1 // SPDX-License-Identifier: GPL-2.0
3 * trace task wakeup timings
8 * Based on code from the latency_tracer, that is:
10 * Copyright (C) 2004-2006 Ingo Molnar
11 * Copyright (C) 2004 Nadia Yvette Chambers
13 #include <linux/module.h>
14 #include <linux/kallsyms.h>
15 #include <linux/uaccess.h>
16 #include <linux/ftrace.h>
17 #include <linux/sched/rt.h>
18 #include <linux/sched/deadline.h>
19 #include <trace/events/sched.h>
22 static struct trace_array *wakeup_trace;
23 static int __read_mostly tracer_enabled;
25 static struct task_struct *wakeup_task;
26 static int wakeup_cpu;
27 static int wakeup_current_cpu;
28 static unsigned wakeup_prio = -1;
31 static int tracing_dl = 0;
33 static arch_spinlock_t wakeup_lock =
34 (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
36 static void wakeup_reset(struct trace_array *tr);
37 static void __wakeup_reset(struct trace_array *tr);
38 static int start_func_tracer(struct trace_array *tr, int graph);
39 static void stop_func_tracer(struct trace_array *tr, int graph);
41 static int save_flags;
43 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
44 # define is_graph(tr) ((tr)->trace_flags & TRACE_ITER_DISPLAY_GRAPH)
46 # define is_graph(tr) false
49 #ifdef CONFIG_FUNCTION_TRACER
51 static bool function_enabled;
54 * Prologue for the wakeup function tracers.
56 * Returns 1 if it is OK to continue, and preemption
57 * is disabled and data->disabled is incremented.
58 * 0 if the trace is to be ignored, and preemption
59 * is not disabled and data->disabled is
62 * Note, this function is also used outside this ifdef but
63 * inside the #ifdef of the function graph tracer below.
64 * This is OK, since the function graph tracer is
65 * dependent on the function tracer.
68 func_prolog_preempt_disable(struct trace_array *tr,
69 struct trace_array_cpu **data,
75 if (likely(!wakeup_task))
78 *pc = preempt_count();
79 preempt_disable_notrace();
81 cpu = raw_smp_processor_id();
82 if (cpu != wakeup_current_cpu)
85 *data = per_cpu_ptr(tr->array_buffer.data, cpu);
86 disabled = atomic_inc_return(&(*data)->disabled);
87 if (unlikely(disabled != 1))
93 atomic_dec(&(*data)->disabled);
96 preempt_enable_notrace();
100 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
102 static int wakeup_display_graph(struct trace_array *tr, int set)
104 if (!(is_graph(tr) ^ set))
107 stop_func_tracer(tr, !set);
109 wakeup_reset(wakeup_trace);
112 return start_func_tracer(tr, set);
115 static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
117 struct trace_array *tr = wakeup_trace;
118 struct trace_array_cpu *data;
122 if (ftrace_graph_ignore_func(trace))
125 * Do not trace a function if it's filtered by set_graph_notrace.
126 * Make the index of ret stack negative to indicate that it should
127 * ignore further functions. But it needs its own ret stack entry
128 * to recover the original index in order to continue tracing after
129 * returning from the function.
131 if (ftrace_graph_notrace_addr(trace->func))
134 if (!func_prolog_preempt_disable(tr, &data, &pc))
137 local_save_flags(flags);
138 ret = __trace_graph_entry(tr, trace, flags, pc);
139 atomic_dec(&data->disabled);
140 preempt_enable_notrace();
145 static void wakeup_graph_return(struct ftrace_graph_ret *trace)
147 struct trace_array *tr = wakeup_trace;
148 struct trace_array_cpu *data;
152 ftrace_graph_addr_finish(trace);
154 if (!func_prolog_preempt_disable(tr, &data, &pc))
157 local_save_flags(flags);
158 __trace_graph_return(tr, trace, flags, pc);
159 atomic_dec(&data->disabled);
161 preempt_enable_notrace();
165 static struct fgraph_ops fgraph_wakeup_ops = {
166 .entryfunc = &wakeup_graph_entry,
167 .retfunc = &wakeup_graph_return,
170 static void wakeup_trace_open(struct trace_iterator *iter)
172 if (is_graph(iter->tr))
173 graph_trace_open(iter);
176 static void wakeup_trace_close(struct trace_iterator *iter)
179 graph_trace_close(iter);
182 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
183 TRACE_GRAPH_PRINT_CPU | \
184 TRACE_GRAPH_PRINT_REL_TIME | \
185 TRACE_GRAPH_PRINT_DURATION | \
186 TRACE_GRAPH_PRINT_OVERHEAD | \
187 TRACE_GRAPH_PRINT_IRQS)
189 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
192 * In graph mode call the graph tracer output function,
193 * otherwise go with the TRACE_FN event handler
195 if (is_graph(iter->tr))
196 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
198 return TRACE_TYPE_UNHANDLED;
201 static void wakeup_print_header(struct seq_file *s)
203 if (is_graph(wakeup_trace))
204 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
206 trace_default_header(s);
208 #endif /* else CONFIG_FUNCTION_GRAPH_TRACER */
211 * wakeup uses its own tracer function to keep the overhead down:
214 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip,
215 struct ftrace_ops *op, struct ftrace_regs *fregs)
217 struct trace_array *tr = wakeup_trace;
218 struct trace_array_cpu *data;
222 if (!func_prolog_preempt_disable(tr, &data, &pc))
225 local_irq_save(flags);
226 trace_function(tr, ip, parent_ip, flags, pc);
227 local_irq_restore(flags);
229 atomic_dec(&data->disabled);
230 preempt_enable_notrace();
233 static int register_wakeup_function(struct trace_array *tr, int graph, int set)
237 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
238 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
242 ret = register_ftrace_graph(&fgraph_wakeup_ops);
244 ret = register_ftrace_function(tr->ops);
247 function_enabled = true;
252 static void unregister_wakeup_function(struct trace_array *tr, int graph)
254 if (!function_enabled)
258 unregister_ftrace_graph(&fgraph_wakeup_ops);
260 unregister_ftrace_function(tr->ops);
262 function_enabled = false;
265 static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
267 if (!(mask & TRACE_ITER_FUNCTION))
271 register_wakeup_function(tr, is_graph(tr), 1);
273 unregister_wakeup_function(tr, is_graph(tr));
276 #else /* CONFIG_FUNCTION_TRACER */
277 static int register_wakeup_function(struct trace_array *tr, int graph, int set)
281 static void unregister_wakeup_function(struct trace_array *tr, int graph) { }
282 static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
286 #endif /* else CONFIG_FUNCTION_TRACER */
288 #ifndef CONFIG_FUNCTION_GRAPH_TRACER
289 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
291 return TRACE_TYPE_UNHANDLED;
294 static void wakeup_trace_open(struct trace_iterator *iter) { }
295 static void wakeup_trace_close(struct trace_iterator *iter) { }
297 static void wakeup_print_header(struct seq_file *s)
299 trace_default_header(s);
301 #endif /* !CONFIG_FUNCTION_GRAPH_TRACER */
304 __trace_function(struct trace_array *tr,
305 unsigned long ip, unsigned long parent_ip,
306 unsigned long flags, int pc)
309 trace_graph_function(tr, ip, parent_ip, flags, pc);
311 trace_function(tr, ip, parent_ip, flags, pc);
314 static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set)
316 struct tracer *tracer = tr->current_trace;
318 if (wakeup_function_set(tr, mask, set))
321 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
322 if (mask & TRACE_ITER_DISPLAY_GRAPH)
323 return wakeup_display_graph(tr, set);
326 return trace_keep_overwrite(tracer, mask, set);
329 static int start_func_tracer(struct trace_array *tr, int graph)
333 ret = register_wakeup_function(tr, graph, 0);
335 if (!ret && tracing_is_enabled())
343 static void stop_func_tracer(struct trace_array *tr, int graph)
347 unregister_wakeup_function(tr, graph);
351 * Should this new latency be reported/recorded?
353 static bool report_latency(struct trace_array *tr, u64 delta)
355 if (tracing_thresh) {
356 if (delta < tracing_thresh)
359 if (delta <= tr->max_latency)
366 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
368 if (task != wakeup_task)
371 wakeup_current_cpu = cpu;
375 tracing_sched_switch_trace(struct trace_array *tr,
376 struct task_struct *prev,
377 struct task_struct *next,
378 unsigned long flags, int pc)
380 struct trace_event_call *call = &event_context_switch;
381 struct trace_buffer *buffer = tr->array_buffer.buffer;
382 struct ring_buffer_event *event;
383 struct ctx_switch_entry *entry;
385 event = trace_buffer_lock_reserve(buffer, TRACE_CTX,
386 sizeof(*entry), flags, pc);
389 entry = ring_buffer_event_data(event);
390 entry->prev_pid = prev->pid;
391 entry->prev_prio = prev->prio;
392 entry->prev_state = task_state_index(prev);
393 entry->next_pid = next->pid;
394 entry->next_prio = next->prio;
395 entry->next_state = task_state_index(next);
396 entry->next_cpu = task_cpu(next);
398 if (!call_filter_check_discard(call, entry, buffer, event))
399 trace_buffer_unlock_commit(tr, buffer, event, flags, pc);
403 tracing_sched_wakeup_trace(struct trace_array *tr,
404 struct task_struct *wakee,
405 struct task_struct *curr,
406 unsigned long flags, int pc)
408 struct trace_event_call *call = &event_wakeup;
409 struct ring_buffer_event *event;
410 struct ctx_switch_entry *entry;
411 struct trace_buffer *buffer = tr->array_buffer.buffer;
413 event = trace_buffer_lock_reserve(buffer, TRACE_WAKE,
414 sizeof(*entry), flags, pc);
417 entry = ring_buffer_event_data(event);
418 entry->prev_pid = curr->pid;
419 entry->prev_prio = curr->prio;
420 entry->prev_state = task_state_index(curr);
421 entry->next_pid = wakee->pid;
422 entry->next_prio = wakee->prio;
423 entry->next_state = task_state_index(wakee);
424 entry->next_cpu = task_cpu(wakee);
426 if (!call_filter_check_discard(call, entry, buffer, event))
427 trace_buffer_unlock_commit(tr, buffer, event, flags, pc);
431 probe_wakeup_sched_switch(void *ignore, bool preempt,
432 struct task_struct *prev, struct task_struct *next)
434 struct trace_array_cpu *data;
441 tracing_record_cmdline(prev);
443 if (unlikely(!tracer_enabled))
447 * When we start a new trace, we set wakeup_task to NULL
448 * and then set tracer_enabled = 1. We want to make sure
449 * that another CPU does not see the tracer_enabled = 1
450 * and the wakeup_task with an older task, that might
451 * actually be the same as next.
455 if (next != wakeup_task)
458 pc = preempt_count();
460 /* disable local data, not wakeup_cpu data */
461 cpu = raw_smp_processor_id();
462 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
463 if (likely(disabled != 1))
466 local_irq_save(flags);
467 arch_spin_lock(&wakeup_lock);
469 /* We could race with grabbing wakeup_lock */
470 if (unlikely(!tracer_enabled || next != wakeup_task))
473 /* The task we are waiting for is waking up */
474 data = per_cpu_ptr(wakeup_trace->array_buffer.data, wakeup_cpu);
476 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
477 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
478 __trace_stack(wakeup_trace, flags, 0, pc);
480 T0 = data->preempt_timestamp;
481 T1 = ftrace_now(cpu);
484 if (!report_latency(wakeup_trace, delta))
487 if (likely(!is_tracing_stopped())) {
488 wakeup_trace->max_latency = delta;
489 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu, NULL);
493 __wakeup_reset(wakeup_trace);
494 arch_spin_unlock(&wakeup_lock);
495 local_irq_restore(flags);
497 atomic_dec(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
500 static void __wakeup_reset(struct trace_array *tr)
507 put_task_struct(wakeup_task);
512 static void wakeup_reset(struct trace_array *tr)
516 tracing_reset_online_cpus(&tr->array_buffer);
518 local_irq_save(flags);
519 arch_spin_lock(&wakeup_lock);
521 arch_spin_unlock(&wakeup_lock);
522 local_irq_restore(flags);
526 probe_wakeup(void *ignore, struct task_struct *p)
528 struct trace_array_cpu *data;
529 int cpu = smp_processor_id();
534 if (likely(!tracer_enabled))
537 tracing_record_cmdline(p);
538 tracing_record_cmdline(current);
541 * Semantic is like this:
542 * - wakeup tracer handles all tasks in the system, independently
543 * from their scheduling class;
544 * - wakeup_rt tracer handles tasks belonging to sched_dl and
546 * - wakeup_dl handles tasks belonging to sched_dl class only.
548 if (tracing_dl || (wakeup_dl && !dl_task(p)) ||
549 (wakeup_rt && !dl_task(p) && !rt_task(p)) ||
550 (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio)))
553 pc = preempt_count();
554 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
555 if (unlikely(disabled != 1))
558 /* interrupts should be off from try_to_wake_up */
559 arch_spin_lock(&wakeup_lock);
561 /* check for races. */
562 if (!tracer_enabled || tracing_dl ||
563 (!dl_task(p) && p->prio >= wakeup_prio))
566 /* reset the trace */
567 __wakeup_reset(wakeup_trace);
569 wakeup_cpu = task_cpu(p);
570 wakeup_current_cpu = wakeup_cpu;
571 wakeup_prio = p->prio;
574 * Once you start tracing a -deadline task, don't bother tracing
575 * another task until the first one wakes up.
582 wakeup_task = get_task_struct(p);
584 local_save_flags(flags);
586 data = per_cpu_ptr(wakeup_trace->array_buffer.data, wakeup_cpu);
587 data->preempt_timestamp = ftrace_now(cpu);
588 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
589 __trace_stack(wakeup_trace, flags, 0, pc);
592 * We must be careful in using CALLER_ADDR2. But since wake_up
593 * is not called by an assembly function (where as schedule is)
594 * it should be safe to use it here.
596 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
599 arch_spin_unlock(&wakeup_lock);
601 atomic_dec(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
604 static void start_wakeup_tracer(struct trace_array *tr)
608 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
610 pr_info("wakeup trace: Couldn't activate tracepoint"
611 " probe to kernel_sched_wakeup\n");
615 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
617 pr_info("wakeup trace: Couldn't activate tracepoint"
618 " probe to kernel_sched_wakeup_new\n");
622 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
624 pr_info("sched trace: Couldn't activate tracepoint"
625 " probe to kernel_sched_switch\n");
626 goto fail_deprobe_wake_new;
629 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
631 pr_info("wakeup trace: Couldn't activate tracepoint"
632 " probe to kernel_sched_migrate_task\n");
633 goto fail_deprobe_sched_switch;
639 * Don't let the tracer_enabled = 1 show up before
640 * the wakeup_task is reset. This may be overkill since
641 * wakeup_reset does a spin_unlock after setting the
642 * wakeup_task to NULL, but I want to be safe.
643 * This is a slow path anyway.
647 if (start_func_tracer(tr, is_graph(tr)))
648 printk(KERN_ERR "failed to start wakeup tracer\n");
651 fail_deprobe_sched_switch:
652 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
653 fail_deprobe_wake_new:
654 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
656 unregister_trace_sched_wakeup(probe_wakeup, NULL);
659 static void stop_wakeup_tracer(struct trace_array *tr)
662 stop_func_tracer(tr, is_graph(tr));
663 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
664 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
665 unregister_trace_sched_wakeup(probe_wakeup, NULL);
666 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
669 static bool wakeup_busy;
671 static int __wakeup_tracer_init(struct trace_array *tr)
673 save_flags = tr->trace_flags;
675 /* non overwrite screws up the latency tracers */
676 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
677 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
681 ftrace_init_array_ops(tr, wakeup_tracer_call);
682 start_wakeup_tracer(tr);
688 static int wakeup_tracer_init(struct trace_array *tr)
695 return __wakeup_tracer_init(tr);
698 static int wakeup_rt_tracer_init(struct trace_array *tr)
705 return __wakeup_tracer_init(tr);
708 static int wakeup_dl_tracer_init(struct trace_array *tr)
715 return __wakeup_tracer_init(tr);
718 static void wakeup_tracer_reset(struct trace_array *tr)
720 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
721 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
723 stop_wakeup_tracer(tr);
724 /* make sure we put back any tasks we are tracing */
727 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
728 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
729 ftrace_reset_array_ops(tr);
733 static void wakeup_tracer_start(struct trace_array *tr)
739 static void wakeup_tracer_stop(struct trace_array *tr)
744 static struct tracer wakeup_tracer __read_mostly =
747 .init = wakeup_tracer_init,
748 .reset = wakeup_tracer_reset,
749 .start = wakeup_tracer_start,
750 .stop = wakeup_tracer_stop,
752 .print_header = wakeup_print_header,
753 .print_line = wakeup_print_line,
754 .flag_changed = wakeup_flag_changed,
755 #ifdef CONFIG_FTRACE_SELFTEST
756 .selftest = trace_selftest_startup_wakeup,
758 .open = wakeup_trace_open,
759 .close = wakeup_trace_close,
760 .allow_instances = true,
764 static struct tracer wakeup_rt_tracer __read_mostly =
767 .init = wakeup_rt_tracer_init,
768 .reset = wakeup_tracer_reset,
769 .start = wakeup_tracer_start,
770 .stop = wakeup_tracer_stop,
772 .print_header = wakeup_print_header,
773 .print_line = wakeup_print_line,
774 .flag_changed = wakeup_flag_changed,
775 #ifdef CONFIG_FTRACE_SELFTEST
776 .selftest = trace_selftest_startup_wakeup,
778 .open = wakeup_trace_open,
779 .close = wakeup_trace_close,
780 .allow_instances = true,
784 static struct tracer wakeup_dl_tracer __read_mostly =
787 .init = wakeup_dl_tracer_init,
788 .reset = wakeup_tracer_reset,
789 .start = wakeup_tracer_start,
790 .stop = wakeup_tracer_stop,
792 .print_header = wakeup_print_header,
793 .print_line = wakeup_print_line,
794 .flag_changed = wakeup_flag_changed,
795 #ifdef CONFIG_FTRACE_SELFTEST
796 .selftest = trace_selftest_startup_wakeup,
798 .open = wakeup_trace_open,
799 .close = wakeup_trace_close,
800 .allow_instances = true,
804 __init static int init_wakeup_tracer(void)
808 ret = register_tracer(&wakeup_tracer);
812 ret = register_tracer(&wakeup_rt_tracer);
816 ret = register_tracer(&wakeup_dl_tracer);
822 core_initcall(init_wakeup_tracer);