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 u64 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, type: TRACE_HWLAT, len: sizeof(*entry),
138 trace_ctx: 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 and with
197 * hwlat_data.lock held.
198 */
199static int get_sample(void)
200{
201 struct hwlat_kthread_data *kdata = get_cpu_data();
202 struct trace_array *tr = hwlat_trace;
203 struct hwlat_sample s;
204 time_type start, t1, t2, last_t2;
205 s64 diff, outer_diff, total, last_total = 0;
206 u64 sample = 0;
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(tv: &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 <= hwlat_data.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 hwlat_data.count++;
289 s.seqnum = hwlat_data.count;
290 s.duration = sample;
291 s.outer_duration = outer_sample;
292 s.nmi_total_ts = kdata->nmi_total_ts;
293 s.nmi_count = kdata->nmi_count;
294 s.count = count;
295 trace_hwlat_sample(sample: &s);
296
297 latency = max(sample, outer_sample);
298
299 /* Keep a running maximum ever recorded hardware latency */
300 if (latency > tr->max_latency) {
301 tr->max_latency = latency;
302 latency_fsnotify(tr);
303 }
304 }
305
306out:
307 return ret;
308}
309
310static struct cpumask save_cpumask;
311
312static void move_to_next_cpu(void)
313{
314 struct cpumask *current_mask = &save_cpumask;
315 struct trace_array *tr = hwlat_trace;
316 int next_cpu;
317
318 /*
319 * If for some reason the user modifies the CPU affinity
320 * of this thread, then stop migrating for the duration
321 * of the current test.
322 */
323 if (!cpumask_equal(src1p: current_mask, current->cpus_ptr))
324 goto change_mode;
325
326 cpus_read_lock();
327 cpumask_and(dstp: current_mask, cpu_online_mask, src2p: tr->tracing_cpumask);
328 next_cpu = cpumask_next_wrap(raw_smp_processor_id(), src: current_mask);
329 cpus_read_unlock();
330
331 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
332 goto change_mode;
333
334 cpumask_clear(dstp: current_mask);
335 cpumask_set_cpu(cpu: next_cpu, dstp: current_mask);
336
337 set_cpus_allowed_ptr(current, new_mask: current_mask);
338 return;
339
340 change_mode:
341 hwlat_data.thread_mode = MODE_NONE;
342 pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
343}
344
345/*
346 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
347 *
348 * Used to periodically sample the CPU TSC via a call to get_sample. We
349 * disable interrupts, which does (intentionally) introduce latency since we
350 * need to ensure nothing else might be running (and thus preempting).
351 * Obviously this should never be used in production environments.
352 *
353 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
354 */
355static int kthread_fn(void *data)
356{
357 u64 interval;
358
359 while (!kthread_should_stop()) {
360
361 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
362 move_to_next_cpu();
363
364 local_irq_disable();
365 get_sample();
366 local_irq_enable();
367
368 mutex_lock(&hwlat_data.lock);
369 interval = hwlat_data.sample_window - hwlat_data.sample_width;
370 mutex_unlock(lock: &hwlat_data.lock);
371
372 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
373
374 /* Always sleep for at least 1ms */
375 if (interval < 1)
376 interval = 1;
377
378 if (msleep_interruptible(msecs: interval))
379 break;
380 }
381
382 return 0;
383}
384
385/*
386 * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
387 *
388 * This kicks the running hardware latency sampling/detector kernel thread and
389 * tells it to stop sampling now. Use this on unload and at system shutdown.
390 */
391static void stop_single_kthread(void)
392{
393 struct hwlat_kthread_data *kdata = get_cpu_data();
394 struct task_struct *kthread;
395
396 cpus_read_lock();
397 kthread = kdata->kthread;
398
399 if (!kthread)
400 goto out_put_cpus;
401
402 kthread_stop(k: kthread);
403 kdata->kthread = NULL;
404
405out_put_cpus:
406 cpus_read_unlock();
407}
408
409
410/*
411 * start_single_kthread - Kick off the hardware latency sampling/detector kthread
412 *
413 * This starts the kernel thread that will sit and sample the CPU timestamp
414 * counter (TSC or similar) and look for potential hardware latencies.
415 */
416static int start_single_kthread(struct trace_array *tr)
417{
418 struct hwlat_kthread_data *kdata = get_cpu_data();
419 struct cpumask *current_mask = &save_cpumask;
420 struct task_struct *kthread;
421 int next_cpu;
422
423 cpus_read_lock();
424 if (kdata->kthread)
425 goto out_put_cpus;
426
427 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
428 if (IS_ERR(ptr: kthread)) {
429 pr_err(BANNER "could not start sampling thread\n");
430 cpus_read_unlock();
431 return -ENOMEM;
432 }
433
434 /* Just pick the first CPU on first iteration */
435 cpumask_and(dstp: current_mask, cpu_online_mask, src2p: tr->tracing_cpumask);
436
437 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN) {
438 next_cpu = cpumask_first(srcp: current_mask);
439 cpumask_clear(dstp: current_mask);
440 cpumask_set_cpu(cpu: next_cpu, dstp: current_mask);
441
442 }
443
444 set_cpus_allowed_ptr(p: kthread, new_mask: current_mask);
445
446 kdata->kthread = kthread;
447 wake_up_process(tsk: kthread);
448
449out_put_cpus:
450 cpus_read_unlock();
451 return 0;
452}
453
454/*
455 * stop_cpu_kthread - Stop a hwlat cpu kthread
456 */
457static void stop_cpu_kthread(unsigned int cpu)
458{
459 struct task_struct *kthread;
460
461 kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
462 if (kthread)
463 kthread_stop(k: kthread);
464 per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
465}
466
467/*
468 * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
469 *
470 * This kicks the running hardware latency sampling/detector kernel threads and
471 * tells it to stop sampling now. Use this on unload and at system shutdown.
472 */
473static void stop_per_cpu_kthreads(void)
474{
475 unsigned int cpu;
476
477 cpus_read_lock();
478 for_each_online_cpu(cpu)
479 stop_cpu_kthread(cpu);
480 cpus_read_unlock();
481}
482
483/*
484 * start_cpu_kthread - Start a hwlat cpu kthread
485 */
486static int start_cpu_kthread(unsigned int cpu)
487{
488 struct task_struct *kthread;
489
490 /* Do not start a new hwlatd thread if it is already running */
491 if (per_cpu(hwlat_per_cpu_data, cpu).kthread)
492 return 0;
493
494 kthread = kthread_run_on_cpu(threadfn: kthread_fn, NULL, cpu, namefmt: "hwlatd/%u");
495 if (IS_ERR(ptr: kthread)) {
496 pr_err(BANNER "could not start sampling thread\n");
497 return -ENOMEM;
498 }
499
500 per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
501
502 return 0;
503}
504
505#ifdef CONFIG_HOTPLUG_CPU
506static void hwlat_hotplug_workfn(struct work_struct *dummy)
507{
508 struct trace_array *tr = hwlat_trace;
509 unsigned int cpu = smp_processor_id();
510
511 mutex_lock(&trace_types_lock);
512 mutex_lock(&hwlat_data.lock);
513 cpus_read_lock();
514
515 if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
516 goto out_unlock;
517
518 if (!cpu_online(cpu))
519 goto out_unlock;
520 if (!cpumask_test_cpu(cpu, cpumask: tr->tracing_cpumask))
521 goto out_unlock;
522
523 start_cpu_kthread(cpu);
524
525out_unlock:
526 cpus_read_unlock();
527 mutex_unlock(lock: &hwlat_data.lock);
528 mutex_unlock(lock: &trace_types_lock);
529}
530
531static DECLARE_WORK(hwlat_hotplug_work, hwlat_hotplug_workfn);
532
533/*
534 * hwlat_cpu_init - CPU hotplug online callback function
535 */
536static int hwlat_cpu_init(unsigned int cpu)
537{
538 schedule_work_on(cpu, work: &hwlat_hotplug_work);
539 return 0;
540}
541
542/*
543 * hwlat_cpu_die - CPU hotplug offline callback function
544 */
545static int hwlat_cpu_die(unsigned int cpu)
546{
547 stop_cpu_kthread(cpu);
548 return 0;
549}
550
551static void hwlat_init_hotplug_support(void)
552{
553 int ret;
554
555 ret = cpuhp_setup_state(state: CPUHP_AP_ONLINE_DYN, name: "trace/hwlat:online",
556 startup: hwlat_cpu_init, teardown: hwlat_cpu_die);
557 if (ret < 0)
558 pr_warn(BANNER "Error to init cpu hotplug support\n");
559
560 return;
561}
562#else /* CONFIG_HOTPLUG_CPU */
563static void hwlat_init_hotplug_support(void)
564{
565 return;
566}
567#endif /* CONFIG_HOTPLUG_CPU */
568
569/*
570 * start_per_cpu_kthreads - Kick off the hardware latency sampling/detector kthreads
571 *
572 * This starts the kernel threads that will sit on potentially all cpus and
573 * sample the CPU timestamp counter (TSC or similar) and look for potential
574 * hardware latencies.
575 */
576static int start_per_cpu_kthreads(struct trace_array *tr)
577{
578 struct cpumask *current_mask = &save_cpumask;
579 unsigned int cpu;
580 int retval;
581
582 cpus_read_lock();
583 /*
584 * Run only on CPUs in which hwlat is allowed to run.
585 */
586 cpumask_and(dstp: current_mask, cpu_online_mask, src2p: tr->tracing_cpumask);
587
588 for_each_cpu(cpu, current_mask) {
589 retval = start_cpu_kthread(cpu);
590 if (retval)
591 goto out_error;
592 }
593 cpus_read_unlock();
594
595 return 0;
596
597out_error:
598 cpus_read_unlock();
599 stop_per_cpu_kthreads();
600 return retval;
601}
602
603static void *s_mode_start(struct seq_file *s, loff_t *pos)
604{
605 int mode = *pos;
606
607 mutex_lock(&hwlat_data.lock);
608
609 if (mode >= MODE_MAX)
610 return NULL;
611
612 return pos;
613}
614
615static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
616{
617 int mode = ++(*pos);
618
619 if (mode >= MODE_MAX)
620 return NULL;
621
622 return pos;
623}
624
625static int s_mode_show(struct seq_file *s, void *v)
626{
627 loff_t *pos = v;
628 int mode = *pos;
629
630 if (mode == hwlat_data.thread_mode)
631 seq_printf(m: s, fmt: "[%s]", thread_mode_str[mode]);
632 else
633 seq_printf(m: s, fmt: "%s", thread_mode_str[mode]);
634
635 if (mode < MODE_MAX - 1) /* if mode is any but last */
636 seq_puts(m: s, s: " ");
637
638 return 0;
639}
640
641static void s_mode_stop(struct seq_file *s, void *v)
642{
643 seq_puts(m: s, s: "\n");
644 mutex_unlock(lock: &hwlat_data.lock);
645}
646
647static const struct seq_operations thread_mode_seq_ops = {
648 .start = s_mode_start,
649 .next = s_mode_next,
650 .show = s_mode_show,
651 .stop = s_mode_stop
652};
653
654static int hwlat_mode_open(struct inode *inode, struct file *file)
655{
656 return seq_open(file, &thread_mode_seq_ops);
657};
658
659static void hwlat_tracer_start(struct trace_array *tr);
660static void hwlat_tracer_stop(struct trace_array *tr);
661
662/**
663 * hwlat_mode_write - Write function for "mode" entry
664 * @filp: The active open file structure
665 * @ubuf: The user buffer that contains the value to write
666 * @cnt: The maximum number of bytes to write to "file"
667 * @ppos: The current position in @file
668 *
669 * This function provides a write implementation for the "mode" interface
670 * to the hardware latency detector. hwlatd has different operation modes.
671 * The "none" sets the allowed cpumask for a single hwlatd thread at the
672 * startup and lets the scheduler handle the migration. The default mode is
673 * the "round-robin" one, in which a single hwlatd thread runs, migrating
674 * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
675 * creates one hwlatd thread per allowed CPU.
676 */
677static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
678 size_t cnt, loff_t *ppos)
679{
680 struct trace_array *tr = hwlat_trace;
681 const char *mode;
682 char buf[64];
683 int ret, i;
684
685 if (cnt >= sizeof(buf))
686 return -EINVAL;
687
688 if (copy_from_user(to: buf, from: ubuf, n: cnt))
689 return -EFAULT;
690
691 buf[cnt] = 0;
692
693 mode = strstrip(str: buf);
694
695 ret = -EINVAL;
696
697 /*
698 * trace_types_lock is taken to avoid concurrency on start/stop
699 * and hwlat_busy.
700 */
701 mutex_lock(&trace_types_lock);
702 if (hwlat_busy)
703 hwlat_tracer_stop(tr);
704
705 mutex_lock(&hwlat_data.lock);
706
707 for (i = 0; i < MODE_MAX; i++) {
708 if (strcmp(mode, thread_mode_str[i]) == 0) {
709 hwlat_data.thread_mode = i;
710 ret = cnt;
711 }
712 }
713
714 mutex_unlock(lock: &hwlat_data.lock);
715
716 if (hwlat_busy)
717 hwlat_tracer_start(tr);
718 mutex_unlock(lock: &trace_types_lock);
719
720 *ppos += cnt;
721
722
723
724 return ret;
725}
726
727/*
728 * The width parameter is read/write using the generic trace_min_max_param
729 * method. The *val is protected by the hwlat_data lock and is upper
730 * bounded by the window parameter.
731 */
732static struct trace_min_max_param hwlat_width = {
733 .lock = &hwlat_data.lock,
734 .val = &hwlat_data.sample_width,
735 .max = &hwlat_data.sample_window,
736 .min = NULL,
737};
738
739/*
740 * The window parameter is read/write using the generic trace_min_max_param
741 * method. The *val is protected by the hwlat_data lock and is lower
742 * bounded by the width parameter.
743 */
744static struct trace_min_max_param hwlat_window = {
745 .lock = &hwlat_data.lock,
746 .val = &hwlat_data.sample_window,
747 .max = NULL,
748 .min = &hwlat_data.sample_width,
749};
750
751static const struct file_operations thread_mode_fops = {
752 .open = hwlat_mode_open,
753 .read = seq_read,
754 .llseek = seq_lseek,
755 .release = seq_release,
756 .write = hwlat_mode_write
757};
758/**
759 * init_tracefs - A function to initialize the tracefs interface files
760 *
761 * This function creates entries in tracefs for "hwlat_detector".
762 * It creates the hwlat_detector directory in the tracing directory,
763 * and within that directory is the count, width and window files to
764 * change and view those values.
765 */
766static int init_tracefs(void)
767{
768 int ret;
769 struct dentry *top_dir;
770
771 ret = tracing_init_dentry();
772 if (ret)
773 return -ENOMEM;
774
775 top_dir = tracefs_create_dir(name: "hwlat_detector", NULL);
776 if (!top_dir)
777 return -ENOMEM;
778
779 hwlat_sample_window = tracefs_create_file(name: "window", TRACE_MODE_WRITE,
780 parent: top_dir,
781 data: &hwlat_window,
782 fops: &trace_min_max_fops);
783 if (!hwlat_sample_window)
784 goto err;
785
786 hwlat_sample_width = tracefs_create_file(name: "width", TRACE_MODE_WRITE,
787 parent: top_dir,
788 data: &hwlat_width,
789 fops: &trace_min_max_fops);
790 if (!hwlat_sample_width)
791 goto err;
792
793 hwlat_thread_mode = trace_create_file(name: "mode", TRACE_MODE_WRITE,
794 parent: top_dir,
795 NULL,
796 fops: &thread_mode_fops);
797 if (!hwlat_thread_mode)
798 goto err;
799
800 return 0;
801
802 err:
803 tracefs_remove(dentry: top_dir);
804 return -ENOMEM;
805}
806
807static void hwlat_tracer_start(struct trace_array *tr)
808{
809 int err;
810
811 if (hwlat_data.thread_mode == MODE_PER_CPU)
812 err = start_per_cpu_kthreads(tr);
813 else
814 err = start_single_kthread(tr);
815 if (err)
816 pr_err(BANNER "Cannot start hwlat kthread\n");
817}
818
819static void hwlat_tracer_stop(struct trace_array *tr)
820{
821 if (hwlat_data.thread_mode == MODE_PER_CPU)
822 stop_per_cpu_kthreads();
823 else
824 stop_single_kthread();
825}
826
827static int hwlat_tracer_init(struct trace_array *tr)
828{
829 /* Only allow one instance to enable this */
830 if (hwlat_busy)
831 return -EBUSY;
832
833 hwlat_trace = tr;
834
835 hwlat_data.count = 0;
836 tr->max_latency = 0;
837 save_tracing_thresh = tracing_thresh;
838
839 /* tracing_thresh is in nsecs, we speak in usecs */
840 if (!tracing_thresh)
841 tracing_thresh = last_tracing_thresh;
842
843 if (tracer_tracing_is_on(tr))
844 hwlat_tracer_start(tr);
845
846 hwlat_busy = true;
847
848 return 0;
849}
850
851static void hwlat_tracer_reset(struct trace_array *tr)
852{
853 hwlat_tracer_stop(tr);
854
855 /* the tracing threshold is static between runs */
856 last_tracing_thresh = tracing_thresh;
857
858 tracing_thresh = save_tracing_thresh;
859 hwlat_busy = false;
860}
861
862static struct tracer hwlat_tracer __read_mostly =
863{
864 .name = "hwlat",
865 .init = hwlat_tracer_init,
866 .reset = hwlat_tracer_reset,
867 .start = hwlat_tracer_start,
868 .stop = hwlat_tracer_stop,
869 .allow_instances = true,
870};
871
872__init static int init_hwlat_tracer(void)
873{
874 int ret;
875
876 mutex_init(&hwlat_data.lock);
877
878 ret = register_tracer(type: &hwlat_tracer);
879 if (ret)
880 return ret;
881
882 hwlat_init_hotplug_support();
883
884 init_tracefs();
885
886 return 0;
887}
888late_initcall(init_hwlat_tracer);
889

source code of linux/kernel/trace/trace_hwlat.c