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 (!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);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * trace irqs off critical timings
4 *
5 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
6 * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
7 *
8 * From code in the latency_tracer, that is:
9 *
10 * Copyright (C) 2004-2006 Ingo Molnar
11 * Copyright (C) 2004 Nadia Yvette Chambers
12 */
13#include <linux/kallsyms.h>
14#include <linux/uaccess.h>
15#include <linux/module.h>
16#include <linux/ftrace.h>
17#include <linux/kprobes.h>
18
19#include "trace.h"
20
21#include <trace/events/preemptirq.h>
22
23#if defined(CONFIG_IRQSOFF_TRACER) || defined(CONFIG_PREEMPT_TRACER)
24static struct trace_array *irqsoff_trace __read_mostly;
25static int tracer_enabled __read_mostly;
26
27static DEFINE_PER_CPU(int, tracing_cpu);
28
29static DEFINE_RAW_SPINLOCK(max_trace_lock);
30
31enum {
32 TRACER_IRQS_OFF = (1 << 1),
33 TRACER_PREEMPT_OFF = (1 << 2),
34};
35
36static int trace_type __read_mostly;
37
38static int save_flags;
39
40static void stop_irqsoff_tracer(struct trace_array *tr, int graph);
41static int start_irqsoff_tracer(struct trace_array *tr, int graph);
42
43#ifdef CONFIG_PREEMPT_TRACER
44static inline int
45preempt_trace(int pc)
46{
47 return ((trace_type & TRACER_PREEMPT_OFF) && pc);
48}
49#else
50# define preempt_trace(pc) (0)
51#endif
52
53#ifdef CONFIG_IRQSOFF_TRACER
54static inline int
55irq_trace(void)
56{
57 return ((trace_type & TRACER_IRQS_OFF) &&
58 irqs_disabled());
59}
60#else
61# define irq_trace() (0)
62#endif
63
64#ifdef CONFIG_FUNCTION_GRAPH_TRACER
65static int irqsoff_display_graph(struct trace_array *tr, int set);
66# define is_graph(tr) ((tr)->trace_flags & TRACE_ITER_DISPLAY_GRAPH)
67#else
68static inline int irqsoff_display_graph(struct trace_array *tr, int set)
69{
70 return -EINVAL;
71}
72# define is_graph(tr) false
73#endif
74
75/*
76 * Sequence count - we record it when starting a measurement and
77 * skip the latency if the sequence has changed - some other section
78 * did a maximum and could disturb our measurement with serial console
79 * printouts, etc. Truly coinciding maximum latencies should be rare
80 * and what happens together happens separately as well, so this doesn't
81 * decrease the validity of the maximum found:
82 */
83static __cacheline_aligned_in_smp unsigned long max_sequence;
84
85#ifdef CONFIG_FUNCTION_TRACER
86/*
87 * Prologue for the preempt and irqs off function tracers.
88 *
89 * Returns 1 if it is OK to continue, and data->disabled is
90 * incremented.
91 * 0 if the trace is to be ignored, and data->disabled
92 * is kept the same.
93 *
94 * Note, this function is also used outside this ifdef but
95 * inside the #ifdef of the function graph tracer below.
96 * This is OK, since the function graph tracer is
97 * dependent on the function tracer.
98 */
99static int func_prolog_dec(struct trace_array *tr,
100 struct trace_array_cpu **data,
101 unsigned long *flags)
102{
103 long disabled;
104 int cpu;
105
106 /*
107 * Does not matter if we preempt. We test the flags
108 * afterward, to see if irqs are disabled or not.
109 * If we preempt and get a false positive, the flags
110 * test will fail.
111 */
112 cpu = raw_smp_processor_id();
113 if (likely(!per_cpu(tracing_cpu, cpu)))
114 return 0;
115
116 local_save_flags(*flags);
117 /*
118 * Slight chance to get a false positive on tracing_cpu,
119 * although I'm starting to think there isn't a chance.
120 * Leave this for now just to be paranoid.
121 */
122 if (!irqs_disabled_flags(*flags) && !preempt_count())
123 return 0;
124
125 *data = per_cpu_ptr(tr->array_buffer.data, cpu);
126 disabled = atomic_inc_return(&(*data)->disabled);
127
128 if (likely(disabled == 1))
129 return 1;
130
131 atomic_dec(&(*data)->disabled);
132
133 return 0;
134}
135
136/*
137 * irqsoff uses its own tracer function to keep the overhead down:
138 */
139static void
140irqsoff_tracer_call(unsigned long ip, unsigned long parent_ip,
141 struct ftrace_ops *op, struct ftrace_regs *fregs)
142{
143 struct trace_array *tr = irqsoff_trace;
144 struct trace_array_cpu *data;
145 unsigned long flags;
146 unsigned int trace_ctx;
147
148 if (!func_prolog_dec(tr, &data, &flags))
149 return;
150
151 trace_ctx = tracing_gen_ctx_flags(flags);
152
153 trace_function(tr, ip, parent_ip, trace_ctx);
154
155 atomic_dec(&data->disabled);
156}
157#endif /* CONFIG_FUNCTION_TRACER */
158
159#ifdef CONFIG_FUNCTION_GRAPH_TRACER
160static int irqsoff_display_graph(struct trace_array *tr, int set)
161{
162 int cpu;
163
164 if (!(is_graph(tr) ^ set))
165 return 0;
166
167 stop_irqsoff_tracer(irqsoff_trace, !set);
168
169 for_each_possible_cpu(cpu)
170 per_cpu(tracing_cpu, cpu) = 0;
171
172 tr->max_latency = 0;
173 tracing_reset_online_cpus(&irqsoff_trace->array_buffer);
174
175 return start_irqsoff_tracer(irqsoff_trace, set);
176}
177
178static int irqsoff_graph_entry(struct ftrace_graph_ent *trace,
179 struct fgraph_ops *gops)
180{
181 struct trace_array *tr = irqsoff_trace;
182 struct trace_array_cpu *data;
183 unsigned long flags;
184 unsigned int trace_ctx;
185 u64 *calltime;
186 int ret;
187
188 if (ftrace_graph_ignore_func(gops, trace))
189 return 0;
190 /*
191 * Do not trace a function if it's filtered by set_graph_notrace.
192 * Make the index of ret stack negative to indicate that it should
193 * ignore further functions. But it needs its own ret stack entry
194 * to recover the original index in order to continue tracing after
195 * returning from the function.
196 */
197 if (ftrace_graph_notrace_addr(trace->func))
198 return 1;
199
200 if (!func_prolog_dec(tr, &data, &flags))
201 return 0;
202
203 calltime = fgraph_reserve_data(gops->idx, sizeof(*calltime));
204 if (!calltime)
205 return 0;
206
207 *calltime = trace_clock_local();
208
209 trace_ctx = tracing_gen_ctx_flags(flags);
210 ret = __trace_graph_entry(tr, trace, trace_ctx);
211 atomic_dec(&data->disabled);
212
213 return ret;
214}
215
216static void irqsoff_graph_return(struct ftrace_graph_ret *trace,
217 struct fgraph_ops *gops)
218{
219 struct trace_array *tr = irqsoff_trace;
220 struct trace_array_cpu *data;
221 unsigned long flags;
222 unsigned int trace_ctx;
223 u64 *calltime;
224 int size;
225
226 ftrace_graph_addr_finish(gops, trace);
227
228 if (!func_prolog_dec(tr, &data, &flags))
229 return;
230
231 calltime = fgraph_retrieve_data(gops->idx, &size);
232 if (!calltime)
233 return;
234 trace->calltime = *calltime;
235
236 trace_ctx = tracing_gen_ctx_flags(flags);
237 __trace_graph_return(tr, trace, trace_ctx);
238 atomic_dec(&data->disabled);
239}
240
241static struct fgraph_ops fgraph_ops = {
242 .entryfunc = &irqsoff_graph_entry,
243 .retfunc = &irqsoff_graph_return,
244};
245
246static void irqsoff_trace_open(struct trace_iterator *iter)
247{
248 if (is_graph(iter->tr))
249 graph_trace_open(iter);
250 else
251 iter->private = NULL;
252}
253
254static void irqsoff_trace_close(struct trace_iterator *iter)
255{
256 if (iter->private)
257 graph_trace_close(iter);
258}
259
260#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_CPU | \
261 TRACE_GRAPH_PRINT_PROC | \
262 TRACE_GRAPH_PRINT_REL_TIME | \
263 TRACE_GRAPH_PRINT_DURATION)
264
265static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
266{
267 /*
268 * In graph mode call the graph tracer output function,
269 * otherwise go with the TRACE_FN event handler
270 */
271 if (is_graph(iter->tr))
272 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
273
274 return TRACE_TYPE_UNHANDLED;
275}
276
277static void irqsoff_print_header(struct seq_file *s)
278{
279 struct trace_array *tr = irqsoff_trace;
280
281 if (is_graph(tr))
282 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
283 else
284 trace_default_header(s);
285}
286
287static void
288__trace_function(struct trace_array *tr,
289 unsigned long ip, unsigned long parent_ip,
290 unsigned int trace_ctx)
291{
292 if (is_graph(tr))
293 trace_graph_function(tr, ip, parent_ip, trace_ctx);
294 else
295 trace_function(tr, ip, parent_ip, trace_ctx);
296}
297
298#else
299#define __trace_function trace_function
300
301static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
302{
303 return TRACE_TYPE_UNHANDLED;
304}
305
306static void irqsoff_trace_open(struct trace_iterator *iter) { }
307static void irqsoff_trace_close(struct trace_iterator *iter) { }
308
309#ifdef CONFIG_FUNCTION_TRACER
310static void irqsoff_print_header(struct seq_file *s)
311{
312 trace_default_header(s);
313}
314#else
315static void irqsoff_print_header(struct seq_file *s)
316{
317 trace_latency_header(s);
318}
319#endif /* CONFIG_FUNCTION_TRACER */
320#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
321
322/*
323 * Should this new latency be reported/recorded?
324 */
325static bool report_latency(struct trace_array *tr, u64 delta)
326{
327 if (tracing_thresh) {
328 if (delta < tracing_thresh)
329 return false;
330 } else {
331 if (delta <= tr->max_latency)
332 return false;
333 }
334 return true;
335}
336
337static void
338check_critical_timing(struct trace_array *tr,
339 struct trace_array_cpu *data,
340 unsigned long parent_ip,
341 int cpu)
342{
343 u64 T0, T1, delta;
344 unsigned long flags;
345 unsigned int trace_ctx;
346
347 T0 = data->preempt_timestamp;
348 T1 = ftrace_now(cpu);
349 delta = T1-T0;
350
351 trace_ctx = tracing_gen_ctx();
352
353 if (!report_latency(tr, delta))
354 goto out;
355
356 raw_spin_lock_irqsave(&max_trace_lock, flags);
357
358 /* check if we are still the max latency */
359 if (!report_latency(tr, delta))
360 goto out_unlock;
361
362 __trace_function(tr, CALLER_ADDR0, parent_ip, trace_ctx);
363 /* Skip 5 functions to get to the irq/preempt enable function */
364 __trace_stack(tr, trace_ctx, 5);
365
366 if (data->critical_sequence != max_sequence)
367 goto out_unlock;
368
369 data->critical_end = parent_ip;
370
371 if (likely(!is_tracing_stopped())) {
372 tr->max_latency = delta;
373 update_max_tr_single(tr, current, cpu);
374 }
375
376 max_sequence++;
377
378out_unlock:
379 raw_spin_unlock_irqrestore(&max_trace_lock, flags);
380
381out:
382 data->critical_sequence = max_sequence;
383 data->preempt_timestamp = ftrace_now(cpu);
384 __trace_function(tr, CALLER_ADDR0, parent_ip, trace_ctx);
385}
386
387static nokprobe_inline void
388start_critical_timing(unsigned long ip, unsigned long parent_ip)
389{
390 int cpu;
391 struct trace_array *tr = irqsoff_trace;
392 struct trace_array_cpu *data;
393
394 if (!tracer_enabled || !tracing_is_enabled())
395 return;
396
397 cpu = raw_smp_processor_id();
398
399 if (per_cpu(tracing_cpu, cpu))
400 return;
401
402 data = per_cpu_ptr(tr->array_buffer.data, cpu);
403
404 if (unlikely(!data) || atomic_read(&data->disabled))
405 return;
406
407 atomic_inc(&data->disabled);
408
409 data->critical_sequence = max_sequence;
410 data->preempt_timestamp = ftrace_now(cpu);
411 data->critical_start = parent_ip ? : ip;
412
413 __trace_function(tr, ip, parent_ip, tracing_gen_ctx());
414
415 per_cpu(tracing_cpu, cpu) = 1;
416
417 atomic_dec(&data->disabled);
418}
419
420static nokprobe_inline void
421stop_critical_timing(unsigned long ip, unsigned long parent_ip)
422{
423 int cpu;
424 struct trace_array *tr = irqsoff_trace;
425 struct trace_array_cpu *data;
426 unsigned int trace_ctx;
427
428 cpu = raw_smp_processor_id();
429 /* Always clear the tracing cpu on stopping the trace */
430 if (unlikely(per_cpu(tracing_cpu, cpu)))
431 per_cpu(tracing_cpu, cpu) = 0;
432 else
433 return;
434
435 if (!tracer_enabled || !tracing_is_enabled())
436 return;
437
438 data = per_cpu_ptr(tr->array_buffer.data, cpu);
439
440 if (unlikely(!data) ||
441 !data->critical_start || atomic_read(&data->disabled))
442 return;
443
444 atomic_inc(&data->disabled);
445
446 trace_ctx = tracing_gen_ctx();
447 __trace_function(tr, ip, parent_ip, trace_ctx);
448 check_critical_timing(tr, data, parent_ip ? : ip, cpu);
449 data->critical_start = 0;
450 atomic_dec(&data->disabled);
451}
452
453/* start and stop critical timings used to for stoppage (in idle) */
454void start_critical_timings(void)
455{
456 if (preempt_trace(preempt_count()) || irq_trace())
457 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
458}
459EXPORT_SYMBOL_GPL(start_critical_timings);
460NOKPROBE_SYMBOL(start_critical_timings);
461
462void stop_critical_timings(void)
463{
464 if (preempt_trace(preempt_count()) || irq_trace())
465 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
466}
467EXPORT_SYMBOL_GPL(stop_critical_timings);
468NOKPROBE_SYMBOL(stop_critical_timings);
469
470#ifdef CONFIG_FUNCTION_TRACER
471static bool function_enabled;
472
473static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
474{
475 int ret;
476
477 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
478 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
479 return 0;
480
481 if (graph)
482 ret = register_ftrace_graph(&fgraph_ops);
483 else
484 ret = register_ftrace_function(tr->ops);
485
486 if (!ret)
487 function_enabled = true;
488
489 return ret;
490}
491
492static void unregister_irqsoff_function(struct trace_array *tr, int graph)
493{
494 if (!function_enabled)
495 return;
496
497 if (graph)
498 unregister_ftrace_graph(&fgraph_ops);
499 else
500 unregister_ftrace_function(tr->ops);
501
502 function_enabled = false;
503}
504
505static int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
506{
507 if (!(mask & TRACE_ITER_FUNCTION))
508 return 0;
509
510 if (set)
511 register_irqsoff_function(tr, is_graph(tr), 1);
512 else
513 unregister_irqsoff_function(tr, is_graph(tr));
514 return 1;
515}
516#else
517static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
518{
519 return 0;
520}
521static void unregister_irqsoff_function(struct trace_array *tr, int graph) { }
522static inline int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
523{
524 return 0;
525}
526#endif /* CONFIG_FUNCTION_TRACER */
527
528static int irqsoff_flag_changed(struct trace_array *tr, u32 mask, int set)
529{
530 struct tracer *tracer = tr->current_trace;
531
532 if (irqsoff_function_set(tr, mask, set))
533 return 0;
534
535#ifdef CONFIG_FUNCTION_GRAPH_TRACER
536 if (mask & TRACE_ITER_DISPLAY_GRAPH)
537 return irqsoff_display_graph(tr, set);
538#endif
539
540 return trace_keep_overwrite(tracer, mask, set);
541}
542
543static int start_irqsoff_tracer(struct trace_array *tr, int graph)
544{
545 int ret;
546
547 ret = register_irqsoff_function(tr, graph, 0);
548
549 if (!ret && tracing_is_enabled())
550 tracer_enabled = 1;
551 else
552 tracer_enabled = 0;
553
554 return ret;
555}
556
557static void stop_irqsoff_tracer(struct trace_array *tr, int graph)
558{
559 tracer_enabled = 0;
560
561 unregister_irqsoff_function(tr, graph);
562}
563
564static bool irqsoff_busy;
565
566static int __irqsoff_tracer_init(struct trace_array *tr)
567{
568 if (irqsoff_busy)
569 return -EBUSY;
570
571 save_flags = tr->trace_flags;
572
573 /* non overwrite screws up the latency tracers */
574 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
575 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
576 /* without pause, we will produce garbage if another latency occurs */
577 set_tracer_flag(tr, TRACE_ITER_PAUSE_ON_TRACE, 1);
578
579 tr->max_latency = 0;
580 irqsoff_trace = tr;
581 /* make sure that the tracer is visible */
582 smp_wmb();
583
584 ftrace_init_array_ops(tr, irqsoff_tracer_call);
585
586 /* Only toplevel instance supports graph tracing */
587 if (start_irqsoff_tracer(tr, (tr->flags & TRACE_ARRAY_FL_GLOBAL &&
588 is_graph(tr))))
589 printk(KERN_ERR "failed to start irqsoff tracer\n");
590
591 irqsoff_busy = true;
592 return 0;
593}
594
595static void __irqsoff_tracer_reset(struct trace_array *tr)
596{
597 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
598 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
599 int pause_flag = save_flags & TRACE_ITER_PAUSE_ON_TRACE;
600
601 stop_irqsoff_tracer(tr, is_graph(tr));
602
603 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
604 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
605 set_tracer_flag(tr, TRACE_ITER_PAUSE_ON_TRACE, pause_flag);
606 ftrace_reset_array_ops(tr);
607
608 irqsoff_busy = false;
609}
610
611static void irqsoff_tracer_start(struct trace_array *tr)
612{
613 tracer_enabled = 1;
614}
615
616static void irqsoff_tracer_stop(struct trace_array *tr)
617{
618 tracer_enabled = 0;
619}
620
621#ifdef CONFIG_IRQSOFF_TRACER
622/*
623 * We are only interested in hardirq on/off events:
624 */
625void tracer_hardirqs_on(unsigned long a0, unsigned long a1)
626{
627 if (!preempt_trace(preempt_count()) && irq_trace())
628 stop_critical_timing(a0, a1);
629}
630NOKPROBE_SYMBOL(tracer_hardirqs_on);
631
632void tracer_hardirqs_off(unsigned long a0, unsigned long a1)
633{
634 if (!preempt_trace(preempt_count()) && irq_trace())
635 start_critical_timing(a0, a1);
636}
637NOKPROBE_SYMBOL(tracer_hardirqs_off);
638
639static int irqsoff_tracer_init(struct trace_array *tr)
640{
641 trace_type = TRACER_IRQS_OFF;
642
643 return __irqsoff_tracer_init(tr);
644}
645
646static void irqsoff_tracer_reset(struct trace_array *tr)
647{
648 __irqsoff_tracer_reset(tr);
649}
650
651static struct tracer irqsoff_tracer __read_mostly =
652{
653 .name = "irqsoff",
654 .init = irqsoff_tracer_init,
655 .reset = irqsoff_tracer_reset,
656 .start = irqsoff_tracer_start,
657 .stop = irqsoff_tracer_stop,
658 .print_max = true,
659 .print_header = irqsoff_print_header,
660 .print_line = irqsoff_print_line,
661 .flag_changed = irqsoff_flag_changed,
662#ifdef CONFIG_FTRACE_SELFTEST
663 .selftest = trace_selftest_startup_irqsoff,
664#endif
665 .open = irqsoff_trace_open,
666 .close = irqsoff_trace_close,
667 .allow_instances = true,
668 .use_max_tr = true,
669};
670#endif /* CONFIG_IRQSOFF_TRACER */
671
672#ifdef CONFIG_PREEMPT_TRACER
673void tracer_preempt_on(unsigned long a0, unsigned long a1)
674{
675 if (preempt_trace(preempt_count()) && !irq_trace())
676 stop_critical_timing(a0, a1);
677}
678
679void tracer_preempt_off(unsigned long a0, unsigned long a1)
680{
681 if (preempt_trace(preempt_count()) && !irq_trace())
682 start_critical_timing(a0, a1);
683}
684
685static int preemptoff_tracer_init(struct trace_array *tr)
686{
687 trace_type = TRACER_PREEMPT_OFF;
688
689 return __irqsoff_tracer_init(tr);
690}
691
692static void preemptoff_tracer_reset(struct trace_array *tr)
693{
694 __irqsoff_tracer_reset(tr);
695}
696
697static struct tracer preemptoff_tracer __read_mostly =
698{
699 .name = "preemptoff",
700 .init = preemptoff_tracer_init,
701 .reset = preemptoff_tracer_reset,
702 .start = irqsoff_tracer_start,
703 .stop = irqsoff_tracer_stop,
704 .print_max = true,
705 .print_header = irqsoff_print_header,
706 .print_line = irqsoff_print_line,
707 .flag_changed = irqsoff_flag_changed,
708#ifdef CONFIG_FTRACE_SELFTEST
709 .selftest = trace_selftest_startup_preemptoff,
710#endif
711 .open = irqsoff_trace_open,
712 .close = irqsoff_trace_close,
713 .allow_instances = true,
714 .use_max_tr = true,
715};
716#endif /* CONFIG_PREEMPT_TRACER */
717
718#if defined(CONFIG_IRQSOFF_TRACER) && defined(CONFIG_PREEMPT_TRACER)
719
720static int preemptirqsoff_tracer_init(struct trace_array *tr)
721{
722 trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
723
724 return __irqsoff_tracer_init(tr);
725}
726
727static void preemptirqsoff_tracer_reset(struct trace_array *tr)
728{
729 __irqsoff_tracer_reset(tr);
730}
731
732static struct tracer preemptirqsoff_tracer __read_mostly =
733{
734 .name = "preemptirqsoff",
735 .init = preemptirqsoff_tracer_init,
736 .reset = preemptirqsoff_tracer_reset,
737 .start = irqsoff_tracer_start,
738 .stop = irqsoff_tracer_stop,
739 .print_max = true,
740 .print_header = irqsoff_print_header,
741 .print_line = irqsoff_print_line,
742 .flag_changed = irqsoff_flag_changed,
743#ifdef CONFIG_FTRACE_SELFTEST
744 .selftest = trace_selftest_startup_preemptirqsoff,
745#endif
746 .open = irqsoff_trace_open,
747 .close = irqsoff_trace_close,
748 .allow_instances = true,
749 .use_max_tr = true,
750};
751#endif
752
753__init static int init_irqsoff_tracer(void)
754{
755#ifdef CONFIG_IRQSOFF_TRACER
756 register_tracer(&irqsoff_tracer);
757#endif
758#ifdef CONFIG_PREEMPT_TRACER
759 register_tracer(&preemptoff_tracer);
760#endif
761#if defined(CONFIG_IRQSOFF_TRACER) && defined(CONFIG_PREEMPT_TRACER)
762 register_tracer(&preemptirqsoff_tracer);
763#endif
764
765 return 0;
766}
767core_initcall(init_irqsoff_tracer);
768#endif /* IRQSOFF_TRACER || PREEMPTOFF_TRACER */