Skip to contentSkip to conversation

Some new features for the preempt/irqsoff tracers

PATCH Superseded Viktor Rosendahl (BMW) 3 participants 8 messages Oct 15, 2019 – Oct 16, 2019

Subsystems (4)

v9 of 10 Superseded Latest is v10 →

lore
Share & tools
lore
#0 lore raw

v9 cover Some new features for the preempt/irqsoff tracers

Hello all,

Changes in v9:

- [PATCH 1/4]:
* Rebased the patch for v5.4-rc3+
* Dropped support for the hwlat tracer. It seems to me, both from testing and looking at the code, that hwlat doesn't actually update /sys/kernel/debug/tracing/tracing_max_latency. I think it would be a bit weird to send fsnotify for changes to a file that are not visible to userspace. Also, the hwlat use case is not very interesting from my perspective because hwlat works in non-overwrite mode by default. For me, the effort needed to make it work properly would be more than what it's worth.
* Because of the changes mentioned above, I removed the Reviewed-by tag.

- [PATCH 2/4]:
* No change.

- [PATCH 3/4]:
* Removed the mention of hwlat from the commit and help messages.

- [PACTH 4/4]:
* Changed the trace-option to a kernel parameter. This would hopefully make it easier to remove it if it becomes obsolete.

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 (BMW) (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

 .../admin-guide/kernel-parameters.txt         |    4 +
 include/linux/irqflags.h                      |   22 +
 kernel/printk/printk.c                        |    6 +-
 kernel/trace/Kconfig                          |    6 +-
 kernel/trace/preemptirq_delay_test.c          |  144 +-
 kernel/trace/trace.c                          |   75 +-
 kernel/trace/trace.h                          |   18 +
 kernel/trace/trace_irqsoff.c                  |   32 +
 tools/Makefile                                |   14 +-
 tools/trace/Makefile                          |   20 +
 tools/trace/latency-collector.c               | 1212 +++++++++++++++++
 11 files changed, 1521 insertions(+), 32 deletions(-)
 create mode 100644 tools/trace/Makefile
 create mode 100644 tools/trace/latency-collector.c
Signature · 1 line
2.17.1
#1 lore raw

PATCH v9 1/4 ftrace: Implement fs notification for tracing_max_latency

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 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().

This patch creates a new workqueue. The reason for doing this is that I wanted to use the WQ_UNBOUND and WQ_HIGHPRI flags. 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.

---
 kernel/trace/trace.c | 75 ++++++++++++++++++++++++++++++++++++++++++--
 kernel/trace/trace.h | 18 +++++++++++
 2 files changed, 91 insertions(+), 2 deletions(-)
kernel/trace/trace.c +73 -2
index 6a0ee9178365..523e3e57f1d4 100644
@@ -45,6 +45,9 @@
 #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"
@@ -1497,6 +1500,74 @@ static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt)
 }
 
 unsigned long __read_mostly	tracing_thresh;
+static const struct file_operations tracing_max_lat_fops;
+
+#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
+	defined(CONFIG_FSNOTIFY)
+
+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
 /*
@@ -1536,6 +1607,7 @@ __update_max_tr(struct trace_array *tr, struct task_struct *tsk, int cpu)
 
 	/* record this tasks comm */
 	tracing_record_cmdline(tsk);
+	latency_fsnotify(tr);
 }
 
 /**
@@ -8585,8 +8657,7 @@ 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))
kernel/trace/trace.h +18 -0
index d685c61085c0..2fb9b7353a99 100644
@@ -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;
 	/*
@@ -786,6 +793,17 @@ 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
+
+static void latency_fsnotify(struct trace_array *tr) { }
+
+#endif
+
 #ifdef CONFIG_STACKTRACE
 void __trace_stack(struct trace_array *tr, unsigned long flags, int skip,
 		   int pc);
Signature · 1 line
2.17.1
#2 lore raw

Re: [PATCH v9 1/4] ftrace: Implement fs notification for tracing_max_latency

Quoted 57 lines · Viktor Rosendahl (BMW)

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 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().

This patch creates a new workqueue. The reason for doing this is that I wanted to use the WQ_UNBOUND and WQ_HIGHPRI flags. 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.

---
 kernel/trace/trace.c | 75 ++++++++++++++++++++++++++++++++++++++++++--
 kernel/trace/trace.h | 18 +++++++++++
 2 files changed, 91 insertions(+), 2 deletions(-)
kernel/trace/trace.c +3 -0
index 6a0ee9178365..523e3e57f1d4 100644
@@ -45,6 +45,9 @@
 #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"

kernel/trace/trace.c static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt) · line 1506

Quoted 7 lines
kernel/trace/trace.c +4 -0
@@ -1497,6 +1500,74 @@ static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt)
 }
 
 unsigned long __read_mostly	tracing_thresh;
+static const struct file_operations tracing_max_lat_fops;
+
+#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
+	defined(CONFIG_FSNOTIFY)

Something bothers me. If you dropped support for HWLAT_TRACER as you mentioned in the cover letter, then why does this #if look for the CONFIG option?

(and similar comment on the rest of the patch..)

thanks,

- Joel

Quoted 131 lines
kernel/trace/trace.c +66 -2
+
+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
 /*
@@ -1536,6 +1607,7 @@ __update_max_tr(struct trace_array *tr, struct task_struct *tsk, int cpu)
 
 	/* record this tasks comm */
 	tracing_record_cmdline(tsk);
+	latency_fsnotify(tr);
 }
 
 /**
@@ -8585,8 +8657,7 @@ 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))
kernel/trace/trace.h +18 -0
index d685c61085c0..2fb9b7353a99 100644
@@ -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;
 	/*
@@ -786,6 +793,17 @@ 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
+
+static void latency_fsnotify(struct trace_array *tr) { }
+
+#endif
+
 #ifdef CONFIG_STACKTRACE
 void __trace_stack(struct trace_array *tr, unsigned long flags, int skip,
 		   int pc);
Signature · 1 line
2.17.1
maintainer
#3 lore raw

Re: [PATCH v9 1/4] ftrace: Implement fs notification for tracing_max_latency

Quoted 10 lines · On Tue, 15 Oct 2019 20:41:20 -0400 Joel Fernandes <[email protected]>…

kernel/trace/trace.c static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt) · line 1506

Quoted 4 lines
kernel/trace/trace.c +4 -0
+static const struct file_operations tracing_max_lat_fops;
+
+#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
+	defined(CONFIG_FSNOTIFY)  

Something bothers me. If you dropped support for HWLAT_TRACER as you mentioned in the cover letter, then why does this #if look for the CONFIG option?

(and similar comment on the rest of the patch..)

Also, if you have a complex if statement like this, it is better to create a new define to set it in a header file:

#if (defined(CONFIG_TRACER_MAX_TRACE) || defined(CONFIG_HWLAT_TRACER)) && \
defined(CONFIG_FSNOTIFY)

# define HAVE_TRACER_FSNOTIFY
#endif

And then just use that:

#ifdef HAVE_TRACER_FSNOTIFY
...
#endif

It keeps things a bit more manageable.

-- Steve

#4 lore raw

Re: [PATCH v9 1/4] ftrace: Implement fs notification for tracing_max_latency

Quoted 3 lines

Something bothers me. If you dropped support for HWLAT_TRACER as you mentioned in the cover letter, then why does this #if look for the CONFIG option?

You are right. I forgot to change those #if statements; I also overlooked to give the -i option to grep when I searched for remaining references to hwlat.

Quoted 1 line

(and similar comment on the rest of the patch..)

Below is a new version of the patch. I will not send out v10 yet, as I expect that there might be more comments.

I did not define a new macro, since it is now much simpler after removing the HWLAT_TRACER stuff from those #if statements.

I made the latency_fsnotify() static and removed its declaration from kernel/trace/trace.h, since it's only used in kernel/trace/trace.c.

best regards,

Viktor

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 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().

This patch creates a new workqueue. The reason for doing this is that I wanted to use the WQ_UNBOUND and WQ_HIGHPRI flags. 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.

---
 kernel/trace/trace.c | 73 ++++++++++++++++++++++++++++++++++++++++++--
 kernel/trace/trace.h |  7 +++++
 2 files changed, 78 insertions(+), 2 deletions(-)
kernel/trace/trace.c +71 -2
index 6a0ee9178365..2124646dfb2a 100644
@@ -45,6 +45,9 @@
 #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"
@@ -1497,6 +1500,70 @@ static ssize_t trace_seq_to_buffer(struct trace_seq *s, void *buf, size_t cnt)
 }
 
 unsigned long __read_mostly	tracing_thresh;
+static const struct file_operations tracing_max_lat_fops;
+
+#if defined(CONFIG_TRACER_MAX_TRACE) && defined(CONFIG_FSNOTIFY)
+
+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);
+
+static 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);
+}
+
+
+#else /* defined(CONFIG_TRACER_MAX_TRACE) && defined(CONFIG_FSNOTIFY) */
+
+#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
 /*
@@ -1536,6 +1603,9 @@ __update_max_tr(struct trace_array *tr, struct task_struct *tsk, int cpu)
 
 	/* record this tasks comm */
 	tracing_record_cmdline(tsk);
+#ifdef CONFIG_FSNOTIFY
+	latency_fsnotify(tr);
+#endif
 }
 
 /**
@@ -8585,8 +8655,7 @@ 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))
kernel/trace/trace.h +7 -0
index d685c61085c0..8564c72ea7b5 100644
@@ -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;
+#endif
+#if defined(CONFIG_TRACER_MAX_TRACE) && defined(CONFIG_FSNOTIFY)
+	struct dentry		*d_max_latency;
+	struct work_struct	fsnotify_work;
+	struct irq_work		fsnotify_irqwork;
 #endif
 	struct trace_pid_list	__rcu *filtered_pids;
 	/*
Signature · 1 line
2.17.1
#5 lore raw

PATCH v9 2/4 preemptirq_delay_test: Add the burst feature and a sysfs trigger

In reply to Viktor Rosendahl (BMW)

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.

---
 kernel/trace/Kconfig                 |   6 +-
 kernel/trace/preemptirq_delay_test.c | 144 +++++++++++++++++++++++----
 2 files changed, 128 insertions(+), 22 deletions(-)
kernel/trace/Kconfig +3 -3
index e08527f50d2a..2a58380ea310 100644
@@ -752,9 +752,9 @@ 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
 
kernel/trace/preemptirq_delay_test.c +125 -19

180 lines not on this page · raw message

Signature · 1 line
2.17.1
#6 lore raw

PATCH v9 3/4 Add the latency-collector to tools

In reply to Viktor Rosendahl (BMW)

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, and wakeup 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
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

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

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

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.

---
 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
tools/Makefile +8 -6
index 7e42f7b8bfa7..957762a8b811 100644
@@ -30,6 +30,7 @@ help:
 	@echo '  selftests              - various kernel selftests'
 	@echo '  spi                    - spi tools'
 	@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
tools/trace/Makefile new file +20 -0
new file mode 100644
index 000000000000..59cd483ab01f
@@ -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)
tools/trace/latency-collector.c new file +1212 -0

1,212 lines not on this page · raw message

Signature · 1 line
2.17.1
#7 lore raw

PATCH v9 4/4 ftrace: Add an option for tracing console latencies

In reply to Viktor Rosendahl (BMW)

This new kernel parameter "trace_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.

---
 .../admin-guide/kernel-parameters.txt         |  4 +++
 include/linux/irqflags.h                      | 22 +++++++++++++
 kernel/printk/printk.c                        |  6 ++--
 kernel/trace/trace_irqsoff.c                  | 32 +++++++++++++++++++
 4 files changed, 62 insertions(+), 2 deletions(-)
Documentation/admin-guide/kernel-parameters.txt +4 -0
index a84a83f8881e..3f96a380c528 100644
@@ -4768,6 +4768,10 @@
 	trace_buf_size=nn[KMG]
 			[FTRACE] will set tracing buffer size on each cpu.
 
+	trace_console_latency=[0|1]
+			[FTRACE] Trace the console latencies if 1 is given. Do
+			not trace the latencies if 0 or no parameter is given.
+
 	trace_event=[event-list]
 			[FTRACE] Set and start specified trace events in order
 			to facilitate early boot debugging. The event-list is a
include/linux/irqflags.h +22 -0
index 21619c92c377..3de891723331 100644
@@ -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
 
 /*
kernel/printk/printk.c +4 -2
index ca65327a6de8..f27e96273453 100644
@@ -2369,6 +2369,7 @@ 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) {
@@ -2469,9 +2470,10 @@ 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);
kernel/trace/trace_irqsoff.c +32 -0
index a745b0cee5d3..84876192cc73 100644
@@ -14,6 +14,7 @@
 #include <linux/uaccess.h>
 #include <linux/module.h>
 #include <linux/ftrace.h>
+#include <linux/init.h>
 #include <linux/kprobes.h>
 
 #include "trace.h"
@@ -456,6 +457,37 @@ void stop_critical_timings(void)
 EXPORT_SYMBOL_GPL(stop_critical_timings);
 NOKPROBE_SYMBOL(stop_critical_timings);
 
+static bool no_console_latency __read_mostly = true;
+
+static int __init trace_console_latency(char *str)
+{
+	int param;
+
+	get_option(&str, &param);
+	no_console_latency = !(param == 1);
+
+	return 0;
+}
+
+early_param("trace_console_latency", trace_console_latency);
+
+bool console_tracing_disabled(void)
+{
+	int pc = preempt_count();
+
+	/*
+	 * If tracing is disabled, then the question of whether to trace console
+	 * latencies is moot. By always returning false here we save the caller
+	 * the calls to start/stop_critical_timings(). These calls would not do
+	 * anything anyway.
+	 */
+	if (!preempt_trace(pc) && !irq_trace())
+		return false;
+
+	return no_console_latency;
+}
+EXPORT_SYMBOL_GPL(console_tracing_disabled);
+
 #ifdef CONFIG_FUNCTION_TRACER
 static bool function_enabled;
 
Signature · 1 line
2.17.1