1// SPDX-License-Identifier: GPL-2.0
2/*
3 * trace_hwlatdetect.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 <clark@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
57/* sampling thread*/
58static struct task_struct *hwlat_kthread;
59
60static struct dentry *hwlat_sample_width; /* sample width us */
61static struct dentry *hwlat_sample_window; /* sample window us */
62
63/* Save the previous tracing_thresh value */
64static unsigned long save_tracing_thresh;
65
66/* NMI timestamp counters */
67static u64 nmi_ts_start;
68static u64 nmi_total_ts;
69static int nmi_count;
70static int nmi_cpu;
71
72/* Tells NMIs to call back to the hwlat tracer to record timestamps */
73bool trace_hwlat_callback_enabled;
74
75/* If the user changed threshold, remember it */
76static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
77
78/* Individual latency samples are stored here when detected. */
79struct hwlat_sample {
80 u64 seqnum; /* unique sequence */
81 u64 duration; /* delta */
82 u64 outer_duration; /* delta (outer loop) */
83 u64 nmi_total_ts; /* Total time spent in NMIs */
84 struct timespec64 timestamp; /* wall time */
85 int nmi_count; /* # NMIs during this sample */
86};
87
88/* keep the global state somewhere. */
89static struct hwlat_data {
90
91 struct mutex lock; /* protect changes */
92
93 u64 count; /* total since reset */
94
95 u64 sample_window; /* total sampling window (on+off) */
96 u64 sample_width; /* active sampling portion of window */
97
98} hwlat_data = {
99 .sample_window = DEFAULT_SAMPLE_WINDOW,
100 .sample_width = DEFAULT_SAMPLE_WIDTH,
101};
102
103static void trace_hwlat_sample(struct hwlat_sample *sample)
104{
105 struct trace_array *tr = hwlat_trace;
106 struct trace_event_call *call = &event_hwlat;
107 struct ring_buffer *buffer = tr->trace_buffer.buffer;
108 struct ring_buffer_event *event;
109 struct hwlat_entry *entry;
110 unsigned long flags;
111 int pc;
112
113 pc = preempt_count();
114 local_save_flags(flags);
115
116 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
117 flags, pc);
118 if (!event)
119 return;
120 entry = ring_buffer_event_data(event);
121 entry->seqnum = sample->seqnum;
122 entry->duration = sample->duration;
123 entry->outer_duration = sample->outer_duration;
124 entry->timestamp = sample->timestamp;
125 entry->nmi_total_ts = sample->nmi_total_ts;
126 entry->nmi_count = sample->nmi_count;
127
128 if (!call_filter_check_discard(call, entry, buffer, event))
129 trace_buffer_unlock_commit_nostack(buffer, event);
130}
131
132/* Macros to encapsulate the time capturing infrastructure */
133#define time_type u64
134#define time_get() trace_clock_local()
135#define time_to_us(x) div_u64(x, 1000)
136#define time_sub(a, b) ((a) - (b))
137#define init_time(a, b) (a = b)
138#define time_u64(a) a
139
140void trace_hwlat_callback(bool enter)
141{
142 if (smp_processor_id() != nmi_cpu)
143 return;
144
145 /*
146 * Currently trace_clock_local() calls sched_clock() and the
147 * generic version is not NMI safe.
148 */
149 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
150 if (enter)
151 nmi_ts_start = time_get();
152 else
153 nmi_total_ts = time_get() - nmi_ts_start;
154 }
155
156 if (enter)
157 nmi_count++;
158}
159
160/**
161 * get_sample - sample the CPU TSC and look for likely hardware latencies
162 *
163 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
164 * hardware-induced latency. Called with interrupts disabled and with
165 * hwlat_data.lock held.
166 */
167static int get_sample(void)
168{
169 struct trace_array *tr = hwlat_trace;
170 time_type start, t1, t2, last_t2;
171 s64 diff, total, last_total = 0;
172 u64 sample = 0;
173 u64 thresh = tracing_thresh;
174 u64 outer_sample = 0;
175 int ret = -1;
176
177 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
178
179 nmi_cpu = smp_processor_id();
180 nmi_total_ts = 0;
181 nmi_count = 0;
182 /* Make sure NMIs see this first */
183 barrier();
184
185 trace_hwlat_callback_enabled = true;
186
187 init_time(last_t2, 0);
188 start = time_get(); /* start timestamp */
189
190 do {
191
192 t1 = time_get(); /* we'll look for a discontinuity */
193 t2 = time_get();
194
195 if (time_u64(last_t2)) {
196 /* Check the delta from outer loop (t2 to next t1) */
197 diff = time_to_us(time_sub(t1, last_t2));
198 /* This shouldn't happen */
199 if (diff < 0) {
200 pr_err(BANNER "time running backwards\n");
201 goto out;
202 }
203 if (diff > outer_sample)
204 outer_sample = diff;
205 }
206 last_t2 = t2;
207
208 total = time_to_us(time_sub(t2, start)); /* sample width */
209
210 /* Check for possible overflows */
211 if (total < last_total) {
212 pr_err("Time total overflowed\n");
213 break;
214 }
215 last_total = total;
216
217 /* This checks the inner loop (t1 to t2) */
218 diff = time_to_us(time_sub(t2, t1)); /* current diff */
219
220 /* This shouldn't happen */
221 if (diff < 0) {
222 pr_err(BANNER "time running backwards\n");
223 goto out;
224 }
225
226 if (diff > sample)
227 sample = diff; /* only want highest value */
228
229 } while (total <= hwlat_data.sample_width);
230
231 barrier(); /* finish the above in the view for NMIs */
232 trace_hwlat_callback_enabled = false;
233 barrier(); /* Make sure nmi_total_ts is no longer updated */
234
235 ret = 0;
236
237 /* If we exceed the threshold value, we have found a hardware latency */
238 if (sample > thresh || outer_sample > thresh) {
239 struct hwlat_sample s;
240
241 ret = 1;
242
243 /* We read in microseconds */
244 if (nmi_total_ts)
245 do_div(nmi_total_ts, NSEC_PER_USEC);
246
247 hwlat_data.count++;
248 s.seqnum = hwlat_data.count;
249 s.duration = sample;
250 s.outer_duration = outer_sample;
251 ktime_get_real_ts64(&s.timestamp);
252 s.nmi_total_ts = nmi_total_ts;
253 s.nmi_count = nmi_count;
254 trace_hwlat_sample(&s);
255
256 /* Keep a running maximum ever recorded hardware latency */
257 if (sample > tr->max_latency)
258 tr->max_latency = sample;
259 }
260
261out:
262 return ret;
263}
264
265static struct cpumask save_cpumask;
266static bool disable_migrate;
267
268static void move_to_next_cpu(void)
269{
270 struct cpumask *current_mask = &save_cpumask;
271 int next_cpu;
272
273 if (disable_migrate)
274 return;
275 /*
276 * If for some reason the user modifies the CPU affinity
277 * of this thread, than stop migrating for the duration
278 * of the current test.
279 */
280 if (!cpumask_equal(current_mask, &current->cpus_allowed))
281 goto disable;
282
283 get_online_cpus();
284 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
285 next_cpu = cpumask_next(smp_processor_id(), current_mask);
286 put_online_cpus();
287
288 if (next_cpu >= nr_cpu_ids)
289 next_cpu = cpumask_first(current_mask);
290
291 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
292 goto disable;
293
294 cpumask_clear(current_mask);
295 cpumask_set_cpu(next_cpu, current_mask);
296
297 sched_setaffinity(0, current_mask);
298 return;
299
300 disable:
301 disable_migrate = true;
302}
303
304/*
305 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
306 *
307 * Used to periodically sample the CPU TSC via a call to get_sample. We
308 * disable interrupts, which does (intentionally) introduce latency since we
309 * need to ensure nothing else might be running (and thus preempting).
310 * Obviously this should never be used in production environments.
311 *
312 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
313 */
314static int kthread_fn(void *data)
315{
316 u64 interval;
317
318 while (!kthread_should_stop()) {
319
320 move_to_next_cpu();
321
322 local_irq_disable();
323 get_sample();
324 local_irq_enable();
325
326 mutex_lock(&hwlat_data.lock);
327 interval = hwlat_data.sample_window - hwlat_data.sample_width;
328 mutex_unlock(&hwlat_data.lock);
329
330 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
331
332 /* Always sleep for at least 1ms */
333 if (interval < 1)
334 interval = 1;
335
336 if (msleep_interruptible(interval))
337 break;
338 }
339
340 return 0;
341}
342
343/**
344 * start_kthread - Kick off the hardware latency sampling/detector kthread
345 *
346 * This starts the kernel thread that will sit and sample the CPU timestamp
347 * counter (TSC or similar) and look for potential hardware latencies.
348 */
349static int start_kthread(struct trace_array *tr)
350{
351 struct cpumask *current_mask = &save_cpumask;
352 struct task_struct *kthread;
353 int next_cpu;
354
355 if (WARN_ON(hwlat_kthread))
356 return 0;
357
358 /* Just pick the first CPU on first iteration */
359 current_mask = &save_cpumask;
360 get_online_cpus();
361 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
362 put_online_cpus();
363 next_cpu = cpumask_first(current_mask);
364
365 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
366 if (IS_ERR(kthread)) {
367 pr_err(BANNER "could not start sampling thread\n");
368 return -ENOMEM;
369 }
370
371 cpumask_clear(current_mask);
372 cpumask_set_cpu(next_cpu, current_mask);
373 sched_setaffinity(kthread->pid, current_mask);
374
375 hwlat_kthread = kthread;
376 wake_up_process(kthread);
377
378 return 0;
379}
380
381/**
382 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
383 *
384 * This kicks the running hardware latency sampling/detector kernel thread and
385 * tells it to stop sampling now. Use this on unload and at system shutdown.
386 */
387static void stop_kthread(void)
388{
389 if (!hwlat_kthread)
390 return;
391 kthread_stop(hwlat_kthread);
392 hwlat_kthread = NULL;
393}
394
395/*
396 * hwlat_read - Wrapper read function for reading both window and width
397 * @filp: The active open file structure
398 * @ubuf: The userspace provided buffer to read value into
399 * @cnt: The maximum number of bytes to read
400 * @ppos: The current "file" position
401 *
402 * This function provides a generic read implementation for the global state
403 * "hwlat_data" structure filesystem entries.
404 */
405static ssize_t hwlat_read(struct file *filp, char __user *ubuf,
406 size_t cnt, loff_t *ppos)
407{
408 char buf[U64STR_SIZE];
409 u64 *entry = filp->private_data;
410 u64 val;
411 int len;
412
413 if (!entry)
414 return -EFAULT;
415
416 if (cnt > sizeof(buf))
417 cnt = sizeof(buf);
418
419 val = *entry;
420
421 len = snprintf(buf, sizeof(buf), "%llu\n", val);
422
423 return simple_read_from_buffer(ubuf, cnt, ppos, buf, len);
424}
425
426/**
427 * hwlat_width_write - Write function for "width" entry
428 * @filp: The active open file structure
429 * @ubuf: The user buffer that contains the value to write
430 * @cnt: The maximum number of bytes to write to "file"
431 * @ppos: The current position in @file
432 *
433 * This function provides a write implementation for the "width" interface
434 * to the hardware latency detector. It can be used to configure
435 * for how many us of the total window us we will actively sample for any
436 * hardware-induced latency periods. Obviously, it is not possible to
437 * sample constantly and have the system respond to a sample reader, or,
438 * worse, without having the system appear to have gone out to lunch. It
439 * is enforced that width is less that the total window size.
440 */
441static ssize_t
442hwlat_width_write(struct file *filp, const char __user *ubuf,
443 size_t cnt, loff_t *ppos)
444{
445 u64 val;
446 int err;
447
448 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
449 if (err)
450 return err;
451
452 mutex_lock(&hwlat_data.lock);
453 if (val < hwlat_data.sample_window)
454 hwlat_data.sample_width = val;
455 else
456 err = -EINVAL;
457 mutex_unlock(&hwlat_data.lock);
458
459 if (err)
460 return err;
461
462 return cnt;
463}
464
465/**
466 * hwlat_window_write - Write function for "window" entry
467 * @filp: The active open file structure
468 * @ubuf: The user buffer that contains the value to write
469 * @cnt: The maximum number of bytes to write to "file"
470 * @ppos: The current position in @file
471 *
472 * This function provides a write implementation for the "window" interface
473 * to the hardware latency detetector. The window is the total time
474 * in us that will be considered one sample period. Conceptually, windows
475 * occur back-to-back and contain a sample width period during which
476 * actual sampling occurs. Can be used to write a new total window size. It
477 * is enfoced that any value written must be greater than the sample width
478 * size, or an error results.
479 */
480static ssize_t
481hwlat_window_write(struct file *filp, const char __user *ubuf,
482 size_t cnt, loff_t *ppos)
483{
484 u64 val;
485 int err;
486
487 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
488 if (err)
489 return err;
490
491 mutex_lock(&hwlat_data.lock);
492 if (hwlat_data.sample_width < val)
493 hwlat_data.sample_window = val;
494 else
495 err = -EINVAL;
496 mutex_unlock(&hwlat_data.lock);
497
498 if (err)
499 return err;
500
501 return cnt;
502}
503
504static const struct file_operations width_fops = {
505 .open = tracing_open_generic,
506 .read = hwlat_read,
507 .write = hwlat_width_write,
508};
509
510static const struct file_operations window_fops = {
511 .open = tracing_open_generic,
512 .read = hwlat_read,
513 .write = hwlat_window_write,
514};
515
516/**
517 * init_tracefs - A function to initialize the tracefs interface files
518 *
519 * This function creates entries in tracefs for "hwlat_detector".
520 * It creates the hwlat_detector directory in the tracing directory,
521 * and within that directory is the count, width and window files to
522 * change and view those values.
523 */
524static int init_tracefs(void)
525{
526 struct dentry *d_tracer;
527 struct dentry *top_dir;
528
529 d_tracer = tracing_init_dentry();
530 if (IS_ERR(d_tracer))
531 return -ENOMEM;
532
533 top_dir = tracefs_create_dir("hwlat_detector", d_tracer);
534 if (!top_dir)
535 return -ENOMEM;
536
537 hwlat_sample_window = tracefs_create_file("window", 0640,
538 top_dir,
539 &hwlat_data.sample_window,
540 &window_fops);
541 if (!hwlat_sample_window)
542 goto err;
543
544 hwlat_sample_width = tracefs_create_file("width", 0644,
545 top_dir,
546 &hwlat_data.sample_width,
547 &width_fops);
548 if (!hwlat_sample_width)
549 goto err;
550
551 return 0;
552
553 err:
554 tracefs_remove_recursive(top_dir);
555 return -ENOMEM;
556}
557
558static void hwlat_tracer_start(struct trace_array *tr)
559{
560 int err;
561
562 err = start_kthread(tr);
563 if (err)
564 pr_err(BANNER "Cannot start hwlat kthread\n");
565}
566
567static void hwlat_tracer_stop(struct trace_array *tr)
568{
569 stop_kthread();
570}
571
572static bool hwlat_busy;
573
574static int hwlat_tracer_init(struct trace_array *tr)
575{
576 /* Only allow one instance to enable this */
577 if (hwlat_busy)
578 return -EBUSY;
579
580 hwlat_trace = tr;
581
582 disable_migrate = false;
583 hwlat_data.count = 0;
584 tr->max_latency = 0;
585 save_tracing_thresh = tracing_thresh;
586
587 /* tracing_thresh is in nsecs, we speak in usecs */
588 if (!tracing_thresh)
589 tracing_thresh = last_tracing_thresh;
590
591 if (tracer_tracing_is_on(tr))
592 hwlat_tracer_start(tr);
593
594 hwlat_busy = true;
595
596 return 0;
597}
598
599static void hwlat_tracer_reset(struct trace_array *tr)
600{
601 stop_kthread();
602
603 /* the tracing threshold is static between runs */
604 last_tracing_thresh = tracing_thresh;
605
606 tracing_thresh = save_tracing_thresh;
607 hwlat_busy = false;
608}
609
610static struct tracer hwlat_tracer __read_mostly =
611{
612 .name = "hwlat",
613 .init = hwlat_tracer_init,
614 .reset = hwlat_tracer_reset,
615 .start = hwlat_tracer_start,
616 .stop = hwlat_tracer_stop,
617 .allow_instances = true,
618};
619
620__init static int init_hwlat_tracer(void)
621{
622 int ret;
623
624 mutex_init(&hwlat_data.lock);
625
626 ret = register_tracer(&hwlat_tracer);
627 if (ret)
628 return ret;
629
630 init_tracefs();
631
632 return 0;
633}
634late_initcall(init_hwlat_tracer);
635