2 * trace task wakeup timings
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
7 * Based on code from the latency_tracer, that is:
9 * Copyright (C) 2004-2006 Ingo Molnar
10 * Copyright (C) 2004 William Lee Irwin III
12 #include <linux/module.h>
14 #include <linux/debugfs.h>
15 #include <linux/kallsyms.h>
16 #include <linux/uaccess.h>
17 #include <linux/ftrace.h>
18 #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 arch_spinlock_t wakeup_lock =
32 (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
34 static void wakeup_reset(struct trace_array *tr);
35 static void __wakeup_reset(struct trace_array *tr);
36 static int wakeup_graph_entry(struct ftrace_graph_ent *trace);
37 static void wakeup_graph_return(struct ftrace_graph_ret *trace);
39 static int save_lat_flag;
41 #define TRACE_DISPLAY_GRAPH 1
43 static struct tracer_opt trace_opts[] = {
44 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
45 /* display latency trace as call graph */
46 { TRACER_OPT(display-graph, TRACE_DISPLAY_GRAPH) },
51 static struct tracer_flags tracer_flags = {
56 #define is_graph() (tracer_flags.val & TRACE_DISPLAY_GRAPH)
58 #ifdef CONFIG_FUNCTION_TRACER
61 * Prologue for the wakeup function tracers.
63 * Returns 1 if it is OK to continue, and preemption
64 * is disabled and data->disabled is incremented.
65 * 0 if the trace is to be ignored, and preemption
66 * is not disabled and data->disabled is
69 * Note, this function is also used outside this ifdef but
70 * inside the #ifdef of the function graph tracer below.
71 * This is OK, since the function graph tracer is
72 * dependent on the function tracer.
75 func_prolog_preempt_disable(struct trace_array *tr,
76 struct trace_array_cpu **data,
82 if (likely(!wakeup_task))
85 *pc = preempt_count();
86 preempt_disable_notrace();
88 cpu = raw_smp_processor_id();
89 if (cpu != wakeup_current_cpu)
92 *data = tr->data[cpu];
93 disabled = atomic_inc_return(&(*data)->disabled);
94 if (unlikely(disabled != 1))
100 atomic_dec(&(*data)->disabled);
103 preempt_enable_notrace();
108 * wakeup uses its own tracer function to keep the overhead down:
111 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
113 struct trace_array *tr = wakeup_trace;
114 struct trace_array_cpu *data;
118 if (!func_prolog_preempt_disable(tr, &data, &pc))
121 local_irq_save(flags);
122 trace_function(tr, ip, parent_ip, flags, pc);
123 local_irq_restore(flags);
125 atomic_dec(&data->disabled);
126 preempt_enable_notrace();
129 static struct ftrace_ops trace_ops __read_mostly =
131 .func = wakeup_tracer_call,
132 .flags = FTRACE_OPS_FL_GLOBAL,
134 #endif /* CONFIG_FUNCTION_TRACER */
136 static int start_func_tracer(int graph)
141 ret = register_ftrace_function(&trace_ops);
143 ret = register_ftrace_graph(&wakeup_graph_return,
144 &wakeup_graph_entry);
146 if (!ret && tracing_is_enabled())
154 static void stop_func_tracer(int graph)
159 unregister_ftrace_function(&trace_ops);
161 unregister_ftrace_graph();
164 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
165 static int wakeup_set_flag(u32 old_flags, u32 bit, int set)
168 if (!(bit & TRACE_DISPLAY_GRAPH))
171 if (!(is_graph() ^ set))
174 stop_func_tracer(!set);
176 wakeup_reset(wakeup_trace);
177 tracing_max_latency = 0;
179 return start_func_tracer(set);
182 static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
184 struct trace_array *tr = wakeup_trace;
185 struct trace_array_cpu *data;
189 if (!func_prolog_preempt_disable(tr, &data, &pc))
192 local_save_flags(flags);
193 ret = __trace_graph_entry(tr, trace, flags, pc);
194 atomic_dec(&data->disabled);
195 preempt_enable_notrace();
200 static void wakeup_graph_return(struct ftrace_graph_ret *trace)
202 struct trace_array *tr = wakeup_trace;
203 struct trace_array_cpu *data;
207 if (!func_prolog_preempt_disable(tr, &data, &pc))
210 local_save_flags(flags);
211 __trace_graph_return(tr, trace, flags, pc);
212 atomic_dec(&data->disabled);
214 preempt_enable_notrace();
218 static void wakeup_trace_open(struct trace_iterator *iter)
221 graph_trace_open(iter);
224 static void wakeup_trace_close(struct trace_iterator *iter)
227 graph_trace_close(iter);
230 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
231 TRACE_GRAPH_PRINT_ABS_TIME | \
232 TRACE_GRAPH_PRINT_DURATION)
234 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
237 * In graph mode call the graph tracer output function,
238 * otherwise go with the TRACE_FN event handler
241 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
243 return TRACE_TYPE_UNHANDLED;
246 static void wakeup_print_header(struct seq_file *s)
249 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
251 trace_default_header(s);
255 __trace_function(struct trace_array *tr,
256 unsigned long ip, unsigned long parent_ip,
257 unsigned long flags, int pc)
260 trace_graph_function(tr, ip, parent_ip, flags, pc);
262 trace_function(tr, ip, parent_ip, flags, pc);
265 #define __trace_function trace_function
267 static int wakeup_set_flag(u32 old_flags, u32 bit, int set)
272 static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
277 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
279 return TRACE_TYPE_UNHANDLED;
282 static void wakeup_graph_return(struct ftrace_graph_ret *trace) { }
283 static void wakeup_trace_open(struct trace_iterator *iter) { }
284 static void wakeup_trace_close(struct trace_iterator *iter) { }
286 #ifdef CONFIG_FUNCTION_TRACER
287 static void wakeup_print_header(struct seq_file *s)
289 trace_default_header(s);
292 static void wakeup_print_header(struct seq_file *s)
294 trace_latency_header(s);
296 #endif /* CONFIG_FUNCTION_TRACER */
297 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
300 * Should this new latency be reported/recorded?
302 static int report_latency(cycle_t delta)
304 if (tracing_thresh) {
305 if (delta < tracing_thresh)
308 if (delta <= tracing_max_latency)
315 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
317 if (task != wakeup_task)
320 wakeup_current_cpu = cpu;
324 probe_wakeup_sched_switch(void *ignore,
325 struct task_struct *prev, struct task_struct *next)
327 struct trace_array_cpu *data;
328 cycle_t T0, T1, delta;
334 tracing_record_cmdline(prev);
336 if (unlikely(!tracer_enabled))
340 * When we start a new trace, we set wakeup_task to NULL
341 * and then set tracer_enabled = 1. We want to make sure
342 * that another CPU does not see the tracer_enabled = 1
343 * and the wakeup_task with an older task, that might
344 * actually be the same as next.
348 if (next != wakeup_task)
351 pc = preempt_count();
353 /* disable local data, not wakeup_cpu data */
354 cpu = raw_smp_processor_id();
355 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
356 if (likely(disabled != 1))
359 local_irq_save(flags);
360 arch_spin_lock(&wakeup_lock);
362 /* We could race with grabbing wakeup_lock */
363 if (unlikely(!tracer_enabled || next != wakeup_task))
366 /* The task we are waiting for is waking up */
367 data = wakeup_trace->data[wakeup_cpu];
369 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
370 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
372 T0 = data->preempt_timestamp;
373 T1 = ftrace_now(cpu);
376 if (!report_latency(delta))
379 if (likely(!is_tracing_stopped())) {
380 tracing_max_latency = delta;
381 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
385 __wakeup_reset(wakeup_trace);
386 arch_spin_unlock(&wakeup_lock);
387 local_irq_restore(flags);
389 atomic_dec(&wakeup_trace->data[cpu]->disabled);
392 static void __wakeup_reset(struct trace_array *tr)
398 put_task_struct(wakeup_task);
403 static void wakeup_reset(struct trace_array *tr)
407 tracing_reset_online_cpus(tr);
409 local_irq_save(flags);
410 arch_spin_lock(&wakeup_lock);
412 arch_spin_unlock(&wakeup_lock);
413 local_irq_restore(flags);
417 probe_wakeup(void *ignore, struct task_struct *p, int success)
419 struct trace_array_cpu *data;
420 int cpu = smp_processor_id();
425 if (likely(!tracer_enabled))
428 tracing_record_cmdline(p);
429 tracing_record_cmdline(current);
431 if ((wakeup_rt && !rt_task(p)) ||
432 p->prio >= wakeup_prio ||
433 p->prio >= current->prio)
436 pc = preempt_count();
437 disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled);
438 if (unlikely(disabled != 1))
441 /* interrupts should be off from try_to_wake_up */
442 arch_spin_lock(&wakeup_lock);
444 /* check for races. */
445 if (!tracer_enabled || p->prio >= wakeup_prio)
448 /* reset the trace */
449 __wakeup_reset(wakeup_trace);
451 wakeup_cpu = task_cpu(p);
452 wakeup_current_cpu = wakeup_cpu;
453 wakeup_prio = p->prio;
456 get_task_struct(wakeup_task);
458 local_save_flags(flags);
460 data = wakeup_trace->data[wakeup_cpu];
461 data->preempt_timestamp = ftrace_now(cpu);
462 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
465 * We must be careful in using CALLER_ADDR2. But since wake_up
466 * is not called by an assembly function (where as schedule is)
467 * it should be safe to use it here.
469 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
472 arch_spin_unlock(&wakeup_lock);
474 atomic_dec(&wakeup_trace->data[cpu]->disabled);
477 static void start_wakeup_tracer(struct trace_array *tr)
481 ret = register_trace_sched_wakeup(probe_wakeup, NULL);
483 pr_info("wakeup trace: Couldn't activate tracepoint"
484 " probe to kernel_sched_wakeup\n");
488 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
490 pr_info("wakeup trace: Couldn't activate tracepoint"
491 " probe to kernel_sched_wakeup_new\n");
495 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
497 pr_info("sched trace: Couldn't activate tracepoint"
498 " probe to kernel_sched_switch\n");
499 goto fail_deprobe_wake_new;
502 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
504 pr_info("wakeup trace: Couldn't activate tracepoint"
505 " probe to kernel_sched_migrate_task\n");
512 * Don't let the tracer_enabled = 1 show up before
513 * the wakeup_task is reset. This may be overkill since
514 * wakeup_reset does a spin_unlock after setting the
515 * wakeup_task to NULL, but I want to be safe.
516 * This is a slow path anyway.
520 if (start_func_tracer(is_graph()))
521 printk(KERN_ERR "failed to start wakeup tracer\n");
524 fail_deprobe_wake_new:
525 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
527 unregister_trace_sched_wakeup(probe_wakeup, NULL);
530 static void stop_wakeup_tracer(struct trace_array *tr)
533 stop_func_tracer(is_graph());
534 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
535 unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
536 unregister_trace_sched_wakeup(probe_wakeup, NULL);
537 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
540 static int __wakeup_tracer_init(struct trace_array *tr)
542 save_lat_flag = trace_flags & TRACE_ITER_LATENCY_FMT;
543 trace_flags |= TRACE_ITER_LATENCY_FMT;
545 tracing_max_latency = 0;
547 start_wakeup_tracer(tr);
551 static int wakeup_tracer_init(struct trace_array *tr)
554 return __wakeup_tracer_init(tr);
557 static int wakeup_rt_tracer_init(struct trace_array *tr)
560 return __wakeup_tracer_init(tr);
563 static void wakeup_tracer_reset(struct trace_array *tr)
565 stop_wakeup_tracer(tr);
566 /* make sure we put back any tasks we are tracing */
570 trace_flags &= ~TRACE_ITER_LATENCY_FMT;
573 static void wakeup_tracer_start(struct trace_array *tr)
579 static void wakeup_tracer_stop(struct trace_array *tr)
584 static struct tracer wakeup_tracer __read_mostly =
587 .init = wakeup_tracer_init,
588 .reset = wakeup_tracer_reset,
589 .start = wakeup_tracer_start,
590 .stop = wakeup_tracer_stop,
592 .print_header = wakeup_print_header,
593 .print_line = wakeup_print_line,
594 .flags = &tracer_flags,
595 .set_flag = wakeup_set_flag,
596 #ifdef CONFIG_FTRACE_SELFTEST
597 .selftest = trace_selftest_startup_wakeup,
599 .open = wakeup_trace_open,
600 .close = wakeup_trace_close,
604 static struct tracer wakeup_rt_tracer __read_mostly =
607 .init = wakeup_rt_tracer_init,
608 .reset = wakeup_tracer_reset,
609 .start = wakeup_tracer_start,
610 .stop = wakeup_tracer_stop,
611 .wait_pipe = poll_wait_pipe,
613 .print_header = wakeup_print_header,
614 .print_line = wakeup_print_line,
615 .flags = &tracer_flags,
616 .set_flag = wakeup_set_flag,
617 #ifdef CONFIG_FTRACE_SELFTEST
618 .selftest = trace_selftest_startup_wakeup,
620 .open = wakeup_trace_open,
621 .close = wakeup_trace_close,
625 __init static int init_wakeup_tracer(void)
629 ret = register_tracer(&wakeup_tracer);
633 ret = register_tracer(&wakeup_rt_tracer);
639 device_initcall(init_wakeup_tracer);