1 /*
2 * trace_hwlatdetect.c - A simple Hardware Latency detector.
3 *
4 * Use this tracer to detect large system latencies induced by the behavior of
5 * certain underlying system hardware or firmware, independent of Linux itself.
6 * The code was developed originally to detect the presence of SMIs on Intel
7 * and AMD systems, although there is no dependency upon x86 herein.
8 *
9 * The classical example usage of this tracer is in detecting the presence of
10 * SMIs or System Management Interrupts on Intel and AMD systems. An SMI is a
11 * somewhat special form of hardware interrupt spawned from earlier CPU debug
12 * modes in which the (BIOS/EFI/etc.) firmware arranges for the South Bridge
13 * LPC (or other device) to generate a special interrupt under certain
14 * circumstances, for example, upon expiration of a special SMI timer device,
15 * due to certain external thermal readings, on certain I/O address accesses,
16 * and other situations. An SMI hits a special CPU pin, triggers a special
17 * SMI mode (complete with special memory map), and the OS is unaware.
18 *
19 * Although certain hardware-inducing latencies are necessary (for example,
20 * a modern system often requires an SMI handler for correct thermal control
21 * and remote management) they can wreak havoc upon any OS-level performance
22 * guarantees toward low-latency, especially when the OS is not even made
23 * aware of the presence of these interrupts. For this reason, we need a
24 * somewhat brute force mechanism to detect these interrupts. In this case,
25 * we do it by hogging all of the CPU(s) for configurable timer intervals,
26 * sampling the built-in CPU timer, looking for discontiguous readings.
27 *
28 * WARNING: This implementation necessarily introduces latencies. Therefore,
29 * you should NEVER use this tracer while running in a production
30 * environment requiring any kind of low-latency performance
31 * guarantee(s).
32 *
33 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
34 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
35 *
36 * Includes useful feedback from Clark Williams <clark@redhat.com>
37 *
38 * This file is licensed under the terms of the GNU General Public
39 * License version 2. This program is licensed "as is" without any
40 * warranty of any kind, whether express or implied.
41 */
42 #include <linux/kthread.h>
43 #include <linux/tracefs.h>
44 #include <linux/uaccess.h>
45 #include <linux/cpumask.h>
46 #include <linux/delay.h>
47 #include <linux/sched/clock.h>
48 #include "trace.h"
49
50 static struct trace_array *hwlat_trace;
51
52 #define U64STR_SIZE 22 /* 20 digits max */
53
54 #define BANNER "hwlat_detector: "
55 #define DEFAULT_SAMPLE_WINDOW 1000000 /* 1s */
56 #define DEFAULT_SAMPLE_WIDTH 500000 /* 0.5s */
57 #define DEFAULT_LAT_THRESHOLD 10 /* 10us */
58
59 /* sampling thread*/
60 static struct task_struct *hwlat_kthread;
61
62 static struct dentry *hwlat_sample_width; /* sample width us */
63 static struct dentry *hwlat_sample_window; /* sample window us */
64
65 /* Save the previous tracing_thresh value */
66 static unsigned long save_tracing_thresh;
67
68 /* NMI timestamp counters */
69 static u64 nmi_ts_start;
70 static u64 nmi_total_ts;
71 static int nmi_count;
72 static int nmi_cpu;
73
74 /* Tells NMIs to call back to the hwlat tracer to record timestamps */
75 bool trace_hwlat_callback_enabled;
76
77 /* If the user changed threshold, remember it */
78 static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
79
80 /* Individual latency samples are stored here when detected. */
81 struct hwlat_sample {
82 u64 seqnum; /* unique sequence */
83 u64 duration; /* delta */
84 u64 outer_duration; /* delta (outer loop) */
85 u64 nmi_total_ts; /* Total time spent in NMIs */
86 struct timespec64 timestamp; /* wall time */
87 int nmi_count; /* # NMIs during this sample */
88 };
89
90 /* keep the global state somewhere. */
91 static struct hwlat_data {
92
93 struct mutex lock; /* protect changes */
94
95 u64 count; /* total since reset */
96
97 u64 sample_window; /* total sampling window (on+off) */
98 u64 sample_width; /* active sampling portion of window */
99
100 } hwlat_data = {
101 .sample_window = DEFAULT_SAMPLE_WINDOW,
102 .sample_width = DEFAULT_SAMPLE_WIDTH,
103 };
104
trace_hwlat_sample(struct hwlat_sample * sample)105 static void trace_hwlat_sample(struct hwlat_sample *sample)
106 {
107 struct trace_array *tr = hwlat_trace;
108 struct trace_event_call *call = &event_hwlat;
109 struct ring_buffer *buffer = tr->trace_buffer.buffer;
110 struct ring_buffer_event *event;
111 struct hwlat_entry *entry;
112 unsigned long flags;
113 int pc;
114
115 pc = preempt_count();
116 local_save_flags(flags);
117
118 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
119 flags, pc);
120 if (!event)
121 return;
122 entry = ring_buffer_event_data(event);
123 entry->seqnum = sample->seqnum;
124 entry->duration = sample->duration;
125 entry->outer_duration = sample->outer_duration;
126 entry->timestamp = sample->timestamp;
127 entry->nmi_total_ts = sample->nmi_total_ts;
128 entry->nmi_count = sample->nmi_count;
129
130 if (!call_filter_check_discard(call, entry, buffer, event))
131 trace_buffer_unlock_commit_nostack(buffer, event);
132 }
133
134 /* Macros to encapsulate the time capturing infrastructure */
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
trace_hwlat_callback(bool enter)142 void trace_hwlat_callback(bool enter)
143 {
144 if (smp_processor_id() != nmi_cpu)
145 return;
146
147 /*
148 * Currently trace_clock_local() calls sched_clock() and the
149 * generic version is not NMI safe.
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 * get_sample - sample the CPU TSC and look for likely hardware latencies
164 *
165 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
166 * hardware-induced latency. Called with interrupts disabled and with
167 * hwlat_data.lock held.
168 */
get_sample(void)169 static int get_sample(void)
170 {
171 struct trace_array *tr = hwlat_trace;
172 time_type start, t1, t2, last_t2;
173 s64 diff, total, last_total = 0;
174 u64 sample = 0;
175 u64 thresh = tracing_thresh;
176 u64 outer_sample = 0;
177 int ret = -1;
178
179 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
180
181 nmi_cpu = smp_processor_id();
182 nmi_total_ts = 0;
183 nmi_count = 0;
184 /* Make sure NMIs see this first */
185 barrier();
186
187 trace_hwlat_callback_enabled = true;
188
189 init_time(last_t2, 0);
190 start = time_get(); /* start timestamp */
191
192 do {
193
194 t1 = time_get(); /* we'll look for a discontinuity */
195 t2 = time_get();
196
197 if (time_u64(last_t2)) {
198 /* Check the delta from outer loop (t2 to next t1) */
199 diff = time_to_us(time_sub(t1, last_t2));
200 /* This shouldn't happen */
201 if (diff < 0) {
202 pr_err(BANNER "time running backwards\n");
203 goto out;
204 }
205 if (diff > outer_sample)
206 outer_sample = diff;
207 }
208 last_t2 = t2;
209
210 total = time_to_us(time_sub(t2, start)); /* sample width */
211
212 /* Check for possible overflows */
213 if (total < last_total) {
214 pr_err("Time total overflowed\n");
215 break;
216 }
217 last_total = total;
218
219 /* This checks the inner loop (t1 to t2) */
220 diff = time_to_us(time_sub(t2, t1)); /* current diff */
221
222 /* This shouldn't happen */
223 if (diff < 0) {
224 pr_err(BANNER "time running backwards\n");
225 goto out;
226 }
227
228 if (diff > sample)
229 sample = diff; /* only want highest value */
230
231 } while (total <= hwlat_data.sample_width);
232
233 barrier(); /* finish the above in the view for NMIs */
234 trace_hwlat_callback_enabled = false;
235 barrier(); /* Make sure nmi_total_ts is no longer updated */
236
237 ret = 0;
238
239 /* If we exceed the threshold value, we have found a hardware latency */
240 if (sample > thresh || outer_sample > thresh) {
241 struct hwlat_sample s;
242
243 ret = 1;
244
245 /* We read in microseconds */
246 if (nmi_total_ts)
247 do_div(nmi_total_ts, NSEC_PER_USEC);
248
249 hwlat_data.count++;
250 s.seqnum = hwlat_data.count;
251 s.duration = sample;
252 s.outer_duration = outer_sample;
253 ktime_get_real_ts64(&s.timestamp);
254 s.nmi_total_ts = nmi_total_ts;
255 s.nmi_count = nmi_count;
256 trace_hwlat_sample(&s);
257
258 /* Keep a running maximum ever recorded hardware latency */
259 if (sample > tr->max_latency)
260 tr->max_latency = sample;
261 if (outer_sample > tr->max_latency)
262 tr->max_latency = outer_sample;
263 }
264
265 out:
266 return ret;
267 }
268
269 static struct cpumask save_cpumask;
270 static bool disable_migrate;
271
move_to_next_cpu(void)272 static void move_to_next_cpu(void)
273 {
274 struct cpumask *current_mask = &save_cpumask;
275 int next_cpu;
276
277 if (disable_migrate)
278 return;
279 /*
280 * If for some reason the user modifies the CPU affinity
281 * of this thread, than stop migrating for the duration
282 * of the current test.
283 */
284 if (!cpumask_equal(current_mask, ¤t->cpus_allowed))
285 goto disable;
286
287 get_online_cpus();
288 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
289 next_cpu = cpumask_next(smp_processor_id(), current_mask);
290 put_online_cpus();
291
292 if (next_cpu >= nr_cpu_ids)
293 next_cpu = cpumask_first(current_mask);
294
295 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
296 goto disable;
297
298 cpumask_clear(current_mask);
299 cpumask_set_cpu(next_cpu, current_mask);
300
301 sched_setaffinity(0, current_mask);
302 return;
303
304 disable:
305 disable_migrate = true;
306 }
307
308 /*
309 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
310 *
311 * Used to periodically sample the CPU TSC via a call to get_sample. We
312 * disable interrupts, which does (intentionally) introduce latency since we
313 * need to ensure nothing else might be running (and thus preempting).
314 * Obviously this should never be used in production environments.
315 *
316 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
317 */
kthread_fn(void * data)318 static int kthread_fn(void *data)
319 {
320 u64 interval;
321
322 while (!kthread_should_stop()) {
323
324 move_to_next_cpu();
325
326 local_irq_disable();
327 get_sample();
328 local_irq_enable();
329
330 mutex_lock(&hwlat_data.lock);
331 interval = hwlat_data.sample_window - hwlat_data.sample_width;
332 mutex_unlock(&hwlat_data.lock);
333
334 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
335
336 /* Always sleep for at least 1ms */
337 if (interval < 1)
338 interval = 1;
339
340 if (msleep_interruptible(interval))
341 break;
342 }
343
344 return 0;
345 }
346
347 /**
348 * start_kthread - Kick off the hardware latency sampling/detector kthread
349 *
350 * This starts the kernel thread that will sit and sample the CPU timestamp
351 * counter (TSC or similar) and look for potential hardware latencies.
352 */
start_kthread(struct trace_array * tr)353 static int start_kthread(struct trace_array *tr)
354 {
355 struct cpumask *current_mask = &save_cpumask;
356 struct task_struct *kthread;
357 int next_cpu;
358
359 /* Just pick the first CPU on first iteration */
360 current_mask = &save_cpumask;
361 get_online_cpus();
362 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
363 put_online_cpus();
364 next_cpu = cpumask_first(current_mask);
365
366 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
367 if (IS_ERR(kthread)) {
368 pr_err(BANNER "could not start sampling thread\n");
369 return -ENOMEM;
370 }
371
372 cpumask_clear(current_mask);
373 cpumask_set_cpu(next_cpu, current_mask);
374 sched_setaffinity(kthread->pid, current_mask);
375
376 hwlat_kthread = kthread;
377 wake_up_process(kthread);
378
379 return 0;
380 }
381
382 /**
383 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
384 *
385 * This kicks the running hardware latency sampling/detector kernel thread and
386 * tells it to stop sampling now. Use this on unload and at system shutdown.
387 */
stop_kthread(void)388 static void stop_kthread(void)
389 {
390 if (!hwlat_kthread)
391 return;
392 kthread_stop(hwlat_kthread);
393 hwlat_kthread = NULL;
394 }
395
396 /*
397 * hwlat_read - Wrapper read function for reading both window and width
398 * @filp: The active open file structure
399 * @ubuf: The userspace provided buffer to read value into
400 * @cnt: The maximum number of bytes to read
401 * @ppos: The current "file" position
402 *
403 * This function provides a generic read implementation for the global state
404 * "hwlat_data" structure filesystem entries.
405 */
hwlat_read(struct file * filp,char __user * ubuf,size_t cnt,loff_t * ppos)406 static ssize_t hwlat_read(struct file *filp, char __user *ubuf,
407 size_t cnt, loff_t *ppos)
408 {
409 char buf[U64STR_SIZE];
410 u64 *entry = filp->private_data;
411 u64 val;
412 int len;
413
414 if (!entry)
415 return -EFAULT;
416
417 if (cnt > sizeof(buf))
418 cnt = sizeof(buf);
419
420 val = *entry;
421
422 len = snprintf(buf, sizeof(buf), "%llu\n", val);
423
424 return simple_read_from_buffer(ubuf, cnt, ppos, buf, len);
425 }
426
427 /**
428 * hwlat_width_write - Write function for "width" entry
429 * @filp: The active open file structure
430 * @ubuf: The user buffer that contains the value to write
431 * @cnt: The maximum number of bytes to write to "file"
432 * @ppos: The current position in @file
433 *
434 * This function provides a write implementation for the "width" interface
435 * to the hardware latency detector. It can be used to configure
436 * for how many us of the total window us we will actively sample for any
437 * hardware-induced latency periods. Obviously, it is not possible to
438 * sample constantly and have the system respond to a sample reader, or,
439 * worse, without having the system appear to have gone out to lunch. It
440 * is enforced that width is less that the total window size.
441 */
442 static ssize_t
hwlat_width_write(struct file * filp,const char __user * ubuf,size_t cnt,loff_t * ppos)443 hwlat_width_write(struct file *filp, const char __user *ubuf,
444 size_t cnt, loff_t *ppos)
445 {
446 u64 val;
447 int err;
448
449 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
450 if (err)
451 return err;
452
453 mutex_lock(&hwlat_data.lock);
454 if (val < hwlat_data.sample_window)
455 hwlat_data.sample_width = val;
456 else
457 err = -EINVAL;
458 mutex_unlock(&hwlat_data.lock);
459
460 if (err)
461 return err;
462
463 return cnt;
464 }
465
466 /**
467 * hwlat_window_write - Write function for "window" entry
468 * @filp: The active open file structure
469 * @ubuf: The user buffer that contains the value to write
470 * @cnt: The maximum number of bytes to write to "file"
471 * @ppos: The current position in @file
472 *
473 * This function provides a write implementation for the "window" interface
474 * to the hardware latency detetector. The window is the total time
475 * in us that will be considered one sample period. Conceptually, windows
476 * occur back-to-back and contain a sample width period during which
477 * actual sampling occurs. Can be used to write a new total window size. It
478 * is enfoced that any value written must be greater than the sample width
479 * size, or an error results.
480 */
481 static ssize_t
hwlat_window_write(struct file * filp,const char __user * ubuf,size_t cnt,loff_t * ppos)482 hwlat_window_write(struct file *filp, const char __user *ubuf,
483 size_t cnt, loff_t *ppos)
484 {
485 u64 val;
486 int err;
487
488 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
489 if (err)
490 return err;
491
492 mutex_lock(&hwlat_data.lock);
493 if (hwlat_data.sample_width < val)
494 hwlat_data.sample_window = val;
495 else
496 err = -EINVAL;
497 mutex_unlock(&hwlat_data.lock);
498
499 if (err)
500 return err;
501
502 return cnt;
503 }
504
505 static const struct file_operations width_fops = {
506 .open = tracing_open_generic,
507 .read = hwlat_read,
508 .write = hwlat_width_write,
509 };
510
511 static const struct file_operations window_fops = {
512 .open = tracing_open_generic,
513 .read = hwlat_read,
514 .write = hwlat_window_write,
515 };
516
517 /**
518 * init_tracefs - A function to initialize the tracefs interface files
519 *
520 * This function creates entries in tracefs for "hwlat_detector".
521 * It creates the hwlat_detector directory in the tracing directory,
522 * and within that directory is the count, width and window files to
523 * change and view those values.
524 */
init_tracefs(void)525 static int init_tracefs(void)
526 {
527 struct dentry *d_tracer;
528 struct dentry *top_dir;
529
530 d_tracer = tracing_init_dentry();
531 if (IS_ERR(d_tracer))
532 return -ENOMEM;
533
534 top_dir = tracefs_create_dir("hwlat_detector", d_tracer);
535 if (!top_dir)
536 return -ENOMEM;
537
538 hwlat_sample_window = tracefs_create_file("window", 0640,
539 top_dir,
540 &hwlat_data.sample_window,
541 &window_fops);
542 if (!hwlat_sample_window)
543 goto err;
544
545 hwlat_sample_width = tracefs_create_file("width", 0644,
546 top_dir,
547 &hwlat_data.sample_width,
548 &width_fops);
549 if (!hwlat_sample_width)
550 goto err;
551
552 return 0;
553
554 err:
555 tracefs_remove_recursive(top_dir);
556 return -ENOMEM;
557 }
558
hwlat_tracer_start(struct trace_array * tr)559 static void hwlat_tracer_start(struct trace_array *tr)
560 {
561 int err;
562
563 err = start_kthread(tr);
564 if (err)
565 pr_err(BANNER "Cannot start hwlat kthread\n");
566 }
567
hwlat_tracer_stop(struct trace_array * tr)568 static void hwlat_tracer_stop(struct trace_array *tr)
569 {
570 stop_kthread();
571 }
572
573 static bool hwlat_busy;
574
hwlat_tracer_init(struct trace_array * tr)575 static int hwlat_tracer_init(struct trace_array *tr)
576 {
577 /* Only allow one instance to enable this */
578 if (hwlat_busy)
579 return -EBUSY;
580
581 hwlat_trace = tr;
582
583 disable_migrate = false;
584 hwlat_data.count = 0;
585 tr->max_latency = 0;
586 save_tracing_thresh = tracing_thresh;
587
588 /* tracing_thresh is in nsecs, we speak in usecs */
589 if (!tracing_thresh)
590 tracing_thresh = last_tracing_thresh;
591
592 if (tracer_tracing_is_on(tr))
593 hwlat_tracer_start(tr);
594
595 hwlat_busy = true;
596
597 return 0;
598 }
599
hwlat_tracer_reset(struct trace_array * tr)600 static void hwlat_tracer_reset(struct trace_array *tr)
601 {
602 stop_kthread();
603
604 /* the tracing threshold is static between runs */
605 last_tracing_thresh = tracing_thresh;
606
607 tracing_thresh = save_tracing_thresh;
608 hwlat_busy = false;
609 }
610
611 static struct tracer hwlat_tracer __read_mostly =
612 {
613 .name = "hwlat",
614 .init = hwlat_tracer_init,
615 .reset = hwlat_tracer_reset,
616 .start = hwlat_tracer_start,
617 .stop = hwlat_tracer_stop,
618 .allow_instances = true,
619 };
620
init_hwlat_tracer(void)621 __init static int init_hwlat_tracer(void)
622 {
623 int ret;
624
625 mutex_init(&hwlat_data.lock);
626
627 ret = register_tracer(&hwlat_tracer);
628 if (ret)
629 return ret;
630
631 init_tracefs();
632
633 return 0;
634 }
635 late_initcall(init_hwlat_tracer);
636