Merge tag 'vfio-v3.15-rc1' of git://github.com/awilliam/linux-vfio
[cascardo/linux.git] / kernel / trace / trace_sched_wakeup.c
1 /*
2  * trace task wakeup timings
3  *
4  * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6  *
7  * Based on code from the latency_tracer, that is:
8  *
9  *  Copyright (C) 2004-2006 Ingo Molnar
10  *  Copyright (C) 2004 Nadia Yvette Chambers
11  */
12 #include <linux/module.h>
13 #include <linux/fs.h>
14 #include <linux/debugfs.h>
15 #include <linux/kallsyms.h>
16 #include <linux/uaccess.h>
17 #include <linux/ftrace.h>
18 #include <linux/sched/rt.h>
19 #include <linux/sched/deadline.h>
20 #include <trace/events/sched.h>
21 #include "trace.h"
22
23 static struct trace_array       *wakeup_trace;
24 static int __read_mostly        tracer_enabled;
25
26 static struct task_struct       *wakeup_task;
27 static int                      wakeup_cpu;
28 static int                      wakeup_current_cpu;
29 static unsigned                 wakeup_prio = -1;
30 static int                      wakeup_rt;
31 static int                      wakeup_dl;
32 static int                      tracing_dl = 0;
33
34 static arch_spinlock_t wakeup_lock =
35         (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
36
37 static void wakeup_reset(struct trace_array *tr);
38 static void __wakeup_reset(struct trace_array *tr);
39 static int wakeup_graph_entry(struct ftrace_graph_ent *trace);
40 static void wakeup_graph_return(struct ftrace_graph_ret *trace);
41
42 static int save_flags;
43 static bool function_enabled;
44
45 #define TRACE_DISPLAY_GRAPH     1
46
47 static struct tracer_opt trace_opts[] = {
48 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
49         /* display latency trace as call graph */
50         { TRACER_OPT(display-graph, TRACE_DISPLAY_GRAPH) },
51 #endif
52         { } /* Empty entry */
53 };
54
55 static struct tracer_flags tracer_flags = {
56         .val  = 0,
57         .opts = trace_opts,
58 };
59
60 #define is_graph() (tracer_flags.val & TRACE_DISPLAY_GRAPH)
61
62 #ifdef CONFIG_FUNCTION_TRACER
63
64 /*
65  * Prologue for the wakeup function tracers.
66  *
67  * Returns 1 if it is OK to continue, and preemption
68  *            is disabled and data->disabled is incremented.
69  *         0 if the trace is to be ignored, and preemption
70  *            is not disabled and data->disabled is
71  *            kept the same.
72  *
73  * Note, this function is also used outside this ifdef but
74  *  inside the #ifdef of the function graph tracer below.
75  *  This is OK, since the function graph tracer is
76  *  dependent on the function tracer.
77  */
78 static int
79 func_prolog_preempt_disable(struct trace_array *tr,
80                             struct trace_array_cpu **data,
81                             int *pc)
82 {
83         long disabled;
84         int cpu;
85
86         if (likely(!wakeup_task))
87                 return 0;
88
89         *pc = preempt_count();
90         preempt_disable_notrace();
91
92         cpu = raw_smp_processor_id();
93         if (cpu != wakeup_current_cpu)
94                 goto out_enable;
95
96         *data = per_cpu_ptr(tr->trace_buffer.data, cpu);
97         disabled = atomic_inc_return(&(*data)->disabled);
98         if (unlikely(disabled != 1))
99                 goto out;
100
101         return 1;
102
103 out:
104         atomic_dec(&(*data)->disabled);
105
106 out_enable:
107         preempt_enable_notrace();
108         return 0;
109 }
110
111 /*
112  * wakeup uses its own tracer function to keep the overhead down:
113  */
114 static void
115 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip,
116                    struct ftrace_ops *op, struct pt_regs *pt_regs)
117 {
118         struct trace_array *tr = wakeup_trace;
119         struct trace_array_cpu *data;
120         unsigned long flags;
121         int pc;
122
123         if (!func_prolog_preempt_disable(tr, &data, &pc))
124                 return;
125
126         local_irq_save(flags);
127         trace_function(tr, ip, parent_ip, flags, pc);
128         local_irq_restore(flags);
129
130         atomic_dec(&data->disabled);
131         preempt_enable_notrace();
132 }
133
134 static struct ftrace_ops trace_ops __read_mostly =
135 {
136         .func = wakeup_tracer_call,
137         .flags = FTRACE_OPS_FL_GLOBAL | FTRACE_OPS_FL_RECURSION_SAFE,
138 };
139 #endif /* CONFIG_FUNCTION_TRACER */
140
141 static int register_wakeup_function(int graph, int set)
142 {
143         int ret;
144
145         /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
146         if (function_enabled || (!set && !(trace_flags & TRACE_ITER_FUNCTION)))
147                 return 0;
148
149         if (graph)
150                 ret = register_ftrace_graph(&wakeup_graph_return,
151                                             &wakeup_graph_entry);
152         else
153                 ret = register_ftrace_function(&trace_ops);
154
155         if (!ret)
156                 function_enabled = true;
157
158         return ret;
159 }
160
161 static void unregister_wakeup_function(int graph)
162 {
163         if (!function_enabled)
164                 return;
165
166         if (graph)
167                 unregister_ftrace_graph();
168         else
169                 unregister_ftrace_function(&trace_ops);
170
171         function_enabled = false;
172 }
173
174 static void wakeup_function_set(int set)
175 {
176         if (set)
177                 register_wakeup_function(is_graph(), 1);
178         else
179                 unregister_wakeup_function(is_graph());
180 }
181
182 static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set)
183 {
184         struct tracer *tracer = tr->current_trace;
185
186         if (mask & TRACE_ITER_FUNCTION)
187                 wakeup_function_set(set);
188
189         return trace_keep_overwrite(tracer, mask, set);
190 }
191
192 static int start_func_tracer(int graph)
193 {
194         int ret;
195
196         ret = register_wakeup_function(graph, 0);
197
198         if (!ret && tracing_is_enabled())
199                 tracer_enabled = 1;
200         else
201                 tracer_enabled = 0;
202
203         return ret;
204 }
205
206 static void stop_func_tracer(int graph)
207 {
208         tracer_enabled = 0;
209
210         unregister_wakeup_function(graph);
211 }
212
213 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
214 static int
215 wakeup_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
216 {
217
218         if (!(bit & TRACE_DISPLAY_GRAPH))
219                 return -EINVAL;
220
221         if (!(is_graph() ^ set))
222                 return 0;
223
224         stop_func_tracer(!set);
225
226         wakeup_reset(wakeup_trace);
227         tracing_max_latency = 0;
228
229         return start_func_tracer(set);
230 }
231
232 static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
233 {
234         struct trace_array *tr = wakeup_trace;
235         struct trace_array_cpu *data;
236         unsigned long flags;
237         int pc, ret = 0;
238
239         if (!func_prolog_preempt_disable(tr, &data, &pc))
240                 return 0;
241
242         local_save_flags(flags);
243         ret = __trace_graph_entry(tr, trace, flags, pc);
244         atomic_dec(&data->disabled);
245         preempt_enable_notrace();
246
247         return ret;
248 }
249
250 static void wakeup_graph_return(struct ftrace_graph_ret *trace)
251 {
252         struct trace_array *tr = wakeup_trace;
253         struct trace_array_cpu *data;
254         unsigned long flags;
255         int pc;
256
257         if (!func_prolog_preempt_disable(tr, &data, &pc))
258                 return;
259
260         local_save_flags(flags);
261         __trace_graph_return(tr, trace, flags, pc);
262         atomic_dec(&data->disabled);
263
264         preempt_enable_notrace();
265         return;
266 }
267
268 static void wakeup_trace_open(struct trace_iterator *iter)
269 {
270         if (is_graph())
271                 graph_trace_open(iter);
272 }
273
274 static void wakeup_trace_close(struct trace_iterator *iter)
275 {
276         if (iter->private)
277                 graph_trace_close(iter);
278 }
279
280 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \
281                             TRACE_GRAPH_PRINT_ABS_TIME | \
282                             TRACE_GRAPH_PRINT_DURATION)
283
284 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
285 {
286         /*
287          * In graph mode call the graph tracer output function,
288          * otherwise go with the TRACE_FN event handler
289          */
290         if (is_graph())
291                 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
292
293         return TRACE_TYPE_UNHANDLED;
294 }
295
296 static void wakeup_print_header(struct seq_file *s)
297 {
298         if (is_graph())
299                 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
300         else
301                 trace_default_header(s);
302 }
303
304 static void
305 __trace_function(struct trace_array *tr,
306                  unsigned long ip, unsigned long parent_ip,
307                  unsigned long flags, int pc)
308 {
309         if (is_graph())
310                 trace_graph_function(tr, ip, parent_ip, flags, pc);
311         else
312                 trace_function(tr, ip, parent_ip, flags, pc);
313 }
314 #else
315 #define __trace_function trace_function
316
317 static int
318 wakeup_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
319 {
320         return -EINVAL;
321 }
322
323 static int wakeup_graph_entry(struct ftrace_graph_ent *trace)
324 {
325         return -1;
326 }
327
328 static enum print_line_t wakeup_print_line(struct trace_iterator *iter)
329 {
330         return TRACE_TYPE_UNHANDLED;
331 }
332
333 static void wakeup_graph_return(struct ftrace_graph_ret *trace) { }
334 static void wakeup_trace_open(struct trace_iterator *iter) { }
335 static void wakeup_trace_close(struct trace_iterator *iter) { }
336
337 #ifdef CONFIG_FUNCTION_TRACER
338 static void wakeup_print_header(struct seq_file *s)
339 {
340         trace_default_header(s);
341 }
342 #else
343 static void wakeup_print_header(struct seq_file *s)
344 {
345         trace_latency_header(s);
346 }
347 #endif /* CONFIG_FUNCTION_TRACER */
348 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
349
350 /*
351  * Should this new latency be reported/recorded?
352  */
353 static int report_latency(cycle_t delta)
354 {
355         if (tracing_thresh) {
356                 if (delta < tracing_thresh)
357                         return 0;
358         } else {
359                 if (delta <= tracing_max_latency)
360                         return 0;
361         }
362         return 1;
363 }
364
365 static void
366 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu)
367 {
368         if (task != wakeup_task)
369                 return;
370
371         wakeup_current_cpu = cpu;
372 }
373
374 static void notrace
375 probe_wakeup_sched_switch(void *ignore,
376                           struct task_struct *prev, struct task_struct *next)
377 {
378         struct trace_array_cpu *data;
379         cycle_t T0, T1, delta;
380         unsigned long flags;
381         long disabled;
382         int cpu;
383         int pc;
384
385         tracing_record_cmdline(prev);
386
387         if (unlikely(!tracer_enabled))
388                 return;
389
390         /*
391          * When we start a new trace, we set wakeup_task to NULL
392          * and then set tracer_enabled = 1. We want to make sure
393          * that another CPU does not see the tracer_enabled = 1
394          * and the wakeup_task with an older task, that might
395          * actually be the same as next.
396          */
397         smp_rmb();
398
399         if (next != wakeup_task)
400                 return;
401
402         pc = preempt_count();
403
404         /* disable local data, not wakeup_cpu data */
405         cpu = raw_smp_processor_id();
406         disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
407         if (likely(disabled != 1))
408                 goto out;
409
410         local_irq_save(flags);
411         arch_spin_lock(&wakeup_lock);
412
413         /* We could race with grabbing wakeup_lock */
414         if (unlikely(!tracer_enabled || next != wakeup_task))
415                 goto out_unlock;
416
417         /* The task we are waiting for is waking up */
418         data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu);
419
420         __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc);
421         tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc);
422
423         T0 = data->preempt_timestamp;
424         T1 = ftrace_now(cpu);
425         delta = T1-T0;
426
427         if (!report_latency(delta))
428                 goto out_unlock;
429
430         if (likely(!is_tracing_stopped())) {
431                 tracing_max_latency = delta;
432                 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu);
433         }
434
435 out_unlock:
436         __wakeup_reset(wakeup_trace);
437         arch_spin_unlock(&wakeup_lock);
438         local_irq_restore(flags);
439 out:
440         atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
441 }
442
443 static void __wakeup_reset(struct trace_array *tr)
444 {
445         wakeup_cpu = -1;
446         wakeup_prio = -1;
447         tracing_dl = 0;
448
449         if (wakeup_task)
450                 put_task_struct(wakeup_task);
451
452         wakeup_task = NULL;
453 }
454
455 static void wakeup_reset(struct trace_array *tr)
456 {
457         unsigned long flags;
458
459         tracing_reset_online_cpus(&tr->trace_buffer);
460
461         local_irq_save(flags);
462         arch_spin_lock(&wakeup_lock);
463         __wakeup_reset(tr);
464         arch_spin_unlock(&wakeup_lock);
465         local_irq_restore(flags);
466 }
467
468 static void
469 probe_wakeup(void *ignore, struct task_struct *p, int success)
470 {
471         struct trace_array_cpu *data;
472         int cpu = smp_processor_id();
473         unsigned long flags;
474         long disabled;
475         int pc;
476
477         if (likely(!tracer_enabled))
478                 return;
479
480         tracing_record_cmdline(p);
481         tracing_record_cmdline(current);
482
483         /*
484          * Semantic is like this:
485          *  - wakeup tracer handles all tasks in the system, independently
486          *    from their scheduling class;
487          *  - wakeup_rt tracer handles tasks belonging to sched_dl and
488          *    sched_rt class;
489          *  - wakeup_dl handles tasks belonging to sched_dl class only.
490          */
491         if (tracing_dl || (wakeup_dl && !dl_task(p)) ||
492             (wakeup_rt && !dl_task(p) && !rt_task(p)) ||
493             (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio)))
494                 return;
495
496         pc = preempt_count();
497         disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
498         if (unlikely(disabled != 1))
499                 goto out;
500
501         /* interrupts should be off from try_to_wake_up */
502         arch_spin_lock(&wakeup_lock);
503
504         /* check for races. */
505         if (!tracer_enabled || tracing_dl ||
506             (!dl_task(p) && p->prio >= wakeup_prio))
507                 goto out_locked;
508
509         /* reset the trace */
510         __wakeup_reset(wakeup_trace);
511
512         wakeup_cpu = task_cpu(p);
513         wakeup_current_cpu = wakeup_cpu;
514         wakeup_prio = p->prio;
515
516         /*
517          * Once you start tracing a -deadline task, don't bother tracing
518          * another task until the first one wakes up.
519          */
520         if (dl_task(p))
521                 tracing_dl = 1;
522         else
523                 tracing_dl = 0;
524
525         wakeup_task = p;
526         get_task_struct(wakeup_task);
527
528         local_save_flags(flags);
529
530         data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu);
531         data->preempt_timestamp = ftrace_now(cpu);
532         tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc);
533
534         /*
535          * We must be careful in using CALLER_ADDR2. But since wake_up
536          * is not called by an assembly function  (where as schedule is)
537          * it should be safe to use it here.
538          */
539         __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc);
540
541 out_locked:
542         arch_spin_unlock(&wakeup_lock);
543 out:
544         atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled);
545 }
546
547 static void start_wakeup_tracer(struct trace_array *tr)
548 {
549         int ret;
550
551         ret = register_trace_sched_wakeup(probe_wakeup, NULL);
552         if (ret) {
553                 pr_info("wakeup trace: Couldn't activate tracepoint"
554                         " probe to kernel_sched_wakeup\n");
555                 return;
556         }
557
558         ret = register_trace_sched_wakeup_new(probe_wakeup, NULL);
559         if (ret) {
560                 pr_info("wakeup trace: Couldn't activate tracepoint"
561                         " probe to kernel_sched_wakeup_new\n");
562                 goto fail_deprobe;
563         }
564
565         ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL);
566         if (ret) {
567                 pr_info("sched trace: Couldn't activate tracepoint"
568                         " probe to kernel_sched_switch\n");
569                 goto fail_deprobe_wake_new;
570         }
571
572         ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
573         if (ret) {
574                 pr_info("wakeup trace: Couldn't activate tracepoint"
575                         " probe to kernel_sched_migrate_task\n");
576                 return;
577         }
578
579         wakeup_reset(tr);
580
581         /*
582          * Don't let the tracer_enabled = 1 show up before
583          * the wakeup_task is reset. This may be overkill since
584          * wakeup_reset does a spin_unlock after setting the
585          * wakeup_task to NULL, but I want to be safe.
586          * This is a slow path anyway.
587          */
588         smp_wmb();
589
590         if (start_func_tracer(is_graph()))
591                 printk(KERN_ERR "failed to start wakeup tracer\n");
592
593         return;
594 fail_deprobe_wake_new:
595         unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
596 fail_deprobe:
597         unregister_trace_sched_wakeup(probe_wakeup, NULL);
598 }
599
600 static void stop_wakeup_tracer(struct trace_array *tr)
601 {
602         tracer_enabled = 0;
603         stop_func_tracer(is_graph());
604         unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL);
605         unregister_trace_sched_wakeup_new(probe_wakeup, NULL);
606         unregister_trace_sched_wakeup(probe_wakeup, NULL);
607         unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL);
608 }
609
610 static int __wakeup_tracer_init(struct trace_array *tr)
611 {
612         save_flags = trace_flags;
613
614         /* non overwrite screws up the latency tracers */
615         set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
616         set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
617
618         tracing_max_latency = 0;
619         wakeup_trace = tr;
620         start_wakeup_tracer(tr);
621         return 0;
622 }
623
624 static int wakeup_tracer_init(struct trace_array *tr)
625 {
626         wakeup_dl = 0;
627         wakeup_rt = 0;
628         return __wakeup_tracer_init(tr);
629 }
630
631 static int wakeup_rt_tracer_init(struct trace_array *tr)
632 {
633         wakeup_dl = 0;
634         wakeup_rt = 1;
635         return __wakeup_tracer_init(tr);
636 }
637
638 static int wakeup_dl_tracer_init(struct trace_array *tr)
639 {
640         wakeup_dl = 1;
641         wakeup_rt = 0;
642         return __wakeup_tracer_init(tr);
643 }
644
645 static void wakeup_tracer_reset(struct trace_array *tr)
646 {
647         int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
648         int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
649
650         stop_wakeup_tracer(tr);
651         /* make sure we put back any tasks we are tracing */
652         wakeup_reset(tr);
653
654         set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
655         set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
656 }
657
658 static void wakeup_tracer_start(struct trace_array *tr)
659 {
660         wakeup_reset(tr);
661         tracer_enabled = 1;
662 }
663
664 static void wakeup_tracer_stop(struct trace_array *tr)
665 {
666         tracer_enabled = 0;
667 }
668
669 static struct tracer wakeup_tracer __read_mostly =
670 {
671         .name           = "wakeup",
672         .init           = wakeup_tracer_init,
673         .reset          = wakeup_tracer_reset,
674         .start          = wakeup_tracer_start,
675         .stop           = wakeup_tracer_stop,
676         .print_max      = true,
677         .print_header   = wakeup_print_header,
678         .print_line     = wakeup_print_line,
679         .flags          = &tracer_flags,
680         .set_flag       = wakeup_set_flag,
681         .flag_changed   = wakeup_flag_changed,
682 #ifdef CONFIG_FTRACE_SELFTEST
683         .selftest    = trace_selftest_startup_wakeup,
684 #endif
685         .open           = wakeup_trace_open,
686         .close          = wakeup_trace_close,
687         .use_max_tr     = true,
688 };
689
690 static struct tracer wakeup_rt_tracer __read_mostly =
691 {
692         .name           = "wakeup_rt",
693         .init           = wakeup_rt_tracer_init,
694         .reset          = wakeup_tracer_reset,
695         .start          = wakeup_tracer_start,
696         .stop           = wakeup_tracer_stop,
697         .wait_pipe      = poll_wait_pipe,
698         .print_max      = true,
699         .print_header   = wakeup_print_header,
700         .print_line     = wakeup_print_line,
701         .flags          = &tracer_flags,
702         .set_flag       = wakeup_set_flag,
703         .flag_changed   = wakeup_flag_changed,
704 #ifdef CONFIG_FTRACE_SELFTEST
705         .selftest    = trace_selftest_startup_wakeup,
706 #endif
707         .open           = wakeup_trace_open,
708         .close          = wakeup_trace_close,
709         .use_max_tr     = true,
710 };
711
712 static struct tracer wakeup_dl_tracer __read_mostly =
713 {
714         .name           = "wakeup_dl",
715         .init           = wakeup_dl_tracer_init,
716         .reset          = wakeup_tracer_reset,
717         .start          = wakeup_tracer_start,
718         .stop           = wakeup_tracer_stop,
719         .wait_pipe      = poll_wait_pipe,
720         .print_max      = true,
721         .print_header   = wakeup_print_header,
722         .print_line     = wakeup_print_line,
723         .flags          = &tracer_flags,
724         .set_flag       = wakeup_set_flag,
725         .flag_changed   = wakeup_flag_changed,
726 #ifdef CONFIG_FTRACE_SELFTEST
727         .selftest    = trace_selftest_startup_wakeup,
728 #endif
729         .open           = wakeup_trace_open,
730         .close          = wakeup_trace_close,
731         .use_max_tr     = true,
732 };
733
734 __init static int init_wakeup_tracer(void)
735 {
736         int ret;
737
738         ret = register_tracer(&wakeup_tracer);
739         if (ret)
740                 return ret;
741
742         ret = register_tracer(&wakeup_rt_tracer);
743         if (ret)
744                 return ret;
745
746         ret = register_tracer(&wakeup_dl_tracer);
747         if (ret)
748                 return ret;
749
750         return 0;
751 }
752 core_initcall(init_wakeup_tracer);