2 * trace event based perf event profiling/tracing
4 * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra
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 trace_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
);
34 * We checked and allowed to create parent,
35 * allow children without checking.
41 * It's ok to check current process (owner) permissions in here,
42 * because code below is called only via perf_event_open syscall.
45 /* The ftrace function trace is allowed only for root. */
46 if (ftrace_event_is_function(tp_event
)) {
47 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN
))
50 if (!is_sampling_event(p_event
))
54 * We don't allow user space callchains for function trace
55 * event, due to issues with page faults while tracing page
56 * fault handler and its overall trickiness nature.
58 if (!p_event
->attr
.exclude_callchain_user
)
62 * Same reason to disable user stack dump as for user space
65 if (p_event
->attr
.sample_type
& PERF_SAMPLE_STACK_USER
)
69 /* No tracing, just counting, so no obvious leak */
70 if (!(p_event
->attr
.sample_type
& PERF_SAMPLE_RAW
))
73 /* Some events are ok to be traced by non-root users... */
74 if (p_event
->attach_state
== PERF_ATTACH_TASK
) {
75 if (tp_event
->flags
& TRACE_EVENT_FL_CAP_ANY
)
80 * ...otherwise raw tracepoint data can be a severe data leak,
81 * only allow root to have these.
83 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN
))
89 static int perf_trace_event_reg(struct trace_event_call
*tp_event
,
90 struct perf_event
*p_event
)
92 struct hlist_head __percpu
*list
;
96 p_event
->tp_event
= tp_event
;
97 if (tp_event
->perf_refcount
++ > 0)
100 list
= alloc_percpu(struct hlist_head
);
104 for_each_possible_cpu(cpu
)
105 INIT_HLIST_HEAD(per_cpu_ptr(list
, cpu
));
107 tp_event
->perf_events
= list
;
109 if (!total_ref_count
) {
113 for (i
= 0; i
< PERF_NR_CONTEXTS
; i
++) {
114 buf
= (char __percpu
*)alloc_percpu(perf_trace_t
);
118 perf_trace_buf
[i
] = buf
;
122 ret
= tp_event
->class->reg(tp_event
, TRACE_REG_PERF_REGISTER
, NULL
);
130 if (!total_ref_count
) {
133 for (i
= 0; i
< PERF_NR_CONTEXTS
; i
++) {
134 free_percpu(perf_trace_buf
[i
]);
135 perf_trace_buf
[i
] = NULL
;
139 if (!--tp_event
->perf_refcount
) {
140 free_percpu(tp_event
->perf_events
);
141 tp_event
->perf_events
= NULL
;
147 static void perf_trace_event_unreg(struct perf_event
*p_event
)
149 struct trace_event_call
*tp_event
= p_event
->tp_event
;
152 if (--tp_event
->perf_refcount
> 0)
155 tp_event
->class->reg(tp_event
, TRACE_REG_PERF_UNREGISTER
, NULL
);
158 * Ensure our callback won't be called anymore. The buffers
159 * will be freed after that.
161 tracepoint_synchronize_unregister();
163 free_percpu(tp_event
->perf_events
);
164 tp_event
->perf_events
= NULL
;
166 if (!--total_ref_count
) {
167 for (i
= 0; i
< PERF_NR_CONTEXTS
; i
++) {
168 free_percpu(perf_trace_buf
[i
]);
169 perf_trace_buf
[i
] = NULL
;
173 module_put(tp_event
->mod
);
176 static int perf_trace_event_open(struct perf_event
*p_event
)
178 struct trace_event_call
*tp_event
= p_event
->tp_event
;
179 return tp_event
->class->reg(tp_event
, TRACE_REG_PERF_OPEN
, p_event
);
182 static void perf_trace_event_close(struct perf_event
*p_event
)
184 struct trace_event_call
*tp_event
= p_event
->tp_event
;
185 tp_event
->class->reg(tp_event
, TRACE_REG_PERF_CLOSE
, p_event
);
188 static int perf_trace_event_init(struct trace_event_call
*tp_event
,
189 struct perf_event
*p_event
)
193 ret
= perf_trace_event_perm(tp_event
, p_event
);
197 ret
= perf_trace_event_reg(tp_event
, p_event
);
201 ret
= perf_trace_event_open(p_event
);
203 perf_trace_event_unreg(p_event
);
210 int perf_trace_init(struct perf_event
*p_event
)
212 struct trace_event_call
*tp_event
;
213 u64 event_id
= p_event
->attr
.config
;
216 mutex_lock(&event_mutex
);
217 list_for_each_entry(tp_event
, &ftrace_events
, list
) {
218 if (tp_event
->event
.type
== event_id
&&
219 tp_event
->class && tp_event
->class->reg
&&
220 try_module_get(tp_event
->mod
)) {
221 ret
= perf_trace_event_init(tp_event
, p_event
);
223 module_put(tp_event
->mod
);
227 mutex_unlock(&event_mutex
);
232 void perf_trace_destroy(struct perf_event
*p_event
)
234 mutex_lock(&event_mutex
);
235 perf_trace_event_close(p_event
);
236 perf_trace_event_unreg(p_event
);
237 mutex_unlock(&event_mutex
);
240 int perf_trace_add(struct perf_event
*p_event
, int flags
)
242 struct trace_event_call
*tp_event
= p_event
->tp_event
;
243 struct hlist_head __percpu
*pcpu_list
;
244 struct hlist_head
*list
;
246 pcpu_list
= tp_event
->perf_events
;
247 if (WARN_ON_ONCE(!pcpu_list
))
250 if (!(flags
& PERF_EF_START
))
251 p_event
->hw
.state
= PERF_HES_STOPPED
;
253 list
= this_cpu_ptr(pcpu_list
);
254 hlist_add_head_rcu(&p_event
->hlist_entry
, list
);
256 return tp_event
->class->reg(tp_event
, TRACE_REG_PERF_ADD
, p_event
);
259 void perf_trace_del(struct perf_event
*p_event
, int flags
)
261 struct trace_event_call
*tp_event
= p_event
->tp_event
;
262 hlist_del_rcu(&p_event
->hlist_entry
);
263 tp_event
->class->reg(tp_event
, TRACE_REG_PERF_DEL
, p_event
);
266 void *perf_trace_buf_prepare(int size
, unsigned short type
,
267 struct pt_regs
**regs
, int *rctxp
)
269 struct trace_entry
*entry
;
274 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE
% sizeof(unsigned long));
276 if (WARN_ONCE(size
> PERF_MAX_TRACE_SIZE
,
277 "perf buffer not large enough"))
280 pc
= preempt_count();
282 *rctxp
= perf_swevent_get_recursion_context();
287 *regs
= this_cpu_ptr(&__perf_regs
[*rctxp
]);
288 raw_data
= this_cpu_ptr(perf_trace_buf
[*rctxp
]);
290 /* zero the dead bytes from align to not leak stack to user */
291 memset(&raw_data
[size
- sizeof(u64
)], 0, sizeof(u64
));
293 entry
= (struct trace_entry
*)raw_data
;
294 local_save_flags(flags
);
295 tracing_generic_entry_update(entry
, flags
, pc
);
300 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare
);
301 NOKPROBE_SYMBOL(perf_trace_buf_prepare
);
303 #ifdef CONFIG_FUNCTION_TRACER
305 perf_ftrace_function_call(unsigned long ip
, unsigned long parent_ip
,
306 struct ftrace_ops
*ops
, struct pt_regs
*pt_regs
)
308 struct ftrace_entry
*entry
;
309 struct hlist_head
*head
;
313 head
= this_cpu_ptr(event_function
.perf_events
);
314 if (hlist_empty(head
))
317 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
318 sizeof(u64)) - sizeof(u32))
320 BUILD_BUG_ON(ENTRY_SIZE
> PERF_MAX_TRACE_SIZE
);
322 perf_fetch_caller_regs(®s
);
324 entry
= perf_trace_buf_prepare(ENTRY_SIZE
, TRACE_FN
, NULL
, &rctx
);
329 entry
->parent_ip
= parent_ip
;
330 perf_trace_buf_submit(entry
, ENTRY_SIZE
, rctx
, 0,
331 1, ®s
, head
, NULL
);
336 static int perf_ftrace_function_register(struct perf_event
*event
)
338 struct ftrace_ops
*ops
= &event
->ftrace_ops
;
340 ops
->flags
|= FTRACE_OPS_FL_PER_CPU
| FTRACE_OPS_FL_RCU
;
341 ops
->func
= perf_ftrace_function_call
;
342 return register_ftrace_function(ops
);
345 static int perf_ftrace_function_unregister(struct perf_event
*event
)
347 struct ftrace_ops
*ops
= &event
->ftrace_ops
;
348 int ret
= unregister_ftrace_function(ops
);
349 ftrace_free_filter(ops
);
353 static void perf_ftrace_function_enable(struct perf_event
*event
)
355 ftrace_function_local_enable(&event
->ftrace_ops
);
358 static void perf_ftrace_function_disable(struct perf_event
*event
)
360 ftrace_function_local_disable(&event
->ftrace_ops
);
363 int perf_ftrace_event_register(struct trace_event_call
*call
,
364 enum trace_reg type
, void *data
)
367 case TRACE_REG_REGISTER
:
368 case TRACE_REG_UNREGISTER
:
370 case TRACE_REG_PERF_REGISTER
:
371 case TRACE_REG_PERF_UNREGISTER
:
373 case TRACE_REG_PERF_OPEN
:
374 return perf_ftrace_function_register(data
);
375 case TRACE_REG_PERF_CLOSE
:
376 return perf_ftrace_function_unregister(data
);
377 case TRACE_REG_PERF_ADD
:
378 perf_ftrace_function_enable(data
);
380 case TRACE_REG_PERF_DEL
:
381 perf_ftrace_function_disable(data
);
387 #endif /* CONFIG_FUNCTION_TRACER */