NO_SETAFFINITY to migrate_disable
  tracing/osnoise: Skip running osnoise if all instances are off

 kernel/trace/trace_osnoise.c | 99 ++++++++++++++++++++++++++++++++++--
 1 file changed, 96 insertions(+), 3 deletions(-)

-- 
2.38.1

[PATCH 1/2] tracing/osnoise: Switch from PF_NO_SETAFFINITY to migrate_disable

[Daniel Bristot de Oliveira]

Currently, osnoise/timerlat threads run with PF_NO_SETAFFINITY set.
It works well, however, cgroups do not allow PF_NO_SETAFFINITY threads
to be accepted, and this creates a limitation to osnoise/timerlat.

To avoid this limitation, disable migration of the threads as soon
as they start to run, and then clean the PF_NO_SETAFFINITY flag (still)
used during thread creation.

If for some reason a thread migration is requested, e.g., via
sched_settafinity, the tracer thread will notice and exit.

Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Daniel Bristot de Oliveira <bristot@kernel.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: Daniel Bristot de Oliveira <bristot@kernel.org>
---
 kernel/trace/trace_osnoise.c | 68 ++++++++++++++++++++++++++++++++++++
 1 file changed, 68 insertions(+)

diff --git a/kernel/trace/trace_osnoise.c b/kernel/trace/trace_osnoise.c
index e97e3fa5cbed..c265ec5f1726 100644
--- a/kernel/trace/trace_osnoise.c
+++ b/kernel/trace/trace_osnoise.c
@@ -1545,6 +1545,39 @@ static void osnoise_sleep(void)
 	}
 }
 
+/*
+ * osnoise_migration_pending - checks if the task needs to migrate
+ *
+ * osnoise/timerlat threads are per-cpu. If there is a pending request to
+ * migrate the thread away from the current CPU, something bad has happened.
+ * Play the good citizen and leave.
+ *
+ * Returns 0 if it is safe to continue, 1 otherwise.
+ */
+static inline int osnoise_migration_pending(void)
+{
+	if (!current->migration_pending)
+		return 0;
+
+	/*
+	 * If migration is pending, there is a task waiting for the
+	 * tracer to enable migration. The tracer does not allow migration,
+	 * thus: taint and leave to unblock the blocked thread.
+	 */
+	osnoise_taint("migration requested to osnoise threads, leaving.");
+
+	/*
+	 * Unset this thread from the threads managed by the interface.
+	 * The tracers are responsible for cleaning their env before
+	 * exiting.
+	 */
+	mutex_lock(&interface_lock);
+	this_cpu_osn_var()->kthread = NULL;
+	mutex_unlock(&interface_lock);
+
+	return 1;
+}
+
 /*
  * osnoise_main - The osnoise detection kernel thread
  *
@@ -1553,12 +1586,29 @@ static void osnoise_sleep(void)
  */
 static int osnoise_main(void *data)
 {
+	unsigned long flags;
+
+	/*
+	 * This thread was created pinned to the CPU using PF_NO_SETAFFINITY.
+	 * The problem is that cgroup does not allow PF_NO_SETAFFINITY thread.
+	 *
+	 * To work around this limitation, disable migration and remove the
+	 * flag.
+	 */
+	migrate_disable();
+	raw_spin_lock_irqsave(&current->pi_lock, flags);
+	current->flags &= ~(PF_NO_SETAFFINITY);
+	raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 
 	while (!kthread_should_stop()) {
+		if (osnoise_migration_pending())
+			break;
+
 		run_osnoise();
 		osnoise_sleep();
 	}
 
+	migrate_enable();
 	return 0;
 }
 
@@ -1706,6 +1756,7 @@ static int timerlat_main(void *data)
 	struct timerlat_variables *tlat = this_cpu_tmr_var();
 	struct timerlat_sample s;
 	struct sched_param sp;
+	unsigned long flags;
 	u64 now, diff;
 
 	/*
@@ -1714,6 +1765,18 @@ static int timerlat_main(void *data)
 	sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
 	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 
+	/*
+	 * This thread was created pinned to the CPU using PF_NO_SETAFFINITY.
+	 * The problem is that cgroup does not allow PF_NO_SETAFFINITY thread.
+	 *
+	 * To work around this limitation, disable migration and remove the
+	 * flag.
+	 */
+	migrate_disable();
+	raw_spin_lock_irqsave(&current->pi_lock, flags);
+	current->flags &= ~(PF_NO_SETAFFINITY);
+	raw_spin_unlock_irqrestore(&current->pi_lock, flags);
+
 	tlat->count = 0;
 	tlat->tracing_thread = false;
 
@@ -1731,6 +1794,7 @@ static int timerlat_main(void *data)
 	osn_var->sampling = 1;
 
 	while (!kthread_should_stop()) {
+
 		now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
 		diff = now - tlat->abs_period;
 
@@ -1749,10 +1813,14 @@ static int timerlat_main(void *data)
 			if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
 				osnoise_stop_tracing();
 
+		if (osnoise_migration_pending())
+			break;
+
 		wait_next_period(tlat);
 	}
 
 	hrtimer_cancel(&tlat->timer);
+	migrate_enable();
 	return 0;
 }
 #else /* CONFIG_TIMERLAT_TRACER */
-- 
2.38.1

[PATCH 2/2] tracing/osnoise: Skip running osnoise if all instances are off

[Daniel Bristot de Oliveira]

In the case of all tracing instances being off, sleep for the entire period.

 Q: Why not kill all threads so?
 A: It is valid and useful to start the threads with tracing off.
For example, rtla disables tracing, starts the tracer, applies the
scheduling setup to the threads, e.g., sched priority and cgroup,
and then begin tracing with all set.

Skipping the period helps to speed up rtla setup and save the
trace after a stop tracing.

Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Daniel Bristot de Oliveira <bristot@kernel.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: Daniel Bristot de Oliveira <bristot@kernel.org>
---
 kernel/trace/trace_osnoise.c | 31 ++++++++++++++++++++++++++++---
 1 file changed, 28 insertions(+), 3 deletions(-)

diff --git a/kernel/trace/trace_osnoise.c b/kernel/trace/trace_osnoise.c
index c265ec5f1726..220172cb874d 100644
--- a/kernel/trace/trace_osnoise.c
+++ b/kernel/trace/trace_osnoise.c
@@ -1285,6 +1285,22 @@ static __always_inline void osnoise_stop_tracing(void)
 	rcu_read_unlock();
 }
 
+/*
+ * osnoise_has_tracing_on - Check if there is at least one instance on
+ */
+static __always_inline int osnoise_has_tracing_on(void)
+{
+	struct osnoise_instance *inst;
+	int trace_is_on = 0;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(inst, &osnoise_instances, list)
+		trace_is_on += tracer_tracing_is_on(inst->tr);
+	rcu_read_unlock();
+
+	return trace_is_on;
+}
+
 /*
  * notify_new_max_latency - Notify a new max latency via fsnotify interface.
  */
@@ -1517,13 +1533,16 @@ static struct cpumask save_cpumask;
 /*
  * osnoise_sleep - sleep until the next period
  */
-static void osnoise_sleep(void)
+static void osnoise_sleep(bool skip_period)
 {
 	u64 interval;
 	ktime_t wake_time;
 
 	mutex_lock(&interface_lock);
-	interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
+	if (skip_period)
+		interval = osnoise_data.sample_period;
+	else
+		interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
 	mutex_unlock(&interface_lock);
 
 	/*
@@ -1604,8 +1623,14 @@ static int osnoise_main(void *data)
 		if (osnoise_migration_pending())
 			break;
 
+		/* skip a period if tracing is off on all instances */
+		if (!osnoise_has_tracing_on()) {
+			osnoise_sleep(true);
+			continue;
+		}
+
 		run_osnoise();
-		osnoise_sleep();
+		osnoise_sleep(false);
 	}
 
 	migrate_enable();
-- 
2.38.1

Subject: [PATCH 0/2] osnoise improvements

[Daniel Bristot de Oliveira]

Oops....

On 5/11/23 18:45, Daniel Bristot de Oliveira wrote:
> Subject: [PATCH 0/2] osnoise improvements