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 12 static 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 */ 18 typedef 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) */ 22 static int total_ref_count; 23 24 static 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 /* No tracing, just counting, so no obvious leak */ 48 if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW)) 49 return 0; 50 51 /* Some events are ok to be traced by non-root users... */ 52 if (p_event->attach_state == PERF_ATTACH_TASK) { 53 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY) 54 return 0; 55 } 56 57 /* 58 * ...otherwise raw tracepoint data can be a severe data leak, 59 * only allow root to have these. 60 */ 61 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN)) 62 return -EPERM; 63 64 return 0; 65 } 66 67 static int perf_trace_event_reg(struct ftrace_event_call *tp_event, 68 struct perf_event *p_event) 69 { 70 struct hlist_head __percpu *list; 71 int ret = -ENOMEM; 72 int cpu; 73 74 p_event->tp_event = tp_event; 75 if (tp_event->perf_refcount++ > 0) 76 return 0; 77 78 list = alloc_percpu(struct hlist_head); 79 if (!list) 80 goto fail; 81 82 for_each_possible_cpu(cpu) 83 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu)); 84 85 tp_event->perf_events = list; 86 87 if (!total_ref_count) { 88 char __percpu *buf; 89 int i; 90 91 for (i = 0; i < PERF_NR_CONTEXTS; i++) { 92 buf = (char __percpu *)alloc_percpu(perf_trace_t); 93 if (!buf) 94 goto fail; 95 96 perf_trace_buf[i] = buf; 97 } 98 } 99 100 ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL); 101 if (ret) 102 goto fail; 103 104 total_ref_count++; 105 return 0; 106 107 fail: 108 if (!total_ref_count) { 109 int i; 110 111 for (i = 0; i < PERF_NR_CONTEXTS; i++) { 112 free_percpu(perf_trace_buf[i]); 113 perf_trace_buf[i] = NULL; 114 } 115 } 116 117 if (!--tp_event->perf_refcount) { 118 free_percpu(tp_event->perf_events); 119 tp_event->perf_events = NULL; 120 } 121 122 return ret; 123 } 124 125 static void perf_trace_event_unreg(struct perf_event *p_event) 126 { 127 struct ftrace_event_call *tp_event = p_event->tp_event; 128 int i; 129 130 if (--tp_event->perf_refcount > 0) 131 goto out; 132 133 tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL); 134 135 /* 136 * Ensure our callback won't be called anymore. The buffers 137 * will be freed after that. 138 */ 139 tracepoint_synchronize_unregister(); 140 141 free_percpu(tp_event->perf_events); 142 tp_event->perf_events = NULL; 143 144 if (!--total_ref_count) { 145 for (i = 0; i < PERF_NR_CONTEXTS; i++) { 146 free_percpu(perf_trace_buf[i]); 147 perf_trace_buf[i] = NULL; 148 } 149 } 150 out: 151 module_put(tp_event->mod); 152 } 153 154 static int perf_trace_event_open(struct perf_event *p_event) 155 { 156 struct ftrace_event_call *tp_event = p_event->tp_event; 157 return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event); 158 } 159 160 static void perf_trace_event_close(struct perf_event *p_event) 161 { 162 struct ftrace_event_call *tp_event = p_event->tp_event; 163 tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event); 164 } 165 166 static int perf_trace_event_init(struct ftrace_event_call *tp_event, 167 struct perf_event *p_event) 168 { 169 int ret; 170 171 ret = perf_trace_event_perm(tp_event, p_event); 172 if (ret) 173 return ret; 174 175 ret = perf_trace_event_reg(tp_event, p_event); 176 if (ret) 177 return ret; 178 179 ret = perf_trace_event_open(p_event); 180 if (ret) { 181 perf_trace_event_unreg(p_event); 182 return ret; 183 } 184 185 return 0; 186 } 187 188 int perf_trace_init(struct perf_event *p_event) 189 { 190 struct ftrace_event_call *tp_event; 191 u64 event_id = p_event->attr.config; 192 int ret = -EINVAL; 193 194 mutex_lock(&event_mutex); 195 list_for_each_entry(tp_event, &ftrace_events, list) { 196 if (tp_event->event.type == event_id && 197 tp_event->class && tp_event->class->reg && 198 try_module_get(tp_event->mod)) { 199 ret = perf_trace_event_init(tp_event, p_event); 200 if (ret) 201 module_put(tp_event->mod); 202 break; 203 } 204 } 205 mutex_unlock(&event_mutex); 206 207 return ret; 208 } 209 210 void perf_trace_destroy(struct perf_event *p_event) 211 { 212 mutex_lock(&event_mutex); 213 perf_trace_event_close(p_event); 214 perf_trace_event_unreg(p_event); 215 mutex_unlock(&event_mutex); 216 } 217 218 int perf_trace_add(struct perf_event *p_event, int flags) 219 { 220 struct ftrace_event_call *tp_event = p_event->tp_event; 221 struct hlist_head __percpu *pcpu_list; 222 struct hlist_head *list; 223 224 pcpu_list = tp_event->perf_events; 225 if (WARN_ON_ONCE(!pcpu_list)) 226 return -EINVAL; 227 228 if (!(flags & PERF_EF_START)) 229 p_event->hw.state = PERF_HES_STOPPED; 230 231 list = this_cpu_ptr(pcpu_list); 232 hlist_add_head_rcu(&p_event->hlist_entry, list); 233 234 return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event); 235 } 236 237 void perf_trace_del(struct perf_event *p_event, int flags) 238 { 239 struct ftrace_event_call *tp_event = p_event->tp_event; 240 hlist_del_rcu(&p_event->hlist_entry); 241 tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event); 242 } 243 244 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type, 245 struct pt_regs *regs, int *rctxp) 246 { 247 struct trace_entry *entry; 248 unsigned long flags; 249 char *raw_data; 250 int pc; 251 252 BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long)); 253 254 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE, 255 "perf buffer not large enough")) 256 return NULL; 257 258 pc = preempt_count(); 259 260 *rctxp = perf_swevent_get_recursion_context(); 261 if (*rctxp < 0) 262 return NULL; 263 264 raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]); 265 266 /* zero the dead bytes from align to not leak stack to user */ 267 memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64)); 268 269 entry = (struct trace_entry *)raw_data; 270 local_save_flags(flags); 271 tracing_generic_entry_update(entry, flags, pc); 272 entry->type = type; 273 274 return raw_data; 275 } 276 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare); 277 278 #ifdef CONFIG_FUNCTION_TRACER 279 static void 280 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip, 281 struct ftrace_ops *ops, struct pt_regs *pt_regs) 282 { 283 struct ftrace_entry *entry; 284 struct hlist_head *head; 285 struct pt_regs regs; 286 int rctx; 287 288 head = this_cpu_ptr(event_function.perf_events); 289 if (hlist_empty(head)) 290 return; 291 292 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \ 293 sizeof(u64)) - sizeof(u32)) 294 295 BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE); 296 297 perf_fetch_caller_regs(®s); 298 299 entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx); 300 if (!entry) 301 return; 302 303 entry->ip = ip; 304 entry->parent_ip = parent_ip; 305 perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0, 306 1, ®s, head, NULL); 307 308 #undef ENTRY_SIZE 309 } 310 311 static int perf_ftrace_function_register(struct perf_event *event) 312 { 313 struct ftrace_ops *ops = &event->ftrace_ops; 314 315 ops->flags |= FTRACE_OPS_FL_CONTROL; 316 ops->func = perf_ftrace_function_call; 317 return register_ftrace_function(ops); 318 } 319 320 static int perf_ftrace_function_unregister(struct perf_event *event) 321 { 322 struct ftrace_ops *ops = &event->ftrace_ops; 323 int ret = unregister_ftrace_function(ops); 324 ftrace_free_filter(ops); 325 return ret; 326 } 327 328 static void perf_ftrace_function_enable(struct perf_event *event) 329 { 330 ftrace_function_local_enable(&event->ftrace_ops); 331 } 332 333 static void perf_ftrace_function_disable(struct perf_event *event) 334 { 335 ftrace_function_local_disable(&event->ftrace_ops); 336 } 337 338 int perf_ftrace_event_register(struct ftrace_event_call *call, 339 enum trace_reg type, void *data) 340 { 341 switch (type) { 342 case TRACE_REG_REGISTER: 343 case TRACE_REG_UNREGISTER: 344 break; 345 case TRACE_REG_PERF_REGISTER: 346 case TRACE_REG_PERF_UNREGISTER: 347 return 0; 348 case TRACE_REG_PERF_OPEN: 349 return perf_ftrace_function_register(data); 350 case TRACE_REG_PERF_CLOSE: 351 return perf_ftrace_function_unregister(data); 352 case TRACE_REG_PERF_ADD: 353 perf_ftrace_function_enable(data); 354 return 0; 355 case TRACE_REG_PERF_DEL: 356 perf_ftrace_function_disable(data); 357 return 0; 358 } 359 360 return -EINVAL; 361 } 362 #endif /* CONFIG_FUNCTION_TRACER */ 363