2 * trace event based perf event profiling/tracing
4 * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com>
5 * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com>
8 #include <linux/module.h>
9 #include <linux/kprobes.h>
12 static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
15 * Force it to be aligned to unsigned long to avoid misaligned accesses
18 typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
21 /* Count the events in use (per event id, not per instance) */
22 static int total_ref_count;
24 static int perf_trace_event_perm(struct ftrace_event_call *tp_event,
25 struct perf_event *p_event)
27 if (tp_event->perf_perm) {
28 int ret = tp_event->perf_perm(tp_event, p_event);
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))
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.
43 if (!p_event->attr.exclude_callchain_user)
47 * Same reason to disable user stack dump as for user space
50 if (p_event->attr.sample_type & PERF_SAMPLE_STACK_USER)
54 /* No tracing, just counting, so no obvious leak */
55 if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
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)
65 * ...otherwise raw tracepoint data can be a severe data leak,
66 * only allow root to have these.
68 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
74 static int perf_trace_event_reg(struct ftrace_event_call *tp_event,
75 struct perf_event *p_event)
77 struct hlist_head __percpu *list;
81 p_event->tp_event = tp_event;
82 if (tp_event->perf_refcount++ > 0)
85 list = alloc_percpu(struct hlist_head);
89 for_each_possible_cpu(cpu)
90 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
92 tp_event->perf_events = list;
94 if (!total_ref_count) {
98 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
99 buf = (char __percpu *)alloc_percpu(perf_trace_t);
103 perf_trace_buf[i] = buf;
107 ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
115 if (!total_ref_count) {
118 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
119 free_percpu(perf_trace_buf[i]);
120 perf_trace_buf[i] = NULL;
124 if (!--tp_event->perf_refcount) {
125 free_percpu(tp_event->perf_events);
126 tp_event->perf_events = NULL;
132 static void perf_trace_event_unreg(struct perf_event *p_event)
134 struct ftrace_event_call *tp_event = p_event->tp_event;
137 if (--tp_event->perf_refcount > 0)
140 tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
143 * Ensure our callback won't be called anymore. The buffers
144 * will be freed after that.
146 tracepoint_synchronize_unregister();
148 free_percpu(tp_event->perf_events);
149 tp_event->perf_events = NULL;
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;
158 module_put(tp_event->mod);
161 static int perf_trace_event_open(struct perf_event *p_event)
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);
167 static void perf_trace_event_close(struct perf_event *p_event)
169 struct ftrace_event_call *tp_event = p_event->tp_event;
170 tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
173 static int perf_trace_event_init(struct ftrace_event_call *tp_event,
174 struct perf_event *p_event)
178 ret = perf_trace_event_perm(tp_event, p_event);
182 ret = perf_trace_event_reg(tp_event, p_event);
186 ret = perf_trace_event_open(p_event);
188 perf_trace_event_unreg(p_event);
195 int perf_trace_init(struct perf_event *p_event)
197 struct ftrace_event_call *tp_event;
198 u64 event_id = p_event->attr.config;
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);
208 module_put(tp_event->mod);
212 mutex_unlock(&event_mutex);
217 void perf_trace_destroy(struct perf_event *p_event)
219 mutex_lock(&event_mutex);
220 perf_trace_event_close(p_event);
221 perf_trace_event_unreg(p_event);
222 mutex_unlock(&event_mutex);
225 int perf_trace_add(struct perf_event *p_event, int flags)
227 struct ftrace_event_call *tp_event = p_event->tp_event;
228 struct hlist_head __percpu *pcpu_list;
229 struct hlist_head *list;
231 pcpu_list = tp_event->perf_events;
232 if (WARN_ON_ONCE(!pcpu_list))
235 if (!(flags & PERF_EF_START))
236 p_event->hw.state = PERF_HES_STOPPED;
238 list = this_cpu_ptr(pcpu_list);
239 hlist_add_head_rcu(&p_event->hlist_entry, list);
241 return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
244 void perf_trace_del(struct perf_event *p_event, int flags)
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);
251 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
252 struct pt_regs *regs, int *rctxp)
254 struct trace_entry *entry;
259 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
261 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
262 "perf buffer not large enough"))
265 pc = preempt_count();
267 *rctxp = perf_swevent_get_recursion_context();
271 raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
273 /* zero the dead bytes from align to not leak stack to user */
274 memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
276 entry = (struct trace_entry *)raw_data;
277 local_save_flags(flags);
278 tracing_generic_entry_update(entry, flags, pc);
283 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);
285 #ifdef CONFIG_FUNCTION_TRACER
287 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
288 struct ftrace_ops *ops, struct pt_regs *pt_regs)
290 struct ftrace_entry *entry;
291 struct hlist_head *head;
295 head = this_cpu_ptr(event_function.perf_events);
296 if (hlist_empty(head))
299 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
300 sizeof(u64)) - sizeof(u32))
302 BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
304 perf_fetch_caller_regs(®s);
306 entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx);
311 entry->parent_ip = parent_ip;
312 perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0,
313 1, ®s, head, NULL);
318 static int perf_ftrace_function_register(struct perf_event *event)
320 struct ftrace_ops *ops = &event->ftrace_ops;
322 ops->flags |= FTRACE_OPS_FL_CONTROL;
323 ops->func = perf_ftrace_function_call;
324 return register_ftrace_function(ops);
327 static int perf_ftrace_function_unregister(struct perf_event *event)
329 struct ftrace_ops *ops = &event->ftrace_ops;
330 int ret = unregister_ftrace_function(ops);
331 ftrace_free_filter(ops);
335 static void perf_ftrace_function_enable(struct perf_event *event)
337 ftrace_function_local_enable(&event->ftrace_ops);
340 static void perf_ftrace_function_disable(struct perf_event *event)
342 ftrace_function_local_disable(&event->ftrace_ops);
345 int perf_ftrace_event_register(struct ftrace_event_call *call,
346 enum trace_reg type, void *data)
349 case TRACE_REG_REGISTER:
350 case TRACE_REG_UNREGISTER:
352 case TRACE_REG_PERF_REGISTER:
353 case TRACE_REG_PERF_UNREGISTER:
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);
362 case TRACE_REG_PERF_DEL:
363 perf_ftrace_function_disable(data);
369 #endif /* CONFIG_FUNCTION_TRACER */