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         /* The ftrace function trace is allowed only for root. */
  28         if (ftrace_event_is_function(tp_event) &&
  29             perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  30                 return -EPERM;
  31
  32         /* No tracing, just counting, so no obvious leak */
  33         if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
  34                 return 0;
  35
  36         /* Some events are ok to be traced by non-root users... */
  37         if (p_event->attach_state == PERF_ATTACH_TASK) {
  38                 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
  39                         return 0;
  40         }
  41
  42         /*
  43          * ...otherwise raw tracepoint data can be a severe data leak,
  44          * only allow root to have these.
  45          */
  46         if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  47                 return -EPERM;
  48
  49         return 0;
  50 }
  51
  52 static int perf_trace_event_reg(struct ftrace_event_call *tp_event,
  53                                 struct perf_event *p_event)
  54 {
  55         struct hlist_head __percpu *list;
  56         int ret = -ENOMEM;
  57         int cpu;
  58
  59         p_event->tp_event = tp_event;
  60         if (tp_event->perf_refcount++ > 0)
  61                 return 0;
  62
  63         list = alloc_percpu(struct hlist_head);
  64         if (!list)
  65                 goto fail;
  66
  67         for_each_possible_cpu(cpu)
  68                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
  69
  70         tp_event->perf_events = list;
  71
  72         if (!total_ref_count) {
  73                 char __percpu *buf;
  74                 int i;
  75
  76                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
  77                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
  78                         if (!buf)
  79                                 goto fail;
  80
  81                         perf_trace_buf[i] = buf;
  82                 }
  83         }
  84
  85         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
  86         if (ret)
  87                 goto fail;
  88
  89         total_ref_count++;
  90         return 0;
  91
  92 fail:
  93         if (!total_ref_count) {
  94                 int i;
  95
  96                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
  97                         free_percpu(perf_trace_buf[i]);
  98                         perf_trace_buf[i] = NULL;
  99                 }
 100         }
 101
 102         if (!--tp_event->perf_refcount) {
 103                 free_percpu(tp_event->perf_events);
 104                 tp_event->perf_events = NULL;
 105         }
 106
 107         return ret;
 108 }
 109
 110 static void perf_trace_event_unreg(struct perf_event *p_event)
 111 {
 112         struct ftrace_event_call *tp_event = p_event->tp_event;
 113         int i;
 114
 115         if (--tp_event->perf_refcount > 0)
 116                 goto out;
 117
 118         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
 119
 120         /*
 121          * Ensure our callback won't be called anymore. The buffers
 122          * will be freed after that.
 123          */
 124         tracepoint_synchronize_unregister();
 125
 126         free_percpu(tp_event->perf_events);
 127         tp_event->perf_events = NULL;
 128
 129         if (!--total_ref_count) {
 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 out:
 136         module_put(tp_event->mod);
 137 }
 138
 139 static int perf_trace_event_open(struct perf_event *p_event)
 140 {
 141         struct ftrace_event_call *tp_event = p_event->tp_event;
 142         return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event);
 143 }
 144
 145 static void perf_trace_event_close(struct perf_event *p_event)
 146 {
 147         struct ftrace_event_call *tp_event = p_event->tp_event;
 148         tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
 149 }
 150
 151 static int perf_trace_event_init(struct ftrace_event_call *tp_event,
 152                                  struct perf_event *p_event)
 153 {
 154         int ret;
 155
 156         ret = perf_trace_event_perm(tp_event, p_event);
 157         if (ret)
 158                 return ret;
 159
 160         ret = perf_trace_event_reg(tp_event, p_event);
 161         if (ret)
 162                 return ret;
 163
 164         ret = perf_trace_event_open(p_event);
 165         if (ret) {
 166                 perf_trace_event_unreg(p_event);
 167                 return ret;
 168         }
 169
 170         return 0;
 171 }
 172
 173 int perf_trace_init(struct perf_event *p_event)
 174 {
 175         struct ftrace_event_call *tp_event;
 176         int event_id = p_event->attr.config;
 177         int ret = -EINVAL;
 178
 179         mutex_lock(&event_mutex);
 180         list_for_each_entry(tp_event, &ftrace_events, list) {
 181                 if (tp_event->event.type == event_id &&
 182                     tp_event->class && tp_event->class->reg &&
 183                     try_module_get(tp_event->mod)) {
 184                         ret = perf_trace_event_init(tp_event, p_event);
 185                         if (ret)
 186                                 module_put(tp_event->mod);
 187                         break;
 188                 }
 189         }
 190         mutex_unlock(&event_mutex);
 191
 192         return ret;
 193 }
 194
 195 void perf_trace_destroy(struct perf_event *p_event)
 196 {
 197         mutex_lock(&event_mutex);
 198         perf_trace_event_close(p_event);
 199         perf_trace_event_unreg(p_event);
 200         mutex_unlock(&event_mutex);
 201 }
 202
 203 int perf_trace_add(struct perf_event *p_event, int flags)
 204 {
 205         struct ftrace_event_call *tp_event = p_event->tp_event;
 206         struct hlist_head __percpu *pcpu_list;
 207         struct hlist_head *list;
 208
 209         pcpu_list = tp_event->perf_events;
 210         if (WARN_ON_ONCE(!pcpu_list))
 211                 return -EINVAL;
 212
 213         if (!(flags & PERF_EF_START))
 214                 p_event->hw.state = PERF_HES_STOPPED;
 215
 216         list = this_cpu_ptr(pcpu_list);
 217         hlist_add_head_rcu(&p_event->hlist_entry, list);
 218
 219         return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
 220 }
 221
 222 void perf_trace_del(struct perf_event *p_event, int flags)
 223 {
 224         struct ftrace_event_call *tp_event = p_event->tp_event;
 225         hlist_del_rcu(&p_event->hlist_entry);
 226         tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event);
 227 }
 228
 229 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 230                                        struct pt_regs *regs, int *rctxp)
 231 {
 232         struct trace_entry *entry;
 233         unsigned long flags;
 234         char *raw_data;
 235         int pc;
 236
 237         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 238
 239         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 240                         "perf buffer not large enough"))
 241                 return NULL;
 242
 243         pc = preempt_count();
 244
 245         *rctxp = perf_swevent_get_recursion_context();
 246         if (*rctxp < 0)
 247                 return NULL;
 248
 249         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
 250
 251         /* zero the dead bytes from align to not leak stack to user */
 252         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 253
 254         entry = (struct trace_entry *)raw_data;
 255         local_save_flags(flags);
 256         tracing_generic_entry_update(entry, flags, pc);
 257         entry->type = type;
 258
 259         return raw_data;
 260 }
 261 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);
 262
 263 #ifdef CONFIG_FUNCTION_TRACER
 264 static void
 265 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 266                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 267 {
 268         struct ftrace_entry *entry;
 269         struct hlist_head *head;
 270         struct pt_regs regs;
 271         int rctx;
 272
 273         head = this_cpu_ptr(event_function.perf_events);
 274         if (hlist_empty(head))
 275                 return;
 276
 277 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 278                     sizeof(u64)) - sizeof(u32))
 279
 280         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 281
 282         perf_fetch_caller_regs(&regs);
 283
 284         entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx);
 285         if (!entry)
 286                 return;
 287
 288         entry->ip = ip;
 289         entry->parent_ip = parent_ip;
 290         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0,
 291                               1, &regs, head, NULL);
 292
 293 #undef ENTRY_SIZE
 294 }
 295
 296 static int perf_ftrace_function_register(struct perf_event *event)
 297 {
 298         struct ftrace_ops *ops = &event->ftrace_ops;
 299
 300         ops->flags |= FTRACE_OPS_FL_CONTROL;
 301         ops->func = perf_ftrace_function_call;
 302         return register_ftrace_function(ops);
 303 }
 304
 305 static int perf_ftrace_function_unregister(struct perf_event *event)
 306 {
 307         struct ftrace_ops *ops = &event->ftrace_ops;
 308         int ret = unregister_ftrace_function(ops);
 309         ftrace_free_filter(ops);
 310         return ret;
 311 }
 312
 313 static void perf_ftrace_function_enable(struct perf_event *event)
 314 {
 315         ftrace_function_local_enable(&event->ftrace_ops);
 316 }
 317
 318 static void perf_ftrace_function_disable(struct perf_event *event)
 319 {
 320         ftrace_function_local_disable(&event->ftrace_ops);
 321 }
 322
 323 int perf_ftrace_event_register(struct ftrace_event_call *call,
 324                                enum trace_reg type, void *data)
 325 {
 326         switch (type) {
 327         case TRACE_REG_REGISTER:
 328         case TRACE_REG_UNREGISTER:
 329                 break;
 330         case TRACE_REG_PERF_REGISTER:
 331         case TRACE_REG_PERF_UNREGISTER:
 332                 return 0;
 333         case TRACE_REG_PERF_OPEN:
 334                 return perf_ftrace_function_register(data);
 335         case TRACE_REG_PERF_CLOSE:
 336                 return perf_ftrace_function_unregister(data);
 337         case TRACE_REG_PERF_ADD:
 338                 perf_ftrace_function_enable(data);
 339                 return 0;
 340         case TRACE_REG_PERF_DEL:
 341                 perf_ftrace_function_disable(data);
 342                 return 0;
 343         }
 344
 345         return -EINVAL;
 346 }
 347 #endif /* CONFIG_FUNCTION_TRACER */