xref: /openbmc/linux/kernel/trace/trace_hwlat.c (revision fac59652993f075d57860769c99045b3ca18780d)
1  // SPDX-License-Identifier: GPL-2.0
2  /*
3   * trace_hwlat.c - A simple Hardware Latency detector.
4   *
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.
9   *
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.
19   *
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.
28   *
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
32   *          guarantee(s).
33   *
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>
36   *
37   * Includes useful feedback from Clark Williams <williams@redhat.com>
38   *
39   */
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>
46  #include "trace.h"
47  
48  static struct trace_array	*hwlat_trace;
49  
50  #define U64STR_SIZE		22			/* 20 digits max */
51  
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 */
56  
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 */
60  
61  enum {
62  	MODE_NONE = 0,
63  	MODE_ROUND_ROBIN,
64  	MODE_PER_CPU,
65  	MODE_MAX
66  };
67  static char *thread_mode_str[] = { "none", "round-robin", "per-cpu" };
68  
69  /* Save the previous tracing_thresh value */
70  static unsigned long save_tracing_thresh;
71  
72  /* runtime kthread data */
73  struct hwlat_kthread_data {
74  	struct task_struct	*kthread;
75  	/* NMI timestamp counters */
76  	u64			nmi_ts_start;
77  	u64			nmi_total_ts;
78  	int			nmi_count;
79  	int			nmi_cpu;
80  };
81  
82  static struct hwlat_kthread_data hwlat_single_cpu_data;
83  static DEFINE_PER_CPU(struct hwlat_kthread_data, hwlat_per_cpu_data);
84  
85  /* Tells NMIs to call back to the hwlat tracer to record timestamps */
86  bool trace_hwlat_callback_enabled;
87  
88  /* If the user changed threshold, remember it */
89  static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
90  
91  /* Individual latency samples are stored here when detected. */
92  struct hwlat_sample {
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 */
100  };
101  
102  /* keep the global state somewhere. */
103  static struct hwlat_data {
104  
105  	struct mutex lock;		/* protect changes */
106  
107  	u64	count;			/* total since reset */
108  
109  	u64	sample_window;		/* total sampling window (on+off) */
110  	u64	sample_width;		/* active sampling portion of window */
111  
112  	int	thread_mode;		/* thread mode */
113  
114  } hwlat_data = {
115  	.sample_window		= DEFAULT_SAMPLE_WINDOW,
116  	.sample_width		= DEFAULT_SAMPLE_WIDTH,
117  	.thread_mode		= MODE_ROUND_ROBIN
118  };
119  
get_cpu_data(void)120  static struct hwlat_kthread_data *get_cpu_data(void)
121  {
122  	if (hwlat_data.thread_mode == MODE_PER_CPU)
123  		return this_cpu_ptr(&hwlat_per_cpu_data);
124  	else
125  		return &hwlat_single_cpu_data;
126  }
127  
128  static bool hwlat_busy;
129  
trace_hwlat_sample(struct hwlat_sample * sample)130  static void trace_hwlat_sample(struct hwlat_sample *sample)
131  {
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;
137  
138  	event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
139  					  tracing_gen_ctx());
140  	if (!event)
141  		return;
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;
150  
151  	if (!call_filter_check_discard(call, entry, buffer, event))
152  		trace_buffer_unlock_commit_nostack(buffer, event);
153  }
154  
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
162  
trace_hwlat_callback(bool enter)163  void trace_hwlat_callback(bool enter)
164  {
165  	struct hwlat_kthread_data *kdata = get_cpu_data();
166  
167  	if (!kdata->kthread)
168  		return;
169  
170  	/*
171  	 * Currently trace_clock_local() calls sched_clock() and the
172  	 * generic version is not NMI safe.
173  	 */
174  	if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
175  		if (enter)
176  			kdata->nmi_ts_start = time_get();
177  		else
178  			kdata->nmi_total_ts += time_get() - kdata->nmi_ts_start;
179  	}
180  
181  	if (enter)
182  		kdata->nmi_count++;
183  }
184  
185  /*
186   * hwlat_err - report a hwlat error.
187   */
188  #define hwlat_err(msg) ({							\
189  	struct trace_array *tr = hwlat_trace;					\
190  										\
191  	trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg);	\
192  })
193  
194  /**
195   * get_sample - sample the CPU TSC and look for likely hardware latencies
196   *
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.
200   */
get_sample(void)201  static int get_sample(void)
202  {
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;
208  	u64 sample = 0;
209  	u64 thresh = tracing_thresh;
210  	u64 outer_sample = 0;
211  	int ret = -1;
212  	unsigned int count = 0;
213  
214  	do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
215  
216  	kdata->nmi_total_ts = 0;
217  	kdata->nmi_count = 0;
218  	/* Make sure NMIs see this first */
219  	barrier();
220  
221  	trace_hwlat_callback_enabled = true;
222  
223  	init_time(last_t2, 0);
224  	start = time_get(); /* start timestamp */
225  	outer_diff = 0;
226  
227  	do {
228  
229  		t1 = time_get();	/* we'll look for a discontinuity */
230  		t2 = time_get();
231  
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");
238  				goto out;
239  			}
240  			if (outer_diff > outer_sample)
241  				outer_sample = outer_diff;
242  		}
243  		last_t2 = t2;
244  
245  		total = time_to_us(time_sub(t2, start)); /* sample width */
246  
247  		/* Check for possible overflows */
248  		if (total < last_total) {
249  			hwlat_err("Time total overflowed\n");
250  			break;
251  		}
252  		last_total = total;
253  
254  		/* This checks the inner loop (t1 to t2) */
255  		diff = time_to_us(time_sub(t2, t1));     /* current diff */
256  
257  		if (diff > thresh || outer_diff > thresh) {
258  			if (!count)
259  				ktime_get_real_ts64(&s.timestamp);
260  			count++;
261  		}
262  
263  		/* This shouldn't happen */
264  		if (diff < 0) {
265  			hwlat_err(BANNER "time running backwards\n");
266  			goto out;
267  		}
268  
269  		if (diff > sample)
270  			sample = diff; /* only want highest value */
271  
272  	} while (total <= hwlat_data.sample_width);
273  
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 */
277  
278  	ret = 0;
279  
280  	/* If we exceed the threshold value, we have found a hardware latency */
281  	if (sample > thresh || outer_sample > thresh) {
282  		u64 latency;
283  
284  		ret = 1;
285  
286  		/* We read in microseconds */
287  		if (kdata->nmi_total_ts)
288  			do_div(kdata->nmi_total_ts, NSEC_PER_USEC);
289  
290  		hwlat_data.count++;
291  		s.seqnum = hwlat_data.count;
292  		s.duration = sample;
293  		s.outer_duration = outer_sample;
294  		s.nmi_total_ts = kdata->nmi_total_ts;
295  		s.nmi_count = kdata->nmi_count;
296  		s.count = count;
297  		trace_hwlat_sample(&s);
298  
299  		latency = max(sample, outer_sample);
300  
301  		/* Keep a running maximum ever recorded hardware latency */
302  		if (latency > tr->max_latency) {
303  			tr->max_latency = latency;
304  			latency_fsnotify(tr);
305  		}
306  	}
307  
308  out:
309  	return ret;
310  }
311  
312  static struct cpumask save_cpumask;
313  
move_to_next_cpu(void)314  static void move_to_next_cpu(void)
315  {
316  	struct cpumask *current_mask = &save_cpumask;
317  	struct trace_array *tr = hwlat_trace;
318  	int next_cpu;
319  
320  	/*
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.
324  	 */
325  	if (!cpumask_equal(current_mask, current->cpus_ptr))
326  		goto change_mode;
327  
328  	cpus_read_lock();
329  	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
330  	next_cpu = cpumask_next(raw_smp_processor_id(), current_mask);
331  	cpus_read_unlock();
332  
333  	if (next_cpu >= nr_cpu_ids)
334  		next_cpu = cpumask_first(current_mask);
335  
336  	if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
337  		goto change_mode;
338  
339  	cpumask_clear(current_mask);
340  	cpumask_set_cpu(next_cpu, current_mask);
341  
342  	set_cpus_allowed_ptr(current, current_mask);
343  	return;
344  
345   change_mode:
346  	hwlat_data.thread_mode = MODE_NONE;
347  	pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
348  }
349  
350  /*
351   * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
352   *
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.
357   *
358   * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
359   */
kthread_fn(void * data)360  static int kthread_fn(void *data)
361  {
362  	u64 interval;
363  
364  	while (!kthread_should_stop()) {
365  
366  		if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
367  			move_to_next_cpu();
368  
369  		local_irq_disable();
370  		get_sample();
371  		local_irq_enable();
372  
373  		mutex_lock(&hwlat_data.lock);
374  		interval = hwlat_data.sample_window - hwlat_data.sample_width;
375  		mutex_unlock(&hwlat_data.lock);
376  
377  		do_div(interval, USEC_PER_MSEC); /* modifies interval value */
378  
379  		/* Always sleep for at least 1ms */
380  		if (interval < 1)
381  			interval = 1;
382  
383  		if (msleep_interruptible(interval))
384  			break;
385  	}
386  
387  	return 0;
388  }
389  
390  /*
391   * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
392   *
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.
395   */
stop_single_kthread(void)396  static void stop_single_kthread(void)
397  {
398  	struct hwlat_kthread_data *kdata = get_cpu_data();
399  	struct task_struct *kthread;
400  
401  	cpus_read_lock();
402  	kthread = kdata->kthread;
403  
404  	if (!kthread)
405  		goto out_put_cpus;
406  
407  	kthread_stop(kthread);
408  	kdata->kthread = NULL;
409  
410  out_put_cpus:
411  	cpus_read_unlock();
412  }
413  
414  
415  /*
416   * start_single_kthread - Kick off the hardware latency sampling/detector kthread
417   *
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.
420   */
start_single_kthread(struct trace_array * tr)421  static int start_single_kthread(struct trace_array *tr)
422  {
423  	struct hwlat_kthread_data *kdata = get_cpu_data();
424  	struct cpumask *current_mask = &save_cpumask;
425  	struct task_struct *kthread;
426  	int next_cpu;
427  
428  	cpus_read_lock();
429  	if (kdata->kthread)
430  		goto out_put_cpus;
431  
432  	kthread = kthread_create(kthread_fn, NULL, "hwlatd");
433  	if (IS_ERR(kthread)) {
434  		pr_err(BANNER "could not start sampling thread\n");
435  		cpus_read_unlock();
436  		return -ENOMEM;
437  	}
438  
439  	/* Just pick the first CPU on first iteration */
440  	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
441  
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);
446  
447  	}
448  
449  	set_cpus_allowed_ptr(kthread, current_mask);
450  
451  	kdata->kthread = kthread;
452  	wake_up_process(kthread);
453  
454  out_put_cpus:
455  	cpus_read_unlock();
456  	return 0;
457  }
458  
459  /*
460   * stop_cpu_kthread - Stop a hwlat cpu kthread
461   */
stop_cpu_kthread(unsigned int cpu)462  static void stop_cpu_kthread(unsigned int cpu)
463  {
464  	struct task_struct *kthread;
465  
466  	kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
467  	if (kthread)
468  		kthread_stop(kthread);
469  	per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
470  }
471  
472  /*
473   * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
474   *
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.
477   */
stop_per_cpu_kthreads(void)478  static void stop_per_cpu_kthreads(void)
479  {
480  	unsigned int cpu;
481  
482  	cpus_read_lock();
483  	for_each_online_cpu(cpu)
484  		stop_cpu_kthread(cpu);
485  	cpus_read_unlock();
486  }
487  
488  /*
489   * start_cpu_kthread - Start a hwlat cpu kthread
490   */
start_cpu_kthread(unsigned int cpu)491  static int start_cpu_kthread(unsigned int cpu)
492  {
493  	struct task_struct *kthread;
494  
495  	/* Do not start a new hwlatd thread if it is already running */
496  	if (per_cpu(hwlat_per_cpu_data, cpu).kthread)
497  		return 0;
498  
499  	kthread = kthread_run_on_cpu(kthread_fn, NULL, cpu, "hwlatd/%u");
500  	if (IS_ERR(kthread)) {
501  		pr_err(BANNER "could not start sampling thread\n");
502  		return -ENOMEM;
503  	}
504  
505  	per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
506  
507  	return 0;
508  }
509  
510  #ifdef CONFIG_HOTPLUG_CPU
hwlat_hotplug_workfn(struct work_struct * dummy)511  static void hwlat_hotplug_workfn(struct work_struct *dummy)
512  {
513  	struct trace_array *tr = hwlat_trace;
514  	unsigned int cpu = smp_processor_id();
515  
516  	mutex_lock(&trace_types_lock);
517  	mutex_lock(&hwlat_data.lock);
518  	cpus_read_lock();
519  
520  	if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
521  		goto out_unlock;
522  
523  	if (!cpu_online(cpu))
524  		goto out_unlock;
525  	if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
526  		goto out_unlock;
527  
528  	start_cpu_kthread(cpu);
529  
530  out_unlock:
531  	cpus_read_unlock();
532  	mutex_unlock(&hwlat_data.lock);
533  	mutex_unlock(&trace_types_lock);
534  }
535  
536  static DECLARE_WORK(hwlat_hotplug_work, hwlat_hotplug_workfn);
537  
538  /*
539   * hwlat_cpu_init - CPU hotplug online callback function
540   */
hwlat_cpu_init(unsigned int cpu)541  static int hwlat_cpu_init(unsigned int cpu)
542  {
543  	schedule_work_on(cpu, &hwlat_hotplug_work);
544  	return 0;
545  }
546  
547  /*
548   * hwlat_cpu_die - CPU hotplug offline callback function
549   */
hwlat_cpu_die(unsigned int cpu)550  static int hwlat_cpu_die(unsigned int cpu)
551  {
552  	stop_cpu_kthread(cpu);
553  	return 0;
554  }
555  
hwlat_init_hotplug_support(void)556  static void hwlat_init_hotplug_support(void)
557  {
558  	int ret;
559  
560  	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/hwlat:online",
561  				hwlat_cpu_init, hwlat_cpu_die);
562  	if (ret < 0)
563  		pr_warn(BANNER "Error to init cpu hotplug support\n");
564  
565  	return;
566  }
567  #else /* CONFIG_HOTPLUG_CPU */
hwlat_init_hotplug_support(void)568  static void hwlat_init_hotplug_support(void)
569  {
570  	return;
571  }
572  #endif /* CONFIG_HOTPLUG_CPU */
573  
574  /*
575   * start_per_cpu_kthreads - Kick off the hardware latency sampling/detector kthreads
576   *
577   * This starts the kernel threads that will sit on potentially all cpus and
578   * sample the CPU timestamp counter (TSC or similar) and look for potential
579   * hardware latencies.
580   */
start_per_cpu_kthreads(struct trace_array * tr)581  static int start_per_cpu_kthreads(struct trace_array *tr)
582  {
583  	struct cpumask *current_mask = &save_cpumask;
584  	unsigned int cpu;
585  	int retval;
586  
587  	cpus_read_lock();
588  	/*
589  	 * Run only on CPUs in which hwlat is allowed to run.
590  	 */
591  	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
592  
593  	for_each_cpu(cpu, current_mask) {
594  		retval = start_cpu_kthread(cpu);
595  		if (retval)
596  			goto out_error;
597  	}
598  	cpus_read_unlock();
599  
600  	return 0;
601  
602  out_error:
603  	cpus_read_unlock();
604  	stop_per_cpu_kthreads();
605  	return retval;
606  }
607  
s_mode_start(struct seq_file * s,loff_t * pos)608  static void *s_mode_start(struct seq_file *s, loff_t *pos)
609  {
610  	int mode = *pos;
611  
612  	mutex_lock(&hwlat_data.lock);
613  
614  	if (mode >= MODE_MAX)
615  		return NULL;
616  
617  	return pos;
618  }
619  
s_mode_next(struct seq_file * s,void * v,loff_t * pos)620  static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
621  {
622  	int mode = ++(*pos);
623  
624  	if (mode >= MODE_MAX)
625  		return NULL;
626  
627  	return pos;
628  }
629  
s_mode_show(struct seq_file * s,void * v)630  static int s_mode_show(struct seq_file *s, void *v)
631  {
632  	loff_t *pos = v;
633  	int mode = *pos;
634  
635  	if (mode == hwlat_data.thread_mode)
636  		seq_printf(s, "[%s]", thread_mode_str[mode]);
637  	else
638  		seq_printf(s, "%s", thread_mode_str[mode]);
639  
640  	if (mode < MODE_MAX - 1) /* if mode is any but last */
641  		seq_puts(s, " ");
642  
643  	return 0;
644  }
645  
s_mode_stop(struct seq_file * s,void * v)646  static void s_mode_stop(struct seq_file *s, void *v)
647  {
648  	seq_puts(s, "\n");
649  	mutex_unlock(&hwlat_data.lock);
650  }
651  
652  static const struct seq_operations thread_mode_seq_ops = {
653  	.start		= s_mode_start,
654  	.next		= s_mode_next,
655  	.show		= s_mode_show,
656  	.stop		= s_mode_stop
657  };
658  
hwlat_mode_open(struct inode * inode,struct file * file)659  static int hwlat_mode_open(struct inode *inode, struct file *file)
660  {
661  	return seq_open(file, &thread_mode_seq_ops);
662  };
663  
664  static void hwlat_tracer_start(struct trace_array *tr);
665  static void hwlat_tracer_stop(struct trace_array *tr);
666  
667  /**
668   * hwlat_mode_write - Write function for "mode" entry
669   * @filp: The active open file structure
670   * @ubuf: The user buffer that contains the value to write
671   * @cnt: The maximum number of bytes to write to "file"
672   * @ppos: The current position in @file
673   *
674   * This function provides a write implementation for the "mode" interface
675   * to the hardware latency detector. hwlatd has different operation modes.
676   * The "none" sets the allowed cpumask for a single hwlatd thread at the
677   * startup and lets the scheduler handle the migration. The default mode is
678   * the "round-robin" one, in which a single hwlatd thread runs, migrating
679   * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
680   * creates one hwlatd thread per allowed CPU.
681   */
hwlat_mode_write(struct file * filp,const char __user * ubuf,size_t cnt,loff_t * ppos)682  static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
683  				 size_t cnt, loff_t *ppos)
684  {
685  	struct trace_array *tr = hwlat_trace;
686  	const char *mode;
687  	char buf[64];
688  	int ret, i;
689  
690  	if (cnt >= sizeof(buf))
691  		return -EINVAL;
692  
693  	if (copy_from_user(buf, ubuf, cnt))
694  		return -EFAULT;
695  
696  	buf[cnt] = 0;
697  
698  	mode = strstrip(buf);
699  
700  	ret = -EINVAL;
701  
702  	/*
703  	 * trace_types_lock is taken to avoid concurrency on start/stop
704  	 * and hwlat_busy.
705  	 */
706  	mutex_lock(&trace_types_lock);
707  	if (hwlat_busy)
708  		hwlat_tracer_stop(tr);
709  
710  	mutex_lock(&hwlat_data.lock);
711  
712  	for (i = 0; i < MODE_MAX; i++) {
713  		if (strcmp(mode, thread_mode_str[i]) == 0) {
714  			hwlat_data.thread_mode = i;
715  			ret = cnt;
716  		}
717  	}
718  
719  	mutex_unlock(&hwlat_data.lock);
720  
721  	if (hwlat_busy)
722  		hwlat_tracer_start(tr);
723  	mutex_unlock(&trace_types_lock);
724  
725  	*ppos += cnt;
726  
727  
728  
729  	return ret;
730  }
731  
732  /*
733   * The width parameter is read/write using the generic trace_min_max_param
734   * method. The *val is protected by the hwlat_data lock and is upper
735   * bounded by the window parameter.
736   */
737  static struct trace_min_max_param hwlat_width = {
738  	.lock		= &hwlat_data.lock,
739  	.val		= &hwlat_data.sample_width,
740  	.max		= &hwlat_data.sample_window,
741  	.min		= NULL,
742  };
743  
744  /*
745   * The window parameter is read/write using the generic trace_min_max_param
746   * method. The *val is protected by the hwlat_data lock and is lower
747   * bounded by the width parameter.
748   */
749  static struct trace_min_max_param hwlat_window = {
750  	.lock		= &hwlat_data.lock,
751  	.val		= &hwlat_data.sample_window,
752  	.max		= NULL,
753  	.min		= &hwlat_data.sample_width,
754  };
755  
756  static const struct file_operations thread_mode_fops = {
757  	.open		= hwlat_mode_open,
758  	.read		= seq_read,
759  	.llseek		= seq_lseek,
760  	.release	= seq_release,
761  	.write		= hwlat_mode_write
762  };
763  /**
764   * init_tracefs - A function to initialize the tracefs interface files
765   *
766   * This function creates entries in tracefs for "hwlat_detector".
767   * It creates the hwlat_detector directory in the tracing directory,
768   * and within that directory is the count, width and window files to
769   * change and view those values.
770   */
init_tracefs(void)771  static int init_tracefs(void)
772  {
773  	int ret;
774  	struct dentry *top_dir;
775  
776  	ret = tracing_init_dentry();
777  	if (ret)
778  		return -ENOMEM;
779  
780  	top_dir = tracefs_create_dir("hwlat_detector", NULL);
781  	if (!top_dir)
782  		return -ENOMEM;
783  
784  	hwlat_sample_window = tracefs_create_file("window", TRACE_MODE_WRITE,
785  						  top_dir,
786  						  &hwlat_window,
787  						  &trace_min_max_fops);
788  	if (!hwlat_sample_window)
789  		goto err;
790  
791  	hwlat_sample_width = tracefs_create_file("width", TRACE_MODE_WRITE,
792  						 top_dir,
793  						 &hwlat_width,
794  						 &trace_min_max_fops);
795  	if (!hwlat_sample_width)
796  		goto err;
797  
798  	hwlat_thread_mode = trace_create_file("mode", TRACE_MODE_WRITE,
799  					      top_dir,
800  					      NULL,
801  					      &thread_mode_fops);
802  	if (!hwlat_thread_mode)
803  		goto err;
804  
805  	return 0;
806  
807   err:
808  	tracefs_remove(top_dir);
809  	return -ENOMEM;
810  }
811  
hwlat_tracer_start(struct trace_array * tr)812  static void hwlat_tracer_start(struct trace_array *tr)
813  {
814  	int err;
815  
816  	if (hwlat_data.thread_mode == MODE_PER_CPU)
817  		err = start_per_cpu_kthreads(tr);
818  	else
819  		err = start_single_kthread(tr);
820  	if (err)
821  		pr_err(BANNER "Cannot start hwlat kthread\n");
822  }
823  
hwlat_tracer_stop(struct trace_array * tr)824  static void hwlat_tracer_stop(struct trace_array *tr)
825  {
826  	if (hwlat_data.thread_mode == MODE_PER_CPU)
827  		stop_per_cpu_kthreads();
828  	else
829  		stop_single_kthread();
830  }
831  
hwlat_tracer_init(struct trace_array * tr)832  static int hwlat_tracer_init(struct trace_array *tr)
833  {
834  	/* Only allow one instance to enable this */
835  	if (hwlat_busy)
836  		return -EBUSY;
837  
838  	hwlat_trace = tr;
839  
840  	hwlat_data.count = 0;
841  	tr->max_latency = 0;
842  	save_tracing_thresh = tracing_thresh;
843  
844  	/* tracing_thresh is in nsecs, we speak in usecs */
845  	if (!tracing_thresh)
846  		tracing_thresh = last_tracing_thresh;
847  
848  	if (tracer_tracing_is_on(tr))
849  		hwlat_tracer_start(tr);
850  
851  	hwlat_busy = true;
852  
853  	return 0;
854  }
855  
hwlat_tracer_reset(struct trace_array * tr)856  static void hwlat_tracer_reset(struct trace_array *tr)
857  {
858  	hwlat_tracer_stop(tr);
859  
860  	/* the tracing threshold is static between runs */
861  	last_tracing_thresh = tracing_thresh;
862  
863  	tracing_thresh = save_tracing_thresh;
864  	hwlat_busy = false;
865  }
866  
867  static struct tracer hwlat_tracer __read_mostly =
868  {
869  	.name		= "hwlat",
870  	.init		= hwlat_tracer_init,
871  	.reset		= hwlat_tracer_reset,
872  	.start		= hwlat_tracer_start,
873  	.stop		= hwlat_tracer_stop,
874  	.allow_instances = true,
875  };
876  
init_hwlat_tracer(void)877  __init static int init_hwlat_tracer(void)
878  {
879  	int ret;
880  
881  	mutex_init(&hwlat_data.lock);
882  
883  	ret = register_tracer(&hwlat_tracer);
884  	if (ret)
885  		return ret;
886  
887  	hwlat_init_hotplug_support();
888  
889  	init_tracefs();
890  
891  	return 0;
892  }
893  late_initcall(init_hwlat_tracer);
894