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 <clark@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 */
58 static struct task_struct
*hwlat_kthread
;
60 static struct dentry
*hwlat_sample_width
; /* sample width us */
61 static struct dentry
*hwlat_sample_window
; /* sample window us */
63 /* Save the previous tracing_thresh value */
64 static unsigned long save_tracing_thresh
;
66 /* NMI timestamp counters */
67 static u64 nmi_ts_start
;
68 static u64 nmi_total_ts
;
72 /* Tells NMIs to call back to the hwlat tracer to record timestamps */
73 bool trace_hwlat_callback_enabled
;
75 /* If the user changed threshold, remember it */
76 static u64 last_tracing_thresh
= DEFAULT_LAT_THRESHOLD
* NSEC_PER_USEC
;
78 /* Individual latency samples are stored here when detected. */
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 int count
; /* # of iteratons over threash */
89 /* keep the global state somewhere. */
90 static struct hwlat_data
{
92 struct mutex lock
; /* protect changes */
94 u64 count
; /* total since reset */
96 u64 sample_window
; /* total sampling window (on+off) */
97 u64 sample_width
; /* active sampling portion of window */
100 .sample_window
= DEFAULT_SAMPLE_WINDOW
,
101 .sample_width
= DEFAULT_SAMPLE_WIDTH
,
104 static void trace_hwlat_sample(struct hwlat_sample
*sample
)
106 struct trace_array
*tr
= hwlat_trace
;
107 struct trace_event_call
*call
= &event_hwlat
;
108 struct trace_buffer
*buffer
= tr
->array_buffer
.buffer
;
109 struct ring_buffer_event
*event
;
110 struct hwlat_entry
*entry
;
112 event
= trace_buffer_lock_reserve(buffer
, TRACE_HWLAT
, sizeof(*entry
),
116 entry
= ring_buffer_event_data(event
);
117 entry
->seqnum
= sample
->seqnum
;
118 entry
->duration
= sample
->duration
;
119 entry
->outer_duration
= sample
->outer_duration
;
120 entry
->timestamp
= sample
->timestamp
;
121 entry
->nmi_total_ts
= sample
->nmi_total_ts
;
122 entry
->nmi_count
= sample
->nmi_count
;
123 entry
->count
= sample
->count
;
125 if (!call_filter_check_discard(call
, entry
, buffer
, event
))
126 trace_buffer_unlock_commit_nostack(buffer
, event
);
129 /* Macros to encapsulate the time capturing infrastructure */
130 #define time_type u64
131 #define time_get() trace_clock_local()
132 #define time_to_us(x) div_u64(x, 1000)
133 #define time_sub(a, b) ((a) - (b))
134 #define init_time(a, b) (a = b)
135 #define time_u64(a) a
137 void trace_hwlat_callback(bool enter
)
139 if (smp_processor_id() != nmi_cpu
)
143 * Currently trace_clock_local() calls sched_clock() and the
144 * generic version is not NMI safe.
146 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK
)) {
148 nmi_ts_start
= time_get();
150 nmi_total_ts
+= time_get() - nmi_ts_start
;
158 * get_sample - sample the CPU TSC and look for likely hardware latencies
160 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
161 * hardware-induced latency. Called with interrupts disabled and with
162 * hwlat_data.lock held.
164 static int get_sample(void)
166 struct trace_array
*tr
= hwlat_trace
;
167 struct hwlat_sample s
;
168 time_type start
, t1
, t2
, last_t2
;
169 s64 diff
, outer_diff
, total
, last_total
= 0;
171 u64 thresh
= tracing_thresh
;
172 u64 outer_sample
= 0;
174 unsigned int count
= 0;
176 do_div(thresh
, NSEC_PER_USEC
); /* modifies interval value */
178 nmi_cpu
= smp_processor_id();
181 /* Make sure NMIs see this first */
184 trace_hwlat_callback_enabled
= true;
186 init_time(last_t2
, 0);
187 start
= time_get(); /* start timestamp */
192 t1
= time_get(); /* we'll look for a discontinuity */
195 if (time_u64(last_t2
)) {
196 /* Check the delta from outer loop (t2 to next t1) */
197 outer_diff
= time_to_us(time_sub(t1
, last_t2
));
198 /* This shouldn't happen */
199 if (outer_diff
< 0) {
200 pr_err(BANNER
"time running backwards\n");
203 if (outer_diff
> outer_sample
)
204 outer_sample
= outer_diff
;
208 total
= time_to_us(time_sub(t2
, start
)); /* sample width */
210 /* Check for possible overflows */
211 if (total
< last_total
) {
212 pr_err("Time total overflowed\n");
217 /* This checks the inner loop (t1 to t2) */
218 diff
= time_to_us(time_sub(t2
, t1
)); /* current diff */
220 if (diff
> thresh
|| outer_diff
> thresh
) {
222 ktime_get_real_ts64(&s
.timestamp
);
226 /* This shouldn't happen */
228 pr_err(BANNER
"time running backwards\n");
233 sample
= diff
; /* only want highest value */
235 } while (total
<= hwlat_data
.sample_width
);
237 barrier(); /* finish the above in the view for NMIs */
238 trace_hwlat_callback_enabled
= false;
239 barrier(); /* Make sure nmi_total_ts is no longer updated */
243 /* If we exceed the threshold value, we have found a hardware latency */
244 if (sample
> thresh
|| outer_sample
> thresh
) {
249 /* We read in microseconds */
251 do_div(nmi_total_ts
, NSEC_PER_USEC
);
254 s
.seqnum
= hwlat_data
.count
;
256 s
.outer_duration
= outer_sample
;
257 s
.nmi_total_ts
= nmi_total_ts
;
258 s
.nmi_count
= nmi_count
;
260 trace_hwlat_sample(&s
);
262 latency
= max(sample
, outer_sample
);
264 /* Keep a running maximum ever recorded hardware latency */
265 if (latency
> tr
->max_latency
) {
266 tr
->max_latency
= latency
;
267 latency_fsnotify(tr
);
275 static struct cpumask save_cpumask
;
276 static bool disable_migrate
;
278 static void move_to_next_cpu(void)
280 struct cpumask
*current_mask
= &save_cpumask
;
281 struct trace_array
*tr
= hwlat_trace
;
287 * If for some reason the user modifies the CPU affinity
288 * of this thread, then stop migrating for the duration
289 * of the current test.
291 if (!cpumask_equal(current_mask
, current
->cpus_ptr
))
295 cpumask_and(current_mask
, cpu_online_mask
, tr
->tracing_cpumask
);
296 next_cpu
= cpumask_next(smp_processor_id(), current_mask
);
299 if (next_cpu
>= nr_cpu_ids
)
300 next_cpu
= cpumask_first(current_mask
);
302 if (next_cpu
>= nr_cpu_ids
) /* Shouldn't happen! */
305 cpumask_clear(current_mask
);
306 cpumask_set_cpu(next_cpu
, current_mask
);
308 sched_setaffinity(0, current_mask
);
312 disable_migrate
= true;
316 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
318 * Used to periodically sample the CPU TSC via a call to get_sample. We
319 * disable interrupts, which does (intentionally) introduce latency since we
320 * need to ensure nothing else might be running (and thus preempting).
321 * Obviously this should never be used in production environments.
323 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
325 static int kthread_fn(void *data
)
329 while (!kthread_should_stop()) {
337 mutex_lock(&hwlat_data
.lock
);
338 interval
= hwlat_data
.sample_window
- hwlat_data
.sample_width
;
339 mutex_unlock(&hwlat_data
.lock
);
341 do_div(interval
, USEC_PER_MSEC
); /* modifies interval value */
343 /* Always sleep for at least 1ms */
347 if (msleep_interruptible(interval
))
355 * start_kthread - Kick off the hardware latency sampling/detector kthread
357 * This starts the kernel thread that will sit and sample the CPU timestamp
358 * counter (TSC or similar) and look for potential hardware latencies.
360 static int start_kthread(struct trace_array
*tr
)
362 struct cpumask
*current_mask
= &save_cpumask
;
363 struct task_struct
*kthread
;
369 /* Just pick the first CPU on first iteration */
371 cpumask_and(current_mask
, cpu_online_mask
, tr
->tracing_cpumask
);
373 next_cpu
= cpumask_first(current_mask
);
375 kthread
= kthread_create(kthread_fn
, NULL
, "hwlatd");
376 if (IS_ERR(kthread
)) {
377 pr_err(BANNER
"could not start sampling thread\n");
381 cpumask_clear(current_mask
);
382 cpumask_set_cpu(next_cpu
, current_mask
);
383 sched_setaffinity(kthread
->pid
, current_mask
);
385 hwlat_kthread
= kthread
;
386 wake_up_process(kthread
);
392 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
394 * This kicks the running hardware latency sampling/detector kernel thread and
395 * tells it to stop sampling now. Use this on unload and at system shutdown.
397 static void stop_kthread(void)
401 kthread_stop(hwlat_kthread
);
402 hwlat_kthread
= NULL
;
406 * hwlat_read - Wrapper read function for reading both window and width
407 * @filp: The active open file structure
408 * @ubuf: The userspace provided buffer to read value into
409 * @cnt: The maximum number of bytes to read
410 * @ppos: The current "file" position
412 * This function provides a generic read implementation for the global state
413 * "hwlat_data" structure filesystem entries.
415 static ssize_t
hwlat_read(struct file
*filp
, char __user
*ubuf
,
416 size_t cnt
, loff_t
*ppos
)
418 char buf
[U64STR_SIZE
];
419 u64
*entry
= filp
->private_data
;
426 if (cnt
> sizeof(buf
))
431 len
= snprintf(buf
, sizeof(buf
), "%llu\n", val
);
433 return simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, len
);
437 * hwlat_width_write - Write function for "width" entry
438 * @filp: The active open file structure
439 * @ubuf: The user buffer that contains the value to write
440 * @cnt: The maximum number of bytes to write to "file"
441 * @ppos: The current position in @file
443 * This function provides a write implementation for the "width" interface
444 * to the hardware latency detector. It can be used to configure
445 * for how many us of the total window us we will actively sample for any
446 * hardware-induced latency periods. Obviously, it is not possible to
447 * sample constantly and have the system respond to a sample reader, or,
448 * worse, without having the system appear to have gone out to lunch. It
449 * is enforced that width is less that the total window size.
452 hwlat_width_write(struct file
*filp
, const char __user
*ubuf
,
453 size_t cnt
, loff_t
*ppos
)
458 err
= kstrtoull_from_user(ubuf
, cnt
, 10, &val
);
462 mutex_lock(&hwlat_data
.lock
);
463 if (val
< hwlat_data
.sample_window
)
464 hwlat_data
.sample_width
= val
;
467 mutex_unlock(&hwlat_data
.lock
);
476 * hwlat_window_write - Write function for "window" entry
477 * @filp: The active open file structure
478 * @ubuf: The user buffer that contains the value to write
479 * @cnt: The maximum number of bytes to write to "file"
480 * @ppos: The current position in @file
482 * This function provides a write implementation for the "window" interface
483 * to the hardware latency detector. The window is the total time
484 * in us that will be considered one sample period. Conceptually, windows
485 * occur back-to-back and contain a sample width period during which
486 * actual sampling occurs. Can be used to write a new total window size. It
487 * is enforced that any value written must be greater than the sample width
488 * size, or an error results.
491 hwlat_window_write(struct file
*filp
, const char __user
*ubuf
,
492 size_t cnt
, loff_t
*ppos
)
497 err
= kstrtoull_from_user(ubuf
, cnt
, 10, &val
);
501 mutex_lock(&hwlat_data
.lock
);
502 if (hwlat_data
.sample_width
< val
)
503 hwlat_data
.sample_window
= val
;
506 mutex_unlock(&hwlat_data
.lock
);
514 static const struct file_operations width_fops
= {
515 .open
= tracing_open_generic
,
517 .write
= hwlat_width_write
,
520 static const struct file_operations window_fops
= {
521 .open
= tracing_open_generic
,
523 .write
= hwlat_window_write
,
527 * init_tracefs - A function to initialize the tracefs interface files
529 * This function creates entries in tracefs for "hwlat_detector".
530 * It creates the hwlat_detector directory in the tracing directory,
531 * and within that directory is the count, width and window files to
532 * change and view those values.
534 static int init_tracefs(void)
537 struct dentry
*top_dir
;
539 ret
= tracing_init_dentry();
543 top_dir
= tracefs_create_dir("hwlat_detector", NULL
);
547 hwlat_sample_window
= tracefs_create_file("window", 0640,
549 &hwlat_data
.sample_window
,
551 if (!hwlat_sample_window
)
554 hwlat_sample_width
= tracefs_create_file("width", 0644,
556 &hwlat_data
.sample_width
,
558 if (!hwlat_sample_width
)
564 tracefs_remove(top_dir
);
568 static void hwlat_tracer_start(struct trace_array
*tr
)
572 err
= start_kthread(tr
);
574 pr_err(BANNER
"Cannot start hwlat kthread\n");
577 static void hwlat_tracer_stop(struct trace_array
*tr
)
582 static bool hwlat_busy
;
584 static int hwlat_tracer_init(struct trace_array
*tr
)
586 /* Only allow one instance to enable this */
592 disable_migrate
= false;
593 hwlat_data
.count
= 0;
595 save_tracing_thresh
= tracing_thresh
;
597 /* tracing_thresh is in nsecs, we speak in usecs */
599 tracing_thresh
= last_tracing_thresh
;
601 if (tracer_tracing_is_on(tr
))
602 hwlat_tracer_start(tr
);
609 static void hwlat_tracer_reset(struct trace_array
*tr
)
613 /* the tracing threshold is static between runs */
614 last_tracing_thresh
= tracing_thresh
;
616 tracing_thresh
= save_tracing_thresh
;
620 static struct tracer hwlat_tracer __read_mostly
=
623 .init
= hwlat_tracer_init
,
624 .reset
= hwlat_tracer_reset
,
625 .start
= hwlat_tracer_start
,
626 .stop
= hwlat_tracer_stop
,
627 .allow_instances
= true,
630 __init
static int init_hwlat_tracer(void)
634 mutex_init(&hwlat_data
.lock
);
636 ret
= register_tracer(&hwlat_tracer
);
644 late_initcall(init_hwlat_tracer
);