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;
29 static bool wakeup_rt;
30 static bool wakeup_dl;
31 static bool tracing_dl;
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,
70 unsigned int *trace_ctx)
75 if (likely(!wakeup_task))
78 *trace_ctx = tracing_gen_ctx();
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,
116 struct fgraph_ops *gops)
118 struct trace_array *tr = wakeup_trace;
119 struct trace_array_cpu *data;
120 unsigned int trace_ctx;
123 if (ftrace_graph_ignore_func(gops, trace))
126 * Do not trace a function if it's filtered by set_graph_notrace.
127 * Make the index of ret stack negative to indicate that it should
128 * ignore further functions. But it needs its own ret stack entry
129 * to recover the original index in order to continue tracing after
130 * returning from the function.
132 if (ftrace_graph_notrace_addr(trace->func))
135 if (!func_prolog_preempt_disable(tr, &data, &trace_ctx))
138 ret = __trace_graph_entry(tr, trace, trace_ctx);
139 atomic_dec(&data->disabled);
140 preempt_enable_notrace();
145 static void wakeup_graph_return(struct ftrace_graph_ret *trace,
146 struct fgraph_ops *gops)
148 struct trace_array *tr = wakeup_trace;
149 struct trace_array_cpu *data;
150 unsigned int trace_ctx;
152 ftrace_graph_addr_finish(gops, trace);
154 if (!func_prolog_preempt_disable(tr, &data, &trace_ctx))
157 __trace_graph_return(tr, trace, trace_ctx);
158 atomic_dec(&data->disabled);
160 preempt_enable_notrace();
164 static struct fgraph_ops fgraph_wakeup_ops = {
165 .entryfunc = &wakeup_graph_entry,
166 .retfunc = &wakeup_graph_return,
169 static void wakeup_trace_open(struct trace_iterator *iter)
171 if (is_graph(iter->tr))
172 graph_trace_open(iter);
174 iter->private = NULL;
177 static void wakeup_trace_close(struct trace_iterator *iter)
180 graph_trace_close(iter);
183 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
184 TRACE_GRAPH_PRINT_CPU | \
185 TRACE_GRAPH_PRINT_REL_TIME | \
186 TRACE_GRAPH_PRINT_DURATION | \
187 TRACE_GRAPH_PRINT_OVERHEAD | \
188 TRACE_GRAPH_PRINT_IRQS)
190 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
193 * In graph mode call the graph tracer output function,
194 * otherwise go with the TRACE_FN event handler
196 if (is_graph(iter->tr))
197 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
199 return TRACE_TYPE_UNHANDLED;
202 static void wakeup_print_header(struct seq_file *s)
204 if (is_graph(wakeup_trace))
205 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
207 trace_default_header(s);
209 #endif /* else CONFIG_FUNCTION_GRAPH_TRACER */
212 * wakeup uses its own tracer function to keep the overhead down:
215 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip,
216 struct ftrace_ops *op, struct ftrace_regs *fregs)
218 struct trace_array *tr = wakeup_trace;
219 struct trace_array_cpu *data;
221 unsigned int trace_ctx;
223 if (!func_prolog_preempt_disable(tr, &data, &trace_ctx))
226 local_irq_save(flags);
227 trace_function(tr, ip, parent_ip, trace_ctx);
228 local_irq_restore(flags);
230 atomic_dec(&data->disabled);
231 preempt_enable_notrace();
234 static int register_wakeup_function(struct trace_array *tr, int graph, int set)
238 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
239 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
243 ret = register_ftrace_graph(&fgraph_wakeup_ops);
245 ret = register_ftrace_function(tr->ops);
248 function_enabled = true;
253 static void unregister_wakeup_function(struct trace_array *tr, int graph)
255 if (!function_enabled)
259 unregister_ftrace_graph(&fgraph_wakeup_ops);
261 unregister_ftrace_function(tr->ops);
263 function_enabled = false;
266 static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
268 if (!(mask & TRACE_ITER_FUNCTION))
272 register_wakeup_function(tr, is_graph(tr), 1);
274 unregister_wakeup_function(tr, is_graph(tr));
277 #else /* CONFIG_FUNCTION_TRACER */
278 static int register_wakeup_function(struct trace_array *tr, int graph, int set)
282 static void unregister_wakeup_function(struct trace_array *tr, int graph) { }
283 static int wakeup_function_set(struct trace_array *tr, u32 mask, int set)
287 #endif /* else CONFIG_FUNCTION_TRACER */
289 #ifndef CONFIG_FUNCTION_GRAPH_TRACER
290 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
292 return TRACE_TYPE_UNHANDLED;
295 static void wakeup_trace_open(struct trace_iterator *iter) { }
296 static void wakeup_trace_close(struct trace_iterator *iter) { }
298 static void wakeup_print_header(struct seq_file *s)
300 trace_default_header(s);
302 #endif /* !CONFIG_FUNCTION_GRAPH_TRACER */
305 __trace_function(struct trace_array *tr,
306 unsigned long ip, unsigned long parent_ip,
307 unsigned int trace_ctx)
310 trace_graph_function(tr, ip, parent_ip, trace_ctx);
312 trace_function(tr, ip, parent_ip, trace_ctx);
315 static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set)
317 struct tracer *tracer = tr->current_trace;
319 if (wakeup_function_set(tr, mask, set))
322 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
323 if (mask & TRACE_ITER_DISPLAY_GRAPH)
324 return wakeup_display_graph(tr, set);
327 return trace_keep_overwrite(tracer, mask, set);
330 static int start_func_tracer(struct trace_array *tr, int graph)
334 ret = register_wakeup_function(tr, graph, 0);
336 if (!ret && tracing_is_enabled())
344 static void stop_func_tracer(struct trace_array *tr, int graph)
348 unregister_wakeup_function(tr, graph);
352 * Should this new latency be reported/recorded?
354 static bool report_latency(struct trace_array *tr, u64 delta)
356 if (tracing_thresh) {
357 if (delta < tracing_thresh)
360 if (delta <= tr->max_latency)
367 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
369 if (task != wakeup_task)
372 wakeup_current_cpu = cpu;
376 tracing_sched_switch_trace(struct trace_array *tr,
377 struct task_struct *prev,
378 struct task_struct *next,
379 unsigned int trace_ctx)
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), trace_ctx);
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 trace_buffer_unlock_commit(tr, buffer, event, trace_ctx);
402 tracing_sched_wakeup_trace(struct trace_array *tr,
403 struct task_struct *wakee,
404 struct task_struct *curr,
405 unsigned int trace_ctx)
407 struct ring_buffer_event *event;
408 struct ctx_switch_entry *entry;
409 struct trace_buffer *buffer = tr->array_buffer.buffer;
411 event = trace_buffer_lock_reserve(buffer, TRACE_WAKE,
412 sizeof(*entry), trace_ctx);
415 entry = ring_buffer_event_data(event);
416 entry->prev_pid = curr->pid;
417 entry->prev_prio = curr->prio;
418 entry->prev_state = task_state_index(curr);
419 entry->next_pid = wakee->pid;
420 entry->next_prio = wakee->prio;
421 entry->next_state = task_state_index(wakee);
422 entry->next_cpu = task_cpu(wakee);
424 trace_buffer_unlock_commit(tr, buffer, event, trace_ctx);
428 probe_wakeup_sched_switch(void *ignore, bool preempt,
429 struct task_struct *prev, struct task_struct *next,
430 unsigned int prev_state)
432 struct trace_array_cpu *data;
437 unsigned int trace_ctx;
439 tracing_record_cmdline(prev);
441 if (unlikely(!tracer_enabled))
445 * When we start a new trace, we set wakeup_task to NULL
446 * and then set tracer_enabled = 1. We want to make sure
447 * that another CPU does not see the tracer_enabled = 1
448 * and the wakeup_task with an older task, that might
449 * actually be the same as next.
453 if (next != wakeup_task)
456 /* disable local data, not wakeup_cpu data */
457 cpu = raw_smp_processor_id();
458 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
459 if (likely(disabled != 1))
462 local_irq_save(flags);
463 trace_ctx = tracing_gen_ctx_flags(flags);
465 arch_spin_lock(&wakeup_lock);
467 /* We could race with grabbing wakeup_lock */
468 if (unlikely(!tracer_enabled || next != wakeup_task))
471 /* The task we are waiting for is waking up */
472 data = per_cpu_ptr(wakeup_trace->array_buffer.data, wakeup_cpu);
474 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, trace_ctx);
475 tracing_sched_switch_trace(wakeup_trace, prev, next, trace_ctx);
476 __trace_stack(wakeup_trace, trace_ctx, 0);
478 T0 = data->preempt_timestamp;
479 T1 = ftrace_now(cpu);
482 if (!report_latency(wakeup_trace, delta))
485 if (likely(!is_tracing_stopped())) {
486 wakeup_trace->max_latency = delta;
487 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu, NULL);
491 __wakeup_reset(wakeup_trace);
492 arch_spin_unlock(&wakeup_lock);
493 local_irq_restore(flags);
495 atomic_dec(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
498 static void __wakeup_reset(struct trace_array *tr)
505 put_task_struct(wakeup_task);
510 static void wakeup_reset(struct trace_array *tr)
514 tracing_reset_online_cpus(&tr->array_buffer);
516 local_irq_save(flags);
517 arch_spin_lock(&wakeup_lock);
519 arch_spin_unlock(&wakeup_lock);
520 local_irq_restore(flags);
524 probe_wakeup(void *ignore, struct task_struct *p)
526 struct trace_array_cpu *data;
527 int cpu = smp_processor_id();
529 unsigned int trace_ctx;
531 if (likely(!tracer_enabled))
534 tracing_record_cmdline(p);
535 tracing_record_cmdline(current);
538 * Semantic is like this:
539 * - wakeup tracer handles all tasks in the system, independently
540 * from their scheduling class;
541 * - wakeup_rt tracer handles tasks belonging to sched_dl and
543 * - wakeup_dl handles tasks belonging to sched_dl class only.
545 if (tracing_dl || (wakeup_dl && !dl_task(p)) ||
546 (wakeup_rt && !rt_or_dl_task(p)) ||
547 (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio)))
550 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
551 if (unlikely(disabled != 1))
554 trace_ctx = tracing_gen_ctx();
556 /* interrupts should be off from try_to_wake_up */
557 arch_spin_lock(&wakeup_lock);
559 /* check for races. */
560 if (!tracer_enabled || tracing_dl ||
561 (!dl_task(p) && p->prio >= wakeup_prio))
564 /* reset the trace */
565 __wakeup_reset(wakeup_trace);
567 wakeup_cpu = task_cpu(p);
568 wakeup_current_cpu = wakeup_cpu;
569 wakeup_prio = p->prio;
572 * Once you start tracing a -deadline task, don't bother tracing
573 * another task until the first one wakes up.
580 wakeup_task = get_task_struct(p);
582 data = per_cpu_ptr(wakeup_trace->array_buffer.data, wakeup_cpu);
583 data->preempt_timestamp = ftrace_now(cpu);
584 tracing_sched_wakeup_trace(wakeup_trace, p, current, trace_ctx);
585 __trace_stack(wakeup_trace, trace_ctx, 0);
588 * We must be careful in using CALLER_ADDR2. But since wake_up
589 * is not called by an assembly function (where as schedule is)
590 * it should be safe to use it here.
592 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, trace_ctx);
595 arch_spin_unlock(&wakeup_lock);
597 atomic_dec(&per_cpu_ptr(wakeup_trace->array_buffer.data, cpu)->disabled);
600 static void start_wakeup_tracer(struct trace_array *tr)
604 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
606 pr_info("wakeup trace: Couldn't activate tracepoint"
607 " probe to kernel_sched_wakeup\n");
611 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
613 pr_info("wakeup trace: Couldn't activate tracepoint"
614 " probe to kernel_sched_wakeup_new\n");
618 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
620 pr_info("sched trace: Couldn't activate tracepoint"
621 " probe to kernel_sched_switch\n");
622 goto fail_deprobe_wake_new;
625 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
627 pr_info("wakeup trace: Couldn't activate tracepoint"
628 " probe to kernel_sched_migrate_task\n");
629 goto fail_deprobe_sched_switch;
635 * Don't let the tracer_enabled = 1 show up before
636 * the wakeup_task is reset. This may be overkill since
637 * wakeup_reset does a spin_unlock after setting the
638 * wakeup_task to NULL, but I want to be safe.
639 * This is a slow path anyway.
643 if (start_func_tracer(tr, is_graph(tr)))
644 printk(KERN_ERR "failed to start wakeup tracer\n");
647 fail_deprobe_sched_switch:
648 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
649 fail_deprobe_wake_new:
650 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
652 unregister_trace_sched_wakeup(probe_wakeup, NULL);
655 static void stop_wakeup_tracer(struct trace_array *tr)
658 stop_func_tracer(tr, is_graph(tr));
659 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
660 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
661 unregister_trace_sched_wakeup(probe_wakeup, NULL);
662 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
665 static bool wakeup_busy;
667 static int __wakeup_tracer_init(struct trace_array *tr)
669 save_flags = tr->trace_flags;
671 /* non overwrite screws up the latency tracers */
672 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
673 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
677 ftrace_init_array_ops(tr, wakeup_tracer_call);
678 start_wakeup_tracer(tr);
684 static int wakeup_tracer_init(struct trace_array *tr)
691 return __wakeup_tracer_init(tr);
694 static int wakeup_rt_tracer_init(struct trace_array *tr)
701 return __wakeup_tracer_init(tr);
704 static int wakeup_dl_tracer_init(struct trace_array *tr)
711 return __wakeup_tracer_init(tr);
714 static void wakeup_tracer_reset(struct trace_array *tr)
716 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
717 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
719 stop_wakeup_tracer(tr);
720 /* make sure we put back any tasks we are tracing */
723 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
724 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
725 ftrace_reset_array_ops(tr);
729 static void wakeup_tracer_start(struct trace_array *tr)
735 static void wakeup_tracer_stop(struct trace_array *tr)
740 static struct tracer wakeup_tracer __read_mostly =
743 .init = wakeup_tracer_init,
744 .reset = wakeup_tracer_reset,
745 .start = wakeup_tracer_start,
746 .stop = wakeup_tracer_stop,
748 .print_header = wakeup_print_header,
749 .print_line = wakeup_print_line,
750 .flag_changed = wakeup_flag_changed,
751 #ifdef CONFIG_FTRACE_SELFTEST
752 .selftest = trace_selftest_startup_wakeup,
754 .open = wakeup_trace_open,
755 .close = wakeup_trace_close,
756 .allow_instances = true,
760 static struct tracer wakeup_rt_tracer __read_mostly =
763 .init = wakeup_rt_tracer_init,
764 .reset = wakeup_tracer_reset,
765 .start = wakeup_tracer_start,
766 .stop = wakeup_tracer_stop,
768 .print_header = wakeup_print_header,
769 .print_line = wakeup_print_line,
770 .flag_changed = wakeup_flag_changed,
771 #ifdef CONFIG_FTRACE_SELFTEST
772 .selftest = trace_selftest_startup_wakeup,
774 .open = wakeup_trace_open,
775 .close = wakeup_trace_close,
776 .allow_instances = true,
780 static struct tracer wakeup_dl_tracer __read_mostly =
783 .init = wakeup_dl_tracer_init,
784 .reset = wakeup_tracer_reset,
785 .start = wakeup_tracer_start,
786 .stop = wakeup_tracer_stop,
788 .print_header = wakeup_print_header,
789 .print_line = wakeup_print_line,
790 .flag_changed = wakeup_flag_changed,
791 #ifdef CONFIG_FTRACE_SELFTEST
792 .selftest = trace_selftest_startup_wakeup,
794 .open = wakeup_trace_open,
795 .close = wakeup_trace_close,
796 .allow_instances = true,
800 __init static int init_wakeup_tracer(void)
804 ret = register_tracer(&wakeup_tracer);
808 ret = register_tracer(&wakeup_rt_tracer);
812 ret = register_tracer(&wakeup_dl_tracer);
818 core_initcall(init_wakeup_tracer);