git.oblomov.eu Git - linux-2.6/blob - kernel/trace/trace_sched_wakeup.c

   1 /*
   2  * trace task wakeup timings
   3  *
   4  * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
   5  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
   6  *
   7  * Based on code from the latency_tracer, that is:
   8  *
   9  *  Copyright (C) 2004-2006 Ingo Molnar
  10  *  Copyright (C) 2004 William Lee Irwin III
  11  */
  12 #include <linux/module.h>
  13 #include <linux/fs.h>
  14 #include <linux/debugfs.h>
  15 #include <linux/kallsyms.h>
  16 #include <linux/uaccess.h>
  17 #include <linux/ftrace.h>
  18 #include <linux/marker.h>
  19
  20 #include "trace.h"
  21
  22 static struct trace_array       *wakeup_trace;
  23 static int __read_mostly        tracer_enabled;
  24
  25 static struct task_struct       *wakeup_task;
  26 static int                      wakeup_cpu;
  27 static unsigned                 wakeup_prio = -1;
  28
  29 static DEFINE_SPINLOCK(wakeup_lock);
  30
  31 static void __wakeup_reset(struct trace_array *tr);
  32
  33 #ifdef CONFIG_FTRACE
  34 /*
  35  * irqsoff uses its own tracer function to keep the overhead down:
  36  */
  37 static void
  38 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip)
  39 {
  40         struct trace_array *tr = wakeup_trace;
  41         struct trace_array_cpu *data;
  42         unsigned long flags;
  43         long disabled;
  44         int resched;
  45         int cpu;
  46
  47         if (likely(!wakeup_task))
  48                 return;
  49
  50         resched = need_resched();
  51         preempt_disable_notrace();
  52
  53         cpu = raw_smp_processor_id();
  54         data = tr->data[cpu];
  55         disabled = atomic_inc_return(&data->disabled);
  56         if (unlikely(disabled != 1))
  57                 goto out;
  58
  59         spin_lock_irqsave(&wakeup_lock, flags);
  60
  61         if (unlikely(!wakeup_task))
  62                 goto unlock;
  63
  64         /*
  65          * The task can't disappear because it needs to
  66          * wake up first, and we have the wakeup_lock.
  67          */
  68         if (task_cpu(wakeup_task) != cpu)
  69                 goto unlock;
  70
  71         trace_function(tr, data, ip, parent_ip, flags);
  72
  73  unlock:
  74         spin_unlock_irqrestore(&wakeup_lock, flags);
  75
  76  out:
  77         atomic_dec(&data->disabled);
  78
  79         /*
  80          * To prevent recursion from the scheduler, if the
  81          * resched flag was set before we entered, then
  82          * don't reschedule.
  83          */
  84         if (resched)
  85                 preempt_enable_no_resched_notrace();
  86         else
  87                 preempt_enable_notrace();
  88 }
  89
  90 static struct ftrace_ops trace_ops __read_mostly =
  91 {
  92         .func = wakeup_tracer_call,
  93 };
  94 #endif /* CONFIG_FTRACE */
  95
  96 /*
  97  * Should this new latency be reported/recorded?
  98  */
  99 static int report_latency(cycle_t delta)
 100 {
 101         if (tracing_thresh) {
 102                 if (delta < tracing_thresh)
 103                         return 0;
 104         } else {
 105                 if (delta <= tracing_max_latency)
 106                         return 0;
 107         }
 108         return 1;
 109 }
 110
 111 static void notrace
 112 wakeup_sched_switch(void *private, void *rq, struct task_struct *prev,
 113         struct task_struct *next)
 114 {
 115         unsigned long latency = 0, t0 = 0, t1 = 0;
 116         struct trace_array **ptr = private;
 117         struct trace_array *tr = *ptr;
 118         struct trace_array_cpu *data;
 119         cycle_t T0, T1, delta;
 120         unsigned long flags;
 121         long disabled;
 122         int cpu;
 123
 124         if (unlikely(!tracer_enabled))
 125                 return;
 126
 127         /*
 128          * When we start a new trace, we set wakeup_task to NULL
 129          * and then set tracer_enabled = 1. We want to make sure
 130          * that another CPU does not see the tracer_enabled = 1
 131          * and the wakeup_task with an older task, that might
 132          * actually be the same as next.
 133          */
 134         smp_rmb();
 135
 136         if (next != wakeup_task)
 137                 return;
 138
 139         /* The task we are waiting for is waking up */
 140         data = tr->data[wakeup_cpu];
 141
 142         /* disable local data, not wakeup_cpu data */
 143         cpu = raw_smp_processor_id();
 144         disabled = atomic_inc_return(&tr->data[cpu]->disabled);
 145         if (likely(disabled != 1))
 146                 goto out;
 147
 148         spin_lock_irqsave(&wakeup_lock, flags);
 149
 150         /* We could race with grabbing wakeup_lock */
 151         if (unlikely(!tracer_enabled || next != wakeup_task))
 152                 goto out_unlock;
 153
 154         trace_function(tr, data, CALLER_ADDR1, CALLER_ADDR2, flags);
 155
 156         /*
 157          * usecs conversion is slow so we try to delay the conversion
 158          * as long as possible:
 159          */
 160         T0 = data->preempt_timestamp;
 161         T1 = ftrace_now(cpu);
 162         delta = T1-T0;
 163
 164         if (!report_latency(delta))
 165                 goto out_unlock;
 166
 167         latency = nsecs_to_usecs(delta);
 168
 169         tracing_max_latency = delta;
 170         t0 = nsecs_to_usecs(T0);
 171         t1 = nsecs_to_usecs(T1);
 172
 173         update_max_tr(tr, wakeup_task, wakeup_cpu);
 174
 175 out_unlock:
 176         __wakeup_reset(tr);
 177         spin_unlock_irqrestore(&wakeup_lock, flags);
 178 out:
 179         atomic_dec(&tr->data[cpu]->disabled);
 180 }
 181
 182 static notrace void
 183 sched_switch_callback(void *probe_data, void *call_data,
 184                       const char *format, va_list *args)
 185 {
 186         struct task_struct *prev;
 187         struct task_struct *next;
 188         struct rq *__rq;
 189
 190         /* skip prev_pid %d next_pid %d prev_state %ld */
 191         (void)va_arg(*args, int);
 192         (void)va_arg(*args, int);
 193         (void)va_arg(*args, long);
 194         __rq = va_arg(*args, typeof(__rq));
 195         prev = va_arg(*args, typeof(prev));
 196         next = va_arg(*args, typeof(next));
 197
 198         tracing_record_cmdline(prev);
 199
 200         /*
 201          * If tracer_switch_func only points to the local
 202          * switch func, it still needs the ptr passed to it.
 203          */
 204         wakeup_sched_switch(probe_data, __rq, prev, next);
 205 }
 206
 207 static void __wakeup_reset(struct trace_array *tr)
 208 {
 209         struct trace_array_cpu *data;
 210         int cpu;
 211
 212         assert_spin_locked(&wakeup_lock);
 213
 214         for_each_possible_cpu(cpu) {
 215                 data = tr->data[cpu];
 216                 tracing_reset(data);
 217         }
 218
 219         wakeup_cpu = -1;
 220         wakeup_prio = -1;
 221
 222         if (wakeup_task)
 223                 put_task_struct(wakeup_task);
 224
 225         wakeup_task = NULL;
 226 }
 227
 228 static void wakeup_reset(struct trace_array *tr)
 229 {
 230         unsigned long flags;
 231
 232         spin_lock_irqsave(&wakeup_lock, flags);
 233         __wakeup_reset(tr);
 234         spin_unlock_irqrestore(&wakeup_lock, flags);
 235 }
 236
 237 static void
 238 wakeup_check_start(struct trace_array *tr, struct task_struct *p,
 239                    struct task_struct *curr)
 240 {
 241         int cpu = smp_processor_id();
 242         unsigned long flags;
 243         long disabled;
 244
 245         if (likely(!rt_task(p)) ||
 246                         p->prio >= wakeup_prio ||
 247                         p->prio >= curr->prio)
 248                 return;
 249
 250         disabled = atomic_inc_return(&tr->data[cpu]->disabled);
 251         if (unlikely(disabled != 1))
 252                 goto out;
 253
 254         /* interrupts should be off from try_to_wake_up */
 255         spin_lock(&wakeup_lock);
 256
 257         /* check for races. */
 258         if (!tracer_enabled || p->prio >= wakeup_prio)
 259                 goto out_locked;
 260
 261         /* reset the trace */
 262         __wakeup_reset(tr);
 263
 264         wakeup_cpu = task_cpu(p);
 265         wakeup_prio = p->prio;
 266
 267         wakeup_task = p;
 268         get_task_struct(wakeup_task);
 269
 270         local_save_flags(flags);
 271
 272         tr->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu);
 273         trace_function(tr, tr->data[wakeup_cpu],
 274                        CALLER_ADDR1, CALLER_ADDR2, flags);
 275
 276 out_locked:
 277         spin_unlock(&wakeup_lock);
 278 out:
 279         atomic_dec(&tr->data[cpu]->disabled);
 280 }
 281
 282 static notrace void
 283 wake_up_callback(void *probe_data, void *call_data,
 284                  const char *format, va_list *args)
 285 {
 286         struct trace_array **ptr = probe_data;
 287         struct trace_array *tr = *ptr;
 288         struct task_struct *curr;
 289         struct task_struct *task;
 290         struct rq *__rq;
 291
 292         if (likely(!tracer_enabled))
 293                 return;
 294
 295         /* Skip pid %d state %ld */
 296         (void)va_arg(*args, int);
 297         (void)va_arg(*args, long);
 298         /* now get the meat: "rq %p task %p rq->curr %p" */
 299         __rq = va_arg(*args, typeof(__rq));
 300         task = va_arg(*args, typeof(task));
 301         curr = va_arg(*args, typeof(curr));
 302
 303         tracing_record_cmdline(task);
 304         tracing_record_cmdline(curr);
 305
 306         wakeup_check_start(tr, task, curr);
 307 }
 308
 309 static void start_wakeup_tracer(struct trace_array *tr)
 310 {
 311         int ret;
 312
 313         ret = marker_probe_register("kernel_sched_wakeup",
 314                         "pid %d state %ld ## rq %p task %p rq->curr %p",
 315                         wake_up_callback,
 316                         &wakeup_trace);
 317         if (ret) {
 318                 pr_info("wakeup trace: Couldn't add marker"
 319                         " probe to kernel_sched_wakeup\n");
 320                 return;
 321         }
 322
 323         ret = marker_probe_register("kernel_sched_wakeup_new",
 324                         "pid %d state %ld ## rq %p task %p rq->curr %p",
 325                         wake_up_callback,
 326                         &wakeup_trace);
 327         if (ret) {
 328                 pr_info("wakeup trace: Couldn't add marker"
 329                         " probe to kernel_sched_wakeup_new\n");
 330                 goto fail_deprobe;
 331         }
 332
 333         ret = marker_probe_register("kernel_sched_schedule",
 334                 "prev_pid %d next_pid %d prev_state %ld "
 335                 "## rq %p prev %p next %p",
 336                 sched_switch_callback,
 337                 &wakeup_trace);
 338         if (ret) {
 339                 pr_info("sched trace: Couldn't add marker"
 340                         " probe to kernel_sched_schedule\n");
 341                 goto fail_deprobe_wake_new;
 342         }
 343
 344         wakeup_reset(tr);
 345
 346         /*
 347          * Don't let the tracer_enabled = 1 show up before
 348          * the wakeup_task is reset. This may be overkill since
 349          * wakeup_reset does a spin_unlock after setting the
 350          * wakeup_task to NULL, but I want to be safe.
 351          * This is a slow path anyway.
 352          */
 353         smp_wmb();
 354
 355         register_ftrace_function(&trace_ops);
 356
 357         tracer_enabled = 1;
 358
 359         return;
 360 fail_deprobe_wake_new:
 361         marker_probe_unregister("kernel_sched_wakeup_new",
 362                                 wake_up_callback,
 363                                 &wakeup_trace);
 364 fail_deprobe:
 365         marker_probe_unregister("kernel_sched_wakeup",
 366                                 wake_up_callback,
 367                                 &wakeup_trace);
 368 }
 369
 370 static void stop_wakeup_tracer(struct trace_array *tr)
 371 {
 372         tracer_enabled = 0;
 373         unregister_ftrace_function(&trace_ops);
 374         marker_probe_unregister("kernel_sched_schedule",
 375                                 sched_switch_callback,
 376                                 &wakeup_trace);
 377         marker_probe_unregister("kernel_sched_wakeup_new",
 378                                 wake_up_callback,
 379                                 &wakeup_trace);
 380         marker_probe_unregister("kernel_sched_wakeup",
 381                                 wake_up_callback,
 382                                 &wakeup_trace);
 383 }
 384
 385 static void wakeup_tracer_init(struct trace_array *tr)
 386 {
 387         wakeup_trace = tr;
 388
 389         if (tr->ctrl)
 390                 start_wakeup_tracer(tr);
 391 }
 392
 393 static void wakeup_tracer_reset(struct trace_array *tr)
 394 {
 395         if (tr->ctrl) {
 396                 stop_wakeup_tracer(tr);
 397                 /* make sure we put back any tasks we are tracing */
 398                 wakeup_reset(tr);
 399         }
 400 }
 401
 402 static void wakeup_tracer_ctrl_update(struct trace_array *tr)
 403 {
 404         if (tr->ctrl)
 405                 start_wakeup_tracer(tr);
 406         else
 407                 stop_wakeup_tracer(tr);
 408 }
 409
 410 static void wakeup_tracer_open(struct trace_iterator *iter)
 411 {
 412         /* stop the trace while dumping */
 413         if (iter->tr->ctrl)
 414                 stop_wakeup_tracer(iter->tr);
 415 }
 416
 417 static void wakeup_tracer_close(struct trace_iterator *iter)
 418 {
 419         /* forget about any processes we were recording */
 420         if (iter->tr->ctrl)
 421                 start_wakeup_tracer(iter->tr);
 422 }
 423
 424 static struct tracer wakeup_tracer __read_mostly =
 425 {
 426         .name           = "wakeup",
 427         .init           = wakeup_tracer_init,
 428         .reset          = wakeup_tracer_reset,
 429         .open           = wakeup_tracer_open,
 430         .close          = wakeup_tracer_close,
 431         .ctrl_update    = wakeup_tracer_ctrl_update,
 432         .print_max      = 1,
 433 #ifdef CONFIG_FTRACE_SELFTEST
 434         .selftest    = trace_selftest_startup_wakeup,
 435 #endif
 436 };
 437
 438 __init static int init_wakeup_tracer(void)
 439 {
 440         int ret;
 441
 442         ret = register_tracer(&wakeup_tracer);
 443         if (ret)
 444                 return ret;
 445
 446         return 0;
 447 }
 448 device_initcall(init_wakeup_tracer);