Commit | Line | Data |
---|---|---|
ac199db0 | 1 | /* |
97d5a220 | 2 | * trace event based perf event profiling/tracing |
ac199db0 PZ |
3 | * |
4 | * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com> | |
c530665c | 5 | * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com> |
ac199db0 PZ |
6 | */ |
7 | ||
558e6547 | 8 | #include <linux/module.h> |
430ad5a6 | 9 | #include <linux/kprobes.h> |
ac199db0 PZ |
10 | #include "trace.h" |
11 | ||
6016ee13 | 12 | static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS]; |
20ab4425 | 13 | |
eb1e7961 FW |
14 | /* |
15 | * Force it to be aligned to unsigned long to avoid misaligned accesses | |
16 | * suprises | |
17 | */ | |
18 | typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)]) | |
19 | perf_trace_t; | |
ce71b9df | 20 | |
20ab4425 | 21 | /* Count the events in use (per event id, not per instance) */ |
97d5a220 | 22 | static int total_ref_count; |
20ab4425 | 23 | |
61c32659 FW |
24 | static int perf_trace_event_perm(struct ftrace_event_call *tp_event, |
25 | struct perf_event *p_event) | |
26 | { | |
ced39002 JO |
27 | /* The ftrace function trace is allowed only for root. */ |
28 | if (ftrace_event_is_function(tp_event) && | |
29 | perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
30 | return -EPERM; | |
31 | ||
61c32659 FW |
32 | /* No tracing, just counting, so no obvious leak */ |
33 | if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW)) | |
34 | return 0; | |
35 | ||
36 | /* Some events are ok to be traced by non-root users... */ | |
37 | if (p_event->attach_state == PERF_ATTACH_TASK) { | |
38 | if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY) | |
39 | return 0; | |
40 | } | |
41 | ||
42 | /* | |
43 | * ...otherwise raw tracepoint data can be a severe data leak, | |
44 | * only allow root to have these. | |
45 | */ | |
46 | if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN)) | |
47 | return -EPERM; | |
48 | ||
49 | return 0; | |
50 | } | |
51 | ||
ceec0b6f JO |
52 | static int perf_trace_event_reg(struct ftrace_event_call *tp_event, |
53 | struct perf_event *p_event) | |
e5e25cf4 | 54 | { |
6016ee13 | 55 | struct hlist_head __percpu *list; |
ceec0b6f | 56 | int ret = -ENOMEM; |
1c024eca | 57 | int cpu; |
20ab4425 | 58 | |
1c024eca PZ |
59 | p_event->tp_event = tp_event; |
60 | if (tp_event->perf_refcount++ > 0) | |
e5e25cf4 FW |
61 | return 0; |
62 | ||
1c024eca PZ |
63 | list = alloc_percpu(struct hlist_head); |
64 | if (!list) | |
65 | goto fail; | |
66 | ||
67 | for_each_possible_cpu(cpu) | |
68 | INIT_HLIST_HEAD(per_cpu_ptr(list, cpu)); | |
20ab4425 | 69 | |
1c024eca | 70 | tp_event->perf_events = list; |
e5e25cf4 | 71 | |
97d5a220 | 72 | if (!total_ref_count) { |
6016ee13 | 73 | char __percpu *buf; |
b7e2ecef | 74 | int i; |
20ab4425 | 75 | |
7ae07ea3 | 76 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
6016ee13 | 77 | buf = (char __percpu *)alloc_percpu(perf_trace_t); |
b7e2ecef | 78 | if (!buf) |
1c024eca | 79 | goto fail; |
20ab4425 | 80 | |
1c024eca | 81 | perf_trace_buf[i] = buf; |
b7e2ecef | 82 | } |
20ab4425 FW |
83 | } |
84 | ||
ceec0b6f | 85 | ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL); |
1c024eca PZ |
86 | if (ret) |
87 | goto fail; | |
20ab4425 | 88 | |
1c024eca PZ |
89 | total_ref_count++; |
90 | return 0; | |
91 | ||
92 | fail: | |
97d5a220 | 93 | if (!total_ref_count) { |
b7e2ecef PZ |
94 | int i; |
95 | ||
7ae07ea3 | 96 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
b7e2ecef PZ |
97 | free_percpu(perf_trace_buf[i]); |
98 | perf_trace_buf[i] = NULL; | |
99 | } | |
fe8e5b5a | 100 | } |
1c024eca PZ |
101 | |
102 | if (!--tp_event->perf_refcount) { | |
103 | free_percpu(tp_event->perf_events); | |
104 | tp_event->perf_events = NULL; | |
fe8e5b5a | 105 | } |
20ab4425 FW |
106 | |
107 | return ret; | |
e5e25cf4 FW |
108 | } |
109 | ||
ceec0b6f JO |
110 | static void perf_trace_event_unreg(struct perf_event *p_event) |
111 | { | |
112 | struct ftrace_event_call *tp_event = p_event->tp_event; | |
113 | int i; | |
114 | ||
115 | if (--tp_event->perf_refcount > 0) | |
116 | goto out; | |
117 | ||
118 | tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL); | |
119 | ||
120 | /* | |
121 | * Ensure our callback won't be called anymore. The buffers | |
122 | * will be freed after that. | |
123 | */ | |
124 | tracepoint_synchronize_unregister(); | |
125 | ||
126 | free_percpu(tp_event->perf_events); | |
127 | tp_event->perf_events = NULL; | |
128 | ||
129 | if (!--total_ref_count) { | |
130 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { | |
131 | free_percpu(perf_trace_buf[i]); | |
132 | perf_trace_buf[i] = NULL; | |
133 | } | |
134 | } | |
135 | out: | |
136 | module_put(tp_event->mod); | |
137 | } | |
138 | ||
139 | static int perf_trace_event_open(struct perf_event *p_event) | |
140 | { | |
141 | struct ftrace_event_call *tp_event = p_event->tp_event; | |
142 | return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event); | |
143 | } | |
144 | ||
145 | static void perf_trace_event_close(struct perf_event *p_event) | |
146 | { | |
147 | struct ftrace_event_call *tp_event = p_event->tp_event; | |
148 | tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event); | |
149 | } | |
150 | ||
151 | static int perf_trace_event_init(struct ftrace_event_call *tp_event, | |
152 | struct perf_event *p_event) | |
153 | { | |
154 | int ret; | |
155 | ||
156 | ret = perf_trace_event_perm(tp_event, p_event); | |
157 | if (ret) | |
158 | return ret; | |
159 | ||
160 | ret = perf_trace_event_reg(tp_event, p_event); | |
161 | if (ret) | |
162 | return ret; | |
163 | ||
164 | ret = perf_trace_event_open(p_event); | |
165 | if (ret) { | |
166 | perf_trace_event_unreg(p_event); | |
167 | return ret; | |
168 | } | |
169 | ||
170 | return 0; | |
171 | } | |
172 | ||
1c024eca | 173 | int perf_trace_init(struct perf_event *p_event) |
ac199db0 | 174 | { |
1c024eca PZ |
175 | struct ftrace_event_call *tp_event; |
176 | int event_id = p_event->attr.config; | |
20c8928a | 177 | int ret = -EINVAL; |
ac199db0 | 178 | |
20c8928a | 179 | mutex_lock(&event_mutex); |
1c024eca | 180 | list_for_each_entry(tp_event, &ftrace_events, list) { |
ff5f149b | 181 | if (tp_event->event.type == event_id && |
a1d0ce82 | 182 | tp_event->class && tp_event->class->reg && |
1c024eca PZ |
183 | try_module_get(tp_event->mod)) { |
184 | ret = perf_trace_event_init(tp_event, p_event); | |
9cb627d5 LZ |
185 | if (ret) |
186 | module_put(tp_event->mod); | |
20c8928a LZ |
187 | break; |
188 | } | |
ac199db0 | 189 | } |
20c8928a | 190 | mutex_unlock(&event_mutex); |
ac199db0 | 191 | |
20c8928a | 192 | return ret; |
ac199db0 PZ |
193 | } |
194 | ||
ceec0b6f JO |
195 | void perf_trace_destroy(struct perf_event *p_event) |
196 | { | |
197 | mutex_lock(&event_mutex); | |
198 | perf_trace_event_close(p_event); | |
199 | perf_trace_event_unreg(p_event); | |
200 | mutex_unlock(&event_mutex); | |
201 | } | |
202 | ||
a4eaf7f1 | 203 | int perf_trace_add(struct perf_event *p_event, int flags) |
e5e25cf4 | 204 | { |
1c024eca | 205 | struct ftrace_event_call *tp_event = p_event->tp_event; |
6016ee13 | 206 | struct hlist_head __percpu *pcpu_list; |
1c024eca | 207 | struct hlist_head *list; |
20ab4425 | 208 | |
6016ee13 NK |
209 | pcpu_list = tp_event->perf_events; |
210 | if (WARN_ON_ONCE(!pcpu_list)) | |
1c024eca | 211 | return -EINVAL; |
20ab4425 | 212 | |
a4eaf7f1 PZ |
213 | if (!(flags & PERF_EF_START)) |
214 | p_event->hw.state = PERF_HES_STOPPED; | |
215 | ||
6016ee13 | 216 | list = this_cpu_ptr(pcpu_list); |
1c024eca | 217 | hlist_add_head_rcu(&p_event->hlist_entry, list); |
20ab4425 | 218 | |
489c75c3 | 219 | return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event); |
1c024eca | 220 | } |
20ab4425 | 221 | |
a4eaf7f1 | 222 | void perf_trace_del(struct perf_event *p_event, int flags) |
1c024eca | 223 | { |
489c75c3 | 224 | struct ftrace_event_call *tp_event = p_event->tp_event; |
1c024eca | 225 | hlist_del_rcu(&p_event->hlist_entry); |
489c75c3 | 226 | tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event); |
e5e25cf4 FW |
227 | } |
228 | ||
97d5a220 | 229 | __kprobes void *perf_trace_buf_prepare(int size, unsigned short type, |
b7e2ecef | 230 | struct pt_regs *regs, int *rctxp) |
430ad5a6 XG |
231 | { |
232 | struct trace_entry *entry; | |
87f44bbc | 233 | unsigned long flags; |
1c024eca | 234 | char *raw_data; |
b7e2ecef | 235 | int pc; |
430ad5a6 | 236 | |
eb1e7961 FW |
237 | BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long)); |
238 | ||
430ad5a6 XG |
239 | pc = preempt_count(); |
240 | ||
430ad5a6 XG |
241 | *rctxp = perf_swevent_get_recursion_context(); |
242 | if (*rctxp < 0) | |
1c024eca | 243 | return NULL; |
430ad5a6 | 244 | |
3771f077 | 245 | raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]); |
430ad5a6 XG |
246 | |
247 | /* zero the dead bytes from align to not leak stack to user */ | |
eb1e7961 | 248 | memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64)); |
430ad5a6 XG |
249 | |
250 | entry = (struct trace_entry *)raw_data; | |
87f44bbc PZ |
251 | local_save_flags(flags); |
252 | tracing_generic_entry_update(entry, flags, pc); | |
430ad5a6 XG |
253 | entry->type = type; |
254 | ||
255 | return raw_data; | |
430ad5a6 | 256 | } |
97d5a220 | 257 | EXPORT_SYMBOL_GPL(perf_trace_buf_prepare); |
ced39002 JO |
258 | |
259 | #ifdef CONFIG_FUNCTION_TRACER | |
260 | static void | |
2f5f6ad9 | 261 | perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip, |
a1e2e31d | 262 | struct ftrace_ops *ops, struct pt_regs *pt_regs) |
ced39002 JO |
263 | { |
264 | struct ftrace_entry *entry; | |
265 | struct hlist_head *head; | |
266 | struct pt_regs regs; | |
267 | int rctx; | |
268 | ||
269 | #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \ | |
270 | sizeof(u64)) - sizeof(u32)) | |
271 | ||
272 | BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE); | |
273 | ||
274 | perf_fetch_caller_regs(®s); | |
275 | ||
276 | entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx); | |
277 | if (!entry) | |
278 | return; | |
279 | ||
280 | entry->ip = ip; | |
281 | entry->parent_ip = parent_ip; | |
282 | ||
283 | head = this_cpu_ptr(event_function.perf_events); | |
284 | perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0, | |
e6dab5ff | 285 | 1, ®s, head, NULL); |
ced39002 JO |
286 | |
287 | #undef ENTRY_SIZE | |
288 | } | |
289 | ||
290 | static int perf_ftrace_function_register(struct perf_event *event) | |
291 | { | |
292 | struct ftrace_ops *ops = &event->ftrace_ops; | |
293 | ||
294 | ops->flags |= FTRACE_OPS_FL_CONTROL; | |
295 | ops->func = perf_ftrace_function_call; | |
296 | return register_ftrace_function(ops); | |
297 | } | |
298 | ||
299 | static int perf_ftrace_function_unregister(struct perf_event *event) | |
300 | { | |
301 | struct ftrace_ops *ops = &event->ftrace_ops; | |
5500fa51 JO |
302 | int ret = unregister_ftrace_function(ops); |
303 | ftrace_free_filter(ops); | |
304 | return ret; | |
ced39002 JO |
305 | } |
306 | ||
307 | static void perf_ftrace_function_enable(struct perf_event *event) | |
308 | { | |
309 | ftrace_function_local_enable(&event->ftrace_ops); | |
310 | } | |
311 | ||
312 | static void perf_ftrace_function_disable(struct perf_event *event) | |
313 | { | |
314 | ftrace_function_local_disable(&event->ftrace_ops); | |
315 | } | |
316 | ||
317 | int perf_ftrace_event_register(struct ftrace_event_call *call, | |
318 | enum trace_reg type, void *data) | |
319 | { | |
320 | switch (type) { | |
321 | case TRACE_REG_REGISTER: | |
322 | case TRACE_REG_UNREGISTER: | |
323 | break; | |
324 | case TRACE_REG_PERF_REGISTER: | |
325 | case TRACE_REG_PERF_UNREGISTER: | |
326 | return 0; | |
327 | case TRACE_REG_PERF_OPEN: | |
328 | return perf_ftrace_function_register(data); | |
329 | case TRACE_REG_PERF_CLOSE: | |
330 | return perf_ftrace_function_unregister(data); | |
331 | case TRACE_REG_PERF_ADD: | |
332 | perf_ftrace_function_enable(data); | |
333 | return 0; | |
334 | case TRACE_REG_PERF_DEL: | |
335 | perf_ftrace_function_disable(data); | |
336 | return 0; | |
337 | } | |
338 | ||
339 | return -EINVAL; | |
340 | } | |
341 | #endif /* CONFIG_FUNCTION_TRACER */ |