Loading...
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{
180 struct trace_array *tr = irqsoff_trace;
181 struct trace_array_cpu *data;
182 unsigned long flags;
183 unsigned int trace_ctx;
184 int ret;
185
186 if (ftrace_graph_ignore_func(trace))
187 return 0;
188 /*
189 * Do not trace a function if it's filtered by set_graph_notrace.
190 * Make the index of ret stack negative to indicate that it should
191 * ignore further functions. But it needs its own ret stack entry
192 * to recover the original index in order to continue tracing after
193 * returning from the function.
194 */
195 if (ftrace_graph_notrace_addr(trace->func))
196 return 1;
197
198 if (!func_prolog_dec(tr, &data, &flags))
199 return 0;
200
201 trace_ctx = tracing_gen_ctx_flags(flags);
202 ret = __trace_graph_entry(tr, trace, trace_ctx);
203 atomic_dec(&data->disabled);
204
205 return ret;
206}
207
208static void irqsoff_graph_return(struct ftrace_graph_ret *trace)
209{
210 struct trace_array *tr = irqsoff_trace;
211 struct trace_array_cpu *data;
212 unsigned long flags;
213 unsigned int trace_ctx;
214
215 ftrace_graph_addr_finish(trace);
216
217 if (!func_prolog_dec(tr, &data, &flags))
218 return;
219
220 trace_ctx = tracing_gen_ctx_flags(flags);
221 __trace_graph_return(tr, trace, trace_ctx);
222 atomic_dec(&data->disabled);
223}
224
225static struct fgraph_ops fgraph_ops = {
226 .entryfunc = &irqsoff_graph_entry,
227 .retfunc = &irqsoff_graph_return,
228};
229
230static void irqsoff_trace_open(struct trace_iterator *iter)
231{
232 if (is_graph(iter->tr))
233 graph_trace_open(iter);
234 else
235 iter->private = NULL;
236}
237
238static void irqsoff_trace_close(struct trace_iterator *iter)
239{
240 if (iter->private)
241 graph_trace_close(iter);
242}
243
244#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_CPU | \
245 TRACE_GRAPH_PRINT_PROC | \
246 TRACE_GRAPH_PRINT_REL_TIME | \
247 TRACE_GRAPH_PRINT_DURATION)
248
249static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
250{
251 /*
252 * In graph mode call the graph tracer output function,
253 * otherwise go with the TRACE_FN event handler
254 */
255 if (is_graph(iter->tr))
256 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
257
258 return TRACE_TYPE_UNHANDLED;
259}
260
261static void irqsoff_print_header(struct seq_file *s)
262{
263 struct trace_array *tr = irqsoff_trace;
264
265 if (is_graph(tr))
266 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
267 else
268 trace_default_header(s);
269}
270
271static void
272__trace_function(struct trace_array *tr,
273 unsigned long ip, unsigned long parent_ip,
274 unsigned int trace_ctx)
275{
276 if (is_graph(tr))
277 trace_graph_function(tr, ip, parent_ip, trace_ctx);
278 else
279 trace_function(tr, ip, parent_ip, trace_ctx);
280}
281
282#else
283#define __trace_function trace_function
284
285static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
286{
287 return TRACE_TYPE_UNHANDLED;
288}
289
290static void irqsoff_trace_open(struct trace_iterator *iter) { }
291static void irqsoff_trace_close(struct trace_iterator *iter) { }
292
293#ifdef CONFIG_FUNCTION_TRACER
294static void irqsoff_print_header(struct seq_file *s)
295{
296 trace_default_header(s);
297}
298#else
299static void irqsoff_print_header(struct seq_file *s)
300{
301 trace_latency_header(s);
302}
303#endif /* CONFIG_FUNCTION_TRACER */
304#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
305
306/*
307 * Should this new latency be reported/recorded?
308 */
309static bool report_latency(struct trace_array *tr, u64 delta)
310{
311 if (tracing_thresh) {
312 if (delta < tracing_thresh)
313 return false;
314 } else {
315 if (delta <= tr->max_latency)
316 return false;
317 }
318 return true;
319}
320
321static void
322check_critical_timing(struct trace_array *tr,
323 struct trace_array_cpu *data,
324 unsigned long parent_ip,
325 int cpu)
326{
327 u64 T0, T1, delta;
328 unsigned long flags;
329 unsigned int trace_ctx;
330
331 T0 = data->preempt_timestamp;
332 T1 = ftrace_now(cpu);
333 delta = T1-T0;
334
335 trace_ctx = tracing_gen_ctx();
336
337 if (!report_latency(tr, delta))
338 goto out;
339
340 raw_spin_lock_irqsave(&max_trace_lock, flags);
341
342 /* check if we are still the max latency */
343 if (!report_latency(tr, delta))
344 goto out_unlock;
345
346 __trace_function(tr, CALLER_ADDR0, parent_ip, trace_ctx);
347 /* Skip 5 functions to get to the irq/preempt enable function */
348 __trace_stack(tr, trace_ctx, 5);
349
350 if (data->critical_sequence != max_sequence)
351 goto out_unlock;
352
353 data->critical_end = parent_ip;
354
355 if (likely(!is_tracing_stopped())) {
356 tr->max_latency = delta;
357 update_max_tr_single(tr, current, cpu);
358 }
359
360 max_sequence++;
361
362out_unlock:
363 raw_spin_unlock_irqrestore(&max_trace_lock, flags);
364
365out:
366 data->critical_sequence = max_sequence;
367 data->preempt_timestamp = ftrace_now(cpu);
368 __trace_function(tr, CALLER_ADDR0, parent_ip, trace_ctx);
369}
370
371static nokprobe_inline void
372start_critical_timing(unsigned long ip, unsigned long parent_ip)
373{
374 int cpu;
375 struct trace_array *tr = irqsoff_trace;
376 struct trace_array_cpu *data;
377
378 if (!tracer_enabled || !tracing_is_enabled())
379 return;
380
381 cpu = raw_smp_processor_id();
382
383 if (per_cpu(tracing_cpu, cpu))
384 return;
385
386 data = per_cpu_ptr(tr->array_buffer.data, cpu);
387
388 if (unlikely(!data) || atomic_read(&data->disabled))
389 return;
390
391 atomic_inc(&data->disabled);
392
393 data->critical_sequence = max_sequence;
394 data->preempt_timestamp = ftrace_now(cpu);
395 data->critical_start = parent_ip ? : ip;
396
397 __trace_function(tr, ip, parent_ip, tracing_gen_ctx());
398
399 per_cpu(tracing_cpu, cpu) = 1;
400
401 atomic_dec(&data->disabled);
402}
403
404static nokprobe_inline void
405stop_critical_timing(unsigned long ip, unsigned long parent_ip)
406{
407 int cpu;
408 struct trace_array *tr = irqsoff_trace;
409 struct trace_array_cpu *data;
410 unsigned int trace_ctx;
411
412 cpu = raw_smp_processor_id();
413 /* Always clear the tracing cpu on stopping the trace */
414 if (unlikely(per_cpu(tracing_cpu, cpu)))
415 per_cpu(tracing_cpu, cpu) = 0;
416 else
417 return;
418
419 if (!tracer_enabled || !tracing_is_enabled())
420 return;
421
422 data = per_cpu_ptr(tr->array_buffer.data, cpu);
423
424 if (unlikely(!data) ||
425 !data->critical_start || atomic_read(&data->disabled))
426 return;
427
428 atomic_inc(&data->disabled);
429
430 trace_ctx = tracing_gen_ctx();
431 __trace_function(tr, ip, parent_ip, trace_ctx);
432 check_critical_timing(tr, data, parent_ip ? : ip, cpu);
433 data->critical_start = 0;
434 atomic_dec(&data->disabled);
435}
436
437/* start and stop critical timings used to for stoppage (in idle) */
438void start_critical_timings(void)
439{
440 if (preempt_trace(preempt_count()) || irq_trace())
441 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
442}
443EXPORT_SYMBOL_GPL(start_critical_timings);
444NOKPROBE_SYMBOL(start_critical_timings);
445
446void stop_critical_timings(void)
447{
448 if (preempt_trace(preempt_count()) || irq_trace())
449 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
450}
451EXPORT_SYMBOL_GPL(stop_critical_timings);
452NOKPROBE_SYMBOL(stop_critical_timings);
453
454#ifdef CONFIG_FUNCTION_TRACER
455static bool function_enabled;
456
457static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
458{
459 int ret;
460
461 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
462 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
463 return 0;
464
465 if (graph)
466 ret = register_ftrace_graph(&fgraph_ops);
467 else
468 ret = register_ftrace_function(tr->ops);
469
470 if (!ret)
471 function_enabled = true;
472
473 return ret;
474}
475
476static void unregister_irqsoff_function(struct trace_array *tr, int graph)
477{
478 if (!function_enabled)
479 return;
480
481 if (graph)
482 unregister_ftrace_graph(&fgraph_ops);
483 else
484 unregister_ftrace_function(tr->ops);
485
486 function_enabled = false;
487}
488
489static int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
490{
491 if (!(mask & TRACE_ITER_FUNCTION))
492 return 0;
493
494 if (set)
495 register_irqsoff_function(tr, is_graph(tr), 1);
496 else
497 unregister_irqsoff_function(tr, is_graph(tr));
498 return 1;
499}
500#else
501static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
502{
503 return 0;
504}
505static void unregister_irqsoff_function(struct trace_array *tr, int graph) { }
506static inline int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
507{
508 return 0;
509}
510#endif /* CONFIG_FUNCTION_TRACER */
511
512static int irqsoff_flag_changed(struct trace_array *tr, u32 mask, int set)
513{
514 struct tracer *tracer = tr->current_trace;
515
516 if (irqsoff_function_set(tr, mask, set))
517 return 0;
518
519#ifdef CONFIG_FUNCTION_GRAPH_TRACER
520 if (mask & TRACE_ITER_DISPLAY_GRAPH)
521 return irqsoff_display_graph(tr, set);
522#endif
523
524 return trace_keep_overwrite(tracer, mask, set);
525}
526
527static int start_irqsoff_tracer(struct trace_array *tr, int graph)
528{
529 int ret;
530
531 ret = register_irqsoff_function(tr, graph, 0);
532
533 if (!ret && tracing_is_enabled())
534 tracer_enabled = 1;
535 else
536 tracer_enabled = 0;
537
538 return ret;
539}
540
541static void stop_irqsoff_tracer(struct trace_array *tr, int graph)
542{
543 tracer_enabled = 0;
544
545 unregister_irqsoff_function(tr, graph);
546}
547
548static bool irqsoff_busy;
549
550static int __irqsoff_tracer_init(struct trace_array *tr)
551{
552 if (irqsoff_busy)
553 return -EBUSY;
554
555 save_flags = tr->trace_flags;
556
557 /* non overwrite screws up the latency tracers */
558 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
559 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
560 /* without pause, we will produce garbage if another latency occurs */
561 set_tracer_flag(tr, TRACE_ITER_PAUSE_ON_TRACE, 1);
562
563 tr->max_latency = 0;
564 irqsoff_trace = tr;
565 /* make sure that the tracer is visible */
566 smp_wmb();
567
568 ftrace_init_array_ops(tr, irqsoff_tracer_call);
569
570 /* Only toplevel instance supports graph tracing */
571 if (start_irqsoff_tracer(tr, (tr->flags & TRACE_ARRAY_FL_GLOBAL &&
572 is_graph(tr))))
573 printk(KERN_ERR "failed to start irqsoff tracer\n");
574
575 irqsoff_busy = true;
576 return 0;
577}
578
579static void __irqsoff_tracer_reset(struct trace_array *tr)
580{
581 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
582 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
583 int pause_flag = save_flags & TRACE_ITER_PAUSE_ON_TRACE;
584
585 stop_irqsoff_tracer(tr, is_graph(tr));
586
587 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
588 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
589 set_tracer_flag(tr, TRACE_ITER_PAUSE_ON_TRACE, pause_flag);
590 ftrace_reset_array_ops(tr);
591
592 irqsoff_busy = false;
593}
594
595static void irqsoff_tracer_start(struct trace_array *tr)
596{
597 tracer_enabled = 1;
598}
599
600static void irqsoff_tracer_stop(struct trace_array *tr)
601{
602 tracer_enabled = 0;
603}
604
605#ifdef CONFIG_IRQSOFF_TRACER
606/*
607 * We are only interested in hardirq on/off events:
608 */
609void tracer_hardirqs_on(unsigned long a0, unsigned long a1)
610{
611 if (!preempt_trace(preempt_count()) && irq_trace())
612 stop_critical_timing(a0, a1);
613}
614NOKPROBE_SYMBOL(tracer_hardirqs_on);
615
616void tracer_hardirqs_off(unsigned long a0, unsigned long a1)
617{
618 if (!preempt_trace(preempt_count()) && irq_trace())
619 start_critical_timing(a0, a1);
620}
621NOKPROBE_SYMBOL(tracer_hardirqs_off);
622
623static int irqsoff_tracer_init(struct trace_array *tr)
624{
625 trace_type = TRACER_IRQS_OFF;
626
627 return __irqsoff_tracer_init(tr);
628}
629
630static void irqsoff_tracer_reset(struct trace_array *tr)
631{
632 __irqsoff_tracer_reset(tr);
633}
634
635static struct tracer irqsoff_tracer __read_mostly =
636{
637 .name = "irqsoff",
638 .init = irqsoff_tracer_init,
639 .reset = irqsoff_tracer_reset,
640 .start = irqsoff_tracer_start,
641 .stop = irqsoff_tracer_stop,
642 .print_max = true,
643 .print_header = irqsoff_print_header,
644 .print_line = irqsoff_print_line,
645 .flag_changed = irqsoff_flag_changed,
646#ifdef CONFIG_FTRACE_SELFTEST
647 .selftest = trace_selftest_startup_irqsoff,
648#endif
649 .open = irqsoff_trace_open,
650 .close = irqsoff_trace_close,
651 .allow_instances = true,
652 .use_max_tr = true,
653};
654#endif /* CONFIG_IRQSOFF_TRACER */
655
656#ifdef CONFIG_PREEMPT_TRACER
657void tracer_preempt_on(unsigned long a0, unsigned long a1)
658{
659 if (preempt_trace(preempt_count()) && !irq_trace())
660 stop_critical_timing(a0, a1);
661}
662
663void tracer_preempt_off(unsigned long a0, unsigned long a1)
664{
665 if (preempt_trace(preempt_count()) && !irq_trace())
666 start_critical_timing(a0, a1);
667}
668
669static int preemptoff_tracer_init(struct trace_array *tr)
670{
671 trace_type = TRACER_PREEMPT_OFF;
672
673 return __irqsoff_tracer_init(tr);
674}
675
676static void preemptoff_tracer_reset(struct trace_array *tr)
677{
678 __irqsoff_tracer_reset(tr);
679}
680
681static struct tracer preemptoff_tracer __read_mostly =
682{
683 .name = "preemptoff",
684 .init = preemptoff_tracer_init,
685 .reset = preemptoff_tracer_reset,
686 .start = irqsoff_tracer_start,
687 .stop = irqsoff_tracer_stop,
688 .print_max = true,
689 .print_header = irqsoff_print_header,
690 .print_line = irqsoff_print_line,
691 .flag_changed = irqsoff_flag_changed,
692#ifdef CONFIG_FTRACE_SELFTEST
693 .selftest = trace_selftest_startup_preemptoff,
694#endif
695 .open = irqsoff_trace_open,
696 .close = irqsoff_trace_close,
697 .allow_instances = true,
698 .use_max_tr = true,
699};
700#endif /* CONFIG_PREEMPT_TRACER */
701
702#if defined(CONFIG_IRQSOFF_TRACER) && defined(CONFIG_PREEMPT_TRACER)
703
704static int preemptirqsoff_tracer_init(struct trace_array *tr)
705{
706 trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
707
708 return __irqsoff_tracer_init(tr);
709}
710
711static void preemptirqsoff_tracer_reset(struct trace_array *tr)
712{
713 __irqsoff_tracer_reset(tr);
714}
715
716static struct tracer preemptirqsoff_tracer __read_mostly =
717{
718 .name = "preemptirqsoff",
719 .init = preemptirqsoff_tracer_init,
720 .reset = preemptirqsoff_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_preemptirqsoff,
729#endif
730 .open = irqsoff_trace_open,
731 .close = irqsoff_trace_close,
732 .allow_instances = true,
733 .use_max_tr = true,
734};
735#endif
736
737__init static int init_irqsoff_tracer(void)
738{
739#ifdef CONFIG_IRQSOFF_TRACER
740 register_tracer(&irqsoff_tracer);
741#endif
742#ifdef CONFIG_PREEMPT_TRACER
743 register_tracer(&preemptoff_tracer);
744#endif
745#if defined(CONFIG_IRQSOFF_TRACER) && defined(CONFIG_PREEMPT_TRACER)
746 register_tracer(&preemptirqsoff_tracer);
747#endif
748
749 return 0;
750}
751core_initcall(init_irqsoff_tracer);
752#endif /* IRQSOFF_TRACER || PREEMPTOFF_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 (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);