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