Merge branch 'tracing/ftrace'; commit 'v2.6.29-rc2' into tracing/core
[linux-2.6] / kernel / trace / trace_output.c
1 /*
2  * trace_output.c
3  *
4  * Copyright (C) 2008 Red Hat Inc, Steven Rostedt <srostedt@redhat.com>
5  *
6  */
7
8 #include <linux/module.h>
9 #include <linux/mutex.h>
10 #include <linux/ftrace.h>
11
12 #include "trace_output.h"
13
14 /* must be a power of 2 */
15 #define EVENT_HASHSIZE  128
16
17 static DEFINE_MUTEX(trace_event_mutex);
18 static struct hlist_head event_hash[EVENT_HASHSIZE] __read_mostly;
19
20 static int next_event_type = __TRACE_LAST_TYPE + 1;
21
22 /**
23  * trace_seq_printf - sequence printing of trace information
24  * @s: trace sequence descriptor
25  * @fmt: printf format string
26  *
27  * The tracer may use either sequence operations or its own
28  * copy to user routines. To simplify formating of a trace
29  * trace_seq_printf is used to store strings into a special
30  * buffer (@s). Then the output may be either used by
31  * the sequencer or pulled into another buffer.
32  */
33 int
34 trace_seq_printf(struct trace_seq *s, const char *fmt, ...)
35 {
36         int len = (PAGE_SIZE - 1) - s->len;
37         va_list ap;
38         int ret;
39
40         if (!len)
41                 return 0;
42
43         va_start(ap, fmt);
44         ret = vsnprintf(s->buffer + s->len, len, fmt, ap);
45         va_end(ap);
46
47         /* If we can't write it all, don't bother writing anything */
48         if (ret >= len)
49                 return 0;
50
51         s->len += ret;
52
53         return len;
54 }
55
56 /**
57  * trace_seq_puts - trace sequence printing of simple string
58  * @s: trace sequence descriptor
59  * @str: simple string to record
60  *
61  * The tracer may use either the sequence operations or its own
62  * copy to user routines. This function records a simple string
63  * into a special buffer (@s) for later retrieval by a sequencer
64  * or other mechanism.
65  */
66 int trace_seq_puts(struct trace_seq *s, const char *str)
67 {
68         int len = strlen(str);
69
70         if (len > ((PAGE_SIZE - 1) - s->len))
71                 return 0;
72
73         memcpy(s->buffer + s->len, str, len);
74         s->len += len;
75
76         return len;
77 }
78
79 int trace_seq_putc(struct trace_seq *s, unsigned char c)
80 {
81         if (s->len >= (PAGE_SIZE - 1))
82                 return 0;
83
84         s->buffer[s->len++] = c;
85
86         return 1;
87 }
88
89 int trace_seq_putmem(struct trace_seq *s, void *mem, size_t len)
90 {
91         if (len > ((PAGE_SIZE - 1) - s->len))
92                 return 0;
93
94         memcpy(s->buffer + s->len, mem, len);
95         s->len += len;
96
97         return len;
98 }
99
100 int trace_seq_putmem_hex(struct trace_seq *s, void *mem, size_t len)
101 {
102         unsigned char hex[HEX_CHARS];
103         unsigned char *data = mem;
104         int i, j;
105
106 #ifdef __BIG_ENDIAN
107         for (i = 0, j = 0; i < len; i++) {
108 #else
109         for (i = len-1, j = 0; i >= 0; i--) {
110 #endif
111                 hex[j++] = hex_asc_hi(data[i]);
112                 hex[j++] = hex_asc_lo(data[i]);
113         }
114         hex[j++] = ' ';
115
116         return trace_seq_putmem(s, hex, j);
117 }
118
119 int trace_seq_path(struct trace_seq *s, struct path *path)
120 {
121         unsigned char *p;
122
123         if (s->len >= (PAGE_SIZE - 1))
124                 return 0;
125         p = d_path(path, s->buffer + s->len, PAGE_SIZE - s->len);
126         if (!IS_ERR(p)) {
127                 p = mangle_path(s->buffer + s->len, p, "\n");
128                 if (p) {
129                         s->len = p - s->buffer;
130                         return 1;
131                 }
132         } else {
133                 s->buffer[s->len++] = '?';
134                 return 1;
135         }
136
137         return 0;
138 }
139
140 #ifdef CONFIG_KRETPROBES
141 static inline const char *kretprobed(const char *name)
142 {
143         static const char tramp_name[] = "kretprobe_trampoline";
144         int size = sizeof(tramp_name);
145
146         if (strncmp(tramp_name, name, size) == 0)
147                 return "[unknown/kretprobe'd]";
148         return name;
149 }
150 #else
151 static inline const char *kretprobed(const char *name)
152 {
153         return name;
154 }
155 #endif /* CONFIG_KRETPROBES */
156
157 static int
158 seq_print_sym_short(struct trace_seq *s, const char *fmt, unsigned long address)
159 {
160 #ifdef CONFIG_KALLSYMS
161         char str[KSYM_SYMBOL_LEN];
162         const char *name;
163
164         kallsyms_lookup(address, NULL, NULL, NULL, str);
165
166         name = kretprobed(str);
167
168         return trace_seq_printf(s, fmt, name);
169 #endif
170         return 1;
171 }
172
173 static int
174 seq_print_sym_offset(struct trace_seq *s, const char *fmt,
175                      unsigned long address)
176 {
177 #ifdef CONFIG_KALLSYMS
178         char str[KSYM_SYMBOL_LEN];
179         const char *name;
180
181         sprint_symbol(str, address);
182         name = kretprobed(str);
183
184         return trace_seq_printf(s, fmt, name);
185 #endif
186         return 1;
187 }
188
189 #ifndef CONFIG_64BIT
190 # define IP_FMT "%08lx"
191 #else
192 # define IP_FMT "%016lx"
193 #endif
194
195 int seq_print_user_ip(struct trace_seq *s, struct mm_struct *mm,
196                       unsigned long ip, unsigned long sym_flags)
197 {
198         struct file *file = NULL;
199         unsigned long vmstart = 0;
200         int ret = 1;
201
202         if (mm) {
203                 const struct vm_area_struct *vma;
204
205                 down_read(&mm->mmap_sem);
206                 vma = find_vma(mm, ip);
207                 if (vma) {
208                         file = vma->vm_file;
209                         vmstart = vma->vm_start;
210                 }
211                 if (file) {
212                         ret = trace_seq_path(s, &file->f_path);
213                         if (ret)
214                                 ret = trace_seq_printf(s, "[+0x%lx]",
215                                                        ip - vmstart);
216                 }
217                 up_read(&mm->mmap_sem);
218         }
219         if (ret && ((sym_flags & TRACE_ITER_SYM_ADDR) || !file))
220                 ret = trace_seq_printf(s, " <" IP_FMT ">", ip);
221         return ret;
222 }
223
224 int
225 seq_print_userip_objs(const struct userstack_entry *entry, struct trace_seq *s,
226                       unsigned long sym_flags)
227 {
228         struct mm_struct *mm = NULL;
229         int ret = 1;
230         unsigned int i;
231
232         if (trace_flags & TRACE_ITER_SYM_USEROBJ) {
233                 struct task_struct *task;
234                 /*
235                  * we do the lookup on the thread group leader,
236                  * since individual threads might have already quit!
237                  */
238                 rcu_read_lock();
239                 task = find_task_by_vpid(entry->ent.tgid);
240                 if (task)
241                         mm = get_task_mm(task);
242                 rcu_read_unlock();
243         }
244
245         for (i = 0; i < FTRACE_STACK_ENTRIES; i++) {
246                 unsigned long ip = entry->caller[i];
247
248                 if (ip == ULONG_MAX || !ret)
249                         break;
250                 if (i && ret)
251                         ret = trace_seq_puts(s, " <- ");
252                 if (!ip) {
253                         if (ret)
254                                 ret = trace_seq_puts(s, "??");
255                         continue;
256                 }
257                 if (!ret)
258                         break;
259                 if (ret)
260                         ret = seq_print_user_ip(s, mm, ip, sym_flags);
261         }
262
263         if (mm)
264                 mmput(mm);
265         return ret;
266 }
267
268 int
269 seq_print_ip_sym(struct trace_seq *s, unsigned long ip, unsigned long sym_flags)
270 {
271         int ret;
272
273         if (!ip)
274                 return trace_seq_printf(s, "0");
275
276         if (sym_flags & TRACE_ITER_SYM_OFFSET)
277                 ret = seq_print_sym_offset(s, "%s", ip);
278         else
279                 ret = seq_print_sym_short(s, "%s", ip);
280
281         if (!ret)
282                 return 0;
283
284         if (sym_flags & TRACE_ITER_SYM_ADDR)
285                 ret = trace_seq_printf(s, " <" IP_FMT ">", ip);
286         return ret;
287 }
288
289 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
290
291 static int task_state_char(unsigned long state)
292 {
293         int bit = state ? __ffs(state) + 1 : 0;
294
295         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
296 }
297
298 /**
299  * ftrace_find_event - find a registered event
300  * @type: the type of event to look for
301  *
302  * Returns an event of type @type otherwise NULL
303  */
304 struct trace_event *ftrace_find_event(int type)
305 {
306         struct trace_event *event;
307         struct hlist_node *n;
308         unsigned key;
309
310         key = type & (EVENT_HASHSIZE - 1);
311
312         hlist_for_each_entry_rcu(event, n, &event_hash[key], node) {
313                 if (event->type == type)
314                         return event;
315         }
316
317         return NULL;
318 }
319
320 /**
321  * register_ftrace_event - register output for an event type
322  * @event: the event type to register
323  *
324  * Event types are stored in a hash and this hash is used to
325  * find a way to print an event. If the @event->type is set
326  * then it will use that type, otherwise it will assign a
327  * type to use.
328  *
329  * If you assign your own type, please make sure it is added
330  * to the trace_type enum in trace.h, to avoid collisions
331  * with the dynamic types.
332  *
333  * Returns the event type number or zero on error.
334  */
335 int register_ftrace_event(struct trace_event *event)
336 {
337         unsigned key;
338         int ret = 0;
339
340         mutex_lock(&trace_event_mutex);
341
342         if (!event->type)
343                 event->type = next_event_type++;
344         else if (event->type > __TRACE_LAST_TYPE) {
345                 printk(KERN_WARNING "Need to add type to trace.h\n");
346                 WARN_ON(1);
347         }
348
349         if (ftrace_find_event(event->type))
350                 goto out;
351
352         key = event->type & (EVENT_HASHSIZE - 1);
353
354         hlist_add_head_rcu(&event->node, &event_hash[key]);
355
356         ret = event->type;
357  out:
358         mutex_unlock(&trace_event_mutex);
359
360         return ret;
361 }
362
363 /**
364  * unregister_ftrace_event - remove a no longer used event
365  * @event: the event to remove
366  */
367 int unregister_ftrace_event(struct trace_event *event)
368 {
369         mutex_lock(&trace_event_mutex);
370         hlist_del(&event->node);
371         mutex_unlock(&trace_event_mutex);
372
373         return 0;
374 }
375
376 /*
377  * Standard events
378  */
379
380 int
381 trace_nop_print(struct trace_seq *s, struct trace_entry *entry, int flags)
382 {
383         return 0;
384 }
385
386 /* TRACE_FN */
387 static int
388 trace_fn_latency(struct trace_seq *s, struct trace_entry *entry, int flags)
389 {
390         struct ftrace_entry *field;
391
392         trace_assign_type(field, entry);
393
394         if (!seq_print_ip_sym(s, field->ip, flags))
395                 goto partial;
396         if (!trace_seq_puts(s, " ("))
397                 goto partial;
398         if (!seq_print_ip_sym(s, field->parent_ip, flags))
399                 goto partial;
400         if (!trace_seq_puts(s, ")\n"))
401                 goto partial;
402
403         return 0;
404
405  partial:
406         return TRACE_TYPE_PARTIAL_LINE;
407 }
408
409 static int
410 trace_fn_trace(struct trace_seq *s, struct trace_entry *entry, int flags)
411 {
412         struct ftrace_entry *field;
413
414         trace_assign_type(field, entry);
415
416         if (!seq_print_ip_sym(s, field->ip, flags))
417                 goto partial;
418
419         if ((flags & TRACE_ITER_PRINT_PARENT) && field->parent_ip) {
420                 if (!trace_seq_printf(s, " <-"))
421                         goto partial;
422                 if (!seq_print_ip_sym(s,
423                                       field->parent_ip,
424                                       flags))
425                         goto partial;
426         }
427         if (!trace_seq_printf(s, "\n"))
428                 goto partial;
429
430         return 0;
431
432  partial:
433         return TRACE_TYPE_PARTIAL_LINE;
434 }
435
436 static int
437 trace_fn_raw(struct trace_seq *s, struct trace_entry *entry, int flags)
438 {
439         struct ftrace_entry *field;
440
441         trace_assign_type(field, entry);
442
443         if (!trace_seq_printf(s, "%lx %lx\n",
444                               field->ip,
445                               field->parent_ip))
446                 return TRACE_TYPE_PARTIAL_LINE;
447
448         return 0;
449 }
450
451 static int
452 trace_fn_hex(struct trace_seq *s, struct trace_entry *entry, int flags)
453 {
454         struct ftrace_entry *field;
455
456         trace_assign_type(field, entry);
457
458         SEQ_PUT_HEX_FIELD_RET(s, field->ip);
459         SEQ_PUT_HEX_FIELD_RET(s, field->parent_ip);
460
461         return 0;
462 }
463
464 static int
465 trace_fn_bin(struct trace_seq *s, struct trace_entry *entry, int flags)
466 {
467         struct ftrace_entry *field;
468
469         trace_assign_type(field, entry);
470
471         SEQ_PUT_FIELD_RET(s, field->ip);
472         SEQ_PUT_FIELD_RET(s, field->parent_ip);
473
474         return 0;
475 }
476
477 static struct trace_event trace_fn_event = {
478         .type           = TRACE_FN,
479         .trace          = trace_fn_trace,
480         .latency_trace  = trace_fn_latency,
481         .raw            = trace_fn_raw,
482         .hex            = trace_fn_hex,
483         .binary         = trace_fn_bin,
484 };
485
486 /* TRACE_CTX an TRACE_WAKE */
487 static int
488 trace_ctxwake_print(struct trace_seq *s, struct trace_entry *entry, int flags,
489                     char *delim)
490 {
491         struct ctx_switch_entry *field;
492         char *comm;
493         int S, T;
494
495         trace_assign_type(field, entry);
496
497         T = task_state_char(field->next_state);
498         S = task_state_char(field->prev_state);
499         comm = trace_find_cmdline(field->next_pid);
500         if (!trace_seq_printf(s, " %5d:%3d:%c %s [%03d] %5d:%3d:%c %s\n",
501                               field->prev_pid,
502                               field->prev_prio,
503                               S, delim,
504                               field->next_cpu,
505                               field->next_pid,
506                               field->next_prio,
507                               T, comm))
508                 return TRACE_TYPE_PARTIAL_LINE;
509
510         return 0;
511 }
512
513 static int
514 trace_ctx_print(struct trace_seq *s, struct trace_entry *entry, int flags)
515 {
516         return trace_ctxwake_print(s, entry, flags, "==>");
517 }
518
519 static int
520 trace_wake_print(struct trace_seq *s, struct trace_entry *entry, int flags)
521 {
522         return trace_ctxwake_print(s, entry, flags, "  +");
523 }
524
525 static int
526 trace_ctxwake_raw(struct trace_seq *s, struct trace_entry *entry, int flags,
527                   char S)
528 {
529         struct ctx_switch_entry *field;
530         int T;
531
532         trace_assign_type(field, entry);
533
534         if (!S)
535                 task_state_char(field->prev_state);
536         T = task_state_char(field->next_state);
537         if (!trace_seq_printf(s, "%d %d %c %d %d %d %c\n",
538                               field->prev_pid,
539                               field->prev_prio,
540                               S,
541                               field->next_cpu,
542                               field->next_pid,
543                               field->next_prio,
544                               T))
545                 return TRACE_TYPE_PARTIAL_LINE;
546
547         return 0;
548 }
549
550 static int
551 trace_ctx_raw(struct trace_seq *s, struct trace_entry *entry, int flags)
552 {
553         return trace_ctxwake_raw(s, entry, flags, 0);
554 }
555
556 static int
557 trace_wake_raw(struct trace_seq *s, struct trace_entry *entry, int flags)
558 {
559         return trace_ctxwake_raw(s, entry, flags, '+');
560 }
561
562
563 static int
564 trace_ctxwake_hex(struct trace_seq *s, struct trace_entry *entry, int flags,
565                   char S)
566 {
567         struct ctx_switch_entry *field;
568         int T;
569
570         trace_assign_type(field, entry);
571
572         if (!S)
573                 task_state_char(field->prev_state);
574         T = task_state_char(field->next_state);
575
576         SEQ_PUT_HEX_FIELD_RET(s, field->prev_pid);
577         SEQ_PUT_HEX_FIELD_RET(s, field->prev_prio);
578         SEQ_PUT_HEX_FIELD_RET(s, S);
579         SEQ_PUT_HEX_FIELD_RET(s, field->next_cpu);
580         SEQ_PUT_HEX_FIELD_RET(s, field->next_pid);
581         SEQ_PUT_HEX_FIELD_RET(s, field->next_prio);
582         SEQ_PUT_HEX_FIELD_RET(s, T);
583
584         return 0;
585 }
586
587 static int
588 trace_ctx_hex(struct trace_seq *s, struct trace_entry *entry, int flags)
589 {
590         return trace_ctxwake_hex(s, entry, flags, 0);
591 }
592
593 static int
594 trace_wake_hex(struct trace_seq *s, struct trace_entry *entry, int flags)
595 {
596         return trace_ctxwake_hex(s, entry, flags, '+');
597 }
598
599 static int
600 trace_ctxwake_bin(struct trace_seq *s, struct trace_entry *entry, int flags)
601 {
602         struct ctx_switch_entry *field;
603
604         trace_assign_type(field, entry);
605
606         SEQ_PUT_FIELD_RET(s, field->prev_pid);
607         SEQ_PUT_FIELD_RET(s, field->prev_prio);
608         SEQ_PUT_FIELD_RET(s, field->prev_state);
609         SEQ_PUT_FIELD_RET(s, field->next_pid);
610         SEQ_PUT_FIELD_RET(s, field->next_prio);
611         SEQ_PUT_FIELD_RET(s, field->next_state);
612
613         return 0;
614 }
615
616 static struct trace_event trace_ctx_event = {
617         .type           = TRACE_CTX,
618         .trace          = trace_ctx_print,
619         .latency_trace  = trace_ctx_print,
620         .raw            = trace_ctx_raw,
621         .hex            = trace_ctx_hex,
622         .binary         = trace_ctxwake_bin,
623 };
624
625 static struct trace_event trace_wake_event = {
626         .type           = TRACE_WAKE,
627         .trace          = trace_wake_print,
628         .latency_trace  = trace_wake_print,
629         .raw            = trace_wake_raw,
630         .hex            = trace_wake_hex,
631         .binary         = trace_ctxwake_bin,
632 };
633
634 /* TRACE_SPECIAL */
635 static int
636 trace_special_print(struct trace_seq *s, struct trace_entry *entry, int flags)
637 {
638         struct special_entry *field;
639
640         trace_assign_type(field, entry);
641
642         if (!trace_seq_printf(s, "# %ld %ld %ld\n",
643                               field->arg1,
644                               field->arg2,
645                               field->arg3))
646                 return TRACE_TYPE_PARTIAL_LINE;
647
648         return 0;
649 }
650
651 static int
652 trace_special_hex(struct trace_seq *s, struct trace_entry *entry, int flags)
653 {
654         struct special_entry *field;
655
656         trace_assign_type(field, entry);
657
658         SEQ_PUT_HEX_FIELD_RET(s, field->arg1);
659         SEQ_PUT_HEX_FIELD_RET(s, field->arg2);
660         SEQ_PUT_HEX_FIELD_RET(s, field->arg3);
661
662         return 0;
663 }
664
665 static int
666 trace_special_bin(struct trace_seq *s, struct trace_entry *entry, int flags)
667 {
668         struct special_entry *field;
669
670         trace_assign_type(field, entry);
671
672         SEQ_PUT_FIELD_RET(s, field->arg1);
673         SEQ_PUT_FIELD_RET(s, field->arg2);
674         SEQ_PUT_FIELD_RET(s, field->arg3);
675
676         return 0;
677 }
678
679 static struct trace_event trace_special_event = {
680         .type           = TRACE_SPECIAL,
681         .trace          = trace_special_print,
682         .latency_trace  = trace_special_print,
683         .raw            = trace_special_print,
684         .hex            = trace_special_hex,
685         .binary         = trace_special_bin,
686 };
687
688 /* TRACE_STACK */
689
690 static int
691 trace_stack_print(struct trace_seq *s, struct trace_entry *entry, int flags)
692 {
693         struct stack_entry *field;
694         int i;
695
696         trace_assign_type(field, entry);
697
698         for (i = 0; i < FTRACE_STACK_ENTRIES; i++) {
699                 if (i) {
700                         if (!trace_seq_puts(s, " <= "))
701                                 goto partial;
702
703                         if (!seq_print_ip_sym(s, field->caller[i], flags))
704                                 goto partial;
705                 }
706                 if (!trace_seq_puts(s, "\n"))
707                         goto partial;
708         }
709
710         return 0;
711
712  partial:
713         return TRACE_TYPE_PARTIAL_LINE;
714 }
715
716 static struct trace_event trace_stack_event = {
717         .type           = TRACE_STACK,
718         .trace          = trace_stack_print,
719         .latency_trace  = trace_stack_print,
720         .raw            = trace_special_print,
721         .hex            = trace_special_hex,
722         .binary         = trace_special_bin,
723 };
724
725 /* TRACE_USER_STACK */
726 static int
727 trace_user_stack_print(struct trace_seq *s, struct trace_entry *entry,
728                        int flags)
729 {
730         struct userstack_entry *field;
731
732         trace_assign_type(field, entry);
733
734         if (!seq_print_userip_objs(field, s, flags))
735                 goto partial;
736
737         if (!trace_seq_putc(s, '\n'))
738                 goto partial;
739
740         return 0;
741
742  partial:
743         return TRACE_TYPE_PARTIAL_LINE;
744 }
745
746 static struct trace_event trace_user_stack_event = {
747         .type           = TRACE_USER_STACK,
748         .trace          = trace_user_stack_print,
749         .latency_trace  = trace_user_stack_print,
750         .raw            = trace_special_print,
751         .hex            = trace_special_hex,
752         .binary         = trace_special_bin,
753 };
754
755 /* TRACE_PRINT */
756 static int
757 trace_print_print(struct trace_seq *s, struct trace_entry *entry, int flags)
758 {
759         struct print_entry *field;
760
761         trace_assign_type(field, entry);
762
763         if (!seq_print_ip_sym(s, field->ip, flags))
764                 goto partial;
765
766         if (!trace_seq_printf(s, ": %s", field->buf))
767                 goto partial;
768
769         return 0;
770
771  partial:
772         return TRACE_TYPE_PARTIAL_LINE;
773 }
774
775 static int
776 trace_print_raw(struct trace_seq *s, struct trace_entry *entry, int flags)
777 {
778         struct print_entry *field;
779
780         trace_assign_type(field, entry);
781
782         if (!trace_seq_printf(s, "# %lx %s", field->ip, field->buf))
783                 goto partial;
784
785         return 0;
786
787  partial:
788         return TRACE_TYPE_PARTIAL_LINE;
789 }
790
791 static struct trace_event trace_print_event = {
792         .type           = TRACE_PRINT,
793         .trace          = trace_print_print,
794         .latency_trace  = trace_print_print,
795         .raw            = trace_print_raw,
796         .hex            = trace_nop_print,
797         .binary         = trace_nop_print,
798 };
799
800 static struct trace_event *events[] __initdata = {
801         &trace_fn_event,
802         &trace_ctx_event,
803         &trace_wake_event,
804         &trace_special_event,
805         &trace_stack_event,
806         &trace_user_stack_event,
807         &trace_print_event,
808         NULL
809 };
810
811 __init static int init_events(void)
812 {
813         struct trace_event *event;
814         int i, ret;
815
816         for (i = 0; events[i]; i++) {
817                 event = events[i];
818
819                 ret = register_ftrace_event(event);
820                 if (!ret) {
821                         printk(KERN_WARNING "event %d failed to register\n",
822                                event->type);
823                         WARN_ON_ONCE(1);
824                 }
825         }
826
827         return 0;
828 }
829 device_initcall(init_events);