2 * trace_hwlatdetect.c - A simple Hardware Latency detector.
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.
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.
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.
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
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>
36 * Includes useful feedback from Clark Williams <clark@redhat.com>
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.
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>
50 static struct trace_array
*hwlat_trace
;
52 #define U64STR_SIZE 22 /* 20 digits max */
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 */
60 static struct task_struct
*hwlat_kthread
;
62 static struct dentry
*hwlat_sample_width
; /* sample width us */
63 static struct dentry
*hwlat_sample_window
; /* sample window us */
65 /* Save the previous tracing_thresh value */
66 static unsigned long save_tracing_thresh
;
68 /* NMI timestamp counters */
69 static u64 nmi_ts_start
;
70 static u64 nmi_total_ts
;
74 /* Tells NMIs to call back to the hwlat tracer to record timestamps */
75 bool trace_hwlat_callback_enabled
;
77 /* If the user changed threshold, remember it */
78 static u64 last_tracing_thresh
= DEFAULT_LAT_THRESHOLD
* NSEC_PER_USEC
;
80 /* Individual latency samples are stored here when detected. */
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 */
90 /* keep the global state somewhere. */
91 static struct hwlat_data
{
93 struct mutex lock
; /* protect changes */
95 u64 count
; /* total since reset */
97 u64 sample_window
; /* total sampling window (on+off) */
98 u64 sample_width
; /* active sampling portion of window */
101 .sample_window
= DEFAULT_SAMPLE_WINDOW
,
102 .sample_width
= DEFAULT_SAMPLE_WIDTH
,
105 static void trace_hwlat_sample(struct hwlat_sample
*sample
)
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
;
115 pc
= preempt_count();
116 local_save_flags(flags
);
118 event
= trace_buffer_lock_reserve(buffer
, TRACE_HWLAT
, sizeof(*entry
),
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
;
130 if (!call_filter_check_discard(call
, entry
, buffer
, event
))
131 trace_buffer_unlock_commit_nostack(buffer
, event
);
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
142 void trace_hwlat_callback(bool enter
)
144 if (smp_processor_id() != nmi_cpu
)
148 * Currently trace_clock_local() calls sched_clock() and the
149 * generic version is not NMI safe.
151 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK
)) {
153 nmi_ts_start
= time_get();
155 nmi_total_ts
= time_get() - nmi_ts_start
;
163 * get_sample - sample the CPU TSC and look for likely hardware latencies
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.
169 static int get_sample(void)
171 struct trace_array
*tr
= hwlat_trace
;
172 time_type start
, t1
, t2
, last_t2
;
173 s64 diff
, total
, last_total
= 0;
175 u64 thresh
= tracing_thresh
;
176 u64 outer_sample
= 0;
179 do_div(thresh
, NSEC_PER_USEC
); /* modifies interval value */
181 nmi_cpu
= smp_processor_id();
184 /* Make sure NMIs see this first */
187 trace_hwlat_callback_enabled
= true;
189 init_time(last_t2
, 0);
190 start
= time_get(); /* start timestamp */
194 t1
= time_get(); /* we'll look for a discontinuity */
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 */
202 pr_err(BANNER
"time running backwards\n");
205 if (diff
> outer_sample
)
210 total
= time_to_us(time_sub(t2
, start
)); /* sample width */
212 /* Check for possible overflows */
213 if (total
< last_total
) {
214 pr_err("Time total overflowed\n");
219 /* This checks the inner loop (t1 to t2) */
220 diff
= time_to_us(time_sub(t2
, t1
)); /* current diff */
222 /* This shouldn't happen */
224 pr_err(BANNER
"time running backwards\n");
229 sample
= diff
; /* only want highest value */
231 } while (total
<= hwlat_data
.sample_width
);
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 */
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
;
245 /* We read in microseconds */
247 do_div(nmi_total_ts
, NSEC_PER_USEC
);
250 s
.seqnum
= hwlat_data
.count
;
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
);
258 /* Keep a running maximum ever recorded hardware latency */
259 if (sample
> tr
->max_latency
)
260 tr
->max_latency
= sample
;
267 static struct cpumask save_cpumask
;
268 static bool disable_migrate
;
270 static void move_to_next_cpu(void)
272 struct cpumask
*current_mask
= &save_cpumask
;
278 * If for some reason the user modifies the CPU affinity
279 * of this thread, than stop migrating for the duration
280 * of the current test.
282 if (!cpumask_equal(current_mask
, ¤t
->cpus_allowed
))
286 cpumask_and(current_mask
, cpu_online_mask
, tracing_buffer_mask
);
287 next_cpu
= cpumask_next(smp_processor_id(), current_mask
);
290 if (next_cpu
>= nr_cpu_ids
)
291 next_cpu
= cpumask_first(current_mask
);
293 if (next_cpu
>= nr_cpu_ids
) /* Shouldn't happen! */
296 cpumask_clear(current_mask
);
297 cpumask_set_cpu(next_cpu
, current_mask
);
299 sched_setaffinity(0, current_mask
);
303 disable_migrate
= true;
307 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
309 * Used to periodically sample the CPU TSC via a call to get_sample. We
310 * disable interrupts, which does (intentionally) introduce latency since we
311 * need to ensure nothing else might be running (and thus preempting).
312 * Obviously this should never be used in production environments.
314 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
316 static int kthread_fn(void *data
)
320 while (!kthread_should_stop()) {
328 mutex_lock(&hwlat_data
.lock
);
329 interval
= hwlat_data
.sample_window
- hwlat_data
.sample_width
;
330 mutex_unlock(&hwlat_data
.lock
);
332 do_div(interval
, USEC_PER_MSEC
); /* modifies interval value */
334 /* Always sleep for at least 1ms */
338 if (msleep_interruptible(interval
))
346 * start_kthread - Kick off the hardware latency sampling/detector kthread
348 * This starts the kernel thread that will sit and sample the CPU timestamp
349 * counter (TSC or similar) and look for potential hardware latencies.
351 static int start_kthread(struct trace_array
*tr
)
353 struct cpumask
*current_mask
= &save_cpumask
;
354 struct task_struct
*kthread
;
357 /* Just pick the first CPU on first iteration */
358 current_mask
= &save_cpumask
;
360 cpumask_and(current_mask
, cpu_online_mask
, tracing_buffer_mask
);
362 next_cpu
= cpumask_first(current_mask
);
364 kthread
= kthread_create(kthread_fn
, NULL
, "hwlatd");
365 if (IS_ERR(kthread
)) {
366 pr_err(BANNER
"could not start sampling thread\n");
370 cpumask_clear(current_mask
);
371 cpumask_set_cpu(next_cpu
, current_mask
);
372 sched_setaffinity(kthread
->pid
, current_mask
);
374 hwlat_kthread
= kthread
;
375 wake_up_process(kthread
);
381 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
383 * This kicks the running hardware latency sampling/detector kernel thread and
384 * tells it to stop sampling now. Use this on unload and at system shutdown.
386 static void stop_kthread(void)
390 kthread_stop(hwlat_kthread
);
391 hwlat_kthread
= NULL
;
395 * hwlat_read - Wrapper read function for reading both window and width
396 * @filp: The active open file structure
397 * @ubuf: The userspace provided buffer to read value into
398 * @cnt: The maximum number of bytes to read
399 * @ppos: The current "file" position
401 * This function provides a generic read implementation for the global state
402 * "hwlat_data" structure filesystem entries.
404 static ssize_t
hwlat_read(struct file
*filp
, char __user
*ubuf
,
405 size_t cnt
, loff_t
*ppos
)
407 char buf
[U64STR_SIZE
];
408 u64
*entry
= filp
->private_data
;
415 if (cnt
> sizeof(buf
))
420 len
= snprintf(buf
, sizeof(buf
), "%llu\n", val
);
422 return simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, len
);
426 * hwlat_width_write - Write function for "width" entry
427 * @filp: The active open file structure
428 * @ubuf: The user buffer that contains the value to write
429 * @cnt: The maximum number of bytes to write to "file"
430 * @ppos: The current position in @file
432 * This function provides a write implementation for the "width" interface
433 * to the hardware latency detector. It can be used to configure
434 * for how many us of the total window us we will actively sample for any
435 * hardware-induced latency periods. Obviously, it is not possible to
436 * sample constantly and have the system respond to a sample reader, or,
437 * worse, without having the system appear to have gone out to lunch. It
438 * is enforced that width is less that the total window size.
441 hwlat_width_write(struct file
*filp
, const char __user
*ubuf
,
442 size_t cnt
, loff_t
*ppos
)
447 err
= kstrtoull_from_user(ubuf
, cnt
, 10, &val
);
451 mutex_lock(&hwlat_data
.lock
);
452 if (val
< hwlat_data
.sample_window
)
453 hwlat_data
.sample_width
= val
;
456 mutex_unlock(&hwlat_data
.lock
);
465 * hwlat_window_write - Write function for "window" entry
466 * @filp: The active open file structure
467 * @ubuf: The user buffer that contains the value to write
468 * @cnt: The maximum number of bytes to write to "file"
469 * @ppos: The current position in @file
471 * This function provides a write implementation for the "window" interface
472 * to the hardware latency detetector. The window is the total time
473 * in us that will be considered one sample period. Conceptually, windows
474 * occur back-to-back and contain a sample width period during which
475 * actual sampling occurs. Can be used to write a new total window size. It
476 * is enfoced that any value written must be greater than the sample width
477 * size, or an error results.
480 hwlat_window_write(struct file
*filp
, const char __user
*ubuf
,
481 size_t cnt
, loff_t
*ppos
)
486 err
= kstrtoull_from_user(ubuf
, cnt
, 10, &val
);
490 mutex_lock(&hwlat_data
.lock
);
491 if (hwlat_data
.sample_width
< val
)
492 hwlat_data
.sample_window
= val
;
495 mutex_unlock(&hwlat_data
.lock
);
503 static const struct file_operations width_fops
= {
504 .open
= tracing_open_generic
,
506 .write
= hwlat_width_write
,
509 static const struct file_operations window_fops
= {
510 .open
= tracing_open_generic
,
512 .write
= hwlat_window_write
,
516 * init_tracefs - A function to initialize the tracefs interface files
518 * This function creates entries in tracefs for "hwlat_detector".
519 * It creates the hwlat_detector directory in the tracing directory,
520 * and within that directory is the count, width and window files to
521 * change and view those values.
523 static int init_tracefs(void)
525 struct dentry
*d_tracer
;
526 struct dentry
*top_dir
;
528 d_tracer
= tracing_init_dentry();
529 if (IS_ERR(d_tracer
))
532 top_dir
= tracefs_create_dir("hwlat_detector", d_tracer
);
536 hwlat_sample_window
= tracefs_create_file("window", 0640,
538 &hwlat_data
.sample_window
,
540 if (!hwlat_sample_window
)
543 hwlat_sample_width
= tracefs_create_file("width", 0644,
545 &hwlat_data
.sample_width
,
547 if (!hwlat_sample_width
)
553 tracefs_remove_recursive(top_dir
);
557 static void hwlat_tracer_start(struct trace_array
*tr
)
561 err
= start_kthread(tr
);
563 pr_err(BANNER
"Cannot start hwlat kthread\n");
566 static void hwlat_tracer_stop(struct trace_array
*tr
)
571 static bool hwlat_busy
;
573 static int hwlat_tracer_init(struct trace_array
*tr
)
575 /* Only allow one instance to enable this */
581 disable_migrate
= false;
582 hwlat_data
.count
= 0;
584 save_tracing_thresh
= tracing_thresh
;
586 /* tracing_thresh is in nsecs, we speak in usecs */
588 tracing_thresh
= last_tracing_thresh
;
590 if (tracer_tracing_is_on(tr
))
591 hwlat_tracer_start(tr
);
598 static void hwlat_tracer_reset(struct trace_array
*tr
)
602 /* the tracing threshold is static between runs */
603 last_tracing_thresh
= tracing_thresh
;
605 tracing_thresh
= save_tracing_thresh
;
609 static struct tracer hwlat_tracer __read_mostly
=
612 .init
= hwlat_tracer_init
,
613 .reset
= hwlat_tracer_reset
,
614 .start
= hwlat_tracer_start
,
615 .stop
= hwlat_tracer_stop
,
616 .allow_instances
= true,
619 __init
static int init_hwlat_tracer(void)
623 mutex_init(&hwlat_data
.lock
);
625 ret
= register_tracer(&hwlat_tracer
);
633 late_initcall(init_hwlat_tracer
);