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>
49 static struct trace_array *hwlat_trace;
51 #define U64STR_SIZE 22 /* 20 digits max */
53 #define BANNER "hwlat_detector: "
54 #define DEFAULT_SAMPLE_WINDOW 1000000 /* 1s */
55 #define DEFAULT_SAMPLE_WIDTH 500000 /* 0.5s */
56 #define DEFAULT_LAT_THRESHOLD 10 /* 10us */
59 static struct task_struct *hwlat_kthread;
61 static struct dentry *hwlat_sample_width; /* sample width us */
62 static struct dentry *hwlat_sample_window; /* sample window us */
64 /* Save the previous tracing_thresh value */
65 static unsigned long save_tracing_thresh;
67 /* If the user changed threshold, remember it */
68 static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
70 /* Individual latency samples are stored here when detected. */
72 u64 seqnum; /* unique sequence */
73 u64 duration; /* delta */
74 u64 outer_duration; /* delta (outer loop) */
75 struct timespec timestamp; /* wall time */
78 /* keep the global state somewhere. */
79 static struct hwlat_data {
81 struct mutex lock; /* protect changes */
83 u64 count; /* total since reset */
85 u64 sample_window; /* total sampling window (on+off) */
86 u64 sample_width; /* active sampling portion of window */
89 .sample_window = DEFAULT_SAMPLE_WINDOW,
90 .sample_width = DEFAULT_SAMPLE_WIDTH,
93 static void trace_hwlat_sample(struct hwlat_sample *sample)
95 struct trace_array *tr = hwlat_trace;
96 struct trace_event_call *call = &event_hwlat;
97 struct ring_buffer *buffer = tr->trace_buffer.buffer;
98 struct ring_buffer_event *event;
99 struct hwlat_entry *entry;
103 pc = preempt_count();
104 local_save_flags(flags);
106 event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
110 entry = ring_buffer_event_data(event);
111 entry->seqnum = sample->seqnum;
112 entry->duration = sample->duration;
113 entry->outer_duration = sample->outer_duration;
114 entry->timestamp = sample->timestamp;
116 if (!call_filter_check_discard(call, entry, buffer, event))
117 __buffer_unlock_commit(buffer, event);
120 /* Macros to encapsulate the time capturing infrastructure */
121 #define time_type u64
122 #define time_get() trace_clock_local()
123 #define time_to_us(x) div_u64(x, 1000)
124 #define time_sub(a, b) ((a) - (b))
125 #define init_time(a, b) (a = b)
126 #define time_u64(a) a
129 * get_sample - sample the CPU TSC and look for likely hardware latencies
131 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
132 * hardware-induced latency. Called with interrupts disabled and with
133 * hwlat_data.lock held.
135 static int get_sample(void)
137 struct trace_array *tr = hwlat_trace;
138 time_type start, t1, t2, last_t2;
139 s64 diff, total, last_total = 0;
141 u64 thresh = tracing_thresh;
142 u64 outer_sample = 0;
145 do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
147 init_time(last_t2, 0);
148 start = time_get(); /* start timestamp */
152 t1 = time_get(); /* we'll look for a discontinuity */
155 if (time_u64(last_t2)) {
156 /* Check the delta from outer loop (t2 to next t1) */
157 diff = time_to_us(time_sub(t1, last_t2));
158 /* This shouldn't happen */
160 pr_err(BANNER "time running backwards\n");
163 if (diff > outer_sample)
168 total = time_to_us(time_sub(t2, start)); /* sample width */
170 /* Check for possible overflows */
171 if (total < last_total) {
172 pr_err("Time total overflowed\n");
177 /* This checks the inner loop (t1 to t2) */
178 diff = time_to_us(time_sub(t2, t1)); /* current diff */
180 /* This shouldn't happen */
182 pr_err(BANNER "time running backwards\n");
187 sample = diff; /* only want highest value */
189 } while (total <= hwlat_data.sample_width);
193 /* If we exceed the threshold value, we have found a hardware latency */
194 if (sample > thresh || outer_sample > thresh) {
195 struct hwlat_sample s;
200 s.seqnum = hwlat_data.count;
202 s.outer_duration = outer_sample;
203 s.timestamp = CURRENT_TIME;
204 trace_hwlat_sample(&s);
206 /* Keep a running maximum ever recorded hardware latency */
207 if (sample > tr->max_latency)
208 tr->max_latency = sample;
215 static struct cpumask save_cpumask;
216 static bool disable_migrate;
218 static void move_to_next_cpu(void)
220 static struct cpumask *current_mask;
226 /* Just pick the first CPU on first iteration */
228 current_mask = &save_cpumask;
230 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
232 next_cpu = cpumask_first(current_mask);
237 * If for some reason the user modifies the CPU affinity
238 * of this thread, than stop migrating for the duration
239 * of the current test.
241 if (!cpumask_equal(current_mask, ¤t->cpus_allowed))
245 cpumask_and(current_mask, cpu_online_mask, tracing_buffer_mask);
246 next_cpu = cpumask_next(smp_processor_id(), current_mask);
249 if (next_cpu >= nr_cpu_ids)
250 next_cpu = cpumask_first(current_mask);
253 if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
256 cpumask_clear(current_mask);
257 cpumask_set_cpu(next_cpu, current_mask);
259 sched_setaffinity(0, current_mask);
263 disable_migrate = true;
267 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
269 * Used to periodically sample the CPU TSC via a call to get_sample. We
270 * disable interrupts, which does (intentionally) introduce latency since we
271 * need to ensure nothing else might be running (and thus preempting).
272 * Obviously this should never be used in production environments.
274 * Currently this runs on which ever CPU it was scheduled on, but most
275 * real-world hardware latency situations occur across several CPUs,
276 * but we might later generalize this if we find there are any actualy
277 * systems with alternate SMI delivery or other hardware latencies.
279 static int kthread_fn(void *data)
283 while (!kthread_should_stop()) {
291 mutex_lock(&hwlat_data.lock);
292 interval = hwlat_data.sample_window - hwlat_data.sample_width;
293 mutex_unlock(&hwlat_data.lock);
295 do_div(interval, USEC_PER_MSEC); /* modifies interval value */
297 /* Always sleep for at least 1ms */
301 if (msleep_interruptible(interval))
309 * start_kthread - Kick off the hardware latency sampling/detector kthread
311 * This starts the kernel thread that will sit and sample the CPU timestamp
312 * counter (TSC or similar) and look for potential hardware latencies.
314 static int start_kthread(struct trace_array *tr)
316 struct task_struct *kthread;
318 kthread = kthread_create(kthread_fn, NULL, "hwlatd");
319 if (IS_ERR(kthread)) {
320 pr_err(BANNER "could not start sampling thread\n");
323 hwlat_kthread = kthread;
324 wake_up_process(kthread);
330 * stop_kthread - Inform the hardware latency samping/detector kthread to stop
332 * This kicks the running hardware latency sampling/detector kernel thread and
333 * tells it to stop sampling now. Use this on unload and at system shutdown.
335 static void stop_kthread(void)
339 kthread_stop(hwlat_kthread);
340 hwlat_kthread = NULL;
344 * hwlat_read - Wrapper read function for reading both window and width
345 * @filp: The active open file structure
346 * @ubuf: The userspace provided buffer to read value into
347 * @cnt: The maximum number of bytes to read
348 * @ppos: The current "file" position
350 * This function provides a generic read implementation for the global state
351 * "hwlat_data" structure filesystem entries.
353 static ssize_t hwlat_read(struct file *filp, char __user *ubuf,
354 size_t cnt, loff_t *ppos)
356 char buf[U64STR_SIZE];
357 u64 *entry = filp->private_data;
364 if (cnt > sizeof(buf))
369 len = snprintf(buf, sizeof(buf), "%llu\n", val);
371 return simple_read_from_buffer(ubuf, cnt, ppos, buf, len);
375 * hwlat_width_write - Write function for "width" entry
376 * @filp: The active open file structure
377 * @ubuf: The user buffer that contains the value to write
378 * @cnt: The maximum number of bytes to write to "file"
379 * @ppos: The current position in @file
381 * This function provides a write implementation for the "width" interface
382 * to the hardware latency detector. It can be used to configure
383 * for how many us of the total window us we will actively sample for any
384 * hardware-induced latency periods. Obviously, it is not possible to
385 * sample constantly and have the system respond to a sample reader, or,
386 * worse, without having the system appear to have gone out to lunch. It
387 * is enforced that width is less that the total window size.
390 hwlat_width_write(struct file *filp, const char __user *ubuf,
391 size_t cnt, loff_t *ppos)
396 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
400 mutex_lock(&hwlat_data.lock);
401 if (val < hwlat_data.sample_window)
402 hwlat_data.sample_width = val;
405 mutex_unlock(&hwlat_data.lock);
414 * hwlat_window_write - Write function for "window" entry
415 * @filp: The active open file structure
416 * @ubuf: The user buffer that contains the value to write
417 * @cnt: The maximum number of bytes to write to "file"
418 * @ppos: The current position in @file
420 * This function provides a write implementation for the "window" interface
421 * to the hardware latency detetector. The window is the total time
422 * in us that will be considered one sample period. Conceptually, windows
423 * occur back-to-back and contain a sample width period during which
424 * actual sampling occurs. Can be used to write a new total window size. It
425 * is enfoced that any value written must be greater than the sample width
426 * size, or an error results.
429 hwlat_window_write(struct file *filp, const char __user *ubuf,
430 size_t cnt, loff_t *ppos)
435 err = kstrtoull_from_user(ubuf, cnt, 10, &val);
439 mutex_lock(&hwlat_data.lock);
440 if (hwlat_data.sample_width < val)
441 hwlat_data.sample_window = val;
444 mutex_unlock(&hwlat_data.lock);
452 static const struct file_operations width_fops = {
453 .open = tracing_open_generic,
455 .write = hwlat_width_write,
458 static const struct file_operations window_fops = {
459 .open = tracing_open_generic,
461 .write = hwlat_window_write,
465 * init_tracefs - A function to initialize the tracefs interface files
467 * This function creates entries in tracefs for "hwlat_detector".
468 * It creates the hwlat_detector directory in the tracing directory,
469 * and within that directory is the count, width and window files to
470 * change and view those values.
472 static int init_tracefs(void)
474 struct dentry *d_tracer;
475 struct dentry *top_dir;
477 d_tracer = tracing_init_dentry();
478 if (IS_ERR(d_tracer))
481 top_dir = tracefs_create_dir("hwlat_detector", d_tracer);
485 hwlat_sample_window = tracefs_create_file("window", 0640,
487 &hwlat_data.sample_window,
489 if (!hwlat_sample_window)
492 hwlat_sample_width = tracefs_create_file("width", 0644,
494 &hwlat_data.sample_width,
496 if (!hwlat_sample_width)
502 tracefs_remove_recursive(top_dir);
506 static void hwlat_tracer_start(struct trace_array *tr)
510 err = start_kthread(tr);
512 pr_err(BANNER "Cannot start hwlat kthread\n");
515 static void hwlat_tracer_stop(struct trace_array *tr)
520 static bool hwlat_busy;
522 static int hwlat_tracer_init(struct trace_array *tr)
524 /* Only allow one instance to enable this */
530 disable_migrate = false;
531 hwlat_data.count = 0;
533 save_tracing_thresh = tracing_thresh;
535 /* tracing_thresh is in nsecs, we speak in usecs */
537 tracing_thresh = last_tracing_thresh;
539 if (tracer_tracing_is_on(tr))
540 hwlat_tracer_start(tr);
547 static void hwlat_tracer_reset(struct trace_array *tr)
551 /* the tracing threshold is static between runs */
552 last_tracing_thresh = tracing_thresh;
554 tracing_thresh = save_tracing_thresh;
558 static struct tracer hwlat_tracer __read_mostly =
561 .init = hwlat_tracer_init,
562 .reset = hwlat_tracer_reset,
563 .start = hwlat_tracer_start,
564 .stop = hwlat_tracer_stop,
565 .allow_instances = true,
568 __init static int init_hwlat_tracer(void)
572 mutex_init(&hwlat_data.lock);
574 ret = register_tracer(&hwlat_tracer);
582 late_initcall(init_hwlat_tracer);