1 // SPDX-License-Identifier: GPL-2.0
3 * trace_hwlat.c - A simple Hardware Latency detector.
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.
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.
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.
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
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>
37 * Includes useful feedback from Clark Williams <williams@redhat.com>
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>
48 static struct trace_array *hwlat_trace;
50 #define U64STR_SIZE 22 /* 20 digits max */
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 */
57 static struct dentry *hwlat_sample_width; /* sample width us */
58 static struct dentry *hwlat_sample_window; /* sample window us */
59 static struct dentry *hwlat_thread_mode; /* hwlat thread mode */
67 static char *thread_mode_str[] = { "none", "round-robin", "per-cpu" };
69 /* Save the previous tracing_thresh value */
70 static unsigned long save_tracing_thresh;
72 /* runtime kthread data */
73 struct hwlat_kthread_data {
74 struct task_struct *kthread;
75 /* NMI timestamp counters */
82 static struct hwlat_kthread_data hwlat_single_cpu_data;
83 static DEFINE_PER_CPU(struct hwlat_kthread_data, hwlat_per_cpu_data);
85 /* Tells NMIs to call back to the hwlat tracer to record timestamps */
86 bool trace_hwlat_callback_enabled;
88 /* If the user changed threshold, remember it */
89 static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
91 /* Individual latency samples are stored here when detected. */
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 */
102 /* keep the global state somewhere. */
103 static struct hwlat_data {
105 struct mutex lock; /* protect changes */
107 u64 count; /* total since reset */
109 u64 sample_window; /* total sampling window (on+off) */
110 u64 sample_width; /* active sampling portion of window */
112 int thread_mode; /* thread mode */
115 .sample_window = DEFAULT_SAMPLE_WINDOW,
116 .sample_width = DEFAULT_SAMPLE_WIDTH,
117 .thread_mode = MODE_ROUND_ROBIN
120 static struct hwlat_kthread_data *get_cpu_data(void)
122 if (hwlat_data.thread_mode == MODE_PER_CPU)
123 return this_cpu_ptr(&hwlat_per_cpu_data);
125 return &hwlat_single_cpu_data;
128 static bool hwlat_busy;
130 static void trace_hwlat_sample(struct hwlat_sample *sample)
132 struct trace_array *tr = hwlat_trace;
133 struct trace_event_call *call = &event_hwlat;
134 struct trace_buffer *buffer = tr->array_buffer.buffer;
135 struct ring_buffer_event *event;
136 struct hwlat_entry *entry;
138 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
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;
147 entry->nmi_total_ts = sample->nmi_total_ts;
148 entry->nmi_count = sample->nmi_count;
149 entry->count = sample->count;
151 if (!call_filter_check_discard(call, entry, buffer, event))
152 trace_buffer_unlock_commit_nostack(buffer, event);
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
163 void trace_hwlat_callback(bool enter)
165 struct hwlat_kthread_data *kdata = get_cpu_data();
171 * Currently trace_clock_local() calls sched_clock() and the
172 * generic version is not NMI safe.
174 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
176 kdata->nmi_ts_start = time_get();
178 kdata->nmi_total_ts += time_get() - kdata->nmi_ts_start;
186 * hwlat_err - report a hwlat error.
188 #define hwlat_err(msg) ({ \
189 struct trace_array *tr = hwlat_trace; \
191 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg); \
195 * get_sample - sample the CPU TSC and look for likely hardware latencies
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.
201 static int get_sample(void)
203 struct hwlat_kthread_data *kdata = get_cpu_data();
204 struct trace_array *tr = hwlat_trace;
205 struct hwlat_sample s;
206 time_type start, t1, t2, last_t2;
207 s64 diff, outer_diff, total, last_total = 0;
209 u64 thresh = tracing_thresh;
210 u64 outer_sample = 0;
212 unsigned int count = 0;
214 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
216 kdata->nmi_total_ts = 0;
217 kdata->nmi_count = 0;
218 /* Make sure NMIs see this first */
221 trace_hwlat_callback_enabled = true;
223 init_time(last_t2, 0);
224 start = time_get(); /* start timestamp */
229 t1 = time_get(); /* we'll look for a discontinuity */
232 if (time_u64(last_t2)) {
233 /* Check the delta from outer loop (t2 to next t1) */
234 outer_diff = time_to_us(time_sub(t1, last_t2));
235 /* This shouldn't happen */
236 if (outer_diff < 0) {
237 hwlat_err(BANNER "time running backwards\n");
240 if (outer_diff > outer_sample)
241 outer_sample = outer_diff;
245 total = time_to_us(time_sub(t2, start)); /* sample width */
247 /* Check for possible overflows */
248 if (total < last_total) {
249 hwlat_err("Time total overflowed\n");
254 /* This checks the inner loop (t1 to t2) */
255 diff = time_to_us(time_sub(t2, t1)); /* current diff */
257 if (diff > thresh || outer_diff > thresh) {
259 ktime_get_real_ts64(&s.timestamp);
263 /* This shouldn't happen */
265 hwlat_err(BANNER "time running backwards\n");
270 sample = diff; /* only want highest value */
272 } while (total <= hwlat_data.sample_width);
274 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 */
280 /* If we exceed the threshold value, we have found a hardware latency */
281 if (sample > thresh || outer_sample > thresh) {
286 /* We read in microseconds */
287 if (kdata->nmi_total_ts)
288 do_div(kdata->nmi_total_ts, NSEC_PER_USEC);
291 s.seqnum = hwlat_data.count;
293 s.outer_duration = outer_sample;
294 s.nmi_total_ts = kdata->nmi_total_ts;
295 s.nmi_count = kdata->nmi_count;
297 trace_hwlat_sample(&s);
299 latency = max(sample, outer_sample);
301 /* Keep a running maximum ever recorded hardware latency */
302 if (latency > tr->max_latency) {
303 tr->max_latency = latency;
304 latency_fsnotify(tr);
312 static struct cpumask save_cpumask;
314 static void move_to_next_cpu(void)
316 struct cpumask *current_mask = &save_cpumask;
317 struct trace_array *tr = hwlat_trace;
321 * If for some reason the user modifies the CPU affinity
322 * of this thread, then stop migrating for the duration
323 * of the current test.
325 if (!cpumask_equal(current_mask, current->cpus_ptr))
329 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
330 next_cpu = cpumask_next(raw_smp_processor_id(), current_mask);
333 if (next_cpu >= nr_cpu_ids)
334 next_cpu = cpumask_first(current_mask);
336 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
339 cpumask_clear(current_mask);
340 cpumask_set_cpu(next_cpu, current_mask);
342 sched_setaffinity(0, current_mask);
346 hwlat_data.thread_mode = MODE_NONE;
347 pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
351 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
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.
358 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
360 static int kthread_fn(void *data)
364 while (!kthread_should_stop()) {
366 if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
373 mutex_lock(&hwlat_data.lock);
374 interval = hwlat_data.sample_window - hwlat_data.sample_width;
375 mutex_unlock(&hwlat_data.lock);
377 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
379 /* Always sleep for at least 1ms */
383 if (msleep_interruptible(interval))
391 * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
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.
396 static void stop_single_kthread(void)
398 struct hwlat_kthread_data *kdata = get_cpu_data();
399 struct task_struct *kthread;
402 kthread = kdata->kthread;
407 kthread_stop(kthread);
408 kdata->kthread = NULL;
416 * start_single_kthread - Kick off the hardware latency sampling/detector kthread
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.
421 static int start_single_kthread(struct trace_array *tr)
423 struct hwlat_kthread_data *kdata = get_cpu_data();
424 struct cpumask *current_mask = &save_cpumask;
425 struct task_struct *kthread;
432 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
433 if (IS_ERR(kthread)) {
434 pr_err(BANNER "could not start sampling thread\n");
439 /* Just pick the first CPU on first iteration */
440 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
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);
449 sched_setaffinity(kthread->pid, current_mask);
451 kdata->kthread = kthread;
452 wake_up_process(kthread);
460 * stop_cpu_kthread - Stop a hwlat cpu kthread
462 static void stop_cpu_kthread(unsigned int cpu)
464 struct task_struct *kthread;
466 kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
468 kthread_stop(kthread);
469 per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
473 * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
475 * This kicks the running hardware latency sampling/detector kernel threads and
476 * tells it to stop sampling now. Use this on unload and at system shutdown.
478 static void stop_per_cpu_kthreads(void)
483 for_each_online_cpu(cpu)
484 stop_cpu_kthread(cpu);
489 * start_cpu_kthread - Start a hwlat cpu kthread
491 static int start_cpu_kthread(unsigned int cpu)
493 struct task_struct *kthread;
495 kthread = kthread_run_on_cpu(kthread_fn, NULL, cpu, "hwlatd/%u");
496 if (IS_ERR(kthread)) {
497 pr_err(BANNER "could not start sampling thread\n");
501 per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
506 #ifdef CONFIG_HOTPLUG_CPU
507 static void hwlat_hotplug_workfn(struct work_struct *dummy)
509 struct trace_array *tr = hwlat_trace;
510 unsigned int cpu = smp_processor_id();
512 mutex_lock(&trace_types_lock);
513 mutex_lock(&hwlat_data.lock);
516 if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
519 if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
522 start_cpu_kthread(cpu);
526 mutex_unlock(&hwlat_data.lock);
527 mutex_unlock(&trace_types_lock);
530 static DECLARE_WORK(hwlat_hotplug_work, hwlat_hotplug_workfn);
533 * hwlat_cpu_init - CPU hotplug online callback function
535 static int hwlat_cpu_init(unsigned int cpu)
537 schedule_work_on(cpu, &hwlat_hotplug_work);
542 * hwlat_cpu_die - CPU hotplug offline callback function
544 static int hwlat_cpu_die(unsigned int cpu)
546 stop_cpu_kthread(cpu);
550 static void hwlat_init_hotplug_support(void)
554 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/hwlat:online",
555 hwlat_cpu_init, hwlat_cpu_die);
557 pr_warn(BANNER "Error to init cpu hotplug support\n");
561 #else /* CONFIG_HOTPLUG_CPU */
562 static void hwlat_init_hotplug_support(void)
566 #endif /* CONFIG_HOTPLUG_CPU */
569 * start_per_cpu_kthreads - Kick off the hardware latency sampling/detector kthreads
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.
575 static int start_per_cpu_kthreads(struct trace_array *tr)
577 struct cpumask *current_mask = &save_cpumask;
583 * Run only on CPUs in which hwlat is allowed to run.
585 cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
587 for_each_online_cpu(cpu)
588 per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
590 for_each_cpu(cpu, current_mask) {
591 retval = start_cpu_kthread(cpu);
601 stop_per_cpu_kthreads();
605 static void *s_mode_start(struct seq_file *s, loff_t *pos)
609 mutex_lock(&hwlat_data.lock);
611 if (mode >= MODE_MAX)
617 static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
621 if (mode >= MODE_MAX)
627 static int s_mode_show(struct seq_file *s, void *v)
632 if (mode == hwlat_data.thread_mode)
633 seq_printf(s, "[%s]", thread_mode_str[mode]);
635 seq_printf(s, "%s", thread_mode_str[mode]);
637 if (mode != MODE_MAX)
643 static void s_mode_stop(struct seq_file *s, void *v)
646 mutex_unlock(&hwlat_data.lock);
649 static const struct seq_operations thread_mode_seq_ops = {
650 .start = s_mode_start,
656 static int hwlat_mode_open(struct inode *inode, struct file *file)
658 return seq_open(file, &thread_mode_seq_ops);
661 static void hwlat_tracer_start(struct trace_array *tr);
662 static void hwlat_tracer_stop(struct trace_array *tr);
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
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
676 * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
677 * creates one hwlatd thread per allowed CPU.
679 static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
680 size_t cnt, loff_t *ppos)
682 struct trace_array *tr = hwlat_trace;
687 if (cnt >= sizeof(buf))
690 if (copy_from_user(buf, ubuf, cnt))
695 mode = strstrip(buf);
700 * trace_types_lock is taken to avoid concurrency on start/stop
703 mutex_lock(&trace_types_lock);
705 hwlat_tracer_stop(tr);
707 mutex_lock(&hwlat_data.lock);
709 for (i = 0; i < MODE_MAX; i++) {
710 if (strcmp(mode, thread_mode_str[i]) == 0) {
711 hwlat_data.thread_mode = i;
716 mutex_unlock(&hwlat_data.lock);
719 hwlat_tracer_start(tr);
720 mutex_unlock(&trace_types_lock);
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.
734 static struct trace_min_max_param hwlat_width = {
735 .lock = &hwlat_data.lock,
736 .val = &hwlat_data.sample_width,
737 .max = &hwlat_data.sample_window,
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.
746 static struct trace_min_max_param hwlat_window = {
747 .lock = &hwlat_data.lock,
748 .val = &hwlat_data.sample_window,
750 .min = &hwlat_data.sample_width,
753 static const struct file_operations thread_mode_fops = {
754 .open = hwlat_mode_open,
757 .release = seq_release,
758 .write = hwlat_mode_write
761 * init_tracefs - A function to initialize the tracefs interface files
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.
768 static int init_tracefs(void)
771 struct dentry *top_dir;
773 ret = tracing_init_dentry();
777 top_dir = tracefs_create_dir("hwlat_detector", NULL);
781 hwlat_sample_window = tracefs_create_file("window", TRACE_MODE_WRITE,
784 &trace_min_max_fops);
785 if (!hwlat_sample_window)
788 hwlat_sample_width = tracefs_create_file("width", TRACE_MODE_WRITE,
791 &trace_min_max_fops);
792 if (!hwlat_sample_width)
795 hwlat_thread_mode = trace_create_file("mode", TRACE_MODE_WRITE,
799 if (!hwlat_thread_mode)
805 tracefs_remove(top_dir);
809 static void hwlat_tracer_start(struct trace_array *tr)
813 if (hwlat_data.thread_mode == MODE_PER_CPU)
814 err = start_per_cpu_kthreads(tr);
816 err = start_single_kthread(tr);
818 pr_err(BANNER "Cannot start hwlat kthread\n");
821 static void hwlat_tracer_stop(struct trace_array *tr)
823 if (hwlat_data.thread_mode == MODE_PER_CPU)
824 stop_per_cpu_kthreads();
826 stop_single_kthread();
829 static int hwlat_tracer_init(struct trace_array *tr)
831 /* Only allow one instance to enable this */
837 hwlat_data.count = 0;
839 save_tracing_thresh = tracing_thresh;
841 /* tracing_thresh is in nsecs, we speak in usecs */
843 tracing_thresh = last_tracing_thresh;
845 if (tracer_tracing_is_on(tr))
846 hwlat_tracer_start(tr);
853 static void hwlat_tracer_reset(struct trace_array *tr)
855 hwlat_tracer_stop(tr);
857 /* the tracing threshold is static between runs */
858 last_tracing_thresh = tracing_thresh;
860 tracing_thresh = save_tracing_thresh;
864 static struct tracer hwlat_tracer __read_mostly =
867 .init = hwlat_tracer_init,
868 .reset = hwlat_tracer_reset,
869 .start = hwlat_tracer_start,
870 .stop = hwlat_tracer_stop,
871 .allow_instances = true,
874 __init static int init_hwlat_tracer(void)
878 mutex_init(&hwlat_data.lock);
880 ret = register_tracer(&hwlat_tracer);
884 hwlat_init_hotplug_support();
890 late_initcall(init_hwlat_tracer);