1 // SPDX-License-Identifier: GPL-2.0
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.
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>
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.
16 * Copyright (C) 2021 Daniel Bristot de Oliveira, Red Hat, Inc. <bristot@redhat.com>
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>
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 */
35 #include <trace/events/irq.h>
36 #include <trace/events/sched.h>
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/osnoise.h>
44 #define BANNER "osnoise: "
45 #define DEFAULT_SAMPLE_PERIOD 1000000 /* 1s */
46 #define DEFAULT_SAMPLE_RUNTIME 1000000 /* 1s */
48 #define DEFAULT_TIMERLAT_PERIOD 1000 /* 1ms */
49 #define DEFAULT_TIMERLAT_PRIO 95 /* FIFO 95 */
52 * trace_array of the enabled osnoise/timerlat instances.
54 struct osnoise_instance {
55 struct list_head list;
56 struct trace_array *tr;
59 static struct list_head osnoise_instances;
61 static bool osnoise_has_registered_instances(void)
63 return !!list_first_or_null_rcu(&osnoise_instances,
64 struct osnoise_instance,
69 * osnoise_instance_registered - check if a tr is already registered
71 static int osnoise_instance_registered(struct trace_array *tr)
73 struct osnoise_instance *inst;
77 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
87 * osnoise_register_instance - register a new trace instance
89 * Register a trace_array *tr in the list of instances running
90 * osnoise/timerlat tracers.
92 static int osnoise_register_instance(struct trace_array *tr)
94 struct osnoise_instance *inst;
97 * register/unregister serialization is provided by trace's
100 lockdep_assert_held(&trace_types_lock);
102 inst = kmalloc(sizeof(*inst), GFP_KERNEL);
106 INIT_LIST_HEAD_RCU(&inst->list);
108 list_add_tail_rcu(&inst->list, &osnoise_instances);
114 * osnoise_unregister_instance - unregister a registered trace instance
116 * Remove the trace_array *tr from the list of instances running
117 * osnoise/timerlat tracers.
119 static void osnoise_unregister_instance(struct trace_array *tr)
121 struct osnoise_instance *inst;
125 * register/unregister serialization is provided by trace's
128 lockdep_assert_held(&trace_types_lock);
130 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
131 if (inst->tr == tr) {
132 list_del_rcu(&inst->list);
162 #define IRQ_CONTEXT 0
163 #define THREAD_CONTEXT 1
165 * sofirq runtime info.
174 * thread runtime info.
183 * Runtime information: this structure saves the runtime information used by
184 * one sampling thread.
186 struct osnoise_variables {
187 struct task_struct *kthread;
192 struct osn_softirq softirq;
193 struct osn_thread thread;
198 * Per-cpu runtime information.
200 DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var);
203 * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU
205 static inline struct osnoise_variables *this_cpu_osn_var(void)
207 return this_cpu_ptr(&per_cpu_osnoise_var);
210 #ifdef CONFIG_TIMERLAT_TRACER
212 * Runtime information for the timer mode.
214 struct timerlat_variables {
215 struct task_struct *kthread;
216 struct hrtimer timer;
223 DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var);
226 * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU
228 static inline struct timerlat_variables *this_cpu_tmr_var(void)
230 return this_cpu_ptr(&per_cpu_timerlat_var);
234 * tlat_var_reset - Reset the values of the given timerlat_variables
236 static inline void tlat_var_reset(void)
238 struct timerlat_variables *tlat_var;
241 * So far, all the values are initialized as 0, so
242 * zeroing the structure is perfect.
244 for_each_cpu(cpu, cpu_online_mask) {
245 tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu);
246 memset(tlat_var, 0, sizeof(*tlat_var));
249 #else /* CONFIG_TIMERLAT_TRACER */
250 #define tlat_var_reset() do {} while (0)
251 #endif /* CONFIG_TIMERLAT_TRACER */
254 * osn_var_reset - Reset the values of the given osnoise_variables
256 static inline void osn_var_reset(void)
258 struct osnoise_variables *osn_var;
262 * So far, all the values are initialized as 0, so
263 * zeroing the structure is perfect.
265 for_each_cpu(cpu, cpu_online_mask) {
266 osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu);
267 memset(osn_var, 0, sizeof(*osn_var));
272 * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables
274 static inline void osn_var_reset_all(void)
281 * Tells NMIs to call back to the osnoise tracer to record timestamps.
283 bool trace_osnoise_callback_enabled;
286 * osnoise sample structure definition. Used to store the statistics of a
289 struct osnoise_sample {
290 u64 runtime; /* runtime */
291 u64 noise; /* noise */
292 u64 max_sample; /* max single noise sample */
293 int hw_count; /* # HW (incl. hypervisor) interference */
294 int nmi_count; /* # NMIs during this sample */
295 int irq_count; /* # IRQs during this sample */
296 int softirq_count; /* # softirqs during this sample */
297 int thread_count; /* # threads during this sample */
300 #ifdef CONFIG_TIMERLAT_TRACER
302 * timerlat sample structure definition. Used to store the statistics of
305 struct timerlat_sample {
306 u64 timer_latency; /* timer_latency */
307 unsigned int seqnum; /* unique sequence */
308 int context; /* timer context */
313 * Protect the interface.
315 struct mutex interface_lock;
320 static struct osnoise_data {
321 u64 sample_period; /* total sampling period */
322 u64 sample_runtime; /* active sampling portion of period */
323 u64 stop_tracing; /* stop trace in the internal operation (loop/irq) */
324 u64 stop_tracing_total; /* stop trace in the final operation (report/thread) */
325 #ifdef CONFIG_TIMERLAT_TRACER
326 u64 timerlat_period; /* timerlat period */
327 u64 print_stack; /* print IRQ stack if total > */
328 int timerlat_tracer; /* timerlat tracer */
330 bool tainted; /* infor users and developers about a problem */
332 .sample_period = DEFAULT_SAMPLE_PERIOD,
333 .sample_runtime = DEFAULT_SAMPLE_RUNTIME,
335 .stop_tracing_total = 0,
336 #ifdef CONFIG_TIMERLAT_TRACER
338 .timerlat_period = DEFAULT_TIMERLAT_PERIOD,
339 .timerlat_tracer = 0,
343 #ifdef CONFIG_TIMERLAT_TRACER
344 static inline bool timerlat_enabled(void)
346 return osnoise_data.timerlat_tracer;
349 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
351 struct timerlat_variables *tlat_var = this_cpu_tmr_var();
353 * If the timerlat is enabled, but the irq handler did
354 * not run yet enabling timerlat_tracer, do not trace.
356 if (!tlat_var->tracing_thread) {
357 osn_var->softirq.arrival_time = 0;
358 osn_var->softirq.delta_start = 0;
364 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
366 struct timerlat_variables *tlat_var = this_cpu_tmr_var();
368 * If the timerlat is enabled, but the irq handler did
369 * not run yet enabling timerlat_tracer, do not trace.
371 if (!tlat_var->tracing_thread) {
372 osn_var->thread.delta_start = 0;
373 osn_var->thread.arrival_time = 0;
378 #else /* CONFIG_TIMERLAT_TRACER */
379 static inline bool timerlat_enabled(void)
384 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
388 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
394 #ifdef CONFIG_PREEMPT_RT
396 * Print the osnoise header info.
398 static void print_osnoise_headers(struct seq_file *s)
400 if (osnoise_data.tainted)
401 seq_puts(s, "# osnoise is tainted!\n");
403 seq_puts(s, "# _-------=> irqs-off\n");
404 seq_puts(s, "# / _------=> need-resched\n");
405 seq_puts(s, "# | / _-----=> need-resched-lazy\n");
406 seq_puts(s, "# || / _----=> hardirq/softirq\n");
407 seq_puts(s, "# ||| / _---=> preempt-depth\n");
408 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n");
409 seq_puts(s, "# ||||| / _-=> migrate-disable\n");
411 seq_puts(s, "# |||||| / ");
412 seq_puts(s, " MAX\n");
414 seq_puts(s, "# ||||| / ");
415 seq_puts(s, " SINGLE Interference counters:\n");
417 seq_puts(s, "# ||||||| RUNTIME ");
418 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n");
420 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP IN US ");
421 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n");
423 seq_puts(s, "# | | | ||||||| | | ");
424 seq_puts(s, " | | | | | | | |\n");
426 #else /* CONFIG_PREEMPT_RT */
427 static void print_osnoise_headers(struct seq_file *s)
429 if (osnoise_data.tainted)
430 seq_puts(s, "# osnoise is tainted!\n");
432 seq_puts(s, "# _-----=> irqs-off\n");
433 seq_puts(s, "# / _----=> need-resched\n");
434 seq_puts(s, "# | / _---=> hardirq/softirq\n");
435 seq_puts(s, "# || / _--=> preempt-depth\n");
436 seq_puts(s, "# ||| / _-=> migrate-disable ");
437 seq_puts(s, " MAX\n");
438 seq_puts(s, "# |||| / delay ");
439 seq_puts(s, " SINGLE Interference counters:\n");
441 seq_puts(s, "# ||||| RUNTIME ");
442 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n");
444 seq_puts(s, "# TASK-PID CPU# ||||| TIMESTAMP IN US ");
445 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n");
447 seq_puts(s, "# | | | ||||| | | ");
448 seq_puts(s, " | | | | | | | |\n");
450 #endif /* CONFIG_PREEMPT_RT */
453 * osnoise_taint - report an osnoise error.
455 #define osnoise_taint(msg) ({ \
456 struct osnoise_instance *inst; \
457 struct trace_buffer *buffer; \
460 list_for_each_entry_rcu(inst, &osnoise_instances, list) { \
461 buffer = inst->tr->array_buffer.buffer; \
462 trace_array_printk_buf(buffer, _THIS_IP_, msg); \
465 osnoise_data.tainted = true; \
469 * Record an osnoise_sample into the tracer buffer.
472 __trace_osnoise_sample(struct osnoise_sample *sample, struct trace_buffer *buffer)
474 struct trace_event_call *call = &event_osnoise;
475 struct ring_buffer_event *event;
476 struct osnoise_entry *entry;
478 event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry),
482 entry = ring_buffer_event_data(event);
483 entry->runtime = sample->runtime;
484 entry->noise = sample->noise;
485 entry->max_sample = sample->max_sample;
486 entry->hw_count = sample->hw_count;
487 entry->nmi_count = sample->nmi_count;
488 entry->irq_count = sample->irq_count;
489 entry->softirq_count = sample->softirq_count;
490 entry->thread_count = sample->thread_count;
492 if (!call_filter_check_discard(call, entry, buffer, event))
493 trace_buffer_unlock_commit_nostack(buffer, event);
497 * Record an osnoise_sample on all osnoise instances.
499 static void trace_osnoise_sample(struct osnoise_sample *sample)
501 struct osnoise_instance *inst;
502 struct trace_buffer *buffer;
505 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
506 buffer = inst->tr->array_buffer.buffer;
507 __trace_osnoise_sample(sample, buffer);
512 #ifdef CONFIG_TIMERLAT_TRACER
514 * Print the timerlat header info.
516 #ifdef CONFIG_PREEMPT_RT
517 static void print_timerlat_headers(struct seq_file *s)
519 seq_puts(s, "# _-------=> irqs-off\n");
520 seq_puts(s, "# / _------=> need-resched\n");
521 seq_puts(s, "# | / _-----=> need-resched-lazy\n");
522 seq_puts(s, "# || / _----=> hardirq/softirq\n");
523 seq_puts(s, "# ||| / _---=> preempt-depth\n");
524 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n");
525 seq_puts(s, "# ||||| / _-=> migrate-disable\n");
526 seq_puts(s, "# |||||| /\n");
527 seq_puts(s, "# ||||||| ACTIVATION\n");
528 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP ID ");
529 seq_puts(s, " CONTEXT LATENCY\n");
530 seq_puts(s, "# | | | ||||||| | | ");
531 seq_puts(s, " | |\n");
533 #else /* CONFIG_PREEMPT_RT */
534 static void print_timerlat_headers(struct seq_file *s)
536 seq_puts(s, "# _-----=> irqs-off\n");
537 seq_puts(s, "# / _----=> need-resched\n");
538 seq_puts(s, "# | / _---=> hardirq/softirq\n");
539 seq_puts(s, "# || / _--=> preempt-depth\n");
540 seq_puts(s, "# ||| / _-=> migrate-disable\n");
541 seq_puts(s, "# |||| / delay\n");
542 seq_puts(s, "# ||||| ACTIVATION\n");
543 seq_puts(s, "# TASK-PID CPU# ||||| TIMESTAMP ID ");
544 seq_puts(s, " CONTEXT LATENCY\n");
545 seq_puts(s, "# | | | ||||| | | ");
546 seq_puts(s, " | |\n");
548 #endif /* CONFIG_PREEMPT_RT */
551 __trace_timerlat_sample(struct timerlat_sample *sample, struct trace_buffer *buffer)
553 struct trace_event_call *call = &event_osnoise;
554 struct ring_buffer_event *event;
555 struct timerlat_entry *entry;
557 event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry),
561 entry = ring_buffer_event_data(event);
562 entry->seqnum = sample->seqnum;
563 entry->context = sample->context;
564 entry->timer_latency = sample->timer_latency;
566 if (!call_filter_check_discard(call, entry, buffer, event))
567 trace_buffer_unlock_commit_nostack(buffer, event);
571 * Record an timerlat_sample into the tracer buffer.
573 static void trace_timerlat_sample(struct timerlat_sample *sample)
575 struct osnoise_instance *inst;
576 struct trace_buffer *buffer;
579 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
580 buffer = inst->tr->array_buffer.buffer;
581 __trace_timerlat_sample(sample, buffer);
586 #ifdef CONFIG_STACKTRACE
588 #define MAX_CALLS 256
591 * Stack trace will take place only at IRQ level, so, no need
592 * to control nesting here.
597 unsigned long calls[MAX_CALLS];
600 static DEFINE_PER_CPU(struct trace_stack, trace_stack);
603 * timerlat_save_stack - save a stack trace without printing
605 * Save the current stack trace without printing. The
606 * stack will be printed later, after the end of the measurement.
608 static void timerlat_save_stack(int skip)
610 unsigned int size, nr_entries;
611 struct trace_stack *fstack;
613 fstack = this_cpu_ptr(&trace_stack);
615 size = ARRAY_SIZE(fstack->calls);
617 nr_entries = stack_trace_save(fstack->calls, size, skip);
619 fstack->stack_size = nr_entries * sizeof(unsigned long);
620 fstack->nr_entries = nr_entries;
627 __timerlat_dump_stack(struct trace_buffer *buffer, struct trace_stack *fstack, unsigned int size)
629 struct trace_event_call *call = &event_osnoise;
630 struct ring_buffer_event *event;
631 struct stack_entry *entry;
633 event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size,
638 entry = ring_buffer_event_data(event);
640 memcpy(&entry->caller, fstack->calls, size);
641 entry->size = fstack->nr_entries;
643 if (!call_filter_check_discard(call, entry, buffer, event))
644 trace_buffer_unlock_commit_nostack(buffer, event);
648 * timerlat_dump_stack - dump a stack trace previously saved
650 static void timerlat_dump_stack(u64 latency)
652 struct osnoise_instance *inst;
653 struct trace_buffer *buffer;
654 struct trace_stack *fstack;
658 * trace only if latency > print_stack config, if enabled.
660 if (!osnoise_data.print_stack || osnoise_data.print_stack > latency)
663 preempt_disable_notrace();
664 fstack = this_cpu_ptr(&trace_stack);
665 size = fstack->stack_size;
668 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
669 buffer = inst->tr->array_buffer.buffer;
670 __timerlat_dump_stack(buffer, fstack, size);
674 preempt_enable_notrace();
676 #else /* CONFIG_STACKTRACE */
677 #define timerlat_dump_stack(u64 latency) do {} while (0)
678 #define timerlat_save_stack(a) do {} while (0)
679 #endif /* CONFIG_STACKTRACE */
680 #endif /* CONFIG_TIMERLAT_TRACER */
683 * Macros to encapsulate the time capturing infrastructure.
685 #define time_get() trace_clock_local()
686 #define time_to_us(x) div_u64(x, 1000)
687 #define time_sub(a, b) ((a) - (b))
690 * cond_move_irq_delta_start - Forward the delta_start of a running IRQ
692 * If an IRQ is preempted by an NMI, its delta_start is pushed forward
693 * to discount the NMI interference.
695 * See get_int_safe_duration().
698 cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration)
700 if (osn_var->irq.delta_start)
701 osn_var->irq.delta_start += duration;
704 #ifndef CONFIG_PREEMPT_RT
706 * cond_move_softirq_delta_start - Forward the delta_start of a running softirq.
708 * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed
709 * forward to discount the interference.
711 * See get_int_safe_duration().
714 cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration)
716 if (osn_var->softirq.delta_start)
717 osn_var->softirq.delta_start += duration;
719 #else /* CONFIG_PREEMPT_RT */
720 #define cond_move_softirq_delta_start(osn_var, duration) do {} while (0)
724 * cond_move_thread_delta_start - Forward the delta_start of a running thread
726 * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start
727 * is pushed forward to discount the interference.
729 * See get_int_safe_duration().
732 cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration)
734 if (osn_var->thread.delta_start)
735 osn_var->thread.delta_start += duration;
739 * get_int_safe_duration - Get the duration of a window
741 * The irq, softirq and thread varaibles need to have its duration without
742 * the interference from higher priority interrupts. Instead of keeping a
743 * variable to discount the interrupt interference from these variables, the
744 * starting time of these variables are pushed forward with the interrupt's
745 * duration. In this way, a single variable is used to:
747 * - Know if a given window is being measured.
748 * - Account its duration.
749 * - Discount the interference.
751 * To avoid getting inconsistent values, e.g.,:
755 * delta_start -= int duration;
757 * duration = now - delta_start;
759 * result: negative duration if the variable duration before the
760 * interrupt was smaller than the interrupt execution.
762 * A counter of interrupts is used. If the counter increased, try
763 * to capture an interference safe duration.
766 get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start)
768 u64 int_counter, now;
772 int_counter = local_read(&osn_var->int_counter);
773 /* synchronize with interrupts */
777 duration = (now - *delta_start);
779 /* synchronize with interrupts */
781 } while (int_counter != local_read(&osn_var->int_counter));
784 * This is an evidence of race conditions that cause
785 * a value to be "discounted" too much.
788 osnoise_taint("Negative duration!\n");
797 * set_int_safe_time - Save the current time on *time, aware of interference
799 * Get the time, taking into consideration a possible interference from
800 * higher priority interrupts.
802 * See get_int_safe_duration() for an explanation.
805 set_int_safe_time(struct osnoise_variables *osn_var, u64 *time)
810 int_counter = local_read(&osn_var->int_counter);
811 /* synchronize with interrupts */
816 /* synchronize with interrupts */
818 } while (int_counter != local_read(&osn_var->int_counter));
823 #ifdef CONFIG_TIMERLAT_TRACER
825 * copy_int_safe_time - Copy *src into *desc aware of interference
828 copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src)
833 int_counter = local_read(&osn_var->int_counter);
834 /* synchronize with interrupts */
839 /* synchronize with interrupts */
841 } while (int_counter != local_read(&osn_var->int_counter));
845 #endif /* CONFIG_TIMERLAT_TRACER */
848 * trace_osnoise_callback - NMI entry/exit callback
850 * This function is called at the entry and exit NMI code. The bool enter
851 * distinguishes between either case. This function is used to note a NMI
852 * occurrence, compute the noise caused by the NMI, and to remove the noise
853 * it is potentially causing on other interference variables.
855 void trace_osnoise_callback(bool enter)
857 struct osnoise_variables *osn_var = this_cpu_osn_var();
860 if (!osn_var->sampling)
864 * Currently trace_clock_local() calls sched_clock() and the
865 * generic version is not NMI safe.
867 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
869 osn_var->nmi.delta_start = time_get();
870 local_inc(&osn_var->int_counter);
872 duration = time_get() - osn_var->nmi.delta_start;
874 trace_nmi_noise(osn_var->nmi.delta_start, duration);
876 cond_move_irq_delta_start(osn_var, duration);
877 cond_move_softirq_delta_start(osn_var, duration);
878 cond_move_thread_delta_start(osn_var, duration);
883 osn_var->nmi.count++;
887 * osnoise_trace_irq_entry - Note the starting of an IRQ
889 * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs,
890 * it is safe to use a single variable (ons_var->irq) to save the statistics.
891 * The arrival_time is used to report... the arrival time. The delta_start
892 * is used to compute the duration at the IRQ exit handler. See
893 * cond_move_irq_delta_start().
895 void osnoise_trace_irq_entry(int id)
897 struct osnoise_variables *osn_var = this_cpu_osn_var();
899 if (!osn_var->sampling)
902 * This value will be used in the report, but not to compute
903 * the execution time, so it is safe to get it unsafe.
905 osn_var->irq.arrival_time = time_get();
906 set_int_safe_time(osn_var, &osn_var->irq.delta_start);
907 osn_var->irq.count++;
909 local_inc(&osn_var->int_counter);
913 * osnoise_irq_exit - Note the end of an IRQ, sava data and trace
915 * Computes the duration of the IRQ noise, and trace it. Also discounts the
916 * interference from other sources of noise could be currently being accounted.
918 void osnoise_trace_irq_exit(int id, const char *desc)
920 struct osnoise_variables *osn_var = this_cpu_osn_var();
923 if (!osn_var->sampling)
926 duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start);
927 trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration);
928 osn_var->irq.arrival_time = 0;
929 cond_move_softirq_delta_start(osn_var, duration);
930 cond_move_thread_delta_start(osn_var, duration);
934 * trace_irqentry_callback - Callback to the irq:irq_entry traceevent
936 * Used to note the starting of an IRQ occurece.
938 static void trace_irqentry_callback(void *data, int irq,
939 struct irqaction *action)
941 osnoise_trace_irq_entry(irq);
945 * trace_irqexit_callback - Callback to the irq:irq_exit traceevent
947 * Used to note the end of an IRQ occurece.
949 static void trace_irqexit_callback(void *data, int irq,
950 struct irqaction *action, int ret)
952 osnoise_trace_irq_exit(irq, action->name);
956 * arch specific register function.
958 int __weak osnoise_arch_register(void)
964 * arch specific unregister function.
966 void __weak osnoise_arch_unregister(void)
972 * hook_irq_events - Hook IRQ handling events
974 * This function hooks the IRQ related callbacks to the respective trace
977 static int hook_irq_events(void)
981 ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL);
985 ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL);
987 goto out_unregister_entry;
989 ret = osnoise_arch_register();
996 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
997 out_unregister_entry:
998 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1004 * unhook_irq_events - Unhook IRQ handling events
1006 * This function unhooks the IRQ related callbacks to the respective trace
1009 static void unhook_irq_events(void)
1011 osnoise_arch_unregister();
1012 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
1013 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1016 #ifndef CONFIG_PREEMPT_RT
1018 * trace_softirq_entry_callback - Note the starting of a softirq
1020 * Save the starting time of a softirq. As softirqs are non-preemptive to
1021 * other softirqs, it is safe to use a single variable (ons_var->softirq)
1022 * to save the statistics. The arrival_time is used to report... the
1023 * arrival time. The delta_start is used to compute the duration at the
1024 * softirq exit handler. See cond_move_softirq_delta_start().
1026 static void trace_softirq_entry_callback(void *data, unsigned int vec_nr)
1028 struct osnoise_variables *osn_var = this_cpu_osn_var();
1030 if (!osn_var->sampling)
1033 * This value will be used in the report, but not to compute
1034 * the execution time, so it is safe to get it unsafe.
1036 osn_var->softirq.arrival_time = time_get();
1037 set_int_safe_time(osn_var, &osn_var->softirq.delta_start);
1038 osn_var->softirq.count++;
1040 local_inc(&osn_var->int_counter);
1044 * trace_softirq_exit_callback - Note the end of an softirq
1046 * Computes the duration of the softirq noise, and trace it. Also discounts the
1047 * interference from other sources of noise could be currently being accounted.
1049 static void trace_softirq_exit_callback(void *data, unsigned int vec_nr)
1051 struct osnoise_variables *osn_var = this_cpu_osn_var();
1054 if (!osn_var->sampling)
1057 if (unlikely(timerlat_enabled()))
1058 if (!timerlat_softirq_exit(osn_var))
1061 duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start);
1062 trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration);
1063 cond_move_thread_delta_start(osn_var, duration);
1064 osn_var->softirq.arrival_time = 0;
1068 * hook_softirq_events - Hook softirq handling events
1070 * This function hooks the softirq related callbacks to the respective trace
1073 static int hook_softirq_events(void)
1077 ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1081 ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1083 goto out_unreg_entry;
1088 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1094 * unhook_softirq_events - Unhook softirq handling events
1096 * This function hooks the softirq related callbacks to the respective trace
1099 static void unhook_softirq_events(void)
1101 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1102 unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1104 #else /* CONFIG_PREEMPT_RT */
1106 * softirq are threads on the PREEMPT_RT mode.
1108 static int hook_softirq_events(void)
1112 static void unhook_softirq_events(void)
1118 * thread_entry - Record the starting of a thread noise window
1120 * It saves the context switch time for a noisy thread, and increments
1121 * the interference counters.
1124 thread_entry(struct osnoise_variables *osn_var, struct task_struct *t)
1126 if (!osn_var->sampling)
1129 * The arrival time will be used in the report, but not to compute
1130 * the execution time, so it is safe to get it unsafe.
1132 osn_var->thread.arrival_time = time_get();
1134 set_int_safe_time(osn_var, &osn_var->thread.delta_start);
1136 osn_var->thread.count++;
1137 local_inc(&osn_var->int_counter);
1141 * thread_exit - Report the end of a thread noise window
1143 * It computes the total noise from a thread, tracing if needed.
1146 thread_exit(struct osnoise_variables *osn_var, struct task_struct *t)
1150 if (!osn_var->sampling)
1153 if (unlikely(timerlat_enabled()))
1154 if (!timerlat_thread_exit(osn_var))
1157 duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start);
1159 trace_thread_noise(t, osn_var->thread.arrival_time, duration);
1161 osn_var->thread.arrival_time = 0;
1165 * trace_sched_switch - sched:sched_switch trace event handler
1167 * This function is hooked to the sched:sched_switch trace event, and it is
1168 * used to record the beginning and to report the end of a thread noise window.
1171 trace_sched_switch_callback(void *data, bool preempt, struct task_struct *p,
1172 struct task_struct *n)
1174 struct osnoise_variables *osn_var = this_cpu_osn_var();
1176 if (p->pid != osn_var->pid)
1177 thread_exit(osn_var, p);
1179 if (n->pid != osn_var->pid)
1180 thread_entry(osn_var, n);
1184 * hook_thread_events - Hook the insturmentation for thread noise
1186 * Hook the osnoise tracer callbacks to handle the noise from other
1187 * threads on the necessary kernel events.
1189 static int hook_thread_events(void)
1193 ret = register_trace_sched_switch(trace_sched_switch_callback, NULL);
1201 * unhook_thread_events - *nhook the insturmentation for thread noise
1203 * Unook the osnoise tracer callbacks to handle the noise from other
1204 * threads on the necessary kernel events.
1206 static void unhook_thread_events(void)
1208 unregister_trace_sched_switch(trace_sched_switch_callback, NULL);
1212 * save_osn_sample_stats - Save the osnoise_sample statistics
1214 * Save the osnoise_sample statistics before the sampling phase. These
1215 * values will be used later to compute the diff betwneen the statistics
1216 * before and after the osnoise sampling.
1219 save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1221 s->nmi_count = osn_var->nmi.count;
1222 s->irq_count = osn_var->irq.count;
1223 s->softirq_count = osn_var->softirq.count;
1224 s->thread_count = osn_var->thread.count;
1228 * diff_osn_sample_stats - Compute the osnoise_sample statistics
1230 * After a sample period, compute the difference on the osnoise_sample
1231 * statistics. The struct osnoise_sample *s contains the statistics saved via
1232 * save_osn_sample_stats() before the osnoise sampling.
1235 diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1237 s->nmi_count = osn_var->nmi.count - s->nmi_count;
1238 s->irq_count = osn_var->irq.count - s->irq_count;
1239 s->softirq_count = osn_var->softirq.count - s->softirq_count;
1240 s->thread_count = osn_var->thread.count - s->thread_count;
1244 * osnoise_stop_tracing - Stop tracing and the tracer.
1246 static __always_inline void osnoise_stop_tracing(void)
1248 struct osnoise_instance *inst;
1249 struct trace_array *tr;
1252 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1254 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_,
1255 "stop tracing hit on cpu %d\n", smp_processor_id());
1257 tracer_tracing_off(tr);
1263 * notify_new_max_latency - Notify a new max latency via fsnotify interface.
1265 static void notify_new_max_latency(u64 latency)
1267 struct osnoise_instance *inst;
1268 struct trace_array *tr;
1271 list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1273 if (tr->max_latency < latency) {
1274 tr->max_latency = latency;
1275 latency_fsnotify(tr);
1282 * run_osnoise - Sample the time and look for osnoise
1284 * Used to capture the time, looking for potential osnoise latency repeatedly.
1285 * Different from hwlat_detector, it is called with preemption and interrupts
1286 * enabled. This allows irqs, softirqs and threads to run, interfering on the
1287 * osnoise sampling thread, as they would do with a regular thread.
1289 static int run_osnoise(void)
1291 struct osnoise_variables *osn_var = this_cpu_osn_var();
1292 u64 start, sample, last_sample;
1293 u64 last_int_count, int_count;
1294 s64 noise = 0, max_noise = 0;
1295 s64 total, last_total = 0;
1296 struct osnoise_sample s;
1297 unsigned int threshold;
1298 u64 runtime, stop_in;
1304 * Considers the current thread as the workload.
1306 osn_var->pid = current->pid;
1309 * Save the current stats for the diff
1311 save_osn_sample_stats(osn_var, &s);
1314 * if threshold is 0, use the default value of 5 us.
1316 threshold = tracing_thresh ? : 5000;
1319 * Make sure NMIs see sampling first
1321 osn_var->sampling = true;
1325 * Transform the *_us config to nanoseconds to avoid the
1326 * division on the main loop.
1328 runtime = osnoise_data.sample_runtime * NSEC_PER_USEC;
1329 stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC;
1339 last_int_count = set_int_safe_time(osn_var, &last_sample);
1345 int_count = set_int_safe_time(osn_var, &sample);
1347 noise = time_sub(sample, last_sample);
1350 * This shouldn't happen.
1353 osnoise_taint("negative noise!");
1360 total = time_sub(sample, start);
1363 * Check for possible overflows.
1365 if (total < last_total) {
1366 osnoise_taint("total overflow!");
1372 if (noise >= threshold) {
1373 int interference = int_count - last_int_count;
1375 if (noise > max_noise)
1383 trace_sample_threshold(last_sample, noise, interference);
1385 if (osnoise_data.stop_tracing)
1386 if (noise > stop_in)
1387 osnoise_stop_tracing();
1391 * For the non-preemptive kernel config: let threads runs, if
1396 last_sample = sample;
1397 last_int_count = int_count;
1399 } while (total < runtime && !kthread_should_stop());
1402 * Finish the above in the view for interrupts.
1406 osn_var->sampling = false;
1409 * Make sure sampling data is no longer updated.
1416 s.noise = time_to_us(sum_noise);
1417 s.runtime = time_to_us(total);
1418 s.max_sample = time_to_us(max_noise);
1419 s.hw_count = hw_count;
1421 /* Save interference stats info */
1422 diff_osn_sample_stats(osn_var, &s);
1424 trace_osnoise_sample(&s);
1426 notify_new_max_latency(max_noise);
1428 if (osnoise_data.stop_tracing_total)
1429 if (s.noise > osnoise_data.stop_tracing_total)
1430 osnoise_stop_tracing();
1437 static struct cpumask osnoise_cpumask;
1438 static struct cpumask save_cpumask;
1441 * osnoise_main - The osnoise detection kernel thread
1443 * Calls run_osnoise() function to measure the osnoise for the configured runtime,
1446 static int osnoise_main(void *data)
1450 while (!kthread_should_stop()) {
1454 mutex_lock(&interface_lock);
1455 interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
1456 mutex_unlock(&interface_lock);
1458 do_div(interval, USEC_PER_MSEC);
1461 * differently from hwlat_detector, the osnoise tracer can run
1462 * without a pause because preemption is on.
1465 /* Let synchronize_rcu_tasks() make progress */
1466 cond_resched_tasks_rcu_qs();
1470 if (msleep_interruptible(interval))
1477 #ifdef CONFIG_TIMERLAT_TRACER
1479 * timerlat_irq - hrtimer handler for timerlat.
1481 static enum hrtimer_restart timerlat_irq(struct hrtimer *timer)
1483 struct osnoise_variables *osn_var = this_cpu_osn_var();
1484 struct timerlat_variables *tlat;
1485 struct timerlat_sample s;
1490 * I am not sure if the timer was armed for this CPU. So, get
1491 * the timerlat struct from the timer itself, not from this
1494 tlat = container_of(timer, struct timerlat_variables, timer);
1496 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1499 * Enable the osnoise: events for thread an softirq.
1501 tlat->tracing_thread = true;
1503 osn_var->thread.arrival_time = time_get();
1506 * A hardirq is running: the timer IRQ. It is for sure preempting
1507 * a thread, and potentially preempting a softirq.
1509 * At this point, it is not interesting to know the duration of the
1510 * preempted thread (and maybe softirq), but how much time they will
1511 * delay the beginning of the execution of the timer thread.
1513 * To get the correct (net) delay added by the softirq, its delta_start
1514 * is set as the IRQ one. In this way, at the return of the IRQ, the delta
1515 * start of the sofitrq will be zeroed, accounting then only the time
1518 * The thread follows the same principle. However, if a softirq is
1519 * running, the thread needs to receive the softirq delta_start. The
1520 * reason being is that the softirq will be the last to be unfolded,
1521 * resseting the thread delay to zero.
1523 * The PREEMPT_RT is a special case, though. As softirqs run as threads
1524 * on RT, moving the thread is enough.
1526 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && osn_var->softirq.delta_start) {
1527 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1528 &osn_var->softirq.delta_start);
1530 copy_int_safe_time(osn_var, &osn_var->softirq.delta_start,
1531 &osn_var->irq.delta_start);
1533 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1534 &osn_var->irq.delta_start);
1538 * Compute the current time with the expected time.
1540 diff = now - tlat->abs_period;
1543 s.seqnum = tlat->count;
1544 s.timer_latency = diff;
1545 s.context = IRQ_CONTEXT;
1547 trace_timerlat_sample(&s);
1549 notify_new_max_latency(diff);
1551 if (osnoise_data.stop_tracing)
1552 if (time_to_us(diff) >= osnoise_data.stop_tracing)
1553 osnoise_stop_tracing();
1555 wake_up_process(tlat->kthread);
1557 if (osnoise_data.print_stack)
1558 timerlat_save_stack(0);
1560 return HRTIMER_NORESTART;
1564 * wait_next_period - Wait for the next period for timerlat
1566 static int wait_next_period(struct timerlat_variables *tlat)
1568 ktime_t next_abs_period, now;
1569 u64 rel_period = osnoise_data.timerlat_period * 1000;
1571 now = hrtimer_cb_get_time(&tlat->timer);
1572 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1575 * Save the next abs_period.
1577 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1580 * If the new abs_period is in the past, skip the activation.
1582 while (ktime_compare(now, next_abs_period) > 0) {
1583 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1584 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1587 set_current_state(TASK_INTERRUPTIBLE);
1589 hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD);
1595 * timerlat_main- Timerlat main
1597 static int timerlat_main(void *data)
1599 struct osnoise_variables *osn_var = this_cpu_osn_var();
1600 struct timerlat_variables *tlat = this_cpu_tmr_var();
1601 struct timerlat_sample s;
1602 struct sched_param sp;
1606 * Make the thread RT, that is how cyclictest is usually used.
1608 sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
1609 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1612 tlat->tracing_thread = false;
1614 hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1615 tlat->timer.function = timerlat_irq;
1616 tlat->kthread = current;
1617 osn_var->pid = current->pid;
1619 * Anotate the arrival time.
1621 tlat->abs_period = hrtimer_cb_get_time(&tlat->timer);
1623 wait_next_period(tlat);
1625 osn_var->sampling = 1;
1627 while (!kthread_should_stop()) {
1628 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1629 diff = now - tlat->abs_period;
1631 s.seqnum = tlat->count;
1632 s.timer_latency = diff;
1633 s.context = THREAD_CONTEXT;
1635 trace_timerlat_sample(&s);
1637 timerlat_dump_stack(time_to_us(diff));
1639 tlat->tracing_thread = false;
1640 if (osnoise_data.stop_tracing_total)
1641 if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
1642 osnoise_stop_tracing();
1644 wait_next_period(tlat);
1647 hrtimer_cancel(&tlat->timer);
1650 #else /* CONFIG_TIMERLAT_TRACER */
1651 static int timerlat_main(void *data)
1655 #endif /* CONFIG_TIMERLAT_TRACER */
1658 * stop_kthread - stop a workload thread
1660 static void stop_kthread(unsigned int cpu)
1662 struct task_struct *kthread;
1664 kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread;
1666 kthread_stop(kthread);
1667 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1671 * stop_per_cpu_kthread - Stop per-cpu threads
1673 * Stop the osnoise sampling htread. Use this on unload and at system
1676 static void stop_per_cpu_kthreads(void)
1682 for_each_online_cpu(cpu)
1689 * start_kthread - Start a workload tread
1691 static int start_kthread(unsigned int cpu)
1693 struct task_struct *kthread;
1694 void *main = osnoise_main;
1697 if (timerlat_enabled()) {
1698 snprintf(comm, 24, "timerlat/%d", cpu);
1699 main = timerlat_main;
1701 snprintf(comm, 24, "osnoise/%d", cpu);
1704 kthread = kthread_create_on_cpu(main, NULL, cpu, comm);
1706 if (IS_ERR(kthread)) {
1707 pr_err(BANNER "could not start sampling thread\n");
1708 stop_per_cpu_kthreads();
1712 per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread;
1713 wake_up_process(kthread);
1719 * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads
1721 * This starts the kernel thread that will look for osnoise on many
1724 static int start_per_cpu_kthreads(void)
1726 struct cpumask *current_mask = &save_cpumask;
1732 * Run only on online CPUs in which osnoise is allowed to run.
1734 cpumask_and(current_mask, cpu_online_mask, &osnoise_cpumask);
1736 for_each_possible_cpu(cpu)
1737 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1739 for_each_cpu(cpu, current_mask) {
1740 retval = start_kthread(cpu);
1742 stop_per_cpu_kthreads();
1752 #ifdef CONFIG_HOTPLUG_CPU
1753 static void osnoise_hotplug_workfn(struct work_struct *dummy)
1755 unsigned int cpu = smp_processor_id();
1757 mutex_lock(&trace_types_lock);
1759 if (!osnoise_has_registered_instances())
1760 goto out_unlock_trace;
1762 mutex_lock(&interface_lock);
1765 if (!cpumask_test_cpu(cpu, &osnoise_cpumask))
1772 mutex_unlock(&interface_lock);
1774 mutex_unlock(&trace_types_lock);
1777 static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn);
1780 * osnoise_cpu_init - CPU hotplug online callback function
1782 static int osnoise_cpu_init(unsigned int cpu)
1784 schedule_work_on(cpu, &osnoise_hotplug_work);
1789 * osnoise_cpu_die - CPU hotplug offline callback function
1791 static int osnoise_cpu_die(unsigned int cpu)
1797 static void osnoise_init_hotplug_support(void)
1801 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online",
1802 osnoise_cpu_init, osnoise_cpu_die);
1804 pr_warn(BANNER "Error to init cpu hotplug support\n");
1808 #else /* CONFIG_HOTPLUG_CPU */
1809 static void osnoise_init_hotplug_support(void)
1813 #endif /* CONFIG_HOTPLUG_CPU */
1816 * osnoise_cpus_read - Read function for reading the "cpus" file
1817 * @filp: The active open file structure
1818 * @ubuf: The userspace provided buffer to read value into
1819 * @cnt: The maximum number of bytes to read
1820 * @ppos: The current "file" position
1822 * Prints the "cpus" output into the user-provided buffer.
1825 osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count,
1831 mutex_lock(&interface_lock);
1833 len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1;
1834 mask_str = kmalloc(len, GFP_KERNEL);
1840 len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask));
1846 count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
1851 mutex_unlock(&interface_lock);
1857 * osnoise_cpus_write - Write function for "cpus" entry
1858 * @filp: The active open file structure
1859 * @ubuf: The user buffer that contains the value to write
1860 * @cnt: The maximum number of bytes to write to "file"
1861 * @ppos: The current position in @file
1863 * This function provides a write implementation for the "cpus"
1864 * interface to the osnoise trace. By default, it lists all CPUs,
1865 * in this way, allowing osnoise threads to run on any online CPU
1866 * of the system. It serves to restrict the execution of osnoise to the
1867 * set of CPUs writing via this interface. Why not use "tracing_cpumask"?
1868 * Because the user might be interested in tracing what is running on
1869 * other CPUs. For instance, one might run osnoise in one HT CPU
1870 * while observing what is running on the sibling HT CPU.
1873 osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
1876 cpumask_var_t osnoise_cpumask_new;
1883 if (copy_from_user(buf, ubuf, count))
1886 if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL))
1889 err = cpulist_parse(buf, osnoise_cpumask_new);
1894 * trace_types_lock is taken to avoid concurrency on start/stop.
1896 mutex_lock(&trace_types_lock);
1897 running = osnoise_has_registered_instances();
1899 stop_per_cpu_kthreads();
1901 mutex_lock(&interface_lock);
1903 * osnoise_cpumask is read by CPU hotplug operations.
1907 cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
1910 mutex_unlock(&interface_lock);
1913 start_per_cpu_kthreads();
1914 mutex_unlock(&trace_types_lock);
1916 free_cpumask_var(osnoise_cpumask_new);
1920 free_cpumask_var(osnoise_cpumask_new);
1926 * osnoise/runtime_us: cannot be greater than the period.
1928 static struct trace_min_max_param osnoise_runtime = {
1929 .lock = &interface_lock,
1930 .val = &osnoise_data.sample_runtime,
1931 .max = &osnoise_data.sample_period,
1936 * osnoise/period_us: cannot be smaller than the runtime.
1938 static struct trace_min_max_param osnoise_period = {
1939 .lock = &interface_lock,
1940 .val = &osnoise_data.sample_period,
1942 .min = &osnoise_data.sample_runtime,
1946 * osnoise/stop_tracing_us: no limit.
1948 static struct trace_min_max_param osnoise_stop_tracing_in = {
1949 .lock = &interface_lock,
1950 .val = &osnoise_data.stop_tracing,
1956 * osnoise/stop_tracing_total_us: no limit.
1958 static struct trace_min_max_param osnoise_stop_tracing_total = {
1959 .lock = &interface_lock,
1960 .val = &osnoise_data.stop_tracing_total,
1965 #ifdef CONFIG_TIMERLAT_TRACER
1967 * osnoise/print_stack: print the stacktrace of the IRQ handler if the total
1968 * latency is higher than val.
1970 static struct trace_min_max_param osnoise_print_stack = {
1971 .lock = &interface_lock,
1972 .val = &osnoise_data.print_stack,
1978 * osnoise/timerlat_period: min 100 us, max 1 s
1980 u64 timerlat_min_period = 100;
1981 u64 timerlat_max_period = 1000000;
1982 static struct trace_min_max_param timerlat_period = {
1983 .lock = &interface_lock,
1984 .val = &osnoise_data.timerlat_period,
1985 .max = &timerlat_max_period,
1986 .min = &timerlat_min_period,
1990 static const struct file_operations cpus_fops = {
1991 .open = tracing_open_generic,
1992 .read = osnoise_cpus_read,
1993 .write = osnoise_cpus_write,
1994 .llseek = generic_file_llseek,
1997 #ifdef CONFIG_TIMERLAT_TRACER
1998 #ifdef CONFIG_STACKTRACE
1999 static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2003 tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir,
2004 &osnoise_print_stack, &trace_min_max_fops);
2010 #else /* CONFIG_STACKTRACE */
2011 static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2015 #endif /* CONFIG_STACKTRACE */
2018 * init_timerlat_tracefs - A function to initialize the timerlat interface files
2020 static int init_timerlat_tracefs(struct dentry *top_dir)
2024 tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir,
2025 &timerlat_period, &trace_min_max_fops);
2029 return init_timerlat_stack_tracefs(top_dir);
2031 #else /* CONFIG_TIMERLAT_TRACER */
2032 static int init_timerlat_tracefs(struct dentry *top_dir)
2036 #endif /* CONFIG_TIMERLAT_TRACER */
2039 * init_tracefs - A function to initialize the tracefs interface files
2041 * This function creates entries in tracefs for "osnoise" and "timerlat".
2042 * It creates these directories in the tracing directory, and within that
2043 * directory the use can change and view the configs.
2045 static int init_tracefs(void)
2047 struct dentry *top_dir;
2051 ret = tracing_init_dentry();
2055 top_dir = tracefs_create_dir("osnoise", NULL);
2059 tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir,
2060 &osnoise_period, &trace_min_max_fops);
2064 tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir,
2065 &osnoise_runtime, &trace_min_max_fops);
2069 tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir,
2070 &osnoise_stop_tracing_in, &trace_min_max_fops);
2074 tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir,
2075 &osnoise_stop_tracing_total, &trace_min_max_fops);
2079 tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops);
2083 ret = init_timerlat_tracefs(top_dir);
2090 tracefs_remove(top_dir);
2094 static int osnoise_hook_events(void)
2099 * Trace is already hooked, we are re-enabling from
2102 if (trace_osnoise_callback_enabled)
2105 retval = hook_irq_events();
2109 retval = hook_softirq_events();
2111 goto out_unhook_irq;
2113 retval = hook_thread_events();
2120 unhook_softirq_events();
2122 unhook_irq_events();
2127 * osnoise_workload_start - start the workload and hook to events
2129 static int osnoise_workload_start(void)
2134 * Instances need to be registered after calling workload
2135 * start. Hence, if there is already an instance, the
2136 * workload was already registered. Otherwise, this
2137 * code is on the way to register the first instance,
2138 * and the workload will start.
2140 if (osnoise_has_registered_instances())
2143 osn_var_reset_all();
2145 retval = osnoise_hook_events();
2150 * Make sure that ftrace_nmi_enter/exit() see reset values
2151 * before enabling trace_osnoise_callback_enabled.
2154 trace_osnoise_callback_enabled = true;
2156 retval = start_per_cpu_kthreads();
2158 unhook_irq_events();
2166 * osnoise_workload_stop - stop the workload and unhook the events
2168 static void osnoise_workload_stop(void)
2171 * Instances need to be unregistered before calling
2172 * stop. Hence, if there is a registered instance, more
2173 * than one instance is running, and the workload will not
2174 * yet stop. Otherwise, this code is on the way to disable
2175 * the last instance, and the workload can stop.
2177 if (osnoise_has_registered_instances())
2180 trace_osnoise_callback_enabled = false;
2182 * Make sure that ftrace_nmi_enter/exit() see
2183 * trace_osnoise_callback_enabled as false before continuing.
2187 stop_per_cpu_kthreads();
2189 unhook_irq_events();
2190 unhook_softirq_events();
2191 unhook_thread_events();
2194 static void osnoise_tracer_start(struct trace_array *tr)
2199 * If the instance is already registered, there is no need to
2200 * register it again.
2202 if (osnoise_instance_registered(tr))
2205 retval = osnoise_workload_start();
2207 pr_err(BANNER "Error starting osnoise tracer\n");
2209 osnoise_register_instance(tr);
2212 static void osnoise_tracer_stop(struct trace_array *tr)
2214 osnoise_unregister_instance(tr);
2215 osnoise_workload_stop();
2218 static int osnoise_tracer_init(struct trace_array *tr)
2221 * Only allow osnoise tracer if timerlat tracer is not running
2224 if (timerlat_enabled())
2227 tr->max_latency = 0;
2229 osnoise_tracer_start(tr);
2233 static void osnoise_tracer_reset(struct trace_array *tr)
2235 osnoise_tracer_stop(tr);
2238 static struct tracer osnoise_tracer __read_mostly = {
2240 .init = osnoise_tracer_init,
2241 .reset = osnoise_tracer_reset,
2242 .start = osnoise_tracer_start,
2243 .stop = osnoise_tracer_stop,
2244 .print_header = print_osnoise_headers,
2245 .allow_instances = true,
2248 #ifdef CONFIG_TIMERLAT_TRACER
2249 static void timerlat_tracer_start(struct trace_array *tr)
2254 * If the instance is already registered, there is no need to
2255 * register it again.
2257 if (osnoise_instance_registered(tr))
2260 retval = osnoise_workload_start();
2262 pr_err(BANNER "Error starting timerlat tracer\n");
2264 osnoise_register_instance(tr);
2269 static void timerlat_tracer_stop(struct trace_array *tr)
2273 osnoise_unregister_instance(tr);
2276 * Instruct the threads to stop only if this is the last instance.
2278 if (!osnoise_has_registered_instances()) {
2279 for_each_online_cpu(cpu)
2280 per_cpu(per_cpu_osnoise_var, cpu).sampling = 0;
2283 osnoise_workload_stop();
2286 static int timerlat_tracer_init(struct trace_array *tr)
2289 * Only allow timerlat tracer if osnoise tracer is not running already.
2291 if (osnoise_has_registered_instances() && !osnoise_data.timerlat_tracer)
2295 * If this is the first instance, set timerlat_tracer to block
2296 * osnoise tracer start.
2298 if (!osnoise_has_registered_instances())
2299 osnoise_data.timerlat_tracer = 1;
2301 tr->max_latency = 0;
2302 timerlat_tracer_start(tr);
2307 static void timerlat_tracer_reset(struct trace_array *tr)
2309 timerlat_tracer_stop(tr);
2312 * If this is the last instance, reset timerlat_tracer allowing
2313 * osnoise to be started.
2315 if (!osnoise_has_registered_instances())
2316 osnoise_data.timerlat_tracer = 0;
2319 static struct tracer timerlat_tracer __read_mostly = {
2321 .init = timerlat_tracer_init,
2322 .reset = timerlat_tracer_reset,
2323 .start = timerlat_tracer_start,
2324 .stop = timerlat_tracer_stop,
2325 .print_header = print_timerlat_headers,
2326 .allow_instances = true,
2329 __init static int init_timerlat_tracer(void)
2331 return register_tracer(&timerlat_tracer);
2333 #else /* CONFIG_TIMERLAT_TRACER */
2334 __init static int init_timerlat_tracer(void)
2338 #endif /* CONFIG_TIMERLAT_TRACER */
2340 __init static int init_osnoise_tracer(void)
2344 mutex_init(&interface_lock);
2346 cpumask_copy(&osnoise_cpumask, cpu_all_mask);
2348 ret = register_tracer(&osnoise_tracer);
2350 pr_err(BANNER "Error registering osnoise!\n");
2354 ret = init_timerlat_tracer();
2356 pr_err(BANNER "Error registering timerlat!\n");
2360 osnoise_init_hotplug_support();
2362 INIT_LIST_HEAD_RCU(&osnoise_instances);
2368 late_initcall(init_osnoise_tracer);