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	/* 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
 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
 52static int perf_trace_event_reg(struct ftrace_event_call *tp_event,
 53				struct perf_event *p_event)
 54{
 55	struct hlist_head __percpu *list;
 56	int ret = -ENOMEM;
 57	int cpu;
 58
 59	p_event->tp_event = tp_event;
 60	if (tp_event->perf_refcount++ > 0)
 61		return 0;
 62
 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));
 69
 70	tp_event->perf_events = list;
 71
 72	if (!total_ref_count) {
 73		char __percpu *buf;
 74		int i;
 75
 76		for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 77			buf = (char __percpu *)alloc_percpu(perf_trace_t);
 78			if (!buf)
 79				goto fail;
 80
 81			perf_trace_buf[i] = buf;
 82		}
 83	}
 84
 85	ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
 86	if (ret)
 87		goto fail;
 88
 89	total_ref_count++;
 90	return 0;
 91
 92fail:
 93	if (!total_ref_count) {
 94		int i;
 95
 96		for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 97			free_percpu(perf_trace_buf[i]);
 98			perf_trace_buf[i] = NULL;
 99		}
100	}
101
102	if (!--tp_event->perf_refcount) {
103		free_percpu(tp_event->perf_events);
104		tp_event->perf_events = NULL;
105	}
106
107	return ret;
108}
109
110static 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	}
135out:
136	module_put(tp_event->mod);
137}
138
139static 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
145static 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
151static 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
173int perf_trace_init(struct perf_event *p_event)
174{
175	struct ftrace_event_call *tp_event;
176	int event_id = p_event->attr.config;
177	int ret = -EINVAL;
178
179	mutex_lock(&event_mutex);
180	list_for_each_entry(tp_event, &ftrace_events, list) {
181		if (tp_event->event.type == event_id &&
182		    tp_event->class && tp_event->class->reg &&
183		    try_module_get(tp_event->mod)) {
184			ret = perf_trace_event_init(tp_event, p_event);
185			if (ret)
186				module_put(tp_event->mod);
187			break;
188		}
189	}
190	mutex_unlock(&event_mutex);
191
192	return ret;
193}
194
195void 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
203int perf_trace_add(struct perf_event *p_event, int flags)
204{
205	struct ftrace_event_call *tp_event = p_event->tp_event;
206	struct hlist_head __percpu *pcpu_list;
207	struct hlist_head *list;
208
209	pcpu_list = tp_event->perf_events;
210	if (WARN_ON_ONCE(!pcpu_list))
211		return -EINVAL;
212
213	if (!(flags & PERF_EF_START))
214		p_event->hw.state = PERF_HES_STOPPED;
215
216	list = this_cpu_ptr(pcpu_list);
217	hlist_add_head_rcu(&p_event->hlist_entry, list);
218
219	return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
220}
221
222void perf_trace_del(struct perf_event *p_event, int flags)
223{
224	struct ftrace_event_call *tp_event = p_event->tp_event;
225	hlist_del_rcu(&p_event->hlist_entry);
226	tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event);
227}
228
229__kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
230				       struct pt_regs *regs, int *rctxp)
231{
232	struct trace_entry *entry;
233	unsigned long flags;
234	char *raw_data;
235	int pc;
236
237	BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
238
 
 
 
 
239	pc = preempt_count();
240
241	*rctxp = perf_swevent_get_recursion_context();
242	if (*rctxp < 0)
243		return NULL;
244
245	raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
246
247	/* zero the dead bytes from align to not leak stack to user */
248	memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
249
250	entry = (struct trace_entry *)raw_data;
251	local_save_flags(flags);
252	tracing_generic_entry_update(entry, flags, pc);
253	entry->type = type;
254
255	return raw_data;
256}
257EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);
258
259#ifdef CONFIG_FUNCTION_TRACER
260static void
261perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip)
 
262{
263	struct ftrace_entry *entry;
264	struct hlist_head *head;
265	struct pt_regs regs;
266	int rctx;
267
 
 
 
 
268#define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
269		    sizeof(u64)) - sizeof(u32))
270
271	BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
272
273	perf_fetch_caller_regs(&regs);
274
275	entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx);
276	if (!entry)
277		return;
278
279	entry->ip = ip;
280	entry->parent_ip = parent_ip;
281
282	head = this_cpu_ptr(event_function.perf_events);
283	perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0,
284			      1, &regs, head);
285
286#undef ENTRY_SIZE
287}
288
289static int perf_ftrace_function_register(struct perf_event *event)
290{
291	struct ftrace_ops *ops = &event->ftrace_ops;
292
293	ops->flags |= FTRACE_OPS_FL_CONTROL;
294	ops->func = perf_ftrace_function_call;
295	return register_ftrace_function(ops);
296}
297
298static int perf_ftrace_function_unregister(struct perf_event *event)
299{
300	struct ftrace_ops *ops = &event->ftrace_ops;
301	int ret = unregister_ftrace_function(ops);
302	ftrace_free_filter(ops);
303	return ret;
304}
305
306static void perf_ftrace_function_enable(struct perf_event *event)
307{
308	ftrace_function_local_enable(&event->ftrace_ops);
309}
310
311static void perf_ftrace_function_disable(struct perf_event *event)
312{
313	ftrace_function_local_disable(&event->ftrace_ops);
314}
315
316int perf_ftrace_event_register(struct ftrace_event_call *call,
317			       enum trace_reg type, void *data)
318{
319	switch (type) {
320	case TRACE_REG_REGISTER:
321	case TRACE_REG_UNREGISTER:
322		break;
323	case TRACE_REG_PERF_REGISTER:
324	case TRACE_REG_PERF_UNREGISTER:
325		return 0;
326	case TRACE_REG_PERF_OPEN:
327		return perf_ftrace_function_register(data);
328	case TRACE_REG_PERF_CLOSE:
329		return perf_ftrace_function_unregister(data);
330	case TRACE_REG_PERF_ADD:
331		perf_ftrace_function_enable(data);
332		return 0;
333	case TRACE_REG_PERF_DEL:
334		perf_ftrace_function_disable(data);
335		return 0;
336	}
337
338	return -EINVAL;
339}
340#endif /* CONFIG_FUNCTION_TRACER */

  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(&regs);
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, &regs, 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 */