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// 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 pt_regs *pt_regs)
142{
143 struct trace_array *tr = irqsoff_trace;
144 struct trace_array_cpu *data;
145 unsigned long flags;
146
147 if (!func_prolog_dec(tr, &data, &flags))
148 return;
149
150 trace_function(tr, ip, parent_ip, flags, preempt_count());
151
152 atomic_dec(&data->disabled);
153}
154#endif /* CONFIG_FUNCTION_TRACER */
155
156#ifdef CONFIG_FUNCTION_GRAPH_TRACER
157static int irqsoff_display_graph(struct trace_array *tr, int set)
158{
159 int cpu;
160
161 if (!(is_graph(tr) ^ set))
162 return 0;
163
164 stop_irqsoff_tracer(irqsoff_trace, !set);
165
166 for_each_possible_cpu(cpu)
167 per_cpu(tracing_cpu, cpu) = 0;
168
169 tr->max_latency = 0;
170 tracing_reset_online_cpus(&irqsoff_trace->array_buffer);
171
172 return start_irqsoff_tracer(irqsoff_trace, set);
173}
174
175static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
176{
177 struct trace_array *tr = irqsoff_trace;
178 struct trace_array_cpu *data;
179 unsigned long flags;
180 int ret;
181 int pc;
182
183 if (ftrace_graph_ignore_func(trace))
184 return 0;
185 /*
186 * Do not trace a function if it's filtered by set_graph_notrace.
187 * Make the index of ret stack negative to indicate that it should
188 * ignore further functions. But it needs its own ret stack entry
189 * to recover the original index in order to continue tracing after
190 * returning from the function.
191 */
192 if (ftrace_graph_notrace_addr(trace->func))
193 return 1;
194
195 if (!func_prolog_dec(tr, &data, &flags))
196 return 0;
197
198 pc = preempt_count();
199 ret = __trace_graph_entry(tr, trace, flags, pc);
200 atomic_dec(&data->disabled);
201
202 return ret;
203}
204
205static void irqsoff_graph_return(struct ftrace_graph_ret *trace)
206{
207 struct trace_array *tr = irqsoff_trace;
208 struct trace_array_cpu *data;
209 unsigned long flags;
210 int pc;
211
212 ftrace_graph_addr_finish(trace);
213
214 if (!func_prolog_dec(tr, &data, &flags))
215 return;
216
217 pc = preempt_count();
218 __trace_graph_return(tr, trace, flags, pc);
219 atomic_dec(&data->disabled);
220}
221
222static struct fgraph_ops fgraph_ops = {
223 .entryfunc = &irqsoff_graph_entry,
224 .retfunc = &irqsoff_graph_return,
225};
226
227static void irqsoff_trace_open(struct trace_iterator *iter)
228{
229 if (is_graph(iter->tr))
230 graph_trace_open(iter);
231
232}
233
234static void irqsoff_trace_close(struct trace_iterator *iter)
235{
236 if (iter->private)
237 graph_trace_close(iter);
238}
239
240#define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_CPU | \
241 TRACE_GRAPH_PRINT_PROC | \
242 TRACE_GRAPH_PRINT_REL_TIME | \
243 TRACE_GRAPH_PRINT_DURATION)
244
245static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
246{
247 /*
248 * In graph mode call the graph tracer output function,
249 * otherwise go with the TRACE_FN event handler
250 */
251 if (is_graph(iter->tr))
252 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
253
254 return TRACE_TYPE_UNHANDLED;
255}
256
257static void irqsoff_print_header(struct seq_file *s)
258{
259 struct trace_array *tr = irqsoff_trace;
260
261 if (is_graph(tr))
262 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
263 else
264 trace_default_header(s);
265}
266
267static void
268__trace_function(struct trace_array *tr,
269 unsigned long ip, unsigned long parent_ip,
270 unsigned long flags, int pc)
271{
272 if (is_graph(tr))
273 trace_graph_function(tr, ip, parent_ip, flags, pc);
274 else
275 trace_function(tr, ip, parent_ip, flags, pc);
276}
277
278#else
279#define __trace_function trace_function
280
281static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
282{
283 return TRACE_TYPE_UNHANDLED;
284}
285
286static void irqsoff_trace_open(struct trace_iterator *iter) { }
287static void irqsoff_trace_close(struct trace_iterator *iter) { }
288
289#ifdef CONFIG_FUNCTION_TRACER
290static void irqsoff_print_header(struct seq_file *s)
291{
292 trace_default_header(s);
293}
294#else
295static void irqsoff_print_header(struct seq_file *s)
296{
297 trace_latency_header(s);
298}
299#endif /* CONFIG_FUNCTION_TRACER */
300#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
301
302/*
303 * Should this new latency be reported/recorded?
304 */
305static bool report_latency(struct trace_array *tr, u64 delta)
306{
307 if (tracing_thresh) {
308 if (delta < tracing_thresh)
309 return false;
310 } else {
311 if (delta <= tr->max_latency)
312 return false;
313 }
314 return true;
315}
316
317static void
318check_critical_timing(struct trace_array *tr,
319 struct trace_array_cpu *data,
320 unsigned long parent_ip,
321 int cpu)
322{
323 u64 T0, T1, delta;
324 unsigned long flags;
325 int pc;
326
327 T0 = data->preempt_timestamp;
328 T1 = ftrace_now(cpu);
329 delta = T1-T0;
330
331 local_save_flags(flags);
332
333 pc = preempt_count();
334
335 if (!report_latency(tr, delta))
336 goto out;
337
338 raw_spin_lock_irqsave(&max_trace_lock, flags);
339
340 /* check if we are still the max latency */
341 if (!report_latency(tr, delta))
342 goto out_unlock;
343
344 __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
345 /* Skip 5 functions to get to the irq/preempt enable function */
346 __trace_stack(tr, flags, 5, pc);
347
348 if (data->critical_sequence != max_sequence)
349 goto out_unlock;
350
351 data->critical_end = parent_ip;
352
353 if (likely(!is_tracing_stopped())) {
354 tr->max_latency = delta;
355 update_max_tr_single(tr, current, cpu);
356 }
357
358 max_sequence++;
359
360out_unlock:
361 raw_spin_unlock_irqrestore(&max_trace_lock, flags);
362
363out:
364 data->critical_sequence = max_sequence;
365 data->preempt_timestamp = ftrace_now(cpu);
366 __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
367}
368
369static nokprobe_inline void
370start_critical_timing(unsigned long ip, unsigned long parent_ip, int pc)
371{
372 int cpu;
373 struct trace_array *tr = irqsoff_trace;
374 struct trace_array_cpu *data;
375 unsigned long flags;
376
377 if (!tracer_enabled || !tracing_is_enabled())
378 return;
379
380 cpu = raw_smp_processor_id();
381
382 if (per_cpu(tracing_cpu, cpu))
383 return;
384
385 data = per_cpu_ptr(tr->array_buffer.data, cpu);
386
387 if (unlikely(!data) || atomic_read(&data->disabled))
388 return;
389
390 atomic_inc(&data->disabled);
391
392 data->critical_sequence = max_sequence;
393 data->preempt_timestamp = ftrace_now(cpu);
394 data->critical_start = parent_ip ? : ip;
395
396 local_save_flags(flags);
397
398 __trace_function(tr, ip, parent_ip, flags, pc);
399
400 per_cpu(tracing_cpu, cpu) = 1;
401
402 atomic_dec(&data->disabled);
403}
404
405static nokprobe_inline void
406stop_critical_timing(unsigned long ip, unsigned long parent_ip, int pc)
407{
408 int cpu;
409 struct trace_array *tr = irqsoff_trace;
410 struct trace_array_cpu *data;
411 unsigned long flags;
412
413 cpu = raw_smp_processor_id();
414 /* Always clear the tracing cpu on stopping the trace */
415 if (unlikely(per_cpu(tracing_cpu, cpu)))
416 per_cpu(tracing_cpu, cpu) = 0;
417 else
418 return;
419
420 if (!tracer_enabled || !tracing_is_enabled())
421 return;
422
423 data = per_cpu_ptr(tr->array_buffer.data, cpu);
424
425 if (unlikely(!data) ||
426 !data->critical_start || atomic_read(&data->disabled))
427 return;
428
429 atomic_inc(&data->disabled);
430
431 local_save_flags(flags);
432 __trace_function(tr, ip, parent_ip, flags, pc);
433 check_critical_timing(tr, data, parent_ip ? : ip, cpu);
434 data->critical_start = 0;
435 atomic_dec(&data->disabled);
436}
437
438/* start and stop critical timings used to for stoppage (in idle) */
439void start_critical_timings(void)
440{
441 int pc = preempt_count();
442
443 if (preempt_trace(pc) || irq_trace())
444 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1, pc);
445}
446EXPORT_SYMBOL_GPL(start_critical_timings);
447NOKPROBE_SYMBOL(start_critical_timings);
448
449void stop_critical_timings(void)
450{
451 int pc = preempt_count();
452
453 if (preempt_trace(pc) || irq_trace())
454 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1, pc);
455}
456EXPORT_SYMBOL_GPL(stop_critical_timings);
457NOKPROBE_SYMBOL(stop_critical_timings);
458
459#ifdef CONFIG_FUNCTION_TRACER
460static bool function_enabled;
461
462static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
463{
464 int ret;
465
466 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */
467 if (function_enabled || (!set && !(tr->trace_flags & TRACE_ITER_FUNCTION)))
468 return 0;
469
470 if (graph)
471 ret = register_ftrace_graph(&fgraph_ops);
472 else
473 ret = register_ftrace_function(tr->ops);
474
475 if (!ret)
476 function_enabled = true;
477
478 return ret;
479}
480
481static void unregister_irqsoff_function(struct trace_array *tr, int graph)
482{
483 if (!function_enabled)
484 return;
485
486 if (graph)
487 unregister_ftrace_graph(&fgraph_ops);
488 else
489 unregister_ftrace_function(tr->ops);
490
491 function_enabled = false;
492}
493
494static int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
495{
496 if (!(mask & TRACE_ITER_FUNCTION))
497 return 0;
498
499 if (set)
500 register_irqsoff_function(tr, is_graph(tr), 1);
501 else
502 unregister_irqsoff_function(tr, is_graph(tr));
503 return 1;
504}
505#else
506static int register_irqsoff_function(struct trace_array *tr, int graph, int set)
507{
508 return 0;
509}
510static void unregister_irqsoff_function(struct trace_array *tr, int graph) { }
511static inline int irqsoff_function_set(struct trace_array *tr, u32 mask, int set)
512{
513 return 0;
514}
515#endif /* CONFIG_FUNCTION_TRACER */
516
517static int irqsoff_flag_changed(struct trace_array *tr, u32 mask, int set)
518{
519 struct tracer *tracer = tr->current_trace;
520
521 if (irqsoff_function_set(tr, mask, set))
522 return 0;
523
524#ifdef CONFIG_FUNCTION_GRAPH_TRACER
525 if (mask & TRACE_ITER_DISPLAY_GRAPH)
526 return irqsoff_display_graph(tr, set);
527#endif
528
529 return trace_keep_overwrite(tracer, mask, set);
530}
531
532static int start_irqsoff_tracer(struct trace_array *tr, int graph)
533{
534 int ret;
535
536 ret = register_irqsoff_function(tr, graph, 0);
537
538 if (!ret && tracing_is_enabled())
539 tracer_enabled = 1;
540 else
541 tracer_enabled = 0;
542
543 return ret;
544}
545
546static void stop_irqsoff_tracer(struct trace_array *tr, int graph)
547{
548 tracer_enabled = 0;
549
550 unregister_irqsoff_function(tr, graph);
551}
552
553static bool irqsoff_busy;
554
555static int __irqsoff_tracer_init(struct trace_array *tr)
556{
557 if (irqsoff_busy)
558 return -EBUSY;
559
560 save_flags = tr->trace_flags;
561
562 /* non overwrite screws up the latency tracers */
563 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1);
564 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1);
565
566 tr->max_latency = 0;
567 irqsoff_trace = tr;
568 /* make sure that the tracer is visible */
569 smp_wmb();
570
571 ftrace_init_array_ops(tr, irqsoff_tracer_call);
572
573 /* Only toplevel instance supports graph tracing */
574 if (start_irqsoff_tracer(tr, (tr->flags & TRACE_ARRAY_FL_GLOBAL &&
575 is_graph(tr))))
576 printk(KERN_ERR "failed to start irqsoff tracer\n");
577
578 irqsoff_busy = true;
579 return 0;
580}
581
582static void __irqsoff_tracer_reset(struct trace_array *tr)
583{
584 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
585 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
586
587 stop_irqsoff_tracer(tr, is_graph(tr));
588
589 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag);
590 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag);
591 ftrace_reset_array_ops(tr);
592
593 irqsoff_busy = false;
594}
595
596static void irqsoff_tracer_start(struct trace_array *tr)
597{
598 tracer_enabled = 1;
599}
600
601static void irqsoff_tracer_stop(struct trace_array *tr)
602{
603 tracer_enabled = 0;
604}
605
606#ifdef CONFIG_IRQSOFF_TRACER
607/*
608 * We are only interested in hardirq on/off events:
609 */
610void tracer_hardirqs_on(unsigned long a0, unsigned long a1)
611{
612 unsigned int pc = preempt_count();
613
614 if (!preempt_trace(pc) && irq_trace())
615 stop_critical_timing(a0, a1, pc);
616}
617NOKPROBE_SYMBOL(tracer_hardirqs_on);
618
619void tracer_hardirqs_off(unsigned long a0, unsigned long a1)
620{
621 unsigned int pc = preempt_count();
622
623 if (!preempt_trace(pc) && irq_trace())
624 start_critical_timing(a0, a1, pc);
625}
626NOKPROBE_SYMBOL(tracer_hardirqs_off);
627
628static int irqsoff_tracer_init(struct trace_array *tr)
629{
630 trace_type = TRACER_IRQS_OFF;
631
632 return __irqsoff_tracer_init(tr);
633}
634
635static void irqsoff_tracer_reset(struct trace_array *tr)
636{
637 __irqsoff_tracer_reset(tr);
638}
639
640static struct tracer irqsoff_tracer __read_mostly =
641{
642 .name = "irqsoff",
643 .init = irqsoff_tracer_init,
644 .reset = irqsoff_tracer_reset,
645 .start = irqsoff_tracer_start,
646 .stop = irqsoff_tracer_stop,
647 .print_max = true,
648 .print_header = irqsoff_print_header,
649 .print_line = irqsoff_print_line,
650 .flag_changed = irqsoff_flag_changed,
651#ifdef CONFIG_FTRACE_SELFTEST
652 .selftest = trace_selftest_startup_irqsoff,
653#endif
654 .open = irqsoff_trace_open,
655 .close = irqsoff_trace_close,
656 .allow_instances = true,
657 .use_max_tr = true,
658};
659#endif /* CONFIG_IRQSOFF_TRACER */
660
661#ifdef CONFIG_PREEMPT_TRACER
662void tracer_preempt_on(unsigned long a0, unsigned long a1)
663{
664 int pc = preempt_count();
665
666 if (preempt_trace(pc) && !irq_trace())
667 stop_critical_timing(a0, a1, pc);
668}
669
670void tracer_preempt_off(unsigned long a0, unsigned long a1)
671{
672 int pc = preempt_count();
673
674 if (preempt_trace(pc) && !irq_trace())
675 start_critical_timing(a0, a1, pc);
676}
677
678static int preemptoff_tracer_init(struct trace_array *tr)
679{
680 trace_type = TRACER_PREEMPT_OFF;
681
682 return __irqsoff_tracer_init(tr);
683}
684
685static void preemptoff_tracer_reset(struct trace_array *tr)
686{
687 __irqsoff_tracer_reset(tr);
688}
689
690static struct tracer preemptoff_tracer __read_mostly =
691{
692 .name = "preemptoff",
693 .init = preemptoff_tracer_init,
694 .reset = preemptoff_tracer_reset,
695 .start = irqsoff_tracer_start,
696 .stop = irqsoff_tracer_stop,
697 .print_max = true,
698 .print_header = irqsoff_print_header,
699 .print_line = irqsoff_print_line,
700 .flag_changed = irqsoff_flag_changed,
701#ifdef CONFIG_FTRACE_SELFTEST
702 .selftest = trace_selftest_startup_preemptoff,
703#endif
704 .open = irqsoff_trace_open,
705 .close = irqsoff_trace_close,
706 .allow_instances = true,
707 .use_max_tr = true,
708};
709#endif /* CONFIG_PREEMPT_TRACER */
710
711#if defined(CONFIG_IRQSOFF_TRACER) && defined(CONFIG_PREEMPT_TRACER)
712
713static int preemptirqsoff_tracer_init(struct trace_array *tr)
714{
715 trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
716
717 return __irqsoff_tracer_init(tr);
718}
719
720static void preemptirqsoff_tracer_reset(struct trace_array *tr)
721{
722 __irqsoff_tracer_reset(tr);
723}
724
725static struct tracer preemptirqsoff_tracer __read_mostly =
726{
727 .name = "preemptirqsoff",
728 .init = preemptirqsoff_tracer_init,
729 .reset = preemptirqsoff_tracer_reset,
730 .start = irqsoff_tracer_start,
731 .stop = irqsoff_tracer_stop,
732 .print_max = true,
733 .print_header = irqsoff_print_header,
734 .print_line = irqsoff_print_line,
735 .flag_changed = irqsoff_flag_changed,
736#ifdef CONFIG_FTRACE_SELFTEST
737 .selftest = trace_selftest_startup_preemptirqsoff,
738#endif
739 .open = irqsoff_trace_open,
740 .close = irqsoff_trace_close,
741 .allow_instances = true,
742 .use_max_tr = true,
743};
744#endif
745
746__init static int init_irqsoff_tracer(void)
747{
748#ifdef CONFIG_IRQSOFF_TRACER
749 register_tracer(&irqsoff_tracer);
750#endif
751#ifdef CONFIG_PREEMPT_TRACER
752 register_tracer(&preemptoff_tracer);
753#endif
754#if defined(CONFIG_IRQSOFF_TRACER) && defined(CONFIG_PREEMPT_TRACER)
755 register_tracer(&preemptirqsoff_tracer);
756#endif
757
758 return 0;
759}
760core_initcall(init_irqsoff_tracer);
761#endif /* IRQSOFF_TRACER || PREEMPTOFF_TRACER */