90c4f70dc9fdf4dd65c55ff51b319d82ed8c0de1
[releases.git] / trace_osnoise.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * OS Noise Tracer: computes the OS Noise suffered by a running thread.
4  * Timerlat Tracer: measures the wakeup latency of a timer triggered IRQ and thread.
5  *
6  * Based on "hwlat_detector" tracer by:
7  *   Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
8  *   Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
9  *   With feedback from Clark Williams <williams@redhat.com>
10  *
11  * And also based on the rtsl tracer presented on:
12  *  DE OLIVEIRA, Daniel Bristot, et al. Demystifying the real-time linux
13  *  scheduling latency. In: 32nd Euromicro Conference on Real-Time Systems
14  *  (ECRTS 2020). Schloss Dagstuhl-Leibniz-Zentrum fur Informatik, 2020.
15  *
16  * Copyright (C) 2021 Daniel Bristot de Oliveira, Red Hat, Inc. <bristot@redhat.com>
17  */
18
19 #include <linux/kthread.h>
20 #include <linux/tracefs.h>
21 #include <linux/uaccess.h>
22 #include <linux/cpumask.h>
23 #include <linux/delay.h>
24 #include <linux/sched/clock.h>
25 #include <uapi/linux/sched/types.h>
26 #include <linux/sched.h>
27 #include "trace.h"
28
29 #ifdef CONFIG_X86_LOCAL_APIC
30 #include <asm/trace/irq_vectors.h>
31 #undef TRACE_INCLUDE_PATH
32 #undef TRACE_INCLUDE_FILE
33 #endif /* CONFIG_X86_LOCAL_APIC */
34
35 #include <trace/events/irq.h>
36 #include <trace/events/sched.h>
37
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/osnoise.h>
40
41 static struct trace_array       *osnoise_trace;
42
43 /*
44  * Default values.
45  */
46 #define BANNER                  "osnoise: "
47 #define DEFAULT_SAMPLE_PERIOD   1000000                 /* 1s */
48 #define DEFAULT_SAMPLE_RUNTIME  1000000                 /* 1s */
49
50 #define DEFAULT_TIMERLAT_PERIOD 1000                    /* 1ms */
51 #define DEFAULT_TIMERLAT_PRIO   95                      /* FIFO 95 */
52
53 /*
54  * NMI runtime info.
55  */
56 struct osn_nmi {
57         u64     count;
58         u64     delta_start;
59 };
60
61 /*
62  * IRQ runtime info.
63  */
64 struct osn_irq {
65         u64     count;
66         u64     arrival_time;
67         u64     delta_start;
68 };
69
70 #define IRQ_CONTEXT     0
71 #define THREAD_CONTEXT  1
72 /*
73  * sofirq runtime info.
74  */
75 struct osn_softirq {
76         u64     count;
77         u64     arrival_time;
78         u64     delta_start;
79 };
80
81 /*
82  * thread runtime info.
83  */
84 struct osn_thread {
85         u64     count;
86         u64     arrival_time;
87         u64     delta_start;
88 };
89
90 /*
91  * Runtime information: this structure saves the runtime information used by
92  * one sampling thread.
93  */
94 struct osnoise_variables {
95         struct task_struct      *kthread;
96         bool                    sampling;
97         pid_t                   pid;
98         struct osn_nmi          nmi;
99         struct osn_irq          irq;
100         struct osn_softirq      softirq;
101         struct osn_thread       thread;
102         local_t                 int_counter;
103 };
104
105 /*
106  * Per-cpu runtime information.
107  */
108 DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var);
109
110 /*
111  * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU
112  */
113 static inline struct osnoise_variables *this_cpu_osn_var(void)
114 {
115         return this_cpu_ptr(&per_cpu_osnoise_var);
116 }
117
118 #ifdef CONFIG_TIMERLAT_TRACER
119 /*
120  * Runtime information for the timer mode.
121  */
122 struct timerlat_variables {
123         struct task_struct      *kthread;
124         struct hrtimer          timer;
125         u64                     rel_period;
126         u64                     abs_period;
127         bool                    tracing_thread;
128         u64                     count;
129 };
130
131 DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var);
132
133 /*
134  * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU
135  */
136 static inline struct timerlat_variables *this_cpu_tmr_var(void)
137 {
138         return this_cpu_ptr(&per_cpu_timerlat_var);
139 }
140
141 /*
142  * tlat_var_reset - Reset the values of the given timerlat_variables
143  */
144 static inline void tlat_var_reset(void)
145 {
146         struct timerlat_variables *tlat_var;
147         int cpu;
148         /*
149          * So far, all the values are initialized as 0, so
150          * zeroing the structure is perfect.
151          */
152         for_each_cpu(cpu, cpu_online_mask) {
153                 tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu);
154                 memset(tlat_var, 0, sizeof(*tlat_var));
155         }
156 }
157 #else /* CONFIG_TIMERLAT_TRACER */
158 #define tlat_var_reset()        do {} while (0)
159 #endif /* CONFIG_TIMERLAT_TRACER */
160
161 /*
162  * osn_var_reset - Reset the values of the given osnoise_variables
163  */
164 static inline void osn_var_reset(void)
165 {
166         struct osnoise_variables *osn_var;
167         int cpu;
168
169         /*
170          * So far, all the values are initialized as 0, so
171          * zeroing the structure is perfect.
172          */
173         for_each_cpu(cpu, cpu_online_mask) {
174                 osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu);
175                 memset(osn_var, 0, sizeof(*osn_var));
176         }
177 }
178
179 /*
180  * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables
181  */
182 static inline void osn_var_reset_all(void)
183 {
184         osn_var_reset();
185         tlat_var_reset();
186 }
187
188 /*
189  * Tells NMIs to call back to the osnoise tracer to record timestamps.
190  */
191 bool trace_osnoise_callback_enabled;
192
193 /*
194  * osnoise sample structure definition. Used to store the statistics of a
195  * sample run.
196  */
197 struct osnoise_sample {
198         u64                     runtime;        /* runtime */
199         u64                     noise;          /* noise */
200         u64                     max_sample;     /* max single noise sample */
201         int                     hw_count;       /* # HW (incl. hypervisor) interference */
202         int                     nmi_count;      /* # NMIs during this sample */
203         int                     irq_count;      /* # IRQs during this sample */
204         int                     softirq_count;  /* # softirqs during this sample */
205         int                     thread_count;   /* # threads during this sample */
206 };
207
208 #ifdef CONFIG_TIMERLAT_TRACER
209 /*
210  * timerlat sample structure definition. Used to store the statistics of
211  * a sample run.
212  */
213 struct timerlat_sample {
214         u64                     timer_latency;  /* timer_latency */
215         unsigned int            seqnum;         /* unique sequence */
216         int                     context;        /* timer context */
217 };
218 #endif
219
220 /*
221  * Protect the interface.
222  */
223 struct mutex interface_lock;
224
225 /*
226  * Tracer data.
227  */
228 static struct osnoise_data {
229         u64     sample_period;          /* total sampling period */
230         u64     sample_runtime;         /* active sampling portion of period */
231         u64     stop_tracing;           /* stop trace in the internal operation (loop/irq) */
232         u64     stop_tracing_total;     /* stop trace in the final operation (report/thread) */
233 #ifdef CONFIG_TIMERLAT_TRACER
234         u64     timerlat_period;        /* timerlat period */
235         u64     print_stack;            /* print IRQ stack if total > */
236         int     timerlat_tracer;        /* timerlat tracer */
237 #endif
238         bool    tainted;                /* infor users and developers about a problem */
239 } osnoise_data = {
240         .sample_period                  = DEFAULT_SAMPLE_PERIOD,
241         .sample_runtime                 = DEFAULT_SAMPLE_RUNTIME,
242         .stop_tracing                   = 0,
243         .stop_tracing_total             = 0,
244 #ifdef CONFIG_TIMERLAT_TRACER
245         .print_stack                    = 0,
246         .timerlat_period                = DEFAULT_TIMERLAT_PERIOD,
247         .timerlat_tracer                = 0,
248 #endif
249 };
250
251 /*
252  * Boolean variable used to inform that the tracer is currently sampling.
253  */
254 static bool osnoise_busy;
255
256 #ifdef CONFIG_PREEMPT_RT
257 /*
258  * Print the osnoise header info.
259  */
260 static void print_osnoise_headers(struct seq_file *s)
261 {
262         if (osnoise_data.tainted)
263                 seq_puts(s, "# osnoise is tainted!\n");
264
265         seq_puts(s, "#                                _-------=> irqs-off\n");
266         seq_puts(s, "#                               / _------=> need-resched\n");
267         seq_puts(s, "#                              | / _-----=> need-resched-lazy\n");
268         seq_puts(s, "#                              || / _----=> hardirq/softirq\n");
269         seq_puts(s, "#                              ||| / _---=> preempt-depth\n");
270         seq_puts(s, "#                              |||| / _--=> preempt-lazy-depth\n");
271         seq_puts(s, "#                              ||||| / _-=> migrate-disable\n");
272
273         seq_puts(s, "#                              |||||| /          ");
274         seq_puts(s, "                                     MAX\n");
275
276         seq_puts(s, "#                              ||||| /                         ");
277         seq_puts(s, "                    SINGLE      Interference counters:\n");
278
279         seq_puts(s, "#                              |||||||               RUNTIME   ");
280         seq_puts(s, "   NOISE  %% OF CPU  NOISE    +-----------------------------+\n");
281
282         seq_puts(s, "#           TASK-PID      CPU# |||||||   TIMESTAMP    IN US    ");
283         seq_puts(s, "   IN US  AVAILABLE  IN US     HW    NMI    IRQ   SIRQ THREAD\n");
284
285         seq_puts(s, "#              | |         |   |||||||      |           |      ");
286         seq_puts(s, "       |    |            |      |      |      |      |      |\n");
287 }
288 #else /* CONFIG_PREEMPT_RT */
289 static void print_osnoise_headers(struct seq_file *s)
290 {
291         if (osnoise_data.tainted)
292                 seq_puts(s, "# osnoise is tainted!\n");
293
294         seq_puts(s, "#                                _-----=> irqs-off\n");
295         seq_puts(s, "#                               / _----=> need-resched\n");
296         seq_puts(s, "#                              | / _---=> hardirq/softirq\n");
297         seq_puts(s, "#                              || / _--=> preempt-depth     ");
298         seq_puts(s, "                       MAX\n");
299
300         seq_puts(s, "#                              || /                         ");
301         seq_puts(s, "                    SINGLE      Interference counters:\n");
302
303         seq_puts(s, "#                              ||||               RUNTIME   ");
304         seq_puts(s, "   NOISE  %% OF CPU  NOISE    +-----------------------------+\n");
305
306         seq_puts(s, "#           TASK-PID      CPU# ||||   TIMESTAMP    IN US    ");
307         seq_puts(s, "   IN US  AVAILABLE  IN US     HW    NMI    IRQ   SIRQ THREAD\n");
308
309         seq_puts(s, "#              | |         |   ||||      |           |      ");
310         seq_puts(s, "       |    |            |      |      |      |      |      |\n");
311 }
312 #endif /* CONFIG_PREEMPT_RT */
313
314 /*
315  * osnoise_taint - report an osnoise error.
316  */
317 #define osnoise_taint(msg) ({                                                   \
318         struct trace_array *tr = osnoise_trace;                                 \
319                                                                                 \
320         trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg);        \
321         osnoise_data.tainted = true;                                            \
322 })
323
324 /*
325  * Record an osnoise_sample into the tracer buffer.
326  */
327 static void trace_osnoise_sample(struct osnoise_sample *sample)
328 {
329         struct trace_array *tr = osnoise_trace;
330         struct trace_buffer *buffer = tr->array_buffer.buffer;
331         struct trace_event_call *call = &event_osnoise;
332         struct ring_buffer_event *event;
333         struct osnoise_entry *entry;
334
335         event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry),
336                                           tracing_gen_ctx());
337         if (!event)
338                 return;
339         entry   = ring_buffer_event_data(event);
340         entry->runtime          = sample->runtime;
341         entry->noise            = sample->noise;
342         entry->max_sample       = sample->max_sample;
343         entry->hw_count         = sample->hw_count;
344         entry->nmi_count        = sample->nmi_count;
345         entry->irq_count        = sample->irq_count;
346         entry->softirq_count    = sample->softirq_count;
347         entry->thread_count     = sample->thread_count;
348
349         if (!call_filter_check_discard(call, entry, buffer, event))
350                 trace_buffer_unlock_commit_nostack(buffer, event);
351 }
352
353 #ifdef CONFIG_TIMERLAT_TRACER
354 /*
355  * Print the timerlat header info.
356  */
357 #ifdef CONFIG_PREEMPT_RT
358 static void print_timerlat_headers(struct seq_file *s)
359 {
360         seq_puts(s, "#                                _-------=> irqs-off\n");
361         seq_puts(s, "#                               / _------=> need-resched\n");
362         seq_puts(s, "#                              | / _-----=> need-resched-lazy\n");
363         seq_puts(s, "#                              || / _----=> hardirq/softirq\n");
364         seq_puts(s, "#                              ||| / _---=> preempt-depth\n");
365         seq_puts(s, "#                              |||| / _--=> preempt-lazy-depth\n");
366         seq_puts(s, "#                              ||||| / _-=> migrate-disable\n");
367         seq_puts(s, "#                              |||||| /\n");
368         seq_puts(s, "#                              |||||||             ACTIVATION\n");
369         seq_puts(s, "#           TASK-PID      CPU# |||||||   TIMESTAMP    ID     ");
370         seq_puts(s, "       CONTEXT                LATENCY\n");
371         seq_puts(s, "#              | |         |   |||||||      |         |      ");
372         seq_puts(s, "            |                       |\n");
373 }
374 #else /* CONFIG_PREEMPT_RT */
375 static void print_timerlat_headers(struct seq_file *s)
376 {
377         seq_puts(s, "#                                _-----=> irqs-off\n");
378         seq_puts(s, "#                               / _----=> need-resched\n");
379         seq_puts(s, "#                              | / _---=> hardirq/softirq\n");
380         seq_puts(s, "#                              || / _--=> preempt-depth\n");
381         seq_puts(s, "#                              || /\n");
382         seq_puts(s, "#                              ||||             ACTIVATION\n");
383         seq_puts(s, "#           TASK-PID      CPU# ||||   TIMESTAMP    ID     ");
384         seq_puts(s, "       CONTEXT                LATENCY\n");
385         seq_puts(s, "#              | |         |   ||||      |         |      ");
386         seq_puts(s, "            |                       |\n");
387 }
388 #endif /* CONFIG_PREEMPT_RT */
389
390 /*
391  * Record an timerlat_sample into the tracer buffer.
392  */
393 static void trace_timerlat_sample(struct timerlat_sample *sample)
394 {
395         struct trace_array *tr = osnoise_trace;
396         struct trace_event_call *call = &event_osnoise;
397         struct trace_buffer *buffer = tr->array_buffer.buffer;
398         struct ring_buffer_event *event;
399         struct timerlat_entry *entry;
400
401         event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry),
402                                           tracing_gen_ctx());
403         if (!event)
404                 return;
405         entry   = ring_buffer_event_data(event);
406         entry->seqnum                   = sample->seqnum;
407         entry->context                  = sample->context;
408         entry->timer_latency            = sample->timer_latency;
409
410         if (!call_filter_check_discard(call, entry, buffer, event))
411                 trace_buffer_unlock_commit_nostack(buffer, event);
412 }
413
414 #ifdef CONFIG_STACKTRACE
415
416 #define MAX_CALLS       256
417
418 /*
419  * Stack trace will take place only at IRQ level, so, no need
420  * to control nesting here.
421  */
422 struct trace_stack {
423         int             stack_size;
424         int             nr_entries;
425         unsigned long   calls[MAX_CALLS];
426 };
427
428 static DEFINE_PER_CPU(struct trace_stack, trace_stack);
429
430 /*
431  * timerlat_save_stack - save a stack trace without printing
432  *
433  * Save the current stack trace without printing. The
434  * stack will be printed later, after the end of the measurement.
435  */
436 static void timerlat_save_stack(int skip)
437 {
438         unsigned int size, nr_entries;
439         struct trace_stack *fstack;
440
441         fstack = this_cpu_ptr(&trace_stack);
442
443         size = ARRAY_SIZE(fstack->calls);
444
445         nr_entries = stack_trace_save(fstack->calls, size, skip);
446
447         fstack->stack_size = nr_entries * sizeof(unsigned long);
448         fstack->nr_entries = nr_entries;
449
450         return;
451
452 }
453 /*
454  * timerlat_dump_stack - dump a stack trace previously saved
455  *
456  * Dump a saved stack trace into the trace buffer.
457  */
458 static void timerlat_dump_stack(void)
459 {
460         struct trace_event_call *call = &event_osnoise;
461         struct trace_array *tr = osnoise_trace;
462         struct trace_buffer *buffer = tr->array_buffer.buffer;
463         struct ring_buffer_event *event;
464         struct trace_stack *fstack;
465         struct stack_entry *entry;
466         unsigned int size;
467
468         preempt_disable_notrace();
469         fstack = this_cpu_ptr(&trace_stack);
470         size = fstack->stack_size;
471
472         event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size,
473                                           tracing_gen_ctx());
474         if (!event)
475                 goto out;
476
477         entry = ring_buffer_event_data(event);
478
479         memcpy(&entry->caller, fstack->calls, size);
480         entry->size = fstack->nr_entries;
481
482         if (!call_filter_check_discard(call, entry, buffer, event))
483                 trace_buffer_unlock_commit_nostack(buffer, event);
484
485 out:
486         preempt_enable_notrace();
487 }
488 #else
489 #define timerlat_dump_stack() do {} while (0)
490 #define timerlat_save_stack(a) do {} while (0)
491 #endif /* CONFIG_STACKTRACE */
492 #endif /* CONFIG_TIMERLAT_TRACER */
493
494 /*
495  * Macros to encapsulate the time capturing infrastructure.
496  */
497 #define time_get()      trace_clock_local()
498 #define time_to_us(x)   div_u64(x, 1000)
499 #define time_sub(a, b)  ((a) - (b))
500
501 /*
502  * cond_move_irq_delta_start - Forward the delta_start of a running IRQ
503  *
504  * If an IRQ is preempted by an NMI, its delta_start is pushed forward
505  * to discount the NMI interference.
506  *
507  * See get_int_safe_duration().
508  */
509 static inline void
510 cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration)
511 {
512         if (osn_var->irq.delta_start)
513                 osn_var->irq.delta_start += duration;
514 }
515
516 #ifndef CONFIG_PREEMPT_RT
517 /*
518  * cond_move_softirq_delta_start - Forward the delta_start of a running softirq.
519  *
520  * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed
521  * forward to discount the interference.
522  *
523  * See get_int_safe_duration().
524  */
525 static inline void
526 cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration)
527 {
528         if (osn_var->softirq.delta_start)
529                 osn_var->softirq.delta_start += duration;
530 }
531 #else /* CONFIG_PREEMPT_RT */
532 #define cond_move_softirq_delta_start(osn_var, duration) do {} while (0)
533 #endif
534
535 /*
536  * cond_move_thread_delta_start - Forward the delta_start of a running thread
537  *
538  * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start
539  * is pushed forward to discount the interference.
540  *
541  * See get_int_safe_duration().
542  */
543 static inline void
544 cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration)
545 {
546         if (osn_var->thread.delta_start)
547                 osn_var->thread.delta_start += duration;
548 }
549
550 /*
551  * get_int_safe_duration - Get the duration of a window
552  *
553  * The irq, softirq and thread varaibles need to have its duration without
554  * the interference from higher priority interrupts. Instead of keeping a
555  * variable to discount the interrupt interference from these variables, the
556  * starting time of these variables are pushed forward with the interrupt's
557  * duration. In this way, a single variable is used to:
558  *
559  *   - Know if a given window is being measured.
560  *   - Account its duration.
561  *   - Discount the interference.
562  *
563  * To avoid getting inconsistent values, e.g.,:
564  *
565  *      now = time_get()
566  *              --->    interrupt!
567  *                      delta_start -= int duration;
568  *              <---
569  *      duration = now - delta_start;
570  *
571  *      result: negative duration if the variable duration before the
572  *      interrupt was smaller than the interrupt execution.
573  *
574  * A counter of interrupts is used. If the counter increased, try
575  * to capture an interference safe duration.
576  */
577 static inline s64
578 get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start)
579 {
580         u64 int_counter, now;
581         s64 duration;
582
583         do {
584                 int_counter = local_read(&osn_var->int_counter);
585                 /* synchronize with interrupts */
586                 barrier();
587
588                 now = time_get();
589                 duration = (now - *delta_start);
590
591                 /* synchronize with interrupts */
592                 barrier();
593         } while (int_counter != local_read(&osn_var->int_counter));
594
595         /*
596          * This is an evidence of race conditions that cause
597          * a value to be "discounted" too much.
598          */
599         if (duration < 0)
600                 osnoise_taint("Negative duration!\n");
601
602         *delta_start = 0;
603
604         return duration;
605 }
606
607 /*
608  *
609  * set_int_safe_time - Save the current time on *time, aware of interference
610  *
611  * Get the time, taking into consideration a possible interference from
612  * higher priority interrupts.
613  *
614  * See get_int_safe_duration() for an explanation.
615  */
616 static u64
617 set_int_safe_time(struct osnoise_variables *osn_var, u64 *time)
618 {
619         u64 int_counter;
620
621         do {
622                 int_counter = local_read(&osn_var->int_counter);
623                 /* synchronize with interrupts */
624                 barrier();
625
626                 *time = time_get();
627
628                 /* synchronize with interrupts */
629                 barrier();
630         } while (int_counter != local_read(&osn_var->int_counter));
631
632         return int_counter;
633 }
634
635 #ifdef CONFIG_TIMERLAT_TRACER
636 /*
637  * copy_int_safe_time - Copy *src into *desc aware of interference
638  */
639 static u64
640 copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src)
641 {
642         u64 int_counter;
643
644         do {
645                 int_counter = local_read(&osn_var->int_counter);
646                 /* synchronize with interrupts */
647                 barrier();
648
649                 *dst = *src;
650
651                 /* synchronize with interrupts */
652                 barrier();
653         } while (int_counter != local_read(&osn_var->int_counter));
654
655         return int_counter;
656 }
657 #endif /* CONFIG_TIMERLAT_TRACER */
658
659 /*
660  * trace_osnoise_callback - NMI entry/exit callback
661  *
662  * This function is called at the entry and exit NMI code. The bool enter
663  * distinguishes between either case. This function is used to note a NMI
664  * occurrence, compute the noise caused by the NMI, and to remove the noise
665  * it is potentially causing on other interference variables.
666  */
667 void trace_osnoise_callback(bool enter)
668 {
669         struct osnoise_variables *osn_var = this_cpu_osn_var();
670         u64 duration;
671
672         if (!osn_var->sampling)
673                 return;
674
675         /*
676          * Currently trace_clock_local() calls sched_clock() and the
677          * generic version is not NMI safe.
678          */
679         if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
680                 if (enter) {
681                         osn_var->nmi.delta_start = time_get();
682                         local_inc(&osn_var->int_counter);
683                 } else {
684                         duration = time_get() - osn_var->nmi.delta_start;
685
686                         trace_nmi_noise(osn_var->nmi.delta_start, duration);
687
688                         cond_move_irq_delta_start(osn_var, duration);
689                         cond_move_softirq_delta_start(osn_var, duration);
690                         cond_move_thread_delta_start(osn_var, duration);
691                 }
692         }
693
694         if (enter)
695                 osn_var->nmi.count++;
696 }
697
698 /*
699  * osnoise_trace_irq_entry - Note the starting of an IRQ
700  *
701  * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs,
702  * it is safe to use a single variable (ons_var->irq) to save the statistics.
703  * The arrival_time is used to report... the arrival time. The delta_start
704  * is used to compute the duration at the IRQ exit handler. See
705  * cond_move_irq_delta_start().
706  */
707 void osnoise_trace_irq_entry(int id)
708 {
709         struct osnoise_variables *osn_var = this_cpu_osn_var();
710
711         if (!osn_var->sampling)
712                 return;
713         /*
714          * This value will be used in the report, but not to compute
715          * the execution time, so it is safe to get it unsafe.
716          */
717         osn_var->irq.arrival_time = time_get();
718         set_int_safe_time(osn_var, &osn_var->irq.delta_start);
719         osn_var->irq.count++;
720
721         local_inc(&osn_var->int_counter);
722 }
723
724 /*
725  * osnoise_irq_exit - Note the end of an IRQ, sava data and trace
726  *
727  * Computes the duration of the IRQ noise, and trace it. Also discounts the
728  * interference from other sources of noise could be currently being accounted.
729  */
730 void osnoise_trace_irq_exit(int id, const char *desc)
731 {
732         struct osnoise_variables *osn_var = this_cpu_osn_var();
733         s64 duration;
734
735         if (!osn_var->sampling)
736                 return;
737
738         duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start);
739         trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration);
740         osn_var->irq.arrival_time = 0;
741         cond_move_softirq_delta_start(osn_var, duration);
742         cond_move_thread_delta_start(osn_var, duration);
743 }
744
745 /*
746  * trace_irqentry_callback - Callback to the irq:irq_entry traceevent
747  *
748  * Used to note the starting of an IRQ occurece.
749  */
750 static void trace_irqentry_callback(void *data, int irq,
751                                     struct irqaction *action)
752 {
753         osnoise_trace_irq_entry(irq);
754 }
755
756 /*
757  * trace_irqexit_callback - Callback to the irq:irq_exit traceevent
758  *
759  * Used to note the end of an IRQ occurece.
760  */
761 static void trace_irqexit_callback(void *data, int irq,
762                                    struct irqaction *action, int ret)
763 {
764         osnoise_trace_irq_exit(irq, action->name);
765 }
766
767 /*
768  * arch specific register function.
769  */
770 int __weak osnoise_arch_register(void)
771 {
772         return 0;
773 }
774
775 /*
776  * arch specific unregister function.
777  */
778 void __weak osnoise_arch_unregister(void)
779 {
780         return;
781 }
782
783 /*
784  * hook_irq_events - Hook IRQ handling events
785  *
786  * This function hooks the IRQ related callbacks to the respective trace
787  * events.
788  */
789 static int hook_irq_events(void)
790 {
791         int ret;
792
793         ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL);
794         if (ret)
795                 goto out_err;
796
797         ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL);
798         if (ret)
799                 goto out_unregister_entry;
800
801         ret = osnoise_arch_register();
802         if (ret)
803                 goto out_irq_exit;
804
805         return 0;
806
807 out_irq_exit:
808         unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
809 out_unregister_entry:
810         unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
811 out_err:
812         return -EINVAL;
813 }
814
815 /*
816  * unhook_irq_events - Unhook IRQ handling events
817  *
818  * This function unhooks the IRQ related callbacks to the respective trace
819  * events.
820  */
821 static void unhook_irq_events(void)
822 {
823         osnoise_arch_unregister();
824         unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
825         unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
826 }
827
828 #ifndef CONFIG_PREEMPT_RT
829 /*
830  * trace_softirq_entry_callback - Note the starting of a softirq
831  *
832  * Save the starting time of a softirq. As softirqs are non-preemptive to
833  * other softirqs, it is safe to use a single variable (ons_var->softirq)
834  * to save the statistics. The arrival_time is used to report... the
835  * arrival time. The delta_start is used to compute the duration at the
836  * softirq exit handler. See cond_move_softirq_delta_start().
837  */
838 static void trace_softirq_entry_callback(void *data, unsigned int vec_nr)
839 {
840         struct osnoise_variables *osn_var = this_cpu_osn_var();
841
842         if (!osn_var->sampling)
843                 return;
844         /*
845          * This value will be used in the report, but not to compute
846          * the execution time, so it is safe to get it unsafe.
847          */
848         osn_var->softirq.arrival_time = time_get();
849         set_int_safe_time(osn_var, &osn_var->softirq.delta_start);
850         osn_var->softirq.count++;
851
852         local_inc(&osn_var->int_counter);
853 }
854
855 /*
856  * trace_softirq_exit_callback - Note the end of an softirq
857  *
858  * Computes the duration of the softirq noise, and trace it. Also discounts the
859  * interference from other sources of noise could be currently being accounted.
860  */
861 static void trace_softirq_exit_callback(void *data, unsigned int vec_nr)
862 {
863         struct osnoise_variables *osn_var = this_cpu_osn_var();
864         s64 duration;
865
866         if (!osn_var->sampling)
867                 return;
868
869 #ifdef CONFIG_TIMERLAT_TRACER
870         /*
871          * If the timerlat is enabled, but the irq handler did
872          * not run yet enabling timerlat_tracer, do not trace.
873          */
874         if (unlikely(osnoise_data.timerlat_tracer)) {
875                 struct timerlat_variables *tlat_var;
876                 tlat_var = this_cpu_tmr_var();
877                 if (!tlat_var->tracing_thread) {
878                         osn_var->softirq.arrival_time = 0;
879                         osn_var->softirq.delta_start = 0;
880                         return;
881                 }
882         }
883 #endif
884
885         duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start);
886         trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration);
887         cond_move_thread_delta_start(osn_var, duration);
888         osn_var->softirq.arrival_time = 0;
889 }
890
891 /*
892  * hook_softirq_events - Hook softirq handling events
893  *
894  * This function hooks the softirq related callbacks to the respective trace
895  * events.
896  */
897 static int hook_softirq_events(void)
898 {
899         int ret;
900
901         ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL);
902         if (ret)
903                 goto out_err;
904
905         ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL);
906         if (ret)
907                 goto out_unreg_entry;
908
909         return 0;
910
911 out_unreg_entry:
912         unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
913 out_err:
914         return -EINVAL;
915 }
916
917 /*
918  * unhook_softirq_events - Unhook softirq handling events
919  *
920  * This function hooks the softirq related callbacks to the respective trace
921  * events.
922  */
923 static void unhook_softirq_events(void)
924 {
925         unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
926         unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL);
927 }
928 #else /* CONFIG_PREEMPT_RT */
929 /*
930  * softirq are threads on the PREEMPT_RT mode.
931  */
932 static int hook_softirq_events(void)
933 {
934         return 0;
935 }
936 static void unhook_softirq_events(void)
937 {
938 }
939 #endif
940
941 /*
942  * thread_entry - Record the starting of a thread noise window
943  *
944  * It saves the context switch time for a noisy thread, and increments
945  * the interference counters.
946  */
947 static void
948 thread_entry(struct osnoise_variables *osn_var, struct task_struct *t)
949 {
950         if (!osn_var->sampling)
951                 return;
952         /*
953          * The arrival time will be used in the report, but not to compute
954          * the execution time, so it is safe to get it unsafe.
955          */
956         osn_var->thread.arrival_time = time_get();
957
958         set_int_safe_time(osn_var, &osn_var->thread.delta_start);
959
960         osn_var->thread.count++;
961         local_inc(&osn_var->int_counter);
962 }
963
964 /*
965  * thread_exit - Report the end of a thread noise window
966  *
967  * It computes the total noise from a thread, tracing if needed.
968  */
969 static void
970 thread_exit(struct osnoise_variables *osn_var, struct task_struct *t)
971 {
972         s64 duration;
973
974         if (!osn_var->sampling)
975                 return;
976
977 #ifdef CONFIG_TIMERLAT_TRACER
978         if (osnoise_data.timerlat_tracer) {
979                 struct timerlat_variables *tlat_var;
980                 tlat_var = this_cpu_tmr_var();
981                 if (!tlat_var->tracing_thread) {
982                         osn_var->thread.delta_start = 0;
983                         osn_var->thread.arrival_time = 0;
984                         return;
985                 }
986         }
987 #endif
988
989         duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start);
990
991         trace_thread_noise(t, osn_var->thread.arrival_time, duration);
992
993         osn_var->thread.arrival_time = 0;
994 }
995
996 /*
997  * trace_sched_switch - sched:sched_switch trace event handler
998  *
999  * This function is hooked to the sched:sched_switch trace event, and it is
1000  * used to record the beginning and to report the end of a thread noise window.
1001  */
1002 static void
1003 trace_sched_switch_callback(void *data, bool preempt, struct task_struct *p,
1004                             struct task_struct *n)
1005 {
1006         struct osnoise_variables *osn_var = this_cpu_osn_var();
1007
1008         if (p->pid != osn_var->pid)
1009                 thread_exit(osn_var, p);
1010
1011         if (n->pid != osn_var->pid)
1012                 thread_entry(osn_var, n);
1013 }
1014
1015 /*
1016  * hook_thread_events - Hook the insturmentation for thread noise
1017  *
1018  * Hook the osnoise tracer callbacks to handle the noise from other
1019  * threads on the necessary kernel events.
1020  */
1021 static int hook_thread_events(void)
1022 {
1023         int ret;
1024
1025         ret = register_trace_sched_switch(trace_sched_switch_callback, NULL);
1026         if (ret)
1027                 return -EINVAL;
1028
1029         return 0;
1030 }
1031
1032 /*
1033  * unhook_thread_events - *nhook the insturmentation for thread noise
1034  *
1035  * Unook the osnoise tracer callbacks to handle the noise from other
1036  * threads on the necessary kernel events.
1037  */
1038 static void unhook_thread_events(void)
1039 {
1040         unregister_trace_sched_switch(trace_sched_switch_callback, NULL);
1041 }
1042
1043 /*
1044  * save_osn_sample_stats - Save the osnoise_sample statistics
1045  *
1046  * Save the osnoise_sample statistics before the sampling phase. These
1047  * values will be used later to compute the diff betwneen the statistics
1048  * before and after the osnoise sampling.
1049  */
1050 static void
1051 save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1052 {
1053         s->nmi_count = osn_var->nmi.count;
1054         s->irq_count = osn_var->irq.count;
1055         s->softirq_count = osn_var->softirq.count;
1056         s->thread_count = osn_var->thread.count;
1057 }
1058
1059 /*
1060  * diff_osn_sample_stats - Compute the osnoise_sample statistics
1061  *
1062  * After a sample period, compute the difference on the osnoise_sample
1063  * statistics. The struct osnoise_sample *s contains the statistics saved via
1064  * save_osn_sample_stats() before the osnoise sampling.
1065  */
1066 static void
1067 diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1068 {
1069         s->nmi_count = osn_var->nmi.count - s->nmi_count;
1070         s->irq_count = osn_var->irq.count - s->irq_count;
1071         s->softirq_count = osn_var->softirq.count - s->softirq_count;
1072         s->thread_count = osn_var->thread.count - s->thread_count;
1073 }
1074
1075 /*
1076  * osnoise_stop_tracing - Stop tracing and the tracer.
1077  */
1078 static __always_inline void osnoise_stop_tracing(void)
1079 {
1080         struct trace_array *tr = osnoise_trace;
1081
1082         trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_,
1083                         "stop tracing hit on cpu %d\n", smp_processor_id());
1084
1085         tracer_tracing_off(tr);
1086 }
1087
1088 /*
1089  * run_osnoise - Sample the time and look for osnoise
1090  *
1091  * Used to capture the time, looking for potential osnoise latency repeatedly.
1092  * Different from hwlat_detector, it is called with preemption and interrupts
1093  * enabled. This allows irqs, softirqs and threads to run, interfering on the
1094  * osnoise sampling thread, as they would do with a regular thread.
1095  */
1096 static int run_osnoise(void)
1097 {
1098         struct osnoise_variables *osn_var = this_cpu_osn_var();
1099         struct trace_array *tr = osnoise_trace;
1100         u64 start, sample, last_sample;
1101         u64 last_int_count, int_count;
1102         s64 noise = 0, max_noise = 0;
1103         s64 total, last_total = 0;
1104         struct osnoise_sample s;
1105         unsigned int threshold;
1106         u64 runtime, stop_in;
1107         u64 sum_noise = 0;
1108         int hw_count = 0;
1109         int ret = -1;
1110
1111         /*
1112          * Considers the current thread as the workload.
1113          */
1114         osn_var->pid = current->pid;
1115
1116         /*
1117          * Save the current stats for the diff
1118          */
1119         save_osn_sample_stats(osn_var, &s);
1120
1121         /*
1122          * if threshold is 0, use the default value of 5 us.
1123          */
1124         threshold = tracing_thresh ? : 5000;
1125
1126         /*
1127          * Make sure NMIs see sampling first
1128          */
1129         osn_var->sampling = true;
1130         barrier();
1131
1132         /*
1133          * Transform the *_us config to nanoseconds to avoid the
1134          * division on the main loop.
1135          */
1136         runtime = osnoise_data.sample_runtime * NSEC_PER_USEC;
1137         stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC;
1138
1139         /*
1140          * Start timestemp
1141          */
1142         start = time_get();
1143
1144         /*
1145          * "previous" loop.
1146          */
1147         last_int_count = set_int_safe_time(osn_var, &last_sample);
1148
1149         do {
1150                 /*
1151                  * Get sample!
1152                  */
1153                 int_count = set_int_safe_time(osn_var, &sample);
1154
1155                 noise = time_sub(sample, last_sample);
1156
1157                 /*
1158                  * This shouldn't happen.
1159                  */
1160                 if (noise < 0) {
1161                         osnoise_taint("negative noise!");
1162                         goto out;
1163                 }
1164
1165                 /*
1166                  * Sample runtime.
1167                  */
1168                 total = time_sub(sample, start);
1169
1170                 /*
1171                  * Check for possible overflows.
1172                  */
1173                 if (total < last_total) {
1174                         osnoise_taint("total overflow!");
1175                         break;
1176                 }
1177
1178                 last_total = total;
1179
1180                 if (noise >= threshold) {
1181                         int interference = int_count - last_int_count;
1182
1183                         if (noise > max_noise)
1184                                 max_noise = noise;
1185
1186                         if (!interference)
1187                                 hw_count++;
1188
1189                         sum_noise += noise;
1190
1191                         trace_sample_threshold(last_sample, noise, interference);
1192
1193                         if (osnoise_data.stop_tracing)
1194                                 if (noise > stop_in)
1195                                         osnoise_stop_tracing();
1196                 }
1197
1198                 /*
1199                  * In some cases, notably when running on a nohz_full CPU with
1200                  * a stopped tick PREEMPT_RCU has no way to account for QSs.
1201                  * This will eventually cause unwarranted noise as PREEMPT_RCU
1202                  * will force preemption as the means of ending the current
1203                  * grace period. We avoid this problem by calling
1204                  * rcu_momentary_dyntick_idle(), which performs a zero duration
1205                  * EQS allowing PREEMPT_RCU to end the current grace period.
1206                  * This call shouldn't be wrapped inside an RCU critical
1207                  * section.
1208                  *
1209                  * Note that in non PREEMPT_RCU kernels QSs are handled through
1210                  * cond_resched()
1211                  */
1212                 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
1213                         local_irq_disable();
1214                         rcu_momentary_dyntick_idle();
1215                         local_irq_enable();
1216                 }
1217
1218                 /*
1219                  * For the non-preemptive kernel config: let threads runs, if
1220                  * they so wish.
1221                  */
1222                 cond_resched();
1223
1224                 last_sample = sample;
1225                 last_int_count = int_count;
1226
1227         } while (total < runtime && !kthread_should_stop());
1228
1229         /*
1230          * Finish the above in the view for interrupts.
1231          */
1232         barrier();
1233
1234         osn_var->sampling = false;
1235
1236         /*
1237          * Make sure sampling data is no longer updated.
1238          */
1239         barrier();
1240
1241         /*
1242          * Save noise info.
1243          */
1244         s.noise = time_to_us(sum_noise);
1245         s.runtime = time_to_us(total);
1246         s.max_sample = time_to_us(max_noise);
1247         s.hw_count = hw_count;
1248
1249         /* Save interference stats info */
1250         diff_osn_sample_stats(osn_var, &s);
1251
1252         trace_osnoise_sample(&s);
1253
1254         /* Keep a running maximum ever recorded osnoise "latency" */
1255         if (max_noise > tr->max_latency) {
1256                 tr->max_latency = max_noise;
1257                 latency_fsnotify(tr);
1258         }
1259
1260         if (osnoise_data.stop_tracing_total)
1261                 if (s.noise > osnoise_data.stop_tracing_total)
1262                         osnoise_stop_tracing();
1263
1264         return 0;
1265 out:
1266         return ret;
1267 }
1268
1269 static struct cpumask osnoise_cpumask;
1270 static struct cpumask save_cpumask;
1271
1272 /*
1273  * osnoise_sleep - sleep until the next period
1274  */
1275 static void osnoise_sleep(void)
1276 {
1277         u64 interval;
1278         ktime_t wake_time;
1279
1280         mutex_lock(&interface_lock);
1281         interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
1282         mutex_unlock(&interface_lock);
1283
1284         /*
1285          * differently from hwlat_detector, the osnoise tracer can run
1286          * without a pause because preemption is on.
1287          */
1288         if (!interval) {
1289                 /* Let synchronize_rcu_tasks() make progress */
1290                 cond_resched_tasks_rcu_qs();
1291                 return;
1292         }
1293
1294         wake_time = ktime_add_us(ktime_get(), interval);
1295         __set_current_state(TASK_INTERRUPTIBLE);
1296
1297         while (schedule_hrtimeout_range(&wake_time, 0, HRTIMER_MODE_ABS)) {
1298                 if (kthread_should_stop())
1299                         break;
1300         }
1301 }
1302
1303 /*
1304  * osnoise_main - The osnoise detection kernel thread
1305  *
1306  * Calls run_osnoise() function to measure the osnoise for the configured runtime,
1307  * every period.
1308  */
1309 static int osnoise_main(void *data)
1310 {
1311
1312         while (!kthread_should_stop()) {
1313                 run_osnoise();
1314                 osnoise_sleep();
1315         }
1316
1317         return 0;
1318 }
1319
1320 #ifdef CONFIG_TIMERLAT_TRACER
1321 /*
1322  * timerlat_irq - hrtimer handler for timerlat.
1323  */
1324 static enum hrtimer_restart timerlat_irq(struct hrtimer *timer)
1325 {
1326         struct osnoise_variables *osn_var = this_cpu_osn_var();
1327         struct trace_array *tr = osnoise_trace;
1328         struct timerlat_variables *tlat;
1329         struct timerlat_sample s;
1330         u64 now;
1331         u64 diff;
1332
1333         /*
1334          * I am not sure if the timer was armed for this CPU. So, get
1335          * the timerlat struct from the timer itself, not from this
1336          * CPU.
1337          */
1338         tlat = container_of(timer, struct timerlat_variables, timer);
1339
1340         now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1341
1342         /*
1343          * Enable the osnoise: events for thread an softirq.
1344          */
1345         tlat->tracing_thread = true;
1346
1347         osn_var->thread.arrival_time = time_get();
1348
1349         /*
1350          * A hardirq is running: the timer IRQ. It is for sure preempting
1351          * a thread, and potentially preempting a softirq.
1352          *
1353          * At this point, it is not interesting to know the duration of the
1354          * preempted thread (and maybe softirq), but how much time they will
1355          * delay the beginning of the execution of the timer thread.
1356          *
1357          * To get the correct (net) delay added by the softirq, its delta_start
1358          * is set as the IRQ one. In this way, at the return of the IRQ, the delta
1359          * start of the sofitrq will be zeroed, accounting then only the time
1360          * after that.
1361          *
1362          * The thread follows the same principle. However, if a softirq is
1363          * running, the thread needs to receive the softirq delta_start. The
1364          * reason being is that the softirq will be the last to be unfolded,
1365          * resseting the thread delay to zero.
1366          */
1367 #ifndef CONFIG_PREEMPT_RT
1368         if (osn_var->softirq.delta_start) {
1369                 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1370                                    &osn_var->softirq.delta_start);
1371
1372                 copy_int_safe_time(osn_var, &osn_var->softirq.delta_start,
1373                                     &osn_var->irq.delta_start);
1374         } else {
1375                 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1376                                     &osn_var->irq.delta_start);
1377         }
1378 #else /* CONFIG_PREEMPT_RT */
1379         /*
1380          * The sofirqs run as threads on RT, so there is not need
1381          * to keep track of it.
1382          */
1383         copy_int_safe_time(osn_var, &osn_var->thread.delta_start, &osn_var->irq.delta_start);
1384 #endif /* CONFIG_PREEMPT_RT */
1385
1386         /*
1387          * Compute the current time with the expected time.
1388          */
1389         diff = now - tlat->abs_period;
1390
1391         tlat->count++;
1392         s.seqnum = tlat->count;
1393         s.timer_latency = diff;
1394         s.context = IRQ_CONTEXT;
1395
1396         trace_timerlat_sample(&s);
1397
1398         /* Keep a running maximum ever recorded os noise "latency" */
1399         if (diff > tr->max_latency) {
1400                 tr->max_latency = diff;
1401                 latency_fsnotify(tr);
1402         }
1403
1404         if (osnoise_data.stop_tracing)
1405                 if (time_to_us(diff) >= osnoise_data.stop_tracing)
1406                         osnoise_stop_tracing();
1407
1408         wake_up_process(tlat->kthread);
1409
1410         if (osnoise_data.print_stack)
1411                 timerlat_save_stack(0);
1412
1413         return HRTIMER_NORESTART;
1414 }
1415
1416 /*
1417  * wait_next_period - Wait for the next period for timerlat
1418  */
1419 static int wait_next_period(struct timerlat_variables *tlat)
1420 {
1421         ktime_t next_abs_period, now;
1422         u64 rel_period = osnoise_data.timerlat_period * 1000;
1423
1424         now = hrtimer_cb_get_time(&tlat->timer);
1425         next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1426
1427         /*
1428          * Save the next abs_period.
1429          */
1430         tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1431
1432         /*
1433          * If the new abs_period is in the past, skip the activation.
1434          */
1435         while (ktime_compare(now, next_abs_period) > 0) {
1436                 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1437                 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1438         }
1439
1440         set_current_state(TASK_INTERRUPTIBLE);
1441
1442         hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD);
1443         schedule();
1444         return 1;
1445 }
1446
1447 /*
1448  * timerlat_main- Timerlat main
1449  */
1450 static int timerlat_main(void *data)
1451 {
1452         struct osnoise_variables *osn_var = this_cpu_osn_var();
1453         struct timerlat_variables *tlat = this_cpu_tmr_var();
1454         struct timerlat_sample s;
1455         struct sched_param sp;
1456         u64 now, diff;
1457
1458         /*
1459          * Make the thread RT, that is how cyclictest is usually used.
1460          */
1461         sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
1462         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1463
1464         tlat->count = 0;
1465         tlat->tracing_thread = false;
1466
1467         hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1468         tlat->timer.function = timerlat_irq;
1469         tlat->kthread = current;
1470         osn_var->pid = current->pid;
1471         /*
1472          * Anotate the arrival time.
1473          */
1474         tlat->abs_period = hrtimer_cb_get_time(&tlat->timer);
1475
1476         wait_next_period(tlat);
1477
1478         osn_var->sampling = 1;
1479
1480         while (!kthread_should_stop()) {
1481                 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1482                 diff = now - tlat->abs_period;
1483
1484                 s.seqnum = tlat->count;
1485                 s.timer_latency = diff;
1486                 s.context = THREAD_CONTEXT;
1487
1488                 trace_timerlat_sample(&s);
1489
1490 #ifdef CONFIG_STACKTRACE
1491                 if (osnoise_data.print_stack)
1492                         if (osnoise_data.print_stack <= time_to_us(diff))
1493                                 timerlat_dump_stack();
1494 #endif /* CONFIG_STACKTRACE */
1495
1496                 tlat->tracing_thread = false;
1497                 if (osnoise_data.stop_tracing_total)
1498                         if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
1499                                 osnoise_stop_tracing();
1500
1501                 wait_next_period(tlat);
1502         }
1503
1504         hrtimer_cancel(&tlat->timer);
1505         return 0;
1506 }
1507 #endif /* CONFIG_TIMERLAT_TRACER */
1508
1509 /*
1510  * stop_kthread - stop a workload thread
1511  */
1512 static void stop_kthread(unsigned int cpu)
1513 {
1514         struct task_struct *kthread;
1515
1516         kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread;
1517         if (kthread)
1518                 kthread_stop(kthread);
1519         per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1520 }
1521
1522 /*
1523  * stop_per_cpu_kthread - Stop per-cpu threads
1524  *
1525  * Stop the osnoise sampling htread. Use this on unload and at system
1526  * shutdown.
1527  */
1528 static void stop_per_cpu_kthreads(void)
1529 {
1530         int cpu;
1531
1532         cpus_read_lock();
1533
1534         for_each_online_cpu(cpu)
1535                 stop_kthread(cpu);
1536
1537         cpus_read_unlock();
1538 }
1539
1540 /*
1541  * start_kthread - Start a workload tread
1542  */
1543 static int start_kthread(unsigned int cpu)
1544 {
1545         struct task_struct *kthread;
1546         void *main = osnoise_main;
1547         char comm[24];
1548
1549 #ifdef CONFIG_TIMERLAT_TRACER
1550         if (osnoise_data.timerlat_tracer) {
1551                 snprintf(comm, 24, "timerlat/%d", cpu);
1552                 main = timerlat_main;
1553         } else {
1554                 snprintf(comm, 24, "osnoise/%d", cpu);
1555         }
1556 #else
1557         snprintf(comm, 24, "osnoise/%d", cpu);
1558 #endif
1559         kthread = kthread_create_on_cpu(main, NULL, cpu, comm);
1560
1561         if (IS_ERR(kthread)) {
1562                 pr_err(BANNER "could not start sampling thread\n");
1563                 stop_per_cpu_kthreads();
1564                 return -ENOMEM;
1565         }
1566
1567         per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread;
1568         wake_up_process(kthread);
1569
1570         return 0;
1571 }
1572
1573 /*
1574  * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads
1575  *
1576  * This starts the kernel thread that will look for osnoise on many
1577  * cpus.
1578  */
1579 static int start_per_cpu_kthreads(struct trace_array *tr)
1580 {
1581         struct cpumask *current_mask = &save_cpumask;
1582         int retval = 0;
1583         int cpu;
1584
1585         cpus_read_lock();
1586         /*
1587          * Run only on CPUs in which trace and osnoise are allowed to run.
1588          */
1589         cpumask_and(current_mask, tr->tracing_cpumask, &osnoise_cpumask);
1590         /*
1591          * And the CPU is online.
1592          */
1593         cpumask_and(current_mask, cpu_online_mask, current_mask);
1594
1595         for_each_possible_cpu(cpu)
1596                 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1597
1598         for_each_cpu(cpu, current_mask) {
1599                 retval = start_kthread(cpu);
1600                 if (retval) {
1601                         cpus_read_unlock();
1602                         stop_per_cpu_kthreads();
1603                         return retval;
1604                 }
1605         }
1606
1607         cpus_read_unlock();
1608
1609         return retval;
1610 }
1611
1612 #ifdef CONFIG_HOTPLUG_CPU
1613 static void osnoise_hotplug_workfn(struct work_struct *dummy)
1614 {
1615         struct trace_array *tr = osnoise_trace;
1616         unsigned int cpu = smp_processor_id();
1617
1618
1619         mutex_lock(&trace_types_lock);
1620
1621         if (!osnoise_busy)
1622                 goto out_unlock_trace;
1623
1624         mutex_lock(&interface_lock);
1625         cpus_read_lock();
1626
1627         if (!cpumask_test_cpu(cpu, &osnoise_cpumask))
1628                 goto out_unlock;
1629
1630         if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
1631                 goto out_unlock;
1632
1633         start_kthread(cpu);
1634
1635 out_unlock:
1636         cpus_read_unlock();
1637         mutex_unlock(&interface_lock);
1638 out_unlock_trace:
1639         mutex_unlock(&trace_types_lock);
1640 }
1641
1642 static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn);
1643
1644 /*
1645  * osnoise_cpu_init - CPU hotplug online callback function
1646  */
1647 static int osnoise_cpu_init(unsigned int cpu)
1648 {
1649         schedule_work_on(cpu, &osnoise_hotplug_work);
1650         return 0;
1651 }
1652
1653 /*
1654  * osnoise_cpu_die - CPU hotplug offline callback function
1655  */
1656 static int osnoise_cpu_die(unsigned int cpu)
1657 {
1658         stop_kthread(cpu);
1659         return 0;
1660 }
1661
1662 static void osnoise_init_hotplug_support(void)
1663 {
1664         int ret;
1665
1666         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online",
1667                                 osnoise_cpu_init, osnoise_cpu_die);
1668         if (ret < 0)
1669                 pr_warn(BANNER "Error to init cpu hotplug support\n");
1670
1671         return;
1672 }
1673 #else /* CONFIG_HOTPLUG_CPU */
1674 static void osnoise_init_hotplug_support(void)
1675 {
1676         return;
1677 }
1678 #endif /* CONFIG_HOTPLUG_CPU */
1679
1680 /*
1681  * osnoise_cpus_read - Read function for reading the "cpus" file
1682  * @filp: The active open file structure
1683  * @ubuf: The userspace provided buffer to read value into
1684  * @cnt: The maximum number of bytes to read
1685  * @ppos: The current "file" position
1686  *
1687  * Prints the "cpus" output into the user-provided buffer.
1688  */
1689 static ssize_t
1690 osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count,
1691                   loff_t *ppos)
1692 {
1693         char *mask_str;
1694         int len;
1695
1696         mutex_lock(&interface_lock);
1697
1698         len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1;
1699         mask_str = kmalloc(len, GFP_KERNEL);
1700         if (!mask_str) {
1701                 count = -ENOMEM;
1702                 goto out_unlock;
1703         }
1704
1705         len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask));
1706         if (len >= count) {
1707                 count = -EINVAL;
1708                 goto out_free;
1709         }
1710
1711         count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
1712
1713 out_free:
1714         kfree(mask_str);
1715 out_unlock:
1716         mutex_unlock(&interface_lock);
1717
1718         return count;
1719 }
1720
1721 static void osnoise_tracer_start(struct trace_array *tr);
1722 static void osnoise_tracer_stop(struct trace_array *tr);
1723
1724 /*
1725  * osnoise_cpus_write - Write function for "cpus" entry
1726  * @filp: The active open file structure
1727  * @ubuf: The user buffer that contains the value to write
1728  * @cnt: The maximum number of bytes to write to "file"
1729  * @ppos: The current position in @file
1730  *
1731  * This function provides a write implementation for the "cpus"
1732  * interface to the osnoise trace. By default, it lists all  CPUs,
1733  * in this way, allowing osnoise threads to run on any online CPU
1734  * of the system. It serves to restrict the execution of osnoise to the
1735  * set of CPUs writing via this interface. Note that osnoise also
1736  * respects the "tracing_cpumask." Hence, osnoise threads will run only
1737  * on the set of CPUs allowed here AND on "tracing_cpumask." Why not
1738  * have just "tracing_cpumask?" Because the user might be interested
1739  * in tracing what is running on other CPUs. For instance, one might
1740  * run osnoise in one HT CPU while observing what is running on the
1741  * sibling HT CPU.
1742  */
1743 static ssize_t
1744 osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
1745                    loff_t *ppos)
1746 {
1747         struct trace_array *tr = osnoise_trace;
1748         cpumask_var_t osnoise_cpumask_new;
1749         int running, err;
1750         char buf[256];
1751
1752         if (count >= 256)
1753                 return -EINVAL;
1754
1755         if (copy_from_user(buf, ubuf, count))
1756                 return -EFAULT;
1757
1758         if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL))
1759                 return -ENOMEM;
1760
1761         err = cpulist_parse(buf, osnoise_cpumask_new);
1762         if (err)
1763                 goto err_free;
1764
1765         /*
1766          * trace_types_lock is taken to avoid concurrency on start/stop
1767          * and osnoise_busy.
1768          */
1769         mutex_lock(&trace_types_lock);
1770         running = osnoise_busy;
1771         if (running)
1772                 osnoise_tracer_stop(tr);
1773
1774         mutex_lock(&interface_lock);
1775         /*
1776          * osnoise_cpumask is read by CPU hotplug operations.
1777          */
1778         cpus_read_lock();
1779
1780         cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
1781
1782         cpus_read_unlock();
1783         mutex_unlock(&interface_lock);
1784
1785         if (running)
1786                 osnoise_tracer_start(tr);
1787         mutex_unlock(&trace_types_lock);
1788
1789         free_cpumask_var(osnoise_cpumask_new);
1790         return count;
1791
1792 err_free:
1793         free_cpumask_var(osnoise_cpumask_new);
1794
1795         return err;
1796 }
1797
1798 /*
1799  * osnoise/runtime_us: cannot be greater than the period.
1800  */
1801 static struct trace_min_max_param osnoise_runtime = {
1802         .lock   = &interface_lock,
1803         .val    = &osnoise_data.sample_runtime,
1804         .max    = &osnoise_data.sample_period,
1805         .min    = NULL,
1806 };
1807
1808 /*
1809  * osnoise/period_us: cannot be smaller than the runtime.
1810  */
1811 static struct trace_min_max_param osnoise_period = {
1812         .lock   = &interface_lock,
1813         .val    = &osnoise_data.sample_period,
1814         .max    = NULL,
1815         .min    = &osnoise_data.sample_runtime,
1816 };
1817
1818 /*
1819  * osnoise/stop_tracing_us: no limit.
1820  */
1821 static struct trace_min_max_param osnoise_stop_tracing_in = {
1822         .lock   = &interface_lock,
1823         .val    = &osnoise_data.stop_tracing,
1824         .max    = NULL,
1825         .min    = NULL,
1826 };
1827
1828 /*
1829  * osnoise/stop_tracing_total_us: no limit.
1830  */
1831 static struct trace_min_max_param osnoise_stop_tracing_total = {
1832         .lock   = &interface_lock,
1833         .val    = &osnoise_data.stop_tracing_total,
1834         .max    = NULL,
1835         .min    = NULL,
1836 };
1837
1838 #ifdef CONFIG_TIMERLAT_TRACER
1839 /*
1840  * osnoise/print_stack: print the stacktrace of the IRQ handler if the total
1841  * latency is higher than val.
1842  */
1843 static struct trace_min_max_param osnoise_print_stack = {
1844         .lock   = &interface_lock,
1845         .val    = &osnoise_data.print_stack,
1846         .max    = NULL,
1847         .min    = NULL,
1848 };
1849
1850 /*
1851  * osnoise/timerlat_period: min 100 us, max 1 s
1852  */
1853 u64 timerlat_min_period = 100;
1854 u64 timerlat_max_period = 1000000;
1855 static struct trace_min_max_param timerlat_period = {
1856         .lock   = &interface_lock,
1857         .val    = &osnoise_data.timerlat_period,
1858         .max    = &timerlat_max_period,
1859         .min    = &timerlat_min_period,
1860 };
1861 #endif
1862
1863 static const struct file_operations cpus_fops = {
1864         .open           = tracing_open_generic,
1865         .read           = osnoise_cpus_read,
1866         .write          = osnoise_cpus_write,
1867         .llseek         = generic_file_llseek,
1868 };
1869
1870 /*
1871  * init_tracefs - A function to initialize the tracefs interface files
1872  *
1873  * This function creates entries in tracefs for "osnoise" and "timerlat".
1874  * It creates these directories in the tracing directory, and within that
1875  * directory the use can change and view the configs.
1876  */
1877 static int init_tracefs(void)
1878 {
1879         struct dentry *top_dir;
1880         struct dentry *tmp;
1881         int ret;
1882
1883         ret = tracing_init_dentry();
1884         if (ret)
1885                 return -ENOMEM;
1886
1887         top_dir = tracefs_create_dir("osnoise", NULL);
1888         if (!top_dir)
1889                 return 0;
1890
1891         tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir,
1892                                   &osnoise_period, &trace_min_max_fops);
1893         if (!tmp)
1894                 goto err;
1895
1896         tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir,
1897                                   &osnoise_runtime, &trace_min_max_fops);
1898         if (!tmp)
1899                 goto err;
1900
1901         tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir,
1902                                   &osnoise_stop_tracing_in, &trace_min_max_fops);
1903         if (!tmp)
1904                 goto err;
1905
1906         tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir,
1907                                   &osnoise_stop_tracing_total, &trace_min_max_fops);
1908         if (!tmp)
1909                 goto err;
1910
1911         tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops);
1912         if (!tmp)
1913                 goto err;
1914 #ifdef CONFIG_TIMERLAT_TRACER
1915 #ifdef CONFIG_STACKTRACE
1916         tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir,
1917                                   &osnoise_print_stack, &trace_min_max_fops);
1918         if (!tmp)
1919                 goto err;
1920 #endif
1921
1922         tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir,
1923                                   &timerlat_period, &trace_min_max_fops);
1924         if (!tmp)
1925                 goto err;
1926 #endif
1927
1928         return 0;
1929
1930 err:
1931         tracefs_remove(top_dir);
1932         return -ENOMEM;
1933 }
1934
1935 static int osnoise_hook_events(void)
1936 {
1937         int retval;
1938
1939         /*
1940          * Trace is already hooked, we are re-enabling from
1941          * a stop_tracing_*.
1942          */
1943         if (trace_osnoise_callback_enabled)
1944                 return 0;
1945
1946         retval = hook_irq_events();
1947         if (retval)
1948                 return -EINVAL;
1949
1950         retval = hook_softirq_events();
1951         if (retval)
1952                 goto out_unhook_irq;
1953
1954         retval = hook_thread_events();
1955         /*
1956          * All fine!
1957          */
1958         if (!retval)
1959                 return 0;
1960
1961         unhook_softirq_events();
1962 out_unhook_irq:
1963         unhook_irq_events();
1964         return -EINVAL;
1965 }
1966
1967 static void osnoise_unhook_events(void)
1968 {
1969         unhook_thread_events();
1970         unhook_softirq_events();
1971         unhook_irq_events();
1972 }
1973
1974 static int __osnoise_tracer_start(struct trace_array *tr)
1975 {
1976         int retval;
1977
1978         osn_var_reset_all();
1979
1980         retval = osnoise_hook_events();
1981         if (retval)
1982                 return retval;
1983         /*
1984          * Make sure NMIs see reseted values.
1985          */
1986         barrier();
1987         trace_osnoise_callback_enabled = true;
1988
1989         retval = start_per_cpu_kthreads(tr);
1990         if (retval) {
1991                 trace_osnoise_callback_enabled = false;
1992                 /*
1993                  * Make sure that ftrace_nmi_enter/exit() see
1994                  * trace_osnoise_callback_enabled as false before continuing.
1995                  */
1996                 barrier();
1997
1998                 osnoise_unhook_events();
1999                 return retval;
2000         }
2001
2002         osnoise_busy = true;
2003
2004         return 0;
2005 }
2006
2007 static void osnoise_tracer_start(struct trace_array *tr)
2008 {
2009         int retval;
2010
2011         if (osnoise_busy)
2012                 return;
2013
2014         retval = __osnoise_tracer_start(tr);
2015         if (retval)
2016                 pr_err(BANNER "Error starting osnoise tracer\n");
2017
2018 }
2019
2020 static void osnoise_tracer_stop(struct trace_array *tr)
2021 {
2022         if (!osnoise_busy)
2023                 return;
2024
2025         trace_osnoise_callback_enabled = false;
2026         barrier();
2027
2028         stop_per_cpu_kthreads();
2029
2030         osnoise_unhook_events();
2031
2032         osnoise_busy = false;
2033 }
2034
2035 static int osnoise_tracer_init(struct trace_array *tr)
2036 {
2037
2038         /* Only allow one instance to enable this */
2039         if (osnoise_busy)
2040                 return -EBUSY;
2041
2042         osnoise_trace = tr;
2043         tr->max_latency = 0;
2044
2045         osnoise_tracer_start(tr);
2046
2047         return 0;
2048 }
2049
2050 static void osnoise_tracer_reset(struct trace_array *tr)
2051 {
2052         osnoise_tracer_stop(tr);
2053 }
2054
2055 static struct tracer osnoise_tracer __read_mostly = {
2056         .name           = "osnoise",
2057         .init           = osnoise_tracer_init,
2058         .reset          = osnoise_tracer_reset,
2059         .start          = osnoise_tracer_start,
2060         .stop           = osnoise_tracer_stop,
2061         .print_header   = print_osnoise_headers,
2062         .allow_instances = true,
2063 };
2064
2065 #ifdef CONFIG_TIMERLAT_TRACER
2066 static void timerlat_tracer_start(struct trace_array *tr)
2067 {
2068         int retval;
2069
2070         if (osnoise_busy)
2071                 return;
2072
2073         osnoise_data.timerlat_tracer = 1;
2074
2075         retval = __osnoise_tracer_start(tr);
2076         if (retval)
2077                 goto out_err;
2078
2079         return;
2080 out_err:
2081         pr_err(BANNER "Error starting timerlat tracer\n");
2082 }
2083
2084 static void timerlat_tracer_stop(struct trace_array *tr)
2085 {
2086         int cpu;
2087
2088         if (!osnoise_busy)
2089                 return;
2090
2091         for_each_online_cpu(cpu)
2092                 per_cpu(per_cpu_osnoise_var, cpu).sampling = 0;
2093
2094         osnoise_tracer_stop(tr);
2095
2096         osnoise_data.timerlat_tracer = 0;
2097 }
2098
2099 static int timerlat_tracer_init(struct trace_array *tr)
2100 {
2101         /* Only allow one instance to enable this */
2102         if (osnoise_busy)
2103                 return -EBUSY;
2104
2105         osnoise_trace = tr;
2106
2107         tr->max_latency = 0;
2108
2109         timerlat_tracer_start(tr);
2110
2111         return 0;
2112 }
2113
2114 static void timerlat_tracer_reset(struct trace_array *tr)
2115 {
2116         timerlat_tracer_stop(tr);
2117 }
2118
2119 static struct tracer timerlat_tracer __read_mostly = {
2120         .name           = "timerlat",
2121         .init           = timerlat_tracer_init,
2122         .reset          = timerlat_tracer_reset,
2123         .start          = timerlat_tracer_start,
2124         .stop           = timerlat_tracer_stop,
2125         .print_header   = print_timerlat_headers,
2126         .allow_instances = true,
2127 };
2128 #endif /* CONFIG_TIMERLAT_TRACER */
2129
2130 __init static int init_osnoise_tracer(void)
2131 {
2132         int ret;
2133
2134         mutex_init(&interface_lock);
2135
2136         cpumask_copy(&osnoise_cpumask, cpu_all_mask);
2137
2138         ret = register_tracer(&osnoise_tracer);
2139         if (ret) {
2140                 pr_err(BANNER "Error registering osnoise!\n");
2141                 return ret;
2142         }
2143
2144 #ifdef CONFIG_TIMERLAT_TRACER
2145         ret = register_tracer(&timerlat_tracer);
2146         if (ret) {
2147                 pr_err(BANNER "Error registering timerlat\n");
2148                 return ret;
2149         }
2150 #endif
2151         osnoise_init_hotplug_support();
2152
2153         init_tracefs();
2154
2155         return 0;
2156 }
2157 late_initcall(init_osnoise_tracer);