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
 244         if (!(flags & PERF_EF_START))
 245                 p_event->hw.state = PERF_HES_STOPPED;
 246
 247         /*
 248          * If TRACE_REG_PERF_ADD returns false; no custom action was performed
 249          * and we need to take the default action of enqueueing our event on
 250          * the right per-cpu hlist.
 251          */
 252         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
 253                 struct hlist_head __percpu *pcpu_list;
 254                 struct hlist_head *list;
 255
 256                 pcpu_list = tp_event->perf_events;
 257                 if (WARN_ON_ONCE(!pcpu_list))
 258                         return -EINVAL;
 259
 260                 list = this_cpu_ptr(pcpu_list);
 261                 hlist_add_head_rcu(&p_event->hlist_entry, list);
 262         }
 263
 264         return 0;
 265 }
 266
 267 void perf_trace_del(struct perf_event *p_event, int flags)
 268 {
 269         struct trace_event_call *tp_event = p_event->tp_event;
 270
 271         /*
 272          * If TRACE_REG_PERF_DEL returns false; no custom action was performed
 273          * and we need to take the default action of dequeueing our event from
 274          * the right per-cpu hlist.
 275          */
 276         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
 277                 hlist_del_rcu(&p_event->hlist_entry);
 278 }
 279
 280 void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
 281 {
 282         char *raw_data;
 283         int rctx;
 284
 285         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 286
 287         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 288                       "perf buffer not large enough"))
 289                 return NULL;
 290
 291         *rctxp = rctx = perf_swevent_get_recursion_context();
 292         if (rctx < 0)
 293                 return NULL;
 294
 295         if (regs)
 296                 *regs = this_cpu_ptr(&__perf_regs[rctx]);
 297         raw_data = this_cpu_ptr(perf_trace_buf[rctx]);
 298
 299         /* zero the dead bytes from align to not leak stack to user */
 300         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 301         return raw_data;
 302 }
 303 EXPORT_SYMBOL_GPL(perf_trace_buf_alloc);
 304 NOKPROBE_SYMBOL(perf_trace_buf_alloc);
 305
 306 void perf_trace_buf_update(void *record, u16 type)
 307 {
 308         struct trace_entry *entry = record;
 309         int pc = preempt_count();
 310         unsigned long flags;
 311
 312         local_save_flags(flags);
 313         tracing_generic_entry_update(entry, flags, pc);
 314         entry->type = type;
 315 }
 316 NOKPROBE_SYMBOL(perf_trace_buf_update);
 317
 318 #ifdef CONFIG_FUNCTION_TRACER
 319 static void
 320 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 321                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 322 {
 323         struct ftrace_entry *entry;
 324         struct perf_event *event;
 325         struct hlist_head head;
 326         struct pt_regs regs;
 327         int rctx;
 328
 329         if ((unsigned long)ops->private != smp_processor_id())
 330                 return;
 331
 332         event = container_of(ops, struct perf_event, ftrace_ops);
 333
 334         /*
 335          * @event->hlist entry is NULL (per INIT_HLIST_NODE), and all
 336          * the perf code does is hlist_for_each_entry_rcu(), so we can
 337          * get away with simply setting the @head.first pointer in order
 338          * to create a singular list.
 339          */
 340         head.first = &event->hlist_entry;
 341
 342 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 343                     sizeof(u64)) - sizeof(u32))
 344
 345         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 346
 347         memset(&regs, 0, sizeof(regs));
 348         perf_fetch_caller_regs(&regs);
 349
 350         entry = perf_trace_buf_alloc(ENTRY_SIZE, NULL, &rctx);
 351         if (!entry)
 352                 return;
 353
 354         entry->ip = ip;
 355         entry->parent_ip = parent_ip;
 356         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
 357                               1, &regs, &head, NULL);
 358
 359 #undef ENTRY_SIZE
 360 }
 361
 362 static int perf_ftrace_function_register(struct perf_event *event)
 363 {
 364         struct ftrace_ops *ops = &event->ftrace_ops;
 365
 366         ops->flags   = FTRACE_OPS_FL_RCU;
 367         ops->func    = perf_ftrace_function_call;
 368         ops->private = (void *)(unsigned long)nr_cpu_ids;
 369
 370         return register_ftrace_function(ops);
 371 }
 372
 373 static int perf_ftrace_function_unregister(struct perf_event *event)
 374 {
 375         struct ftrace_ops *ops = &event->ftrace_ops;
 376         int ret = unregister_ftrace_function(ops);
 377         ftrace_free_filter(ops);
 378         return ret;
 379 }
 380
 381 int perf_ftrace_event_register(struct trace_event_call *call,
 382                                enum trace_reg type, void *data)
 383 {
 384         struct perf_event *event = data;
 385
 386         switch (type) {
 387         case TRACE_REG_REGISTER:
 388         case TRACE_REG_UNREGISTER:
 389                 break;
 390         case TRACE_REG_PERF_REGISTER:
 391         case TRACE_REG_PERF_UNREGISTER:
 392                 return 0;
 393         case TRACE_REG_PERF_OPEN:
 394                 return perf_ftrace_function_register(data);
 395         case TRACE_REG_PERF_CLOSE:
 396                 return perf_ftrace_function_unregister(data);
 397         case TRACE_REG_PERF_ADD:
 398                 event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();
 399                 return 1;
 400         case TRACE_REG_PERF_DEL:
 401                 event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
 402                 return 1;
 403         }
 404
 405         return -EINVAL;
 406 }
 407 #endif /* CONFIG_FUNCTION_TRACER */