trace_hwlat.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * trace_hwlat.c - A simple Hardware Latency detector.
  4 *
  5 * Use this tracer to detect large system latencies induced by the behavior of
  6 * certain underlying system hardware or firmware, independent of Linux itself.
  7 * The code was developed originally to detect the presence of SMIs on Intel
  8 * and AMD systems, although there is no dependency upon x86 herein.
  9 *
 10 * The classical example usage of this tracer is in detecting the presence of
 11 * SMIs or System Management Interrupts on Intel and AMD systems. An SMI is a
 12 * somewhat special form of hardware interrupt spawned from earlier CPU debug
 13 * modes in which the (BIOS/EFI/etc.) firmware arranges for the South Bridge
 14 * LPC (or other device) to generate a special interrupt under certain
 15 * circumstances, for example, upon expiration of a special SMI timer device,
 16 * due to certain external thermal readings, on certain I/O address accesses,
 17 * and other situations. An SMI hits a special CPU pin, triggers a special
 18 * SMI mode (complete with special memory map), and the OS is unaware.
 19 *
 20 * Although certain hardware-inducing latencies are necessary (for example,
 21 * a modern system often requires an SMI handler for correct thermal control
 22 * and remote management) they can wreak havoc upon any OS-level performance
 23 * guarantees toward low-latency, especially when the OS is not even made
 24 * aware of the presence of these interrupts. For this reason, we need a
 25 * somewhat brute force mechanism to detect these interrupts. In this case,
 26 * we do it by hogging all of the CPU(s) for configurable timer intervals,
 27 * sampling the built-in CPU timer, looking for discontiguous readings.
 28 *
 29 * WARNING: This implementation necessarily introduces latencies. Therefore,
 30 *          you should NEVER use this tracer while running in a production
 31 *          environment requiring any kind of low-latency performance
 32 *          guarantee(s).
 33 *
 34 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
 35 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
 36 *
 37 * Includes useful feedback from Clark Williams <williams@redhat.com>
 38 *
 39 */
 40#include <linux/kthread.h>
 41#include <linux/tracefs.h>
 42#include <linux/uaccess.h>
 43#include <linux/cpumask.h>
 44#include <linux/delay.h>
 45#include <linux/sched/clock.h>
 46#include "trace.h"
 47
 48static struct trace_array	*hwlat_trace;
 49
 50#define U64STR_SIZE		22			/* 20 digits max */
 51
 52#define BANNER			"hwlat_detector: "
 53#define DEFAULT_SAMPLE_WINDOW	1000000			/* 1s */
 54#define DEFAULT_SAMPLE_WIDTH	500000			/* 0.5s */
 55#define DEFAULT_LAT_THRESHOLD	10			/* 10us */
 56
 57static struct dentry *hwlat_sample_width;	/* sample width us */
 58static struct dentry *hwlat_sample_window;	/* sample window us */
 59static struct dentry *hwlat_thread_mode;	/* hwlat thread mode */
 60
 61enum {
 62	MODE_NONE = 0,
 63	MODE_ROUND_ROBIN,
 64	MODE_PER_CPU,
 65	MODE_MAX
 66};
 67static char *thread_mode_str[] = { "none", "round-robin", "per-cpu" };
 68
 69/* Save the previous tracing_thresh value */
 70static unsigned long save_tracing_thresh;
 71
 72/* runtime kthread data */
 73struct hwlat_kthread_data {
 74	struct task_struct	*kthread;
 75	/* NMI timestamp counters */
 76	u64			nmi_ts_start;
 77	u64			nmi_total_ts;
 78	int			nmi_count;
 79	int			nmi_cpu;
 80};
 81
 82static struct hwlat_kthread_data hwlat_single_cpu_data;
 83static DEFINE_PER_CPU(struct hwlat_kthread_data, hwlat_per_cpu_data);
 84
 85/* Tells NMIs to call back to the hwlat tracer to record timestamps */
 86bool trace_hwlat_callback_enabled;
 87
 88/* If the user changed threshold, remember it */
 89static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
 90
 91/* Individual latency samples are stored here when detected. */
 92struct hwlat_sample {
 93	u64			seqnum;		/* unique sequence */
 94	u64			duration;	/* delta */
 95	u64			outer_duration;	/* delta (outer loop) */
 96	u64			nmi_total_ts;	/* Total time spent in NMIs */
 97	struct timespec64	timestamp;	/* wall time */
 98	int			nmi_count;	/* # NMIs during this sample */
 99	int			count;		/* # of iterations over thresh */
100};
101
102/* keep the global state somewhere. */
103static struct hwlat_data {
104
105	struct mutex	lock;		/* protect changes */
106
107	atomic64_t	count;		/* total since reset */
108
109	u64	sample_window;		/* total sampling window (on+off) */
110	u64	sample_width;		/* active sampling portion of window */
111
112	int	thread_mode;		/* thread mode */
113
114} hwlat_data = {
115	.sample_window		= DEFAULT_SAMPLE_WINDOW,
116	.sample_width		= DEFAULT_SAMPLE_WIDTH,
117	.thread_mode		= MODE_ROUND_ROBIN
118};
119
120static struct hwlat_kthread_data *get_cpu_data(void)
121{
122	if (hwlat_data.thread_mode == MODE_PER_CPU)
123		return this_cpu_ptr(&hwlat_per_cpu_data);
124	else
125		return &hwlat_single_cpu_data;
126}
127
128static bool hwlat_busy;
129
130static void trace_hwlat_sample(struct hwlat_sample *sample)
131{
132	struct trace_array *tr = hwlat_trace;
133	struct trace_buffer *buffer = tr->array_buffer.buffer;
134	struct ring_buffer_event *event;
135	struct hwlat_entry *entry;
136
137	event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
138					  tracing_gen_ctx());
139	if (!event)
140		return;
141	entry	= ring_buffer_event_data(event);
142	entry->seqnum			= sample->seqnum;
143	entry->duration			= sample->duration;
144	entry->outer_duration		= sample->outer_duration;
145	entry->timestamp		= sample->timestamp;
146	entry->nmi_total_ts		= sample->nmi_total_ts;
147	entry->nmi_count		= sample->nmi_count;
148	entry->count			= sample->count;
149
150	trace_buffer_unlock_commit_nostack(buffer, event);
151}
152
153/* Macros to encapsulate the time capturing infrastructure */
154#define time_type	u64
155#define time_get()	trace_clock_local()
156#define time_to_us(x)	div_u64(x, 1000)
157#define time_sub(a, b)	((a) - (b))
158#define init_time(a, b)	(a = b)
159#define time_u64(a)	a
160
161void trace_hwlat_callback(bool enter)
162{
163	struct hwlat_kthread_data *kdata = get_cpu_data();
164
165	if (!kdata->kthread)
166		return;
167
168	/*
169	 * Currently trace_clock_local() calls sched_clock() and the
170	 * generic version is not NMI safe.
171	 */
172	if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
173		if (enter)
174			kdata->nmi_ts_start = time_get();
175		else
176			kdata->nmi_total_ts += time_get() - kdata->nmi_ts_start;
177	}
178
179	if (enter)
180		kdata->nmi_count++;
181}
182
183/*
184 * hwlat_err - report a hwlat error.
185 */
186#define hwlat_err(msg) ({							\
187	struct trace_array *tr = hwlat_trace;					\
188										\
189	trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg);	\
190})
191
192/**
193 * get_sample - sample the CPU TSC and look for likely hardware latencies
194 *
195 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
196 * hardware-induced latency. Called with interrupts disabled.
197 */
198static int get_sample(void)
199{
200	struct hwlat_kthread_data *kdata = get_cpu_data();
201	struct trace_array *tr = hwlat_trace;
202	struct hwlat_sample s;
203	time_type start, t1, t2, last_t2;
204	s64 diff, outer_diff, total, last_total = 0;
205	u64 sample = 0;
206	u64 sample_width = READ_ONCE(hwlat_data.sample_width);
207	u64 thresh = tracing_thresh;
208	u64 outer_sample = 0;
209	int ret = -1;
210	unsigned int count = 0;
211
212	do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
213
214	kdata->nmi_total_ts = 0;
215	kdata->nmi_count = 0;
216	/* Make sure NMIs see this first */
217	barrier();
218
219	trace_hwlat_callback_enabled = true;
220
221	init_time(last_t2, 0);
222	start = time_get(); /* start timestamp */
223	outer_diff = 0;
224
225	do {
226
227		t1 = time_get();	/* we'll look for a discontinuity */
228		t2 = time_get();
229
230		if (time_u64(last_t2)) {
231			/* Check the delta from outer loop (t2 to next t1) */
232			outer_diff = time_to_us(time_sub(t1, last_t2));
233			/* This shouldn't happen */
234			if (outer_diff < 0) {
235				hwlat_err(BANNER "time running backwards\n");
236				goto out;
237			}
238			if (outer_diff > outer_sample)
239				outer_sample = outer_diff;
240		}
241		last_t2 = t2;
242
243		total = time_to_us(time_sub(t2, start)); /* sample width */
244
245		/* Check for possible overflows */
246		if (total < last_total) {
247			hwlat_err("Time total overflowed\n");
248			break;
249		}
250		last_total = total;
251
252		/* This checks the inner loop (t1 to t2) */
253		diff = time_to_us(time_sub(t2, t1));     /* current diff */
254
255		if (diff > thresh || outer_diff > thresh) {
256			if (!count)
257				ktime_get_real_ts64(&s.timestamp);
258			count++;
259		}
260
261		/* This shouldn't happen */
262		if (diff < 0) {
263			hwlat_err(BANNER "time running backwards\n");
264			goto out;
265		}
266
267		if (diff > sample)
268			sample = diff; /* only want highest value */
269
270	} while (total <= sample_width);
271
272	barrier(); /* finish the above in the view for NMIs */
273	trace_hwlat_callback_enabled = false;
274	barrier(); /* Make sure nmi_total_ts is no longer updated */
275
276	ret = 0;
277
278	/* If we exceed the threshold value, we have found a hardware latency */
279	if (sample > thresh || outer_sample > thresh) {
280		u64 latency;
281
282		ret = 1;
283
284		/* We read in microseconds */
285		if (kdata->nmi_total_ts)
286			do_div(kdata->nmi_total_ts, NSEC_PER_USEC);
287
288		s.seqnum = atomic64_inc_return(&hwlat_data.count);
289		s.duration = sample;
290		s.outer_duration = outer_sample;
291		s.nmi_total_ts = kdata->nmi_total_ts;
292		s.nmi_count = kdata->nmi_count;
293		s.count = count;
294		trace_hwlat_sample(&s);
295
296		latency = max(sample, outer_sample);
297
298		/* Keep a running maximum ever recorded hardware latency */
299		if (latency > tr->max_latency) {
300			tr->max_latency = latency;
301			latency_fsnotify(tr);
302		}
303	}
304
305out:
306	return ret;
307}
308
309static struct cpumask save_cpumask;
310
311static void move_to_next_cpu(void)
312{
313	struct cpumask *current_mask = &save_cpumask;
314	struct trace_array *tr = hwlat_trace;
315	int next_cpu;
316
317	/*
318	 * If for some reason the user modifies the CPU affinity
319	 * of this thread, then stop migrating for the duration
320	 * of the current test.
321	 */
322	if (!cpumask_equal(current_mask, current->cpus_ptr))
323		goto change_mode;
324
325	cpus_read_lock();
326	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
327	next_cpu = cpumask_next_wrap(raw_smp_processor_id(), current_mask);
328	cpus_read_unlock();
329
330	if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
331		goto change_mode;
332
333	cpumask_clear(current_mask);
334	cpumask_set_cpu(next_cpu, current_mask);
335
336	set_cpus_allowed_ptr(current, current_mask);
337	return;
338
339 change_mode:
340	hwlat_data.thread_mode = MODE_NONE;
341	pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
342}
343
344/*
345 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
346 *
347 * Used to periodically sample the CPU TSC via a call to get_sample. We
348 * disable interrupts, which does (intentionally) introduce latency since we
349 * need to ensure nothing else might be running (and thus preempting).
350 * Obviously this should never be used in production environments.
351 *
352 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
353 */
354static int kthread_fn(void *data)
355{
356	u64 interval;
357
358	while (!kthread_should_stop()) {
359
360		if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
361			move_to_next_cpu();
362
363		local_irq_disable();
364		get_sample();
365		local_irq_enable();
366
367		mutex_lock(&hwlat_data.lock);
368		interval = hwlat_data.sample_window - hwlat_data.sample_width;
369		mutex_unlock(&hwlat_data.lock);
370
371		do_div(interval, USEC_PER_MSEC); /* modifies interval value */
372
373		/* Always sleep for at least 1ms */
374		if (interval < 1)
375			interval = 1;
376
377		if (msleep_interruptible(interval))
378			break;
379	}
380
381	return 0;
382}
383
384/*
385 * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
386 *
387 * This kicks the running hardware latency sampling/detector kernel thread and
388 * tells it to stop sampling now. Use this on unload and at system shutdown.
389 */
390static void stop_single_kthread(void)
391{
392	struct hwlat_kthread_data *kdata = get_cpu_data();
393	struct task_struct *kthread;
394
395	cpus_read_lock();
396	kthread = kdata->kthread;
397
398	if (!kthread)
399		goto out_put_cpus;
400
401	kthread_stop(kthread);
402	kdata->kthread = NULL;
403
404out_put_cpus:
405	cpus_read_unlock();
406}
407
408
409/*
410 * start_single_kthread - Kick off the hardware latency sampling/detector kthread
411 *
412 * This starts the kernel thread that will sit and sample the CPU timestamp
413 * counter (TSC or similar) and look for potential hardware latencies.
414 */
415static int start_single_kthread(struct trace_array *tr)
416{
417	struct hwlat_kthread_data *kdata = get_cpu_data();
418	struct cpumask *current_mask = &save_cpumask;
419	struct task_struct *kthread;
420	int next_cpu;
421
422	cpus_read_lock();
423	if (kdata->kthread)
424		goto out_put_cpus;
425
426	kthread = kthread_create(kthread_fn, NULL, "hwlatd");
427	if (IS_ERR(kthread)) {
428		pr_err(BANNER "could not start sampling thread\n");
429		cpus_read_unlock();
430		return -ENOMEM;
431	}
432
433	/* Just pick the first CPU on first iteration */
434	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
435
436	if (hwlat_data.thread_mode == MODE_ROUND_ROBIN) {
437		next_cpu = cpumask_first(current_mask);
438		cpumask_clear(current_mask);
439		cpumask_set_cpu(next_cpu, current_mask);
440
441	}
442
443	set_cpus_allowed_ptr(kthread, current_mask);
444
445	kdata->kthread = kthread;
446	wake_up_process(kthread);
447
448out_put_cpus:
449	cpus_read_unlock();
450	return 0;
451}
452
453/*
454 * stop_cpu_kthread - Stop a hwlat cpu kthread
455 */
456static void stop_cpu_kthread(unsigned int cpu)
457{
458	struct task_struct *kthread;
459
460	kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
461	if (kthread)
462		kthread_stop(kthread);
463	per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
464}
465
466/*
467 * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
468 *
469 * This kicks the running hardware latency sampling/detector kernel threads and
470 * tells it to stop sampling now. Use this on unload and at system shutdown.
471 */
472static void stop_per_cpu_kthreads(void)
473{
474	unsigned int cpu;
475
476	cpus_read_lock();
477	for_each_online_cpu(cpu)
478		stop_cpu_kthread(cpu);
479	cpus_read_unlock();
480}
481
482/*
483 * start_cpu_kthread - Start a hwlat cpu kthread
484 */
485static int start_cpu_kthread(unsigned int cpu)
486{
487	struct task_struct *kthread;
488
489	/* Do not start a new hwlatd thread if it is already running */
490	if (per_cpu(hwlat_per_cpu_data, cpu).kthread)
491		return 0;
492
493	kthread = kthread_run_on_cpu(kthread_fn, NULL, cpu, "hwlatd/%u");
494	if (IS_ERR(kthread)) {
495		pr_err(BANNER "could not start sampling thread\n");
496		return -ENOMEM;
497	}
498
499	per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
500
501	return 0;
502}
503
504#ifdef CONFIG_HOTPLUG_CPU
505static void hwlat_hotplug_workfn(struct work_struct *dummy)
506{
507	struct trace_array *tr = hwlat_trace;
508	unsigned int cpu = smp_processor_id();
509
510	mutex_lock(&trace_types_lock);
511	mutex_lock(&hwlat_data.lock);
512	cpus_read_lock();
513
514	if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
515		goto out_unlock;
516
517	if (!cpu_online(cpu))
518		goto out_unlock;
519	if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
520		goto out_unlock;
521
522	start_cpu_kthread(cpu);
523
524out_unlock:
525	cpus_read_unlock();
526	mutex_unlock(&hwlat_data.lock);
527	mutex_unlock(&trace_types_lock);
528}
529
530static DECLARE_WORK(hwlat_hotplug_work, hwlat_hotplug_workfn);
531
532/*
533 * hwlat_cpu_init - CPU hotplug online callback function
534 */
535static int hwlat_cpu_init(unsigned int cpu)
536{
537	schedule_work_on(cpu, &hwlat_hotplug_work);
538	return 0;
539}
540
541/*
542 * hwlat_cpu_die - CPU hotplug offline callback function
543 */
544static int hwlat_cpu_die(unsigned int cpu)
545{
546	stop_cpu_kthread(cpu);
547	return 0;
548}
549
550static void hwlat_init_hotplug_support(void)
551{
552	int ret;
553
554	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/hwlat:online",
555				hwlat_cpu_init, hwlat_cpu_die);
556	if (ret < 0)
557		pr_warn(BANNER "Error to init cpu hotplug support\n");
558
559	return;
560}
561#else /* CONFIG_HOTPLUG_CPU */
562static void hwlat_init_hotplug_support(void)
563{
564	return;
565}
566#endif /* CONFIG_HOTPLUG_CPU */
567
568/*
569 * start_per_cpu_kthreads - Kick off the hardware latency sampling/detector kthreads
570 *
571 * This starts the kernel threads that will sit on potentially all cpus and
572 * sample the CPU timestamp counter (TSC or similar) and look for potential
573 * hardware latencies.
574 */
575static int start_per_cpu_kthreads(struct trace_array *tr)
576{
577	struct cpumask *current_mask = &save_cpumask;
578	unsigned int cpu;
579	int retval;
580
581	cpus_read_lock();
582	/*
583	 * Run only on CPUs in which hwlat is allowed to run.
584	 */
585	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
586
587	for_each_cpu(cpu, current_mask) {
588		retval = start_cpu_kthread(cpu);
589		if (retval)
590			goto out_error;
591	}
592	cpus_read_unlock();
593
594	return 0;
595
596out_error:
597	cpus_read_unlock();
598	stop_per_cpu_kthreads();
599	return retval;
600}
601
602static void *s_mode_start(struct seq_file *s, loff_t *pos)
603{
604	int mode = *pos;
605
606	mutex_lock(&hwlat_data.lock);
607
608	if (mode >= MODE_MAX)
609		return NULL;
610
611	return pos;
612}
613
614static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
615{
616	int mode = ++(*pos);
617
618	if (mode >= MODE_MAX)
619		return NULL;
620
621	return pos;
622}
623
624static int s_mode_show(struct seq_file *s, void *v)
625{
626	loff_t *pos = v;
627	int mode = *pos;
628
629	if (mode == hwlat_data.thread_mode)
630		seq_printf(s, "[%s]", thread_mode_str[mode]);
631	else
632		seq_printf(s, "%s", thread_mode_str[mode]);
633
634	if (mode < MODE_MAX - 1) /* if mode is any but last */
635		seq_puts(s, " ");
636
637	return 0;
638}
639
640static void s_mode_stop(struct seq_file *s, void *v)
641{
642	seq_puts(s, "\n");
643	mutex_unlock(&hwlat_data.lock);
644}
645
646static const struct seq_operations thread_mode_seq_ops = {
647	.start		= s_mode_start,
648	.next		= s_mode_next,
649	.show		= s_mode_show,
650	.stop		= s_mode_stop
651};
652
653static int hwlat_mode_open(struct inode *inode, struct file *file)
654{
655	return seq_open(file, &thread_mode_seq_ops);
656};
657
658static void hwlat_tracer_start(struct trace_array *tr);
659static void hwlat_tracer_stop(struct trace_array *tr);
660
661/**
662 * hwlat_mode_write - Write function for "mode" entry
663 * @filp: The active open file structure
664 * @ubuf: The user buffer that contains the value to write
665 * @cnt: The maximum number of bytes to write to "file"
666 * @ppos: The current position in @file
667 *
668 * This function provides a write implementation for the "mode" interface
669 * to the hardware latency detector. hwlatd has different operation modes.
670 * The "none" sets the allowed cpumask for a single hwlatd thread at the
671 * startup and lets the scheduler handle the migration. The default mode is
672 * the "round-robin" one, in which a single hwlatd thread runs, migrating
673 * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
674 * creates one hwlatd thread per allowed CPU.
675 */
676static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
677				 size_t cnt, loff_t *ppos)
678{
679	struct trace_array *tr = hwlat_trace;
680	const char *mode;
681	char buf[64];
682	int ret, i;
683
684	if (cnt >= sizeof(buf))
685		return -EINVAL;
686
687	if (copy_from_user(buf, ubuf, cnt))
688		return -EFAULT;
689
690	buf[cnt] = 0;
691
692	mode = strstrip(buf);
693
694	ret = -EINVAL;
695
696	/*
697	 * trace_types_lock is taken to avoid concurrency on start/stop
698	 * and hwlat_busy.
699	 */
700	mutex_lock(&trace_types_lock);
701	if (hwlat_busy)
702		hwlat_tracer_stop(tr);
703
704	mutex_lock(&hwlat_data.lock);
705
706	for (i = 0; i < MODE_MAX; i++) {
707		if (strcmp(mode, thread_mode_str[i]) == 0) {
708			hwlat_data.thread_mode = i;
709			ret = cnt;
710		}
711	}
712
713	mutex_unlock(&hwlat_data.lock);
714
715	if (hwlat_busy)
716		hwlat_tracer_start(tr);
717	mutex_unlock(&trace_types_lock);
718
719	*ppos += cnt;
720
721
722
723	return ret;
724}
725
726/*
727 * The width parameter is read/write using the generic trace_min_max_param
728 * method. The *val is protected by the hwlat_data lock and is upper
729 * bounded by the window parameter.
730 */
731static struct trace_min_max_param hwlat_width = {
732	.lock		= &hwlat_data.lock,
733	.val		= &hwlat_data.sample_width,
734	.max		= &hwlat_data.sample_window,
735	.min		= NULL,
736};
737
738/*
739 * The window parameter is read/write using the generic trace_min_max_param
740 * method. The *val is protected by the hwlat_data lock and is lower
741 * bounded by the width parameter.
742 */
743static struct trace_min_max_param hwlat_window = {
744	.lock		= &hwlat_data.lock,
745	.val		= &hwlat_data.sample_window,
746	.max		= NULL,
747	.min		= &hwlat_data.sample_width,
748};
749
750static const struct file_operations thread_mode_fops = {
751	.open		= hwlat_mode_open,
752	.read		= seq_read,
753	.llseek		= seq_lseek,
754	.release	= seq_release,
755	.write		= hwlat_mode_write
756};
757/**
758 * init_tracefs - A function to initialize the tracefs interface files
759 *
760 * This function creates entries in tracefs for "hwlat_detector".
761 * It creates the hwlat_detector directory in the tracing directory,
762 * and within that directory is the count, width and window files to
763 * change and view those values.
764 */
765static int init_tracefs(void)
766{
767	int ret;
768	struct dentry *top_dir;
769
770	ret = tracing_init_dentry();
771	if (ret)
772		return -ENOMEM;
773
774	top_dir = tracefs_create_dir("hwlat_detector", NULL);
775	if (!top_dir)
776		return -ENOMEM;
777
778	hwlat_sample_window = tracefs_create_file("window", TRACE_MODE_WRITE,
779						  top_dir,
780						  &hwlat_window,
781						  &trace_min_max_fops);
782	if (!hwlat_sample_window)
783		goto err;
784
785	hwlat_sample_width = tracefs_create_file("width", TRACE_MODE_WRITE,
786						 top_dir,
787						 &hwlat_width,
788						 &trace_min_max_fops);
789	if (!hwlat_sample_width)
790		goto err;
791
792	hwlat_thread_mode = trace_create_file("mode", TRACE_MODE_WRITE,
793					      top_dir,
794					      NULL,
795					      &thread_mode_fops);
796	if (!hwlat_thread_mode)
797		goto err;
798
799	return 0;
800
801 err:
802	tracefs_remove(top_dir);
803	return -ENOMEM;
804}
805
806static void hwlat_tracer_start(struct trace_array *tr)
807{
808	int err;
809
810	if (hwlat_data.thread_mode == MODE_PER_CPU)
811		err = start_per_cpu_kthreads(tr);
812	else
813		err = start_single_kthread(tr);
814	if (err)
815		pr_err(BANNER "Cannot start hwlat kthread\n");
816}
817
818static void hwlat_tracer_stop(struct trace_array *tr)
819{
820	if (hwlat_data.thread_mode == MODE_PER_CPU)
821		stop_per_cpu_kthreads();
822	else
823		stop_single_kthread();
824}
825
826static int hwlat_tracer_init(struct trace_array *tr)
827{
828	/* Only allow one instance to enable this */
829	if (hwlat_busy)
830		return -EBUSY;
831
832	hwlat_trace = tr;
833
834	atomic64_set(&hwlat_data.count, 0);
835	tr->max_latency = 0;
836	save_tracing_thresh = tracing_thresh;
837
838	/* tracing_thresh is in nsecs, we speak in usecs */
839	if (!tracing_thresh)
840		tracing_thresh = last_tracing_thresh;
841
842	if (tracer_tracing_is_on(tr))
843		hwlat_tracer_start(tr);
844
845	hwlat_busy = true;
846
847	return 0;
848}
849
850static void hwlat_tracer_reset(struct trace_array *tr)
851{
852	hwlat_tracer_stop(tr);
853
854	/* the tracing threshold is static between runs */
855	last_tracing_thresh = tracing_thresh;
856
857	tracing_thresh = save_tracing_thresh;
858	hwlat_busy = false;
859}
860
861static struct tracer hwlat_tracer __read_mostly =
862{
863	.name		= "hwlat",
864	.init		= hwlat_tracer_init,
865	.reset		= hwlat_tracer_reset,
866	.start		= hwlat_tracer_start,
867	.stop		= hwlat_tracer_stop,
868	.allow_instances = true,
869};
870
871__init static int init_hwlat_tracer(void)
872{
873	int ret;
874
875	mutex_init(&hwlat_data.lock);
876
877	ret = register_tracer(&hwlat_tracer);
878	if (ret)
879		return ret;
880
881	hwlat_init_hotplug_support();
882
883	init_tracefs();
884
885	return 0;
886}
887late_initcall(init_hwlat_tracer);