kernel/trace/trace_sched_switch.c

   1 /*
   2  * trace context switch
   3  *
   4  * Copyright (C) 2007 Steven Rostedt <srostedt@redhat.com>
   5  *
   6  */
   7 #include <linux/module.h>
   8 #include <linux/fs.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/kallsyms.h>
  11 #include <linux/uaccess.h>
  12 #include <linux/marker.h>
  13 #include <linux/ftrace.h>
  14
  15 #include "trace.h"
  16
  17 static struct trace_array       *ctx_trace;
  18 static int __read_mostly        tracer_enabled;
  19 static atomic_t                 sched_ref;
  20
  21 static void
  22 sched_switch_func(void *private, void *__rq, struct task_struct *prev,
  23                         struct task_struct *next)
  24 {
  25         struct trace_array **ptr = private;
  26         struct trace_array *tr = *ptr;
  27         struct trace_array_cpu *data;
  28         unsigned long flags;
  29         long disabled;
  30         int cpu;
  31
  32         if (!tracer_enabled)
  33                 return;
  34
  35         local_irq_save(flags);
  36         cpu = raw_smp_processor_id();
  37         data = tr->data[cpu];
  38         disabled = atomic_inc_return(&data->disabled);
  39
  40         if (likely(disabled == 1))
  41                 tracing_sched_switch_trace(tr, data, prev, next, flags);
  42
  43         atomic_dec(&data->disabled);
  44         local_irq_restore(flags);
  45 }
  46
  47 static notrace void
  48 sched_switch_callback(void *probe_data, void *call_data,
  49                       const char *format, va_list *args)
  50 {
  51         struct task_struct *prev;
  52         struct task_struct *next;
  53         struct rq *__rq;
  54
  55         if (!atomic_read(&sched_ref))
  56                 return;
  57
  58         /* skip prev_pid %d next_pid %d prev_state %ld */
  59         (void)va_arg(*args, int);
  60         (void)va_arg(*args, int);
  61         (void)va_arg(*args, long);
  62         __rq = va_arg(*args, typeof(__rq));
  63         prev = va_arg(*args, typeof(prev));
  64         next = va_arg(*args, typeof(next));
  65
  66         tracing_record_cmdline(prev);
  67
  68         /*
  69          * If tracer_switch_func only points to the local
  70          * switch func, it still needs the ptr passed to it.
  71          */
  72         sched_switch_func(probe_data, __rq, prev, next);
  73 }
  74
  75 static void
  76 wakeup_func(void *private, void *__rq, struct task_struct *wakee, struct
  77                         task_struct *curr)
  78 {
  79         struct trace_array **ptr = private;
  80         struct trace_array *tr = *ptr;
  81         struct trace_array_cpu *data;
  82         unsigned long flags;
  83         long disabled;
  84         int cpu;
  85
  86         if (!tracer_enabled)
  87                 return;
  88
  89         tracing_record_cmdline(curr);
  90
  91         local_irq_save(flags);
  92         cpu = raw_smp_processor_id();
  93         data = tr->data[cpu];
  94         disabled = atomic_inc_return(&data->disabled);
  95
  96         if (likely(disabled == 1))
  97                 tracing_sched_wakeup_trace(tr, data, wakee, curr, flags);
  98
  99         atomic_dec(&data->disabled);
 100         local_irq_restore(flags);
 101 }
 102
 103 static notrace void
 104 wake_up_callback(void *probe_data, void *call_data,
 105                  const char *format, va_list *args)
 106 {
 107         struct task_struct *curr;
 108         struct task_struct *task;
 109         struct rq *__rq;
 110
 111         if (likely(!tracer_enabled))
 112                 return;
 113
 114         /* Skip pid %d state %ld */
 115         (void)va_arg(*args, int);
 116         (void)va_arg(*args, long);
 117         /* now get the meat: "rq %p task %p rq->curr %p" */
 118         __rq = va_arg(*args, typeof(__rq));
 119         task = va_arg(*args, typeof(task));
 120         curr = va_arg(*args, typeof(curr));
 121
 122         tracing_record_cmdline(task);
 123         tracing_record_cmdline(curr);
 124
 125         wakeup_func(probe_data, __rq, task, curr);
 126 }
 127
 128 void
 129 ftrace_special(unsigned long arg1, unsigned long arg2, unsigned long arg3)
 130 {
 131         struct trace_array *tr = ctx_trace;
 132         struct trace_array_cpu *data;
 133         unsigned long flags;
 134         long disabled;
 135         int cpu;
 136
 137         if (!tracer_enabled)
 138                 return;
 139
 140         local_irq_save(flags);
 141         cpu = raw_smp_processor_id();
 142         data = tr->data[cpu];
 143         disabled = atomic_inc_return(&data->disabled);
 144
 145         if (likely(disabled == 1))
 146                 __trace_special(tr, data, arg1, arg2, arg3);
 147
 148         atomic_dec(&data->disabled);
 149         local_irq_restore(flags);
 150 }
 151
 152 static void sched_switch_reset(struct trace_array *tr)
 153 {
 154         int cpu;
 155
 156         tr->time_start = ftrace_now(tr->cpu);
 157
 158         for_each_online_cpu(cpu)
 159                 tracing_reset(tr->data[cpu]);
 160 }
 161
 162 static int tracing_sched_register(void)
 163 {
 164         int ret;
 165
 166         ret = marker_probe_register("kernel_sched_wakeup",
 167                         "pid %d state %ld ## rq %p task %p rq->curr %p",
 168                         wake_up_callback,
 169                         &ctx_trace);
 170         if (ret) {
 171                 pr_info("wakeup trace: Couldn't add marker"
 172                         " probe to kernel_sched_wakeup\n");
 173                 return ret;
 174         }
 175
 176         ret = marker_probe_register("kernel_sched_wakeup_new",
 177                         "pid %d state %ld ## rq %p task %p rq->curr %p",
 178                         wake_up_callback,
 179                         &ctx_trace);
 180         if (ret) {
 181                 pr_info("wakeup trace: Couldn't add marker"
 182                         " probe to kernel_sched_wakeup_new\n");
 183                 goto fail_deprobe;
 184         }
 185
 186         ret = marker_probe_register("kernel_sched_schedule",
 187                 "prev_pid %d next_pid %d prev_state %ld "
 188                 "## rq %p prev %p next %p",
 189                 sched_switch_callback,
 190                 &ctx_trace);
 191         if (ret) {
 192                 pr_info("sched trace: Couldn't add marker"
 193                         " probe to kernel_sched_schedule\n");
 194                 goto fail_deprobe_wake_new;
 195         }
 196
 197         return ret;
 198 fail_deprobe_wake_new:
 199         marker_probe_unregister("kernel_sched_wakeup_new",
 200                                 wake_up_callback,
 201                                 &ctx_trace);
 202 fail_deprobe:
 203         marker_probe_unregister("kernel_sched_wakeup",
 204                                 wake_up_callback,
 205                                 &ctx_trace);
 206         return ret;
 207 }
 208
 209 static void tracing_sched_unregister(void)
 210 {
 211         marker_probe_unregister("kernel_sched_schedule",
 212                                 sched_switch_callback,
 213                                 &ctx_trace);
 214         marker_probe_unregister("kernel_sched_wakeup_new",
 215                                 wake_up_callback,
 216                                 &ctx_trace);
 217         marker_probe_unregister("kernel_sched_wakeup",
 218                                 wake_up_callback,
 219                                 &ctx_trace);
 220 }
 221
 222 void tracing_start_sched_switch(void)
 223 {
 224         long ref;
 225
 226         ref = atomic_inc_return(&sched_ref);
 227         if (ref == 1)
 228                 tracing_sched_register();
 229 }
 230
 231 void tracing_stop_sched_switch(void)
 232 {
 233         long ref;
 234
 235         ref = atomic_dec_and_test(&sched_ref);
 236         if (ref)
 237                 tracing_sched_unregister();
 238 }
 239
 240 static void start_sched_trace(struct trace_array *tr)
 241 {
 242         sched_switch_reset(tr);
 243         atomic_inc(&trace_record_cmdline_enabled);
 244         tracer_enabled = 1;
 245         tracing_start_sched_switch();
 246 }
 247
 248 static void stop_sched_trace(struct trace_array *tr)
 249 {
 250         tracing_stop_sched_switch();
 251         atomic_dec(&trace_record_cmdline_enabled);
 252         tracer_enabled = 0;
 253 }
 254
 255 static void sched_switch_trace_init(struct trace_array *tr)
 256 {
 257         ctx_trace = tr;
 258
 259         if (tr->ctrl)
 260                 start_sched_trace(tr);
 261 }
 262
 263 static void sched_switch_trace_reset(struct trace_array *tr)
 264 {
 265         if (tr->ctrl)
 266                 stop_sched_trace(tr);
 267 }
 268
 269 static void sched_switch_trace_ctrl_update(struct trace_array *tr)
 270 {
 271         /* When starting a new trace, reset the buffers */
 272         if (tr->ctrl)
 273                 start_sched_trace(tr);
 274         else
 275                 stop_sched_trace(tr);
 276 }
 277
 278 static struct tracer sched_switch_trace __read_mostly =
 279 {
 280         .name           = "sched_switch",
 281         .init           = sched_switch_trace_init,
 282         .reset          = sched_switch_trace_reset,
 283         .ctrl_update    = sched_switch_trace_ctrl_update,
 284 #ifdef CONFIG_FTRACE_SELFTEST
 285         .selftest    = trace_selftest_startup_sched_switch,
 286 #endif
 287 };
 288
 289 __init static int init_sched_switch_trace(void)
 290 {
 291         int ret = 0;
 292
 293         if (atomic_read(&sched_ref))
 294                 ret = tracing_sched_register();
 295         if (ret) {
 296                 pr_info("error registering scheduler trace\n");
 297                 return ret;
 298         }
 299         return register_tracer(&sched_switch_trace);
 300 }
 301 device_initcall(init_sched_switch_trace);