OSDN Git Service

tracing/osnoise: Force quiescent states while tracing
[tomoyo/tomoyo-test1.git] / kernel / trace / 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 /*
42  * Default values.
43  */
44 #define BANNER                  "osnoise: "
45 #define DEFAULT_SAMPLE_PERIOD   1000000                 /* 1s */
46 #define DEFAULT_SAMPLE_RUNTIME  1000000                 /* 1s */
47
48 #define DEFAULT_TIMERLAT_PERIOD 1000                    /* 1ms */
49 #define DEFAULT_TIMERLAT_PRIO   95                      /* FIFO 95 */
50
51 /*
52  * trace_array of the enabled osnoise/timerlat instances.
53  */
54 struct osnoise_instance {
55         struct list_head        list;
56         struct trace_array      *tr;
57 };
58
59 static struct list_head osnoise_instances;
60
61 static bool osnoise_has_registered_instances(void)
62 {
63         return !!list_first_or_null_rcu(&osnoise_instances,
64                                         struct osnoise_instance,
65                                         list);
66 }
67
68 /*
69  * osnoise_instance_registered - check if a tr is already registered
70  */
71 static int osnoise_instance_registered(struct trace_array *tr)
72 {
73         struct osnoise_instance *inst;
74         int found = 0;
75
76         rcu_read_lock();
77         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
78                 if (inst->tr == tr)
79                         found = 1;
80         }
81         rcu_read_unlock();
82
83         return found;
84 }
85
86 /*
87  * osnoise_register_instance - register a new trace instance
88  *
89  * Register a trace_array *tr in the list of instances running
90  * osnoise/timerlat tracers.
91  */
92 static int osnoise_register_instance(struct trace_array *tr)
93 {
94         struct osnoise_instance *inst;
95
96         /*
97          * register/unregister serialization is provided by trace's
98          * trace_types_lock.
99          */
100         lockdep_assert_held(&trace_types_lock);
101
102         inst = kmalloc(sizeof(*inst), GFP_KERNEL);
103         if (!inst)
104                 return -ENOMEM;
105
106         INIT_LIST_HEAD_RCU(&inst->list);
107         inst->tr = tr;
108         list_add_tail_rcu(&inst->list, &osnoise_instances);
109
110         return 0;
111 }
112
113 /*
114  *  osnoise_unregister_instance - unregister a registered trace instance
115  *
116  * Remove the trace_array *tr from the list of instances running
117  * osnoise/timerlat tracers.
118  */
119 static void osnoise_unregister_instance(struct trace_array *tr)
120 {
121         struct osnoise_instance *inst;
122         int found = 0;
123
124         /*
125          * register/unregister serialization is provided by trace's
126          * trace_types_lock.
127          */
128         lockdep_assert_held(&trace_types_lock);
129
130         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
131                 if (inst->tr == tr) {
132                         list_del_rcu(&inst->list);
133                         found = 1;
134                         break;
135                 }
136         }
137
138         if (!found)
139                 return;
140
141         kvfree_rcu(inst);
142 }
143
144 /*
145  * NMI runtime info.
146  */
147 struct osn_nmi {
148         u64     count;
149         u64     delta_start;
150 };
151
152 /*
153  * IRQ runtime info.
154  */
155 struct osn_irq {
156         u64     count;
157         u64     arrival_time;
158         u64     delta_start;
159 };
160
161 #define IRQ_CONTEXT     0
162 #define THREAD_CONTEXT  1
163 /*
164  * sofirq runtime info.
165  */
166 struct osn_softirq {
167         u64     count;
168         u64     arrival_time;
169         u64     delta_start;
170 };
171
172 /*
173  * thread runtime info.
174  */
175 struct osn_thread {
176         u64     count;
177         u64     arrival_time;
178         u64     delta_start;
179 };
180
181 /*
182  * Runtime information: this structure saves the runtime information used by
183  * one sampling thread.
184  */
185 struct osnoise_variables {
186         struct task_struct      *kthread;
187         bool                    sampling;
188         pid_t                   pid;
189         struct osn_nmi          nmi;
190         struct osn_irq          irq;
191         struct osn_softirq      softirq;
192         struct osn_thread       thread;
193         local_t                 int_counter;
194 };
195
196 /*
197  * Per-cpu runtime information.
198  */
199 DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var);
200
201 /*
202  * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU
203  */
204 static inline struct osnoise_variables *this_cpu_osn_var(void)
205 {
206         return this_cpu_ptr(&per_cpu_osnoise_var);
207 }
208
209 #ifdef CONFIG_TIMERLAT_TRACER
210 /*
211  * Runtime information for the timer mode.
212  */
213 struct timerlat_variables {
214         struct task_struct      *kthread;
215         struct hrtimer          timer;
216         u64                     rel_period;
217         u64                     abs_period;
218         bool                    tracing_thread;
219         u64                     count;
220 };
221
222 DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var);
223
224 /*
225  * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU
226  */
227 static inline struct timerlat_variables *this_cpu_tmr_var(void)
228 {
229         return this_cpu_ptr(&per_cpu_timerlat_var);
230 }
231
232 /*
233  * tlat_var_reset - Reset the values of the given timerlat_variables
234  */
235 static inline void tlat_var_reset(void)
236 {
237         struct timerlat_variables *tlat_var;
238         int cpu;
239         /*
240          * So far, all the values are initialized as 0, so
241          * zeroing the structure is perfect.
242          */
243         for_each_cpu(cpu, cpu_online_mask) {
244                 tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu);
245                 memset(tlat_var, 0, sizeof(*tlat_var));
246         }
247 }
248 #else /* CONFIG_TIMERLAT_TRACER */
249 #define tlat_var_reset()        do {} while (0)
250 #endif /* CONFIG_TIMERLAT_TRACER */
251
252 /*
253  * osn_var_reset - Reset the values of the given osnoise_variables
254  */
255 static inline void osn_var_reset(void)
256 {
257         struct osnoise_variables *osn_var;
258         int cpu;
259
260         /*
261          * So far, all the values are initialized as 0, so
262          * zeroing the structure is perfect.
263          */
264         for_each_cpu(cpu, cpu_online_mask) {
265                 osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu);
266                 memset(osn_var, 0, sizeof(*osn_var));
267         }
268 }
269
270 /*
271  * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables
272  */
273 static inline void osn_var_reset_all(void)
274 {
275         osn_var_reset();
276         tlat_var_reset();
277 }
278
279 /*
280  * Tells NMIs to call back to the osnoise tracer to record timestamps.
281  */
282 bool trace_osnoise_callback_enabled;
283
284 /*
285  * osnoise sample structure definition. Used to store the statistics of a
286  * sample run.
287  */
288 struct osnoise_sample {
289         u64                     runtime;        /* runtime */
290         u64                     noise;          /* noise */
291         u64                     max_sample;     /* max single noise sample */
292         int                     hw_count;       /* # HW (incl. hypervisor) interference */
293         int                     nmi_count;      /* # NMIs during this sample */
294         int                     irq_count;      /* # IRQs during this sample */
295         int                     softirq_count;  /* # softirqs during this sample */
296         int                     thread_count;   /* # threads during this sample */
297 };
298
299 #ifdef CONFIG_TIMERLAT_TRACER
300 /*
301  * timerlat sample structure definition. Used to store the statistics of
302  * a sample run.
303  */
304 struct timerlat_sample {
305         u64                     timer_latency;  /* timer_latency */
306         unsigned int            seqnum;         /* unique sequence */
307         int                     context;        /* timer context */
308 };
309 #endif
310
311 /*
312  * Protect the interface.
313  */
314 struct mutex interface_lock;
315
316 /*
317  * Tracer data.
318  */
319 static struct osnoise_data {
320         u64     sample_period;          /* total sampling period */
321         u64     sample_runtime;         /* active sampling portion of period */
322         u64     stop_tracing;           /* stop trace in the internal operation (loop/irq) */
323         u64     stop_tracing_total;     /* stop trace in the final operation (report/thread) */
324 #ifdef CONFIG_TIMERLAT_TRACER
325         u64     timerlat_period;        /* timerlat period */
326         u64     print_stack;            /* print IRQ stack if total > */
327         int     timerlat_tracer;        /* timerlat tracer */
328 #endif
329         bool    tainted;                /* infor users and developers about a problem */
330 } osnoise_data = {
331         .sample_period                  = DEFAULT_SAMPLE_PERIOD,
332         .sample_runtime                 = DEFAULT_SAMPLE_RUNTIME,
333         .stop_tracing                   = 0,
334         .stop_tracing_total             = 0,
335 #ifdef CONFIG_TIMERLAT_TRACER
336         .print_stack                    = 0,
337         .timerlat_period                = DEFAULT_TIMERLAT_PERIOD,
338         .timerlat_tracer                = 0,
339 #endif
340 };
341
342 #ifdef CONFIG_TIMERLAT_TRACER
343 static inline bool timerlat_enabled(void)
344 {
345         return osnoise_data.timerlat_tracer;
346 }
347
348 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
349 {
350         struct timerlat_variables *tlat_var = this_cpu_tmr_var();
351         /*
352          * If the timerlat is enabled, but the irq handler did
353          * not run yet enabling timerlat_tracer, do not trace.
354          */
355         if (!tlat_var->tracing_thread) {
356                 osn_var->softirq.arrival_time = 0;
357                 osn_var->softirq.delta_start = 0;
358                 return 0;
359         }
360         return 1;
361 }
362
363 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
364 {
365         struct timerlat_variables *tlat_var = this_cpu_tmr_var();
366         /*
367          * If the timerlat is enabled, but the irq handler did
368          * not run yet enabling timerlat_tracer, do not trace.
369          */
370         if (!tlat_var->tracing_thread) {
371                 osn_var->thread.delta_start = 0;
372                 osn_var->thread.arrival_time = 0;
373                 return 0;
374         }
375         return 1;
376 }
377 #else /* CONFIG_TIMERLAT_TRACER */
378 static inline bool timerlat_enabled(void)
379 {
380         return false;
381 }
382
383 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
384 {
385         return 1;
386 }
387 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
388 {
389         return 1;
390 }
391 #endif
392
393 #ifdef CONFIG_PREEMPT_RT
394 /*
395  * Print the osnoise header info.
396  */
397 static void print_osnoise_headers(struct seq_file *s)
398 {
399         if (osnoise_data.tainted)
400                 seq_puts(s, "# osnoise is tainted!\n");
401
402         seq_puts(s, "#                                _-------=> irqs-off\n");
403         seq_puts(s, "#                               / _------=> need-resched\n");
404         seq_puts(s, "#                              | / _-----=> need-resched-lazy\n");
405         seq_puts(s, "#                              || / _----=> hardirq/softirq\n");
406         seq_puts(s, "#                              ||| / _---=> preempt-depth\n");
407         seq_puts(s, "#                              |||| / _--=> preempt-lazy-depth\n");
408         seq_puts(s, "#                              ||||| / _-=> migrate-disable\n");
409
410         seq_puts(s, "#                              |||||| /          ");
411         seq_puts(s, "                                     MAX\n");
412
413         seq_puts(s, "#                              ||||| /                         ");
414         seq_puts(s, "                    SINGLE      Interference counters:\n");
415
416         seq_puts(s, "#                              |||||||               RUNTIME   ");
417         seq_puts(s, "   NOISE  %% OF CPU  NOISE    +-----------------------------+\n");
418
419         seq_puts(s, "#           TASK-PID      CPU# |||||||   TIMESTAMP    IN US    ");
420         seq_puts(s, "   IN US  AVAILABLE  IN US     HW    NMI    IRQ   SIRQ THREAD\n");
421
422         seq_puts(s, "#              | |         |   |||||||      |           |      ");
423         seq_puts(s, "       |    |            |      |      |      |      |      |\n");
424 }
425 #else /* CONFIG_PREEMPT_RT */
426 static void print_osnoise_headers(struct seq_file *s)
427 {
428         if (osnoise_data.tainted)
429                 seq_puts(s, "# osnoise is tainted!\n");
430
431         seq_puts(s, "#                                _-----=> irqs-off\n");
432         seq_puts(s, "#                               / _----=> need-resched\n");
433         seq_puts(s, "#                              | / _---=> hardirq/softirq\n");
434         seq_puts(s, "#                              || / _--=> preempt-depth\n");
435         seq_puts(s, "#                              ||| / _-=> migrate-disable     ");
436         seq_puts(s, "                    MAX\n");
437         seq_puts(s, "#                              |||| /     delay               ");
438         seq_puts(s, "                    SINGLE      Interference counters:\n");
439
440         seq_puts(s, "#                              |||||               RUNTIME   ");
441         seq_puts(s, "   NOISE  %% OF CPU  NOISE    +-----------------------------+\n");
442
443         seq_puts(s, "#           TASK-PID      CPU# |||||   TIMESTAMP    IN US    ");
444         seq_puts(s, "   IN US  AVAILABLE  IN US     HW    NMI    IRQ   SIRQ THREAD\n");
445
446         seq_puts(s, "#              | |         |   |||||      |           |      ");
447         seq_puts(s, "       |    |            |      |      |      |      |      |\n");
448 }
449 #endif /* CONFIG_PREEMPT_RT */
450
451 /*
452  * osnoise_taint - report an osnoise error.
453  */
454 #define osnoise_taint(msg) ({                                                   \
455         struct osnoise_instance *inst;                                          \
456         struct trace_buffer *buffer;                                            \
457                                                                                 \
458         rcu_read_lock();                                                        \
459         list_for_each_entry_rcu(inst, &osnoise_instances, list) {               \
460                 buffer = inst->tr->array_buffer.buffer;                         \
461                 trace_array_printk_buf(buffer, _THIS_IP_, msg);                 \
462         }                                                                       \
463         rcu_read_unlock();                                                      \
464         osnoise_data.tainted = true;                                            \
465 })
466
467 /*
468  * Record an osnoise_sample into the tracer buffer.
469  */
470 static void
471 __trace_osnoise_sample(struct osnoise_sample *sample, struct trace_buffer *buffer)
472 {
473         struct trace_event_call *call = &event_osnoise;
474         struct ring_buffer_event *event;
475         struct osnoise_entry *entry;
476
477         event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry),
478                                           tracing_gen_ctx());
479         if (!event)
480                 return;
481         entry   = ring_buffer_event_data(event);
482         entry->runtime          = sample->runtime;
483         entry->noise            = sample->noise;
484         entry->max_sample       = sample->max_sample;
485         entry->hw_count         = sample->hw_count;
486         entry->nmi_count        = sample->nmi_count;
487         entry->irq_count        = sample->irq_count;
488         entry->softirq_count    = sample->softirq_count;
489         entry->thread_count     = sample->thread_count;
490
491         if (!call_filter_check_discard(call, entry, buffer, event))
492                 trace_buffer_unlock_commit_nostack(buffer, event);
493 }
494
495 /*
496  * Record an osnoise_sample on all osnoise instances.
497  */
498 static void trace_osnoise_sample(struct osnoise_sample *sample)
499 {
500         struct osnoise_instance *inst;
501         struct trace_buffer *buffer;
502
503         rcu_read_lock();
504         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
505                 buffer = inst->tr->array_buffer.buffer;
506                 __trace_osnoise_sample(sample, buffer);
507         }
508         rcu_read_unlock();
509 }
510
511 #ifdef CONFIG_TIMERLAT_TRACER
512 /*
513  * Print the timerlat header info.
514  */
515 #ifdef CONFIG_PREEMPT_RT
516 static void print_timerlat_headers(struct seq_file *s)
517 {
518         seq_puts(s, "#                                _-------=> irqs-off\n");
519         seq_puts(s, "#                               / _------=> need-resched\n");
520         seq_puts(s, "#                              | / _-----=> need-resched-lazy\n");
521         seq_puts(s, "#                              || / _----=> hardirq/softirq\n");
522         seq_puts(s, "#                              ||| / _---=> preempt-depth\n");
523         seq_puts(s, "#                              |||| / _--=> preempt-lazy-depth\n");
524         seq_puts(s, "#                              ||||| / _-=> migrate-disable\n");
525         seq_puts(s, "#                              |||||| /\n");
526         seq_puts(s, "#                              |||||||             ACTIVATION\n");
527         seq_puts(s, "#           TASK-PID      CPU# |||||||   TIMESTAMP    ID     ");
528         seq_puts(s, "       CONTEXT                LATENCY\n");
529         seq_puts(s, "#              | |         |   |||||||      |         |      ");
530         seq_puts(s, "            |                       |\n");
531 }
532 #else /* CONFIG_PREEMPT_RT */
533 static void print_timerlat_headers(struct seq_file *s)
534 {
535         seq_puts(s, "#                                _-----=> irqs-off\n");
536         seq_puts(s, "#                               / _----=> need-resched\n");
537         seq_puts(s, "#                              | / _---=> hardirq/softirq\n");
538         seq_puts(s, "#                              || / _--=> preempt-depth\n");
539         seq_puts(s, "#                              ||| / _-=> migrate-disable\n");
540         seq_puts(s, "#                              |||| /     delay\n");
541         seq_puts(s, "#                              |||||            ACTIVATION\n");
542         seq_puts(s, "#           TASK-PID      CPU# |||||   TIMESTAMP   ID      ");
543         seq_puts(s, "      CONTEXT                 LATENCY\n");
544         seq_puts(s, "#              | |         |   |||||      |         |      ");
545         seq_puts(s, "            |                       |\n");
546 }
547 #endif /* CONFIG_PREEMPT_RT */
548
549 static void
550 __trace_timerlat_sample(struct timerlat_sample *sample, struct trace_buffer *buffer)
551 {
552         struct trace_event_call *call = &event_osnoise;
553         struct ring_buffer_event *event;
554         struct timerlat_entry *entry;
555
556         event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry),
557                                           tracing_gen_ctx());
558         if (!event)
559                 return;
560         entry   = ring_buffer_event_data(event);
561         entry->seqnum                   = sample->seqnum;
562         entry->context                  = sample->context;
563         entry->timer_latency            = sample->timer_latency;
564
565         if (!call_filter_check_discard(call, entry, buffer, event))
566                 trace_buffer_unlock_commit_nostack(buffer, event);
567 }
568
569 /*
570  * Record an timerlat_sample into the tracer buffer.
571  */
572 static void trace_timerlat_sample(struct timerlat_sample *sample)
573 {
574         struct osnoise_instance *inst;
575         struct trace_buffer *buffer;
576
577         rcu_read_lock();
578         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
579                 buffer = inst->tr->array_buffer.buffer;
580                 __trace_timerlat_sample(sample, buffer);
581         }
582         rcu_read_unlock();
583 }
584
585 #ifdef CONFIG_STACKTRACE
586
587 #define MAX_CALLS       256
588
589 /*
590  * Stack trace will take place only at IRQ level, so, no need
591  * to control nesting here.
592  */
593 struct trace_stack {
594         int             stack_size;
595         int             nr_entries;
596         unsigned long   calls[MAX_CALLS];
597 };
598
599 static DEFINE_PER_CPU(struct trace_stack, trace_stack);
600
601 /*
602  * timerlat_save_stack - save a stack trace without printing
603  *
604  * Save the current stack trace without printing. The
605  * stack will be printed later, after the end of the measurement.
606  */
607 static void timerlat_save_stack(int skip)
608 {
609         unsigned int size, nr_entries;
610         struct trace_stack *fstack;
611
612         fstack = this_cpu_ptr(&trace_stack);
613
614         size = ARRAY_SIZE(fstack->calls);
615
616         nr_entries = stack_trace_save(fstack->calls, size, skip);
617
618         fstack->stack_size = nr_entries * sizeof(unsigned long);
619         fstack->nr_entries = nr_entries;
620
621         return;
622
623 }
624
625 static void
626 __timerlat_dump_stack(struct trace_buffer *buffer, struct trace_stack *fstack, unsigned int size)
627 {
628         struct trace_event_call *call = &event_osnoise;
629         struct ring_buffer_event *event;
630         struct stack_entry *entry;
631
632         event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size,
633                                           tracing_gen_ctx());
634         if (!event)
635                 return;
636
637         entry = ring_buffer_event_data(event);
638
639         memcpy(&entry->caller, fstack->calls, size);
640         entry->size = fstack->nr_entries;
641
642         if (!call_filter_check_discard(call, entry, buffer, event))
643                 trace_buffer_unlock_commit_nostack(buffer, event);
644 }
645
646 /*
647  * timerlat_dump_stack - dump a stack trace previously saved
648  */
649 static void timerlat_dump_stack(u64 latency)
650 {
651         struct osnoise_instance *inst;
652         struct trace_buffer *buffer;
653         struct trace_stack *fstack;
654         unsigned int size;
655
656         /*
657          * trace only if latency > print_stack config, if enabled.
658          */
659         if (!osnoise_data.print_stack || osnoise_data.print_stack > latency)
660                 return;
661
662         preempt_disable_notrace();
663         fstack = this_cpu_ptr(&trace_stack);
664         size = fstack->stack_size;
665
666         rcu_read_lock();
667         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
668                 buffer = inst->tr->array_buffer.buffer;
669                 __timerlat_dump_stack(buffer, fstack, size);
670
671         }
672         rcu_read_unlock();
673         preempt_enable_notrace();
674 }
675 #else /* CONFIG_STACKTRACE */
676 #define timerlat_dump_stack(u64 latency) do {} while (0)
677 #define timerlat_save_stack(a) do {} while (0)
678 #endif /* CONFIG_STACKTRACE */
679 #endif /* CONFIG_TIMERLAT_TRACER */
680
681 /*
682  * Macros to encapsulate the time capturing infrastructure.
683  */
684 #define time_get()      trace_clock_local()
685 #define time_to_us(x)   div_u64(x, 1000)
686 #define time_sub(a, b)  ((a) - (b))
687
688 /*
689  * cond_move_irq_delta_start - Forward the delta_start of a running IRQ
690  *
691  * If an IRQ is preempted by an NMI, its delta_start is pushed forward
692  * to discount the NMI interference.
693  *
694  * See get_int_safe_duration().
695  */
696 static inline void
697 cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration)
698 {
699         if (osn_var->irq.delta_start)
700                 osn_var->irq.delta_start += duration;
701 }
702
703 #ifndef CONFIG_PREEMPT_RT
704 /*
705  * cond_move_softirq_delta_start - Forward the delta_start of a running softirq.
706  *
707  * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed
708  * forward to discount the interference.
709  *
710  * See get_int_safe_duration().
711  */
712 static inline void
713 cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration)
714 {
715         if (osn_var->softirq.delta_start)
716                 osn_var->softirq.delta_start += duration;
717 }
718 #else /* CONFIG_PREEMPT_RT */
719 #define cond_move_softirq_delta_start(osn_var, duration) do {} while (0)
720 #endif
721
722 /*
723  * cond_move_thread_delta_start - Forward the delta_start of a running thread
724  *
725  * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start
726  * is pushed forward to discount the interference.
727  *
728  * See get_int_safe_duration().
729  */
730 static inline void
731 cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration)
732 {
733         if (osn_var->thread.delta_start)
734                 osn_var->thread.delta_start += duration;
735 }
736
737 /*
738  * get_int_safe_duration - Get the duration of a window
739  *
740  * The irq, softirq and thread varaibles need to have its duration without
741  * the interference from higher priority interrupts. Instead of keeping a
742  * variable to discount the interrupt interference from these variables, the
743  * starting time of these variables are pushed forward with the interrupt's
744  * duration. In this way, a single variable is used to:
745  *
746  *   - Know if a given window is being measured.
747  *   - Account its duration.
748  *   - Discount the interference.
749  *
750  * To avoid getting inconsistent values, e.g.,:
751  *
752  *      now = time_get()
753  *              --->    interrupt!
754  *                      delta_start -= int duration;
755  *              <---
756  *      duration = now - delta_start;
757  *
758  *      result: negative duration if the variable duration before the
759  *      interrupt was smaller than the interrupt execution.
760  *
761  * A counter of interrupts is used. If the counter increased, try
762  * to capture an interference safe duration.
763  */
764 static inline s64
765 get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start)
766 {
767         u64 int_counter, now;
768         s64 duration;
769
770         do {
771                 int_counter = local_read(&osn_var->int_counter);
772                 /* synchronize with interrupts */
773                 barrier();
774
775                 now = time_get();
776                 duration = (now - *delta_start);
777
778                 /* synchronize with interrupts */
779                 barrier();
780         } while (int_counter != local_read(&osn_var->int_counter));
781
782         /*
783          * This is an evidence of race conditions that cause
784          * a value to be "discounted" too much.
785          */
786         if (duration < 0)
787                 osnoise_taint("Negative duration!\n");
788
789         *delta_start = 0;
790
791         return duration;
792 }
793
794 /*
795  *
796  * set_int_safe_time - Save the current time on *time, aware of interference
797  *
798  * Get the time, taking into consideration a possible interference from
799  * higher priority interrupts.
800  *
801  * See get_int_safe_duration() for an explanation.
802  */
803 static u64
804 set_int_safe_time(struct osnoise_variables *osn_var, u64 *time)
805 {
806         u64 int_counter;
807
808         do {
809                 int_counter = local_read(&osn_var->int_counter);
810                 /* synchronize with interrupts */
811                 barrier();
812
813                 *time = time_get();
814
815                 /* synchronize with interrupts */
816                 barrier();
817         } while (int_counter != local_read(&osn_var->int_counter));
818
819         return int_counter;
820 }
821
822 #ifdef CONFIG_TIMERLAT_TRACER
823 /*
824  * copy_int_safe_time - Copy *src into *desc aware of interference
825  */
826 static u64
827 copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src)
828 {
829         u64 int_counter;
830
831         do {
832                 int_counter = local_read(&osn_var->int_counter);
833                 /* synchronize with interrupts */
834                 barrier();
835
836                 *dst = *src;
837
838                 /* synchronize with interrupts */
839                 barrier();
840         } while (int_counter != local_read(&osn_var->int_counter));
841
842         return int_counter;
843 }
844 #endif /* CONFIG_TIMERLAT_TRACER */
845
846 /*
847  * trace_osnoise_callback - NMI entry/exit callback
848  *
849  * This function is called at the entry and exit NMI code. The bool enter
850  * distinguishes between either case. This function is used to note a NMI
851  * occurrence, compute the noise caused by the NMI, and to remove the noise
852  * it is potentially causing on other interference variables.
853  */
854 void trace_osnoise_callback(bool enter)
855 {
856         struct osnoise_variables *osn_var = this_cpu_osn_var();
857         u64 duration;
858
859         if (!osn_var->sampling)
860                 return;
861
862         /*
863          * Currently trace_clock_local() calls sched_clock() and the
864          * generic version is not NMI safe.
865          */
866         if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
867                 if (enter) {
868                         osn_var->nmi.delta_start = time_get();
869                         local_inc(&osn_var->int_counter);
870                 } else {
871                         duration = time_get() - osn_var->nmi.delta_start;
872
873                         trace_nmi_noise(osn_var->nmi.delta_start, duration);
874
875                         cond_move_irq_delta_start(osn_var, duration);
876                         cond_move_softirq_delta_start(osn_var, duration);
877                         cond_move_thread_delta_start(osn_var, duration);
878                 }
879         }
880
881         if (enter)
882                 osn_var->nmi.count++;
883 }
884
885 /*
886  * osnoise_trace_irq_entry - Note the starting of an IRQ
887  *
888  * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs,
889  * it is safe to use a single variable (ons_var->irq) to save the statistics.
890  * The arrival_time is used to report... the arrival time. The delta_start
891  * is used to compute the duration at the IRQ exit handler. See
892  * cond_move_irq_delta_start().
893  */
894 void osnoise_trace_irq_entry(int id)
895 {
896         struct osnoise_variables *osn_var = this_cpu_osn_var();
897
898         if (!osn_var->sampling)
899                 return;
900         /*
901          * This value will be used in the report, but not to compute
902          * the execution time, so it is safe to get it unsafe.
903          */
904         osn_var->irq.arrival_time = time_get();
905         set_int_safe_time(osn_var, &osn_var->irq.delta_start);
906         osn_var->irq.count++;
907
908         local_inc(&osn_var->int_counter);
909 }
910
911 /*
912  * osnoise_irq_exit - Note the end of an IRQ, sava data and trace
913  *
914  * Computes the duration of the IRQ noise, and trace it. Also discounts the
915  * interference from other sources of noise could be currently being accounted.
916  */
917 void osnoise_trace_irq_exit(int id, const char *desc)
918 {
919         struct osnoise_variables *osn_var = this_cpu_osn_var();
920         int duration;
921
922         if (!osn_var->sampling)
923                 return;
924
925         duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start);
926         trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration);
927         osn_var->irq.arrival_time = 0;
928         cond_move_softirq_delta_start(osn_var, duration);
929         cond_move_thread_delta_start(osn_var, duration);
930 }
931
932 /*
933  * trace_irqentry_callback - Callback to the irq:irq_entry traceevent
934  *
935  * Used to note the starting of an IRQ occurece.
936  */
937 static void trace_irqentry_callback(void *data, int irq,
938                                     struct irqaction *action)
939 {
940         osnoise_trace_irq_entry(irq);
941 }
942
943 /*
944  * trace_irqexit_callback - Callback to the irq:irq_exit traceevent
945  *
946  * Used to note the end of an IRQ occurece.
947  */
948 static void trace_irqexit_callback(void *data, int irq,
949                                    struct irqaction *action, int ret)
950 {
951         osnoise_trace_irq_exit(irq, action->name);
952 }
953
954 /*
955  * arch specific register function.
956  */
957 int __weak osnoise_arch_register(void)
958 {
959         return 0;
960 }
961
962 /*
963  * arch specific unregister function.
964  */
965 void __weak osnoise_arch_unregister(void)
966 {
967         return;
968 }
969
970 /*
971  * hook_irq_events - Hook IRQ handling events
972  *
973  * This function hooks the IRQ related callbacks to the respective trace
974  * events.
975  */
976 static int hook_irq_events(void)
977 {
978         int ret;
979
980         ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL);
981         if (ret)
982                 goto out_err;
983
984         ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL);
985         if (ret)
986                 goto out_unregister_entry;
987
988         ret = osnoise_arch_register();
989         if (ret)
990                 goto out_irq_exit;
991
992         return 0;
993
994 out_irq_exit:
995         unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
996 out_unregister_entry:
997         unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
998 out_err:
999         return -EINVAL;
1000 }
1001
1002 /*
1003  * unhook_irq_events - Unhook IRQ handling events
1004  *
1005  * This function unhooks the IRQ related callbacks to the respective trace
1006  * events.
1007  */
1008 static void unhook_irq_events(void)
1009 {
1010         osnoise_arch_unregister();
1011         unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
1012         unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1013 }
1014
1015 #ifndef CONFIG_PREEMPT_RT
1016 /*
1017  * trace_softirq_entry_callback - Note the starting of a softirq
1018  *
1019  * Save the starting time of a softirq. As softirqs are non-preemptive to
1020  * other softirqs, it is safe to use a single variable (ons_var->softirq)
1021  * to save the statistics. The arrival_time is used to report... the
1022  * arrival time. The delta_start is used to compute the duration at the
1023  * softirq exit handler. See cond_move_softirq_delta_start().
1024  */
1025 static void trace_softirq_entry_callback(void *data, unsigned int vec_nr)
1026 {
1027         struct osnoise_variables *osn_var = this_cpu_osn_var();
1028
1029         if (!osn_var->sampling)
1030                 return;
1031         /*
1032          * This value will be used in the report, but not to compute
1033          * the execution time, so it is safe to get it unsafe.
1034          */
1035         osn_var->softirq.arrival_time = time_get();
1036         set_int_safe_time(osn_var, &osn_var->softirq.delta_start);
1037         osn_var->softirq.count++;
1038
1039         local_inc(&osn_var->int_counter);
1040 }
1041
1042 /*
1043  * trace_softirq_exit_callback - Note the end of an softirq
1044  *
1045  * Computes the duration of the softirq noise, and trace it. Also discounts the
1046  * interference from other sources of noise could be currently being accounted.
1047  */
1048 static void trace_softirq_exit_callback(void *data, unsigned int vec_nr)
1049 {
1050         struct osnoise_variables *osn_var = this_cpu_osn_var();
1051         int duration;
1052
1053         if (!osn_var->sampling)
1054                 return;
1055
1056         if (unlikely(timerlat_enabled()))
1057                 if (!timerlat_softirq_exit(osn_var))
1058                         return;
1059
1060         duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start);
1061         trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration);
1062         cond_move_thread_delta_start(osn_var, duration);
1063         osn_var->softirq.arrival_time = 0;
1064 }
1065
1066 /*
1067  * hook_softirq_events - Hook softirq handling events
1068  *
1069  * This function hooks the softirq related callbacks to the respective trace
1070  * events.
1071  */
1072 static int hook_softirq_events(void)
1073 {
1074         int ret;
1075
1076         ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1077         if (ret)
1078                 goto out_err;
1079
1080         ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1081         if (ret)
1082                 goto out_unreg_entry;
1083
1084         return 0;
1085
1086 out_unreg_entry:
1087         unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1088 out_err:
1089         return -EINVAL;
1090 }
1091
1092 /*
1093  * unhook_softirq_events - Unhook softirq handling events
1094  *
1095  * This function hooks the softirq related callbacks to the respective trace
1096  * events.
1097  */
1098 static void unhook_softirq_events(void)
1099 {
1100         unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1101         unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1102 }
1103 #else /* CONFIG_PREEMPT_RT */
1104 /*
1105  * softirq are threads on the PREEMPT_RT mode.
1106  */
1107 static int hook_softirq_events(void)
1108 {
1109         return 0;
1110 }
1111 static void unhook_softirq_events(void)
1112 {
1113 }
1114 #endif
1115
1116 /*
1117  * thread_entry - Record the starting of a thread noise window
1118  *
1119  * It saves the context switch time for a noisy thread, and increments
1120  * the interference counters.
1121  */
1122 static void
1123 thread_entry(struct osnoise_variables *osn_var, struct task_struct *t)
1124 {
1125         if (!osn_var->sampling)
1126                 return;
1127         /*
1128          * The arrival time will be used in the report, but not to compute
1129          * the execution time, so it is safe to get it unsafe.
1130          */
1131         osn_var->thread.arrival_time = time_get();
1132
1133         set_int_safe_time(osn_var, &osn_var->thread.delta_start);
1134
1135         osn_var->thread.count++;
1136         local_inc(&osn_var->int_counter);
1137 }
1138
1139 /*
1140  * thread_exit - Report the end of a thread noise window
1141  *
1142  * It computes the total noise from a thread, tracing if needed.
1143  */
1144 static void
1145 thread_exit(struct osnoise_variables *osn_var, struct task_struct *t)
1146 {
1147         int duration;
1148
1149         if (!osn_var->sampling)
1150                 return;
1151
1152         if (unlikely(timerlat_enabled()))
1153                 if (!timerlat_thread_exit(osn_var))
1154                         return;
1155
1156         duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start);
1157
1158         trace_thread_noise(t, osn_var->thread.arrival_time, duration);
1159
1160         osn_var->thread.arrival_time = 0;
1161 }
1162
1163 /*
1164  * trace_sched_switch - sched:sched_switch trace event handler
1165  *
1166  * This function is hooked to the sched:sched_switch trace event, and it is
1167  * used to record the beginning and to report the end of a thread noise window.
1168  */
1169 static void
1170 trace_sched_switch_callback(void *data, bool preempt, struct task_struct *p,
1171                             struct task_struct *n)
1172 {
1173         struct osnoise_variables *osn_var = this_cpu_osn_var();
1174
1175         if (p->pid != osn_var->pid)
1176                 thread_exit(osn_var, p);
1177
1178         if (n->pid != osn_var->pid)
1179                 thread_entry(osn_var, n);
1180 }
1181
1182 /*
1183  * hook_thread_events - Hook the insturmentation for thread noise
1184  *
1185  * Hook the osnoise tracer callbacks to handle the noise from other
1186  * threads on the necessary kernel events.
1187  */
1188 static int hook_thread_events(void)
1189 {
1190         int ret;
1191
1192         ret = register_trace_sched_switch(trace_sched_switch_callback, NULL);
1193         if (ret)
1194                 return -EINVAL;
1195
1196         return 0;
1197 }
1198
1199 /*
1200  * unhook_thread_events - *nhook the insturmentation for thread noise
1201  *
1202  * Unook the osnoise tracer callbacks to handle the noise from other
1203  * threads on the necessary kernel events.
1204  */
1205 static void unhook_thread_events(void)
1206 {
1207         unregister_trace_sched_switch(trace_sched_switch_callback, NULL);
1208 }
1209
1210 /*
1211  * save_osn_sample_stats - Save the osnoise_sample statistics
1212  *
1213  * Save the osnoise_sample statistics before the sampling phase. These
1214  * values will be used later to compute the diff betwneen the statistics
1215  * before and after the osnoise sampling.
1216  */
1217 static void
1218 save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1219 {
1220         s->nmi_count = osn_var->nmi.count;
1221         s->irq_count = osn_var->irq.count;
1222         s->softirq_count = osn_var->softirq.count;
1223         s->thread_count = osn_var->thread.count;
1224 }
1225
1226 /*
1227  * diff_osn_sample_stats - Compute the osnoise_sample statistics
1228  *
1229  * After a sample period, compute the difference on the osnoise_sample
1230  * statistics. The struct osnoise_sample *s contains the statistics saved via
1231  * save_osn_sample_stats() before the osnoise sampling.
1232  */
1233 static void
1234 diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1235 {
1236         s->nmi_count = osn_var->nmi.count - s->nmi_count;
1237         s->irq_count = osn_var->irq.count - s->irq_count;
1238         s->softirq_count = osn_var->softirq.count - s->softirq_count;
1239         s->thread_count = osn_var->thread.count - s->thread_count;
1240 }
1241
1242 /*
1243  * osnoise_stop_tracing - Stop tracing and the tracer.
1244  */
1245 static __always_inline void osnoise_stop_tracing(void)
1246 {
1247         struct osnoise_instance *inst;
1248         struct trace_array *tr;
1249
1250         rcu_read_lock();
1251         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1252                 tr = inst->tr;
1253                 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_,
1254                                 "stop tracing hit on cpu %d\n", smp_processor_id());
1255
1256                 tracer_tracing_off(tr);
1257         }
1258         rcu_read_unlock();
1259 }
1260
1261 /*
1262  * notify_new_max_latency - Notify a new max latency via fsnotify interface.
1263  */
1264 static void notify_new_max_latency(u64 latency)
1265 {
1266         struct osnoise_instance *inst;
1267         struct trace_array *tr;
1268
1269         rcu_read_lock();
1270         list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1271                 tr = inst->tr;
1272                 if (tr->max_latency < latency) {
1273                         tr->max_latency = latency;
1274                         latency_fsnotify(tr);
1275                 }
1276         }
1277         rcu_read_unlock();
1278 }
1279
1280 /*
1281  * run_osnoise - Sample the time and look for osnoise
1282  *
1283  * Used to capture the time, looking for potential osnoise latency repeatedly.
1284  * Different from hwlat_detector, it is called with preemption and interrupts
1285  * enabled. This allows irqs, softirqs and threads to run, interfering on the
1286  * osnoise sampling thread, as they would do with a regular thread.
1287  */
1288 static int run_osnoise(void)
1289 {
1290         struct osnoise_variables *osn_var = this_cpu_osn_var();
1291         u64 start, sample, last_sample;
1292         u64 last_int_count, int_count;
1293         s64 noise = 0, max_noise = 0;
1294         s64 total, last_total = 0;
1295         struct osnoise_sample s;
1296         unsigned int threshold;
1297         u64 runtime, stop_in;
1298         u64 sum_noise = 0;
1299         int hw_count = 0;
1300         int ret = -1;
1301
1302         /*
1303          * Considers the current thread as the workload.
1304          */
1305         osn_var->pid = current->pid;
1306
1307         /*
1308          * Save the current stats for the diff
1309          */
1310         save_osn_sample_stats(osn_var, &s);
1311
1312         /*
1313          * if threshold is 0, use the default value of 5 us.
1314          */
1315         threshold = tracing_thresh ? : 5000;
1316
1317         /*
1318          * Make sure NMIs see sampling first
1319          */
1320         osn_var->sampling = true;
1321         barrier();
1322
1323         /*
1324          * Transform the *_us config to nanoseconds to avoid the
1325          * division on the main loop.
1326          */
1327         runtime = osnoise_data.sample_runtime * NSEC_PER_USEC;
1328         stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC;
1329
1330         /*
1331          * Start timestemp
1332          */
1333         start = time_get();
1334
1335         /*
1336          * "previous" loop.
1337          */
1338         last_int_count = set_int_safe_time(osn_var, &last_sample);
1339
1340         do {
1341                 /*
1342                  * Get sample!
1343                  */
1344                 int_count = set_int_safe_time(osn_var, &sample);
1345
1346                 noise = time_sub(sample, last_sample);
1347
1348                 /*
1349                  * This shouldn't happen.
1350                  */
1351                 if (noise < 0) {
1352                         osnoise_taint("negative noise!");
1353                         goto out;
1354                 }
1355
1356                 /*
1357                  * Sample runtime.
1358                  */
1359                 total = time_sub(sample, start);
1360
1361                 /*
1362                  * Check for possible overflows.
1363                  */
1364                 if (total < last_total) {
1365                         osnoise_taint("total overflow!");
1366                         break;
1367                 }
1368
1369                 last_total = total;
1370
1371                 if (noise >= threshold) {
1372                         int interference = int_count - last_int_count;
1373
1374                         if (noise > max_noise)
1375                                 max_noise = noise;
1376
1377                         if (!interference)
1378                                 hw_count++;
1379
1380                         sum_noise += noise;
1381
1382                         trace_sample_threshold(last_sample, noise, interference);
1383
1384                         if (osnoise_data.stop_tracing)
1385                                 if (noise > stop_in)
1386                                         osnoise_stop_tracing();
1387                 }
1388
1389                 /*
1390                  * In some cases, notably when running on a nohz_full CPU with
1391                  * a stopped tick PREEMPT_RCU has no way to account for QSs.
1392                  * This will eventually cause unwarranted noise as PREEMPT_RCU
1393                  * will force preemption as the means of ending the current
1394                  * grace period. We avoid this problem by calling
1395                  * rcu_momentary_dyntick_idle(), which performs a zero duration
1396                  * EQS allowing PREEMPT_RCU to end the current grace period.
1397                  * This call shouldn't be wrapped inside an RCU critical
1398                  * section.
1399                  *
1400                  * Note that in non PREEMPT_RCU kernels QSs are handled through
1401                  * cond_resched()
1402                  */
1403                 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
1404                         local_irq_disable();
1405                         rcu_momentary_dyntick_idle();
1406                         local_irq_enable();
1407                 }
1408
1409                 /*
1410                  * For the non-preemptive kernel config: let threads runs, if
1411                  * they so wish.
1412                  */
1413                 cond_resched();
1414
1415                 last_sample = sample;
1416                 last_int_count = int_count;
1417
1418         } while (total < runtime && !kthread_should_stop());
1419
1420         /*
1421          * Finish the above in the view for interrupts.
1422          */
1423         barrier();
1424
1425         osn_var->sampling = false;
1426
1427         /*
1428          * Make sure sampling data is no longer updated.
1429          */
1430         barrier();
1431
1432         /*
1433          * Save noise info.
1434          */
1435         s.noise = time_to_us(sum_noise);
1436         s.runtime = time_to_us(total);
1437         s.max_sample = time_to_us(max_noise);
1438         s.hw_count = hw_count;
1439
1440         /* Save interference stats info */
1441         diff_osn_sample_stats(osn_var, &s);
1442
1443         trace_osnoise_sample(&s);
1444
1445         notify_new_max_latency(max_noise);
1446
1447         if (osnoise_data.stop_tracing_total)
1448                 if (s.noise > osnoise_data.stop_tracing_total)
1449                         osnoise_stop_tracing();
1450
1451         return 0;
1452 out:
1453         return ret;
1454 }
1455
1456 static struct cpumask osnoise_cpumask;
1457 static struct cpumask save_cpumask;
1458
1459 /*
1460  * osnoise_sleep - sleep until the next period
1461  */
1462 static void osnoise_sleep(void)
1463 {
1464         u64 interval;
1465         ktime_t wake_time;
1466
1467         mutex_lock(&interface_lock);
1468         interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
1469         mutex_unlock(&interface_lock);
1470
1471         /*
1472          * differently from hwlat_detector, the osnoise tracer can run
1473          * without a pause because preemption is on.
1474          */
1475         if (!interval) {
1476                 /* Let synchronize_rcu_tasks() make progress */
1477                 cond_resched_tasks_rcu_qs();
1478                 return;
1479         }
1480
1481         wake_time = ktime_add_us(ktime_get(), interval);
1482         __set_current_state(TASK_INTERRUPTIBLE);
1483
1484         while (schedule_hrtimeout_range(&wake_time, 0, HRTIMER_MODE_ABS)) {
1485                 if (kthread_should_stop())
1486                         break;
1487         }
1488 }
1489
1490 /*
1491  * osnoise_main - The osnoise detection kernel thread
1492  *
1493  * Calls run_osnoise() function to measure the osnoise for the configured runtime,
1494  * every period.
1495  */
1496 static int osnoise_main(void *data)
1497 {
1498
1499         while (!kthread_should_stop()) {
1500                 run_osnoise();
1501                 osnoise_sleep();
1502         }
1503
1504         return 0;
1505 }
1506
1507 #ifdef CONFIG_TIMERLAT_TRACER
1508 /*
1509  * timerlat_irq - hrtimer handler for timerlat.
1510  */
1511 static enum hrtimer_restart timerlat_irq(struct hrtimer *timer)
1512 {
1513         struct osnoise_variables *osn_var = this_cpu_osn_var();
1514         struct timerlat_variables *tlat;
1515         struct timerlat_sample s;
1516         u64 now;
1517         u64 diff;
1518
1519         /*
1520          * I am not sure if the timer was armed for this CPU. So, get
1521          * the timerlat struct from the timer itself, not from this
1522          * CPU.
1523          */
1524         tlat = container_of(timer, struct timerlat_variables, timer);
1525
1526         now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1527
1528         /*
1529          * Enable the osnoise: events for thread an softirq.
1530          */
1531         tlat->tracing_thread = true;
1532
1533         osn_var->thread.arrival_time = time_get();
1534
1535         /*
1536          * A hardirq is running: the timer IRQ. It is for sure preempting
1537          * a thread, and potentially preempting a softirq.
1538          *
1539          * At this point, it is not interesting to know the duration of the
1540          * preempted thread (and maybe softirq), but how much time they will
1541          * delay the beginning of the execution of the timer thread.
1542          *
1543          * To get the correct (net) delay added by the softirq, its delta_start
1544          * is set as the IRQ one. In this way, at the return of the IRQ, the delta
1545          * start of the sofitrq will be zeroed, accounting then only the time
1546          * after that.
1547          *
1548          * The thread follows the same principle. However, if a softirq is
1549          * running, the thread needs to receive the softirq delta_start. The
1550          * reason being is that the softirq will be the last to be unfolded,
1551          * resseting the thread delay to zero.
1552          *
1553          * The PREEMPT_RT is a special case, though. As softirqs run as threads
1554          * on RT, moving the thread is enough.
1555          */
1556         if (!IS_ENABLED(CONFIG_PREEMPT_RT) && osn_var->softirq.delta_start) {
1557                 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1558                                    &osn_var->softirq.delta_start);
1559
1560                 copy_int_safe_time(osn_var, &osn_var->softirq.delta_start,
1561                                     &osn_var->irq.delta_start);
1562         } else {
1563                 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1564                                     &osn_var->irq.delta_start);
1565         }
1566
1567         /*
1568          * Compute the current time with the expected time.
1569          */
1570         diff = now - tlat->abs_period;
1571
1572         tlat->count++;
1573         s.seqnum = tlat->count;
1574         s.timer_latency = diff;
1575         s.context = IRQ_CONTEXT;
1576
1577         trace_timerlat_sample(&s);
1578
1579         notify_new_max_latency(diff);
1580
1581         if (osnoise_data.stop_tracing)
1582                 if (time_to_us(diff) >= osnoise_data.stop_tracing)
1583                         osnoise_stop_tracing();
1584
1585         wake_up_process(tlat->kthread);
1586
1587         if (osnoise_data.print_stack)
1588                 timerlat_save_stack(0);
1589
1590         return HRTIMER_NORESTART;
1591 }
1592
1593 /*
1594  * wait_next_period - Wait for the next period for timerlat
1595  */
1596 static int wait_next_period(struct timerlat_variables *tlat)
1597 {
1598         ktime_t next_abs_period, now;
1599         u64 rel_period = osnoise_data.timerlat_period * 1000;
1600
1601         now = hrtimer_cb_get_time(&tlat->timer);
1602         next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1603
1604         /*
1605          * Save the next abs_period.
1606          */
1607         tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1608
1609         /*
1610          * If the new abs_period is in the past, skip the activation.
1611          */
1612         while (ktime_compare(now, next_abs_period) > 0) {
1613                 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1614                 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1615         }
1616
1617         set_current_state(TASK_INTERRUPTIBLE);
1618
1619         hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD);
1620         schedule();
1621         return 1;
1622 }
1623
1624 /*
1625  * timerlat_main- Timerlat main
1626  */
1627 static int timerlat_main(void *data)
1628 {
1629         struct osnoise_variables *osn_var = this_cpu_osn_var();
1630         struct timerlat_variables *tlat = this_cpu_tmr_var();
1631         struct timerlat_sample s;
1632         struct sched_param sp;
1633         u64 now, diff;
1634
1635         /*
1636          * Make the thread RT, that is how cyclictest is usually used.
1637          */
1638         sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
1639         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1640
1641         tlat->count = 0;
1642         tlat->tracing_thread = false;
1643
1644         hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1645         tlat->timer.function = timerlat_irq;
1646         tlat->kthread = current;
1647         osn_var->pid = current->pid;
1648         /*
1649          * Anotate the arrival time.
1650          */
1651         tlat->abs_period = hrtimer_cb_get_time(&tlat->timer);
1652
1653         wait_next_period(tlat);
1654
1655         osn_var->sampling = 1;
1656
1657         while (!kthread_should_stop()) {
1658                 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1659                 diff = now - tlat->abs_period;
1660
1661                 s.seqnum = tlat->count;
1662                 s.timer_latency = diff;
1663                 s.context = THREAD_CONTEXT;
1664
1665                 trace_timerlat_sample(&s);
1666
1667                 timerlat_dump_stack(time_to_us(diff));
1668
1669                 tlat->tracing_thread = false;
1670                 if (osnoise_data.stop_tracing_total)
1671                         if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
1672                                 osnoise_stop_tracing();
1673
1674                 wait_next_period(tlat);
1675         }
1676
1677         hrtimer_cancel(&tlat->timer);
1678         return 0;
1679 }
1680 #else /* CONFIG_TIMERLAT_TRACER */
1681 static int timerlat_main(void *data)
1682 {
1683         return 0;
1684 }
1685 #endif /* CONFIG_TIMERLAT_TRACER */
1686
1687 /*
1688  * stop_kthread - stop a workload thread
1689  */
1690 static void stop_kthread(unsigned int cpu)
1691 {
1692         struct task_struct *kthread;
1693
1694         kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread;
1695         if (kthread)
1696                 kthread_stop(kthread);
1697         per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1698 }
1699
1700 /*
1701  * stop_per_cpu_kthread - Stop per-cpu threads
1702  *
1703  * Stop the osnoise sampling htread. Use this on unload and at system
1704  * shutdown.
1705  */
1706 static void stop_per_cpu_kthreads(void)
1707 {
1708         int cpu;
1709
1710         cpus_read_lock();
1711
1712         for_each_online_cpu(cpu)
1713                 stop_kthread(cpu);
1714
1715         cpus_read_unlock();
1716 }
1717
1718 /*
1719  * start_kthread - Start a workload tread
1720  */
1721 static int start_kthread(unsigned int cpu)
1722 {
1723         struct task_struct *kthread;
1724         void *main = osnoise_main;
1725         char comm[24];
1726
1727         if (timerlat_enabled()) {
1728                 snprintf(comm, 24, "timerlat/%d", cpu);
1729                 main = timerlat_main;
1730         } else {
1731                 snprintf(comm, 24, "osnoise/%d", cpu);
1732         }
1733
1734         kthread = kthread_run_on_cpu(main, NULL, cpu, comm);
1735
1736         if (IS_ERR(kthread)) {
1737                 pr_err(BANNER "could not start sampling thread\n");
1738                 stop_per_cpu_kthreads();
1739                 return -ENOMEM;
1740         }
1741
1742         per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread;
1743
1744         return 0;
1745 }
1746
1747 /*
1748  * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads
1749  *
1750  * This starts the kernel thread that will look for osnoise on many
1751  * cpus.
1752  */
1753 static int start_per_cpu_kthreads(void)
1754 {
1755         struct cpumask *current_mask = &save_cpumask;
1756         int retval = 0;
1757         int cpu;
1758
1759         cpus_read_lock();
1760         /*
1761          * Run only on online CPUs in which osnoise is allowed to run.
1762          */
1763         cpumask_and(current_mask, cpu_online_mask, &osnoise_cpumask);
1764
1765         for_each_possible_cpu(cpu)
1766                 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1767
1768         for_each_cpu(cpu, current_mask) {
1769                 retval = start_kthread(cpu);
1770                 if (retval) {
1771                         stop_per_cpu_kthreads();
1772                         break;
1773                 }
1774         }
1775
1776         cpus_read_unlock();
1777
1778         return retval;
1779 }
1780
1781 #ifdef CONFIG_HOTPLUG_CPU
1782 static void osnoise_hotplug_workfn(struct work_struct *dummy)
1783 {
1784         unsigned int cpu = smp_processor_id();
1785
1786         mutex_lock(&trace_types_lock);
1787
1788         if (!osnoise_has_registered_instances())
1789                 goto out_unlock_trace;
1790
1791         mutex_lock(&interface_lock);
1792         cpus_read_lock();
1793
1794         if (!cpumask_test_cpu(cpu, &osnoise_cpumask))
1795                 goto out_unlock;
1796
1797         start_kthread(cpu);
1798
1799 out_unlock:
1800         cpus_read_unlock();
1801         mutex_unlock(&interface_lock);
1802 out_unlock_trace:
1803         mutex_unlock(&trace_types_lock);
1804 }
1805
1806 static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn);
1807
1808 /*
1809  * osnoise_cpu_init - CPU hotplug online callback function
1810  */
1811 static int osnoise_cpu_init(unsigned int cpu)
1812 {
1813         schedule_work_on(cpu, &osnoise_hotplug_work);
1814         return 0;
1815 }
1816
1817 /*
1818  * osnoise_cpu_die - CPU hotplug offline callback function
1819  */
1820 static int osnoise_cpu_die(unsigned int cpu)
1821 {
1822         stop_kthread(cpu);
1823         return 0;
1824 }
1825
1826 static void osnoise_init_hotplug_support(void)
1827 {
1828         int ret;
1829
1830         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online",
1831                                 osnoise_cpu_init, osnoise_cpu_die);
1832         if (ret < 0)
1833                 pr_warn(BANNER "Error to init cpu hotplug support\n");
1834
1835         return;
1836 }
1837 #else /* CONFIG_HOTPLUG_CPU */
1838 static void osnoise_init_hotplug_support(void)
1839 {
1840         return;
1841 }
1842 #endif /* CONFIG_HOTPLUG_CPU */
1843
1844 /*
1845  * osnoise_cpus_read - Read function for reading the "cpus" file
1846  * @filp: The active open file structure
1847  * @ubuf: The userspace provided buffer to read value into
1848  * @cnt: The maximum number of bytes to read
1849  * @ppos: The current "file" position
1850  *
1851  * Prints the "cpus" output into the user-provided buffer.
1852  */
1853 static ssize_t
1854 osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count,
1855                   loff_t *ppos)
1856 {
1857         char *mask_str;
1858         int len;
1859
1860         mutex_lock(&interface_lock);
1861
1862         len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1;
1863         mask_str = kmalloc(len, GFP_KERNEL);
1864         if (!mask_str) {
1865                 count = -ENOMEM;
1866                 goto out_unlock;
1867         }
1868
1869         len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask));
1870         if (len >= count) {
1871                 count = -EINVAL;
1872                 goto out_free;
1873         }
1874
1875         count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
1876
1877 out_free:
1878         kfree(mask_str);
1879 out_unlock:
1880         mutex_unlock(&interface_lock);
1881
1882         return count;
1883 }
1884
1885 /*
1886  * osnoise_cpus_write - Write function for "cpus" entry
1887  * @filp: The active open file structure
1888  * @ubuf: The user buffer that contains the value to write
1889  * @cnt: The maximum number of bytes to write to "file"
1890  * @ppos: The current position in @file
1891  *
1892  * This function provides a write implementation for the "cpus"
1893  * interface to the osnoise trace. By default, it lists all  CPUs,
1894  * in this way, allowing osnoise threads to run on any online CPU
1895  * of the system. It serves to restrict the execution of osnoise to the
1896  * set of CPUs writing via this interface. Why not use "tracing_cpumask"?
1897  * Because the user might be interested in tracing what is running on
1898  * other CPUs. For instance, one might run osnoise in one HT CPU
1899  * while observing what is running on the sibling HT CPU.
1900  */
1901 static ssize_t
1902 osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
1903                    loff_t *ppos)
1904 {
1905         cpumask_var_t osnoise_cpumask_new;
1906         int running, err;
1907         char buf[256];
1908
1909         if (count >= 256)
1910                 return -EINVAL;
1911
1912         if (copy_from_user(buf, ubuf, count))
1913                 return -EFAULT;
1914
1915         if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL))
1916                 return -ENOMEM;
1917
1918         err = cpulist_parse(buf, osnoise_cpumask_new);
1919         if (err)
1920                 goto err_free;
1921
1922         /*
1923          * trace_types_lock is taken to avoid concurrency on start/stop.
1924          */
1925         mutex_lock(&trace_types_lock);
1926         running = osnoise_has_registered_instances();
1927         if (running)
1928                 stop_per_cpu_kthreads();
1929
1930         mutex_lock(&interface_lock);
1931         /*
1932          * osnoise_cpumask is read by CPU hotplug operations.
1933          */
1934         cpus_read_lock();
1935
1936         cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
1937
1938         cpus_read_unlock();
1939         mutex_unlock(&interface_lock);
1940
1941         if (running)
1942                 start_per_cpu_kthreads();
1943         mutex_unlock(&trace_types_lock);
1944
1945         free_cpumask_var(osnoise_cpumask_new);
1946         return count;
1947
1948 err_free:
1949         free_cpumask_var(osnoise_cpumask_new);
1950
1951         return err;
1952 }
1953
1954 /*
1955  * osnoise/runtime_us: cannot be greater than the period.
1956  */
1957 static struct trace_min_max_param osnoise_runtime = {
1958         .lock   = &interface_lock,
1959         .val    = &osnoise_data.sample_runtime,
1960         .max    = &osnoise_data.sample_period,
1961         .min    = NULL,
1962 };
1963
1964 /*
1965  * osnoise/period_us: cannot be smaller than the runtime.
1966  */
1967 static struct trace_min_max_param osnoise_period = {
1968         .lock   = &interface_lock,
1969         .val    = &osnoise_data.sample_period,
1970         .max    = NULL,
1971         .min    = &osnoise_data.sample_runtime,
1972 };
1973
1974 /*
1975  * osnoise/stop_tracing_us: no limit.
1976  */
1977 static struct trace_min_max_param osnoise_stop_tracing_in = {
1978         .lock   = &interface_lock,
1979         .val    = &osnoise_data.stop_tracing,
1980         .max    = NULL,
1981         .min    = NULL,
1982 };
1983
1984 /*
1985  * osnoise/stop_tracing_total_us: no limit.
1986  */
1987 static struct trace_min_max_param osnoise_stop_tracing_total = {
1988         .lock   = &interface_lock,
1989         .val    = &osnoise_data.stop_tracing_total,
1990         .max    = NULL,
1991         .min    = NULL,
1992 };
1993
1994 #ifdef CONFIG_TIMERLAT_TRACER
1995 /*
1996  * osnoise/print_stack: print the stacktrace of the IRQ handler if the total
1997  * latency is higher than val.
1998  */
1999 static struct trace_min_max_param osnoise_print_stack = {
2000         .lock   = &interface_lock,
2001         .val    = &osnoise_data.print_stack,
2002         .max    = NULL,
2003         .min    = NULL,
2004 };
2005
2006 /*
2007  * osnoise/timerlat_period: min 100 us, max 1 s
2008  */
2009 u64 timerlat_min_period = 100;
2010 u64 timerlat_max_period = 1000000;
2011 static struct trace_min_max_param timerlat_period = {
2012         .lock   = &interface_lock,
2013         .val    = &osnoise_data.timerlat_period,
2014         .max    = &timerlat_max_period,
2015         .min    = &timerlat_min_period,
2016 };
2017 #endif
2018
2019 static const struct file_operations cpus_fops = {
2020         .open           = tracing_open_generic,
2021         .read           = osnoise_cpus_read,
2022         .write          = osnoise_cpus_write,
2023         .llseek         = generic_file_llseek,
2024 };
2025
2026 #ifdef CONFIG_TIMERLAT_TRACER
2027 #ifdef CONFIG_STACKTRACE
2028 static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2029 {
2030         struct dentry *tmp;
2031
2032         tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir,
2033                                   &osnoise_print_stack, &trace_min_max_fops);
2034         if (!tmp)
2035                 return -ENOMEM;
2036
2037         return 0;
2038 }
2039 #else /* CONFIG_STACKTRACE */
2040 static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2041 {
2042         return 0;
2043 }
2044 #endif /* CONFIG_STACKTRACE */
2045
2046 /*
2047  * init_timerlat_tracefs - A function to initialize the timerlat interface files
2048  */
2049 static int init_timerlat_tracefs(struct dentry *top_dir)
2050 {
2051         struct dentry *tmp;
2052
2053         tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir,
2054                                   &timerlat_period, &trace_min_max_fops);
2055         if (!tmp)
2056                 return -ENOMEM;
2057
2058         return init_timerlat_stack_tracefs(top_dir);
2059 }
2060 #else /* CONFIG_TIMERLAT_TRACER */
2061 static int init_timerlat_tracefs(struct dentry *top_dir)
2062 {
2063         return 0;
2064 }
2065 #endif /* CONFIG_TIMERLAT_TRACER */
2066
2067 /*
2068  * init_tracefs - A function to initialize the tracefs interface files
2069  *
2070  * This function creates entries in tracefs for "osnoise" and "timerlat".
2071  * It creates these directories in the tracing directory, and within that
2072  * directory the use can change and view the configs.
2073  */
2074 static int init_tracefs(void)
2075 {
2076         struct dentry *top_dir;
2077         struct dentry *tmp;
2078         int ret;
2079
2080         ret = tracing_init_dentry();
2081         if (ret)
2082                 return -ENOMEM;
2083
2084         top_dir = tracefs_create_dir("osnoise", NULL);
2085         if (!top_dir)
2086                 return 0;
2087
2088         tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir,
2089                                   &osnoise_period, &trace_min_max_fops);
2090         if (!tmp)
2091                 goto err;
2092
2093         tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir,
2094                                   &osnoise_runtime, &trace_min_max_fops);
2095         if (!tmp)
2096                 goto err;
2097
2098         tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir,
2099                                   &osnoise_stop_tracing_in, &trace_min_max_fops);
2100         if (!tmp)
2101                 goto err;
2102
2103         tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir,
2104                                   &osnoise_stop_tracing_total, &trace_min_max_fops);
2105         if (!tmp)
2106                 goto err;
2107
2108         tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops);
2109         if (!tmp)
2110                 goto err;
2111
2112         ret = init_timerlat_tracefs(top_dir);
2113         if (ret)
2114                 goto err;
2115
2116         return 0;
2117
2118 err:
2119         tracefs_remove(top_dir);
2120         return -ENOMEM;
2121 }
2122
2123 static int osnoise_hook_events(void)
2124 {
2125         int retval;
2126
2127         /*
2128          * Trace is already hooked, we are re-enabling from
2129          * a stop_tracing_*.
2130          */
2131         if (trace_osnoise_callback_enabled)
2132                 return 0;
2133
2134         retval = hook_irq_events();
2135         if (retval)
2136                 return -EINVAL;
2137
2138         retval = hook_softirq_events();
2139         if (retval)
2140                 goto out_unhook_irq;
2141
2142         retval = hook_thread_events();
2143         /*
2144          * All fine!
2145          */
2146         if (!retval)
2147                 return 0;
2148
2149         unhook_softirq_events();
2150 out_unhook_irq:
2151         unhook_irq_events();
2152         return -EINVAL;
2153 }
2154
2155 static void osnoise_unhook_events(void)
2156 {
2157         unhook_thread_events();
2158         unhook_softirq_events();
2159         unhook_irq_events();
2160 }
2161
2162 /*
2163  * osnoise_workload_start - start the workload and hook to events
2164  */
2165 static int osnoise_workload_start(void)
2166 {
2167         int retval;
2168
2169         /*
2170          * Instances need to be registered after calling workload
2171          * start. Hence, if there is already an instance, the
2172          * workload was already registered. Otherwise, this
2173          * code is on the way to register the first instance,
2174          * and the workload will start.
2175          */
2176         if (osnoise_has_registered_instances())
2177                 return 0;
2178
2179         osn_var_reset_all();
2180
2181         retval = osnoise_hook_events();
2182         if (retval)
2183                 return retval;
2184
2185         /*
2186          * Make sure that ftrace_nmi_enter/exit() see reset values
2187          * before enabling trace_osnoise_callback_enabled.
2188          */
2189         barrier();
2190         trace_osnoise_callback_enabled = true;
2191
2192         retval = start_per_cpu_kthreads();
2193         if (retval) {
2194                 trace_osnoise_callback_enabled = false;
2195                 /*
2196                  * Make sure that ftrace_nmi_enter/exit() see
2197                  * trace_osnoise_callback_enabled as false before continuing.
2198                  */
2199                 barrier();
2200
2201                 osnoise_unhook_events();
2202                 return retval;
2203         }
2204
2205         return 0;
2206 }
2207
2208 /*
2209  * osnoise_workload_stop - stop the workload and unhook the events
2210  */
2211 static void osnoise_workload_stop(void)
2212 {
2213         /*
2214          * Instances need to be unregistered before calling
2215          * stop. Hence, if there is a registered instance, more
2216          * than one instance is running, and the workload will not
2217          * yet stop. Otherwise, this code is on the way to disable
2218          * the last instance, and the workload can stop.
2219          */
2220         if (osnoise_has_registered_instances())
2221                 return;
2222
2223         /*
2224          * If callbacks were already disabled in a previous stop
2225          * call, there is no need to disable then again.
2226          *
2227          * For instance, this happens when tracing is stopped via:
2228          * echo 0 > tracing_on
2229          * echo nop > current_tracer.
2230          */
2231         if (!trace_osnoise_callback_enabled)
2232                 return;
2233
2234         trace_osnoise_callback_enabled = false;
2235         /*
2236          * Make sure that ftrace_nmi_enter/exit() see
2237          * trace_osnoise_callback_enabled as false before continuing.
2238          */
2239         barrier();
2240
2241         stop_per_cpu_kthreads();
2242
2243         osnoise_unhook_events();
2244 }
2245
2246 static void osnoise_tracer_start(struct trace_array *tr)
2247 {
2248         int retval;
2249
2250         /*
2251          * If the instance is already registered, there is no need to
2252          * register it again.
2253          */
2254         if (osnoise_instance_registered(tr))
2255                 return;
2256
2257         retval = osnoise_workload_start();
2258         if (retval)
2259                 pr_err(BANNER "Error starting osnoise tracer\n");
2260
2261         osnoise_register_instance(tr);
2262 }
2263
2264 static void osnoise_tracer_stop(struct trace_array *tr)
2265 {
2266         osnoise_unregister_instance(tr);
2267         osnoise_workload_stop();
2268 }
2269
2270 static int osnoise_tracer_init(struct trace_array *tr)
2271 {
2272         /*
2273          * Only allow osnoise tracer if timerlat tracer is not running
2274          * already.
2275          */
2276         if (timerlat_enabled())
2277                 return -EBUSY;
2278
2279         tr->max_latency = 0;
2280
2281         osnoise_tracer_start(tr);
2282         return 0;
2283 }
2284
2285 static void osnoise_tracer_reset(struct trace_array *tr)
2286 {
2287         osnoise_tracer_stop(tr);
2288 }
2289
2290 static struct tracer osnoise_tracer __read_mostly = {
2291         .name           = "osnoise",
2292         .init           = osnoise_tracer_init,
2293         .reset          = osnoise_tracer_reset,
2294         .start          = osnoise_tracer_start,
2295         .stop           = osnoise_tracer_stop,
2296         .print_header   = print_osnoise_headers,
2297         .allow_instances = true,
2298 };
2299
2300 #ifdef CONFIG_TIMERLAT_TRACER
2301 static void timerlat_tracer_start(struct trace_array *tr)
2302 {
2303         int retval;
2304
2305         /*
2306          * If the instance is already registered, there is no need to
2307          * register it again.
2308          */
2309         if (osnoise_instance_registered(tr))
2310                 return;
2311
2312         retval = osnoise_workload_start();
2313         if (retval)
2314                 pr_err(BANNER "Error starting timerlat tracer\n");
2315
2316         osnoise_register_instance(tr);
2317
2318         return;
2319 }
2320
2321 static void timerlat_tracer_stop(struct trace_array *tr)
2322 {
2323         int cpu;
2324
2325         osnoise_unregister_instance(tr);
2326
2327         /*
2328          * Instruct the threads to stop only if this is the last instance.
2329          */
2330         if (!osnoise_has_registered_instances()) {
2331                 for_each_online_cpu(cpu)
2332                         per_cpu(per_cpu_osnoise_var, cpu).sampling = 0;
2333         }
2334
2335         osnoise_workload_stop();
2336 }
2337
2338 static int timerlat_tracer_init(struct trace_array *tr)
2339 {
2340         /*
2341          * Only allow timerlat tracer if osnoise tracer is not running already.
2342          */
2343         if (osnoise_has_registered_instances() && !osnoise_data.timerlat_tracer)
2344                 return -EBUSY;
2345
2346         /*
2347          * If this is the first instance, set timerlat_tracer to block
2348          * osnoise tracer start.
2349          */
2350         if (!osnoise_has_registered_instances())
2351                 osnoise_data.timerlat_tracer = 1;
2352
2353         tr->max_latency = 0;
2354         timerlat_tracer_start(tr);
2355
2356         return 0;
2357 }
2358
2359 static void timerlat_tracer_reset(struct trace_array *tr)
2360 {
2361         timerlat_tracer_stop(tr);
2362
2363         /*
2364          * If this is the last instance, reset timerlat_tracer allowing
2365          * osnoise to be started.
2366          */
2367         if (!osnoise_has_registered_instances())
2368                 osnoise_data.timerlat_tracer = 0;
2369 }
2370
2371 static struct tracer timerlat_tracer __read_mostly = {
2372         .name           = "timerlat",
2373         .init           = timerlat_tracer_init,
2374         .reset          = timerlat_tracer_reset,
2375         .start          = timerlat_tracer_start,
2376         .stop           = timerlat_tracer_stop,
2377         .print_header   = print_timerlat_headers,
2378         .allow_instances = true,
2379 };
2380
2381 __init static int init_timerlat_tracer(void)
2382 {
2383         return register_tracer(&timerlat_tracer);
2384 }
2385 #else /* CONFIG_TIMERLAT_TRACER */
2386 __init static int init_timerlat_tracer(void)
2387 {
2388         return 0;
2389 }
2390 #endif /* CONFIG_TIMERLAT_TRACER */
2391
2392 __init static int init_osnoise_tracer(void)
2393 {
2394         int ret;
2395
2396         mutex_init(&interface_lock);
2397
2398         cpumask_copy(&osnoise_cpumask, cpu_all_mask);
2399
2400         ret = register_tracer(&osnoise_tracer);
2401         if (ret) {
2402                 pr_err(BANNER "Error registering osnoise!\n");
2403                 return ret;
2404         }
2405
2406         ret = init_timerlat_tracer();
2407         if (ret) {
2408                 pr_err(BANNER "Error registering timerlat!\n");
2409                 return ret;
2410         }
2411
2412         osnoise_init_hotplug_support();
2413
2414         INIT_LIST_HEAD_RCU(&osnoise_instances);
2415
2416         init_tracefs();
2417
2418         return 0;
2419 }
2420 late_initcall(init_osnoise_tracer);