blob: 5a927075977fc8c66b62b9e4e7bca0432fd7e355 [file] [log] [blame]
Peter Zijlstraac199db2009-03-19 20:26:15 +01001/*
Frederic Weisbecker97d5a222010-03-05 05:35:37 +01002 * trace event based perf event profiling/tracing
Peter Zijlstraac199db2009-03-19 20:26:15 +01003 *
Peter Zijlstra90eec102015-11-16 11:08:45 +01004 * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra
Frederic Weisbeckerc5306652010-03-03 07:16:16 +01005 * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com>
Peter Zijlstraac199db2009-03-19 20:26:15 +01006 */
7
Li Zefan558e6542009-08-24 12:19:47 +08008#include <linux/module.h>
Xiao Guangrong430ad5a2010-01-28 09:32:29 +08009#include <linux/kprobes.h>
Peter Zijlstraac199db2009-03-19 20:26:15 +010010#include "trace.h"
11
Namhyung Kim6016ee12010-08-11 12:47:59 +090012static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
Frederic Weisbecker20ab44252009-09-18 06:10:28 +020013
Frederic Weisbeckereb1e7962010-03-23 00:08:59 +010014/*
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;
Frederic Weisbeckerce71b9d2009-11-22 05:26:55 +010020
Frederic Weisbecker20ab44252009-09-18 06:10:28 +020021/* Count the events in use (per event id, not per instance) */
Frederic Weisbecker97d5a222010-03-05 05:35:37 +010022static int total_ref_count;
Frederic Weisbecker20ab44252009-09-18 06:10:28 +020023
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -040024static int perf_trace_event_perm(struct trace_event_call *tp_event,
Frederic Weisbecker61c32652010-11-18 01:39:17 +010025 struct perf_event *p_event)
26{
Peter Zijlstrad5b5f392013-11-14 16:23:04 +010027 if (tp_event->perf_perm) {
28 int ret = tp_event->perf_perm(tp_event, p_event);
29 if (ret)
30 return ret;
31 }
32
Jiri Olsaf4be0732014-07-16 14:33:29 +020033 /*
34 * We checked and allowed to create parent,
35 * allow children without checking.
36 */
37 if (p_event->parent)
38 return 0;
39
40 /*
41 * It's ok to check current process (owner) permissions in here,
42 * because code below is called only via perf_event_open syscall.
43 */
44
Jiri Olsaced39002012-02-15 15:51:52 +010045 /* The ftrace function trace is allowed only for root. */
Jiri Olsacfa77bc2014-03-02 16:56:38 +010046 if (ftrace_event_is_function(tp_event)) {
47 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
48 return -EPERM;
49
50 /*
51 * We don't allow user space callchains for function trace
52 * event, due to issues with page faults while tracing page
53 * fault handler and its overall trickiness nature.
54 */
55 if (!p_event->attr.exclude_callchain_user)
56 return -EINVAL;
Jiri Olsa63c45f42014-03-02 16:56:39 +010057
58 /*
59 * Same reason to disable user stack dump as for user space
60 * callchains above.
61 */
62 if (p_event->attr.sample_type & PERF_SAMPLE_STACK_USER)
63 return -EINVAL;
Jiri Olsacfa77bc2014-03-02 16:56:38 +010064 }
Jiri Olsaced39002012-02-15 15:51:52 +010065
Frederic Weisbecker61c32652010-11-18 01:39:17 +010066 /* No tracing, just counting, so no obvious leak */
67 if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
68 return 0;
69
70 /* Some events are ok to be traced by non-root users... */
71 if (p_event->attach_state == PERF_ATTACH_TASK) {
72 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
73 return 0;
74 }
75
76 /*
77 * ...otherwise raw tracepoint data can be a severe data leak,
78 * only allow root to have these.
79 */
80 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
81 return -EPERM;
82
83 return 0;
84}
85
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -040086static int perf_trace_event_reg(struct trace_event_call *tp_event,
Jiri Olsaceec0b62012-02-15 15:51:49 +010087 struct perf_event *p_event)
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +020088{
Namhyung Kim6016ee12010-08-11 12:47:59 +090089 struct hlist_head __percpu *list;
Jiri Olsaceec0b62012-02-15 15:51:49 +010090 int ret = -ENOMEM;
Peter Zijlstra1c024eca2010-05-19 14:02:22 +020091 int cpu;
Frederic Weisbecker20ab44252009-09-18 06:10:28 +020092
Peter Zijlstra1c024eca2010-05-19 14:02:22 +020093 p_event->tp_event = tp_event;
94 if (tp_event->perf_refcount++ > 0)
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +020095 return 0;
96
Peter Zijlstra1c024eca2010-05-19 14:02:22 +020097 list = alloc_percpu(struct hlist_head);
98 if (!list)
99 goto fail;
100
101 for_each_possible_cpu(cpu)
102 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
103
104 tp_event->perf_events = list;
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +0200105
Frederic Weisbecker97d5a222010-03-05 05:35:37 +0100106 if (!total_ref_count) {
Namhyung Kim6016ee12010-08-11 12:47:59 +0900107 char __percpu *buf;
Peter Zijlstrab7e2ece2010-05-19 10:52:27 +0200108 int i;
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200109
Frederic Weisbecker7ae07ea2010-08-14 20:45:13 +0200110 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
Namhyung Kim6016ee12010-08-11 12:47:59 +0900111 buf = (char __percpu *)alloc_percpu(perf_trace_t);
Peter Zijlstrab7e2ece2010-05-19 10:52:27 +0200112 if (!buf)
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200113 goto fail;
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200114
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200115 perf_trace_buf[i] = buf;
Peter Zijlstrab7e2ece2010-05-19 10:52:27 +0200116 }
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200117 }
118
Jiri Olsaceec0b62012-02-15 15:51:49 +0100119 ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200120 if (ret)
121 goto fail;
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200122
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200123 total_ref_count++;
124 return 0;
125
126fail:
Peter Zijlstrab7e2ece2010-05-19 10:52:27 +0200127 if (!total_ref_count) {
128 int i;
129
Frederic Weisbecker7ae07ea2010-08-14 20:45:13 +0200130 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
Peter Zijlstrab7e2ece2010-05-19 10:52:27 +0200131 free_percpu(perf_trace_buf[i]);
132 perf_trace_buf[i] = NULL;
133 }
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200134 }
135
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200136 if (!--tp_event->perf_refcount) {
137 free_percpu(tp_event->perf_events);
138 tp_event->perf_events = NULL;
Frederic Weisbeckerfe8e5b52009-10-03 14:55:18 +0200139 }
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200140
141 return ret;
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +0200142}
143
Jiri Olsaceec0b62012-02-15 15:51:49 +0100144static void perf_trace_event_unreg(struct perf_event *p_event)
145{
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400146 struct trace_event_call *tp_event = p_event->tp_event;
Jiri Olsaceec0b62012-02-15 15:51:49 +0100147 int i;
148
149 if (--tp_event->perf_refcount > 0)
150 goto out;
151
152 tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
153
154 /*
155 * Ensure our callback won't be called anymore. The buffers
156 * will be freed after that.
157 */
158 tracepoint_synchronize_unregister();
159
160 free_percpu(tp_event->perf_events);
161 tp_event->perf_events = NULL;
162
163 if (!--total_ref_count) {
164 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
165 free_percpu(perf_trace_buf[i]);
166 perf_trace_buf[i] = NULL;
167 }
168 }
169out:
170 module_put(tp_event->mod);
171}
172
173static int perf_trace_event_open(struct perf_event *p_event)
174{
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400175 struct trace_event_call *tp_event = p_event->tp_event;
Jiri Olsaceec0b62012-02-15 15:51:49 +0100176 return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event);
177}
178
179static void perf_trace_event_close(struct perf_event *p_event)
180{
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400181 struct trace_event_call *tp_event = p_event->tp_event;
Jiri Olsaceec0b62012-02-15 15:51:49 +0100182 tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
183}
184
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400185static int perf_trace_event_init(struct trace_event_call *tp_event,
Jiri Olsaceec0b62012-02-15 15:51:49 +0100186 struct perf_event *p_event)
187{
188 int ret;
189
190 ret = perf_trace_event_perm(tp_event, p_event);
191 if (ret)
192 return ret;
193
194 ret = perf_trace_event_reg(tp_event, p_event);
195 if (ret)
196 return ret;
197
198 ret = perf_trace_event_open(p_event);
199 if (ret) {
200 perf_trace_event_unreg(p_event);
201 return ret;
202 }
203
204 return 0;
205}
206
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200207int perf_trace_init(struct perf_event *p_event)
Peter Zijlstraac199db2009-03-19 20:26:15 +0100208{
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400209 struct trace_event_call *tp_event;
Vince Weaver0022ced2013-11-15 12:39:45 -0500210 u64 event_id = p_event->attr.config;
Li Zefan20c89282009-05-06 10:33:45 +0800211 int ret = -EINVAL;
Peter Zijlstraac199db2009-03-19 20:26:15 +0100212
Li Zefan20c89282009-05-06 10:33:45 +0800213 mutex_lock(&event_mutex);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200214 list_for_each_entry(tp_event, &ftrace_events, list) {
Steven Rostedtff5f1492010-05-21 11:49:57 -0400215 if (tp_event->event.type == event_id &&
Steven Rostedta1d0ce82010-06-08 11:22:06 -0400216 tp_event->class && tp_event->class->reg &&
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200217 try_module_get(tp_event->mod)) {
218 ret = perf_trace_event_init(tp_event, p_event);
Li Zefan9cb627d2010-09-01 12:58:43 +0200219 if (ret)
220 module_put(tp_event->mod);
Li Zefan20c89282009-05-06 10:33:45 +0800221 break;
222 }
Peter Zijlstraac199db2009-03-19 20:26:15 +0100223 }
Li Zefan20c89282009-05-06 10:33:45 +0800224 mutex_unlock(&event_mutex);
Peter Zijlstraac199db2009-03-19 20:26:15 +0100225
Li Zefan20c89282009-05-06 10:33:45 +0800226 return ret;
Peter Zijlstraac199db2009-03-19 20:26:15 +0100227}
228
Jiri Olsaceec0b62012-02-15 15:51:49 +0100229void perf_trace_destroy(struct perf_event *p_event)
230{
231 mutex_lock(&event_mutex);
232 perf_trace_event_close(p_event);
233 perf_trace_event_unreg(p_event);
234 mutex_unlock(&event_mutex);
235}
236
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +0200237int perf_trace_add(struct perf_event *p_event, int flags)
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +0200238{
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400239 struct trace_event_call *tp_event = p_event->tp_event;
Namhyung Kim6016ee12010-08-11 12:47:59 +0900240 struct hlist_head __percpu *pcpu_list;
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200241 struct hlist_head *list;
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200242
Namhyung Kim6016ee12010-08-11 12:47:59 +0900243 pcpu_list = tp_event->perf_events;
244 if (WARN_ON_ONCE(!pcpu_list))
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200245 return -EINVAL;
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +0200246
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +0200247 if (!(flags & PERF_EF_START))
248 p_event->hw.state = PERF_HES_STOPPED;
249
Namhyung Kim6016ee12010-08-11 12:47:59 +0900250 list = this_cpu_ptr(pcpu_list);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200251 hlist_add_head_rcu(&p_event->hlist_entry, list);
Frederic Weisbecker20ab44252009-09-18 06:10:28 +0200252
Jiri Olsa489c75c2012-02-15 15:51:50 +0100253 return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
Frederic Weisbeckere5e25cf2009-09-18 00:54:43 +0200254}
255
Peter Zijlstraa4eaf7f2010-06-16 14:37:10 +0200256void perf_trace_del(struct perf_event *p_event, int flags)
Peter Zijlstraac199db2009-03-19 20:26:15 +0100257{
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400258 struct trace_event_call *tp_event = p_event->tp_event;
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200259 hlist_del_rcu(&p_event->hlist_entry);
Jiri Olsa489c75c2012-02-15 15:51:50 +0100260 tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event);
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200261}
Peter Zijlstraac199db2009-03-19 20:26:15 +0100262
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700263void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800264{
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200265 char *raw_data;
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700266 int rctx;
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800267
Frederic Weisbeckereb1e7962010-03-23 00:08:59 +0100268 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
269
Oleg Nesterovcd92bf62013-06-17 19:02:11 +0200270 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700271 "perf buffer not large enough"))
Oleg Nesterovcd92bf62013-06-17 19:02:11 +0200272 return NULL;
273
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700274 *rctxp = rctx = perf_swevent_get_recursion_context();
275 if (rctx < 0)
Peter Zijlstra1c024eca2010-05-19 14:02:22 +0200276 return NULL;
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800277
Peter Zijlstra (Intel)86038c52014-12-16 12:47:34 +0100278 if (regs)
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700279 *regs = this_cpu_ptr(&__perf_regs[rctx]);
280 raw_data = this_cpu_ptr(perf_trace_buf[rctx]);
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800281
282 /* zero the dead bytes from align to not leak stack to user */
Frederic Weisbeckereb1e7962010-03-23 00:08:59 +0100283 memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700284 return raw_data;
285}
286EXPORT_SYMBOL_GPL(perf_trace_buf_alloc);
287NOKPROBE_SYMBOL(perf_trace_buf_alloc);
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800288
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700289void perf_trace_buf_update(void *record, u16 type)
290{
291 struct trace_entry *entry = record;
292 int pc = preempt_count();
293 unsigned long flags;
294
Peter Zijlstra87f44bb2010-05-25 11:02:55 +0200295 local_save_flags(flags);
296 tracing_generic_entry_update(entry, flags, pc);
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800297 entry->type = type;
Xiao Guangrong430ad5a2010-01-28 09:32:29 +0800298}
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700299NOKPROBE_SYMBOL(perf_trace_buf_update);
Jiri Olsaced39002012-02-15 15:51:52 +0100300
301#ifdef CONFIG_FUNCTION_TRACER
302static void
Steven Rostedt2f5f6ad2011-08-08 16:57:47 -0400303perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
Steven Rostedta1e2e312011-08-09 12:50:46 -0400304 struct ftrace_ops *ops, struct pt_regs *pt_regs)
Jiri Olsaced39002012-02-15 15:51:52 +0100305{
306 struct ftrace_entry *entry;
307 struct hlist_head *head;
308 struct pt_regs regs;
309 int rctx;
310
Oleg Nesterovb8ebfd32013-06-17 19:02:04 +0200311 head = this_cpu_ptr(event_function.perf_events);
312 if (hlist_empty(head))
313 return;
314
Jiri Olsaced39002012-02-15 15:51:52 +0100315#define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
316 sizeof(u64)) - sizeof(u32))
317
318 BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
319
Alexei Starovoitovec5e099d2016-04-06 18:43:22 -0700320 memset(&regs, 0, sizeof(regs));
Jiri Olsaced39002012-02-15 15:51:52 +0100321 perf_fetch_caller_regs(&regs);
322
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700323 entry = perf_trace_buf_alloc(ENTRY_SIZE, NULL, &rctx);
Jiri Olsaced39002012-02-15 15:51:52 +0100324 if (!entry)
325 return;
326
327 entry->ip = ip;
328 entry->parent_ip = parent_ip;
Alexei Starovoitov1e1dcd92016-04-06 18:43:24 -0700329 perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
Andrew Vagine6dab5f2012-07-11 18:14:58 +0400330 1, &regs, head, NULL);
Jiri Olsaced39002012-02-15 15:51:52 +0100331
332#undef ENTRY_SIZE
333}
334
335static int perf_ftrace_function_register(struct perf_event *event)
336{
337 struct ftrace_ops *ops = &event->ftrace_ops;
338
Steven Rostedt (Red Hat)ba27f2b2015-11-30 17:23:39 -0500339 ops->flags |= FTRACE_OPS_FL_PER_CPU | FTRACE_OPS_FL_RCU;
Jiri Olsaced39002012-02-15 15:51:52 +0100340 ops->func = perf_ftrace_function_call;
341 return register_ftrace_function(ops);
342}
343
344static int perf_ftrace_function_unregister(struct perf_event *event)
345{
346 struct ftrace_ops *ops = &event->ftrace_ops;
Jiri Olsa5500fa52012-02-15 15:51:54 +0100347 int ret = unregister_ftrace_function(ops);
348 ftrace_free_filter(ops);
349 return ret;
Jiri Olsaced39002012-02-15 15:51:52 +0100350}
351
352static void perf_ftrace_function_enable(struct perf_event *event)
353{
354 ftrace_function_local_enable(&event->ftrace_ops);
355}
356
357static void perf_ftrace_function_disable(struct perf_event *event)
358{
359 ftrace_function_local_disable(&event->ftrace_ops);
360}
361
Steven Rostedt (Red Hat)2425bcb2015-05-05 11:45:27 -0400362int perf_ftrace_event_register(struct trace_event_call *call,
Jiri Olsaced39002012-02-15 15:51:52 +0100363 enum trace_reg type, void *data)
364{
365 switch (type) {
366 case TRACE_REG_REGISTER:
367 case TRACE_REG_UNREGISTER:
368 break;
369 case TRACE_REG_PERF_REGISTER:
370 case TRACE_REG_PERF_UNREGISTER:
371 return 0;
372 case TRACE_REG_PERF_OPEN:
373 return perf_ftrace_function_register(data);
374 case TRACE_REG_PERF_CLOSE:
375 return perf_ftrace_function_unregister(data);
376 case TRACE_REG_PERF_ADD:
377 perf_ftrace_function_enable(data);
378 return 0;
379 case TRACE_REG_PERF_DEL:
380 perf_ftrace_function_disable(data);
381 return 0;
382 }
383
384 return -EINVAL;
385}
386#endif /* CONFIG_FUNCTION_TRACER */