blob: d440ddd5fd8b29adf605075004fc5edd9a7ee17c [file] [log] [blame]
Steven Rostedt (VMware)bcea3f92018-08-16 11:23:53 -04001// SPDX-License-Identifier: GPL-2.0
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -04002/*
Srivatsa S. Bhat (VMware)0c3c86b2019-10-10 11:51:17 -07003 * trace_hwlat.c - A simple Hardware Latency detector.
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -04004 *
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 *
Daniel Bristot de Oliveirabb1b24c2021-06-22 16:42:19 +020037 * Includes useful feedback from Clark Williams <williams@redhat.com>
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040038 *
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040039 */
40#include <linux/kthread.h>
41#include <linux/tracefs.h>
42#include <linux/uaccess.h>
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -040043#include <linux/cpumask.h>
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040044#include <linux/delay.h>
Ingo Molnare6017572017-02-01 16:36:40 +010045#include <linux/sched/clock.h>
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040046#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
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040057static struct dentry *hwlat_sample_width; /* sample width us */
58static struct dentry *hwlat_sample_window; /* sample window us */
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +020059static struct dentry *hwlat_thread_mode; /* hwlat thread mode */
60
61enum {
62 MODE_NONE = 0,
63 MODE_ROUND_ROBIN,
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +020064 MODE_PER_CPU,
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +020065 MODE_MAX
66};
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +020067static char *thread_mode_str[] = { "none", "round-robin", "per-cpu" };
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040068
69/* Save the previous tracing_thresh value */
70static unsigned long save_tracing_thresh;
71
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +020072/* 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
Wang ShaoBo1d628892021-10-21 11:52:25 +080082static struct hwlat_kthread_data hwlat_single_cpu_data;
83static DEFINE_PER_CPU(struct hwlat_kthread_data, hwlat_per_cpu_data);
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -040084
85/* Tells NMIs to call back to the hwlat tracer to record timestamps */
86bool trace_hwlat_callback_enabled;
87
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -040088/* 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 {
Deepa Dinamani51aad0a2017-05-08 15:59:13 -070093 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 */
Ingo Molnarf2cc0202021-03-23 18:49:35 +010099 int count; /* # of iterations over thresh */
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400100};
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
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200112 int thread_mode; /* thread mode */
113
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400114} hwlat_data = {
115 .sample_window = DEFAULT_SAMPLE_WINDOW,
116 .sample_width = DEFAULT_SAMPLE_WIDTH,
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200117 .thread_mode = MODE_ROUND_ROBIN
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400118};
119
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200120static 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
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200128static bool hwlat_busy;
129
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400130static void trace_hwlat_sample(struct hwlat_sample *sample)
131{
132 struct trace_array *tr = hwlat_trace;
133 struct trace_event_call *call = &event_hwlat;
Steven Rostedt (VMware)13292492019-12-13 13:58:57 -0500134 struct trace_buffer *buffer = tr->array_buffer.buffer;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400135 struct ring_buffer_event *event;
136 struct hwlat_entry *entry;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400137
138 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
Sebastian Andrzej Siewior36590c502021-01-25 20:45:08 +0100139 tracing_gen_ctx());
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400140 if (!event)
141 return;
142 entry = ring_buffer_event_data(event);
143 entry->seqnum = sample->seqnum;
144 entry->duration = sample->duration;
145 entry->outer_duration = sample->outer_duration;
146 entry->timestamp = sample->timestamp;
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400147 entry->nmi_total_ts = sample->nmi_total_ts;
148 entry->nmi_count = sample->nmi_count;
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500149 entry->count = sample->count;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400150
151 if (!call_filter_check_discard(call, entry, buffer, event))
Steven Rostedt (Red Hat)52ffabe32016-11-23 20:28:38 -0500152 trace_buffer_unlock_commit_nostack(buffer, event);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400153}
154
155/* Macros to encapsulate the time capturing infrastructure */
156#define time_type u64
157#define time_get() trace_clock_local()
158#define time_to_us(x) div_u64(x, 1000)
159#define time_sub(a, b) ((a) - (b))
160#define init_time(a, b) (a = b)
161#define time_u64(a) a
162
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400163void trace_hwlat_callback(bool enter)
164{
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200165 struct hwlat_kthread_data *kdata = get_cpu_data();
166
167 if (!kdata->kthread)
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400168 return;
169
170 /*
171 * Currently trace_clock_local() calls sched_clock() and the
172 * generic version is not NMI safe.
173 */
174 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
175 if (enter)
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200176 kdata->nmi_ts_start = time_get();
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400177 else
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200178 kdata->nmi_total_ts += time_get() - kdata->nmi_ts_start;
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400179 }
180
181 if (enter)
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200182 kdata->nmi_count++;
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400183}
184
Daniel Bristot de Oliveiraaa892f82021-06-22 16:42:25 +0200185/*
186 * hwlat_err - report a hwlat error.
187 */
188#define hwlat_err(msg) ({ \
189 struct trace_array *tr = hwlat_trace; \
190 \
191 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg); \
192})
193
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400194/**
195 * get_sample - sample the CPU TSC and look for likely hardware latencies
196 *
197 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
198 * hardware-induced latency. Called with interrupts disabled and with
199 * hwlat_data.lock held.
200 */
201static int get_sample(void)
202{
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200203 struct hwlat_kthread_data *kdata = get_cpu_data();
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400204 struct trace_array *tr = hwlat_trace;
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500205 struct hwlat_sample s;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400206 time_type start, t1, t2, last_t2;
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500207 s64 diff, outer_diff, total, last_total = 0;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400208 u64 sample = 0;
209 u64 thresh = tracing_thresh;
210 u64 outer_sample = 0;
211 int ret = -1;
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500212 unsigned int count = 0;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400213
214 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
215
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200216 kdata->nmi_total_ts = 0;
217 kdata->nmi_count = 0;
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400218 /* Make sure NMIs see this first */
219 barrier();
220
221 trace_hwlat_callback_enabled = true;
222
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400223 init_time(last_t2, 0);
224 start = time_get(); /* start timestamp */
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500225 outer_diff = 0;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400226
227 do {
228
229 t1 = time_get(); /* we'll look for a discontinuity */
230 t2 = time_get();
231
232 if (time_u64(last_t2)) {
233 /* Check the delta from outer loop (t2 to next t1) */
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500234 outer_diff = time_to_us(time_sub(t1, last_t2));
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400235 /* This shouldn't happen */
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500236 if (outer_diff < 0) {
Daniel Bristot de Oliveiraaa892f82021-06-22 16:42:25 +0200237 hwlat_err(BANNER "time running backwards\n");
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400238 goto out;
239 }
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500240 if (outer_diff > outer_sample)
241 outer_sample = outer_diff;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400242 }
243 last_t2 = t2;
244
245 total = time_to_us(time_sub(t2, start)); /* sample width */
246
247 /* Check for possible overflows */
248 if (total < last_total) {
Daniel Bristot de Oliveiraaa892f82021-06-22 16:42:25 +0200249 hwlat_err("Time total overflowed\n");
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400250 break;
251 }
252 last_total = total;
253
254 /* This checks the inner loop (t1 to t2) */
255 diff = time_to_us(time_sub(t2, t1)); /* current diff */
256
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500257 if (diff > thresh || outer_diff > thresh) {
258 if (!count)
259 ktime_get_real_ts64(&s.timestamp);
260 count++;
261 }
262
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400263 /* This shouldn't happen */
264 if (diff < 0) {
Daniel Bristot de Oliveiraaa892f82021-06-22 16:42:25 +0200265 hwlat_err(BANNER "time running backwards\n");
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400266 goto out;
267 }
268
269 if (diff > sample)
270 sample = diff; /* only want highest value */
271
272 } while (total <= hwlat_data.sample_width);
273
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400274 barrier(); /* finish the above in the view for NMIs */
275 trace_hwlat_callback_enabled = false;
276 barrier(); /* Make sure nmi_total_ts is no longer updated */
277
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400278 ret = 0;
279
280 /* If we exceed the threshold value, we have found a hardware latency */
281 if (sample > thresh || outer_sample > thresh) {
Viktor Rosendahl (BMW)91edde22019-10-09 00:08:21 +0200282 u64 latency;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400283
284 ret = 1;
285
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400286 /* We read in microseconds */
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200287 if (kdata->nmi_total_ts)
288 do_div(kdata->nmi_total_ts, NSEC_PER_USEC);
Steven Rostedt (Red Hat)7b2c8622016-08-04 12:49:53 -0400289
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400290 hwlat_data.count++;
291 s.seqnum = hwlat_data.count;
292 s.duration = sample;
293 s.outer_duration = outer_sample;
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200294 s.nmi_total_ts = kdata->nmi_total_ts;
295 s.nmi_count = kdata->nmi_count;
Steven Rostedt (VMware)b396bfd2020-02-12 12:21:03 -0500296 s.count = count;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400297 trace_hwlat_sample(&s);
298
Viktor Rosendahl (BMW)91edde22019-10-09 00:08:21 +0200299 latency = max(sample, outer_sample);
300
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400301 /* Keep a running maximum ever recorded hardware latency */
Viktor Rosendahl (BMW)91edde22019-10-09 00:08:21 +0200302 if (latency > tr->max_latency) {
303 tr->max_latency = latency;
304 latency_fsnotify(tr);
305 }
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400306 }
307
308out:
309 return ret;
310}
311
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400312static struct cpumask save_cpumask;
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400313
Steven Rostedt (VMware)f447c192017-01-31 16:48:23 -0500314static void move_to_next_cpu(void)
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400315{
Steven Rostedt (VMware)f447c192017-01-31 16:48:23 -0500316 struct cpumask *current_mask = &save_cpumask;
Kevin Hao96b48332020-07-30 16:23:18 +0800317 struct trace_array *tr = hwlat_trace;
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400318 int next_cpu;
319
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400320 /*
321 * If for some reason the user modifies the CPU affinity
Srivatsa S. Bhat (VMware)0c3c86b2019-10-10 11:51:17 -0700322 * of this thread, then stop migrating for the duration
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400323 * of the current test.
324 */
Sebastian Andrzej Siewior3bd37062019-04-23 16:26:36 +0200325 if (!cpumask_equal(current_mask, current->cpus_ptr))
Daniel Bristot de Oliveira7bb7d802021-06-22 16:42:21 +0200326 goto change_mode;
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400327
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200328 cpus_read_lock();
Kevin Hao96b48332020-07-30 16:23:18 +0800329 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
Steven Rostedt (VMware)51397dc2021-08-04 14:18:48 -0400330 next_cpu = cpumask_next(raw_smp_processor_id(), current_mask);
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200331 cpus_read_unlock();
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400332
333 if (next_cpu >= nr_cpu_ids)
334 next_cpu = cpumask_first(current_mask);
335
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400336 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
Daniel Bristot de Oliveira7bb7d802021-06-22 16:42:21 +0200337 goto change_mode;
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400338
339 cpumask_clear(current_mask);
340 cpumask_set_cpu(next_cpu, current_mask);
341
342 sched_setaffinity(0, current_mask);
343 return;
344
Daniel Bristot de Oliveira7bb7d802021-06-22 16:42:21 +0200345 change_mode:
346 hwlat_data.thread_mode = MODE_NONE;
347 pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400348}
349
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400350/*
351 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
352 *
353 * Used to periodically sample the CPU TSC via a call to get_sample. We
354 * disable interrupts, which does (intentionally) introduce latency since we
355 * need to ensure nothing else might be running (and thus preempting).
356 * Obviously this should never be used in production environments.
357 *
Luiz Capitulino8e0f1142017-02-13 12:25:17 -0500358 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400359 */
360static int kthread_fn(void *data)
361{
362 u64 interval;
363
364 while (!kthread_should_stop()) {
365
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200366 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
367 move_to_next_cpu();
Steven Rostedt (Red Hat)0330f7a2016-07-15 15:48:56 -0400368
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400369 local_irq_disable();
370 get_sample();
371 local_irq_enable();
372
373 mutex_lock(&hwlat_data.lock);
374 interval = hwlat_data.sample_window - hwlat_data.sample_width;
375 mutex_unlock(&hwlat_data.lock);
376
377 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
378
379 /* Always sleep for at least 1ms */
380 if (interval < 1)
381 interval = 1;
382
383 if (msleep_interruptible(interval))
384 break;
385 }
386
387 return 0;
388}
389
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200390/*
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200391 * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
392 *
393 * This kicks the running hardware latency sampling/detector kernel thread and
394 * tells it to stop sampling now. Use this on unload and at system shutdown.
395 */
396static void stop_single_kthread(void)
397{
398 struct hwlat_kthread_data *kdata = get_cpu_data();
Daniel Bristot de Oliveira039a6022021-06-22 16:42:29 +0200399 struct task_struct *kthread;
400
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200401 cpus_read_lock();
Daniel Bristot de Oliveira039a6022021-06-22 16:42:29 +0200402 kthread = kdata->kthread;
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200403
404 if (!kthread)
Daniel Bristot de Oliveira039a6022021-06-22 16:42:29 +0200405 goto out_put_cpus;
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200406
407 kthread_stop(kthread);
408 kdata->kthread = NULL;
Daniel Bristot de Oliveira039a6022021-06-22 16:42:29 +0200409
410out_put_cpus:
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200411 cpus_read_unlock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200412}
413
414
415/*
416 * start_single_kthread - Kick off the hardware latency sampling/detector kthread
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400417 *
418 * This starts the kernel thread that will sit and sample the CPU timestamp
419 * counter (TSC or similar) and look for potential hardware latencies.
420 */
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200421static int start_single_kthread(struct trace_array *tr)
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400422{
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200423 struct hwlat_kthread_data *kdata = get_cpu_data();
Steven Rostedt (VMware)f447c192017-01-31 16:48:23 -0500424 struct cpumask *current_mask = &save_cpumask;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400425 struct task_struct *kthread;
Steven Rostedt (VMware)f447c192017-01-31 16:48:23 -0500426 int next_cpu;
427
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200428 cpus_read_lock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200429 if (kdata->kthread)
Daniel Bristot de Oliveira039a6022021-06-22 16:42:29 +0200430 goto out_put_cpus;
Erica Bugden82fbc8c2018-08-01 12:45:54 +0200431
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400432 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
433 if (IS_ERR(kthread)) {
434 pr_err(BANNER "could not start sampling thread\n");
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200435 cpus_read_unlock();
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400436 return -ENOMEM;
437 }
Steven Rostedt (VMware)f447c192017-01-31 16:48:23 -0500438
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200439 /* Just pick the first CPU on first iteration */
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200440 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200441
442 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN) {
443 next_cpu = cpumask_first(current_mask);
444 cpumask_clear(current_mask);
445 cpumask_set_cpu(next_cpu, current_mask);
446
447 }
448
Steven Rostedt (VMware)f447c192017-01-31 16:48:23 -0500449 sched_setaffinity(kthread->pid, current_mask);
450
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200451 kdata->kthread = kthread;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400452 wake_up_process(kthread);
453
Daniel Bristot de Oliveira039a6022021-06-22 16:42:29 +0200454out_put_cpus:
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200455 cpus_read_unlock();
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400456 return 0;
457}
458
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200459/*
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200460 * stop_cpu_kthread - Stop a hwlat cpu kthread
461 */
462static void stop_cpu_kthread(unsigned int cpu)
463{
464 struct task_struct *kthread;
465
466 kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
467 if (kthread)
468 kthread_stop(kthread);
Daniel Bristot de Oliveiraba998f72021-06-22 16:42:31 +0200469 per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200470}
471
472/*
473 * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400474 *
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200475 * This kicks the running hardware latency sampling/detector kernel threads and
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400476 * tells it to stop sampling now. Use this on unload and at system shutdown.
477 */
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200478static void stop_per_cpu_kthreads(void)
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400479{
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200480 unsigned int cpu;
481
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200482 cpus_read_lock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200483 for_each_online_cpu(cpu)
484 stop_cpu_kthread(cpu);
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200485 cpus_read_unlock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200486}
487
488/*
489 * start_cpu_kthread - Start a hwlat cpu kthread
490 */
491static int start_cpu_kthread(unsigned int cpu)
492{
493 struct task_struct *kthread;
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200494
Cai Huoqingff78f662022-01-14 14:03:10 -0800495 kthread = kthread_run_on_cpu(kthread_fn, NULL, cpu, "hwlatd/%u");
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200496 if (IS_ERR(kthread)) {
497 pr_err(BANNER "could not start sampling thread\n");
498 return -ENOMEM;
499 }
500
501 per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200502
503 return 0;
504}
505
Daniel Bristot de Oliveiraba998f72021-06-22 16:42:31 +0200506#ifdef CONFIG_HOTPLUG_CPU
507static void hwlat_hotplug_workfn(struct work_struct *dummy)
508{
509 struct trace_array *tr = hwlat_trace;
510 unsigned int cpu = smp_processor_id();
511
512 mutex_lock(&trace_types_lock);
513 mutex_lock(&hwlat_data.lock);
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200514 cpus_read_lock();
Daniel Bristot de Oliveiraba998f72021-06-22 16:42:31 +0200515
516 if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
517 goto out_unlock;
518
519 if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
520 goto out_unlock;
521
522 start_cpu_kthread(cpu);
523
524out_unlock:
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200525 cpus_read_unlock();
Daniel Bristot de Oliveiraba998f72021-06-22 16:42:31 +0200526 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
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200568/*
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
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200581 cpus_read_lock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200582 /*
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_online_cpu(cpu)
588 per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
589
590 for_each_cpu(cpu, current_mask) {
591 retval = start_cpu_kthread(cpu);
592 if (retval)
593 goto out_error;
594 }
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200595 cpus_read_unlock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200596
597 return 0;
598
599out_error:
Sebastian Andrzej Siewior99c37d12021-08-03 16:16:19 +0200600 cpus_read_unlock();
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200601 stop_per_cpu_kthreads();
602 return retval;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400603}
604
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200605static void *s_mode_start(struct seq_file *s, loff_t *pos)
606{
607 int mode = *pos;
608
609 mutex_lock(&hwlat_data.lock);
610
611 if (mode >= MODE_MAX)
612 return NULL;
613
614 return pos;
615}
616
617static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
618{
619 int mode = ++(*pos);
620
621 if (mode >= MODE_MAX)
622 return NULL;
623
624 return pos;
625}
626
627static int s_mode_show(struct seq_file *s, void *v)
628{
629 loff_t *pos = v;
630 int mode = *pos;
631
632 if (mode == hwlat_data.thread_mode)
633 seq_printf(s, "[%s]", thread_mode_str[mode]);
634 else
635 seq_printf(s, "%s", thread_mode_str[mode]);
636
637 if (mode != MODE_MAX)
638 seq_puts(s, " ");
639
640 return 0;
641}
642
643static void s_mode_stop(struct seq_file *s, void *v)
644{
645 seq_puts(s, "\n");
646 mutex_unlock(&hwlat_data.lock);
647}
648
649static const struct seq_operations thread_mode_seq_ops = {
650 .start = s_mode_start,
651 .next = s_mode_next,
652 .show = s_mode_show,
653 .stop = s_mode_stop
654};
655
656static int hwlat_mode_open(struct inode *inode, struct file *file)
657{
658 return seq_open(file, &thread_mode_seq_ops);
659};
660
661static void hwlat_tracer_start(struct trace_array *tr);
662static void hwlat_tracer_stop(struct trace_array *tr);
663
664/**
665 * hwlat_mode_write - Write function for "mode" entry
666 * @filp: The active open file structure
667 * @ubuf: The user buffer that contains the value to write
668 * @cnt: The maximum number of bytes to write to "file"
669 * @ppos: The current position in @file
670 *
671 * This function provides a write implementation for the "mode" interface
672 * to the hardware latency detector. hwlatd has different operation modes.
673 * The "none" sets the allowed cpumask for a single hwlatd thread at the
674 * startup and lets the scheduler handle the migration. The default mode is
675 * the "round-robin" one, in which a single hwlatd thread runs, migrating
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200676 * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
677 * creates one hwlatd thread per allowed CPU.
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200678 */
679static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
680 size_t cnt, loff_t *ppos)
681{
682 struct trace_array *tr = hwlat_trace;
683 const char *mode;
684 char buf[64];
685 int ret, i;
686
687 if (cnt >= sizeof(buf))
688 return -EINVAL;
689
690 if (copy_from_user(buf, ubuf, cnt))
691 return -EFAULT;
692
693 buf[cnt] = 0;
694
695 mode = strstrip(buf);
696
697 ret = -EINVAL;
698
699 /*
700 * trace_types_lock is taken to avoid concurrency on start/stop
701 * and hwlat_busy.
702 */
703 mutex_lock(&trace_types_lock);
704 if (hwlat_busy)
705 hwlat_tracer_stop(tr);
706
707 mutex_lock(&hwlat_data.lock);
708
709 for (i = 0; i < MODE_MAX; i++) {
710 if (strcmp(mode, thread_mode_str[i]) == 0) {
711 hwlat_data.thread_mode = i;
712 ret = cnt;
713 }
714 }
715
716 mutex_unlock(&hwlat_data.lock);
717
718 if (hwlat_busy)
719 hwlat_tracer_start(tr);
720 mutex_unlock(&trace_types_lock);
721
722 *ppos += cnt;
723
724
725
726 return ret;
727}
728
Daniel Bristot de Oliveiraf27a1c92021-06-22 16:42:24 +0200729/*
730 * The width parameter is read/write using the generic trace_min_max_param
731 * method. The *val is protected by the hwlat_data lock and is upper
732 * bounded by the window parameter.
733 */
734static struct trace_min_max_param hwlat_width = {
735 .lock = &hwlat_data.lock,
736 .val = &hwlat_data.sample_width,
737 .max = &hwlat_data.sample_window,
738 .min = NULL,
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400739};
740
Daniel Bristot de Oliveiraf27a1c92021-06-22 16:42:24 +0200741/*
742 * The window parameter is read/write using the generic trace_min_max_param
743 * method. The *val is protected by the hwlat_data lock and is lower
744 * bounded by the width parameter.
745 */
746static struct trace_min_max_param hwlat_window = {
747 .lock = &hwlat_data.lock,
748 .val = &hwlat_data.sample_window,
749 .max = NULL,
750 .min = &hwlat_data.sample_width,
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400751};
752
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200753static const struct file_operations thread_mode_fops = {
754 .open = hwlat_mode_open,
755 .read = seq_read,
756 .llseek = seq_lseek,
757 .release = seq_release,
758 .write = hwlat_mode_write
759};
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400760/**
761 * init_tracefs - A function to initialize the tracefs interface files
762 *
763 * This function creates entries in tracefs for "hwlat_detector".
764 * It creates the hwlat_detector directory in the tracing directory,
765 * and within that directory is the count, width and window files to
766 * change and view those values.
767 */
768static int init_tracefs(void)
769{
Wei Yang22c36b12020-07-12 09:10:36 +0800770 int ret;
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400771 struct dentry *top_dir;
772
Wei Yang22c36b12020-07-12 09:10:36 +0800773 ret = tracing_init_dentry();
774 if (ret)
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400775 return -ENOMEM;
776
Wei Yang22c36b12020-07-12 09:10:36 +0800777 top_dir = tracefs_create_dir("hwlat_detector", NULL);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400778 if (!top_dir)
779 return -ENOMEM;
780
Steven Rostedt (VMware)21ccc9c2021-08-18 11:24:51 -0400781 hwlat_sample_window = tracefs_create_file("window", TRACE_MODE_WRITE,
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400782 top_dir,
Daniel Bristot de Oliveiraf27a1c92021-06-22 16:42:24 +0200783 &hwlat_window,
784 &trace_min_max_fops);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400785 if (!hwlat_sample_window)
786 goto err;
787
Steven Rostedt (VMware)21ccc9c2021-08-18 11:24:51 -0400788 hwlat_sample_width = tracefs_create_file("width", TRACE_MODE_WRITE,
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400789 top_dir,
Daniel Bristot de Oliveiraf27a1c92021-06-22 16:42:24 +0200790 &hwlat_width,
791 &trace_min_max_fops);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400792 if (!hwlat_sample_width)
793 goto err;
794
Steven Rostedt (VMware)21ccc9c2021-08-18 11:24:51 -0400795 hwlat_thread_mode = trace_create_file("mode", TRACE_MODE_WRITE,
Daniel Bristot de Oliveira8fa826b2021-06-22 16:42:20 +0200796 top_dir,
797 NULL,
798 &thread_mode_fops);
799 if (!hwlat_thread_mode)
800 goto err;
801
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400802 return 0;
803
804 err:
Al Viroa3d1e7e2019-11-18 09:43:10 -0500805 tracefs_remove(top_dir);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400806 return -ENOMEM;
807}
808
809static void hwlat_tracer_start(struct trace_array *tr)
810{
811 int err;
812
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200813 if (hwlat_data.thread_mode == MODE_PER_CPU)
814 err = start_per_cpu_kthreads(tr);
815 else
816 err = start_single_kthread(tr);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400817 if (err)
818 pr_err(BANNER "Cannot start hwlat kthread\n");
819}
820
821static void hwlat_tracer_stop(struct trace_array *tr)
822{
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200823 if (hwlat_data.thread_mode == MODE_PER_CPU)
824 stop_per_cpu_kthreads();
825 else
826 stop_single_kthread();
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400827}
828
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400829static int hwlat_tracer_init(struct trace_array *tr)
830{
831 /* Only allow one instance to enable this */
832 if (hwlat_busy)
833 return -EBUSY;
834
835 hwlat_trace = tr;
836
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400837 hwlat_data.count = 0;
838 tr->max_latency = 0;
839 save_tracing_thresh = tracing_thresh;
840
841 /* tracing_thresh is in nsecs, we speak in usecs */
842 if (!tracing_thresh)
843 tracing_thresh = last_tracing_thresh;
844
845 if (tracer_tracing_is_on(tr))
846 hwlat_tracer_start(tr);
847
848 hwlat_busy = true;
849
850 return 0;
851}
852
853static void hwlat_tracer_reset(struct trace_array *tr)
854{
Daniel Bristot de Oliveiraf46b1652021-06-22 16:42:22 +0200855 hwlat_tracer_stop(tr);
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400856
857 /* the tracing threshold is static between runs */
858 last_tracing_thresh = tracing_thresh;
859
860 tracing_thresh = save_tracing_thresh;
861 hwlat_busy = false;
862}
863
864static struct tracer hwlat_tracer __read_mostly =
865{
866 .name = "hwlat",
867 .init = hwlat_tracer_init,
868 .reset = hwlat_tracer_reset,
869 .start = hwlat_tracer_start,
870 .stop = hwlat_tracer_stop,
871 .allow_instances = true,
872};
873
874__init static int init_hwlat_tracer(void)
875{
876 int ret;
877
878 mutex_init(&hwlat_data.lock);
879
880 ret = register_tracer(&hwlat_tracer);
881 if (ret)
882 return ret;
883
Daniel Bristot de Oliveiraba998f72021-06-22 16:42:31 +0200884 hwlat_init_hotplug_support();
885
Steven Rostedt (Red Hat)e7c15cd2016-06-23 12:45:36 -0400886 init_tracefs();
887
888 return 0;
889}
890late_initcall(init_hwlat_tracer);