kernel/trace/trace_irqsoff.c

   1 /*
   2  * trace irqs off criticall timings
   3  *
   4  * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
   5  * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com>
   6  *
   7  * From code in 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/kallsyms.h>
  13 #include <linux/debugfs.h>
  14 #include <linux/uaccess.h>
  15 #include <linux/module.h>
  16 #include <linux/ftrace.h>
  17 #include <linux/fs.h>
  18
  19 #include "trace.h"
  20
  21 static struct trace_array               *irqsoff_trace __read_mostly;
  22 static int                              tracer_enabled __read_mostly;
  23
  24 static DEFINE_PER_CPU(int, tracing_cpu);
  25
  26 static DEFINE_SPINLOCK(max_trace_lock);
  27
  28 enum {
  29         TRACER_IRQS_OFF         = (1 << 1),
  30         TRACER_PREEMPT_OFF      = (1 << 2),
  31 };
  32
  33 static int trace_type __read_mostly;
  34
  35 #ifdef CONFIG_PREEMPT_TRACER
  36 static inline int notrace
  37 preempt_trace(void)
  38 {
  39         return ((trace_type & TRACER_PREEMPT_OFF) && preempt_count());
  40 }
  41 #else
  42 # define preempt_trace() (0)
  43 #endif
  44
  45 #ifdef CONFIG_IRQSOFF_TRACER
  46 static inline int notrace
  47 irq_trace(void)
  48 {
  49         return ((trace_type & TRACER_IRQS_OFF) &&
  50                 irqs_disabled());
  51 }
  52 #else
  53 # define irq_trace() (0)
  54 #endif
  55
  56 /*
  57  * Sequence count - we record it when starting a measurement and
  58  * skip the latency if the sequence has changed - some other section
  59  * did a maximum and could disturb our measurement with serial console
  60  * printouts, etc. Truly coinciding maximum latencies should be rare
  61  * and what happens together happens separately as well, so this doesnt
  62  * decrease the validity of the maximum found:
  63  */
  64 static __cacheline_aligned_in_smp       unsigned long max_sequence;
  65
  66 #ifdef CONFIG_FTRACE
  67 /*
  68  * irqsoff uses its own tracer function to keep the overhead down:
  69  */
  70 static void notrace
  71 irqsoff_tracer_call(unsigned long ip, unsigned long parent_ip)
  72 {
  73         struct trace_array *tr = irqsoff_trace;
  74         struct trace_array_cpu *data;
  75         unsigned long flags;
  76         long disabled;
  77         int cpu;
  78
  79         /*
  80          * Does not matter if we preempt. We test the flags
  81          * afterward, to see if irqs are disabled or not.
  82          * If we preempt and get a false positive, the flags
  83          * test will fail.
  84          */
  85         cpu = raw_smp_processor_id();
  86         if (likely(!per_cpu(tracing_cpu, cpu)))
  87                 return;
  88
  89         local_save_flags(flags);
  90         /* slight chance to get a false positive on tracing_cpu */
  91         if (!irqs_disabled_flags(flags))
  92                 return;
  93
  94         data = tr->data[cpu];
  95         disabled = atomic_inc_return(&data->disabled);
  96
  97         if (likely(disabled == 1))
  98                 trace_function(tr, data, ip, parent_ip, flags);
  99
 100         atomic_dec(&data->disabled);
 101 }
 102
 103 static struct ftrace_ops trace_ops __read_mostly =
 104 {
 105         .func = irqsoff_tracer_call,
 106 };
 107 #endif /* CONFIG_FTRACE */
 108
 109 /*
 110  * Should this new latency be reported/recorded?
 111  */
 112 static int notrace report_latency(cycle_t delta)
 113 {
 114         if (tracing_thresh) {
 115                 if (delta < tracing_thresh)
 116                         return 0;
 117         } else {
 118                 if (delta <= tracing_max_latency)
 119                         return 0;
 120         }
 121         return 1;
 122 }
 123
 124 static void notrace
 125 check_critical_timing(struct trace_array *tr,
 126                       struct trace_array_cpu *data,
 127                       unsigned long parent_ip,
 128                       int cpu)
 129 {
 130         unsigned long latency, t0, t1;
 131         cycle_t T0, T1, delta;
 132         unsigned long flags;
 133
 134         /*
 135          * usecs conversion is slow so we try to delay the conversion
 136          * as long as possible:
 137          */
 138         T0 = data->preempt_timestamp;
 139         T1 = ftrace_now(cpu);
 140         delta = T1-T0;
 141
 142         local_save_flags(flags);
 143
 144         if (!report_latency(delta))
 145                 goto out;
 146
 147         spin_lock_irqsave(&max_trace_lock, flags);
 148
 149         /* check if we are still the max latency */
 150         if (!report_latency(delta))
 151                 goto out_unlock;
 152
 153         trace_function(tr, data, CALLER_ADDR0, parent_ip, flags);
 154
 155         latency = nsecs_to_usecs(delta);
 156
 157         if (data->critical_sequence != max_sequence)
 158                 goto out_unlock;
 159
 160         tracing_max_latency = delta;
 161         t0 = nsecs_to_usecs(T0);
 162         t1 = nsecs_to_usecs(T1);
 163
 164         data->critical_end = parent_ip;
 165
 166         update_max_tr_single(tr, current, cpu);
 167
 168         if (tracing_thresh) {
 169                 printk(KERN_INFO "(%16s-%-5d|#%d):"
 170                         " %lu us critical section violates %lu us threshold.\n",
 171                                 current->comm, current->pid,
 172                                 raw_smp_processor_id(),
 173                                 latency, nsecs_to_usecs(tracing_thresh));
 174         } else {
 175                 printk(KERN_INFO "(%16s-%-5d|#%d):"
 176                        " new %lu us maximum-latency critical section.\n",
 177                                 current->comm, current->pid,
 178                                 raw_smp_processor_id(),
 179                                 latency);
 180         }
 181
 182         max_sequence++;
 183
 184 out_unlock:
 185         spin_unlock_irqrestore(&max_trace_lock, flags);
 186
 187 out:
 188         data->critical_sequence = max_sequence;
 189         data->preempt_timestamp = ftrace_now(cpu);
 190         tracing_reset(data);
 191         trace_function(tr, data, CALLER_ADDR0, parent_ip, flags);
 192 }
 193
 194 static inline void notrace
 195 start_critical_timing(unsigned long ip, unsigned long parent_ip)
 196 {
 197         int cpu;
 198         struct trace_array *tr = irqsoff_trace;
 199         struct trace_array_cpu *data;
 200         unsigned long flags;
 201
 202         if (likely(!tracer_enabled))
 203                 return;
 204
 205         if (__get_cpu_var(tracing_cpu))
 206                 return;
 207
 208         cpu = raw_smp_processor_id();
 209         data = tr->data[cpu];
 210
 211         if (unlikely(!data) || unlikely(!head_page(data)) ||
 212             atomic_read(&data->disabled))
 213                 return;
 214
 215         atomic_inc(&data->disabled);
 216
 217         data->critical_sequence = max_sequence;
 218         data->preempt_timestamp = ftrace_now(cpu);
 219         data->critical_start = parent_ip ? : ip;
 220         tracing_reset(data);
 221
 222         local_save_flags(flags);
 223
 224         trace_function(tr, data, ip, parent_ip, flags);
 225
 226         __get_cpu_var(tracing_cpu) = 1;
 227
 228         atomic_dec(&data->disabled);
 229 }
 230
 231 static inline void notrace
 232 stop_critical_timing(unsigned long ip, unsigned long parent_ip)
 233 {
 234         int cpu;
 235         struct trace_array *tr = irqsoff_trace;
 236         struct trace_array_cpu *data;
 237         unsigned long flags;
 238
 239         /* Always clear the tracing cpu on stopping the trace */
 240         if (unlikely(__get_cpu_var(tracing_cpu)))
 241                 __get_cpu_var(tracing_cpu) = 0;
 242         else
 243                 return;
 244
 245         if (!tracer_enabled)
 246                 return;
 247
 248         cpu = raw_smp_processor_id();
 249         data = tr->data[cpu];
 250
 251         if (unlikely(!data) || unlikely(!head_page(data)) ||
 252             !data->critical_start || atomic_read(&data->disabled))
 253                 return;
 254
 255         atomic_inc(&data->disabled);
 256         local_save_flags(flags);
 257         trace_function(tr, data, ip, parent_ip, flags);
 258         check_critical_timing(tr, data, parent_ip ? : ip, cpu);
 259         data->critical_start = 0;
 260         atomic_dec(&data->disabled);
 261 }
 262
 263 /* start and stop critical timings used to for stoppage (in idle) */
 264 void notrace start_critical_timings(void)
 265 {
 266         if (preempt_trace() || irq_trace())
 267                 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 268 }
 269
 270 void notrace stop_critical_timings(void)
 271 {
 272         if (preempt_trace() || irq_trace())
 273                 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 274 }
 275
 276 #ifdef CONFIG_IRQSOFF_TRACER
 277 #ifdef CONFIG_PROVE_LOCKING
 278 void notrace time_hardirqs_on(unsigned long a0, unsigned long a1)
 279 {
 280         if (!preempt_trace() && irq_trace())
 281                 stop_critical_timing(a0, a1);
 282 }
 283
 284 void notrace time_hardirqs_off(unsigned long a0, unsigned long a1)
 285 {
 286         if (!preempt_trace() && irq_trace())
 287                 start_critical_timing(a0, a1);
 288 }
 289
 290 #else /* !CONFIG_PROVE_LOCKING */
 291
 292 /*
 293  * Stubs:
 294  */
 295
 296 void early_boot_irqs_off(void)
 297 {
 298 }
 299
 300 void early_boot_irqs_on(void)
 301 {
 302 }
 303
 304 void trace_softirqs_on(unsigned long ip)
 305 {
 306 }
 307
 308 void trace_softirqs_off(unsigned long ip)
 309 {
 310 }
 311
 312 inline void print_irqtrace_events(struct task_struct *curr)
 313 {
 314 }
 315
 316 /*
 317  * We are only interested in hardirq on/off events:
 318  */
 319 void notrace trace_hardirqs_on(void)
 320 {
 321         if (!preempt_trace() && irq_trace())
 322                 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 323 }
 324 EXPORT_SYMBOL(trace_hardirqs_on);
 325
 326 void notrace trace_hardirqs_off(void)
 327 {
 328         if (!preempt_trace() && irq_trace())
 329                 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 330 }
 331 EXPORT_SYMBOL(trace_hardirqs_off);
 332
 333 void notrace trace_hardirqs_on_caller(unsigned long caller_addr)
 334 {
 335         if (!preempt_trace() && irq_trace())
 336                 stop_critical_timing(CALLER_ADDR0, caller_addr);
 337 }
 338 EXPORT_SYMBOL(trace_hardirqs_on_caller);
 339
 340 void notrace trace_hardirqs_off_caller(unsigned long caller_addr)
 341 {
 342         if (!preempt_trace() && irq_trace())
 343                 start_critical_timing(CALLER_ADDR0, caller_addr);
 344 }
 345 EXPORT_SYMBOL(trace_hardirqs_off_caller);
 346
 347 #endif /* CONFIG_PROVE_LOCKING */
 348 #endif /*  CONFIG_IRQSOFF_TRACER */
 349
 350 #ifdef CONFIG_PREEMPT_TRACER
 351 void notrace trace_preempt_on(unsigned long a0, unsigned long a1)
 352 {
 353         stop_critical_timing(a0, a1);
 354 }
 355
 356 void notrace trace_preempt_off(unsigned long a0, unsigned long a1)
 357 {
 358         start_critical_timing(a0, a1);
 359 }
 360 #endif /* CONFIG_PREEMPT_TRACER */
 361
 362 static void start_irqsoff_tracer(struct trace_array *tr)
 363 {
 364         register_ftrace_function(&trace_ops);
 365         tracer_enabled = 1;
 366 }
 367
 368 static void stop_irqsoff_tracer(struct trace_array *tr)
 369 {
 370         tracer_enabled = 0;
 371         unregister_ftrace_function(&trace_ops);
 372 }
 373
 374 static void __irqsoff_tracer_init(struct trace_array *tr)
 375 {
 376         irqsoff_trace = tr;
 377         /* make sure that the tracer is visibel */
 378         smp_wmb();
 379
 380         if (tr->ctrl)
 381                 start_irqsoff_tracer(tr);
 382 }
 383
 384 static void irqsoff_tracer_reset(struct trace_array *tr)
 385 {
 386         if (tr->ctrl)
 387                 stop_irqsoff_tracer(tr);
 388 }
 389
 390 static void irqsoff_tracer_ctrl_update(struct trace_array *tr)
 391 {
 392         if (tr->ctrl)
 393                 start_irqsoff_tracer(tr);
 394         else
 395                 stop_irqsoff_tracer(tr);
 396 }
 397
 398 static void notrace irqsoff_tracer_open(struct trace_iterator *iter)
 399 {
 400         /* stop the trace while dumping */
 401         if (iter->tr->ctrl)
 402                 stop_irqsoff_tracer(iter->tr);
 403 }
 404
 405 static void notrace irqsoff_tracer_close(struct trace_iterator *iter)
 406 {
 407         if (iter->tr->ctrl)
 408                 start_irqsoff_tracer(iter->tr);
 409 }
 410
 411 #ifdef CONFIG_IRQSOFF_TRACER
 412 static void irqsoff_tracer_init(struct trace_array *tr)
 413 {
 414         trace_type = TRACER_IRQS_OFF;
 415
 416         __irqsoff_tracer_init(tr);
 417 }
 418 static struct tracer irqsoff_tracer __read_mostly =
 419 {
 420         .name           = "irqsoff",
 421         .init           = irqsoff_tracer_init,
 422         .reset          = irqsoff_tracer_reset,
 423         .open           = irqsoff_tracer_open,
 424         .close          = irqsoff_tracer_close,
 425         .ctrl_update    = irqsoff_tracer_ctrl_update,
 426         .print_max      = 1,
 427 #ifdef CONFIG_FTRACE_SELFTEST
 428         .selftest    = trace_selftest_startup_irqsoff,
 429 #endif
 430 };
 431 # define register_irqsoff(trace) register_tracer(&trace)
 432 #else
 433 # define register_irqsoff(trace) do { } while (0)
 434 #endif
 435
 436 #ifdef CONFIG_PREEMPT_TRACER
 437 static void preemptoff_tracer_init(struct trace_array *tr)
 438 {
 439         trace_type = TRACER_PREEMPT_OFF;
 440
 441         __irqsoff_tracer_init(tr);
 442 }
 443
 444 static struct tracer preemptoff_tracer __read_mostly =
 445 {
 446         .name           = "preemptoff",
 447         .init           = preemptoff_tracer_init,
 448         .reset          = irqsoff_tracer_reset,
 449         .open           = irqsoff_tracer_open,
 450         .close          = irqsoff_tracer_close,
 451         .ctrl_update    = irqsoff_tracer_ctrl_update,
 452         .print_max      = 1,
 453 #ifdef CONFIG_FTRACE_SELFTEST
 454         .selftest    = trace_selftest_startup_preemptoff,
 455 #endif
 456 };
 457 # define register_preemptoff(trace) register_tracer(&trace)
 458 #else
 459 # define register_preemptoff(trace) do { } while (0)
 460 #endif
 461
 462 #if defined(CONFIG_IRQSOFF_TRACER) && \
 463         defined(CONFIG_PREEMPT_TRACER)
 464
 465 static void preemptirqsoff_tracer_init(struct trace_array *tr)
 466 {
 467         trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
 468
 469         __irqsoff_tracer_init(tr);
 470 }
 471
 472 static struct tracer preemptirqsoff_tracer __read_mostly =
 473 {
 474         .name           = "preemptirqsoff",
 475         .init           = preemptirqsoff_tracer_init,
 476         .reset          = irqsoff_tracer_reset,
 477         .open           = irqsoff_tracer_open,
 478         .close          = irqsoff_tracer_close,
 479         .ctrl_update    = irqsoff_tracer_ctrl_update,
 480         .print_max      = 1,
 481 #ifdef CONFIG_FTRACE_SELFTEST
 482         .selftest    = trace_selftest_startup_preemptirqsoff,
 483 #endif
 484 };
 485
 486 # define register_preemptirqsoff(trace) register_tracer(&trace)
 487 #else
 488 # define register_preemptirqsoff(trace) do { } while (0)
 489 #endif
 490
 491 __init static int init_irqsoff_tracer(void)
 492 {
 493         register_irqsoff(irqsoff_tracer);
 494         register_preemptoff(preemptoff_tracer);
 495         register_preemptirqsoff(preemptirqsoff_tracer);
 496
 497         return 0;
 498 }
 499 device_initcall(init_irqsoff_tracer);