[PATCH v6 0/4] Some new features for the preempt/irqsoff tracers

[PATCH v6 0/4] Some new features for the preempt/irqsoff tracers

[Viktor Rosendahl]

Hello all,

Changes in v5:
- [PATCH 1/4]:
  * Using the irq_work mechanism instead of the games with linked lists, percpu
    variables and modifications to the scheduling and idle code.

- [PATCH 2/4]:
  * Updated the copyright message, adding my own name and what I did.

- [PATCH 3/4]:
  * Added my own name as author to the copyright message.

- [PACTH 4/4]:
  * Improved the build on sh architecture with gcc 7.4.0 by including
    linux/types.h. The build on sh architecture still fails for me later in the
    build process with an ICE error from the compiler but the same failure also
    happens without any of my patches.

This series is meant to address two issues with the latency tracing.

The first three patches provide a method to trace latencies that always
occurs very close to each other and to differentiate between them, in spite
of the fact that the latency tracers work in overwrite mode.

[PATCH 1/4] This implement fs notification for tracing_max_latency. It
makes it possible for userspace to detect when a new latency has been
detected.

[PATCH 2/4] This extends the preemptirq_delay_test module so that it can be
used to generate a burst of closely occurring latencies.

[PATCH 3/4] This adds a user space program to the tools directory that
utilizes the fs notification feature and a randomized algorithm to print out
any of the latencies in a burst with approximately equal probability.

The last patch is not directly connected but earlier it didn't apply
cleanly on its own. However, now it does, so in principle it could be
applied separately from the others.

[PATCH 4/4] This adds the option console-latency to the trace options. This
makes it possible to enable tracing of console latencies.

best regards,

Viktor

Viktor Rosendahl (4):
  ftrace: Implement fs notification for tracing_max_latency
  preemptirq_delay_test: Add the burst feature and a sysfs trigger
  Add the latency-collector to tools
  ftrace: Add an option for tracing console latencies

 include/linux/irqflags.h             |   22 +
 kernel/printk/printk.c               |    6 +-
 kernel/trace/Kconfig                 |    6 +-
 kernel/trace/preemptirq_delay_test.c |  147 +++-
 kernel/trace/trace.c                 |   75 +-
 kernel/trace/trace.h                 |   19 +
 kernel/trace/trace_hwlat.c           |    4 +-
 kernel/trace/trace_irqsoff.c         |   12 +
 tools/Makefile                       |   14 +-
 tools/trace/Makefile                 |   20 +
 tools/trace/latency-collector.c      | 1212 ++++++++++++++++++++++++++
 11 files changed, 1504 insertions(+), 33 deletions(-)
 create mode 100644 tools/trace/Makefile
 create mode 100644 tools/trace/latency-collector.c

-- 
2.17.1

[PATCH v6 1/4] ftrace: Implement fs notification for tracing_max_latency

[Viktor Rosendahl]

This patch implements the feature that the tracing_max_latency file,
e.g. /sys/kernel/debug/tracing/tracing_max_latency will receive
notifications through the fsnotify framework when a new latency is
available.

One particularly interesting use of this facility is when enabling
threshold tracing, through /sys/kernel/debug/tracing/tracing_thresh,
together with the preempt/irqsoff tracers. This makes it possible to
implement a user space program that can, with equal probability,
obtain traces of latencies that occur immediately after each other in
spite of the fact that the preempt/irqsoff tracers operate in overwrite
mode.

This facility works with the hwlat, preempt/irqsoff, and wakeup
tracers.

The tracers may call the latency_fsnotify() from places such as
__schedule() or do_idle(); this makes it impossible to call
queue_work() directly without risking a deadlock. The same would
happen with a softirq,  kernel thread or tasklet. For this reason we
use the irq_work mechanism to call queue_work().

Signed-off-by: Viktor Rosendahl <viktor.rosendahl@gmail.com>
---
 kernel/trace/trace.c       | 75 +++++++++++++++++++++++++++++++++++++-
 kernel/trace/trace.h       | 18 +++++++++
 kernel/trace/trace_hwlat.c |  4 +-
 3 files changed, 94 insertions(+), 3 deletions(-)

diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index 563e80f9006a..72ac20c4aaa1 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -44,6 +44,9 @@
 #include <linux/trace.h>
 #include <linux/sched/clock.h>
 #include <linux/sched/rt.h>
+#include <linux/fsnotify.h>
+#include <linux/irq_work.h>
+#include <linux/workqueue.h>
 
 #include "trace.h"
 #include "trace_output.h"
@@ -1480,6 +1483,74 @@ static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt)
 
 unsigned long __read_mostly	tracing_thresh;
 
+#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
+	defined(CONFIG_FSNOTIFY)
+
+static const struct file_operations tracing_max_lat_fops;
+static struct workqueue_struct *fsnotify_wq;
+
+static void latency_fsnotify_workfn(struct work_struct *work)
+{
+	struct trace_array *tr = container_of(work, struct trace_array,
+					      fsnotify_work);
+	fsnotify(tr->d_max_latency->d_inode, FS_MODIFY,
+		 tr->d_max_latency->d_inode, FSNOTIFY_EVENT_INODE, NULL, 0);
+}
+
+static void latency_fsnotify_workfn_irq(struct irq_work *iwork)
+{
+	struct trace_array *tr = container_of(iwork, struct trace_array,
+					      fsnotify_irqwork);
+	queue_work(fsnotify_wq, &tr->fsnotify_work);
+}
+
+static void trace_create_maxlat_file(struct trace_array *tr,
+				     struct dentry *d_tracer)
+{
+	INIT_WORK(&tr->fsnotify_work, latency_fsnotify_workfn);
+	init_irq_work(&tr->fsnotify_irqwork, latency_fsnotify_workfn_irq);
+	tr->d_max_latency = trace_create_file("tracing_max_latency", 0644,
+					      d_tracer, &tr->max_latency,
+					      &tracing_max_lat_fops);
+}
+
+__init static int latency_fsnotify_init(void)
+{
+	fsnotify_wq = alloc_workqueue("tr_max_lat_wq",
+				      WQ_UNBOUND | WQ_HIGHPRI, 0);
+	if (!fsnotify_wq) {
+		pr_err("Unable to allocate tr_max_lat_wq\n");
+		return -ENOMEM;
+	}
+	return 0;
+}
+
+late_initcall_sync(latency_fsnotify_init);
+
+void latency_fsnotify(struct trace_array *tr)
+{
+	if (!fsnotify_wq)
+		return;
+	/*
+	 * We cannot call queue_work(&tr->fsnotify_work) from here because it's
+	 * possible that we are called from __schedule() or do_idle(), which
+	 * could cause a deadlock.
+	 */
+	irq_work_queue(&tr->fsnotify_irqwork);
+}
+
+/*
+ * (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
+ *  defined(CONFIG_FSNOTIFY)
+ */
+#else
+
+#define trace_create_maxlat_file(tr, d_tracer)				\
+	trace_create_file("tracing_max_latency", 0644, d_tracer,	\
+			  &tr->max_latency, &tracing_max_lat_fops)
+
+#endif
+
 #ifdef CONFIG_TRACER_MAX_TRACE
 /*
  * Copy the new maximum trace into the separate maximum-trace
@@ -1518,6 +1589,7 @@ __update_max_tr(struct trace_array *tr, struct task_struct *tsk, int cpu)
 
 	/* record this tasks comm */
 	tracing_record_cmdline(tsk);
+	latency_fsnotify(tr);
 }
 
 /**
@@ -8550,8 +8622,7 @@ init_tracer_tracefs(struct trace_array *tr, struct dentry *d_tracer)
 	create_trace_options_dir(tr);
 
 #if defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)
-	trace_create_file("tracing_max_latency", 0644, d_tracer,
-			&tr->max_latency, &tracing_max_lat_fops);
+	trace_create_maxlat_file(tr, d_tracer);
 #endif
 
 	if (ftrace_create_function_files(tr, d_tracer))
diff --git a/kernel/trace/trace.h b/kernel/trace/trace.h
index 005f08629b8b..4913ed1138aa 100644
--- a/kernel/trace/trace.h
+++ b/kernel/trace/trace.h
@@ -16,6 +16,8 @@
 #include <linux/trace_events.h>
 #include <linux/compiler.h>
 #include <linux/glob.h>
+#include <linux/irq_work.h>
+#include <linux/workqueue.h>
 
 #ifdef CONFIG_FTRACE_SYSCALLS
 #include <asm/unistd.h>		/* For NR_SYSCALLS	     */
@@ -264,6 +266,11 @@ struct trace_array {
 #endif
 #if defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)
 	unsigned long		max_latency;
+#ifdef CONFIG_FSNOTIFY
+	struct dentry		*d_max_latency;
+	struct work_struct	fsnotify_work;
+	struct irq_work		fsnotify_irqwork;
+#endif
 #endif
 	struct trace_pid_list	__rcu *filtered_pids;
 	/*
@@ -785,6 +792,17 @@ void update_max_tr_single(struct trace_array *tr,
 			  struct task_struct *tsk, int cpu);
 #endif /* CONFIG_TRACER_MAX_TRACE */
 
+#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
+	defined(CONFIG_FSNOTIFY)
+
+void latency_fsnotify(struct trace_array *tr);
+
+#else
+
+#define latency_fsnotify(tr)     do { } while (0)
+
+#endif
+
 #ifdef CONFIG_STACKTRACE
 void __trace_stack(struct trace_array *tr, unsigned long flags, int skip,
 		   int pc);
diff --git a/kernel/trace/trace_hwlat.c b/kernel/trace/trace_hwlat.c
index fa95139445b2..9c379261ee89 100644
--- a/kernel/trace/trace_hwlat.c
+++ b/kernel/trace/trace_hwlat.c
@@ -254,8 +254,10 @@ static int get_sample(void)
 		trace_hwlat_sample(&s);
 
 		/* Keep a running maximum ever recorded hardware latency */
-		if (sample > tr->max_latency)
+		if (sample > tr->max_latency) {
 			tr->max_latency = sample;
+			latency_fsnotify(tr);
+		}
 	}
 
 out:
-- 
2.17.1

[PATCH v6 2/4] preemptirq_delay_test: Add the burst feature and a sysfs trigger

[Viktor Rosendahl]

This burst feature enables the user to generate a burst of
preempt/irqsoff latencies. This makes it possible to test whether we
are able to detect latencies that systematically occur very close to
each other.

The maximum burst size is 10. We also create 10 identical test
functions, so that we get 10 different backtraces; this is useful
when we want to test whether we can detect all the latencies in a
burst. Otherwise, there would be no easy way of differentiating
between which latency in a burst was captured by the tracer.

In addition, there is a sysfs trigger, so that it's not necessary to
reload the module to repeat the test. The trigger will appear as
/sys/kernel/preemptirq_delay_test/trigger in sysfs.

Signed-off-by: Viktor Rosendahl <viktor.rosendahl@gmail.com>
---
 kernel/trace/Kconfig                 |   6 +-
 kernel/trace/preemptirq_delay_test.c | 147 +++++++++++++++++++++++----
 2 files changed, 131 insertions(+), 22 deletions(-)

diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
index 98da8998c25c..5ff573d844e4 100644
--- a/kernel/trace/Kconfig
+++ b/kernel/trace/Kconfig
@@ -751,9 +751,9 @@ config PREEMPTIRQ_DELAY_TEST
 	  configurable delay. The module busy waits for the duration of the
 	  critical section.
 
-	  For example, the following invocation forces a one-time irq-disabled
-	  critical section for 500us:
-	  modprobe preemptirq_delay_test test_mode=irq delay=500000
+	  For example, the following invocation generates a burst of three
+	  irq-disabled critical sections for 500us:
+	  modprobe preemptirq_delay_test test_mode=irq delay=500 burst_size=3
 
 	  If unsure, say N
 
diff --git a/kernel/trace/preemptirq_delay_test.c b/kernel/trace/preemptirq_delay_test.c
index d8765c952fab..96bf6b4858bc 100644
--- a/kernel/trace/preemptirq_delay_test.c
+++ b/kernel/trace/preemptirq_delay_test.c
@@ -3,6 +3,9 @@
  * Preempt / IRQ disable delay thread to test latency tracers
  *
  * Copyright (C) 2018 Joel Fernandes (Google) <joel@joelfernandes.org>
+ * Copyright (C) 2018, 2019 BMW Car IT GmbH
+ * Author: Viktor Rosendahl (viktor.rosendahl@bmw.de)
+ * - Added the burst feature and the sysfs trigger
  */
 
 #include <linux/trace_clock.h>
@@ -10,18 +13,25 @@
 #include <linux/interrupt.h>
 #include <linux/irq.h>
 #include <linux/kernel.h>
+#include <linux/kobject.h>
 #include <linux/kthread.h>
 #include <linux/module.h>
 #include <linux/printk.h>
 #include <linux/string.h>
+#include <linux/sysfs.h>
 
 static ulong delay = 100;
-static char test_mode[10] = "irq";
+static char test_mode[12] = "irq";
+static uint burst_size = 1;
 
-module_param_named(delay, delay, ulong, S_IRUGO);
-module_param_string(test_mode, test_mode, 10, S_IRUGO);
-MODULE_PARM_DESC(delay, "Period in microseconds (100 uS default)");
-MODULE_PARM_DESC(test_mode, "Mode of the test such as preempt or irq (default irq)");
+module_param_named(delay, delay, ulong, 0444);
+module_param_string(test_mode, test_mode, 12, 0444);
+module_param_named(burst_size, burst_size, uint, 0444);
+MODULE_PARM_DESC(delay, "Period in microseconds (100 us default)");
+MODULE_PARM_DESC(test_mode, "Mode of the test such as preempt, irq, or alternate (default irq)");
+MODULE_PARM_DESC(burst_size, "The size of a burst (default 1)");
+
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
 
 static void busy_wait(ulong time)
 {
@@ -34,37 +44,136 @@ static void busy_wait(ulong time)
 	} while ((end - start) < (time * 1000));
 }
 
-static int preemptirq_delay_run(void *data)
+static __always_inline void irqoff_test(void)
 {
 	unsigned long flags;
+	local_irq_save(flags);
+	busy_wait(delay);
+	local_irq_restore(flags);
+}
 
-	if (!strcmp(test_mode, "irq")) {
-		local_irq_save(flags);
-		busy_wait(delay);
-		local_irq_restore(flags);
-	} else if (!strcmp(test_mode, "preempt")) {
-		preempt_disable();
-		busy_wait(delay);
-		preempt_enable();
+static __always_inline void preemptoff_test(void)
+{
+	preempt_disable();
+	busy_wait(delay);
+	preempt_enable();
+}
+
+static void execute_preemptirqtest(int idx)
+{
+	if (!strcmp(test_mode, "irq"))
+		irqoff_test();
+	else if (!strcmp(test_mode, "preempt"))
+		preemptoff_test();
+	else if (!strcmp(test_mode, "alternate")) {
+		if (idx % 2 == 0)
+			irqoff_test();
+		else
+			preemptoff_test();
 	}
+}
+
+#define DECLARE_TESTFN(POSTFIX)				\
+	static void preemptirqtest_##POSTFIX(int idx)	\
+	{						\
+		execute_preemptirqtest(idx);		\
+	}						\
 
+/*
+ * We create 10 different functions, so that we can get 10 different
+ * backtraces.
+ */
+DECLARE_TESTFN(0)
+DECLARE_TESTFN(1)
+DECLARE_TESTFN(2)
+DECLARE_TESTFN(3)
+DECLARE_TESTFN(4)
+DECLARE_TESTFN(5)
+DECLARE_TESTFN(6)
+DECLARE_TESTFN(7)
+DECLARE_TESTFN(8)
+DECLARE_TESTFN(9)
+
+static void (*testfuncs[])(int)  = {
+	preemptirqtest_0,
+	preemptirqtest_1,
+	preemptirqtest_2,
+	preemptirqtest_3,
+	preemptirqtest_4,
+	preemptirqtest_5,
+	preemptirqtest_6,
+	preemptirqtest_7,
+	preemptirqtest_8,
+	preemptirqtest_9,
+};
+
+#define NR_TEST_FUNCS ARRAY_SIZE(testfuncs)
+
+static int preemptirq_delay_run(void *data)
+{
+	int i;
+	int s = MIN(burst_size, NR_TEST_FUNCS);
+
+	for (i = 0; i < s; i++)
+		(testfuncs[i])(i);
 	return 0;
 }
 
-static int __init preemptirq_delay_init(void)
+static struct task_struct *preemptirq_start_test(void)
 {
 	char task_name[50];
-	struct task_struct *test_task;
 
 	snprintf(task_name, sizeof(task_name), "%s_test", test_mode);
+	return kthread_run(preemptirq_delay_run, NULL, task_name);
+}
+
+
+static ssize_t trigger_store(struct kobject *kobj, struct kobj_attribute *attr,
+			 const char *buf, size_t count)
+{
+	preemptirq_start_test();
+	return count;
+}
+
+static struct kobj_attribute trigger_attribute =
+	__ATTR(trigger, 0200, NULL, trigger_store);
+
+static struct attribute *attrs[] = {
+	&trigger_attribute.attr,
+	NULL,
+};
+
+static struct attribute_group attr_group = {
+	.attrs = attrs,
+};
+
+static struct kobject *preemptirq_delay_kobj;
+
+static int __init preemptirq_delay_init(void)
+{
+	struct task_struct *test_task;
+	int retval;
+
+	test_task = preemptirq_start_test();
+	retval = PTR_ERR_OR_ZERO(test_task);
+	if (retval != 0)
+		return retval;
+
+	preemptirq_delay_kobj = kobject_create_and_add("preemptirq_delay_test",
+						       kernel_kobj);
+	if (!preemptirq_delay_kobj)
+		return -ENOMEM;
+
+	retval = sysfs_create_group(preemptirq_delay_kobj, &attr_group);
+	if (retval)
+		kobject_put(preemptirq_delay_kobj);
 
-	test_task = kthread_run(preemptirq_delay_run, NULL, task_name);
-	return PTR_ERR_OR_ZERO(test_task);
+	return retval;
 }
 
 static void __exit preemptirq_delay_exit(void)
 {
-	return;
+	kobject_put(preemptirq_delay_kobj);
 }
 
 module_init(preemptirq_delay_init)
-- 
2.17.1

[PATCH v6 3/4] Add the latency-collector to tools

[Viktor Rosendahl]

This is a tool that is intended to work around the fact that the
preemptoff, irqsoff, and preemptirqsoff tracers only work in
overwrite mode. The idea is to act randomly in such a way that we
do not systematically lose any latencies, so that if enough testing
is done, all latencies will be captured. If the same burst of
latencies is repeated, then sooner or later we will have captured all
the latencies.

It also works with the wakeup_dl, wakeup_rt, wakeup, and hwlat tracers.
However, in that case it is probably not useful to use the random
sleep functionality.

The reason why it may be desirable to catch all latencies with a long
test campaign is that for some organizations, it's necessary to test
the kernel in the field and not practical for developers to work
iteratively with field testers. Because of cost and project schedules
it is not possible to start a new test campaign every time a latency
problem has been fixed.

It uses inotify to detect changes to /sys/kernel/debug/tracing/trace.
When a latency is detected, it will either sleep or print
immediately, depending on a function that act as an unfair coin
toss.

If immediate print is chosen, it means that we open
/sys/kernel/debug/tracing/trace and thereby cause a blackout period
that will hide any subsequent latencies.

If sleep is chosen, it means that we wait before opening
/sys/kernel/debug/tracing/trace, by default for 1000 ms, to see if
there is another latency during this period. If there is, then we will
lose the previous latency. The coin will be tossed again with a
different probability, and we will either print the new latency, or
possibly a subsequent one.

The probability for the unfair coin toss is chosen so that there
is equal probability to obtain any of the latencies in a burst.
However, this assumes that we make an assumption of how many
latencies there can be. By default  the program assumes that there
are no more than 2 latencies in a burst, the probability of immediate
printout will be:

1/2 and 1

Thus, the probability of getting each of the two latencies will be 1/2.

If we ever find that there is more than one latency in a series,
meaning that we reach the probability of 1, then the table will be
expanded to:

1/3, 1/2, and 1

Thus, we assume that there are no more than three latencies and each
with a probability of 1/3 of being captured. If the probability of 1
is reached in the new table, that is we see more than two closely
occurring latencies, then the table will again be extended, and so
on.

On my systems, it seems like this scheme works fairly well, as
long as the latencies we trace are long enough, 300 us seems to be
enough. This userspace program receive the inotify event at the end
of a latency, and it has time until the end of the next latency
to react, that is to open /sys/kernel/debug/tracing/trace. Thus,
if we trace latencies that are >300 us, then we have at least 300 us
to react.

The minimum latency will of course not be 300 us on all systems, it
will depend on the hardware, kernel version, workload and
configuration.

Example output:

root@myhost:~# echo 100 > /sys/kernel/debug/tracing/tracing_thresh
root@myhost:~# echo preemptirqsoff > /sys/kernel/debug/tracing/current_tracer
root@myhost:~# latency-collector -rvv -a 11 &
[1] 4915
root@myhost:~# Running with policy rr and priority 99. Using 3 print threads.
/sys/kernel/debug/tracing/trace will be printed with random delay
Start size of the probability table:                    11
Print a message when prob. table table changes size:    true
Print a warning when an event has been lost:            true
Sleep time is:                                          1000 ms
Initializing probability table to 11
Thread 4916 runs with policy rr and priority 99
Thread 4917 runs with policy rr and priority 99
Thread 4918 runs with policy rr and priority 99
Thread 4919 runs with policy rr and priority 99

root@myhost:~# modprobe preemptirq_delay_test delay=105 test_mode=alternate burst_size=10
2850.638308 Latency 1 printout skipped due to random delay
2850.638383 Latency 2 printout skipped due to random delay
2850.638497 Latency 3 immediate print
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> BEGIN <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
alternat-4922    6d...    0us!: execute_preemptirqtest <-preemptirqtest_2
alternat-4922    6d...  106us : execute_preemptirqtest <-preemptirqtest_2
alternat-4922    6d...  106us : tracer_hardirqs_on <-preemptirqtest_2
alternat-4922    6d...  108us : <stack trace>
 => preemptirqtest_2
 => preemptirq_delay_run
 => kthread
 => ret_from_fork
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> END <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<

In the example above, we randomly, happened to get the third latency
in a burst burst of 10. If the experiment is repeated enough times,
we will get all 10.

Sometimes, there will be lost latencies, so that we get:

3054.078294 Latency 44 printout skipped due to inotify loop
..or:

3211.957685 Latency 112 printout skipped due to print loop

In my experience, these are not frequent enough to be a problem.

Also, we do not really lose any latency in these cases, it's
more like we get another latency than what was intended by
the design.

Signed-off-by: Viktor Rosendahl <viktor.rosendahl@gmail.com>
---
 tools/Makefile                  |   14 +-
 tools/trace/Makefile            |   20 +
 tools/trace/latency-collector.c | 1212 +++++++++++++++++++++++++++++++
 3 files changed, 1240 insertions(+), 6 deletions(-)
 create mode 100644 tools/trace/Makefile
 create mode 100644 tools/trace/latency-collector.c

diff --git a/tools/Makefile b/tools/Makefile
index 68defd7ecf5d..2b3738e66280 100644
--- a/tools/Makefile
+++ b/tools/Makefile
@@ -30,6 +30,7 @@ help:
 	@echo '  spi                    - spi tools'
 	@echo '  objtool                - an ELF object analysis tool'
 	@echo '  tmon                   - thermal monitoring and tuning tool'
+	@echo '  trace                  - misc tracing tools'
 	@echo '  turbostat              - Intel CPU idle stats and freq reporting tool'
 	@echo '  usb                    - USB testing tools'
 	@echo '  virtio                 - vhost test module'
@@ -63,7 +64,7 @@ acpi: FORCE
 cpupower: FORCE
 	$(call descend,power/$@)
 
-cgroup firewire hv guest spi usb virtio vm bpf iio gpio objtool leds wmi pci firmware debugging: FORCE
+cgroup firewire hv guest spi usb virtio vm bpf iio gpio objtool leds wmi pci firmware debugging trace: FORCE
 	$(call descend,$@)
 
 liblockdep: FORCE
@@ -99,7 +100,7 @@ all: acpi cgroup cpupower gpio hv firewire liblockdep \
 		perf selftests spi turbostat usb \
 		virtio vm bpf x86_energy_perf_policy \
 		tmon freefall iio objtool kvm_stat wmi \
-		pci debugging
+		pci debugging trace
 
 acpi_install:
 	$(call descend,power/$(@:_install=),install)
@@ -107,7 +108,7 @@ acpi_install:
 cpupower_install:
 	$(call descend,power/$(@:_install=),install)
 
-cgroup_install firewire_install gpio_install hv_install iio_install perf_install spi_install usb_install virtio_install vm_install bpf_install objtool_install wmi_install pci_install debugging_install:
+cgroup_install firewire_install gpio_install hv_install iio_install perf_install spi_install usb_install virtio_install vm_install bpf_install objtool_install wmi_install pci_install debugging_install trace_install:
 	$(call descend,$(@:_install=),install)
 
 liblockdep_install:
@@ -133,7 +134,8 @@ install: acpi_install cgroup_install cpupower_install gpio_install \
 		perf_install selftests_install turbostat_install usb_install \
 		virtio_install vm_install bpf_install x86_energy_perf_policy_install \
 		tmon_install freefall_install objtool_install kvm_stat_install \
-		wmi_install pci_install debugging_install intel-speed-select_install
+		wmi_install pci_install debugging_install intel-speed-select_install \
+		trace_install
 
 acpi_clean:
 	$(call descend,power/acpi,clean)
@@ -141,7 +143,7 @@ acpi_clean:
 cpupower_clean:
 	$(call descend,power/cpupower,clean)
 
-cgroup_clean hv_clean firewire_clean spi_clean usb_clean virtio_clean vm_clean wmi_clean bpf_clean iio_clean gpio_clean objtool_clean leds_clean pci_clean firmware_clean debugging_clean:
+cgroup_clean hv_clean firewire_clean spi_clean usb_clean virtio_clean vm_clean wmi_clean bpf_clean iio_clean gpio_clean objtool_clean leds_clean pci_clean firmware_clean debugging_clean trace_clean:
 	$(call descend,$(@:_clean=),clean)
 
 liblockdep_clean:
@@ -180,6 +182,6 @@ clean: acpi_clean cgroup_clean cpupower_clean hv_clean firewire_clean \
 		vm_clean bpf_clean iio_clean x86_energy_perf_policy_clean tmon_clean \
 		freefall_clean build_clean libbpf_clean libsubcmd_clean liblockdep_clean \
 		gpio_clean objtool_clean leds_clean wmi_clean pci_clean firmware_clean debugging_clean \
-		intel-speed-select_clean
+		intel-speed-select_clean trace_clean
 
 .PHONY: FORCE
diff --git a/tools/trace/Makefile b/tools/trace/Makefile
new file mode 100644
index 000000000000..59cd483ab01f
--- /dev/null
+++ b/tools/trace/Makefile
@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: GPL-2.0
+# Makefile for vm tools
+#
+TARGETS = latency-collector
+CFLAGS = -Wall -Wextra -g -O2
+LDFLAGS = -lpthread
+
+all: $(TARGETS)
+
+%: %.c
+	$(CC) $(CFLAGS) -o $@ $< $(LDFLAGS)
+
+clean:
+	$(RM) latency-collector
+
+sbindir ?= /usr/sbin
+
+install: all
+	install -d $(DESTDIR)$(sbindir)
+	install -m 755 -p $(TARGETS) $(DESTDIR)$(sbindir)
diff --git a/tools/trace/latency-collector.c b/tools/trace/latency-collector.c
new file mode 100644
index 000000000000..ae1ef0a94966
--- /dev/null
+++ b/tools/trace/latency-collector.c
@@ -0,0 +1,1212 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2017, 2018, 2019 BMW Car IT GmbH
+ * Author: Viktor Rosendahl (viktor.rosendahl@bmw.de)
+ */
+
+#define _GNU_SOURCE
+#define _POSIX_C_SOURCE 200809L
+
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <err.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <getopt.h>
+#include <sched.h>
+#include <linux/unistd.h>
+#include <sys/inotify.h>
+#include <unistd.h>
+#include <pthread.h>
+
+static const char *prg_name;
+static const char *prg_unknown = "unknown program name";
+
+static int fd_stdout;
+
+/* These are default values */
+static int sched_policy;
+static bool sched_policy_set;
+
+static int sched_pri;
+static bool sched_pri_set;
+
+static bool trace_enable = true;
+static bool use_random_sleep;
+
+static char inotify_buffer[655360];
+
+#define likely(x)      __builtin_expect(!!(x), 1)
+#define unlikely(x)    __builtin_expect(!!(x), 0)
+#define bool2str(x)    (x ? "true":"false")
+
+#define DEFAULT_NR_PRINTER_THREADS (3)
+static unsigned int nr_threads = DEFAULT_NR_PRINTER_THREADS;
+
+#define DEFAULT_TABLE_SIZE (2)
+static unsigned int table_startsize = DEFAULT_TABLE_SIZE;
+
+static int verbosity;
+
+#define verbose_sizechange() (verbosity >= 1)
+#define verbose_lostevent()  (verbosity >= 2)
+
+static const char *debug_tracefile;
+static const char *debug_tracefile_dflt = "/sys/kernel/debug/tracing/trace";
+static const char *debug_maxlat_file;
+static const char *debug_maxlat_dflt =
+	"/sys/kernel/debug/tracing/tracing_max_latency";
+
+
+#define DEV_URANDOM     "/dev/urandom"
+#define RT_DEFAULT_PRI (99)
+#define DEFAULT_PRI    (0)
+
+#define USEC_PER_MSEC (1000L)
+#define NSEC_PER_USEC (1000L)
+#define NSEC_PER_MSEC (USEC_PER_MSEC * NSEC_PER_USEC)
+
+#define MSEC_PER_SEC (1000L)
+#define USEC_PER_SEC (USEC_PER_MSEC * MSEC_PER_SEC)
+#define NSEC_PER_SEC (NSEC_PER_MSEC * MSEC_PER_SEC)
+
+#define SLEEP_TIME_MS_DEFAULT (1000L)
+
+static long sleep_time = (USEC_PER_MSEC * SLEEP_TIME_MS_DEFAULT);
+
+static char queue_full_warning[] =
+"Could not queue trace for printing. It is likely that events happen faster\n"
+"than what they can be printed. Probably partly because of random sleeping\n";
+
+
+struct policy {
+	const char *name;
+	int policy;
+	int default_pri;
+};
+
+static const struct policy policies[] = {
+	{ "other", SCHED_OTHER, DEFAULT_PRI    },
+	{ "batch", SCHED_BATCH, DEFAULT_PRI    },
+	{ "idle",  SCHED_IDLE,  DEFAULT_PRI    },
+	{ "rr",    SCHED_RR,    RT_DEFAULT_PRI },
+	{ "fifo",  SCHED_FIFO,  RT_DEFAULT_PRI },
+	{ NULL,    0,           DEFAULT_PRI    }
+};
+
+struct entry {
+	int ticket;
+	int ticket_completed_ref;
+};
+
+struct print_state {
+	int ticket_counter;
+	int ticket_completed;
+	pthread_mutex_t mutex;
+	pthread_cond_t cond;
+	int cnt;
+	pthread_mutex_t cnt_mutex;
+};
+
+struct short_msg {
+	char buf[160];
+	int len;
+};
+
+static struct print_state printstate;
+
+#define PROB_TABLE_MAX_SIZE (1000)
+
+int probabilities[PROB_TABLE_MAX_SIZE];
+
+struct sleep_table {
+	int *table;
+	int size;
+	pthread_mutex_t mutex;
+};
+
+static struct sleep_table sleeptable;
+
+#define QUEUE_SIZE (10)
+
+struct queue {
+	struct entry entries[QUEUE_SIZE];
+	int next_prod_idx;
+	int next_cons_idx;
+	pthread_mutex_t mutex;
+	pthread_cond_t cond;
+};
+
+#define MAX_THREADS (40)
+
+struct queue printqueue;
+pthread_t printthread[MAX_THREADS];
+pthread_mutex_t print_mtx = PTHREAD_MUTEX_INITIALIZER;
+#define PRINT_BUFFER_SIZE (16 * 1024 * 1024)
+
+static __always_inline void *malloc_or_die(size_t size)
+{
+	void *ptr = malloc(size);
+
+	if (ptr == NULL)
+		err(0, "malloc() failed");
+	return ptr;
+}
+
+static __always_inline void write_or_die(int fd, const char *buf, size_t count)
+{
+	ssize_t r;
+
+	do {
+		r = write(fd, buf, count);
+		if (r < 0) {
+			if (errno == EINTR)
+				continue;
+			err(0, "write() failed");
+		}
+		count -= r;
+		buf += r;
+	} while (count > 0);
+}
+
+static __always_inline void clock_gettime_or_die(clockid_t clk_id,
+						 struct timespec *tp)
+{
+	int r = clock_gettime(clk_id, tp);
+
+	if (r != 0)
+		err(0, "clock_gettime() failed");
+}
+
+static void open_stdout(void)
+{
+	if (setvbuf(stdout, NULL, _IONBF, 0) != 0)
+		err(0, "setvbuf() failed");
+	fd_stdout = fileno(stdout);
+	if (fd_stdout < 0)
+		err(0, "fileno() failed");
+}
+
+static __always_inline void get_time_in_future(struct timespec *future,
+					       long time_us)
+{
+	long nsec;
+
+	clock_gettime_or_die(CLOCK_MONOTONIC, future);
+	future->tv_sec += time_us / USEC_PER_SEC;
+	nsec = future->tv_nsec + (time_us * NSEC_PER_USEC) % NSEC_PER_SEC;
+	if (nsec >= NSEC_PER_SEC) {
+		future->tv_nsec = nsec % NSEC_PER_SEC;
+		future->tv_sec += 1;
+	}
+}
+
+static __always_inline bool time_has_passed(const struct timespec *time)
+{
+	struct timespec now;
+
+	clock_gettime_or_die(CLOCK_MONOTONIC, &now);
+	if (now.tv_sec > time->tv_sec)
+		return true;
+	if (now.tv_sec < time->tv_sec)
+		return false;
+	return (now.tv_nsec >= time->tv_nsec);
+}
+
+static void init_printstate(void)
+{
+	pthread_condattr_t attr;
+
+	printstate.ticket_counter = 0;
+	printstate.ticket_completed = 0;
+	printstate.cnt = 0;
+
+	if (pthread_mutex_init(&printstate.mutex, NULL) != 0)
+		err(0, "pthread_mutex_init() failed");
+
+	if (pthread_mutex_init(&printstate.cnt_mutex, NULL) != 0)
+		err(0, "pthread_mutex_init() failed");
+
+	if (pthread_condattr_init(&attr) != 0)
+		err(0, "pthread_condattr_init()");
+
+	if (pthread_condattr_setclock(&attr, CLOCK_MONOTONIC) != 0)
+		err(0, "pthread_condattr_setclock()");
+
+	if (pthread_cond_init(&printstate.cond, &attr) != 0)
+		err(0, "pthread_cond_init() failed");
+}
+
+static __always_inline void mutex_lock(pthread_mutex_t *mtx)
+{
+	if (pthread_mutex_lock(mtx) != 0)
+		err(0, "pthread_mutex_lock() failed");
+}
+
+static __always_inline void mutex_unlock(pthread_mutex_t *mtx)
+{
+	if (pthread_mutex_unlock(mtx) != 0)
+		err(0, "pthread_mutex_unlock() failed");
+}
+
+static __always_inline void cond_signal(pthread_cond_t *cond)
+{
+	if (pthread_cond_signal(cond) != 0)
+		err(0, "pthread_cond_signal() failed");
+}
+
+static __always_inline void cond_broadcast(pthread_cond_t *cond)
+{
+	if (pthread_cond_broadcast(cond) != 0)
+		err(0, "pthread_cond_broadcast() failed");
+}
+
+static int printstate_next_ticket(struct entry *req)
+{
+	int r;
+
+	r = ++(printstate.ticket_counter);
+	req->ticket = r;
+	req->ticket_completed_ref = printstate.ticket_completed;
+	cond_broadcast(&printstate.cond);
+	return r;
+}
+
+static __always_inline
+void printstate_mark_req_completed(const struct entry *req)
+{
+	if (req->ticket > printstate.ticket_completed)
+		printstate.ticket_completed = req->ticket;
+}
+
+static __always_inline
+bool printstate_has_new_req_arrived(const struct entry *req)
+{
+	return (printstate.ticket_counter != req->ticket);
+}
+
+static __always_inline int printstate_cnt_inc(void)
+{
+	int value;
+
+	mutex_lock(&printstate.cnt_mutex);
+	value = ++printstate.cnt;
+	mutex_unlock(&printstate.cnt_mutex);
+	return value;
+}
+
+static __always_inline int printstate_cnt_dec(void)
+{
+	int value;
+
+	mutex_lock(&printstate.cnt_mutex);
+	value = --printstate.cnt;
+	mutex_unlock(&printstate.cnt_mutex);
+	return value;
+}
+
+static __always_inline int printstate_cnt_read(void)
+{
+	int value;
+
+	mutex_lock(&printstate.cnt_mutex);
+	value = printstate.cnt;
+	mutex_unlock(&printstate.cnt_mutex);
+	return value;
+}
+
+static __always_inline
+bool prev_req_won_race(const struct entry *req)
+{
+	return (printstate.ticket_completed != req->ticket_completed_ref);
+}
+
+static void sleeptable_resize(int size, bool printout, struct short_msg *msg)
+{
+	int bytes;
+
+	if (printout) {
+		msg->len = 0;
+		if (unlikely(size > PROB_TABLE_MAX_SIZE))
+			bytes = snprintf(msg->buf, sizeof(msg->buf),
+"Cannot increase probability table to %d (maximum size reached)\n", size);
+		else
+			bytes = snprintf(msg->buf, sizeof(msg->buf),
+"Increasing probability table to %d\n", size);
+		if (bytes < 0)
+			warn("snprintf() failed");
+		else
+			msg->len = bytes;
+	}
+
+	if (unlikely(size < 0)) {
+		/* Should never happen */
+		errx(0, "Bad program state at %s:%d", __FILE__, __LINE__);
+		return;
+	}
+	sleeptable.size = size;
+	sleeptable.table = &probabilities[PROB_TABLE_MAX_SIZE - size];
+}
+
+static void init_probabilities(void)
+{
+	int i;
+	int j = 1000;
+
+	for (i = 0; i < PROB_TABLE_MAX_SIZE; i++) {
+		probabilities[i] = 1000 / j;
+		j--;
+	}
+	if (pthread_mutex_init(&sleeptable.mutex, NULL) != 0)
+		err(0, "pthread_mutex_init() failed");
+}
+
+static int table_get_probability(const struct entry *req,
+				 struct short_msg *msg)
+{
+	int diff = req->ticket - req->ticket_completed_ref;
+	int rval = 0;
+
+	msg->len = 0;
+	diff--;
+	/* Should never happen...*/
+	if (diff < 0)
+		errx(0, "Programmer assumption error at %s:%d\n", __FILE__,
+		     __LINE__);
+	mutex_lock(&sleeptable.mutex);
+	if (diff >= (sleeptable.size - 1)) {
+		rval = sleeptable.table[sleeptable.size - 1];
+		sleeptable_resize(sleeptable.size + 1, verbose_sizechange(),
+				  msg);
+	} else {
+		rval = sleeptable.table[diff];
+	}
+	mutex_unlock(&sleeptable.mutex);
+	return rval;
+}
+
+static void init_queue(struct queue *q)
+{
+	q->next_prod_idx = 0;
+	q->next_cons_idx = 0;
+	if (pthread_mutex_init(&q->mutex, NULL) != 0)
+		err(0, "pthread_mutex_init() failed");
+	if (pthread_cond_init(&q->cond, NULL) != 0)
+		err(0, "pthread_cond_init() failed");
+}
+
+static __always_inline int queue_len(const struct queue *q)
+{
+	if (q->next_prod_idx >= q->next_cons_idx)
+		return q->next_prod_idx - q->next_cons_idx;
+	else
+		return QUEUE_SIZE - q->next_cons_idx + q->next_prod_idx;
+}
+
+static __always_inline int queue_nr_free(const struct queue *q)
+{
+	int nr_free = QUEUE_SIZE - queue_len(q);
+
+	/*
+	 * If there is only one slot left we will anyway lie and claim that the
+	 * queue is full because adding an element will make it appear empty
+	 */
+	if (nr_free == 1)
+		nr_free = 0;
+	return nr_free;
+}
+
+static __always_inline void queue_idx_inc(int *idx)
+{
+	*idx = (*idx + 1) % QUEUE_SIZE;
+}
+
+static __always_inline void queue_push_to_back(struct queue *q,
+					      const struct entry *e)
+{
+	q->entries[q->next_prod_idx] = *e;
+	queue_idx_inc(&q->next_prod_idx);
+}
+
+static __always_inline struct entry queue_pop_from_front(struct queue *q)
+{
+	struct entry e = q->entries[q->next_cons_idx];
+
+	queue_idx_inc(&q->next_cons_idx);
+	return e;
+}
+
+static __always_inline void queue_cond_signal(struct queue *q)
+{
+	if (pthread_cond_signal(&q->cond) != 0)
+		err(0, "pthread_cond_signal() failed");
+}
+
+static __always_inline void queue_cond_wait(struct queue *q)
+{
+	if (pthread_cond_wait(&q->cond, &q->mutex) != 0)
+		err(0, "pthread_cond_wait() failed");
+}
+
+
+static __always_inline int queue_try_to_add_entry(struct queue *q,
+						  const struct entry *e)
+{
+	int r = 0;
+
+	mutex_lock(&q->mutex);
+	if (queue_nr_free(q) > 0) {
+		queue_push_to_back(q, e);
+		cond_signal(&q->cond);
+	} else
+		r = -1;
+	mutex_unlock(&q->mutex);
+	return r;
+}
+
+static struct entry queue_wait_for_entry(struct queue *q)
+{
+	struct entry e;
+
+	mutex_lock(&q->mutex);
+	while (true) {
+		if (queue_len(&printqueue) > 0) {
+			e = queue_pop_from_front(q);
+			break;
+		}
+		queue_cond_wait(q);
+	}
+	mutex_unlock(&q->mutex);
+
+	return e;
+}
+
+static const struct policy *policy_from_name(const char *name)
+{
+	const struct policy *p = &policies[0];
+
+	while (p->name != NULL) {
+		if (!strcmp(name, p->name))
+			return p;
+		p++;
+	}
+	return NULL;
+}
+
+static const char *policy_name(int policy)
+{
+	const struct policy *p = &policies[0];
+	static const char *rval = "unknown";
+
+	while (p->name != NULL) {
+		if (p->policy == policy)
+			return p->name;
+		p++;
+	}
+	return rval;
+}
+
+static bool toss_coin(struct drand48_data *buffer, unsigned int prob)
+{
+	long r;
+
+	if (lrand48_r(buffer, &r) != 0)
+		err(0, "lrand48_r() failed");
+	r = r % 1000L;
+	if (r < prob)
+		return true;
+	else
+		return false;
+}
+
+
+static long go_to_sleep(const struct entry *req)
+{
+	struct timespec future;
+	long delay = sleep_time;
+
+	get_time_in_future(&future, delay);
+
+	mutex_lock(&printstate.mutex);
+	while (!printstate_has_new_req_arrived(req)) {
+		if (pthread_cond_timedwait(&printstate.cond, &printstate.mutex,
+					   &future) != 0) {
+			if (errno != ETIMEDOUT && errno != 0)
+				err(0, "pthread_cond_timedwait() %d", errno);
+		}
+		if (time_has_passed(&future))
+			break;
+	};
+
+	if (printstate_has_new_req_arrived(req))
+		delay = -1;
+	mutex_unlock(&printstate.mutex);
+
+	return delay;
+}
+
+
+static void set_priority(void)
+{
+	int r;
+	pid_t pid;
+	struct sched_param param;
+
+	memset(&param, 0, sizeof(param));
+	param.sched_priority = sched_pri;
+
+	pid = getpid();
+	r = sched_setscheduler(pid, sched_policy, &param);
+
+	if (r != 0)
+		err(0, "sched_setscheduler() failed");
+}
+
+pid_t latency_collector_gettid(void)
+{
+	return (pid_t) syscall(__NR_gettid);
+}
+
+static void print_priority(void)
+{
+	pid_t tid;
+	int policy;
+	int r;
+	struct sched_param param;
+
+	tid = latency_collector_gettid();
+	r = pthread_getschedparam(pthread_self(), &policy, &param);
+	if (r != 0)
+		err(0, "pthread_getschedparam() failed");
+	mutex_lock(&print_mtx);
+	printf("Thread %d runs with scheduling policy %s and priority %d\n",
+	       tid, policy_name(policy), param.sched_priority);
+	mutex_unlock(&print_mtx);
+}
+
+static __always_inline
+void __print_skipmessage(const struct short_msg *resize_msg,
+			 const struct timespec *timestamp, char *buffer,
+			 size_t bufspace, const struct entry *req, bool excuse,
+			 const char *str)
+{
+	ssize_t bytes = 0;
+	char *p = &buffer[0];
+	long us, sec;
+	int r;
+
+	sec = timestamp->tv_sec;
+	us = timestamp->tv_nsec / 1000;
+
+	if (resize_msg != NULL && resize_msg->len > 0) {
+		strncpy(p, resize_msg->buf, resize_msg->len);
+		bytes += resize_msg->len;
+		p += resize_msg->len;
+		bufspace -= resize_msg->len;
+	}
+
+	if (excuse)
+		r = snprintf(p, bufspace,
+"%ld.%06ld Latency %d printout skipped due to %s\n",
+			     sec, us, req->ticket, str);
+	else
+		r = snprintf(p, bufspace, "%ld.%06ld Latency %d detected\n",
+			    sec, us, req->ticket);
+
+	if (r < 0)
+		warn("snprintf() failed");
+	else
+		bytes += r;
+
+	/* These prints could happen concurrently */
+	mutex_lock(&print_mtx);
+	write_or_die(fd_stdout, buffer, bytes);
+	mutex_unlock(&print_mtx);
+}
+
+static void print_skipmessage(const struct short_msg *resize_msg,
+			      const struct timespec *timestamp, char *buffer,
+			      size_t bufspace, const struct entry *req,
+			      bool excuse)
+{
+	__print_skipmessage(resize_msg, timestamp, buffer, bufspace, req,
+			    excuse, "random delay");
+}
+
+static void print_lostmessage(const struct timespec *timestamp, char *buffer,
+			      size_t bufspace, const struct entry *req,
+			      const char *reason)
+{
+	__print_skipmessage(NULL, timestamp, buffer, bufspace, req, true,
+			    reason);
+}
+
+static void print_tracefile(const struct short_msg *resize_msg,
+			    const struct timespec *timestamp, char *buffer,
+			    size_t bufspace, long slept,
+			    const struct entry *req)
+{
+	static const int reserve = 256;
+	char *p = &buffer[0];
+	ssize_t bytes = 0;
+	ssize_t bytes_tot = 0;
+	long us, sec;
+	long slept_ms;
+	int trace_fd;
+
+	/* Save some space for the final string and final null char */
+	bufspace = bufspace - reserve - 1;
+
+	if (resize_msg != NULL && resize_msg->len > 0) {
+		bytes = resize_msg->len;
+		strncpy(p, resize_msg->buf, bytes);
+		bytes_tot += bytes;
+		p += bytes;
+		bufspace -= bytes;
+	}
+
+	trace_fd = open(debug_tracefile, O_RDONLY);
+
+	if (trace_fd < 0) {
+		warn("open() failed on %s", debug_tracefile);
+		return;
+	}
+
+	sec = timestamp->tv_sec;
+	us = timestamp->tv_nsec / 1000;
+
+	if (slept != 0) {
+		slept_ms = slept / 1000;
+		bytes = snprintf(p, bufspace,
+"%ld.%06ld Latency %d randomly sleep for %ld ms before print\n",
+				 sec, us, req->ticket, slept_ms);
+	} else {
+		bytes = snprintf(p, bufspace,
+				 "%ld.%06ld Latency %d immediate print\n", sec,
+				 us, req->ticket);
+	}
+
+	if (bytes < 0) {
+		warn("snprintf() failed");
+		return;
+	}
+	p += bytes;
+	bufspace -= bytes;
+	bytes_tot += bytes;
+
+	bytes = snprintf(p, bufspace,
+">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> BEGIN <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n"
+		);
+
+	if (bytes < 0) {
+		warn("snprintf() failed");
+		return;
+	}
+
+	p += bytes;
+	bufspace -= bytes;
+	bytes_tot += bytes;
+
+	do {
+		bytes = read(trace_fd, p, bufspace);
+		if (bytes < 0) {
+			if (errno == EINTR)
+				continue;
+			warn("read() failed on %s", debug_tracefile);
+			if (close(trace_fd) != 0)
+				warn("close() failed on %s", debug_tracefile);
+			return;
+		}
+		if (bytes == 0)
+			break;
+		p += bytes;
+		bufspace -= bytes;
+		bytes_tot += bytes;
+	} while (true);
+
+	if (close(trace_fd) != 0)
+		warn("close() failed on %s", debug_tracefile);
+
+	printstate_cnt_dec();
+	/* Add the reserve space back to the budget for the final string */
+	bufspace += reserve;
+
+	bytes = snprintf(p, bufspace,
+			 ">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> END <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n\n");
+
+	if (bytes < 0) {
+		warn("snprintf() failed");
+		return;
+	}
+
+	bytes_tot += bytes;
+
+	/* These prints could happen concurrently */
+	mutex_lock(&print_mtx);
+	write_or_die(fd_stdout, buffer, bytes_tot);
+	mutex_unlock(&print_mtx);
+}
+
+static void do_jittertest(void)
+{
+	int ifd = inotify_init();
+	int wd;
+	const ssize_t bufsize = sizeof(inotify_buffer);
+	const ssize_t istructsize = sizeof(struct inotify_event);
+	char *buf = &inotify_buffer[0];
+	ssize_t nr_read;
+	char *p;
+	int modified;
+	struct inotify_event *event;
+	struct entry req;
+	char *buffer;
+	const size_t bufspace = PRINT_BUFFER_SIZE;
+	struct timespec timestamp;
+
+	print_priority();
+
+	buffer = malloc_or_die(bufspace);
+
+	if (ifd < 0)
+		err(0, "inotify_init() failed!");
+
+	wd = inotify_add_watch(ifd, debug_maxlat_file, IN_MODIFY);
+	if (wd < 0)
+		err(0, "inotify_add_watch() failed!");
+
+	while (true) {
+		modified = 0;
+		nr_read = read(ifd, buf, bufsize);
+		if (nr_read <= 0)
+			err(0, "read() failed on inotify fd!");
+		if (nr_read == bufsize)
+			warnx("inotify() buffer filled, skipping events");
+		if (nr_read < istructsize)
+			errx(0, "read() returned too few bytes on inotify fd");
+
+		for (p = buf; p < buf + nr_read;) {
+			event = (struct inotify_event *) p;
+			if ((event->mask & IN_MODIFY) != 0)
+				modified++;
+			p += istructsize + event->len;
+		}
+		while (modified > 0) {
+			mutex_lock(&printstate.mutex);
+			printstate_next_ticket(&req);
+			if (printstate_cnt_read() > 0) {
+				printstate_mark_req_completed(&req);
+				mutex_unlock(&printstate.mutex);
+				if (verbose_lostevent()) {
+					clock_gettime_or_die(CLOCK_MONOTONIC,
+							     &timestamp);
+					print_lostmessage(&timestamp, buffer,
+							  bufspace, &req,
+							  "inotify loop");
+				}
+				break;
+			}
+			mutex_unlock(&printstate.mutex);
+			if (queue_try_to_add_entry(&printqueue, &req) != 0) {
+				/* These prints could happen concurrently */
+				mutex_lock(&print_mtx);
+				write_or_die(fd_stdout, queue_full_warning,
+					     sizeof(queue_full_warning));
+				mutex_unlock(&print_mtx);
+			}
+			modified--;
+		}
+	}
+}
+
+static void *do_printloop(void *arg)
+{
+	const size_t bufspace = PRINT_BUFFER_SIZE;
+	char *buffer;
+	long *rseed = (long *) arg;
+	struct drand48_data drandbuf;
+	long slept = 0;
+	struct entry req;
+	int prob = 0;
+	struct timespec timestamp;
+	struct short_msg resize_msg;
+
+	print_priority();
+
+	if (srand48_r(*rseed, &drandbuf) != 0)
+		err(0, "srand48_r() failed!\n");
+
+	buffer = malloc_or_die(bufspace);
+
+	while (true) {
+		req = queue_wait_for_entry(&printqueue);
+		clock_gettime_or_die(CLOCK_MONOTONIC, &timestamp);
+		mutex_lock(&printstate.mutex);
+		if (prev_req_won_race(&req)) {
+			printstate_mark_req_completed(&req);
+			mutex_unlock(&printstate.mutex);
+			if (verbose_lostevent())
+				print_lostmessage(&timestamp, buffer, bufspace,
+						  &req, "print loop");
+			continue;
+		}
+		mutex_unlock(&printstate.mutex);
+
+		/*
+		 * Toss a coin to decide if we want to sleep a random amount
+		 * before printing out the backtrace. The reason for this is
+		 * that opening /sys/kernel/debug/tracing/trace will cause a
+		 * blackout of about 430 ms, where no latencies will be noted
+		 * by the latency tracer. Thus by randomly sleeping a random
+		 * amount we try to avoid missing traces systematically due to
+		 * this. With this option we will sometimes get the first
+		 * latency, some other times some of the later ones, in case of
+		 * closely spaced traces.
+		 */
+		if (trace_enable && use_random_sleep) {
+			slept = 0;
+			prob = table_get_probability(&req, &resize_msg);
+			if (!toss_coin(&drandbuf, prob))
+				slept = go_to_sleep(&req);
+			if (slept >= 0) {
+				/* A print is ongoing */
+				printstate_cnt_inc();
+				/*
+				 * We will do the printout below so we have to
+				 * mark it as completed while we still have the
+				 * mutex.
+				 */
+				mutex_lock(&printstate.mutex);
+				printstate_mark_req_completed(&req);
+				mutex_unlock(&printstate.mutex);
+			}
+		}
+		if (trace_enable) {
+			/*
+			 * slept < 0  means that we detected another
+			 * notification in go_to_sleep() above
+			 */
+			if (slept >= 0)
+				/*
+				 * N.B. printstate_cnt_dec(); will be called
+				 * inside print_tracefile()
+				 */
+				print_tracefile(&resize_msg, &timestamp, buffer,
+						bufspace, slept, &req);
+			else
+				print_skipmessage(&resize_msg, &timestamp,
+						  buffer, bufspace, &req, true);
+		} else {
+			print_skipmessage(&resize_msg, &timestamp, buffer,
+					  bufspace, &req, false);
+		}
+	}
+	return NULL;
+}
+
+static void start_printthread(void)
+{
+	unsigned int i;
+	long *seed;
+	int ufd;
+
+	ufd = open(DEV_URANDOM, O_RDONLY);
+	if (nr_threads > MAX_THREADS) {
+		warnx(
+"Number of requested print threads was %d, max number is %d\n",
+		      nr_threads, MAX_THREADS);
+		nr_threads = MAX_THREADS;
+	}
+	for (i = 0; i < nr_threads; i++) {
+		seed = malloc_or_die(sizeof(*seed));
+		if (ufd <  0 ||
+		    read(ufd, seed, sizeof(*seed)) != sizeof(*seed)) {
+			printf(
+"Warning! Using trivial random nummer seed, since %s not available\n",
+			DEV_URANDOM);
+			fflush(stdout);
+			*seed = i;
+		}
+		if (pthread_create(&printthread[i], NULL, do_printloop, seed)
+		    != 0)
+			err(0, "pthread_create()");
+	}
+	if (ufd > 0)
+		close(ufd);
+}
+
+static void show_usage(void)
+{
+	printf(
+"Usage: %s [OPTION]...\n\n"
+"Collect closely occurring latencies from %s\n"
+"when any of the following tracers are enabled: preemptirqsoff, preemptoff,\n"
+"irqsoff, wakeup_dl, wakeup_rt, wakeup, or hwlat. A tracer can be enabled by\n"
+"doing something like this:\n\n"
+
+"echo 1000 > /sys/kernel/debug/tracing/tracing_thresh\n"
+"echo preemptirqsoff > /sys/kernel/debug/tracing/current_tracer\n\n"
+
+"The occurrence of a latency is detected by monitoring the file\n"
+"%s with inotify.\n\n"
+
+"The following options are supported:\n"
+"-c, --policy POL\tRun the program with scheduling policy POL. POL can be\n"
+"\t\t\tother, batch, idle, rr or fifo. The default is rr. When\n"
+"\t\t\tusing rr or fifo, remember that these policies may cause\n"
+"\t\t\tother tasks to experience latencies.\n\n"
+
+"-p, --priority PRI\tRun the program with priority PRI. The acceptable range\n"
+"\t\t\tof PRI depends on the scheduling policy.\n\n"
+
+"-n, --notrace\t\tIf latency is detected, do not print out the content of\n"
+"\t\t\tthe trace file to standard output\n\n"
+
+"-t, --threads NRTHR\tRun NRTHR threads for printing. Default is %d.\n\n"
+
+"-r, --random\t\tArbitrarily sleep a certain amount of time, default\n"
+"\t\t\t%ld ms, before reading the trace file. The\n"
+"\t\t\tprobabilities for sleep are chosen so that the\n"
+"\t\t\tprobability of obtaining any of a cluster of closely\n"
+"\t\t\toccurring latencies are equal, i.e. we will randomly\n"
+"\t\t\tchoose which one we collect from the trace file.\n\n"
+"\t\t\tThis option is probably only useful with the irqsoff,\n"
+"\t\t\tpreemptoff, and preemptirqsoff tracers.\n\n"
+
+"-a, --nrlat NRLAT\tFor the purpose of arbitrary delay, assume that there\n"
+"\t\t\tare no more than NRLAT clustered latencies. If NRLAT\n"
+"\t\t\tlatencies are detected during a run, this value will\n"
+"\t\t\tautomatically be increased to NRLAT + 1 and then to\n"
+"\t\t\tNRLAT + 2 and so on. The default is %d. This option\n"
+"\t\t\timplies -r. We need to know this number in order to\n"
+"\t\t\tbe able to calculate the probabilities of sleeping.\n"
+"\t\t\tSpecifically, the probabilities of not sleeping, i.e. to\n"
+"\t\t\tdo an immediate printout will be:\n\n"
+"\t\t\t1/NRLAT  1/(NRLAT - 1) ... 1/3  1/2  1\n\n"
+"\t\t\tThe probability of sleeping will be:\n\n"
+"\t\t\t1 - P, where P is from the series above\n\n"
+"\t\t\tThis descending probability will cause us to choose\n"
+"\t\t\tan occurrence at random. Observe that the final\n"
+"\t\t\tprobability is 0, it is when we reach this probability\n"
+"\t\t\tthat we increase NRLAT automatically. As an example,\n"
+"\t\t\twith the default value of 2, the probabilities will be:\n\n"
+"\t\t\t1/2  0\n\n"
+"\t\t\tThis means, when a latency is detected we will sleep\n"
+"\t\t\twith 50%% probability. If we ever detect another latency\n"
+"\t\t\tduring the sleep period, then the probability of sleep\n"
+"\t\t\twill be 0%% and the table will be expanded to:\n\n"
+"\t\t\t1/3  1/2  0\n\n"
+
+"-v, --verbose\t\tIncrease the verbosity. If this option is given once,\n"
+"\t\t\tthen print a message every time that the NRLAT value\n"
+"\t\t\tis automatically increased. If this option is given at\n"
+"\t\t\tleast twice, then also print a warning for lost events.\n\n"
+
+"-u, --time TIME\t\tArbitrarily sleep for a specified time TIME ms before\n"
+"\t\t\tprinting out the trace from the trace file. The default\n"
+"\t\t\tis %ld ms. This option implies -r.\n\n"
+
+"-f, --tracefile FILE\tUse FILE as trace file. The default is\n"
+"\t\t\t%s.\n\n"
+
+"-m, --max-lat FILE\tUse FILE as tracing_max_latency file. The default is\n"
+"\t\t\t%s\n\n"
+,
+prg_name, debug_tracefile_dflt, debug_maxlat_dflt, DEFAULT_NR_PRINTER_THREADS,
+SLEEP_TIME_MS_DEFAULT, DEFAULT_TABLE_SIZE, SLEEP_TIME_MS_DEFAULT,
+debug_tracefile_dflt, debug_maxlat_dflt);
+}
+
+static void scan_arguments(int argc, char *argv[])
+{
+	int c;
+	int option_idx = 0;
+
+	debug_tracefile = debug_tracefile_dflt;
+	debug_maxlat_file = debug_maxlat_dflt;
+
+	static struct option long_options[] = {
+		{ "policy",	required_argument,	0, 'c' },
+		{ "priority",	required_argument,	0, 'p' },
+		{ "help",	no_argument,		0, 'h' },
+		{ "notrace",	no_argument,		0, 'n' },
+		{ "random",	no_argument,		0, 'r' },
+		{ "nrlat",	required_argument,	0, 'a' },
+		{ "threads",	required_argument,	0, 't' },
+		{ "time",	required_argument,	0, 'u' },
+		{ "verbose",	no_argument,		0, 'v' },
+		{ "tracefile",	required_argument,	0, 'f' },
+		{ "max-lat",	required_argument,	0, 'm' },
+		{ 0,		0,			0,  0  }
+	};
+	const struct policy *p;
+	int max, min;
+	int value;
+
+	while (true) {
+		c = getopt_long(argc, argv, "c:p:hnra:t:u:vf:m:", long_options,
+				&option_idx);
+		if (c == -1)
+			break;
+
+		switch (c) {
+		case 'c':
+			p = policy_from_name(optarg);
+			if (p != NULL) {
+				sched_policy = p->policy;
+				sched_policy_set = true;
+				if (!sched_pri_set) {
+					sched_pri = p->default_pri;
+					sched_pri_set = true;
+				}
+			} else {
+				warnx("Unknown scheduling %s\n", optarg);
+				show_usage();
+				exit(0);
+			}
+			break;
+		case 'p':
+			sched_pri = atoi(optarg);
+			sched_pri_set = true;
+			break;
+		case 'h':
+			show_usage();
+			exit(0);
+			break;
+		case 'n':
+			trace_enable = false;
+			use_random_sleep = false;
+			break;
+		case 't':
+			value = atoi(optarg);
+			if (value > 0)
+				nr_threads = value;
+			else {
+				warnx("NRTHR must be > 0\n");
+				show_usage();
+				exit(0);
+			}
+			break;
+		case 'u':
+			value = atoi(optarg);
+			if (value < 0) {
+				warnx("TIME must be >= 0\n");
+				show_usage();
+				exit(0);
+			}
+			trace_enable = true;
+			use_random_sleep = true;
+			sleep_time = value * USEC_PER_MSEC;
+			break;
+		case 'v':
+			verbosity++;
+			break;
+		case 'r':
+			trace_enable = true;
+			use_random_sleep = true;
+			break;
+		case 'a':
+			value = atoi(optarg);
+			if (value <= 0) {
+				warnx("NRLAT must be > 0\n");
+				show_usage();
+				exit(0);
+			}
+			trace_enable = true;
+			use_random_sleep = true;
+			table_startsize = value;
+			break;
+		case 'f':
+			debug_tracefile = strdup(optarg);
+			break;
+		case 'm':
+			debug_maxlat_file = strdup(optarg);
+			break;
+		default:
+			show_usage();
+			exit(0);
+			break;
+		}
+	}
+
+	if (!sched_policy_set) {
+		sched_policy = SCHED_RR;
+		sched_policy_set = true;
+		if (!sched_pri_set) {
+			sched_pri = RT_DEFAULT_PRI;
+			sched_pri_set = true;
+		}
+	}
+
+	max = sched_get_priority_max(sched_policy);
+	min = sched_get_priority_min(sched_policy);
+
+	if (sched_pri < min) {
+		printf(
+"ATTENTION: Increasing priority to minimum, which is %d\n", min);
+		sched_pri = min;
+	}
+	if (sched_pri > max) {
+		printf(
+"ATTENTION: Reducing priority to maximum, which is %d\n", max);
+		sched_pri = max;
+	}
+}
+
+static void show_params(void)
+{
+	printf(
+		"Running with scheduling policy %s and priority %d. Using %d print threads.\n",
+		policy_name(sched_policy), sched_pri, nr_threads);
+	if (trace_enable) {
+		if (use_random_sleep) {
+			printf(
+"%s will be printed with random delay\n"
+"Start size of the probability table:\t\t\t%d\n"
+"Print a message when prob. table table changes size:\t%s\n"
+"Print a warning when an event has been lost:\t\t%s\n"
+"Sleep time is:\t\t\t\t\t\t%ld ms\n",
+debug_tracefile,
+table_startsize,
+bool2str(verbose_sizechange()),
+bool2str(verbose_lostevent()),
+sleep_time / USEC_PER_MSEC);
+		} else {
+			printf("%s will be printed immediately\n",
+			       debug_tracefile);
+		}
+	} else {
+		printf("%s will not be printed\n",
+		       debug_tracefile);
+	}
+}
+
+int main(int argc, char *argv[])
+{
+	open_stdout();
+
+	if (argc >= 1)
+		prg_name = argv[0];
+	else
+		prg_name = prg_unknown;
+
+	scan_arguments(argc, argv);
+	show_params();
+
+	init_printstate();
+	if (use_random_sleep) {
+		init_probabilities();
+		if (verbose_sizechange())
+			printf("Initializing probability table to %d\n",
+			       table_startsize);
+		sleeptable_resize(table_startsize, false, NULL);
+	}
+	set_priority();
+	init_queue(&printqueue);
+	start_printthread();
+	do_jittertest();
+
+	return 0;
+}
-- 
2.17.1

[PATCH v6 4/4] ftrace: Add an option for tracing console latencies

[Viktor Rosendahl]

This new trace option "console-latency" will enable the latency
tracers to trace the console latencies. Previously this has always been
implicitely disabled. I guess this is because they are considered
to be well known and unavoidable.

However, for some organizations it may nevertheless be desirable to
trace them. Basically, we want to be able to tell that there are
latencies in the system under test because someone has incorrectly
enabled the serial console.

Signed-off-by: Viktor Rosendahl <viktor.rosendahl@gmail.com>
---
 include/linux/irqflags.h     | 22 ++++++++++++++++++++++
 kernel/printk/printk.c       |  6 ++++--
 kernel/trace/trace.h         |  1 +
 kernel/trace/trace_irqsoff.c | 12 ++++++++++++
 4 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/include/linux/irqflags.h b/include/linux/irqflags.h
index 21619c92c377..3de891723331 100644
--- a/include/linux/irqflags.h
+++ b/include/linux/irqflags.h
@@ -13,6 +13,7 @@
 #define _LINUX_TRACE_IRQFLAGS_H
 
 #include <linux/typecheck.h>
+#include <linux/types.h>
 #include <asm/irqflags.h>
 
 /* Currently trace_softirqs_on/off is used only by lockdep */
@@ -68,9 +69,30 @@ do {						\
 	defined(CONFIG_PREEMPT_TRACER)
  extern void stop_critical_timings(void);
  extern void start_critical_timings(void);
+ extern bool console_tracing_disabled(void);
+
+# define console_stop_critical_timings(flag)		\
+	do {						\
+		typecheck(bool, flag);			\
+		flag = console_tracing_disabled();	\
+		if (flag)				\
+			stop_critical_timings();	\
+	} while (0)
+
+# define console_start_critical_timings(flag)		 \
+	do {						 \
+		typecheck(bool, flag);			 \
+		if (flag)				 \
+			start_critical_timings();	 \
+	} while (0)
+
 #else
 # define stop_critical_timings() do { } while (0)
 # define start_critical_timings() do { } while (0)
+# define console_stop_critical_timings(flag)	\
+	do { typecheck(bool, flag); } while (0)
+# define console_start_critical_timings(flag)	\
+	do { typecheck(bool, flag); } while (0)
 #endif
 
 /*
diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
index 1888f6a3b694..036460a7e4cd 100644
--- a/kernel/printk/printk.c
+++ b/kernel/printk/printk.c
@@ -2359,6 +2359,7 @@ void console_unlock(void)
 	static char ext_text[CONSOLE_EXT_LOG_MAX];
 	static char text[LOG_LINE_MAX + PREFIX_MAX];
 	unsigned long flags;
+	bool cflag;
 	bool do_cond_resched, retry;
 
 	if (console_suspended) {
@@ -2459,9 +2460,10 @@ void console_unlock(void)
 		 */
 		console_lock_spinning_enable();
 
-		stop_critical_timings();	/* don't trace print latency */
+		/* don't trace print latency if it's disabled */
+		console_stop_critical_timings(cflag);
 		call_console_drivers(ext_text, ext_len, text, len);
-		start_critical_timings();
+		console_start_critical_timings(cflag);
 
 		if (console_lock_spinning_disable_and_check()) {
 			printk_safe_exit_irqrestore(flags);
diff --git a/kernel/trace/trace.h b/kernel/trace/trace.h
index 4913ed1138aa..93a8b82c27e4 100644
--- a/kernel/trace/trace.h
+++ b/kernel/trace/trace.h
@@ -1262,6 +1262,7 @@ extern int trace_get_user(struct trace_parser *parser, const char __user *ubuf,
 		C(PRINTK_MSGONLY,	"printk-msg-only"),	\
 		C(CONTEXT_INFO,		"context-info"),   /* Print pid/cpu/time */ \
 		C(LATENCY_FMT,		"latency-format"),	\
+		C(CONSOLE_LATENCY,	"console-latency"),	\
 		C(RECORD_CMD,		"record-cmd"),		\
 		C(RECORD_TGID,		"record-tgid"),		\
 		C(OVERWRITE,		"overwrite"),		\
diff --git a/kernel/trace/trace_irqsoff.c b/kernel/trace/trace_irqsoff.c
index a745b0cee5d3..bc02be207a7a 100644
--- a/kernel/trace/trace_irqsoff.c
+++ b/kernel/trace/trace_irqsoff.c
@@ -456,6 +456,18 @@ void stop_critical_timings(void)
 EXPORT_SYMBOL_GPL(stop_critical_timings);
 NOKPROBE_SYMBOL(stop_critical_timings);
 
+bool console_tracing_disabled(void)
+{
+	struct trace_array *tr = irqsoff_trace;
+	int pc = preempt_count();
+
+	if (!preempt_trace(pc) && !irq_trace())
+		return false;
+
+	return !(tr->trace_flags & TRACE_ITER_CONSOLE_LATENCY);
+}
+EXPORT_SYMBOL_GPL(console_tracing_disabled);
+
 #ifdef CONFIG_FUNCTION_TRACER
 static bool function_enabled;
 
-- 
2.17.1

Re: [PATCH v6 1/4] ftrace: Implement fs notification for tracing_max_latency

[Joel Fernandes]

On Thu, Sep 05, 2019 at 03:25:45PM +0200, Viktor Rosendahl wrote:
> This patch implements the feature that the tracing_max_latency file,
> e.g. /sys/kernel/debug/tracing/tracing_max_latency will receive
> notifications through the fsnotify framework when a new latency is
> available.
> 
> One particularly interesting use of this facility is when enabling
> threshold tracing, through /sys/kernel/debug/tracing/tracing_thresh,
> together with the preempt/irqsoff tracers. This makes it possible to
> implement a user space program that can, with equal probability,
> obtain traces of latencies that occur immediately after each other in
> spite of the fact that the preempt/irqsoff tracers operate in overwrite
> mode.
> 
> This facility works with the hwlat, preempt/irqsoff, and wakeup
> tracers.
> 
> The tracers may call the latency_fsnotify() from places such as
> __schedule() or do_idle(); this makes it impossible to call
> queue_work() directly without risking a deadlock. The same would
> happen with a softirq,  kernel thread or tasklet. For this reason we
> use the irq_work mechanism to call queue_work().
> Signed-off-by: Viktor Rosendahl <viktor.rosendahl@gmail.com>

Looks better now, one comment below:

> ---
>  kernel/trace/trace.c       | 75 +++++++++++++++++++++++++++++++++++++-
>  kernel/trace/trace.h       | 18 +++++++++
>  kernel/trace/trace_hwlat.c |  4 +-
>  3 files changed, 94 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> index 563e80f9006a..72ac20c4aaa1 100644
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -44,6 +44,9 @@
>  #include <linux/trace.h>
>  #include <linux/sched/clock.h>
>  #include <linux/sched/rt.h>
> +#include <linux/fsnotify.h>
> +#include <linux/irq_work.h>
> +#include <linux/workqueue.h>
>  
>  #include "trace.h"
>  #include "trace_output.h"
> @@ -1480,6 +1483,74 @@ static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt)
>  
>  unsigned long __read_mostly	tracing_thresh;
>  
> +#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
> +	defined(CONFIG_FSNOTIFY)
> +
> +static const struct file_operations tracing_max_lat_fops;
> +static struct workqueue_struct *fsnotify_wq;
> +
> +static void latency_fsnotify_workfn(struct work_struct *work)
> +{
> +	struct trace_array *tr = container_of(work, struct trace_array,
> +					      fsnotify_work);
> +	fsnotify(tr->d_max_latency->d_inode, FS_MODIFY,
> +		 tr->d_max_latency->d_inode, FSNOTIFY_EVENT_INODE, NULL, 0);
> +}
> +
> +static void latency_fsnotify_workfn_irq(struct irq_work *iwork)
> +{
> +	struct trace_array *tr = container_of(iwork, struct trace_array,
> +					      fsnotify_irqwork);
> +	queue_work(fsnotify_wq, &tr->fsnotify_work);
> +}
> +
> +static void trace_create_maxlat_file(struct trace_array *tr,
> +				     struct dentry *d_tracer)
> +{
> +	INIT_WORK(&tr->fsnotify_work, latency_fsnotify_workfn);
> +	init_irq_work(&tr->fsnotify_irqwork, latency_fsnotify_workfn_irq);
> +	tr->d_max_latency = trace_create_file("tracing_max_latency", 0644,
> +					      d_tracer, &tr->max_latency,
> +					      &tracing_max_lat_fops);
> +}
> +
> +__init static int latency_fsnotify_init(void)
> +{
> +	fsnotify_wq = alloc_workqueue("tr_max_lat_wq",
> +				      WQ_UNBOUND | WQ_HIGHPRI, 0);
> +	if (!fsnotify_wq) {
> +		pr_err("Unable to allocate tr_max_lat_wq\n");
> +		return -ENOMEM;
> +	}

Why not just use the system workqueue instead of adding another workqueue?

thanks,

 - Joel
 

Re: [PATCH v6 1/4] ftrace: Implement fs notification for tracing_max_latency

[Viktor Rosendahl]

On 9/6/19 4:17 PM, Joel Fernandes wrote:
> On Thu, Sep 05, 2019 at 03:25:45PM +0200, Viktor Rosendahl wrote:
<clip>
>> +
>> +__init static int latency_fsnotify_init(void)
>> +{
>> +	fsnotify_wq = alloc_workqueue("tr_max_lat_wq",
>> +				      WQ_UNBOUND | WQ_HIGHPRI, 0);
>> +	if (!fsnotify_wq) {
>> +		pr_err("Unable to allocate tr_max_lat_wq\n");
>> +		return -ENOMEM;
>> +	}
> 
> Why not just use the system workqueue instead of adding another workqueue?
> 

For the the latency-collector to work properly in the worst case, when a
new latency occurs immediately, the fsnotify must be received in less
time than what the threshold is set to. If we always are slower we will
always lose certain latencies.

My intention was to minimize latency in some important cases, so that
user space receives the notification sooner rather than later.

There doesn't seem to be any system workqueue with WQ_UNBOUND and
WQ_HIGHPRI. My thinking was that WQ_UNBOUND might help with the latency
in some important cases.

If we use:

queue_work(system_highpri_wq, &tr->fsnotify_work);

then the work will (almost) always execute on the same CPU but if we are
unlucky that CPU could be too busy while there could be another CPU in
the system that would be able to process the work soon enough.

queue_work_on() could be used to queue the work on another CPU but it
seems difficult to select the right CPU.

best regards,

Viktor

Re: [PATCH v6 1/4] ftrace: Implement fs notification for tracing_max_latency

[Joel Fernandes]

On Sat, Sep 07, 2019 at 11:12:59PM +0200, Viktor Rosendahl wrote:
> On 9/6/19 4:17 PM, Joel Fernandes wrote:
> > On Thu, Sep 05, 2019 at 03:25:45PM +0200, Viktor Rosendahl wrote:
> <clip>
> > > +
> > > +__init static int latency_fsnotify_init(void)
> > > +{
> > > +	fsnotify_wq = alloc_workqueue("tr_max_lat_wq",
> > > +				      WQ_UNBOUND | WQ_HIGHPRI, 0);
> > > +	if (!fsnotify_wq) {
> > > +		pr_err("Unable to allocate tr_max_lat_wq\n");
> > > +		return -ENOMEM;
> > > +	}
> > 
> > Why not just use the system workqueue instead of adding another workqueue?
> > 
> 
> For the the latency-collector to work properly in the worst case, when a
> new latency occurs immediately, the fsnotify must be received in less
> time than what the threshold is set to. If we always are slower we will
> always lose certain latencies.
> 
> My intention was to minimize latency in some important cases, so that
> user space receives the notification sooner rather than later.
> 
> There doesn't seem to be any system workqueue with WQ_UNBOUND and
> WQ_HIGHPRI. My thinking was that WQ_UNBOUND might help with the latency
> in some important cases.
> 
> If we use:
> 
> queue_work(system_highpri_wq, &tr->fsnotify_work);
> 
> then the work will (almost) always execute on the same CPU but if we are
> unlucky that CPU could be too busy while there could be another CPU in
> the system that would be able to process the work soon enough.
> 
> queue_work_on() could be used to queue the work on another CPU but it
> seems difficult to select the right CPU.

Ok, a separate WQ is fine with me as such since the preempt/irq events are on
a debug kernel anyway.

I'll keep reviewing your patches next few days, I am at the LPC conference so
might be a bit slow. Overall I think the series look like its maturing and
getting close.

thanks,

 - Joel

Re: [PATCH v6 1/4] ftrace: Implement fs notification for tracing_max_latency

[Viktor Rosendahl]

On 9/8/19 1:38 AM, Joel Fernandes wrote:> On Sat, Sep 07, 2019 at 
11:12:59PM +0200, Viktor Rosendahl wrote:
 >> On 9/6/19 4:17 PM, Joel Fernandes wrote:
 >>> On Thu, Sep 05, 2019 at 03:25:45PM +0200, Viktor Rosendahl wrote:
 >> <clip>
 >>>> +
 >>>> +__init static int latency_fsnotify_init(void)
 >>>> +{
 >>>> +	fsnotify_wq = alloc_workqueue("tr_max_lat_wq",
 >>>> +				      WQ_UNBOUND | WQ_HIGHPRI, 0);
 >>>> +	if (!fsnotify_wq) {
 >>>> +		pr_err("Unable to allocate tr_max_lat_wq\n");
 >>>> +		return -ENOMEM;
 >>>> +	}
 >>>
 >>> Why not just use the system workqueue instead of adding another 
workqueue?
 >>>
 >>
 >> For the the latency-collector to work properly in the worst case, when a
 >> new latency occurs immediately, the fsnotify must be received in less
 >> time than what the threshold is set to. If we always are slower we will
 >> always lose certain latencies.
 >>
 >> My intention was to minimize latency in some important cases, so that
 >> user space receives the notification sooner rather than later.
 >>
 >> There doesn't seem to be any system workqueue with WQ_UNBOUND and
 >> WQ_HIGHPRI. My thinking was that WQ_UNBOUND might help with the latency
 >> in some important cases.
 >>
 >> If we use:
 >>
 >> queue_work(system_highpri_wq, &tr->fsnotify_work);
 >>
 >> then the work will (almost) always execute on the same CPU but if we are
 >> unlucky that CPU could be too busy while there could be another CPU in
 >> the system that would be able to process the work soon enough.
 >>
 >> queue_work_on() could be used to queue the work on another CPU but it
 >> seems difficult to select the right CPU.
 >
 > Ok, a separate WQ is fine with me as such since the preempt/irq 
events are on
 > a debug kernel anyway.

Keep in mind that this feature is also enabled by the wakeup tracers and by
hwlat. These are often enabled by production kernels.

I guess it would be possible to add some ifdefs so that we create a new
workqueue only if preempt/irqsoff tracing is enabled in the kernel 
config and
use system_highpri_wq if we only have the wakeup and hwlat tracers in the
config.

However, I don't really like adding yet some more ifdefs to the code.

Since a new workqueue will not necessariy create a new worker thread 
nowadays,
I thought that it would be OK with a new unbound workqueue, which should not
add much to the tendency to create more worker threads.

 >
 > I'll keep reviewing your patches next few days, I am at the LPC 
conference so
 > might be a bit slow. Overall I think the series look like its 
maturing and
 > getting close.
 >

Ok, thanks. Could you let me know when you have looked through it all so 
that
I know when it makes sense to make another reroll of the series?

best regards,

Viktor