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 <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                  * 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
  74 static 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
 114 fail:
 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
 132 static 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         }
 157 out:
 158         module_put(tp_event->mod);
 159 }
 160
 161 static 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
 167 static 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
 173 static 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
 195 int 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
 217 void 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
 225 int 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
 244 void 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 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 }
 283 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);
 284 NOKPROBE_SYMBOL(perf_trace_buf_prepare);
 285
 286 #ifdef CONFIG_FUNCTION_TRACER
 287 static void
 288 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 289                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 290 {
 291         struct ftrace_entry *entry;
 292         struct hlist_head *head;
 293         struct pt_regs regs;
 294         int rctx;
 295
 296         head = this_cpu_ptr(event_function.perf_events);
 297         if (hlist_empty(head))
 298                 return;
 299
 300 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 301                     sizeof(u64)) - sizeof(u32))
 302
 303         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 304
 305         perf_fetch_caller_regs(&regs);
 306
 307         entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx);
 308         if (!entry)
 309                 return;
 310
 311         entry->ip = ip;
 312         entry->parent_ip = parent_ip;
 313         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0,
 314                               1, &regs, head, NULL);
 315
 316 #undef ENTRY_SIZE
 317 }
 318
 319 static int perf_ftrace_function_register(struct perf_event *event)
 320 {
 321         struct ftrace_ops *ops = &event->ftrace_ops;
 322
 323         ops->flags |= FTRACE_OPS_FL_CONTROL;
 324         ops->func = perf_ftrace_function_call;
 325         return register_ftrace_function(ops);
 326 }
 327
 328 static int perf_ftrace_function_unregister(struct perf_event *event)
 329 {
 330         struct ftrace_ops *ops = &event->ftrace_ops;
 331         int ret = unregister_ftrace_function(ops);
 332         ftrace_free_filter(ops);
 333         return ret;
 334 }
 335
 336 static void perf_ftrace_function_enable(struct perf_event *event)
 337 {
 338         ftrace_function_local_enable(&event->ftrace_ops);
 339 }
 340
 341 static void perf_ftrace_function_disable(struct perf_event *event)
 342 {
 343         ftrace_function_local_disable(&event->ftrace_ops);
 344 }
 345
 346 int perf_ftrace_event_register(struct ftrace_event_call *call,
 347                                enum trace_reg type, void *data)
 348 {
 349         switch (type) {
 350         case TRACE_REG_REGISTER:
 351         case TRACE_REG_UNREGISTER:
 352                 break;
 353         case TRACE_REG_PERF_REGISTER:
 354         case TRACE_REG_PERF_UNREGISTER:
 355                 return 0;
 356         case TRACE_REG_PERF_OPEN:
 357                 return perf_ftrace_function_register(data);
 358         case TRACE_REG_PERF_CLOSE:
 359                 return perf_ftrace_function_unregister(data);
 360         case TRACE_REG_PERF_ADD:
 361                 perf_ftrace_function_enable(data);
 362                 return 0;
 363         case TRACE_REG_PERF_DEL:
 364                 perf_ftrace_function_disable(data);
 365                 return 0;
 366         }
 367
 368         return -EINVAL;
 369 }
 370 #endif /* CONFIG_FUNCTION_TRACER */