1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
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
51
52#define BANNER "hwlat_detector: "
53#define DEFAULT_SAMPLE_WINDOW 1000000
54#define DEFAULT_SAMPLE_WIDTH 500000
55#define DEFAULT_LAT_THRESHOLD 10
56
57
58static struct task_struct *hwlat_kthread;
59
60static struct dentry *hwlat_sample_width;
61static struct dentry *hwlat_sample_window;
62
63
64static unsigned long save_tracing_thresh;
65
66
67static u64 nmi_ts_start;
68static u64 nmi_total_ts;
69static int nmi_count;
70static int nmi_cpu;
71
72
73bool trace_hwlat_callback_enabled;
74
75
76static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
77
78
79struct hwlat_sample {
80 u64 seqnum;
81 u64 duration;
82 u64 outer_duration;
83 u64 nmi_total_ts;
84 struct timespec64 timestamp;
85 int nmi_count;
86 int count;
87};
88
89
90static struct hwlat_data {
91
92 struct mutex lock;
93
94 u64 count;
95
96 u64 sample_window;
97 u64 sample_width;
98
99} hwlat_data = {
100 .sample_window = DEFAULT_SAMPLE_WINDOW,
101 .sample_width = DEFAULT_SAMPLE_WIDTH,
102};
103
104static void trace_hwlat_sample(struct hwlat_sample *sample)
105{
106 struct trace_array *tr = hwlat_trace;
107 struct trace_event_call *call = &event_hwlat;
108 struct trace_buffer *buffer = tr->array_buffer.buffer;
109 struct ring_buffer_event *event;
110 struct hwlat_entry *entry;
111 unsigned long flags;
112 int pc;
113
114 pc = preempt_count();
115 local_save_flags(flags);
116
117 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
118 flags, pc);
119 if (!event)
120 return;
121 entry = ring_buffer_event_data(event);
122 entry->seqnum = sample->seqnum;
123 entry->duration = sample->duration;
124 entry->outer_duration = sample->outer_duration;
125 entry->timestamp = sample->timestamp;
126 entry->nmi_total_ts = sample->nmi_total_ts;
127 entry->nmi_count = sample->nmi_count;
128 entry->count = sample->count;
129
130 if (!call_filter_check_discard(call, entry, buffer, event))
131 trace_buffer_unlock_commit_nostack(buffer, event);
132}
133
134
135#define time_type u64
136#define time_get() trace_clock_local()
137#define time_to_us(x) div_u64(x, 1000)
138#define time_sub(a, b) ((a) - (b))
139#define init_time(a, b) (a = b)
140#define time_u64(a) a
141
142void trace_hwlat_callback(bool enter)
143{
144 if (smp_processor_id() != nmi_cpu)
145 return;
146
147
148
149
150
151 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
152 if (enter)
153 nmi_ts_start = time_get();
154 else
155 nmi_total_ts += time_get() - nmi_ts_start;
156 }
157
158 if (enter)
159 nmi_count++;
160}
161
162
163
164
165
166
167
168
169static int get_sample(void)
170{
171 struct trace_array *tr = hwlat_trace;
172 struct hwlat_sample s;
173 time_type start, t1, t2, last_t2;
174 s64 diff, outer_diff, total, last_total = 0;
175 u64 sample = 0;
176 u64 thresh = tracing_thresh;
177 u64 outer_sample = 0;
178 int ret = -1;
179 unsigned int count = 0;
180
181 do_div(thresh, NSEC_PER_USEC);
182
183 nmi_cpu = smp_processor_id();
184 nmi_total_ts = 0;
185 nmi_count = 0;
186
187 barrier();
188
189 trace_hwlat_callback_enabled = true;
190
191 init_time(last_t2, 0);
192 start = time_get();
193 outer_diff = 0;
194
195 do {
196
197 t1 = time_get();
198 t2 = time_get();
199
200 if (time_u64(last_t2)) {
201
202 outer_diff = time_to_us(time_sub(t1, last_t2));
203
204 if (outer_diff < 0) {
205 pr_err(BANNER "time running backwards\n");
206 goto out;
207 }
208 if (outer_diff > outer_sample)
209 outer_sample = outer_diff;
210 }
211 last_t2 = t2;
212
213 total = time_to_us(time_sub(t2, start));
214
215
216 if (total < last_total) {
217 pr_err("Time total overflowed\n");
218 break;
219 }
220 last_total = total;
221
222
223 diff = time_to_us(time_sub(t2, t1));
224
225 if (diff > thresh || outer_diff > thresh) {
226 if (!count)
227 ktime_get_real_ts64(&s.timestamp);
228 count++;
229 }
230
231
232 if (diff < 0) {
233 pr_err(BANNER "time running backwards\n");
234 goto out;
235 }
236
237 if (diff > sample)
238 sample = diff;
239
240 } while (total <= hwlat_data.sample_width);
241
242 barrier();
243 trace_hwlat_callback_enabled = false;
244 barrier();
245
246 ret = 0;
247
248
249 if (sample > thresh || outer_sample > thresh) {
250 u64 latency;
251
252 ret = 1;
253
254
255 if (nmi_total_ts)
256 do_div(nmi_total_ts, NSEC_PER_USEC);
257
258 hwlat_data.count++;
259 s.seqnum = hwlat_data.count;
260 s.duration = sample;
261 s.outer_duration = outer_sample;
262 s.nmi_total_ts = nmi_total_ts;
263 s.nmi_count = nmi_count;
264 s.count = count;
265 trace_hwlat_sample(&s);
266
267 latency = max(sample, outer_sample);
268
269
270 if (latency > tr->max_latency) {
271 tr->max_latency = latency;
272 latency_fsnotify(tr);
273 }
274 }
275
276out:
277 return ret;
278}
279
280static struct cpumask save_cpumask;
281static bool disable_migrate;
282
283static void move_to_next_cpu(void)
284{
285 struct cpumask *current_mask = &save_cpumask;
286 struct trace_array *tr = hwlat_trace;
287 int next_cpu;
288
289 if (disable_migrate)
290 return;
291
292
293
294
295
296 if (!cpumask_equal(current_mask, current->cpus_ptr))
297 goto disable;
298
299 get_online_cpus();
300 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
301 next_cpu = cpumask_next(smp_processor_id(), current_mask);
302 put_online_cpus();
303
304 if (next_cpu >= nr_cpu_ids)
305 next_cpu = cpumask_first(current_mask);
306
307 if (next_cpu >= nr_cpu_ids)
308 goto disable;
309
310 cpumask_clear(current_mask);
311 cpumask_set_cpu(next_cpu, current_mask);
312
313 sched_setaffinity(0, current_mask);
314 return;
315
316 disable:
317 disable_migrate = true;
318}
319
320
321
322
323
324
325
326
327
328
329
330static int kthread_fn(void *data)
331{
332 u64 interval;
333
334 while (!kthread_should_stop()) {
335
336 move_to_next_cpu();
337
338 local_irq_disable();
339 get_sample();
340 local_irq_enable();
341
342 mutex_lock(&hwlat_data.lock);
343 interval = hwlat_data.sample_window - hwlat_data.sample_width;
344 mutex_unlock(&hwlat_data.lock);
345
346 do_div(interval, USEC_PER_MSEC);
347
348
349 if (interval < 1)
350 interval = 1;
351
352 if (msleep_interruptible(interval))
353 break;
354 }
355
356 return 0;
357}
358
359
360
361
362
363
364
365static int start_kthread(struct trace_array *tr)
366{
367 struct cpumask *current_mask = &save_cpumask;
368 struct task_struct *kthread;
369 int next_cpu;
370
371 if (WARN_ON(hwlat_kthread))
372 return 0;
373
374
375 get_online_cpus();
376 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
377 put_online_cpus();
378 next_cpu = cpumask_first(current_mask);
379
380 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
381 if (IS_ERR(kthread)) {
382 pr_err(BANNER "could not start sampling thread\n");
383 return -ENOMEM;
384 }
385
386 cpumask_clear(current_mask);
387 cpumask_set_cpu(next_cpu, current_mask);
388 sched_setaffinity(kthread->pid, current_mask);
389
390 hwlat_kthread = kthread;
391 wake_up_process(kthread);
392
393 return 0;
394}
395
396
397
398
399
400
401
402static void stop_kthread(void)
403{
404 if (!hwlat_kthread)
405 return;
406 kthread_stop(hwlat_kthread);
407 hwlat_kthread = NULL;
408}
409
410
411
412
413
414
415
416
417
418
419
420static ssize_t hwlat_read(struct file *filp, char __user *ubuf,
421 size_t cnt, loff_t *ppos)
422{
423 char buf[U64STR_SIZE];
424 u64 *entry = filp->private_data;
425 u64 val;
426 int len;
427
428 if (!entry)
429 return -EFAULT;
430
431 if (cnt > sizeof(buf))
432 cnt = sizeof(buf);
433
434 val = *entry;
435
436 len = snprintf(buf, sizeof(buf), "%llu\n", val);
437
438 return simple_read_from_buffer(ubuf, cnt, ppos, buf, len);
439}
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456static ssize_t
457hwlat_width_write(struct file *filp, const char __user *ubuf,
458 size_t cnt, loff_t *ppos)
459{
460 u64 val;
461 int err;
462
463 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
464 if (err)
465 return err;
466
467 mutex_lock(&hwlat_data.lock);
468 if (val < hwlat_data.sample_window)
469 hwlat_data.sample_width = val;
470 else
471 err = -EINVAL;
472 mutex_unlock(&hwlat_data.lock);
473
474 if (err)
475 return err;
476
477 return cnt;
478}
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495static ssize_t
496hwlat_window_write(struct file *filp, const char __user *ubuf,
497 size_t cnt, loff_t *ppos)
498{
499 u64 val;
500 int err;
501
502 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
503 if (err)
504 return err;
505
506 mutex_lock(&hwlat_data.lock);
507 if (hwlat_data.sample_width < val)
508 hwlat_data.sample_window = val;
509 else
510 err = -EINVAL;
511 mutex_unlock(&hwlat_data.lock);
512
513 if (err)
514 return err;
515
516 return cnt;
517}
518
519static const struct file_operations width_fops = {
520 .open = tracing_open_generic,
521 .read = hwlat_read,
522 .write = hwlat_width_write,
523};
524
525static const struct file_operations window_fops = {
526 .open = tracing_open_generic,
527 .read = hwlat_read,
528 .write = hwlat_window_write,
529};
530
531
532
533
534
535
536
537
538
539static int init_tracefs(void)
540{
541 struct dentry *d_tracer;
542 struct dentry *top_dir;
543
544 d_tracer = tracing_init_dentry();
545 if (IS_ERR(d_tracer))
546 return -ENOMEM;
547
548 top_dir = tracefs_create_dir("hwlat_detector", d_tracer);
549 if (!top_dir)
550 return -ENOMEM;
551
552 hwlat_sample_window = tracefs_create_file("window", 0640,
553 top_dir,
554 &hwlat_data.sample_window,
555 &window_fops);
556 if (!hwlat_sample_window)
557 goto err;
558
559 hwlat_sample_width = tracefs_create_file("width", 0644,
560 top_dir,
561 &hwlat_data.sample_width,
562 &width_fops);
563 if (!hwlat_sample_width)
564 goto err;
565
566 return 0;
567
568 err:
569 tracefs_remove(top_dir);
570 return -ENOMEM;
571}
572
573static void hwlat_tracer_start(struct trace_array *tr)
574{
575 int err;
576
577 err = start_kthread(tr);
578 if (err)
579 pr_err(BANNER "Cannot start hwlat kthread\n");
580}
581
582static void hwlat_tracer_stop(struct trace_array *tr)
583{
584 stop_kthread();
585}
586
587static bool hwlat_busy;
588
589static int hwlat_tracer_init(struct trace_array *tr)
590{
591
592 if (hwlat_busy)
593 return -EBUSY;
594
595 hwlat_trace = tr;
596
597 disable_migrate = false;
598 hwlat_data.count = 0;
599 tr->max_latency = 0;
600 save_tracing_thresh = tracing_thresh;
601
602
603 if (!tracing_thresh)
604 tracing_thresh = last_tracing_thresh;
605
606 if (tracer_tracing_is_on(tr))
607 hwlat_tracer_start(tr);
608
609 hwlat_busy = true;
610
611 return 0;
612}
613
614static void hwlat_tracer_reset(struct trace_array *tr)
615{
616 stop_kthread();
617
618
619 last_tracing_thresh = tracing_thresh;
620
621 tracing_thresh = save_tracing_thresh;
622 hwlat_busy = false;
623}
624
625static struct tracer hwlat_tracer __read_mostly =
626{
627 .name = "hwlat",
628 .init = hwlat_tracer_init,
629 .reset = hwlat_tracer_reset,
630 .start = hwlat_tracer_start,
631 .stop = hwlat_tracer_stop,
632 .allow_instances = true,
633};
634
635__init static int init_hwlat_tracer(void)
636{
637 int ret;
638
639 mutex_init(&hwlat_data.lock);
640
641 ret = register_tracer(&hwlat_tracer);
642 if (ret)
643 return ret;
644
645 init_tracefs();
646
647 return 0;
648}
649late_initcall(init_hwlat_tracer);
650