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>
41 static struct trace_array *osnoise_trace;
46 #define BANNER "osnoise: "
47 #define DEFAULT_SAMPLE_PERIOD 1000000 /* 1s */
48 #define DEFAULT_SAMPLE_RUNTIME 1000000 /* 1s */
50 #define DEFAULT_TIMERLAT_PERIOD 1000 /* 1ms */
51 #define DEFAULT_TIMERLAT_PRIO 95 /* FIFO 95 */
71 #define THREAD_CONTEXT 1
73 * sofirq runtime info.
82 * thread runtime info.
91 * Runtime information: this structure saves the runtime information used by
92 * one sampling thread.
94 struct osnoise_variables {
95 struct task_struct *kthread;
100 struct osn_softirq softirq;
101 struct osn_thread thread;
106 * Per-cpu runtime information.
108 DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var);
111 * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU
113 static inline struct osnoise_variables *this_cpu_osn_var(void)
115 return this_cpu_ptr(&per_cpu_osnoise_var);
118 #ifdef CONFIG_TIMERLAT_TRACER
120 * Runtime information for the timer mode.
122 struct timerlat_variables {
123 struct task_struct *kthread;
124 struct hrtimer timer;
131 DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var);
134 * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU
136 static inline struct timerlat_variables *this_cpu_tmr_var(void)
138 return this_cpu_ptr(&per_cpu_timerlat_var);
142 * tlat_var_reset - Reset the values of the given timerlat_variables
144 static inline void tlat_var_reset(void)
146 struct timerlat_variables *tlat_var;
149 * So far, all the values are initialized as 0, so
150 * zeroing the structure is perfect.
152 for_each_cpu(cpu, cpu_online_mask) {
153 tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu);
154 memset(tlat_var, 0, sizeof(*tlat_var));
157 #else /* CONFIG_TIMERLAT_TRACER */
158 #define tlat_var_reset() do {} while (0)
159 #endif /* CONFIG_TIMERLAT_TRACER */
162 * osn_var_reset - Reset the values of the given osnoise_variables
164 static inline void osn_var_reset(void)
166 struct osnoise_variables *osn_var;
170 * So far, all the values are initialized as 0, so
171 * zeroing the structure is perfect.
173 for_each_cpu(cpu, cpu_online_mask) {
174 osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu);
175 memset(osn_var, 0, sizeof(*osn_var));
180 * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables
182 static inline void osn_var_reset_all(void)
189 * Tells NMIs to call back to the osnoise tracer to record timestamps.
191 bool trace_osnoise_callback_enabled;
194 * osnoise sample structure definition. Used to store the statistics of a
197 struct osnoise_sample {
198 u64 runtime; /* runtime */
199 u64 noise; /* noise */
200 u64 max_sample; /* max single noise sample */
201 int hw_count; /* # HW (incl. hypervisor) interference */
202 int nmi_count; /* # NMIs during this sample */
203 int irq_count; /* # IRQs during this sample */
204 int softirq_count; /* # softirqs during this sample */
205 int thread_count; /* # threads during this sample */
208 #ifdef CONFIG_TIMERLAT_TRACER
210 * timerlat sample structure definition. Used to store the statistics of
213 struct timerlat_sample {
214 u64 timer_latency; /* timer_latency */
215 unsigned int seqnum; /* unique sequence */
216 int context; /* timer context */
221 * Protect the interface.
223 struct mutex interface_lock;
228 static struct osnoise_data {
229 u64 sample_period; /* total sampling period */
230 u64 sample_runtime; /* active sampling portion of period */
231 u64 stop_tracing; /* stop trace in the internal operation (loop/irq) */
232 u64 stop_tracing_total; /* stop trace in the final operation (report/thread) */
233 #ifdef CONFIG_TIMERLAT_TRACER
234 u64 timerlat_period; /* timerlat period */
235 u64 print_stack; /* print IRQ stack if total > */
236 int timerlat_tracer; /* timerlat tracer */
238 bool tainted; /* infor users and developers about a problem */
240 .sample_period = DEFAULT_SAMPLE_PERIOD,
241 .sample_runtime = DEFAULT_SAMPLE_RUNTIME,
243 .stop_tracing_total = 0,
244 #ifdef CONFIG_TIMERLAT_TRACER
246 .timerlat_period = DEFAULT_TIMERLAT_PERIOD,
247 .timerlat_tracer = 0,
252 * Boolean variable used to inform that the tracer is currently sampling.
254 static bool osnoise_busy;
256 #ifdef CONFIG_PREEMPT_RT
258 * Print the osnoise header info.
260 static void print_osnoise_headers(struct seq_file *s)
262 if (osnoise_data.tainted)
263 seq_puts(s, "# osnoise is tainted!\n");
265 seq_puts(s, "# _-------=> irqs-off\n");
266 seq_puts(s, "# / _------=> need-resched\n");
267 seq_puts(s, "# | / _-----=> need-resched-lazy\n");
268 seq_puts(s, "# || / _----=> hardirq/softirq\n");
269 seq_puts(s, "# ||| / _---=> preempt-depth\n");
270 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n");
271 seq_puts(s, "# ||||| / _-=> migrate-disable\n");
273 seq_puts(s, "# |||||| / ");
274 seq_puts(s, " MAX\n");
276 seq_puts(s, "# ||||| / ");
277 seq_puts(s, " SINGLE Interference counters:\n");
279 seq_puts(s, "# ||||||| RUNTIME ");
280 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n");
282 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP IN US ");
283 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n");
285 seq_puts(s, "# | | | ||||||| | | ");
286 seq_puts(s, " | | | | | | | |\n");
288 #else /* CONFIG_PREEMPT_RT */
289 static void print_osnoise_headers(struct seq_file *s)
291 if (osnoise_data.tainted)
292 seq_puts(s, "# osnoise is tainted!\n");
294 seq_puts(s, "# _-----=> irqs-off\n");
295 seq_puts(s, "# / _----=> need-resched\n");
296 seq_puts(s, "# | / _---=> hardirq/softirq\n");
297 seq_puts(s, "# || / _--=> preempt-depth ");
298 seq_puts(s, " MAX\n");
300 seq_puts(s, "# || / ");
301 seq_puts(s, " SINGLE Interference counters:\n");
303 seq_puts(s, "# |||| RUNTIME ");
304 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n");
306 seq_puts(s, "# TASK-PID CPU# |||| TIMESTAMP IN US ");
307 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n");
309 seq_puts(s, "# | | | |||| | | ");
310 seq_puts(s, " | | | | | | | |\n");
312 #endif /* CONFIG_PREEMPT_RT */
315 * osnoise_taint - report an osnoise error.
317 #define osnoise_taint(msg) ({ \
318 struct trace_array *tr = osnoise_trace; \
320 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg); \
321 osnoise_data.tainted = true; \
325 * Record an osnoise_sample into the tracer buffer.
327 static void trace_osnoise_sample(struct osnoise_sample *sample)
329 struct trace_array *tr = osnoise_trace;
330 struct trace_buffer *buffer = tr->array_buffer.buffer;
331 struct trace_event_call *call = &event_osnoise;
332 struct ring_buffer_event *event;
333 struct osnoise_entry *entry;
335 event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry),
339 entry = ring_buffer_event_data(event);
340 entry->runtime = sample->runtime;
341 entry->noise = sample->noise;
342 entry->max_sample = sample->max_sample;
343 entry->hw_count = sample->hw_count;
344 entry->nmi_count = sample->nmi_count;
345 entry->irq_count = sample->irq_count;
346 entry->softirq_count = sample->softirq_count;
347 entry->thread_count = sample->thread_count;
349 if (!call_filter_check_discard(call, entry, buffer, event))
350 trace_buffer_unlock_commit_nostack(buffer, event);
353 #ifdef CONFIG_TIMERLAT_TRACER
355 * Print the timerlat header info.
357 #ifdef CONFIG_PREEMPT_RT
358 static void print_timerlat_headers(struct seq_file *s)
360 seq_puts(s, "# _-------=> irqs-off\n");
361 seq_puts(s, "# / _------=> need-resched\n");
362 seq_puts(s, "# | / _-----=> need-resched-lazy\n");
363 seq_puts(s, "# || / _----=> hardirq/softirq\n");
364 seq_puts(s, "# ||| / _---=> preempt-depth\n");
365 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n");
366 seq_puts(s, "# ||||| / _-=> migrate-disable\n");
367 seq_puts(s, "# |||||| /\n");
368 seq_puts(s, "# ||||||| ACTIVATION\n");
369 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP ID ");
370 seq_puts(s, " CONTEXT LATENCY\n");
371 seq_puts(s, "# | | | ||||||| | | ");
372 seq_puts(s, " | |\n");
374 #else /* CONFIG_PREEMPT_RT */
375 static void print_timerlat_headers(struct seq_file *s)
377 seq_puts(s, "# _-----=> irqs-off\n");
378 seq_puts(s, "# / _----=> need-resched\n");
379 seq_puts(s, "# | / _---=> hardirq/softirq\n");
380 seq_puts(s, "# || / _--=> preempt-depth\n");
381 seq_puts(s, "# || /\n");
382 seq_puts(s, "# |||| ACTIVATION\n");
383 seq_puts(s, "# TASK-PID CPU# |||| TIMESTAMP ID ");
384 seq_puts(s, " CONTEXT LATENCY\n");
385 seq_puts(s, "# | | | |||| | | ");
386 seq_puts(s, " | |\n");
388 #endif /* CONFIG_PREEMPT_RT */
391 * Record an timerlat_sample into the tracer buffer.
393 static void trace_timerlat_sample(struct timerlat_sample *sample)
395 struct trace_array *tr = osnoise_trace;
396 struct trace_event_call *call = &event_osnoise;
397 struct trace_buffer *buffer = tr->array_buffer.buffer;
398 struct ring_buffer_event *event;
399 struct timerlat_entry *entry;
401 event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry),
405 entry = ring_buffer_event_data(event);
406 entry->seqnum = sample->seqnum;
407 entry->context = sample->context;
408 entry->timer_latency = sample->timer_latency;
410 if (!call_filter_check_discard(call, entry, buffer, event))
411 trace_buffer_unlock_commit_nostack(buffer, event);
414 #ifdef CONFIG_STACKTRACE
416 #define MAX_CALLS 256
419 * Stack trace will take place only at IRQ level, so, no need
420 * to control nesting here.
425 unsigned long calls[MAX_CALLS];
428 static DEFINE_PER_CPU(struct trace_stack, trace_stack);
431 * timerlat_save_stack - save a stack trace without printing
433 * Save the current stack trace without printing. The
434 * stack will be printed later, after the end of the measurement.
436 static void timerlat_save_stack(int skip)
438 unsigned int size, nr_entries;
439 struct trace_stack *fstack;
441 fstack = this_cpu_ptr(&trace_stack);
443 size = ARRAY_SIZE(fstack->calls);
445 nr_entries = stack_trace_save(fstack->calls, size, skip);
447 fstack->stack_size = nr_entries * sizeof(unsigned long);
448 fstack->nr_entries = nr_entries;
454 * timerlat_dump_stack - dump a stack trace previously saved
456 * Dump a saved stack trace into the trace buffer.
458 static void timerlat_dump_stack(void)
460 struct trace_event_call *call = &event_osnoise;
461 struct trace_array *tr = osnoise_trace;
462 struct trace_buffer *buffer = tr->array_buffer.buffer;
463 struct ring_buffer_event *event;
464 struct trace_stack *fstack;
465 struct stack_entry *entry;
468 preempt_disable_notrace();
469 fstack = this_cpu_ptr(&trace_stack);
470 size = fstack->stack_size;
472 event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size,
477 entry = ring_buffer_event_data(event);
479 memcpy(&entry->caller, fstack->calls, size);
480 entry->size = fstack->nr_entries;
482 if (!call_filter_check_discard(call, entry, buffer, event))
483 trace_buffer_unlock_commit_nostack(buffer, event);
486 preempt_enable_notrace();
489 #define timerlat_dump_stack() do {} while (0)
490 #define timerlat_save_stack(a) do {} while (0)
491 #endif /* CONFIG_STACKTRACE */
492 #endif /* CONFIG_TIMERLAT_TRACER */
495 * Macros to encapsulate the time capturing infrastructure.
497 #define time_get() trace_clock_local()
498 #define time_to_us(x) div_u64(x, 1000)
499 #define time_sub(a, b) ((a) - (b))
502 * cond_move_irq_delta_start - Forward the delta_start of a running IRQ
504 * If an IRQ is preempted by an NMI, its delta_start is pushed forward
505 * to discount the NMI interference.
507 * See get_int_safe_duration().
510 cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration)
512 if (osn_var->irq.delta_start)
513 osn_var->irq.delta_start += duration;
516 #ifndef CONFIG_PREEMPT_RT
518 * cond_move_softirq_delta_start - Forward the delta_start of a running softirq.
520 * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed
521 * forward to discount the interference.
523 * See get_int_safe_duration().
526 cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration)
528 if (osn_var->softirq.delta_start)
529 osn_var->softirq.delta_start += duration;
531 #else /* CONFIG_PREEMPT_RT */
532 #define cond_move_softirq_delta_start(osn_var, duration) do {} while (0)
536 * cond_move_thread_delta_start - Forward the delta_start of a running thread
538 * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start
539 * is pushed forward to discount the interference.
541 * See get_int_safe_duration().
544 cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration)
546 if (osn_var->thread.delta_start)
547 osn_var->thread.delta_start += duration;
551 * get_int_safe_duration - Get the duration of a window
553 * The irq, softirq and thread varaibles need to have its duration without
554 * the interference from higher priority interrupts. Instead of keeping a
555 * variable to discount the interrupt interference from these variables, the
556 * starting time of these variables are pushed forward with the interrupt's
557 * duration. In this way, a single variable is used to:
559 * - Know if a given window is being measured.
560 * - Account its duration.
561 * - Discount the interference.
563 * To avoid getting inconsistent values, e.g.,:
567 * delta_start -= int duration;
569 * duration = now - delta_start;
571 * result: negative duration if the variable duration before the
572 * interrupt was smaller than the interrupt execution.
574 * A counter of interrupts is used. If the counter increased, try
575 * to capture an interference safe duration.
578 get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start)
580 u64 int_counter, now;
584 int_counter = local_read(&osn_var->int_counter);
585 /* synchronize with interrupts */
589 duration = (now - *delta_start);
591 /* synchronize with interrupts */
593 } while (int_counter != local_read(&osn_var->int_counter));
596 * This is an evidence of race conditions that cause
597 * a value to be "discounted" too much.
600 osnoise_taint("Negative duration!\n");
609 * set_int_safe_time - Save the current time on *time, aware of interference
611 * Get the time, taking into consideration a possible interference from
612 * higher priority interrupts.
614 * See get_int_safe_duration() for an explanation.
617 set_int_safe_time(struct osnoise_variables *osn_var, u64 *time)
622 int_counter = local_read(&osn_var->int_counter);
623 /* synchronize with interrupts */
628 /* synchronize with interrupts */
630 } while (int_counter != local_read(&osn_var->int_counter));
635 #ifdef CONFIG_TIMERLAT_TRACER
637 * copy_int_safe_time - Copy *src into *desc aware of interference
640 copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src)
645 int_counter = local_read(&osn_var->int_counter);
646 /* synchronize with interrupts */
651 /* synchronize with interrupts */
653 } while (int_counter != local_read(&osn_var->int_counter));
657 #endif /* CONFIG_TIMERLAT_TRACER */
660 * trace_osnoise_callback - NMI entry/exit callback
662 * This function is called at the entry and exit NMI code. The bool enter
663 * distinguishes between either case. This function is used to note a NMI
664 * occurrence, compute the noise caused by the NMI, and to remove the noise
665 * it is potentially causing on other interference variables.
667 void trace_osnoise_callback(bool enter)
669 struct osnoise_variables *osn_var = this_cpu_osn_var();
672 if (!osn_var->sampling)
676 * Currently trace_clock_local() calls sched_clock() and the
677 * generic version is not NMI safe.
679 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
681 osn_var->nmi.delta_start = time_get();
682 local_inc(&osn_var->int_counter);
684 duration = time_get() - osn_var->nmi.delta_start;
686 trace_nmi_noise(osn_var->nmi.delta_start, duration);
688 cond_move_irq_delta_start(osn_var, duration);
689 cond_move_softirq_delta_start(osn_var, duration);
690 cond_move_thread_delta_start(osn_var, duration);
695 osn_var->nmi.count++;
699 * osnoise_trace_irq_entry - Note the starting of an IRQ
701 * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs,
702 * it is safe to use a single variable (ons_var->irq) to save the statistics.
703 * The arrival_time is used to report... the arrival time. The delta_start
704 * is used to compute the duration at the IRQ exit handler. See
705 * cond_move_irq_delta_start().
707 void osnoise_trace_irq_entry(int id)
709 struct osnoise_variables *osn_var = this_cpu_osn_var();
711 if (!osn_var->sampling)
714 * This value will be used in the report, but not to compute
715 * the execution time, so it is safe to get it unsafe.
717 osn_var->irq.arrival_time = time_get();
718 set_int_safe_time(osn_var, &osn_var->irq.delta_start);
719 osn_var->irq.count++;
721 local_inc(&osn_var->int_counter);
725 * osnoise_irq_exit - Note the end of an IRQ, sava data and trace
727 * Computes the duration of the IRQ noise, and trace it. Also discounts the
728 * interference from other sources of noise could be currently being accounted.
730 void osnoise_trace_irq_exit(int id, const char *desc)
732 struct osnoise_variables *osn_var = this_cpu_osn_var();
735 if (!osn_var->sampling)
738 duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start);
739 trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration);
740 osn_var->irq.arrival_time = 0;
741 cond_move_softirq_delta_start(osn_var, duration);
742 cond_move_thread_delta_start(osn_var, duration);
746 * trace_irqentry_callback - Callback to the irq:irq_entry traceevent
748 * Used to note the starting of an IRQ occurece.
750 static void trace_irqentry_callback(void *data, int irq,
751 struct irqaction *action)
753 osnoise_trace_irq_entry(irq);
757 * trace_irqexit_callback - Callback to the irq:irq_exit traceevent
759 * Used to note the end of an IRQ occurece.
761 static void trace_irqexit_callback(void *data, int irq,
762 struct irqaction *action, int ret)
764 osnoise_trace_irq_exit(irq, action->name);
768 * arch specific register function.
770 int __weak osnoise_arch_register(void)
776 * arch specific unregister function.
778 void __weak osnoise_arch_unregister(void)
784 * hook_irq_events - Hook IRQ handling events
786 * This function hooks the IRQ related callbacks to the respective trace
789 static int hook_irq_events(void)
793 ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL);
797 ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL);
799 goto out_unregister_entry;
801 ret = osnoise_arch_register();
808 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
809 out_unregister_entry:
810 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
816 * unhook_irq_events - Unhook IRQ handling events
818 * This function unhooks the IRQ related callbacks to the respective trace
821 static void unhook_irq_events(void)
823 osnoise_arch_unregister();
824 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
825 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
828 #ifndef CONFIG_PREEMPT_RT
830 * trace_softirq_entry_callback - Note the starting of a softirq
832 * Save the starting time of a softirq. As softirqs are non-preemptive to
833 * other softirqs, it is safe to use a single variable (ons_var->softirq)
834 * to save the statistics. The arrival_time is used to report... the
835 * arrival time. The delta_start is used to compute the duration at the
836 * softirq exit handler. See cond_move_softirq_delta_start().
838 static void trace_softirq_entry_callback(void *data, unsigned int vec_nr)
840 struct osnoise_variables *osn_var = this_cpu_osn_var();
842 if (!osn_var->sampling)
845 * This value will be used in the report, but not to compute
846 * the execution time, so it is safe to get it unsafe.
848 osn_var->softirq.arrival_time = time_get();
849 set_int_safe_time(osn_var, &osn_var->softirq.delta_start);
850 osn_var->softirq.count++;
852 local_inc(&osn_var->int_counter);
856 * trace_softirq_exit_callback - Note the end of an softirq
858 * Computes the duration of the softirq noise, and trace it. Also discounts the
859 * interference from other sources of noise could be currently being accounted.
861 static void trace_softirq_exit_callback(void *data, unsigned int vec_nr)
863 struct osnoise_variables *osn_var = this_cpu_osn_var();
866 if (!osn_var->sampling)
869 #ifdef CONFIG_TIMERLAT_TRACER
871 * If the timerlat is enabled, but the irq handler did
872 * not run yet enabling timerlat_tracer, do not trace.
874 if (unlikely(osnoise_data.timerlat_tracer)) {
875 struct timerlat_variables *tlat_var;
876 tlat_var = this_cpu_tmr_var();
877 if (!tlat_var->tracing_thread) {
878 osn_var->softirq.arrival_time = 0;
879 osn_var->softirq.delta_start = 0;
885 duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start);
886 trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration);
887 cond_move_thread_delta_start(osn_var, duration);
888 osn_var->softirq.arrival_time = 0;
892 * hook_softirq_events - Hook softirq handling events
894 * This function hooks the softirq related callbacks to the respective trace
897 static int hook_softirq_events(void)
901 ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL);
905 ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL);
907 goto out_unreg_entry;
912 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
918 * unhook_softirq_events - Unhook softirq handling events
920 * This function hooks the softirq related callbacks to the respective trace
923 static void unhook_softirq_events(void)
925 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
926 unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL);
928 #else /* CONFIG_PREEMPT_RT */
930 * softirq are threads on the PREEMPT_RT mode.
932 static int hook_softirq_events(void)
936 static void unhook_softirq_events(void)
942 * thread_entry - Record the starting of a thread noise window
944 * It saves the context switch time for a noisy thread, and increments
945 * the interference counters.
948 thread_entry(struct osnoise_variables *osn_var, struct task_struct *t)
950 if (!osn_var->sampling)
953 * The arrival time will be used in the report, but not to compute
954 * the execution time, so it is safe to get it unsafe.
956 osn_var->thread.arrival_time = time_get();
958 set_int_safe_time(osn_var, &osn_var->thread.delta_start);
960 osn_var->thread.count++;
961 local_inc(&osn_var->int_counter);
965 * thread_exit - Report the end of a thread noise window
967 * It computes the total noise from a thread, tracing if needed.
970 thread_exit(struct osnoise_variables *osn_var, struct task_struct *t)
974 if (!osn_var->sampling)
977 #ifdef CONFIG_TIMERLAT_TRACER
978 if (osnoise_data.timerlat_tracer) {
979 struct timerlat_variables *tlat_var;
980 tlat_var = this_cpu_tmr_var();
981 if (!tlat_var->tracing_thread) {
982 osn_var->thread.delta_start = 0;
983 osn_var->thread.arrival_time = 0;
989 duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start);
991 trace_thread_noise(t, osn_var->thread.arrival_time, duration);
993 osn_var->thread.arrival_time = 0;
997 * trace_sched_switch - sched:sched_switch trace event handler
999 * This function is hooked to the sched:sched_switch trace event, and it is
1000 * used to record the beginning and to report the end of a thread noise window.
1003 trace_sched_switch_callback(void *data, bool preempt, struct task_struct *p,
1004 struct task_struct *n)
1006 struct osnoise_variables *osn_var = this_cpu_osn_var();
1008 if (p->pid != osn_var->pid)
1009 thread_exit(osn_var, p);
1011 if (n->pid != osn_var->pid)
1012 thread_entry(osn_var, n);
1016 * hook_thread_events - Hook the insturmentation for thread noise
1018 * Hook the osnoise tracer callbacks to handle the noise from other
1019 * threads on the necessary kernel events.
1021 static int hook_thread_events(void)
1025 ret = register_trace_sched_switch(trace_sched_switch_callback, NULL);
1033 * unhook_thread_events - *nhook the insturmentation for thread noise
1035 * Unook the osnoise tracer callbacks to handle the noise from other
1036 * threads on the necessary kernel events.
1038 static void unhook_thread_events(void)
1040 unregister_trace_sched_switch(trace_sched_switch_callback, NULL);
1044 * save_osn_sample_stats - Save the osnoise_sample statistics
1046 * Save the osnoise_sample statistics before the sampling phase. These
1047 * values will be used later to compute the diff betwneen the statistics
1048 * before and after the osnoise sampling.
1051 save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1053 s->nmi_count = osn_var->nmi.count;
1054 s->irq_count = osn_var->irq.count;
1055 s->softirq_count = osn_var->softirq.count;
1056 s->thread_count = osn_var->thread.count;
1060 * diff_osn_sample_stats - Compute the osnoise_sample statistics
1062 * After a sample period, compute the difference on the osnoise_sample
1063 * statistics. The struct osnoise_sample *s contains the statistics saved via
1064 * save_osn_sample_stats() before the osnoise sampling.
1067 diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1069 s->nmi_count = osn_var->nmi.count - s->nmi_count;
1070 s->irq_count = osn_var->irq.count - s->irq_count;
1071 s->softirq_count = osn_var->softirq.count - s->softirq_count;
1072 s->thread_count = osn_var->thread.count - s->thread_count;
1076 * osnoise_stop_tracing - Stop tracing and the tracer.
1078 static __always_inline void osnoise_stop_tracing(void)
1080 struct trace_array *tr = osnoise_trace;
1082 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_,
1083 "stop tracing hit on cpu %d\n", smp_processor_id());
1085 tracer_tracing_off(tr);
1089 * run_osnoise - Sample the time and look for osnoise
1091 * Used to capture the time, looking for potential osnoise latency repeatedly.
1092 * Different from hwlat_detector, it is called with preemption and interrupts
1093 * enabled. This allows irqs, softirqs and threads to run, interfering on the
1094 * osnoise sampling thread, as they would do with a regular thread.
1096 static int run_osnoise(void)
1098 struct osnoise_variables *osn_var = this_cpu_osn_var();
1099 struct trace_array *tr = osnoise_trace;
1100 u64 start, sample, last_sample;
1101 u64 last_int_count, int_count;
1102 s64 noise = 0, max_noise = 0;
1103 s64 total, last_total = 0;
1104 struct osnoise_sample s;
1105 unsigned int threshold;
1106 u64 runtime, stop_in;
1112 * Considers the current thread as the workload.
1114 osn_var->pid = current->pid;
1117 * Save the current stats for the diff
1119 save_osn_sample_stats(osn_var, &s);
1122 * if threshold is 0, use the default value of 5 us.
1124 threshold = tracing_thresh ? : 5000;
1127 * Make sure NMIs see sampling first
1129 osn_var->sampling = true;
1133 * Transform the *_us config to nanoseconds to avoid the
1134 * division on the main loop.
1136 runtime = osnoise_data.sample_runtime * NSEC_PER_USEC;
1137 stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC;
1147 last_int_count = set_int_safe_time(osn_var, &last_sample);
1153 int_count = set_int_safe_time(osn_var, &sample);
1155 noise = time_sub(sample, last_sample);
1158 * This shouldn't happen.
1161 osnoise_taint("negative noise!");
1168 total = time_sub(sample, start);
1171 * Check for possible overflows.
1173 if (total < last_total) {
1174 osnoise_taint("total overflow!");
1180 if (noise >= threshold) {
1181 int interference = int_count - last_int_count;
1183 if (noise > max_noise)
1191 trace_sample_threshold(last_sample, noise, interference);
1193 if (osnoise_data.stop_tracing)
1194 if (noise > stop_in)
1195 osnoise_stop_tracing();
1199 * In some cases, notably when running on a nohz_full CPU with
1200 * a stopped tick PREEMPT_RCU has no way to account for QSs.
1201 * This will eventually cause unwarranted noise as PREEMPT_RCU
1202 * will force preemption as the means of ending the current
1203 * grace period. We avoid this problem by calling
1204 * rcu_momentary_dyntick_idle(), which performs a zero duration
1205 * EQS allowing PREEMPT_RCU to end the current grace period.
1206 * This call shouldn't be wrapped inside an RCU critical
1209 * Note that in non PREEMPT_RCU kernels QSs are handled through
1212 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
1213 local_irq_disable();
1214 rcu_momentary_dyntick_idle();
1219 * For the non-preemptive kernel config: let threads runs, if
1224 last_sample = sample;
1225 last_int_count = int_count;
1227 } while (total < runtime && !kthread_should_stop());
1230 * Finish the above in the view for interrupts.
1234 osn_var->sampling = false;
1237 * Make sure sampling data is no longer updated.
1244 s.noise = time_to_us(sum_noise);
1245 s.runtime = time_to_us(total);
1246 s.max_sample = time_to_us(max_noise);
1247 s.hw_count = hw_count;
1249 /* Save interference stats info */
1250 diff_osn_sample_stats(osn_var, &s);
1252 trace_osnoise_sample(&s);
1254 /* Keep a running maximum ever recorded osnoise "latency" */
1255 if (max_noise > tr->max_latency) {
1256 tr->max_latency = max_noise;
1257 latency_fsnotify(tr);
1260 if (osnoise_data.stop_tracing_total)
1261 if (s.noise > osnoise_data.stop_tracing_total)
1262 osnoise_stop_tracing();
1269 static struct cpumask osnoise_cpumask;
1270 static struct cpumask save_cpumask;
1273 * osnoise_sleep - sleep until the next period
1275 static void osnoise_sleep(void)
1280 mutex_lock(&interface_lock);
1281 interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
1282 mutex_unlock(&interface_lock);
1285 * differently from hwlat_detector, the osnoise tracer can run
1286 * without a pause because preemption is on.
1289 /* Let synchronize_rcu_tasks() make progress */
1290 cond_resched_tasks_rcu_qs();
1294 wake_time = ktime_add_us(ktime_get(), interval);
1295 __set_current_state(TASK_INTERRUPTIBLE);
1297 while (schedule_hrtimeout_range(&wake_time, 0, HRTIMER_MODE_ABS)) {
1298 if (kthread_should_stop())
1304 * osnoise_main - The osnoise detection kernel thread
1306 * Calls run_osnoise() function to measure the osnoise for the configured runtime,
1309 static int osnoise_main(void *data)
1312 while (!kthread_should_stop()) {
1320 #ifdef CONFIG_TIMERLAT_TRACER
1322 * timerlat_irq - hrtimer handler for timerlat.
1324 static enum hrtimer_restart timerlat_irq(struct hrtimer *timer)
1326 struct osnoise_variables *osn_var = this_cpu_osn_var();
1327 struct trace_array *tr = osnoise_trace;
1328 struct timerlat_variables *tlat;
1329 struct timerlat_sample s;
1334 * I am not sure if the timer was armed for this CPU. So, get
1335 * the timerlat struct from the timer itself, not from this
1338 tlat = container_of(timer, struct timerlat_variables, timer);
1340 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1343 * Enable the osnoise: events for thread an softirq.
1345 tlat->tracing_thread = true;
1347 osn_var->thread.arrival_time = time_get();
1350 * A hardirq is running: the timer IRQ. It is for sure preempting
1351 * a thread, and potentially preempting a softirq.
1353 * At this point, it is not interesting to know the duration of the
1354 * preempted thread (and maybe softirq), but how much time they will
1355 * delay the beginning of the execution of the timer thread.
1357 * To get the correct (net) delay added by the softirq, its delta_start
1358 * is set as the IRQ one. In this way, at the return of the IRQ, the delta
1359 * start of the sofitrq will be zeroed, accounting then only the time
1362 * The thread follows the same principle. However, if a softirq is
1363 * running, the thread needs to receive the softirq delta_start. The
1364 * reason being is that the softirq will be the last to be unfolded,
1365 * resseting the thread delay to zero.
1367 #ifndef CONFIG_PREEMPT_RT
1368 if (osn_var->softirq.delta_start) {
1369 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1370 &osn_var->softirq.delta_start);
1372 copy_int_safe_time(osn_var, &osn_var->softirq.delta_start,
1373 &osn_var->irq.delta_start);
1375 copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1376 &osn_var->irq.delta_start);
1378 #else /* CONFIG_PREEMPT_RT */
1380 * The sofirqs run as threads on RT, so there is not need
1381 * to keep track of it.
1383 copy_int_safe_time(osn_var, &osn_var->thread.delta_start, &osn_var->irq.delta_start);
1384 #endif /* CONFIG_PREEMPT_RT */
1387 * Compute the current time with the expected time.
1389 diff = now - tlat->abs_period;
1392 s.seqnum = tlat->count;
1393 s.timer_latency = diff;
1394 s.context = IRQ_CONTEXT;
1396 trace_timerlat_sample(&s);
1398 /* Keep a running maximum ever recorded os noise "latency" */
1399 if (diff > tr->max_latency) {
1400 tr->max_latency = diff;
1401 latency_fsnotify(tr);
1404 if (osnoise_data.stop_tracing)
1405 if (time_to_us(diff) >= osnoise_data.stop_tracing)
1406 osnoise_stop_tracing();
1408 wake_up_process(tlat->kthread);
1410 if (osnoise_data.print_stack)
1411 timerlat_save_stack(0);
1413 return HRTIMER_NORESTART;
1417 * wait_next_period - Wait for the next period for timerlat
1419 static int wait_next_period(struct timerlat_variables *tlat)
1421 ktime_t next_abs_period, now;
1422 u64 rel_period = osnoise_data.timerlat_period * 1000;
1424 now = hrtimer_cb_get_time(&tlat->timer);
1425 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1428 * Save the next abs_period.
1430 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1433 * If the new abs_period is in the past, skip the activation.
1435 while (ktime_compare(now, next_abs_period) > 0) {
1436 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1437 tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1440 set_current_state(TASK_INTERRUPTIBLE);
1442 hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD);
1448 * timerlat_main- Timerlat main
1450 static int timerlat_main(void *data)
1452 struct osnoise_variables *osn_var = this_cpu_osn_var();
1453 struct timerlat_variables *tlat = this_cpu_tmr_var();
1454 struct timerlat_sample s;
1455 struct sched_param sp;
1459 * Make the thread RT, that is how cyclictest is usually used.
1461 sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
1462 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1465 tlat->tracing_thread = false;
1467 hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1468 tlat->timer.function = timerlat_irq;
1469 tlat->kthread = current;
1470 osn_var->pid = current->pid;
1472 * Anotate the arrival time.
1474 tlat->abs_period = hrtimer_cb_get_time(&tlat->timer);
1476 wait_next_period(tlat);
1478 osn_var->sampling = 1;
1480 while (!kthread_should_stop()) {
1481 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1482 diff = now - tlat->abs_period;
1484 s.seqnum = tlat->count;
1485 s.timer_latency = diff;
1486 s.context = THREAD_CONTEXT;
1488 trace_timerlat_sample(&s);
1490 #ifdef CONFIG_STACKTRACE
1491 if (osnoise_data.print_stack)
1492 if (osnoise_data.print_stack <= time_to_us(diff))
1493 timerlat_dump_stack();
1494 #endif /* CONFIG_STACKTRACE */
1496 tlat->tracing_thread = false;
1497 if (osnoise_data.stop_tracing_total)
1498 if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
1499 osnoise_stop_tracing();
1501 wait_next_period(tlat);
1504 hrtimer_cancel(&tlat->timer);
1507 #endif /* CONFIG_TIMERLAT_TRACER */
1510 * stop_kthread - stop a workload thread
1512 static void stop_kthread(unsigned int cpu)
1514 struct task_struct *kthread;
1516 kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread;
1518 kthread_stop(kthread);
1519 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1523 * stop_per_cpu_kthread - Stop per-cpu threads
1525 * Stop the osnoise sampling htread. Use this on unload and at system
1528 static void stop_per_cpu_kthreads(void)
1534 for_each_online_cpu(cpu)
1541 * start_kthread - Start a workload tread
1543 static int start_kthread(unsigned int cpu)
1545 struct task_struct *kthread;
1546 void *main = osnoise_main;
1549 #ifdef CONFIG_TIMERLAT_TRACER
1550 if (osnoise_data.timerlat_tracer) {
1551 snprintf(comm, 24, "timerlat/%d", cpu);
1552 main = timerlat_main;
1554 snprintf(comm, 24, "osnoise/%d", cpu);
1557 snprintf(comm, 24, "osnoise/%d", cpu);
1559 kthread = kthread_create_on_cpu(main, NULL, cpu, comm);
1561 if (IS_ERR(kthread)) {
1562 pr_err(BANNER "could not start sampling thread\n");
1563 stop_per_cpu_kthreads();
1567 per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread;
1568 wake_up_process(kthread);
1574 * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads
1576 * This starts the kernel thread that will look for osnoise on many
1579 static int start_per_cpu_kthreads(struct trace_array *tr)
1581 struct cpumask *current_mask = &save_cpumask;
1587 * Run only on CPUs in which trace and osnoise are allowed to run.
1589 cpumask_and(current_mask, tr->tracing_cpumask, &osnoise_cpumask);
1591 * And the CPU is online.
1593 cpumask_and(current_mask, cpu_online_mask, current_mask);
1595 for_each_possible_cpu(cpu)
1596 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
1598 for_each_cpu(cpu, current_mask) {
1599 retval = start_kthread(cpu);
1602 stop_per_cpu_kthreads();
1612 #ifdef CONFIG_HOTPLUG_CPU
1613 static void osnoise_hotplug_workfn(struct work_struct *dummy)
1615 struct trace_array *tr = osnoise_trace;
1616 unsigned int cpu = smp_processor_id();
1619 mutex_lock(&trace_types_lock);
1622 goto out_unlock_trace;
1624 mutex_lock(&interface_lock);
1627 if (!cpumask_test_cpu(cpu, &osnoise_cpumask))
1630 if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
1637 mutex_unlock(&interface_lock);
1639 mutex_unlock(&trace_types_lock);
1642 static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn);
1645 * osnoise_cpu_init - CPU hotplug online callback function
1647 static int osnoise_cpu_init(unsigned int cpu)
1649 schedule_work_on(cpu, &osnoise_hotplug_work);
1654 * osnoise_cpu_die - CPU hotplug offline callback function
1656 static int osnoise_cpu_die(unsigned int cpu)
1662 static void osnoise_init_hotplug_support(void)
1666 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online",
1667 osnoise_cpu_init, osnoise_cpu_die);
1669 pr_warn(BANNER "Error to init cpu hotplug support\n");
1673 #else /* CONFIG_HOTPLUG_CPU */
1674 static void osnoise_init_hotplug_support(void)
1678 #endif /* CONFIG_HOTPLUG_CPU */
1681 * osnoise_cpus_read - Read function for reading the "cpus" file
1682 * @filp: The active open file structure
1683 * @ubuf: The userspace provided buffer to read value into
1684 * @cnt: The maximum number of bytes to read
1685 * @ppos: The current "file" position
1687 * Prints the "cpus" output into the user-provided buffer.
1690 osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count,
1696 mutex_lock(&interface_lock);
1698 len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1;
1699 mask_str = kmalloc(len, GFP_KERNEL);
1705 len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask));
1711 count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
1716 mutex_unlock(&interface_lock);
1721 static void osnoise_tracer_start(struct trace_array *tr);
1722 static void osnoise_tracer_stop(struct trace_array *tr);
1725 * osnoise_cpus_write - Write function for "cpus" entry
1726 * @filp: The active open file structure
1727 * @ubuf: The user buffer that contains the value to write
1728 * @cnt: The maximum number of bytes to write to "file"
1729 * @ppos: The current position in @file
1731 * This function provides a write implementation for the "cpus"
1732 * interface to the osnoise trace. By default, it lists all CPUs,
1733 * in this way, allowing osnoise threads to run on any online CPU
1734 * of the system. It serves to restrict the execution of osnoise to the
1735 * set of CPUs writing via this interface. Note that osnoise also
1736 * respects the "tracing_cpumask." Hence, osnoise threads will run only
1737 * on the set of CPUs allowed here AND on "tracing_cpumask." Why not
1738 * have just "tracing_cpumask?" Because the user might be interested
1739 * in tracing what is running on other CPUs. For instance, one might
1740 * run osnoise in one HT CPU while observing what is running on the
1744 osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
1747 struct trace_array *tr = osnoise_trace;
1748 cpumask_var_t osnoise_cpumask_new;
1755 if (copy_from_user(buf, ubuf, count))
1758 if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL))
1761 err = cpulist_parse(buf, osnoise_cpumask_new);
1766 * trace_types_lock is taken to avoid concurrency on start/stop
1769 mutex_lock(&trace_types_lock);
1770 running = osnoise_busy;
1772 osnoise_tracer_stop(tr);
1774 mutex_lock(&interface_lock);
1776 * osnoise_cpumask is read by CPU hotplug operations.
1780 cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
1783 mutex_unlock(&interface_lock);
1786 osnoise_tracer_start(tr);
1787 mutex_unlock(&trace_types_lock);
1789 free_cpumask_var(osnoise_cpumask_new);
1793 free_cpumask_var(osnoise_cpumask_new);
1799 * osnoise/runtime_us: cannot be greater than the period.
1801 static struct trace_min_max_param osnoise_runtime = {
1802 .lock = &interface_lock,
1803 .val = &osnoise_data.sample_runtime,
1804 .max = &osnoise_data.sample_period,
1809 * osnoise/period_us: cannot be smaller than the runtime.
1811 static struct trace_min_max_param osnoise_period = {
1812 .lock = &interface_lock,
1813 .val = &osnoise_data.sample_period,
1815 .min = &osnoise_data.sample_runtime,
1819 * osnoise/stop_tracing_us: no limit.
1821 static struct trace_min_max_param osnoise_stop_tracing_in = {
1822 .lock = &interface_lock,
1823 .val = &osnoise_data.stop_tracing,
1829 * osnoise/stop_tracing_total_us: no limit.
1831 static struct trace_min_max_param osnoise_stop_tracing_total = {
1832 .lock = &interface_lock,
1833 .val = &osnoise_data.stop_tracing_total,
1838 #ifdef CONFIG_TIMERLAT_TRACER
1840 * osnoise/print_stack: print the stacktrace of the IRQ handler if the total
1841 * latency is higher than val.
1843 static struct trace_min_max_param osnoise_print_stack = {
1844 .lock = &interface_lock,
1845 .val = &osnoise_data.print_stack,
1851 * osnoise/timerlat_period: min 100 us, max 1 s
1853 u64 timerlat_min_period = 100;
1854 u64 timerlat_max_period = 1000000;
1855 static struct trace_min_max_param timerlat_period = {
1856 .lock = &interface_lock,
1857 .val = &osnoise_data.timerlat_period,
1858 .max = &timerlat_max_period,
1859 .min = &timerlat_min_period,
1863 static const struct file_operations cpus_fops = {
1864 .open = tracing_open_generic,
1865 .read = osnoise_cpus_read,
1866 .write = osnoise_cpus_write,
1867 .llseek = generic_file_llseek,
1871 * init_tracefs - A function to initialize the tracefs interface files
1873 * This function creates entries in tracefs for "osnoise" and "timerlat".
1874 * It creates these directories in the tracing directory, and within that
1875 * directory the use can change and view the configs.
1877 static int init_tracefs(void)
1879 struct dentry *top_dir;
1883 ret = tracing_init_dentry();
1887 top_dir = tracefs_create_dir("osnoise", NULL);
1891 tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir,
1892 &osnoise_period, &trace_min_max_fops);
1896 tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir,
1897 &osnoise_runtime, &trace_min_max_fops);
1901 tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir,
1902 &osnoise_stop_tracing_in, &trace_min_max_fops);
1906 tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir,
1907 &osnoise_stop_tracing_total, &trace_min_max_fops);
1911 tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops);
1914 #ifdef CONFIG_TIMERLAT_TRACER
1915 #ifdef CONFIG_STACKTRACE
1916 tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir,
1917 &osnoise_print_stack, &trace_min_max_fops);
1922 tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir,
1923 &timerlat_period, &trace_min_max_fops);
1931 tracefs_remove(top_dir);
1935 static int osnoise_hook_events(void)
1940 * Trace is already hooked, we are re-enabling from
1943 if (trace_osnoise_callback_enabled)
1946 retval = hook_irq_events();
1950 retval = hook_softirq_events();
1952 goto out_unhook_irq;
1954 retval = hook_thread_events();
1961 unhook_softirq_events();
1963 unhook_irq_events();
1967 static void osnoise_unhook_events(void)
1969 unhook_thread_events();
1970 unhook_softirq_events();
1971 unhook_irq_events();
1974 static int __osnoise_tracer_start(struct trace_array *tr)
1978 osn_var_reset_all();
1980 retval = osnoise_hook_events();
1984 * Make sure NMIs see reseted values.
1987 trace_osnoise_callback_enabled = true;
1989 retval = start_per_cpu_kthreads(tr);
1991 trace_osnoise_callback_enabled = false;
1993 * Make sure that ftrace_nmi_enter/exit() see
1994 * trace_osnoise_callback_enabled as false before continuing.
1998 osnoise_unhook_events();
2002 osnoise_busy = true;
2007 static void osnoise_tracer_start(struct trace_array *tr)
2014 retval = __osnoise_tracer_start(tr);
2016 pr_err(BANNER "Error starting osnoise tracer\n");
2020 static void osnoise_tracer_stop(struct trace_array *tr)
2025 trace_osnoise_callback_enabled = false;
2028 stop_per_cpu_kthreads();
2030 osnoise_unhook_events();
2032 osnoise_busy = false;
2035 static int osnoise_tracer_init(struct trace_array *tr)
2038 /* Only allow one instance to enable this */
2043 tr->max_latency = 0;
2045 osnoise_tracer_start(tr);
2050 static void osnoise_tracer_reset(struct trace_array *tr)
2052 osnoise_tracer_stop(tr);
2055 static struct tracer osnoise_tracer __read_mostly = {
2057 .init = osnoise_tracer_init,
2058 .reset = osnoise_tracer_reset,
2059 .start = osnoise_tracer_start,
2060 .stop = osnoise_tracer_stop,
2061 .print_header = print_osnoise_headers,
2062 .allow_instances = true,
2065 #ifdef CONFIG_TIMERLAT_TRACER
2066 static void timerlat_tracer_start(struct trace_array *tr)
2073 osnoise_data.timerlat_tracer = 1;
2075 retval = __osnoise_tracer_start(tr);
2081 pr_err(BANNER "Error starting timerlat tracer\n");
2084 static void timerlat_tracer_stop(struct trace_array *tr)
2091 for_each_online_cpu(cpu)
2092 per_cpu(per_cpu_osnoise_var, cpu).sampling = 0;
2094 osnoise_tracer_stop(tr);
2096 osnoise_data.timerlat_tracer = 0;
2099 static int timerlat_tracer_init(struct trace_array *tr)
2101 /* Only allow one instance to enable this */
2107 tr->max_latency = 0;
2109 timerlat_tracer_start(tr);
2114 static void timerlat_tracer_reset(struct trace_array *tr)
2116 timerlat_tracer_stop(tr);
2119 static struct tracer timerlat_tracer __read_mostly = {
2121 .init = timerlat_tracer_init,
2122 .reset = timerlat_tracer_reset,
2123 .start = timerlat_tracer_start,
2124 .stop = timerlat_tracer_stop,
2125 .print_header = print_timerlat_headers,
2126 .allow_instances = true,
2128 #endif /* CONFIG_TIMERLAT_TRACER */
2130 __init static int init_osnoise_tracer(void)
2134 mutex_init(&interface_lock);
2136 cpumask_copy(&osnoise_cpumask, cpu_all_mask);
2138 ret = register_tracer(&osnoise_tracer);
2140 pr_err(BANNER "Error registering osnoise!\n");
2144 #ifdef CONFIG_TIMERLAT_TRACER
2145 ret = register_tracer(&timerlat_tracer);
2147 pr_err(BANNER "Error registering timerlat\n");
2151 osnoise_init_hotplug_support();
2157 late_initcall(init_osnoise_tracer);