| 1 | /* |
| 2 | * trace event based perf event profiling/tracing |
| 3 | * |
| 4 | * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra |
| 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 trace_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 | /* |
| 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 | |
| 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)) |
| 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; |
| 57 | |
| 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; |
| 64 | } |
| 65 | |
| 66 | /* 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 | |
| 86 | static int perf_trace_event_reg(struct trace_event_call *tp_event, |
| 87 | struct perf_event *p_event) |
| 88 | { |
| 89 | struct hlist_head __percpu *list; |
| 90 | int ret = -ENOMEM; |
| 91 | int cpu; |
| 92 | |
| 93 | p_event->tp_event = tp_event; |
| 94 | if (tp_event->perf_refcount++ > 0) |
| 95 | return 0; |
| 96 | |
| 97 | 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; |
| 105 | |
| 106 | if (!total_ref_count) { |
| 107 | char __percpu *buf; |
| 108 | int i; |
| 109 | |
| 110 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
| 111 | buf = (char __percpu *)alloc_percpu(perf_trace_t); |
| 112 | if (!buf) |
| 113 | goto fail; |
| 114 | |
| 115 | perf_trace_buf[i] = buf; |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL); |
| 120 | if (ret) |
| 121 | goto fail; |
| 122 | |
| 123 | total_ref_count++; |
| 124 | return 0; |
| 125 | |
| 126 | fail: |
| 127 | if (!total_ref_count) { |
| 128 | int i; |
| 129 | |
| 130 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
| 131 | free_percpu(perf_trace_buf[i]); |
| 132 | perf_trace_buf[i] = NULL; |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | if (!--tp_event->perf_refcount) { |
| 137 | free_percpu(tp_event->perf_events); |
| 138 | tp_event->perf_events = NULL; |
| 139 | } |
| 140 | |
| 141 | return ret; |
| 142 | } |
| 143 | |
| 144 | static void perf_trace_event_unreg(struct perf_event *p_event) |
| 145 | { |
| 146 | struct trace_event_call *tp_event = p_event->tp_event; |
| 147 | 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 | } |
| 169 | out: |
| 170 | module_put(tp_event->mod); |
| 171 | } |
| 172 | |
| 173 | static int perf_trace_event_open(struct perf_event *p_event) |
| 174 | { |
| 175 | struct trace_event_call *tp_event = p_event->tp_event; |
| 176 | return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event); |
| 177 | } |
| 178 | |
| 179 | static void perf_trace_event_close(struct perf_event *p_event) |
| 180 | { |
| 181 | struct trace_event_call *tp_event = p_event->tp_event; |
| 182 | tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event); |
| 183 | } |
| 184 | |
| 185 | static int perf_trace_event_init(struct trace_event_call *tp_event, |
| 186 | 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 | |
| 207 | int perf_trace_init(struct perf_event *p_event) |
| 208 | { |
| 209 | struct trace_event_call *tp_event; |
| 210 | u64 event_id = p_event->attr.config; |
| 211 | int ret = -EINVAL; |
| 212 | |
| 213 | mutex_lock(&event_mutex); |
| 214 | list_for_each_entry(tp_event, &ftrace_events, list) { |
| 215 | if (tp_event->event.type == event_id && |
| 216 | tp_event->class && tp_event->class->reg && |
| 217 | try_module_get(tp_event->mod)) { |
| 218 | ret = perf_trace_event_init(tp_event, p_event); |
| 219 | if (ret) |
| 220 | module_put(tp_event->mod); |
| 221 | break; |
| 222 | } |
| 223 | } |
| 224 | mutex_unlock(&event_mutex); |
| 225 | |
| 226 | return ret; |
| 227 | } |
| 228 | |
| 229 | void 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 | |
| 237 | int perf_trace_add(struct perf_event *p_event, int flags) |
| 238 | { |
| 239 | struct trace_event_call *tp_event = p_event->tp_event; |
| 240 | struct hlist_head __percpu *pcpu_list; |
| 241 | struct hlist_head *list; |
| 242 | |
| 243 | pcpu_list = tp_event->perf_events; |
| 244 | if (WARN_ON_ONCE(!pcpu_list)) |
| 245 | return -EINVAL; |
| 246 | |
| 247 | if (!(flags & PERF_EF_START)) |
| 248 | p_event->hw.state = PERF_HES_STOPPED; |
| 249 | |
| 250 | list = this_cpu_ptr(pcpu_list); |
| 251 | hlist_add_head_rcu(&p_event->hlist_entry, list); |
| 252 | |
| 253 | return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event); |
| 254 | } |
| 255 | |
| 256 | void perf_trace_del(struct perf_event *p_event, int flags) |
| 257 | { |
| 258 | struct trace_event_call *tp_event = p_event->tp_event; |
| 259 | hlist_del_rcu(&p_event->hlist_entry); |
| 260 | tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event); |
| 261 | } |
| 262 | |
| 263 | void *perf_trace_buf_prepare(int size, unsigned short type, |
| 264 | struct pt_regs **regs, int *rctxp) |
| 265 | { |
| 266 | struct trace_entry *entry; |
| 267 | unsigned long flags; |
| 268 | char *raw_data; |
| 269 | int pc; |
| 270 | |
| 271 | BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long)); |
| 272 | |
| 273 | if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE, |
| 274 | "perf buffer not large enough")) |
| 275 | return NULL; |
| 276 | |
| 277 | pc = preempt_count(); |
| 278 | |
| 279 | *rctxp = perf_swevent_get_recursion_context(); |
| 280 | if (*rctxp < 0) |
| 281 | return NULL; |
| 282 | |
| 283 | if (regs) |
| 284 | *regs = this_cpu_ptr(&__perf_regs[*rctxp]); |
| 285 | raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]); |
| 286 | |
| 287 | /* zero the dead bytes from align to not leak stack to user */ |
| 288 | memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64)); |
| 289 | |
| 290 | entry = (struct trace_entry *)raw_data; |
| 291 | local_save_flags(flags); |
| 292 | tracing_generic_entry_update(entry, flags, pc); |
| 293 | entry->type = type; |
| 294 | |
| 295 | return raw_data; |
| 296 | } |
| 297 | EXPORT_SYMBOL_GPL(perf_trace_buf_prepare); |
| 298 | NOKPROBE_SYMBOL(perf_trace_buf_prepare); |
| 299 | |
| 300 | #ifdef CONFIG_FUNCTION_TRACER |
| 301 | static void |
| 302 | perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip, |
| 303 | struct ftrace_ops *ops, struct pt_regs *pt_regs) |
| 304 | { |
| 305 | struct ftrace_entry *entry; |
| 306 | struct hlist_head *head; |
| 307 | struct pt_regs regs; |
| 308 | int rctx; |
| 309 | |
| 310 | head = this_cpu_ptr(event_function.perf_events); |
| 311 | if (hlist_empty(head)) |
| 312 | return; |
| 313 | |
| 314 | #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \ |
| 315 | sizeof(u64)) - sizeof(u32)) |
| 316 | |
| 317 | BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE); |
| 318 | |
| 319 | perf_fetch_caller_regs(®s); |
| 320 | |
| 321 | entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx); |
| 322 | if (!entry) |
| 323 | return; |
| 324 | |
| 325 | entry->ip = ip; |
| 326 | entry->parent_ip = parent_ip; |
| 327 | perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0, |
| 328 | 1, ®s, head, NULL); |
| 329 | |
| 330 | #undef ENTRY_SIZE |
| 331 | } |
| 332 | |
| 333 | static int perf_ftrace_function_register(struct perf_event *event) |
| 334 | { |
| 335 | struct ftrace_ops *ops = &event->ftrace_ops; |
| 336 | |
| 337 | ops->flags |= FTRACE_OPS_FL_PER_CPU | FTRACE_OPS_FL_RCU; |
| 338 | ops->func = perf_ftrace_function_call; |
| 339 | return register_ftrace_function(ops); |
| 340 | } |
| 341 | |
| 342 | static int perf_ftrace_function_unregister(struct perf_event *event) |
| 343 | { |
| 344 | struct ftrace_ops *ops = &event->ftrace_ops; |
| 345 | int ret = unregister_ftrace_function(ops); |
| 346 | ftrace_free_filter(ops); |
| 347 | return ret; |
| 348 | } |
| 349 | |
| 350 | static void perf_ftrace_function_enable(struct perf_event *event) |
| 351 | { |
| 352 | ftrace_function_local_enable(&event->ftrace_ops); |
| 353 | } |
| 354 | |
| 355 | static void perf_ftrace_function_disable(struct perf_event *event) |
| 356 | { |
| 357 | ftrace_function_local_disable(&event->ftrace_ops); |
| 358 | } |
| 359 | |
| 360 | int perf_ftrace_event_register(struct trace_event_call *call, |
| 361 | enum trace_reg type, void *data) |
| 362 | { |
| 363 | switch (type) { |
| 364 | case TRACE_REG_REGISTER: |
| 365 | case TRACE_REG_UNREGISTER: |
| 366 | break; |
| 367 | case TRACE_REG_PERF_REGISTER: |
| 368 | case TRACE_REG_PERF_UNREGISTER: |
| 369 | return 0; |
| 370 | case TRACE_REG_PERF_OPEN: |
| 371 | return perf_ftrace_function_register(data); |
| 372 | case TRACE_REG_PERF_CLOSE: |
| 373 | return perf_ftrace_function_unregister(data); |
| 374 | case TRACE_REG_PERF_ADD: |
| 375 | perf_ftrace_function_enable(data); |
| 376 | return 0; |
| 377 | case TRACE_REG_PERF_DEL: |
| 378 | perf_ftrace_function_disable(data); |
| 379 | return 0; |
| 380 | } |
| 381 | |
| 382 | return -EINVAL; |
| 383 | } |
| 384 | #endif /* CONFIG_FUNCTION_TRACER */ |