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         ftrace(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 = 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): %lu us wakeup latency "
 111                        "violates %lu us threshold.\n"
 112                        " => started at timestamp %lu: ",
 113                                 wakeup_task->comm, wakeup_task->pid,
 114                                 raw_smp_processor_id(),
 115                                 latency, nsecs_to_usecs(tracing_thresh), t0);
 116         } else {
 117                 printk(KERN_INFO "(%16s-%-5d|#%d): new %lu us maximum "
 118                        "wakeup latency.\n => started at timestamp %lu: ",
 119                                 wakeup_task->comm, wakeup_task->pid,
 120                                 cpu, latency, t0);
 121         }
 122
 123         printk(KERN_CONT "   ended at timestamp %lu: ", t1);
 124         dump_stack();
 125         t1 = nsecs_to_usecs(now(cpu));
 126         printk(KERN_CONT "   dump-end timestamp %lu\n\n", t1);
 127
 128 out_unlock:
 129         __wakeup_reset(tr);
 130         spin_unlock_irqrestore(&wakeup_lock, flags);
 131 out:
 132         atomic_dec(&tr->data[cpu]->disabled);
 133 }
 134
 135 static void notrace __wakeup_reset(struct trace_array *tr)
 136 {
 137         struct trace_array_cpu *data;
 138         int cpu;
 139
 140         assert_spin_locked(&wakeup_lock);
 141
 142         for_each_possible_cpu(cpu) {
 143                 data = tr->data[cpu];
 144                 tracing_reset(data);
 145         }
 146
 147         wakeup_cpu = -1;
 148         wakeup_prio = -1;
 149
 150         if (wakeup_task)
 151                 put_task_struct(wakeup_task);
 152
 153         wakeup_task = NULL;
 154 }
 155
 156 static void notrace wakeup_reset(struct trace_array *tr)
 157 {
 158         unsigned long flags;
 159
 160         spin_lock_irqsave(&wakeup_lock, flags);
 161         __wakeup_reset(tr);
 162         spin_unlock_irqrestore(&wakeup_lock, flags);
 163 }
 164
 165 static notrace void
 166 wakeup_check_start(struct trace_array *tr, struct task_struct *p,
 167                    struct task_struct *curr)
 168 {
 169         int cpu = smp_processor_id();
 170         unsigned long flags;
 171         long disabled;
 172
 173         if (likely(!rt_task(p)) ||
 174                         p->prio >= wakeup_prio ||
 175                         p->prio >= curr->prio)
 176                 return;
 177
 178         disabled = atomic_inc_return(&tr->data[cpu]->disabled);
 179         if (unlikely(disabled != 1))
 180                 goto out;
 181
 182         /* interrupts should be off from try_to_wake_up */
 183         spin_lock(&wakeup_lock);
 184
 185         /* check for races. */
 186         if (!tracer_enabled || p->prio >= wakeup_prio)
 187                 goto out_locked;
 188
 189         /* reset the trace */
 190         __wakeup_reset(tr);
 191
 192         wakeup_cpu = task_cpu(p);
 193         wakeup_prio = p->prio;
 194
 195         wakeup_task = p;
 196         get_task_struct(wakeup_task);
 197
 198         local_save_flags(flags);
 199
 200         tr->data[wakeup_cpu]->preempt_timestamp = now(cpu);
 201         ftrace(tr, tr->data[wakeup_cpu], CALLER_ADDR1, CALLER_ADDR2, flags);
 202
 203 out_locked:
 204         spin_unlock(&wakeup_lock);
 205 out:
 206         atomic_dec(&tr->data[cpu]->disabled);
 207 }
 208
 209 notrace void
 210 ftrace_wake_up_task(struct task_struct *wakee, struct task_struct *curr)
 211 {
 212         if (likely(!tracer_enabled))
 213                 return;
 214
 215         wakeup_check_start(wakeup_trace, wakee, curr);
 216 }
 217
 218 notrace void
 219 ftrace_wake_up_new_task(struct task_struct *wakee, struct task_struct *curr)
 220 {
 221         if (likely(!tracer_enabled))
 222                 return;
 223
 224         wakeup_check_start(wakeup_trace, wakee, curr);
 225 }
 226
 227 static notrace void start_wakeup_tracer(struct trace_array *tr)
 228 {
 229         wakeup_reset(tr);
 230
 231         /*
 232          * Don't let the tracer_enabled = 1 show up before
 233          * the wakeup_task is reset. This may be overkill since
 234          * wakeup_reset does a spin_unlock after setting the
 235          * wakeup_task to NULL, but I want to be safe.
 236          * This is a slow path anyway.
 237          */
 238         smp_wmb();
 239
 240         tracer_enabled = 1;
 241
 242         return;
 243 }
 244
 245 static notrace void stop_wakeup_tracer(struct trace_array *tr)
 246 {
 247         tracer_enabled = 0;
 248 }
 249
 250 static notrace void wakeup_tracer_init(struct trace_array *tr)
 251 {
 252         wakeup_trace = tr;
 253
 254         if (tr->ctrl)
 255                 start_wakeup_tracer(tr);
 256 }
 257
 258 static notrace void wakeup_tracer_reset(struct trace_array *tr)
 259 {
 260         if (tr->ctrl) {
 261                 stop_wakeup_tracer(tr);
 262                 /* make sure we put back any tasks we are tracing */
 263                 wakeup_reset(tr);
 264         }
 265 }
 266
 267 static void wakeup_tracer_ctrl_update(struct trace_array *tr)
 268 {
 269         if (tr->ctrl)
 270                 start_wakeup_tracer(tr);
 271         else
 272                 stop_wakeup_tracer(tr);
 273 }
 274
 275 static void notrace wakeup_tracer_open(struct trace_iterator *iter)
 276 {
 277         /* stop the trace while dumping */
 278         if (iter->tr->ctrl)
 279                 stop_wakeup_tracer(iter->tr);
 280 }
 281
 282 static void notrace wakeup_tracer_close(struct trace_iterator *iter)
 283 {
 284         /* forget about any processes we were recording */
 285         if (iter->tr->ctrl)
 286                 start_wakeup_tracer(iter->tr);
 287 }
 288
 289 static struct tracer wakeup_tracer __read_mostly =
 290 {
 291         .name           = "wakeup",
 292         .init           = wakeup_tracer_init,
 293         .reset          = wakeup_tracer_reset,
 294         .open           = wakeup_tracer_open,
 295         .close          = wakeup_tracer_close,
 296         .ctrl_update    = wakeup_tracer_ctrl_update,
 297         .print_max      = 1,
 298 #ifdef CONFIG_FTRACE_SELFTEST
 299         .selftest    = trace_selftest_startup_wakeup,
 300 #endif
 301 };
 302
 303 __init static int init_wakeup_tracer(void)
 304 {
 305         int ret;
 306
 307         ret = register_tracer(&wakeup_tracer);
 308         if (ret)
 309                 return ret;
 310
 311         return 0;
 312 }
 313 device_initcall(init_wakeup_tracer);