projects/linux·blob master · refs

541 lines · 12.6 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * trace event based perf event profiling/tracing
4 *
5 * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra
6 * Copyright (C) 2009-2010 Frederic Weisbecker <[email protected]>
7 */
8
9#include <linux/module.h>
10#include <linux/kprobes.h>
11#include <linux/security.h>
12#include "trace.h"
13#include "trace_probe.h"
14
15static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
16
17/*
18 * Force it to be aligned to unsigned long to avoid misaligned accesses
19 * surprises
20 */
21typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
22 perf_trace_t;
23
24/* Count the events in use (per event id, not per instance) */
25static int total_ref_count;
26
27static int perf_trace_event_perm(struct trace_event_call *tp_event,
28 struct perf_event *p_event)
29{
30 int ret;
31
32 if (tp_event->perf_perm) {
33 ret = tp_event->perf_perm(tp_event, p_event);
34 if (ret)
35 return ret;
36 }
37
38 /*
39 * We checked and allowed to create parent,
40 * allow children without checking.
41 */
42 if (p_event->parent)
43 return 0;
44
45 /*
46 * It's ok to check current process (owner) permissions in here,
47 * because code below is called only via perf_event_open syscall.
48 */
49
50 /* The ftrace function trace is allowed only for root. */
51 if (ftrace_event_is_function(tp_event)) {
52 ret = perf_allow_tracepoint();
53 if (ret)
54 return ret;
55
56 if (!is_sampling_event(p_event))
57 return 0;
58
59 /*
60 * We don't allow user space callchains for function trace
61 * event, due to issues with page faults while tracing page
62 * fault handler and its overall trickiness nature.
63 */
64 if (!p_event->attr.exclude_callchain_user)
65 return -EINVAL;
66
67 /*
68 * Same reason to disable user stack dump as for user space
69 * callchains above.
70 */
71 if (p_event->attr.sample_type & PERF_SAMPLE_STACK_USER)
72 return -EINVAL;
73 }
74
75 /* No tracing, just counting, so no obvious leak */
76 if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
77 return 0;
78
79 /* Some events are ok to be traced by non-root users... */
80 if (p_event->attach_state == PERF_ATTACH_TASK) {
81 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
82 return 0;
83 }
84
85 /*
86 * ...otherwise raw tracepoint data can be a severe data leak,
87 * only allow root to have these.
88 */
89 ret = perf_allow_tracepoint();
90 if (ret)
91 return ret;
92
93 return 0;
94}
95
96static int perf_trace_event_reg(struct trace_event_call *tp_event,
97 struct perf_event *p_event)
98{
99 struct hlist_head __percpu *list;
100 int ret = -ENOMEM;
101 int cpu;
102
103 p_event->tp_event = tp_event;
104 if (tp_event->perf_refcount++ > 0)
105 return 0;
106
107 list = alloc_percpu(struct hlist_head);
108 if (!list)
109 goto fail;
110
111 for_each_possible_cpu(cpu)
112 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
113
114 tp_event->perf_events = list;
115
116 if (!total_ref_count) {
117 char __percpu *buf;
118 int i;
119
120 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
121 buf = (char __percpu *)alloc_percpu(perf_trace_t);
122 if (!buf)
123 goto fail;
124
125 perf_trace_buf[i] = buf;
126 }
127 }
128
129 ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
130 if (ret)
131 goto fail;
132
133 total_ref_count++;
134 return 0;
135
136fail:
137 if (!total_ref_count) {
138 int i;
139
140 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
141 free_percpu(perf_trace_buf[i]);
142 perf_trace_buf[i] = NULL;
143 }
144 }
145
146 if (!--tp_event->perf_refcount) {
147 free_percpu(tp_event->perf_events);
148 tp_event->perf_events = NULL;
149 }
150
151 return ret;
152}
153
154static void perf_trace_event_unreg(struct perf_event *p_event)
155{
156 struct trace_event_call *tp_event = p_event->tp_event;
157 int i;
158
159 if (--tp_event->perf_refcount > 0)
160 return;
161
162 tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
163
164 /*
165 * Ensure our callback won't be called anymore. The buffers
166 * will be freed after that.
167 */
168 tracepoint_synchronize_unregister();
169
170 free_percpu(tp_event->perf_events);
171 tp_event->perf_events = NULL;
172
173 if (!--total_ref_count) {
174 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
175 free_percpu(perf_trace_buf[i]);
176 perf_trace_buf[i] = NULL;
177 }
178 }
179}
180
181static int perf_trace_event_open(struct perf_event *p_event)
182{
183 struct trace_event_call *tp_event = p_event->tp_event;
184 return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event);
185}
186
187static void perf_trace_event_close(struct perf_event *p_event)
188{
189 struct trace_event_call *tp_event = p_event->tp_event;
190 tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
191}
192
193static int perf_trace_event_init(struct trace_event_call *tp_event,
194 struct perf_event *p_event)
195{
196 int ret;
197
198 ret = perf_trace_event_perm(tp_event, p_event);
199 if (ret)
200 return ret;
201
202 ret = perf_trace_event_reg(tp_event, p_event);
203 if (ret)
204 return ret;
205
206 ret = perf_trace_event_open(p_event);
207 if (ret) {
208 perf_trace_event_unreg(p_event);
209 return ret;
210 }
211
212 return 0;
213}
214
215int perf_trace_init(struct perf_event *p_event)
216{
217 struct trace_event_call *tp_event;
218 u64 event_id = p_event->attr.config;
219 int ret = -EINVAL;
220
221 mutex_lock(&event_mutex);
222 list_for_each_entry(tp_event, &ftrace_events, list) {
223 if (tp_event->event.type == event_id &&
224 tp_event->class && tp_event->class->reg &&
225 trace_event_try_get_ref(tp_event)) {
226 ret = perf_trace_event_init(tp_event, p_event);
227 if (ret)
228 trace_event_put_ref(tp_event);
229 break;
230 }
231 }
232 mutex_unlock(&event_mutex);
233
234 return ret;
235}
236
237void perf_trace_destroy(struct perf_event *p_event)
238{
239 mutex_lock(&event_mutex);
240 perf_trace_event_close(p_event);
241 perf_trace_event_unreg(p_event);
242 trace_event_put_ref(p_event->tp_event);
243 mutex_unlock(&event_mutex);
244}
245
246#ifdef CONFIG_KPROBE_EVENTS
247int perf_kprobe_init(struct perf_event *p_event, bool is_retprobe)
248{
249 int ret;
250 char *func = NULL;
251 struct trace_event_call *tp_event;
252
253 if (p_event->attr.kprobe_func) {
254 func = strndup_user(u64_to_user_ptr(p_event->attr.kprobe_func),
255 KSYM_NAME_LEN);
256 if (IS_ERR(func)) {
257 ret = PTR_ERR(func);
258 return (ret == -EINVAL) ? -E2BIG : ret;
259 }
260
261 if (func[0] == '\0') {
262 kfree(func);
263 func = NULL;
264 }
265 }
266
267 tp_event = create_local_trace_kprobe(
268 func, (void *)(unsigned long)(p_event->attr.kprobe_addr),
269 p_event->attr.probe_offset, is_retprobe);
270 if (IS_ERR(tp_event)) {
271 ret = PTR_ERR(tp_event);
272 goto out;
273 }
274
275 mutex_lock(&event_mutex);
276 ret = perf_trace_event_init(tp_event, p_event);
277 if (ret)
278 destroy_local_trace_kprobe(tp_event);
279 mutex_unlock(&event_mutex);
280out:
281 kfree(func);
282 return ret;
283}
284
285void perf_kprobe_destroy(struct perf_event *p_event)
286{
287 mutex_lock(&event_mutex);
288 perf_trace_event_close(p_event);
289 perf_trace_event_unreg(p_event);
290 trace_event_put_ref(p_event->tp_event);
291 mutex_unlock(&event_mutex);
292
293 destroy_local_trace_kprobe(p_event->tp_event);
294}
295#endif /* CONFIG_KPROBE_EVENTS */
296
297#ifdef CONFIG_UPROBE_EVENTS
298int perf_uprobe_init(struct perf_event *p_event,
299 unsigned long ref_ctr_offset, bool is_retprobe)
300{
301 int ret;
302 char *path = NULL;
303 struct trace_event_call *tp_event;
304
305 if (!p_event->attr.uprobe_path)
306 return -EINVAL;
307
308 path = strndup_user(u64_to_user_ptr(p_event->attr.uprobe_path),
309 PATH_MAX);
310 if (IS_ERR(path)) {
311 ret = PTR_ERR(path);
312 return (ret == -EINVAL) ? -E2BIG : ret;
313 }
314 if (path[0] == '\0') {
315 ret = -EINVAL;
316 goto out;
317 }
318
319 tp_event = create_local_trace_uprobe(path, p_event->attr.probe_offset,
320 ref_ctr_offset, is_retprobe);
321 if (IS_ERR(tp_event)) {
322 ret = PTR_ERR(tp_event);
323 goto out;
324 }
325
326 /*
327 * local trace_uprobe need to hold event_mutex to call
328 * uprobe_buffer_enable() and uprobe_buffer_disable().
329 * event_mutex is not required for local trace_kprobes.
330 */
331 mutex_lock(&event_mutex);
332 ret = perf_trace_event_init(tp_event, p_event);
333 if (ret)
334 destroy_local_trace_uprobe(tp_event);
335 mutex_unlock(&event_mutex);
336out:
337 kfree(path);
338 return ret;
339}
340
341void perf_uprobe_destroy(struct perf_event *p_event)
342{
343 mutex_lock(&event_mutex);
344 perf_trace_event_close(p_event);
345 perf_trace_event_unreg(p_event);
346 trace_event_put_ref(p_event->tp_event);
347 mutex_unlock(&event_mutex);
348 destroy_local_trace_uprobe(p_event->tp_event);
349}
350#endif /* CONFIG_UPROBE_EVENTS */
351
352int perf_trace_add(struct perf_event *p_event, int flags)
353{
354 struct trace_event_call *tp_event = p_event->tp_event;
355 struct hw_perf_event *hwc = &p_event->hw;
356
357 if (!(flags & PERF_EF_START))
358 p_event->hw.state = PERF_HES_STOPPED;
359
360 if (is_sampling_event(p_event)) {
361 hwc->last_period = hwc->sample_period;
362 perf_swevent_set_period(p_event);
363 }
364
365 /*
366 * If TRACE_REG_PERF_ADD returns false; no custom action was performed
367 * and we need to take the default action of enqueueing our event on
368 * the right per-cpu hlist.
369 */
370 if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
371 struct hlist_head __percpu *pcpu_list;
372 struct hlist_head *list;
373
374 pcpu_list = tp_event->perf_events;
375 if (WARN_ON_ONCE(!pcpu_list))
376 return -EINVAL;
377
378 list = this_cpu_ptr(pcpu_list);
379 hlist_add_head_rcu(&p_event->hlist_entry, list);
380 }
381
382 return 0;
383}
384
385void perf_trace_del(struct perf_event *p_event, int flags)
386{
387 struct trace_event_call *tp_event = p_event->tp_event;
388
389 /*
390 * If TRACE_REG_PERF_DEL returns false; no custom action was performed
391 * and we need to take the default action of dequeueing our event from
392 * the right per-cpu hlist.
393 */
394 if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
395 hlist_del_rcu(&p_event->hlist_entry);
396}
397
398void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
399{
400 char *raw_data;
401 int rctx;
402
403 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
404
405 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
406 "perf buffer not large enough, wanted %d, have %d",
407 size, PERF_MAX_TRACE_SIZE))
408 return NULL;
409
410 *rctxp = rctx = perf_swevent_get_recursion_context();
411 if (rctx < 0)
412 return NULL;
413
414 if (regs)
415 *regs = this_cpu_ptr(&__perf_regs[rctx]);
416 raw_data = this_cpu_ptr(perf_trace_buf[rctx]);
417
418 /* zero the dead bytes from align to not leak stack to user */
419 memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
420 return raw_data;
421}
422EXPORT_SYMBOL_GPL(perf_trace_buf_alloc);
423NOKPROBE_SYMBOL(perf_trace_buf_alloc);
424
425void perf_trace_buf_update(void *record, u16 type)
426{
427 struct trace_entry *entry = record;
428
429 tracing_generic_entry_update(entry, type, tracing_gen_ctx());
430}
431NOKPROBE_SYMBOL(perf_trace_buf_update);
432
433#ifdef CONFIG_FUNCTION_TRACER
434static void
435perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
436 struct ftrace_ops *ops, struct ftrace_regs *fregs)
437{
438 struct ftrace_entry *entry;
439 struct perf_event *event;
440 struct hlist_head head;
441 struct pt_regs regs;
442 int rctx;
443 int bit;
444
445 if (!rcu_is_watching())
446 return;
447
448 bit = ftrace_test_recursion_trylock(ip, parent_ip);
449 if (bit < 0)
450 return;
451
452 if ((unsigned long)ops->private != smp_processor_id())
453 goto out;
454
455 event = container_of(ops, struct perf_event, ftrace_ops);
456
457 /*
458 * @event->hlist entry is NULL (per INIT_HLIST_NODE), and all
459 * the perf code does is hlist_for_each_entry_rcu(), so we can
460 * get away with simply setting the @head.first pointer in order
461 * to create a singular list.
462 */
463 head.first = &event->hlist_entry;
464
465#define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
466 sizeof(u64)) - sizeof(u32))
467
468 BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
469
470 memset(&regs, 0, sizeof(regs));
471 perf_fetch_caller_regs(&regs);
472
473 entry = perf_trace_buf_alloc(ENTRY_SIZE, NULL, &rctx);
474 if (!entry)
475 goto out;
476
477 entry->ip = ip;
478 entry->parent_ip = parent_ip;
479 perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
480 1, &regs, &head, NULL);
481
482out:
483 ftrace_test_recursion_unlock(bit);
484#undef ENTRY_SIZE
485}
486
487static int perf_ftrace_function_register(struct perf_event *event)
488{
489 struct ftrace_ops *ops = &event->ftrace_ops;
490
491 ops->func = perf_ftrace_function_call;
492 ops->private = (void *)(unsigned long)nr_cpu_ids;
493
494 return register_ftrace_function(ops);
495}
496
497static int perf_ftrace_function_unregister(struct perf_event *event)
498{
499 struct ftrace_ops *ops = &event->ftrace_ops;
500 int ret = 0;
501
502 /*
503 * Perf will call this unconditionally even if the ops is not
504 * enabled. The unregister_ftrace_function() will warn if called
505 * when not enabled. Just bypass the unregistering if ops isn't
506 * enabled here.
507 */
508 if (ops->flags & FTRACE_OPS_FL_ENABLED)
509 ret = unregister_ftrace_function(ops);
510
511 ftrace_free_filter(ops);
512 return ret;
513}
514
515int perf_ftrace_event_register(struct trace_event_call *call,
516 enum trace_reg type, void *data)
517{
518 struct perf_event *event = data;
519
520 switch (type) {
521 case TRACE_REG_REGISTER:
522 case TRACE_REG_UNREGISTER:
523 break;
524 case TRACE_REG_PERF_REGISTER:
525 case TRACE_REG_PERF_UNREGISTER:
526 return 0;
527 case TRACE_REG_PERF_OPEN:
528 return perf_ftrace_function_register(data);
529 case TRACE_REG_PERF_CLOSE:
530 return perf_ftrace_function_unregister(data);
531 case TRACE_REG_PERF_ADD:
532 event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();
533 return 1;
534 case TRACE_REG_PERF_DEL:
535 event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
536 return 1;
537 }
538
539 return -EINVAL;
540}
541#endif /* CONFIG_FUNCTION_TRACER */