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 */
88 /* keep the global state somewhere. */
89 static struct hwlat_data
{
91 struct mutex lock
; /* protect changes */
93 u64 count
; /* total since reset */
95 u64 sample_window
; /* total sampling window (on+off) */
96 u64 sample_width
; /* active sampling portion of window */
99 .sample_window
= DEFAULT_SAMPLE_WINDOW
,
100 .sample_width
= DEFAULT_SAMPLE_WIDTH
,
103 static void trace_hwlat_sample(struct hwlat_sample
*sample
)
105 struct trace_array
*tr
= hwlat_trace
;
106 struct trace_event_call
*call
= &event_hwlat
;
107 struct trace_buffer
*buffer
= tr
->array_buffer
.buffer
;
108 struct ring_buffer_event
*event
;
109 struct hwlat_entry
*entry
;
113 pc
= preempt_count();
114 local_save_flags(flags
);
116 event
= trace_buffer_lock_reserve(buffer
, TRACE_HWLAT
, sizeof(*entry
),
120 entry
= ring_buffer_event_data(event
);
121 entry
->seqnum
= sample
->seqnum
;
122 entry
->duration
= sample
->duration
;
123 entry
->outer_duration
= sample
->outer_duration
;
124 entry
->timestamp
= sample
->timestamp
;
125 entry
->nmi_total_ts
= sample
->nmi_total_ts
;
126 entry
->nmi_count
= sample
->nmi_count
;
128 if (!call_filter_check_discard(call
, entry
, buffer
, event
))
129 trace_buffer_unlock_commit_nostack(buffer
, event
);
132 /* Macros to encapsulate the time capturing infrastructure */
133 #define time_type u64
134 #define time_get() trace_clock_local()
135 #define time_to_us(x) div_u64(x, 1000)
136 #define time_sub(a, b) ((a) - (b))
137 #define init_time(a, b) (a = b)
138 #define time_u64(a) a
140 void trace_hwlat_callback(bool enter
)
142 if (smp_processor_id() != nmi_cpu
)
146 * Currently trace_clock_local() calls sched_clock() and the
147 * generic version is not NMI safe.
149 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK
)) {
151 nmi_ts_start
= time_get();
153 nmi_total_ts
+= time_get() - nmi_ts_start
;
161 * get_sample - sample the CPU TSC and look for likely hardware latencies
163 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
164 * hardware-induced latency. Called with interrupts disabled and with
165 * hwlat_data.lock held.
167 static int get_sample(void)
169 struct trace_array
*tr
= hwlat_trace
;
170 time_type start
, t1
, t2
, last_t2
;
171 s64 diff
, total
, last_total
= 0;
173 u64 thresh
= tracing_thresh
;
174 u64 outer_sample
= 0;
177 do_div(thresh
, NSEC_PER_USEC
); /* modifies interval value */
179 nmi_cpu
= smp_processor_id();
182 /* Make sure NMIs see this first */
185 trace_hwlat_callback_enabled
= true;
187 init_time(last_t2
, 0);
188 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 diff
= time_to_us(time_sub(t1
, last_t2
));
198 /* This shouldn't happen */
200 pr_err(BANNER
"time running backwards\n");
203 if (diff
> outer_sample
)
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 /* This shouldn't happen */
222 pr_err(BANNER
"time running backwards\n");
227 sample
= diff
; /* only want highest value */
229 } while (total
<= hwlat_data
.sample_width
);
231 barrier(); /* finish the above in the view for NMIs */
232 trace_hwlat_callback_enabled
= false;
233 barrier(); /* Make sure nmi_total_ts is no longer updated */
237 /* If we exceed the threshold value, we have found a hardware latency */
238 if (sample
> thresh
|| outer_sample
> thresh
) {
239 struct hwlat_sample s
;
244 /* We read in microseconds */
246 do_div(nmi_total_ts
, NSEC_PER_USEC
);
249 s
.seqnum
= hwlat_data
.count
;
251 s
.outer_duration
= outer_sample
;
252 ktime_get_real_ts64(&s
.timestamp
);
253 s
.nmi_total_ts
= nmi_total_ts
;
254 s
.nmi_count
= nmi_count
;
255 trace_hwlat_sample(&s
);
257 latency
= max(sample
, outer_sample
);
259 /* Keep a running maximum ever recorded hardware latency */
260 if (latency
> tr
->max_latency
) {
261 tr
->max_latency
= latency
;
262 latency_fsnotify(tr
);
270 static struct cpumask save_cpumask
;
271 static bool disable_migrate
;
273 static void move_to_next_cpu(void)
275 struct cpumask
*current_mask
= &save_cpumask
;
281 * If for some reason the user modifies the CPU affinity
282 * of this thread, then stop migrating for the duration
283 * of the current test.
285 if (!cpumask_equal(current_mask
, current
->cpus_ptr
))
289 cpumask_and(current_mask
, cpu_online_mask
, tracing_buffer_mask
);
290 next_cpu
= cpumask_next(smp_processor_id(), current_mask
);
293 if (next_cpu
>= nr_cpu_ids
)
294 next_cpu
= cpumask_first(current_mask
);
296 if (next_cpu
>= nr_cpu_ids
) /* Shouldn't happen! */
299 cpumask_clear(current_mask
);
300 cpumask_set_cpu(next_cpu
, current_mask
);
302 sched_setaffinity(0, current_mask
);
306 disable_migrate
= true;
310 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
312 * Used to periodically sample the CPU TSC via a call to get_sample. We
313 * disable interrupts, which does (intentionally) introduce latency since we
314 * need to ensure nothing else might be running (and thus preempting).
315 * Obviously this should never be used in production environments.
317 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
319 static int kthread_fn(void *data
)
323 while (!kthread_should_stop()) {
331 mutex_lock(&hwlat_data
.lock
);
332 interval
= hwlat_data
.sample_window
- hwlat_data
.sample_width
;
333 mutex_unlock(&hwlat_data
.lock
);
335 do_div(interval
, USEC_PER_MSEC
); /* modifies interval value */
337 /* Always sleep for at least 1ms */
341 if (msleep_interruptible(interval
))
349 * start_kthread - Kick off the hardware latency sampling/detector kthread
351 * This starts the kernel thread that will sit and sample the CPU timestamp
352 * counter (TSC or similar) and look for potential hardware latencies.
354 static int start_kthread(struct trace_array
*tr
)
356 struct cpumask
*current_mask
= &save_cpumask
;
357 struct task_struct
*kthread
;
360 if (WARN_ON(hwlat_kthread
))
363 /* Just pick the first CPU on first iteration */
364 current_mask
= &save_cpumask
;
366 cpumask_and(current_mask
, cpu_online_mask
, tracing_buffer_mask
);
368 next_cpu
= cpumask_first(current_mask
);
370 kthread
= kthread_create(kthread_fn
, NULL
, "hwlatd");
371 if (IS_ERR(kthread
)) {
372 pr_err(BANNER
"could not start sampling thread\n");
376 cpumask_clear(current_mask
);
377 cpumask_set_cpu(next_cpu
, current_mask
);
378 sched_setaffinity(kthread
->pid
, current_mask
);
380 hwlat_kthread
= kthread
;
381 wake_up_process(kthread
);
387 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
389 * This kicks the running hardware latency sampling/detector kernel thread and
390 * tells it to stop sampling now. Use this on unload and at system shutdown.
392 static void stop_kthread(void)
396 kthread_stop(hwlat_kthread
);
397 hwlat_kthread
= NULL
;
401 * hwlat_read - Wrapper read function for reading both window and width
402 * @filp: The active open file structure
403 * @ubuf: The userspace provided buffer to read value into
404 * @cnt: The maximum number of bytes to read
405 * @ppos: The current "file" position
407 * This function provides a generic read implementation for the global state
408 * "hwlat_data" structure filesystem entries.
410 static ssize_t
hwlat_read(struct file
*filp
, char __user
*ubuf
,
411 size_t cnt
, loff_t
*ppos
)
413 char buf
[U64STR_SIZE
];
414 u64
*entry
= filp
->private_data
;
421 if (cnt
> sizeof(buf
))
426 len
= snprintf(buf
, sizeof(buf
), "%llu\n", val
);
428 return simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, len
);
432 * hwlat_width_write - Write function for "width" entry
433 * @filp: The active open file structure
434 * @ubuf: The user buffer that contains the value to write
435 * @cnt: The maximum number of bytes to write to "file"
436 * @ppos: The current position in @file
438 * This function provides a write implementation for the "width" interface
439 * to the hardware latency detector. It can be used to configure
440 * for how many us of the total window us we will actively sample for any
441 * hardware-induced latency periods. Obviously, it is not possible to
442 * sample constantly and have the system respond to a sample reader, or,
443 * worse, without having the system appear to have gone out to lunch. It
444 * is enforced that width is less that the total window size.
447 hwlat_width_write(struct file
*filp
, const char __user
*ubuf
,
448 size_t cnt
, loff_t
*ppos
)
453 err
= kstrtoull_from_user(ubuf
, cnt
, 10, &val
);
457 mutex_lock(&hwlat_data
.lock
);
458 if (val
< hwlat_data
.sample_window
)
459 hwlat_data
.sample_width
= val
;
462 mutex_unlock(&hwlat_data
.lock
);
471 * hwlat_window_write - Write function for "window" entry
472 * @filp: The active open file structure
473 * @ubuf: The user buffer that contains the value to write
474 * @cnt: The maximum number of bytes to write to "file"
475 * @ppos: The current position in @file
477 * This function provides a write implementation for the "window" interface
478 * to the hardware latency detetector. The window is the total time
479 * in us that will be considered one sample period. Conceptually, windows
480 * occur back-to-back and contain a sample width period during which
481 * actual sampling occurs. Can be used to write a new total window size. It
482 * is enfoced that any value written must be greater than the sample width
483 * size, or an error results.
486 hwlat_window_write(struct file
*filp
, const char __user
*ubuf
,
487 size_t cnt
, loff_t
*ppos
)
492 err
= kstrtoull_from_user(ubuf
, cnt
, 10, &val
);
496 mutex_lock(&hwlat_data
.lock
);
497 if (hwlat_data
.sample_width
< val
)
498 hwlat_data
.sample_window
= val
;
501 mutex_unlock(&hwlat_data
.lock
);
509 static const struct file_operations width_fops
= {
510 .open
= tracing_open_generic
,
512 .write
= hwlat_width_write
,
515 static const struct file_operations window_fops
= {
516 .open
= tracing_open_generic
,
518 .write
= hwlat_window_write
,
522 * init_tracefs - A function to initialize the tracefs interface files
524 * This function creates entries in tracefs for "hwlat_detector".
525 * It creates the hwlat_detector directory in the tracing directory,
526 * and within that directory is the count, width and window files to
527 * change and view those values.
529 static int init_tracefs(void)
531 struct dentry
*d_tracer
;
532 struct dentry
*top_dir
;
534 d_tracer
= tracing_init_dentry();
535 if (IS_ERR(d_tracer
))
538 top_dir
= tracefs_create_dir("hwlat_detector", d_tracer
);
542 hwlat_sample_window
= tracefs_create_file("window", 0640,
544 &hwlat_data
.sample_window
,
546 if (!hwlat_sample_window
)
549 hwlat_sample_width
= tracefs_create_file("width", 0644,
551 &hwlat_data
.sample_width
,
553 if (!hwlat_sample_width
)
559 tracefs_remove(top_dir
);
563 static void hwlat_tracer_start(struct trace_array
*tr
)
567 err
= start_kthread(tr
);
569 pr_err(BANNER
"Cannot start hwlat kthread\n");
572 static void hwlat_tracer_stop(struct trace_array
*tr
)
577 static bool hwlat_busy
;
579 static int hwlat_tracer_init(struct trace_array
*tr
)
581 /* Only allow one instance to enable this */
587 disable_migrate
= false;
588 hwlat_data
.count
= 0;
590 save_tracing_thresh
= tracing_thresh
;
592 /* tracing_thresh is in nsecs, we speak in usecs */
594 tracing_thresh
= last_tracing_thresh
;
596 if (tracer_tracing_is_on(tr
))
597 hwlat_tracer_start(tr
);
604 static void hwlat_tracer_reset(struct trace_array
*tr
)
608 /* the tracing threshold is static between runs */
609 last_tracing_thresh
= tracing_thresh
;
611 tracing_thresh
= save_tracing_thresh
;
615 static struct tracer hwlat_tracer __read_mostly
=
618 .init
= hwlat_tracer_init
,
619 .reset
= hwlat_tracer_reset
,
620 .start
= hwlat_tracer_start
,
621 .stop
= hwlat_tracer_stop
,
622 .allow_instances
= true,
625 __init
static int init_hwlat_tracer(void)
629 mutex_init(&hwlat_data
.lock
);
631 ret
= register_tracer(&hwlat_tracer
);
639 late_initcall(init_hwlat_tracer
);