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