Loading...
1/*
2 * trace irqs off critical timings
3 *
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6 *
7 * From code in the latency_tracer, that is:
8 *
9 * Copyright (C) 2004-2006 Ingo Molnar
10 * Copyright (C) 2004 Nadia Yvette Chambers
11 */
12#include <linux/kallsyms.h>
13#include <linux/uaccess.h>
14#include <linux/module.h>
15#include <linux/ftrace.h>
16
17#include "trace.h"
18
19static struct trace_array *irqsoff_trace __read_mostly;
20static int tracer_enabled __read_mostly;
21
22static DEFINE_PER_CPU(int, tracing_cpu);
23
24static DEFINE_RAW_SPINLOCK(max_trace_lock);
25
26enum {
27 TRACER_IRQS_OFF = (1 << 1),
28 TRACER_PREEMPT_OFF = (1 << 2),
29};
30
31static int trace_type __read_mostly;
32
33static int save_flags;
34
35static void stop_irqsoff_tracer(struct trace_array *tr, int graph);
36static int start_irqsoff_tracer(struct trace_array *tr, int graph);
37
38#ifdef CONFIG_PREEMPT_TRACER
39static inline int
40preempt_trace(void)
41{
42 return ((trace_type & TRACER_PREEMPT_OFF) && preempt_count());
43}
44#else
45# define preempt_trace() (0)
46#endif
47
48#ifdef CONFIG_IRQSOFF_TRACER
49static inline int
50irq_trace(void)
51{
52 return ((trace_type & TRACER_IRQS_OFF) &&
53 irqs_disabled());
54}
55#else
56# define irq_trace() (0)
57#endif
58
59#ifdef CONFIG_FUNCTION_GRAPH_TRACER
60static int irqsoff_display_graph(struct trace_array *tr, int set);
61# define is_graph(tr) ((tr)->trace_flags & TRACE_ITER_DISPLAY_GRAPH)
62#else
63static inline int irqsoff_display_graph(struct trace_array *tr, int set)
64{
65 return -EINVAL;
66}
67# define is_graph(tr) false
68#endif
69
70/*
71 * Sequence count - we record it when starting a measurement and
72 * skip the latency if the sequence has changed - some other section
73 * did a maximum and could disturb our measurement with serial console
74 * printouts, etc. Truly coinciding maximum latencies should be rare
75 * and what happens together happens separately as well, so this doesn't
76 * decrease the validity of the maximum found:
77 */
78static __cacheline_aligned_in_smp unsigned long max_sequence;
79
80#ifdef CONFIG_FUNCTION_TRACER
81/*
82 * Prologue for the preempt and irqs off function tracers.
83 *
84 * Returns 1 if it is OK to continue, and data->disabled is
85 * incremented.
86 * 0 if the trace is to be ignored, and data->disabled
87 * is kept the same.
88 *
89 * Note, this function is also used outside this ifdef but
90 * inside the #ifdef of the function graph tracer below.
91 * This is OK, since the function graph tracer is
92 * dependent on the function tracer.
93 */
94static int func_prolog_dec(struct trace_array *tr,
95 struct trace_array_cpu **data,
96 unsigned long *flags)
97{
98 long disabled;
99 int cpu;
100
101 /*
102 * Does not matter if we preempt. We test the flags
103 * afterward, to see if irqs are disabled or not.
104 * If we preempt and get a false positive, the flags
105 * test will fail.
106 */
107 cpu = raw_smp_processor_id();
108 if (likely(!per_cpu(tracing_cpu, cpu)))
109 return 0;
110
111 local_save_flags(*flags);
112 /*
113 * Slight chance to get a false positive on tracing_cpu,
114 * although I'm starting to think there isn't a chance.
115 * Leave this for now just to be paranoid.
116 */
117 if (!irqs_disabled_flags(*flags) && !preempt_count())
118 return 0;
119
120 *data = per_cpu_ptr(tr->trace_buffer.data, cpu);
121 disabled = atomic_inc_return(&(*data)->disabled);
122
123 if (likely(disabled == 1))
124 return 1;
125
126 atomic_dec(&(*data)->disabled);
127
128 return 0;
129}
130
131/*
132 * irqsoff uses its own tracer function to keep the overhead down:
133 */
134static void
135irqsoff_tracer_call(unsigned long ip, unsigned long parent_ip,
136 struct ftrace_ops *op, struct pt_regs *pt_regs)
137{
138 struct trace_array *tr = irqsoff_trace;
139 struct trace_array_cpu *data;
140 unsigned long flags;
141
142 if (!func_prolog_dec(tr, &data, &flags))
143 return;
144
145 trace_function(tr, ip, parent_ip, flags, preempt_count());
146
147 atomic_dec(&data->disabled);
148}
149#endif /* CONFIG_FUNCTION_TRACER */
150
151#ifdef CONFIG_FUNCTION_GRAPH_TRACER
152static int irqsoff_display_graph(struct trace_array *tr, int set)
153{
154 int cpu;
155
156 if (!(is_graph(tr) ^ set))
157 return 0;
158
159 stop_irqsoff_tracer(irqsoff_trace, !set);
160
161 for_each_possible_cpu(cpu)
162 per_cpu(tracing_cpu, cpu) = 0;
163
164 tr->max_latency = 0;
165 tracing_reset_online_cpus(&irqsoff_trace->trace_buffer);
166
167 return start_irqsoff_tracer(irqsoff_trace, set);
168}
169
170static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
171{
172 struct trace_array *tr = irqsoff_trace;
173 struct trace_array_cpu *data;
174 unsigned long flags;
175 int ret;
176 int pc;
177
178 if (ftrace_graph_ignore_func(trace))
179 return 0;
180 /*
181 * Do not trace a function if it's filtered by set_graph_notrace.
182 * Make the index of ret stack negative to indicate that it should
183 * ignore further functions. But it needs its own ret stack entry
184 * to recover the original index in order to continue tracing after
185 * returning from the function.
186 */
187 if (ftrace_graph_notrace_addr(trace->func))
188 return 1;
189
190 if (!func_prolog_dec(tr, &data, &flags))
191 return 0;
192
193 pc = preempt_count();
194 ret = __trace_graph_entry(tr, trace, flags, pc);
195 atomic_dec(&data->disabled);
196
197 return ret;
198}
199
200static void irqsoff_graph_return(struct ftrace_graph_ret *trace)
201{
202 struct trace_array *tr = irqsoff_trace;
203 struct trace_array_cpu *data;
204 unsigned long flags;
205 int pc;
206
207 if (!func_prolog_dec(tr, &data, &flags))
208 return;
209
210 pc = preempt_count();
211 __trace_graph_return(tr, trace, flags, pc);
212 atomic_dec(&data->disabled);
213}
214
215static void irqsoff_trace_open(struct trace_iterator *iter)
216{
217 if (is_graph(iter->tr))
218 graph_trace_open(iter);
219
220}
221
222static void irqsoff_trace_close(struct trace_iterator *iter)
223{
224 if (iter->private)
225 graph_trace_close(iter);
226}
227
228#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_CPU | \
229 TRACE_GRAPH_PRINT_PROC | \
230 TRACE_GRAPH_PRINT_ABS_TIME | \
231 TRACE_GRAPH_PRINT_DURATION)
232
233static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
234{
235 /*
236 * In graph mode call the graph tracer output function,
237 * otherwise go with the TRACE_FN event handler
238 */
239 if (is_graph(iter->tr))
240 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
241
242 return TRACE_TYPE_UNHANDLED;
243}
244
245static void irqsoff_print_header(struct seq_file *s)
246{
247 struct trace_array *tr = irqsoff_trace;
248
249 if (is_graph(tr))
250 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
251 else
252 trace_default_header(s);
253}
254
255static void
256__trace_function(struct trace_array *tr,
257 unsigned long ip, unsigned long parent_ip,
258 unsigned long flags, int pc)
259{
260 if (is_graph(tr))
261 trace_graph_function(tr, ip, parent_ip, flags, pc);
262 else
263 trace_function(tr, ip, parent_ip, flags, pc);
264}
265
266#else
267#define __trace_function trace_function
268
269#ifdef CONFIG_FUNCTION_TRACER
270static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
271{
272 return -1;
273}
274#endif
275
276static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
277{
278 return TRACE_TYPE_UNHANDLED;
279}
280
281static void irqsoff_trace_open(struct trace_iterator *iter) { }
282static void irqsoff_trace_close(struct trace_iterator *iter) { }
283
284#ifdef CONFIG_FUNCTION_TRACER
285static void irqsoff_graph_return(struct ftrace_graph_ret *trace) { }
286static void irqsoff_print_header(struct seq_file *s)
287{
288 trace_default_header(s);
289}
290#else
291static void irqsoff_print_header(struct seq_file *s)
292{
293 trace_latency_header(s);
294}
295#endif /* CONFIG_FUNCTION_TRACER */
296#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
297
298/*
299 * Should this new latency be reported/recorded?
300 */
301static bool report_latency(struct trace_array *tr, u64 delta)
302{
303 if (tracing_thresh) {
304 if (delta < tracing_thresh)
305 return false;
306 } else {
307 if (delta <= tr->max_latency)
308 return false;
309 }
310 return true;
311}
312
313static void
314check_critical_timing(struct trace_array *tr,
315 struct trace_array_cpu *data,
316 unsigned long parent_ip,
317 int cpu)
318{
319 u64 T0, T1, delta;
320 unsigned long flags;
321 int pc;
322
323 T0 = data->preempt_timestamp;
324 T1 = ftrace_now(cpu);
325 delta = T1-T0;
326
327 local_save_flags(flags);
328
329 pc = preempt_count();
330
331 if (!report_latency(tr, delta))
332 goto out;
333
334 raw_spin_lock_irqsave(&max_trace_lock, flags);
335
336 /* check if we are still the max latency */
337 if (!report_latency(tr, delta))
338 goto out_unlock;
339
340 __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
341 /* Skip 5 functions to get to the irq/preempt enable function */
342 __trace_stack(tr, flags, 5, pc);
343
344 if (data->critical_sequence != max_sequence)
345 goto out_unlock;
346
347 data->critical_end = parent_ip;
348
349 if (likely(!is_tracing_stopped())) {
350 tr->max_latency = delta;
351 update_max_tr_single(tr, current, cpu);
352 }
353
354 max_sequence++;
355
356out_unlock:
357 raw_spin_unlock_irqrestore(&max_trace_lock, flags);
358
359out:
360 data->critical_sequence = max_sequence;
361 data->preempt_timestamp = ftrace_now(cpu);
362 __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
363}
364
365static inline void
366start_critical_timing(unsigned long ip, unsigned long parent_ip)
367{
368 int cpu;
369 struct trace_array *tr = irqsoff_trace;
370 struct trace_array_cpu *data;
371 unsigned long flags;
372
373 if (!tracer_enabled || !tracing_is_enabled())
374 return;
375
376 cpu = raw_smp_processor_id();
377
378 if (per_cpu(tracing_cpu, cpu))
379 return;
380
381 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
382
383 if (unlikely(!data) || atomic_read(&data->disabled))
384 return;
385
386 atomic_inc(&data->disabled);
387
388 data->critical_sequence = max_sequence;
389 data->preempt_timestamp = ftrace_now(cpu);
390 data->critical_start = parent_ip ? : ip;
391
392 local_save_flags(flags);
393
394 __trace_function(tr, ip, parent_ip, flags, preempt_count());
395
396 per_cpu(tracing_cpu, cpu) = 1;
397
398 atomic_dec(&data->disabled);
399}
400
401static inline void
402stop_critical_timing(unsigned long ip, unsigned long parent_ip)
403{
404 int cpu;
405 struct trace_array *tr = irqsoff_trace;
406 struct trace_array_cpu *data;
407 unsigned long flags;
408
409 cpu = raw_smp_processor_id();
410 /* Always clear the tracing cpu on stopping the trace */
411 if (unlikely(per_cpu(tracing_cpu, cpu)))
412 per_cpu(tracing_cpu, cpu) = 0;
413 else
414 return;
415
416 if (!tracer_enabled || !tracing_is_enabled())
417 return;
418
419 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
420
421 if (unlikely(!data) ||
422 !data->critical_start || atomic_read(&data->disabled))
423 return;
424
425 atomic_inc(&data->disabled);
426
427 local_save_flags(flags);
428 __trace_function(tr, ip, parent_ip, flags, preempt_count());
429 check_critical_timing(tr, data, parent_ip ? : ip, cpu);
430 data->critical_start = 0;
431 atomic_dec(&data->disabled);
432}
433
434/* start and stop critical timings used to for stoppage (in idle) */
435void start_critical_timings(void)
436{
437 if (preempt_trace() || irq_trace())
438 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
439}
440EXPORT_SYMBOL_GPL(start_critical_timings);
441
442void stop_critical_timings(void)
443{
444 if (preempt_trace() || irq_trace())
445 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
446}
447EXPORT_SYMBOL_GPL(stop_critical_timings);
448
449#ifdef CONFIG_IRQSOFF_TRACER
450#ifdef CONFIG_PROVE_LOCKING
451void time_hardirqs_on(unsigned long a0, unsigned long a1)
452{
453 if (!preempt_trace() && irq_trace())
454 stop_critical_timing(a0, a1);
455}
456
457void time_hardirqs_off(unsigned long a0, unsigned long a1)
458{
459 if (!preempt_trace() && irq_trace())
460 start_critical_timing(a0, a1);
461}
462
463#else /* !CONFIG_PROVE_LOCKING */
464
465/*
466 * Stubs:
467 */
468
469void trace_softirqs_on(unsigned long ip)
470{
471}
472
473void trace_softirqs_off(unsigned long ip)
474{
475}
476
477inline void print_irqtrace_events(struct task_struct *curr)
478{
479}
480
481/*
482 * We are only interested in hardirq on/off events:
483 */
484void trace_hardirqs_on(void)
485{
486 if (!preempt_trace() && irq_trace())
487 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
488}
489EXPORT_SYMBOL(trace_hardirqs_on);
490
491void trace_hardirqs_off(void)
492{
493 if (!preempt_trace() && irq_trace())
494 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
495}
496EXPORT_SYMBOL(trace_hardirqs_off);
497
498__visible void trace_hardirqs_on_caller(unsigned long caller_addr)
499{
500 if (!preempt_trace() && irq_trace())
501 stop_critical_timing(CALLER_ADDR0, caller_addr);
502}
503EXPORT_SYMBOL(trace_hardirqs_on_caller);
504
505__visible void trace_hardirqs_off_caller(unsigned long caller_addr)
506{
507 if (!preempt_trace() && irq_trace())
508 start_critical_timing(CALLER_ADDR0, caller_addr);
509}
510EXPORT_SYMBOL(trace_hardirqs_off_caller);
511
512#endif /* CONFIG_PROVE_LOCKING */
513#endif /* CONFIG_IRQSOFF_TRACER */
514
515#ifdef CONFIG_PREEMPT_TRACER
516void trace_preempt_on(unsigned long a0, unsigned long a1)
517{
518 if (preempt_trace() && !irq_trace())
519 stop_critical_timing(a0, a1);
520}
521
522void trace_preempt_off(unsigned long a0, unsigned long a1)
523{
524 if (preempt_trace() && !irq_trace())
525 start_critical_timing(a0, a1);
526}
527#endif /* CONFIG_PREEMPT_TRACER */
528
529#ifdef CONFIG_FUNCTION_TRACER
530static bool function_enabled;
531
532static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
533{
534 int ret;
535
536 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
537 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
538 return 0;
539
540 if (graph)
541 ret = register_ftrace_graph(&irqsoff_graph_return,
542 &irqsoff_graph_entry);
543 else
544 ret = register_ftrace_function(tr->ops);
545
546 if (!ret)
547 function_enabled = true;
548
549 return ret;
550}
551
552static void unregister_irqsoff_function(struct trace_array *tr, int graph)
553{
554 if (!function_enabled)
555 return;
556
557 if (graph)
558 unregister_ftrace_graph();
559 else
560 unregister_ftrace_function(tr->ops);
561
562 function_enabled = false;
563}
564
565static int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
566{
567 if (!(mask & TRACE_ITER_FUNCTION))
568 return 0;
569
570 if (set)
571 register_irqsoff_function(tr, is_graph(tr), 1);
572 else
573 unregister_irqsoff_function(tr, is_graph(tr));
574 return 1;
575}
576#else
577static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
578{
579 return 0;
580}
581static void unregister_irqsoff_function(struct trace_array *tr, int graph) { }
582static inline int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
583{
584 return 0;
585}
586#endif /* CONFIG_FUNCTION_TRACER */
587
588static int irqsoff_flag_changed(struct trace_array *tr, u32 mask, int set)
589{
590 struct tracer *tracer = tr->current_trace;
591
592 if (irqsoff_function_set(tr, mask, set))
593 return 0;
594
595#ifdef CONFIG_FUNCTION_GRAPH_TRACER
596 if (mask & TRACE_ITER_DISPLAY_GRAPH)
597 return irqsoff_display_graph(tr, set);
598#endif
599
600 return trace_keep_overwrite(tracer, mask, set);
601}
602
603static int start_irqsoff_tracer(struct trace_array *tr, int graph)
604{
605 int ret;
606
607 ret = register_irqsoff_function(tr, graph, 0);
608
609 if (!ret && tracing_is_enabled())
610 tracer_enabled = 1;
611 else
612 tracer_enabled = 0;
613
614 return ret;
615}
616
617static void stop_irqsoff_tracer(struct trace_array *tr, int graph)
618{
619 tracer_enabled = 0;
620
621 unregister_irqsoff_function(tr, graph);
622}
623
624static bool irqsoff_busy;
625
626static int __irqsoff_tracer_init(struct trace_array *tr)
627{
628 if (irqsoff_busy)
629 return -EBUSY;
630
631 save_flags = tr->trace_flags;
632
633 /* non overwrite screws up the latency tracers */
634 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
635 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
636
637 tr->max_latency = 0;
638 irqsoff_trace = tr;
639 /* make sure that the tracer is visible */
640 smp_wmb();
641
642 ftrace_init_array_ops(tr, irqsoff_tracer_call);
643
644 /* Only toplevel instance supports graph tracing */
645 if (start_irqsoff_tracer(tr, (tr->flags & TRACE_ARRAY_FL_GLOBAL &&
646 is_graph(tr))))
647 printk(KERN_ERR "failed to start irqsoff tracer\n");
648
649 irqsoff_busy = true;
650 return 0;
651}
652
653static void irqsoff_tracer_reset(struct trace_array *tr)
654{
655 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
656 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
657
658 stop_irqsoff_tracer(tr, is_graph(tr));
659
660 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
661 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
662 ftrace_reset_array_ops(tr);
663
664 irqsoff_busy = false;
665}
666
667static void irqsoff_tracer_start(struct trace_array *tr)
668{
669 tracer_enabled = 1;
670}
671
672static void irqsoff_tracer_stop(struct trace_array *tr)
673{
674 tracer_enabled = 0;
675}
676
677#ifdef CONFIG_IRQSOFF_TRACER
678static int irqsoff_tracer_init(struct trace_array *tr)
679{
680 trace_type = TRACER_IRQS_OFF;
681
682 return __irqsoff_tracer_init(tr);
683}
684static struct tracer irqsoff_tracer __read_mostly =
685{
686 .name = "irqsoff",
687 .init = irqsoff_tracer_init,
688 .reset = irqsoff_tracer_reset,
689 .start = irqsoff_tracer_start,
690 .stop = irqsoff_tracer_stop,
691 .print_max = true,
692 .print_header = irqsoff_print_header,
693 .print_line = irqsoff_print_line,
694 .flag_changed = irqsoff_flag_changed,
695#ifdef CONFIG_FTRACE_SELFTEST
696 .selftest = trace_selftest_startup_irqsoff,
697#endif
698 .open = irqsoff_trace_open,
699 .close = irqsoff_trace_close,
700 .allow_instances = true,
701 .use_max_tr = true,
702};
703# define register_irqsoff(trace) register_tracer(&trace)
704#else
705# define register_irqsoff(trace) do { } while (0)
706#endif
707
708#ifdef CONFIG_PREEMPT_TRACER
709static int preemptoff_tracer_init(struct trace_array *tr)
710{
711 trace_type = TRACER_PREEMPT_OFF;
712
713 return __irqsoff_tracer_init(tr);
714}
715
716static struct tracer preemptoff_tracer __read_mostly =
717{
718 .name = "preemptoff",
719 .init = preemptoff_tracer_init,
720 .reset = irqsoff_tracer_reset,
721 .start = irqsoff_tracer_start,
722 .stop = irqsoff_tracer_stop,
723 .print_max = true,
724 .print_header = irqsoff_print_header,
725 .print_line = irqsoff_print_line,
726 .flag_changed = irqsoff_flag_changed,
727#ifdef CONFIG_FTRACE_SELFTEST
728 .selftest = trace_selftest_startup_preemptoff,
729#endif
730 .open = irqsoff_trace_open,
731 .close = irqsoff_trace_close,
732 .allow_instances = true,
733 .use_max_tr = true,
734};
735# define register_preemptoff(trace) register_tracer(&trace)
736#else
737# define register_preemptoff(trace) do { } while (0)
738#endif
739
740#if defined(CONFIG_IRQSOFF_TRACER) && \
741 defined(CONFIG_PREEMPT_TRACER)
742
743static int preemptirqsoff_tracer_init(struct trace_array *tr)
744{
745 trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
746
747 return __irqsoff_tracer_init(tr);
748}
749
750static struct tracer preemptirqsoff_tracer __read_mostly =
751{
752 .name = "preemptirqsoff",
753 .init = preemptirqsoff_tracer_init,
754 .reset = irqsoff_tracer_reset,
755 .start = irqsoff_tracer_start,
756 .stop = irqsoff_tracer_stop,
757 .print_max = true,
758 .print_header = irqsoff_print_header,
759 .print_line = irqsoff_print_line,
760 .flag_changed = irqsoff_flag_changed,
761#ifdef CONFIG_FTRACE_SELFTEST
762 .selftest = trace_selftest_startup_preemptirqsoff,
763#endif
764 .open = irqsoff_trace_open,
765 .close = irqsoff_trace_close,
766 .allow_instances = true,
767 .use_max_tr = true,
768};
769
770# define register_preemptirqsoff(trace) register_tracer(&trace)
771#else
772# define register_preemptirqsoff(trace) do { } while (0)
773#endif
774
775__init static int init_irqsoff_tracer(void)
776{
777 register_irqsoff(irqsoff_tracer);
778 register_preemptoff(preemptoff_tracer);
779 register_preemptirqsoff(preemptirqsoff_tracer);
780
781 return 0;
782}
783core_initcall(init_irqsoff_tracer);
1/*
2 * trace irqs off critical timings
3 *
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
6 *
7 * From code in the latency_tracer, that is:
8 *
9 * Copyright (C) 2004-2006 Ingo Molnar
10 * Copyright (C) 2004 Nadia Yvette Chambers
11 */
12#include <linux/kallsyms.h>
13#include <linux/uaccess.h>
14#include <linux/module.h>
15#include <linux/ftrace.h>
16
17#include "trace.h"
18
19static struct trace_array *irqsoff_trace __read_mostly;
20static int tracer_enabled __read_mostly;
21
22static DEFINE_PER_CPU(int, tracing_cpu);
23
24static DEFINE_RAW_SPINLOCK(max_trace_lock);
25
26enum {
27 TRACER_IRQS_OFF = (1 << 1),
28 TRACER_PREEMPT_OFF = (1 << 2),
29};
30
31static int trace_type __read_mostly;
32
33static int save_flags;
34
35static void stop_irqsoff_tracer(struct trace_array *tr, int graph);
36static int start_irqsoff_tracer(struct trace_array *tr, int graph);
37
38#ifdef CONFIG_PREEMPT_TRACER
39static inline int
40preempt_trace(void)
41{
42 return ((trace_type & TRACER_PREEMPT_OFF) && preempt_count());
43}
44#else
45# define preempt_trace() (0)
46#endif
47
48#ifdef CONFIG_IRQSOFF_TRACER
49static inline int
50irq_trace(void)
51{
52 return ((trace_type & TRACER_IRQS_OFF) &&
53 irqs_disabled());
54}
55#else
56# define irq_trace() (0)
57#endif
58
59#ifdef CONFIG_FUNCTION_GRAPH_TRACER
60static int irqsoff_display_graph(struct trace_array *tr, int set);
61# define is_graph(tr) ((tr)->trace_flags & TRACE_ITER_DISPLAY_GRAPH)
62#else
63static inline int irqsoff_display_graph(struct trace_array *tr, int set)
64{
65 return -EINVAL;
66}
67# define is_graph(tr) false
68#endif
69
70/*
71 * Sequence count - we record it when starting a measurement and
72 * skip the latency if the sequence has changed - some other section
73 * did a maximum and could disturb our measurement with serial console
74 * printouts, etc. Truly coinciding maximum latencies should be rare
75 * and what happens together happens separately as well, so this doesn't
76 * decrease the validity of the maximum found:
77 */
78static __cacheline_aligned_in_smp unsigned long max_sequence;
79
80#ifdef CONFIG_FUNCTION_TRACER
81/*
82 * Prologue for the preempt and irqs off function tracers.
83 *
84 * Returns 1 if it is OK to continue, and data->disabled is
85 * incremented.
86 * 0 if the trace is to be ignored, and data->disabled
87 * is kept the same.
88 *
89 * Note, this function is also used outside this ifdef but
90 * inside the #ifdef of the function graph tracer below.
91 * This is OK, since the function graph tracer is
92 * dependent on the function tracer.
93 */
94static int func_prolog_dec(struct trace_array *tr,
95 struct trace_array_cpu **data,
96 unsigned long *flags)
97{
98 long disabled;
99 int cpu;
100
101 /*
102 * Does not matter if we preempt. We test the flags
103 * afterward, to see if irqs are disabled or not.
104 * If we preempt and get a false positive, the flags
105 * test will fail.
106 */
107 cpu = raw_smp_processor_id();
108 if (likely(!per_cpu(tracing_cpu, cpu)))
109 return 0;
110
111 local_save_flags(*flags);
112 /*
113 * Slight chance to get a false positive on tracing_cpu,
114 * although I'm starting to think there isn't a chance.
115 * Leave this for now just to be paranoid.
116 */
117 if (!irqs_disabled_flags(*flags) && !preempt_count())
118 return 0;
119
120 *data = per_cpu_ptr(tr->trace_buffer.data, cpu);
121 disabled = atomic_inc_return(&(*data)->disabled);
122
123 if (likely(disabled == 1))
124 return 1;
125
126 atomic_dec(&(*data)->disabled);
127
128 return 0;
129}
130
131/*
132 * irqsoff uses its own tracer function to keep the overhead down:
133 */
134static void
135irqsoff_tracer_call(unsigned long ip, unsigned long parent_ip,
136 struct ftrace_ops *op, struct pt_regs *pt_regs)
137{
138 struct trace_array *tr = irqsoff_trace;
139 struct trace_array_cpu *data;
140 unsigned long flags;
141
142 if (!func_prolog_dec(tr, &data, &flags))
143 return;
144
145 trace_function(tr, ip, parent_ip, flags, preempt_count());
146
147 atomic_dec(&data->disabled);
148}
149#endif /* CONFIG_FUNCTION_TRACER */
150
151#ifdef CONFIG_FUNCTION_GRAPH_TRACER
152static int irqsoff_display_graph(struct trace_array *tr, int set)
153{
154 int cpu;
155
156 if (!(is_graph(tr) ^ set))
157 return 0;
158
159 stop_irqsoff_tracer(irqsoff_trace, !set);
160
161 for_each_possible_cpu(cpu)
162 per_cpu(tracing_cpu, cpu) = 0;
163
164 tr->max_latency = 0;
165 tracing_reset_online_cpus(&irqsoff_trace->trace_buffer);
166
167 return start_irqsoff_tracer(irqsoff_trace, set);
168}
169
170static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
171{
172 struct trace_array *tr = irqsoff_trace;
173 struct trace_array_cpu *data;
174 unsigned long flags;
175 int ret;
176 int pc;
177
178 if (!func_prolog_dec(tr, &data, &flags))
179 return 0;
180
181 pc = preempt_count();
182 ret = __trace_graph_entry(tr, trace, flags, pc);
183 atomic_dec(&data->disabled);
184
185 return ret;
186}
187
188static void irqsoff_graph_return(struct ftrace_graph_ret *trace)
189{
190 struct trace_array *tr = irqsoff_trace;
191 struct trace_array_cpu *data;
192 unsigned long flags;
193 int pc;
194
195 if (!func_prolog_dec(tr, &data, &flags))
196 return;
197
198 pc = preempt_count();
199 __trace_graph_return(tr, trace, flags, pc);
200 atomic_dec(&data->disabled);
201}
202
203static void irqsoff_trace_open(struct trace_iterator *iter)
204{
205 if (is_graph(iter->tr))
206 graph_trace_open(iter);
207
208}
209
210static void irqsoff_trace_close(struct trace_iterator *iter)
211{
212 if (iter->private)
213 graph_trace_close(iter);
214}
215
216#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_CPU | \
217 TRACE_GRAPH_PRINT_PROC | \
218 TRACE_GRAPH_PRINT_ABS_TIME | \
219 TRACE_GRAPH_PRINT_DURATION)
220
221static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
222{
223 /*
224 * In graph mode call the graph tracer output function,
225 * otherwise go with the TRACE_FN event handler
226 */
227 if (is_graph(iter->tr))
228 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
229
230 return TRACE_TYPE_UNHANDLED;
231}
232
233static void irqsoff_print_header(struct seq_file *s)
234{
235 struct trace_array *tr = irqsoff_trace;
236
237 if (is_graph(tr))
238 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
239 else
240 trace_default_header(s);
241}
242
243static void
244__trace_function(struct trace_array *tr,
245 unsigned long ip, unsigned long parent_ip,
246 unsigned long flags, int pc)
247{
248 if (is_graph(tr))
249 trace_graph_function(tr, ip, parent_ip, flags, pc);
250 else
251 trace_function(tr, ip, parent_ip, flags, pc);
252}
253
254#else
255#define __trace_function trace_function
256
257#ifdef CONFIG_FUNCTION_TRACER
258static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
259{
260 return -1;
261}
262#endif
263
264static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
265{
266 return TRACE_TYPE_UNHANDLED;
267}
268
269static void irqsoff_trace_open(struct trace_iterator *iter) { }
270static void irqsoff_trace_close(struct trace_iterator *iter) { }
271
272#ifdef CONFIG_FUNCTION_TRACER
273static void irqsoff_graph_return(struct ftrace_graph_ret *trace) { }
274static void irqsoff_print_header(struct seq_file *s)
275{
276 trace_default_header(s);
277}
278#else
279static void irqsoff_print_header(struct seq_file *s)
280{
281 trace_latency_header(s);
282}
283#endif /* CONFIG_FUNCTION_TRACER */
284#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
285
286/*
287 * Should this new latency be reported/recorded?
288 */
289static bool report_latency(struct trace_array *tr, cycle_t delta)
290{
291 if (tracing_thresh) {
292 if (delta < tracing_thresh)
293 return false;
294 } else {
295 if (delta <= tr->max_latency)
296 return false;
297 }
298 return true;
299}
300
301static void
302check_critical_timing(struct trace_array *tr,
303 struct trace_array_cpu *data,
304 unsigned long parent_ip,
305 int cpu)
306{
307 cycle_t T0, T1, delta;
308 unsigned long flags;
309 int pc;
310
311 T0 = data->preempt_timestamp;
312 T1 = ftrace_now(cpu);
313 delta = T1-T0;
314
315 local_save_flags(flags);
316
317 pc = preempt_count();
318
319 if (!report_latency(tr, delta))
320 goto out;
321
322 raw_spin_lock_irqsave(&max_trace_lock, flags);
323
324 /* check if we are still the max latency */
325 if (!report_latency(tr, delta))
326 goto out_unlock;
327
328 __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
329 /* Skip 5 functions to get to the irq/preempt enable function */
330 __trace_stack(tr, flags, 5, pc);
331
332 if (data->critical_sequence != max_sequence)
333 goto out_unlock;
334
335 data->critical_end = parent_ip;
336
337 if (likely(!is_tracing_stopped())) {
338 tr->max_latency = delta;
339 update_max_tr_single(tr, current, cpu);
340 }
341
342 max_sequence++;
343
344out_unlock:
345 raw_spin_unlock_irqrestore(&max_trace_lock, flags);
346
347out:
348 data->critical_sequence = max_sequence;
349 data->preempt_timestamp = ftrace_now(cpu);
350 __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
351}
352
353static inline void
354start_critical_timing(unsigned long ip, unsigned long parent_ip)
355{
356 int cpu;
357 struct trace_array *tr = irqsoff_trace;
358 struct trace_array_cpu *data;
359 unsigned long flags;
360
361 if (!tracer_enabled || !tracing_is_enabled())
362 return;
363
364 cpu = raw_smp_processor_id();
365
366 if (per_cpu(tracing_cpu, cpu))
367 return;
368
369 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
370
371 if (unlikely(!data) || atomic_read(&data->disabled))
372 return;
373
374 atomic_inc(&data->disabled);
375
376 data->critical_sequence = max_sequence;
377 data->preempt_timestamp = ftrace_now(cpu);
378 data->critical_start = parent_ip ? : ip;
379
380 local_save_flags(flags);
381
382 __trace_function(tr, ip, parent_ip, flags, preempt_count());
383
384 per_cpu(tracing_cpu, cpu) = 1;
385
386 atomic_dec(&data->disabled);
387}
388
389static inline void
390stop_critical_timing(unsigned long ip, unsigned long parent_ip)
391{
392 int cpu;
393 struct trace_array *tr = irqsoff_trace;
394 struct trace_array_cpu *data;
395 unsigned long flags;
396
397 cpu = raw_smp_processor_id();
398 /* Always clear the tracing cpu on stopping the trace */
399 if (unlikely(per_cpu(tracing_cpu, cpu)))
400 per_cpu(tracing_cpu, cpu) = 0;
401 else
402 return;
403
404 if (!tracer_enabled || !tracing_is_enabled())
405 return;
406
407 data = per_cpu_ptr(tr->trace_buffer.data, cpu);
408
409 if (unlikely(!data) ||
410 !data->critical_start || atomic_read(&data->disabled))
411 return;
412
413 atomic_inc(&data->disabled);
414
415 local_save_flags(flags);
416 __trace_function(tr, ip, parent_ip, flags, preempt_count());
417 check_critical_timing(tr, data, parent_ip ? : ip, cpu);
418 data->critical_start = 0;
419 atomic_dec(&data->disabled);
420}
421
422/* start and stop critical timings used to for stoppage (in idle) */
423void start_critical_timings(void)
424{
425 if (preempt_trace() || irq_trace())
426 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
427}
428EXPORT_SYMBOL_GPL(start_critical_timings);
429
430void stop_critical_timings(void)
431{
432 if (preempt_trace() || irq_trace())
433 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
434}
435EXPORT_SYMBOL_GPL(stop_critical_timings);
436
437#ifdef CONFIG_IRQSOFF_TRACER
438#ifdef CONFIG_PROVE_LOCKING
439void time_hardirqs_on(unsigned long a0, unsigned long a1)
440{
441 if (!preempt_trace() && irq_trace())
442 stop_critical_timing(a0, a1);
443}
444
445void time_hardirqs_off(unsigned long a0, unsigned long a1)
446{
447 if (!preempt_trace() && irq_trace())
448 start_critical_timing(a0, a1);
449}
450
451#else /* !CONFIG_PROVE_LOCKING */
452
453/*
454 * Stubs:
455 */
456
457void trace_softirqs_on(unsigned long ip)
458{
459}
460
461void trace_softirqs_off(unsigned long ip)
462{
463}
464
465inline void print_irqtrace_events(struct task_struct *curr)
466{
467}
468
469/*
470 * We are only interested in hardirq on/off events:
471 */
472void trace_hardirqs_on(void)
473{
474 if (!preempt_trace() && irq_trace())
475 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
476}
477EXPORT_SYMBOL(trace_hardirqs_on);
478
479void trace_hardirqs_off(void)
480{
481 if (!preempt_trace() && irq_trace())
482 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
483}
484EXPORT_SYMBOL(trace_hardirqs_off);
485
486__visible void trace_hardirqs_on_caller(unsigned long caller_addr)
487{
488 if (!preempt_trace() && irq_trace())
489 stop_critical_timing(CALLER_ADDR0, caller_addr);
490}
491EXPORT_SYMBOL(trace_hardirqs_on_caller);
492
493__visible void trace_hardirqs_off_caller(unsigned long caller_addr)
494{
495 if (!preempt_trace() && irq_trace())
496 start_critical_timing(CALLER_ADDR0, caller_addr);
497}
498EXPORT_SYMBOL(trace_hardirqs_off_caller);
499
500#endif /* CONFIG_PROVE_LOCKING */
501#endif /* CONFIG_IRQSOFF_TRACER */
502
503#ifdef CONFIG_PREEMPT_TRACER
504void trace_preempt_on(unsigned long a0, unsigned long a1)
505{
506 if (preempt_trace() && !irq_trace())
507 stop_critical_timing(a0, a1);
508}
509
510void trace_preempt_off(unsigned long a0, unsigned long a1)
511{
512 if (preempt_trace() && !irq_trace())
513 start_critical_timing(a0, a1);
514}
515#endif /* CONFIG_PREEMPT_TRACER */
516
517#ifdef CONFIG_FUNCTION_TRACER
518static bool function_enabled;
519
520static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
521{
522 int ret;
523
524 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
525 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
526 return 0;
527
528 if (graph)
529 ret = register_ftrace_graph(&irqsoff_graph_return,
530 &irqsoff_graph_entry);
531 else
532 ret = register_ftrace_function(tr->ops);
533
534 if (!ret)
535 function_enabled = true;
536
537 return ret;
538}
539
540static void unregister_irqsoff_function(struct trace_array *tr, int graph)
541{
542 if (!function_enabled)
543 return;
544
545 if (graph)
546 unregister_ftrace_graph();
547 else
548 unregister_ftrace_function(tr->ops);
549
550 function_enabled = false;
551}
552
553static int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
554{
555 if (!(mask & TRACE_ITER_FUNCTION))
556 return 0;
557
558 if (set)
559 register_irqsoff_function(tr, is_graph(tr), 1);
560 else
561 unregister_irqsoff_function(tr, is_graph(tr));
562 return 1;
563}
564#else
565static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
566{
567 return 0;
568}
569static void unregister_irqsoff_function(struct trace_array *tr, int graph) { }
570static inline int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
571{
572 return 0;
573}
574#endif /* CONFIG_FUNCTION_TRACER */
575
576static int irqsoff_flag_changed(struct trace_array *tr, u32 mask, int set)
577{
578 struct tracer *tracer = tr->current_trace;
579
580 if (irqsoff_function_set(tr, mask, set))
581 return 0;
582
583#ifdef CONFIG_FUNCTION_GRAPH_TRACER
584 if (mask & TRACE_ITER_DISPLAY_GRAPH)
585 return irqsoff_display_graph(tr, set);
586#endif
587
588 return trace_keep_overwrite(tracer, mask, set);
589}
590
591static int start_irqsoff_tracer(struct trace_array *tr, int graph)
592{
593 int ret;
594
595 ret = register_irqsoff_function(tr, graph, 0);
596
597 if (!ret && tracing_is_enabled())
598 tracer_enabled = 1;
599 else
600 tracer_enabled = 0;
601
602 return ret;
603}
604
605static void stop_irqsoff_tracer(struct trace_array *tr, int graph)
606{
607 tracer_enabled = 0;
608
609 unregister_irqsoff_function(tr, graph);
610}
611
612static bool irqsoff_busy;
613
614static int __irqsoff_tracer_init(struct trace_array *tr)
615{
616 if (irqsoff_busy)
617 return -EBUSY;
618
619 save_flags = tr->trace_flags;
620
621 /* non overwrite screws up the latency tracers */
622 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
623 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
624
625 tr->max_latency = 0;
626 irqsoff_trace = tr;
627 /* make sure that the tracer is visible */
628 smp_wmb();
629
630 ftrace_init_array_ops(tr, irqsoff_tracer_call);
631
632 /* Only toplevel instance supports graph tracing */
633 if (start_irqsoff_tracer(tr, (tr->flags & TRACE_ARRAY_FL_GLOBAL &&
634 is_graph(tr))))
635 printk(KERN_ERR "failed to start irqsoff tracer\n");
636
637 irqsoff_busy = true;
638 return 0;
639}
640
641static void irqsoff_tracer_reset(struct trace_array *tr)
642{
643 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
644 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
645
646 stop_irqsoff_tracer(tr, is_graph(tr));
647
648 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
649 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
650 ftrace_reset_array_ops(tr);
651
652 irqsoff_busy = false;
653}
654
655static void irqsoff_tracer_start(struct trace_array *tr)
656{
657 tracer_enabled = 1;
658}
659
660static void irqsoff_tracer_stop(struct trace_array *tr)
661{
662 tracer_enabled = 0;
663}
664
665#ifdef CONFIG_IRQSOFF_TRACER
666static int irqsoff_tracer_init(struct trace_array *tr)
667{
668 trace_type = TRACER_IRQS_OFF;
669
670 return __irqsoff_tracer_init(tr);
671}
672static struct tracer irqsoff_tracer __read_mostly =
673{
674 .name = "irqsoff",
675 .init = irqsoff_tracer_init,
676 .reset = irqsoff_tracer_reset,
677 .start = irqsoff_tracer_start,
678 .stop = irqsoff_tracer_stop,
679 .print_max = true,
680 .print_header = irqsoff_print_header,
681 .print_line = irqsoff_print_line,
682 .flag_changed = irqsoff_flag_changed,
683#ifdef CONFIG_FTRACE_SELFTEST
684 .selftest = trace_selftest_startup_irqsoff,
685#endif
686 .open = irqsoff_trace_open,
687 .close = irqsoff_trace_close,
688 .allow_instances = true,
689 .use_max_tr = true,
690};
691# define register_irqsoff(trace) register_tracer(&trace)
692#else
693# define register_irqsoff(trace) do { } while (0)
694#endif
695
696#ifdef CONFIG_PREEMPT_TRACER
697static int preemptoff_tracer_init(struct trace_array *tr)
698{
699 trace_type = TRACER_PREEMPT_OFF;
700
701 return __irqsoff_tracer_init(tr);
702}
703
704static struct tracer preemptoff_tracer __read_mostly =
705{
706 .name = "preemptoff",
707 .init = preemptoff_tracer_init,
708 .reset = irqsoff_tracer_reset,
709 .start = irqsoff_tracer_start,
710 .stop = irqsoff_tracer_stop,
711 .print_max = true,
712 .print_header = irqsoff_print_header,
713 .print_line = irqsoff_print_line,
714 .flag_changed = irqsoff_flag_changed,
715#ifdef CONFIG_FTRACE_SELFTEST
716 .selftest = trace_selftest_startup_preemptoff,
717#endif
718 .open = irqsoff_trace_open,
719 .close = irqsoff_trace_close,
720 .allow_instances = true,
721 .use_max_tr = true,
722};
723# define register_preemptoff(trace) register_tracer(&trace)
724#else
725# define register_preemptoff(trace) do { } while (0)
726#endif
727
728#if defined(CONFIG_IRQSOFF_TRACER) && \
729 defined(CONFIG_PREEMPT_TRACER)
730
731static int preemptirqsoff_tracer_init(struct trace_array *tr)
732{
733 trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
734
735 return __irqsoff_tracer_init(tr);
736}
737
738static struct tracer preemptirqsoff_tracer __read_mostly =
739{
740 .name = "preemptirqsoff",
741 .init = preemptirqsoff_tracer_init,
742 .reset = irqsoff_tracer_reset,
743 .start = irqsoff_tracer_start,
744 .stop = irqsoff_tracer_stop,
745 .print_max = true,
746 .print_header = irqsoff_print_header,
747 .print_line = irqsoff_print_line,
748 .flag_changed = irqsoff_flag_changed,
749#ifdef CONFIG_FTRACE_SELFTEST
750 .selftest = trace_selftest_startup_preemptirqsoff,
751#endif
752 .open = irqsoff_trace_open,
753 .close = irqsoff_trace_close,
754 .allow_instances = true,
755 .use_max_tr = true,
756};
757
758# define register_preemptirqsoff(trace) register_tracer(&trace)
759#else
760# define register_preemptirqsoff(trace) do { } while (0)
761#endif
762
763__init static int init_irqsoff_tracer(void)
764{
765 register_irqsoff(irqsoff_tracer);
766 register_preemptoff(preemptoff_tracer);
767 register_preemptirqsoff(preemptirqsoff_tracer);
768
769 return 0;
770}
771core_initcall(init_irqsoff_tracer);