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
  19 #include "trace.h"
  20
  21 static struct trace_array       *wakeup_trace;
  22 static int __read_mostly        tracer_enabled;
  23
  24 static struct task_struct       *wakeup_task;
  25 static int                      wakeup_cpu;
  26 static unsigned                 wakeup_prio = -1;
  27
  28 static DEFINE_SPINLOCK(wakeup_lock);
  29
  30 static void notrace __wakeup_reset(struct trace_array *tr);
  31
  32 /*
  33  * Should this new latency be reported/recorded?
  34  */
  35 static int notrace report_latency(cycle_t delta)
  36 {
  37         if (tracing_thresh) {
  38                 if (delta < tracing_thresh)
  39                         return 0;
  40         } else {
  41                 if (delta <= tracing_max_latency)
  42                         return 0;
  43         }
  44         return 1;
  45 }
  46
  47 void notrace
  48 wakeup_sched_switch(struct task_struct *prev, struct task_struct *next)
  49 {
  50         unsigned long latency = 0, t0 = 0, t1 = 0;
  51         struct trace_array *tr = wakeup_trace;
  52         struct trace_array_cpu *data;
  53         cycle_t T0, T1, delta;
  54         unsigned long flags;
  55         long disabled;
  56         int cpu;
  57
  58         if (unlikely(!tracer_enabled))
  59                 return;
  60
  61         /*
  62          * When we start a new trace, we set wakeup_task to NULL
  63          * and then set tracer_enabled = 1. We want to make sure
  64          * that another CPU does not see the tracer_enabled = 1
  65          * and the wakeup_task with an older task, that might
  66          * actually be the same as next.
  67          */
  68         smp_rmb();
  69
  70         if (next != wakeup_task)
  71                 return;
  72
  73         /* The task we are waitng for is waking up */
  74         data = tr->data[wakeup_cpu];
  75
  76         /* disable local data, not wakeup_cpu data */
  77         cpu = raw_smp_processor_id();
  78         disabled = atomic_inc_return(&tr->data[cpu]->disabled);
  79         if (likely(disabled != 1))
  80                 goto out;
  81
  82         spin_lock_irqsave(&wakeup_lock, flags);
  83
  84         /* We could race with grabbing wakeup_lock */
  85         if (unlikely(!tracer_enabled || next != wakeup_task))
  86                 goto out_unlock;
  87
  88         trace_function(tr, data, CALLER_ADDR1, CALLER_ADDR2, flags);
  89
  90         /*
  91          * usecs conversion is slow so we try to delay the conversion
  92          * as long as possible:
  93          */
  94         T0 = data->preempt_timestamp;
  95         T1 = ftrace_now(cpu);
  96         delta = T1-T0;
  97
  98         if (!report_latency(delta))
  99                 goto out_unlock;
 100
 101         latency = nsecs_to_usecs(delta);
 102
 103         tracing_max_latency = delta;
 104         t0 = nsecs_to_usecs(T0);
 105         t1 = nsecs_to_usecs(T1);
 106
 107         update_max_tr(tr, wakeup_task, wakeup_cpu);
 108
 109         if (tracing_thresh) {
 110                 printk(KERN_INFO "(%16s-%-5d|#%d):"
 111                         " %lu us wakeup latency violates %lu us threshold.\n",
 112                                 wakeup_task->comm, wakeup_task->pid,
 113                                 raw_smp_processor_id(),
 114                                 latency, nsecs_to_usecs(tracing_thresh));
 115         } else {
 116                 printk(KERN_INFO "(%16s-%-5d|#%d):"
 117                         " new %lu us maximum wakeup latency.\n",
 118                                 wakeup_task->comm, wakeup_task->pid,
 119                                 cpu, latency);
 120         }
 121
 122 out_unlock:
 123         __wakeup_reset(tr);
 124         spin_unlock_irqrestore(&wakeup_lock, flags);
 125 out:
 126         atomic_dec(&tr->data[cpu]->disabled);
 127 }
 128
 129 static void notrace __wakeup_reset(struct trace_array *tr)
 130 {
 131         struct trace_array_cpu *data;
 132         int cpu;
 133
 134         assert_spin_locked(&wakeup_lock);
 135
 136         for_each_possible_cpu(cpu) {
 137                 data = tr->data[cpu];
 138                 tracing_reset(data);
 139         }
 140
 141         wakeup_cpu = -1;
 142         wakeup_prio = -1;
 143
 144         if (wakeup_task)
 145                 put_task_struct(wakeup_task);
 146
 147         wakeup_task = NULL;
 148 }
 149
 150 static void notrace wakeup_reset(struct trace_array *tr)
 151 {
 152         unsigned long flags;
 153
 154         spin_lock_irqsave(&wakeup_lock, flags);
 155         __wakeup_reset(tr);
 156         spin_unlock_irqrestore(&wakeup_lock, flags);
 157 }
 158
 159 static notrace void
 160 wakeup_check_start(struct trace_array *tr, struct task_struct *p,
 161                    struct task_struct *curr)
 162 {
 163         int cpu = smp_processor_id();
 164         unsigned long flags;
 165         long disabled;
 166
 167         if (likely(!rt_task(p)) ||
 168                         p->prio >= wakeup_prio ||
 169                         p->prio >= curr->prio)
 170                 return;
 171
 172         disabled = atomic_inc_return(&tr->data[cpu]->disabled);
 173         if (unlikely(disabled != 1))
 174                 goto out;
 175
 176         /* interrupts should be off from try_to_wake_up */
 177         spin_lock(&wakeup_lock);
 178
 179         /* check for races. */
 180         if (!tracer_enabled || p->prio >= wakeup_prio)
 181                 goto out_locked;
 182
 183         /* reset the trace */
 184         __wakeup_reset(tr);
 185
 186         wakeup_cpu = task_cpu(p);
 187         wakeup_prio = p->prio;
 188
 189         wakeup_task = p;
 190         get_task_struct(wakeup_task);
 191
 192         local_save_flags(flags);
 193
 194         tr->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu);
 195         trace_function(tr, tr->data[wakeup_cpu],
 196                        CALLER_ADDR1, CALLER_ADDR2, flags);
 197
 198 out_locked:
 199         spin_unlock(&wakeup_lock);
 200 out:
 201         atomic_dec(&tr->data[cpu]->disabled);
 202 }
 203
 204 notrace void
 205 ftrace_wake_up_task(struct task_struct *wakee, struct task_struct *curr)
 206 {
 207         if (likely(!tracer_enabled))
 208                 return;
 209
 210         wakeup_check_start(wakeup_trace, wakee, curr);
 211 }
 212
 213 notrace void
 214 ftrace_wake_up_new_task(struct task_struct *wakee, struct task_struct *curr)
 215 {
 216         if (likely(!tracer_enabled))
 217                 return;
 218
 219         wakeup_check_start(wakeup_trace, wakee, curr);
 220 }
 221
 222 static notrace void start_wakeup_tracer(struct trace_array *tr)
 223 {
 224         wakeup_reset(tr);
 225
 226         /*
 227          * Don't let the tracer_enabled = 1 show up before
 228          * the wakeup_task is reset. This may be overkill since
 229          * wakeup_reset does a spin_unlock after setting the
 230          * wakeup_task to NULL, but I want to be safe.
 231          * This is a slow path anyway.
 232          */
 233         smp_wmb();
 234
 235         tracer_enabled = 1;
 236
 237         return;
 238 }
 239
 240 static notrace void stop_wakeup_tracer(struct trace_array *tr)
 241 {
 242         tracer_enabled = 0;
 243 }
 244
 245 static notrace void wakeup_tracer_init(struct trace_array *tr)
 246 {
 247         wakeup_trace = tr;
 248
 249         if (tr->ctrl)
 250                 start_wakeup_tracer(tr);
 251 }
 252
 253 static notrace void wakeup_tracer_reset(struct trace_array *tr)
 254 {
 255         if (tr->ctrl) {
 256                 stop_wakeup_tracer(tr);
 257                 /* make sure we put back any tasks we are tracing */
 258                 wakeup_reset(tr);
 259         }
 260 }
 261
 262 static void wakeup_tracer_ctrl_update(struct trace_array *tr)
 263 {
 264         if (tr->ctrl)
 265                 start_wakeup_tracer(tr);
 266         else
 267                 stop_wakeup_tracer(tr);
 268 }
 269
 270 static void notrace wakeup_tracer_open(struct trace_iterator *iter)
 271 {
 272         /* stop the trace while dumping */
 273         if (iter->tr->ctrl)
 274                 stop_wakeup_tracer(iter->tr);
 275 }
 276
 277 static void notrace wakeup_tracer_close(struct trace_iterator *iter)
 278 {
 279         /* forget about any processes we were recording */
 280         if (iter->tr->ctrl)
 281                 start_wakeup_tracer(iter->tr);
 282 }
 283
 284 static struct tracer wakeup_tracer __read_mostly =
 285 {
 286         .name           = "wakeup",
 287         .init           = wakeup_tracer_init,
 288         .reset          = wakeup_tracer_reset,
 289         .open           = wakeup_tracer_open,
 290         .close          = wakeup_tracer_close,
 291         .ctrl_update    = wakeup_tracer_ctrl_update,
 292         .print_max      = 1,
 293 #ifdef CONFIG_FTRACE_SELFTEST
 294         .selftest    = trace_selftest_startup_wakeup,
 295 #endif
 296 };
 297
 298 __init static int init_wakeup_tracer(void)
 299 {
 300         int ret;
 301
 302         ret = register_tracer(&wakeup_tracer);
 303         if (ret)
 304                 return ret;
 305
 306         return 0;
 307 }
 308 device_initcall(init_wakeup_tracer);