kernel/trace/trace_event_perf.c

   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                 if (!is_sampling_event(p_event))
  51                         return 0;
  52
  53                 /*
  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.
  57                  */
  58                 if (!p_event->attr.exclude_callchain_user)
  59                         return -EINVAL;
  60
  61                 /*
  62                  * Same reason to disable user stack dump as for user space
  63                  * callchains above.
  64                  */
  65                 if (p_event->attr.sample_type & PERF_SAMPLE_STACK_USER)
  66                         return -EINVAL;
  67         }
  68
  69         /* No tracing, just counting, so no obvious leak */
  70         if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
  71                 return 0;
  72
  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)
  76                         return 0;
  77         }
  78
  79         /*
  80          * ...otherwise raw tracepoint data can be a severe data leak,
  81          * only allow root to have these.
  82          */
  83         if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  84                 return -EPERM;
  85
  86         return 0;
  87 }
  88
  89 static int perf_trace_event_reg(struct trace_event_call *tp_event,
  90                                 struct perf_event *p_event)
  91 {
  92         struct hlist_head __percpu *list;
  93         int ret = -ENOMEM;
  94         int cpu;
  95
  96         p_event->tp_event = tp_event;
  97         if (tp_event->perf_refcount++ > 0)
  98                 return 0;
  99
 100         list = alloc_percpu(struct hlist_head);
 101         if (!list)
 102                 goto fail;
 103
 104         for_each_possible_cpu(cpu)
 105                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
 106
 107         tp_event->perf_events = list;
 108
 109         if (!total_ref_count) {
 110                 char __percpu *buf;
 111                 int i;
 112
 113                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 114                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
 115                         if (!buf)
 116                                 goto fail;
 117
 118                         perf_trace_buf[i] = buf;
 119                 }
 120         }
 121
 122         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
 123         if (ret)
 124                 goto fail;
 125
 126         total_ref_count++;
 127         return 0;
 128
 129 fail:
 130         if (!total_ref_count) {
 131                 int i;
 132
 133                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 134                         free_percpu(perf_trace_buf[i]);
 135                         perf_trace_buf[i] = NULL;
 136                 }
 137         }
 138
 139         if (!--tp_event->perf_refcount) {
 140                 free_percpu(tp_event->perf_events);
 141                 tp_event->perf_events = NULL;
 142         }
 143
 144         return ret;
 145 }
 146
 147 static void perf_trace_event_unreg(struct perf_event *p_event)
 148 {
 149         struct trace_event_call *tp_event = p_event->tp_event;
 150         int i;
 151
 152         if (--tp_event->perf_refcount > 0)
 153                 goto out;
 154
 155         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
 156
 157         /*
 158          * Ensure our callback won't be called anymore. The buffers
 159          * will be freed after that.
 160          */
 161         tracepoint_synchronize_unregister();
 162
 163         free_percpu(tp_event->perf_events);
 164         tp_event->perf_events = NULL;
 165
 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;
 170                 }
 171         }
 172 out:
 173         module_put(tp_event->mod);
 174 }
 175
 176 static int perf_trace_event_open(struct perf_event *p_event)
 177 {
 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);
 180 }
 181
 182 static void perf_trace_event_close(struct perf_event *p_event)
 183 {
 184         struct trace_event_call *tp_event = p_event->tp_event;
 185         tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
 186 }
 187
 188 static int perf_trace_event_init(struct trace_event_call *tp_event,
 189                                  struct perf_event *p_event)
 190 {
 191         int ret;
 192
 193         ret = perf_trace_event_perm(tp_event, p_event);
 194         if (ret)
 195                 return ret;
 196
 197         ret = perf_trace_event_reg(tp_event, p_event);
 198         if (ret)
 199                 return ret;
 200
 201         ret = perf_trace_event_open(p_event);
 202         if (ret) {
 203                 perf_trace_event_unreg(p_event);
 204                 return ret;
 205         }
 206
 207         return 0;
 208 }
 209
 210 int perf_trace_init(struct perf_event *p_event)
 211 {
 212         struct trace_event_call *tp_event;
 213         u64 event_id = p_event->attr.config;
 214         int ret = -EINVAL;
 215
 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);
 222                         if (ret)
 223                                 module_put(tp_event->mod);
 224                         break;
 225                 }
 226         }
 227         mutex_unlock(&event_mutex);
 228
 229         return ret;
 230 }
 231
 232 void perf_trace_destroy(struct perf_event *p_event)
 233 {
 234         mutex_lock(&event_mutex);
 235         perf_trace_event_close(p_event);
 236         perf_trace_event_unreg(p_event);
 237         mutex_unlock(&event_mutex);
 238 }
 239
 240 int perf_trace_add(struct perf_event *p_event, int flags)
 241 {
 242         struct trace_event_call *tp_event = p_event->tp_event;
 243         struct hlist_head __percpu *pcpu_list;
 244         struct hlist_head *list;
 245
 246         pcpu_list = tp_event->perf_events;
 247         if (WARN_ON_ONCE(!pcpu_list))
 248                 return -EINVAL;
 249
 250         if (!(flags & PERF_EF_START))
 251                 p_event->hw.state = PERF_HES_STOPPED;
 252
 253         list = this_cpu_ptr(pcpu_list);
 254         hlist_add_head_rcu(&p_event->hlist_entry, list);
 255
 256         return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
 257 }
 258
 259 void perf_trace_del(struct perf_event *p_event, int flags)
 260 {
 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);
 264 }
 265
 266 void *perf_trace_buf_prepare(int size, unsigned short type,
 267                              struct pt_regs **regs, int *rctxp)
 268 {
 269         struct trace_entry *entry;
 270         unsigned long flags;
 271         char *raw_data;
 272         int pc;
 273
 274         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 275
 276         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 277                         "perf buffer not large enough"))
 278                 return NULL;
 279
 280         pc = preempt_count();
 281
 282         *rctxp = perf_swevent_get_recursion_context();
 283         if (*rctxp < 0)
 284                 return NULL;
 285
 286         if (regs)
 287                 *regs = this_cpu_ptr(&__perf_regs[*rctxp]);
 288         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
 289
 290         /* zero the dead bytes from align to not leak stack to user */
 291         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 292
 293         entry = (struct trace_entry *)raw_data;
 294         local_save_flags(flags);
 295         tracing_generic_entry_update(entry, flags, pc);
 296         entry->type = type;
 297
 298         return raw_data;
 299 }
 300 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);
 301 NOKPROBE_SYMBOL(perf_trace_buf_prepare);
 302
 303 #ifdef CONFIG_FUNCTION_TRACER
 304 static void
 305 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 306                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 307 {
 308         struct ftrace_entry *entry;
 309         struct hlist_head *head;
 310         struct pt_regs regs;
 311         int rctx;
 312
 313         head = this_cpu_ptr(event_function.perf_events);
 314         if (hlist_empty(head))
 315                 return;
 316
 317 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 318                     sizeof(u64)) - sizeof(u32))
 319
 320         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 321
 322         perf_fetch_caller_regs(&regs);
 323
 324         entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx);
 325         if (!entry)
 326                 return;
 327
 328         entry->ip = ip;
 329         entry->parent_ip = parent_ip;
 330         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0,
 331                               1, &regs, head, NULL);
 332
 333 #undef ENTRY_SIZE
 334 }
 335
 336 static int perf_ftrace_function_register(struct perf_event *event)
 337 {
 338         struct ftrace_ops *ops = &event->ftrace_ops;
 339
 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);
 343 }
 344
 345 static int perf_ftrace_function_unregister(struct perf_event *event)
 346 {
 347         struct ftrace_ops *ops = &event->ftrace_ops;
 348         int ret = unregister_ftrace_function(ops);
 349         ftrace_free_filter(ops);
 350         return ret;
 351 }
 352
 353 static void perf_ftrace_function_enable(struct perf_event *event)
 354 {
 355         ftrace_function_local_enable(&event->ftrace_ops);
 356 }
 357
 358 static void perf_ftrace_function_disable(struct perf_event *event)
 359 {
 360         ftrace_function_local_disable(&event->ftrace_ops);
 361 }
 362
 363 int perf_ftrace_event_register(struct trace_event_call *call,
 364                                enum trace_reg type, void *data)
 365 {
 366         switch (type) {
 367         case TRACE_REG_REGISTER:
 368         case TRACE_REG_UNREGISTER:
 369                 break;
 370         case TRACE_REG_PERF_REGISTER:
 371         case TRACE_REG_PERF_UNREGISTER:
 372                 return 0;
 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);
 379                 return 0;
 380         case TRACE_REG_PERF_DEL:
 381                 perf_ftrace_function_disable(data);
 382                 return 0;
 383         }
 384
 385         return -EINVAL;
 386 }
 387 #endif /* CONFIG_FUNCTION_TRACER */