Loading...
1#include <trace/syscall.h>
2#include <trace/events/syscalls.h>
3#include <linux/syscalls.h>
4#include <linux/slab.h>
5#include <linux/kernel.h>
6#include <linux/module.h> /* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
7#include <linux/ftrace.h>
8#include <linux/perf_event.h>
9#include <asm/syscall.h>
10
11#include "trace_output.h"
12#include "trace.h"
13
14static DEFINE_MUTEX(syscall_trace_lock);
15
16static int syscall_enter_register(struct trace_event_call *event,
17 enum trace_reg type, void *data);
18static int syscall_exit_register(struct trace_event_call *event,
19 enum trace_reg type, void *data);
20
21static struct list_head *
22syscall_get_enter_fields(struct trace_event_call *call)
23{
24 struct syscall_metadata *entry = call->data;
25
26 return &entry->enter_fields;
27}
28
29extern struct syscall_metadata *__start_syscalls_metadata[];
30extern struct syscall_metadata *__stop_syscalls_metadata[];
31
32static struct syscall_metadata **syscalls_metadata;
33
34#ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
35static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
36{
37 /*
38 * Only compare after the "sys" prefix. Archs that use
39 * syscall wrappers may have syscalls symbols aliases prefixed
40 * with ".SyS" or ".sys" instead of "sys", leading to an unwanted
41 * mismatch.
42 */
43 return !strcmp(sym + 3, name + 3);
44}
45#endif
46
47#ifdef ARCH_TRACE_IGNORE_COMPAT_SYSCALLS
48/*
49 * Some architectures that allow for 32bit applications
50 * to run on a 64bit kernel, do not map the syscalls for
51 * the 32bit tasks the same as they do for 64bit tasks.
52 *
53 * *cough*x86*cough*
54 *
55 * In such a case, instead of reporting the wrong syscalls,
56 * simply ignore them.
57 *
58 * For an arch to ignore the compat syscalls it needs to
59 * define ARCH_TRACE_IGNORE_COMPAT_SYSCALLS as well as
60 * define the function arch_trace_is_compat_syscall() to let
61 * the tracing system know that it should ignore it.
62 */
63static int
64trace_get_syscall_nr(struct task_struct *task, struct pt_regs *regs)
65{
66 if (unlikely(arch_trace_is_compat_syscall(regs)))
67 return -1;
68
69 return syscall_get_nr(task, regs);
70}
71#else
72static inline int
73trace_get_syscall_nr(struct task_struct *task, struct pt_regs *regs)
74{
75 return syscall_get_nr(task, regs);
76}
77#endif /* ARCH_TRACE_IGNORE_COMPAT_SYSCALLS */
78
79static __init struct syscall_metadata *
80find_syscall_meta(unsigned long syscall)
81{
82 struct syscall_metadata **start;
83 struct syscall_metadata **stop;
84 char str[KSYM_SYMBOL_LEN];
85
86
87 start = __start_syscalls_metadata;
88 stop = __stop_syscalls_metadata;
89 kallsyms_lookup(syscall, NULL, NULL, NULL, str);
90
91 if (arch_syscall_match_sym_name(str, "sys_ni_syscall"))
92 return NULL;
93
94 for ( ; start < stop; start++) {
95 if ((*start)->name && arch_syscall_match_sym_name(str, (*start)->name))
96 return *start;
97 }
98 return NULL;
99}
100
101static struct syscall_metadata *syscall_nr_to_meta(int nr)
102{
103 if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
104 return NULL;
105
106 return syscalls_metadata[nr];
107}
108
109const char *get_syscall_name(int syscall)
110{
111 struct syscall_metadata *entry;
112
113 entry = syscall_nr_to_meta(syscall);
114 if (!entry)
115 return NULL;
116
117 return entry->name;
118}
119
120static enum print_line_t
121print_syscall_enter(struct trace_iterator *iter, int flags,
122 struct trace_event *event)
123{
124 struct trace_array *tr = iter->tr;
125 struct trace_seq *s = &iter->seq;
126 struct trace_entry *ent = iter->ent;
127 struct syscall_trace_enter *trace;
128 struct syscall_metadata *entry;
129 int i, syscall;
130
131 trace = (typeof(trace))ent;
132 syscall = trace->nr;
133 entry = syscall_nr_to_meta(syscall);
134
135 if (!entry)
136 goto end;
137
138 if (entry->enter_event->event.type != ent->type) {
139 WARN_ON_ONCE(1);
140 goto end;
141 }
142
143 trace_seq_printf(s, "%s(", entry->name);
144
145 for (i = 0; i < entry->nb_args; i++) {
146
147 if (trace_seq_has_overflowed(s))
148 goto end;
149
150 /* parameter types */
151 if (tr->trace_flags & TRACE_ITER_VERBOSE)
152 trace_seq_printf(s, "%s ", entry->types[i]);
153
154 /* parameter values */
155 trace_seq_printf(s, "%s: %lx%s", entry->args[i],
156 trace->args[i],
157 i == entry->nb_args - 1 ? "" : ", ");
158 }
159
160 trace_seq_putc(s, ')');
161end:
162 trace_seq_putc(s, '\n');
163
164 return trace_handle_return(s);
165}
166
167static enum print_line_t
168print_syscall_exit(struct trace_iterator *iter, int flags,
169 struct trace_event *event)
170{
171 struct trace_seq *s = &iter->seq;
172 struct trace_entry *ent = iter->ent;
173 struct syscall_trace_exit *trace;
174 int syscall;
175 struct syscall_metadata *entry;
176
177 trace = (typeof(trace))ent;
178 syscall = trace->nr;
179 entry = syscall_nr_to_meta(syscall);
180
181 if (!entry) {
182 trace_seq_putc(s, '\n');
183 goto out;
184 }
185
186 if (entry->exit_event->event.type != ent->type) {
187 WARN_ON_ONCE(1);
188 return TRACE_TYPE_UNHANDLED;
189 }
190
191 trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
192 trace->ret);
193
194 out:
195 return trace_handle_return(s);
196}
197
198extern char *__bad_type_size(void);
199
200#define SYSCALL_FIELD(type, field, name) \
201 sizeof(type) != sizeof(trace.field) ? \
202 __bad_type_size() : \
203 #type, #name, offsetof(typeof(trace), field), \
204 sizeof(trace.field), is_signed_type(type)
205
206static int __init
207__set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
208{
209 int i;
210 int pos = 0;
211
212 /* When len=0, we just calculate the needed length */
213#define LEN_OR_ZERO (len ? len - pos : 0)
214
215 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
216 for (i = 0; i < entry->nb_args; i++) {
217 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
218 entry->args[i], sizeof(unsigned long),
219 i == entry->nb_args - 1 ? "" : ", ");
220 }
221 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
222
223 for (i = 0; i < entry->nb_args; i++) {
224 pos += snprintf(buf + pos, LEN_OR_ZERO,
225 ", ((unsigned long)(REC->%s))", entry->args[i]);
226 }
227
228#undef LEN_OR_ZERO
229
230 /* return the length of print_fmt */
231 return pos;
232}
233
234static int __init set_syscall_print_fmt(struct trace_event_call *call)
235{
236 char *print_fmt;
237 int len;
238 struct syscall_metadata *entry = call->data;
239
240 if (entry->enter_event != call) {
241 call->print_fmt = "\"0x%lx\", REC->ret";
242 return 0;
243 }
244
245 /* First: called with 0 length to calculate the needed length */
246 len = __set_enter_print_fmt(entry, NULL, 0);
247
248 print_fmt = kmalloc(len + 1, GFP_KERNEL);
249 if (!print_fmt)
250 return -ENOMEM;
251
252 /* Second: actually write the @print_fmt */
253 __set_enter_print_fmt(entry, print_fmt, len + 1);
254 call->print_fmt = print_fmt;
255
256 return 0;
257}
258
259static void __init free_syscall_print_fmt(struct trace_event_call *call)
260{
261 struct syscall_metadata *entry = call->data;
262
263 if (entry->enter_event == call)
264 kfree(call->print_fmt);
265}
266
267static int __init syscall_enter_define_fields(struct trace_event_call *call)
268{
269 struct syscall_trace_enter trace;
270 struct syscall_metadata *meta = call->data;
271 int ret;
272 int i;
273 int offset = offsetof(typeof(trace), args);
274
275 ret = trace_define_field(call, SYSCALL_FIELD(int, nr, __syscall_nr),
276 FILTER_OTHER);
277 if (ret)
278 return ret;
279
280 for (i = 0; i < meta->nb_args; i++) {
281 ret = trace_define_field(call, meta->types[i],
282 meta->args[i], offset,
283 sizeof(unsigned long), 0,
284 FILTER_OTHER);
285 offset += sizeof(unsigned long);
286 }
287
288 return ret;
289}
290
291static int __init syscall_exit_define_fields(struct trace_event_call *call)
292{
293 struct syscall_trace_exit trace;
294 int ret;
295
296 ret = trace_define_field(call, SYSCALL_FIELD(int, nr, __syscall_nr),
297 FILTER_OTHER);
298 if (ret)
299 return ret;
300
301 ret = trace_define_field(call, SYSCALL_FIELD(long, ret, ret),
302 FILTER_OTHER);
303
304 return ret;
305}
306
307static void ftrace_syscall_enter(void *data, struct pt_regs *regs, long id)
308{
309 struct trace_array *tr = data;
310 struct trace_event_file *trace_file;
311 struct syscall_trace_enter *entry;
312 struct syscall_metadata *sys_data;
313 struct ring_buffer_event *event;
314 struct ring_buffer *buffer;
315 unsigned long irq_flags;
316 int pc;
317 int syscall_nr;
318 int size;
319
320 syscall_nr = trace_get_syscall_nr(current, regs);
321 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
322 return;
323
324 /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE) */
325 trace_file = rcu_dereference_sched(tr->enter_syscall_files[syscall_nr]);
326 if (!trace_file)
327 return;
328
329 if (trace_trigger_soft_disabled(trace_file))
330 return;
331
332 sys_data = syscall_nr_to_meta(syscall_nr);
333 if (!sys_data)
334 return;
335
336 size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
337
338 local_save_flags(irq_flags);
339 pc = preempt_count();
340
341 buffer = tr->trace_buffer.buffer;
342 event = trace_buffer_lock_reserve(buffer,
343 sys_data->enter_event->event.type, size, irq_flags, pc);
344 if (!event)
345 return;
346
347 entry = ring_buffer_event_data(event);
348 entry->nr = syscall_nr;
349 syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
350
351 event_trigger_unlock_commit(trace_file, buffer, event, entry,
352 irq_flags, pc);
353}
354
355static void ftrace_syscall_exit(void *data, struct pt_regs *regs, long ret)
356{
357 struct trace_array *tr = data;
358 struct trace_event_file *trace_file;
359 struct syscall_trace_exit *entry;
360 struct syscall_metadata *sys_data;
361 struct ring_buffer_event *event;
362 struct ring_buffer *buffer;
363 unsigned long irq_flags;
364 int pc;
365 int syscall_nr;
366
367 syscall_nr = trace_get_syscall_nr(current, regs);
368 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
369 return;
370
371 /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE()) */
372 trace_file = rcu_dereference_sched(tr->exit_syscall_files[syscall_nr]);
373 if (!trace_file)
374 return;
375
376 if (trace_trigger_soft_disabled(trace_file))
377 return;
378
379 sys_data = syscall_nr_to_meta(syscall_nr);
380 if (!sys_data)
381 return;
382
383 local_save_flags(irq_flags);
384 pc = preempt_count();
385
386 buffer = tr->trace_buffer.buffer;
387 event = trace_buffer_lock_reserve(buffer,
388 sys_data->exit_event->event.type, sizeof(*entry),
389 irq_flags, pc);
390 if (!event)
391 return;
392
393 entry = ring_buffer_event_data(event);
394 entry->nr = syscall_nr;
395 entry->ret = syscall_get_return_value(current, regs);
396
397 event_trigger_unlock_commit(trace_file, buffer, event, entry,
398 irq_flags, pc);
399}
400
401static int reg_event_syscall_enter(struct trace_event_file *file,
402 struct trace_event_call *call)
403{
404 struct trace_array *tr = file->tr;
405 int ret = 0;
406 int num;
407
408 num = ((struct syscall_metadata *)call->data)->syscall_nr;
409 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
410 return -ENOSYS;
411 mutex_lock(&syscall_trace_lock);
412 if (!tr->sys_refcount_enter)
413 ret = register_trace_sys_enter(ftrace_syscall_enter, tr);
414 if (!ret) {
415 rcu_assign_pointer(tr->enter_syscall_files[num], file);
416 tr->sys_refcount_enter++;
417 }
418 mutex_unlock(&syscall_trace_lock);
419 return ret;
420}
421
422static void unreg_event_syscall_enter(struct trace_event_file *file,
423 struct trace_event_call *call)
424{
425 struct trace_array *tr = file->tr;
426 int num;
427
428 num = ((struct syscall_metadata *)call->data)->syscall_nr;
429 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
430 return;
431 mutex_lock(&syscall_trace_lock);
432 tr->sys_refcount_enter--;
433 RCU_INIT_POINTER(tr->enter_syscall_files[num], NULL);
434 if (!tr->sys_refcount_enter)
435 unregister_trace_sys_enter(ftrace_syscall_enter, tr);
436 mutex_unlock(&syscall_trace_lock);
437}
438
439static int reg_event_syscall_exit(struct trace_event_file *file,
440 struct trace_event_call *call)
441{
442 struct trace_array *tr = file->tr;
443 int ret = 0;
444 int num;
445
446 num = ((struct syscall_metadata *)call->data)->syscall_nr;
447 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
448 return -ENOSYS;
449 mutex_lock(&syscall_trace_lock);
450 if (!tr->sys_refcount_exit)
451 ret = register_trace_sys_exit(ftrace_syscall_exit, tr);
452 if (!ret) {
453 rcu_assign_pointer(tr->exit_syscall_files[num], file);
454 tr->sys_refcount_exit++;
455 }
456 mutex_unlock(&syscall_trace_lock);
457 return ret;
458}
459
460static void unreg_event_syscall_exit(struct trace_event_file *file,
461 struct trace_event_call *call)
462{
463 struct trace_array *tr = file->tr;
464 int num;
465
466 num = ((struct syscall_metadata *)call->data)->syscall_nr;
467 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
468 return;
469 mutex_lock(&syscall_trace_lock);
470 tr->sys_refcount_exit--;
471 RCU_INIT_POINTER(tr->exit_syscall_files[num], NULL);
472 if (!tr->sys_refcount_exit)
473 unregister_trace_sys_exit(ftrace_syscall_exit, tr);
474 mutex_unlock(&syscall_trace_lock);
475}
476
477static int __init init_syscall_trace(struct trace_event_call *call)
478{
479 int id;
480 int num;
481
482 num = ((struct syscall_metadata *)call->data)->syscall_nr;
483 if (num < 0 || num >= NR_syscalls) {
484 pr_debug("syscall %s metadata not mapped, disabling ftrace event\n",
485 ((struct syscall_metadata *)call->data)->name);
486 return -ENOSYS;
487 }
488
489 if (set_syscall_print_fmt(call) < 0)
490 return -ENOMEM;
491
492 id = trace_event_raw_init(call);
493
494 if (id < 0) {
495 free_syscall_print_fmt(call);
496 return id;
497 }
498
499 return id;
500}
501
502struct trace_event_functions enter_syscall_print_funcs = {
503 .trace = print_syscall_enter,
504};
505
506struct trace_event_functions exit_syscall_print_funcs = {
507 .trace = print_syscall_exit,
508};
509
510struct trace_event_class __refdata event_class_syscall_enter = {
511 .system = "syscalls",
512 .reg = syscall_enter_register,
513 .define_fields = syscall_enter_define_fields,
514 .get_fields = syscall_get_enter_fields,
515 .raw_init = init_syscall_trace,
516};
517
518struct trace_event_class __refdata event_class_syscall_exit = {
519 .system = "syscalls",
520 .reg = syscall_exit_register,
521 .define_fields = syscall_exit_define_fields,
522 .fields = LIST_HEAD_INIT(event_class_syscall_exit.fields),
523 .raw_init = init_syscall_trace,
524};
525
526unsigned long __init __weak arch_syscall_addr(int nr)
527{
528 return (unsigned long)sys_call_table[nr];
529}
530
531void __init init_ftrace_syscalls(void)
532{
533 struct syscall_metadata *meta;
534 unsigned long addr;
535 int i;
536
537 syscalls_metadata = kcalloc(NR_syscalls, sizeof(*syscalls_metadata),
538 GFP_KERNEL);
539 if (!syscalls_metadata) {
540 WARN_ON(1);
541 return;
542 }
543
544 for (i = 0; i < NR_syscalls; i++) {
545 addr = arch_syscall_addr(i);
546 meta = find_syscall_meta(addr);
547 if (!meta)
548 continue;
549
550 meta->syscall_nr = i;
551 syscalls_metadata[i] = meta;
552 }
553}
554
555#ifdef CONFIG_PERF_EVENTS
556
557static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
558static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
559static int sys_perf_refcount_enter;
560static int sys_perf_refcount_exit;
561
562static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
563{
564 struct syscall_metadata *sys_data;
565 struct syscall_trace_enter *rec;
566 struct hlist_head *head;
567 int syscall_nr;
568 int rctx;
569 int size;
570
571 syscall_nr = trace_get_syscall_nr(current, regs);
572 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
573 return;
574 if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
575 return;
576
577 sys_data = syscall_nr_to_meta(syscall_nr);
578 if (!sys_data)
579 return;
580
581 head = this_cpu_ptr(sys_data->enter_event->perf_events);
582 if (hlist_empty(head))
583 return;
584
585 /* get the size after alignment with the u32 buffer size field */
586 size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
587 size = ALIGN(size + sizeof(u32), sizeof(u64));
588 size -= sizeof(u32);
589
590 rec = (struct syscall_trace_enter *)perf_trace_buf_prepare(size,
591 sys_data->enter_event->event.type, NULL, &rctx);
592 if (!rec)
593 return;
594
595 rec->nr = syscall_nr;
596 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
597 (unsigned long *)&rec->args);
598 perf_trace_buf_submit(rec, size, rctx, 0, 1, regs, head, NULL);
599}
600
601static int perf_sysenter_enable(struct trace_event_call *call)
602{
603 int ret = 0;
604 int num;
605
606 num = ((struct syscall_metadata *)call->data)->syscall_nr;
607
608 mutex_lock(&syscall_trace_lock);
609 if (!sys_perf_refcount_enter)
610 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
611 if (ret) {
612 pr_info("event trace: Could not activate"
613 "syscall entry trace point");
614 } else {
615 set_bit(num, enabled_perf_enter_syscalls);
616 sys_perf_refcount_enter++;
617 }
618 mutex_unlock(&syscall_trace_lock);
619 return ret;
620}
621
622static void perf_sysenter_disable(struct trace_event_call *call)
623{
624 int num;
625
626 num = ((struct syscall_metadata *)call->data)->syscall_nr;
627
628 mutex_lock(&syscall_trace_lock);
629 sys_perf_refcount_enter--;
630 clear_bit(num, enabled_perf_enter_syscalls);
631 if (!sys_perf_refcount_enter)
632 unregister_trace_sys_enter(perf_syscall_enter, NULL);
633 mutex_unlock(&syscall_trace_lock);
634}
635
636static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
637{
638 struct syscall_metadata *sys_data;
639 struct syscall_trace_exit *rec;
640 struct hlist_head *head;
641 int syscall_nr;
642 int rctx;
643 int size;
644
645 syscall_nr = trace_get_syscall_nr(current, regs);
646 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
647 return;
648 if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
649 return;
650
651 sys_data = syscall_nr_to_meta(syscall_nr);
652 if (!sys_data)
653 return;
654
655 head = this_cpu_ptr(sys_data->exit_event->perf_events);
656 if (hlist_empty(head))
657 return;
658
659 /* We can probably do that at build time */
660 size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
661 size -= sizeof(u32);
662
663 rec = (struct syscall_trace_exit *)perf_trace_buf_prepare(size,
664 sys_data->exit_event->event.type, NULL, &rctx);
665 if (!rec)
666 return;
667
668 rec->nr = syscall_nr;
669 rec->ret = syscall_get_return_value(current, regs);
670 perf_trace_buf_submit(rec, size, rctx, 0, 1, regs, head, NULL);
671}
672
673static int perf_sysexit_enable(struct trace_event_call *call)
674{
675 int ret = 0;
676 int num;
677
678 num = ((struct syscall_metadata *)call->data)->syscall_nr;
679
680 mutex_lock(&syscall_trace_lock);
681 if (!sys_perf_refcount_exit)
682 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
683 if (ret) {
684 pr_info("event trace: Could not activate"
685 "syscall exit trace point");
686 } else {
687 set_bit(num, enabled_perf_exit_syscalls);
688 sys_perf_refcount_exit++;
689 }
690 mutex_unlock(&syscall_trace_lock);
691 return ret;
692}
693
694static void perf_sysexit_disable(struct trace_event_call *call)
695{
696 int num;
697
698 num = ((struct syscall_metadata *)call->data)->syscall_nr;
699
700 mutex_lock(&syscall_trace_lock);
701 sys_perf_refcount_exit--;
702 clear_bit(num, enabled_perf_exit_syscalls);
703 if (!sys_perf_refcount_exit)
704 unregister_trace_sys_exit(perf_syscall_exit, NULL);
705 mutex_unlock(&syscall_trace_lock);
706}
707
708#endif /* CONFIG_PERF_EVENTS */
709
710static int syscall_enter_register(struct trace_event_call *event,
711 enum trace_reg type, void *data)
712{
713 struct trace_event_file *file = data;
714
715 switch (type) {
716 case TRACE_REG_REGISTER:
717 return reg_event_syscall_enter(file, event);
718 case TRACE_REG_UNREGISTER:
719 unreg_event_syscall_enter(file, event);
720 return 0;
721
722#ifdef CONFIG_PERF_EVENTS
723 case TRACE_REG_PERF_REGISTER:
724 return perf_sysenter_enable(event);
725 case TRACE_REG_PERF_UNREGISTER:
726 perf_sysenter_disable(event);
727 return 0;
728 case TRACE_REG_PERF_OPEN:
729 case TRACE_REG_PERF_CLOSE:
730 case TRACE_REG_PERF_ADD:
731 case TRACE_REG_PERF_DEL:
732 return 0;
733#endif
734 }
735 return 0;
736}
737
738static int syscall_exit_register(struct trace_event_call *event,
739 enum trace_reg type, void *data)
740{
741 struct trace_event_file *file = data;
742
743 switch (type) {
744 case TRACE_REG_REGISTER:
745 return reg_event_syscall_exit(file, event);
746 case TRACE_REG_UNREGISTER:
747 unreg_event_syscall_exit(file, event);
748 return 0;
749
750#ifdef CONFIG_PERF_EVENTS
751 case TRACE_REG_PERF_REGISTER:
752 return perf_sysexit_enable(event);
753 case TRACE_REG_PERF_UNREGISTER:
754 perf_sysexit_disable(event);
755 return 0;
756 case TRACE_REG_PERF_OPEN:
757 case TRACE_REG_PERF_CLOSE:
758 case TRACE_REG_PERF_ADD:
759 case TRACE_REG_PERF_DEL:
760 return 0;
761#endif
762 }
763 return 0;
764}
1#include <trace/syscall.h>
2#include <trace/events/syscalls.h>
3#include <linux/slab.h>
4#include <linux/kernel.h>
5#include <linux/ftrace.h>
6#include <linux/perf_event.h>
7#include <asm/syscall.h>
8
9#include "trace_output.h"
10#include "trace.h"
11
12static DEFINE_MUTEX(syscall_trace_lock);
13static int sys_refcount_enter;
14static int sys_refcount_exit;
15static DECLARE_BITMAP(enabled_enter_syscalls, NR_syscalls);
16static DECLARE_BITMAP(enabled_exit_syscalls, NR_syscalls);
17
18static int syscall_enter_register(struct ftrace_event_call *event,
19 enum trace_reg type);
20static int syscall_exit_register(struct ftrace_event_call *event,
21 enum trace_reg type);
22
23static int syscall_enter_define_fields(struct ftrace_event_call *call);
24static int syscall_exit_define_fields(struct ftrace_event_call *call);
25
26static struct list_head *
27syscall_get_enter_fields(struct ftrace_event_call *call)
28{
29 struct syscall_metadata *entry = call->data;
30
31 return &entry->enter_fields;
32}
33
34struct trace_event_functions enter_syscall_print_funcs = {
35 .trace = print_syscall_enter,
36};
37
38struct trace_event_functions exit_syscall_print_funcs = {
39 .trace = print_syscall_exit,
40};
41
42struct ftrace_event_class event_class_syscall_enter = {
43 .system = "syscalls",
44 .reg = syscall_enter_register,
45 .define_fields = syscall_enter_define_fields,
46 .get_fields = syscall_get_enter_fields,
47 .raw_init = init_syscall_trace,
48};
49
50struct ftrace_event_class event_class_syscall_exit = {
51 .system = "syscalls",
52 .reg = syscall_exit_register,
53 .define_fields = syscall_exit_define_fields,
54 .fields = LIST_HEAD_INIT(event_class_syscall_exit.fields),
55 .raw_init = init_syscall_trace,
56};
57
58extern struct syscall_metadata *__start_syscalls_metadata[];
59extern struct syscall_metadata *__stop_syscalls_metadata[];
60
61static struct syscall_metadata **syscalls_metadata;
62
63#ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
64static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
65{
66 /*
67 * Only compare after the "sys" prefix. Archs that use
68 * syscall wrappers may have syscalls symbols aliases prefixed
69 * with "SyS" instead of "sys", leading to an unwanted
70 * mismatch.
71 */
72 return !strcmp(sym + 3, name + 3);
73}
74#endif
75
76static __init struct syscall_metadata *
77find_syscall_meta(unsigned long syscall)
78{
79 struct syscall_metadata **start;
80 struct syscall_metadata **stop;
81 char str[KSYM_SYMBOL_LEN];
82
83
84 start = __start_syscalls_metadata;
85 stop = __stop_syscalls_metadata;
86 kallsyms_lookup(syscall, NULL, NULL, NULL, str);
87
88 if (arch_syscall_match_sym_name(str, "sys_ni_syscall"))
89 return NULL;
90
91 for ( ; start < stop; start++) {
92 if ((*start)->name && arch_syscall_match_sym_name(str, (*start)->name))
93 return *start;
94 }
95 return NULL;
96}
97
98static struct syscall_metadata *syscall_nr_to_meta(int nr)
99{
100 if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
101 return NULL;
102
103 return syscalls_metadata[nr];
104}
105
106enum print_line_t
107print_syscall_enter(struct trace_iterator *iter, int flags,
108 struct trace_event *event)
109{
110 struct trace_seq *s = &iter->seq;
111 struct trace_entry *ent = iter->ent;
112 struct syscall_trace_enter *trace;
113 struct syscall_metadata *entry;
114 int i, ret, syscall;
115
116 trace = (typeof(trace))ent;
117 syscall = trace->nr;
118 entry = syscall_nr_to_meta(syscall);
119
120 if (!entry)
121 goto end;
122
123 if (entry->enter_event->event.type != ent->type) {
124 WARN_ON_ONCE(1);
125 goto end;
126 }
127
128 ret = trace_seq_printf(s, "%s(", entry->name);
129 if (!ret)
130 return TRACE_TYPE_PARTIAL_LINE;
131
132 for (i = 0; i < entry->nb_args; i++) {
133 /* parameter types */
134 if (trace_flags & TRACE_ITER_VERBOSE) {
135 ret = trace_seq_printf(s, "%s ", entry->types[i]);
136 if (!ret)
137 return TRACE_TYPE_PARTIAL_LINE;
138 }
139 /* parameter values */
140 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
141 trace->args[i],
142 i == entry->nb_args - 1 ? "" : ", ");
143 if (!ret)
144 return TRACE_TYPE_PARTIAL_LINE;
145 }
146
147 ret = trace_seq_putc(s, ')');
148 if (!ret)
149 return TRACE_TYPE_PARTIAL_LINE;
150
151end:
152 ret = trace_seq_putc(s, '\n');
153 if (!ret)
154 return TRACE_TYPE_PARTIAL_LINE;
155
156 return TRACE_TYPE_HANDLED;
157}
158
159enum print_line_t
160print_syscall_exit(struct trace_iterator *iter, int flags,
161 struct trace_event *event)
162{
163 struct trace_seq *s = &iter->seq;
164 struct trace_entry *ent = iter->ent;
165 struct syscall_trace_exit *trace;
166 int syscall;
167 struct syscall_metadata *entry;
168 int ret;
169
170 trace = (typeof(trace))ent;
171 syscall = trace->nr;
172 entry = syscall_nr_to_meta(syscall);
173
174 if (!entry) {
175 trace_seq_printf(s, "\n");
176 return TRACE_TYPE_HANDLED;
177 }
178
179 if (entry->exit_event->event.type != ent->type) {
180 WARN_ON_ONCE(1);
181 return TRACE_TYPE_UNHANDLED;
182 }
183
184 ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
185 trace->ret);
186 if (!ret)
187 return TRACE_TYPE_PARTIAL_LINE;
188
189 return TRACE_TYPE_HANDLED;
190}
191
192extern char *__bad_type_size(void);
193
194#define SYSCALL_FIELD(type, name) \
195 sizeof(type) != sizeof(trace.name) ? \
196 __bad_type_size() : \
197 #type, #name, offsetof(typeof(trace), name), \
198 sizeof(trace.name), is_signed_type(type)
199
200static
201int __set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
202{
203 int i;
204 int pos = 0;
205
206 /* When len=0, we just calculate the needed length */
207#define LEN_OR_ZERO (len ? len - pos : 0)
208
209 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
210 for (i = 0; i < entry->nb_args; i++) {
211 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
212 entry->args[i], sizeof(unsigned long),
213 i == entry->nb_args - 1 ? "" : ", ");
214 }
215 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
216
217 for (i = 0; i < entry->nb_args; i++) {
218 pos += snprintf(buf + pos, LEN_OR_ZERO,
219 ", ((unsigned long)(REC->%s))", entry->args[i]);
220 }
221
222#undef LEN_OR_ZERO
223
224 /* return the length of print_fmt */
225 return pos;
226}
227
228static int set_syscall_print_fmt(struct ftrace_event_call *call)
229{
230 char *print_fmt;
231 int len;
232 struct syscall_metadata *entry = call->data;
233
234 if (entry->enter_event != call) {
235 call->print_fmt = "\"0x%lx\", REC->ret";
236 return 0;
237 }
238
239 /* First: called with 0 length to calculate the needed length */
240 len = __set_enter_print_fmt(entry, NULL, 0);
241
242 print_fmt = kmalloc(len + 1, GFP_KERNEL);
243 if (!print_fmt)
244 return -ENOMEM;
245
246 /* Second: actually write the @print_fmt */
247 __set_enter_print_fmt(entry, print_fmt, len + 1);
248 call->print_fmt = print_fmt;
249
250 return 0;
251}
252
253static void free_syscall_print_fmt(struct ftrace_event_call *call)
254{
255 struct syscall_metadata *entry = call->data;
256
257 if (entry->enter_event == call)
258 kfree(call->print_fmt);
259}
260
261static int syscall_enter_define_fields(struct ftrace_event_call *call)
262{
263 struct syscall_trace_enter trace;
264 struct syscall_metadata *meta = call->data;
265 int ret;
266 int i;
267 int offset = offsetof(typeof(trace), args);
268
269 ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
270 if (ret)
271 return ret;
272
273 for (i = 0; i < meta->nb_args; i++) {
274 ret = trace_define_field(call, meta->types[i],
275 meta->args[i], offset,
276 sizeof(unsigned long), 0,
277 FILTER_OTHER);
278 offset += sizeof(unsigned long);
279 }
280
281 return ret;
282}
283
284static int syscall_exit_define_fields(struct ftrace_event_call *call)
285{
286 struct syscall_trace_exit trace;
287 int ret;
288
289 ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
290 if (ret)
291 return ret;
292
293 ret = trace_define_field(call, SYSCALL_FIELD(long, ret),
294 FILTER_OTHER);
295
296 return ret;
297}
298
299void ftrace_syscall_enter(void *ignore, struct pt_regs *regs, long id)
300{
301 struct syscall_trace_enter *entry;
302 struct syscall_metadata *sys_data;
303 struct ring_buffer_event *event;
304 struct ring_buffer *buffer;
305 int size;
306 int syscall_nr;
307
308 syscall_nr = syscall_get_nr(current, regs);
309 if (syscall_nr < 0)
310 return;
311 if (!test_bit(syscall_nr, enabled_enter_syscalls))
312 return;
313
314 sys_data = syscall_nr_to_meta(syscall_nr);
315 if (!sys_data)
316 return;
317
318 size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
319
320 event = trace_current_buffer_lock_reserve(&buffer,
321 sys_data->enter_event->event.type, size, 0, 0);
322 if (!event)
323 return;
324
325 entry = ring_buffer_event_data(event);
326 entry->nr = syscall_nr;
327 syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
328
329 if (!filter_current_check_discard(buffer, sys_data->enter_event,
330 entry, event))
331 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
332}
333
334void ftrace_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
335{
336 struct syscall_trace_exit *entry;
337 struct syscall_metadata *sys_data;
338 struct ring_buffer_event *event;
339 struct ring_buffer *buffer;
340 int syscall_nr;
341
342 syscall_nr = syscall_get_nr(current, regs);
343 if (syscall_nr < 0)
344 return;
345 if (!test_bit(syscall_nr, enabled_exit_syscalls))
346 return;
347
348 sys_data = syscall_nr_to_meta(syscall_nr);
349 if (!sys_data)
350 return;
351
352 event = trace_current_buffer_lock_reserve(&buffer,
353 sys_data->exit_event->event.type, sizeof(*entry), 0, 0);
354 if (!event)
355 return;
356
357 entry = ring_buffer_event_data(event);
358 entry->nr = syscall_nr;
359 entry->ret = syscall_get_return_value(current, regs);
360
361 if (!filter_current_check_discard(buffer, sys_data->exit_event,
362 entry, event))
363 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
364}
365
366int reg_event_syscall_enter(struct ftrace_event_call *call)
367{
368 int ret = 0;
369 int num;
370
371 num = ((struct syscall_metadata *)call->data)->syscall_nr;
372 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
373 return -ENOSYS;
374 mutex_lock(&syscall_trace_lock);
375 if (!sys_refcount_enter)
376 ret = register_trace_sys_enter(ftrace_syscall_enter, NULL);
377 if (!ret) {
378 set_bit(num, enabled_enter_syscalls);
379 sys_refcount_enter++;
380 }
381 mutex_unlock(&syscall_trace_lock);
382 return ret;
383}
384
385void unreg_event_syscall_enter(struct ftrace_event_call *call)
386{
387 int num;
388
389 num = ((struct syscall_metadata *)call->data)->syscall_nr;
390 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
391 return;
392 mutex_lock(&syscall_trace_lock);
393 sys_refcount_enter--;
394 clear_bit(num, enabled_enter_syscalls);
395 if (!sys_refcount_enter)
396 unregister_trace_sys_enter(ftrace_syscall_enter, NULL);
397 mutex_unlock(&syscall_trace_lock);
398}
399
400int reg_event_syscall_exit(struct ftrace_event_call *call)
401{
402 int ret = 0;
403 int num;
404
405 num = ((struct syscall_metadata *)call->data)->syscall_nr;
406 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
407 return -ENOSYS;
408 mutex_lock(&syscall_trace_lock);
409 if (!sys_refcount_exit)
410 ret = register_trace_sys_exit(ftrace_syscall_exit, NULL);
411 if (!ret) {
412 set_bit(num, enabled_exit_syscalls);
413 sys_refcount_exit++;
414 }
415 mutex_unlock(&syscall_trace_lock);
416 return ret;
417}
418
419void unreg_event_syscall_exit(struct ftrace_event_call *call)
420{
421 int num;
422
423 num = ((struct syscall_metadata *)call->data)->syscall_nr;
424 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
425 return;
426 mutex_lock(&syscall_trace_lock);
427 sys_refcount_exit--;
428 clear_bit(num, enabled_exit_syscalls);
429 if (!sys_refcount_exit)
430 unregister_trace_sys_exit(ftrace_syscall_exit, NULL);
431 mutex_unlock(&syscall_trace_lock);
432}
433
434int init_syscall_trace(struct ftrace_event_call *call)
435{
436 int id;
437 int num;
438
439 num = ((struct syscall_metadata *)call->data)->syscall_nr;
440 if (num < 0 || num >= NR_syscalls) {
441 pr_debug("syscall %s metadata not mapped, disabling ftrace event\n",
442 ((struct syscall_metadata *)call->data)->name);
443 return -ENOSYS;
444 }
445
446 if (set_syscall_print_fmt(call) < 0)
447 return -ENOMEM;
448
449 id = trace_event_raw_init(call);
450
451 if (id < 0) {
452 free_syscall_print_fmt(call);
453 return id;
454 }
455
456 return id;
457}
458
459unsigned long __init __weak arch_syscall_addr(int nr)
460{
461 return (unsigned long)sys_call_table[nr];
462}
463
464int __init init_ftrace_syscalls(void)
465{
466 struct syscall_metadata *meta;
467 unsigned long addr;
468 int i;
469
470 syscalls_metadata = kzalloc(sizeof(*syscalls_metadata) *
471 NR_syscalls, GFP_KERNEL);
472 if (!syscalls_metadata) {
473 WARN_ON(1);
474 return -ENOMEM;
475 }
476
477 for (i = 0; i < NR_syscalls; i++) {
478 addr = arch_syscall_addr(i);
479 meta = find_syscall_meta(addr);
480 if (!meta)
481 continue;
482
483 meta->syscall_nr = i;
484 syscalls_metadata[i] = meta;
485 }
486
487 return 0;
488}
489core_initcall(init_ftrace_syscalls);
490
491#ifdef CONFIG_PERF_EVENTS
492
493static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
494static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
495static int sys_perf_refcount_enter;
496static int sys_perf_refcount_exit;
497
498static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
499{
500 struct syscall_metadata *sys_data;
501 struct syscall_trace_enter *rec;
502 struct hlist_head *head;
503 int syscall_nr;
504 int rctx;
505 int size;
506
507 syscall_nr = syscall_get_nr(current, regs);
508 if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
509 return;
510
511 sys_data = syscall_nr_to_meta(syscall_nr);
512 if (!sys_data)
513 return;
514
515 /* get the size after alignment with the u32 buffer size field */
516 size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
517 size = ALIGN(size + sizeof(u32), sizeof(u64));
518 size -= sizeof(u32);
519
520 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
521 "perf buffer not large enough"))
522 return;
523
524 rec = (struct syscall_trace_enter *)perf_trace_buf_prepare(size,
525 sys_data->enter_event->event.type, regs, &rctx);
526 if (!rec)
527 return;
528
529 rec->nr = syscall_nr;
530 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
531 (unsigned long *)&rec->args);
532
533 head = this_cpu_ptr(sys_data->enter_event->perf_events);
534 perf_trace_buf_submit(rec, size, rctx, 0, 1, regs, head);
535}
536
537int perf_sysenter_enable(struct ftrace_event_call *call)
538{
539 int ret = 0;
540 int num;
541
542 num = ((struct syscall_metadata *)call->data)->syscall_nr;
543
544 mutex_lock(&syscall_trace_lock);
545 if (!sys_perf_refcount_enter)
546 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
547 if (ret) {
548 pr_info("event trace: Could not activate"
549 "syscall entry trace point");
550 } else {
551 set_bit(num, enabled_perf_enter_syscalls);
552 sys_perf_refcount_enter++;
553 }
554 mutex_unlock(&syscall_trace_lock);
555 return ret;
556}
557
558void perf_sysenter_disable(struct ftrace_event_call *call)
559{
560 int num;
561
562 num = ((struct syscall_metadata *)call->data)->syscall_nr;
563
564 mutex_lock(&syscall_trace_lock);
565 sys_perf_refcount_enter--;
566 clear_bit(num, enabled_perf_enter_syscalls);
567 if (!sys_perf_refcount_enter)
568 unregister_trace_sys_enter(perf_syscall_enter, NULL);
569 mutex_unlock(&syscall_trace_lock);
570}
571
572static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
573{
574 struct syscall_metadata *sys_data;
575 struct syscall_trace_exit *rec;
576 struct hlist_head *head;
577 int syscall_nr;
578 int rctx;
579 int size;
580
581 syscall_nr = syscall_get_nr(current, regs);
582 if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
583 return;
584
585 sys_data = syscall_nr_to_meta(syscall_nr);
586 if (!sys_data)
587 return;
588
589 /* We can probably do that at build time */
590 size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
591 size -= sizeof(u32);
592
593 /*
594 * Impossible, but be paranoid with the future
595 * How to put this check outside runtime?
596 */
597 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
598 "exit event has grown above perf buffer size"))
599 return;
600
601 rec = (struct syscall_trace_exit *)perf_trace_buf_prepare(size,
602 sys_data->exit_event->event.type, regs, &rctx);
603 if (!rec)
604 return;
605
606 rec->nr = syscall_nr;
607 rec->ret = syscall_get_return_value(current, regs);
608
609 head = this_cpu_ptr(sys_data->exit_event->perf_events);
610 perf_trace_buf_submit(rec, size, rctx, 0, 1, regs, head);
611}
612
613int perf_sysexit_enable(struct ftrace_event_call *call)
614{
615 int ret = 0;
616 int num;
617
618 num = ((struct syscall_metadata *)call->data)->syscall_nr;
619
620 mutex_lock(&syscall_trace_lock);
621 if (!sys_perf_refcount_exit)
622 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
623 if (ret) {
624 pr_info("event trace: Could not activate"
625 "syscall exit trace point");
626 } else {
627 set_bit(num, enabled_perf_exit_syscalls);
628 sys_perf_refcount_exit++;
629 }
630 mutex_unlock(&syscall_trace_lock);
631 return ret;
632}
633
634void perf_sysexit_disable(struct ftrace_event_call *call)
635{
636 int num;
637
638 num = ((struct syscall_metadata *)call->data)->syscall_nr;
639
640 mutex_lock(&syscall_trace_lock);
641 sys_perf_refcount_exit--;
642 clear_bit(num, enabled_perf_exit_syscalls);
643 if (!sys_perf_refcount_exit)
644 unregister_trace_sys_exit(perf_syscall_exit, NULL);
645 mutex_unlock(&syscall_trace_lock);
646}
647
648#endif /* CONFIG_PERF_EVENTS */
649
650static int syscall_enter_register(struct ftrace_event_call *event,
651 enum trace_reg type)
652{
653 switch (type) {
654 case TRACE_REG_REGISTER:
655 return reg_event_syscall_enter(event);
656 case TRACE_REG_UNREGISTER:
657 unreg_event_syscall_enter(event);
658 return 0;
659
660#ifdef CONFIG_PERF_EVENTS
661 case TRACE_REG_PERF_REGISTER:
662 return perf_sysenter_enable(event);
663 case TRACE_REG_PERF_UNREGISTER:
664 perf_sysenter_disable(event);
665 return 0;
666#endif
667 }
668 return 0;
669}
670
671static int syscall_exit_register(struct ftrace_event_call *event,
672 enum trace_reg type)
673{
674 switch (type) {
675 case TRACE_REG_REGISTER:
676 return reg_event_syscall_exit(event);
677 case TRACE_REG_UNREGISTER:
678 unreg_event_syscall_exit(event);
679 return 0;
680
681#ifdef CONFIG_PERF_EVENTS
682 case TRACE_REG_PERF_REGISTER:
683 return perf_sysexit_enable(event);
684 case TRACE_REG_PERF_UNREGISTER:
685 perf_sysexit_disable(event);
686 return 0;
687#endif
688 }
689 return 0;
690}