kernel/trace/trace_irqsoff.c

   1 /*
   2  * trace irqs off critical 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_RAW_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 static int save_flags;
  36
  37 static void stop_irqsoff_tracer(struct trace_array *tr, int graph);
  38 static int start_irqsoff_tracer(struct trace_array *tr, int graph);
  39
  40 #ifdef CONFIG_PREEMPT_TRACER
  41 static inline int
  42 preempt_trace(void)
  43 {
  44         return ((trace_type & TRACER_PREEMPT_OFF) && preempt_count());
  45 }
  46 #else
  47 # define preempt_trace() (0)
  48 #endif
  49
  50 #ifdef CONFIG_IRQSOFF_TRACER
  51 static inline int
  52 irq_trace(void)
  53 {
  54         return ((trace_type & TRACER_IRQS_OFF) &&
  55                 irqs_disabled());
  56 }
  57 #else
  58 # define irq_trace() (0)
  59 #endif
  60
  61 #define TRACE_DISPLAY_GRAPH     1
  62
  63 static struct tracer_opt trace_opts[] = {
  64 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
  65         /* display latency trace as call graph */
  66         { TRACER_OPT(display-graph, TRACE_DISPLAY_GRAPH) },
  67 #endif
  68         { } /* Empty entry */
  69 };
  70
  71 static struct tracer_flags tracer_flags = {
  72         .val  = 0,
  73         .opts = trace_opts,
  74 };
  75
  76 #define is_graph() (tracer_flags.val & TRACE_DISPLAY_GRAPH)
  77
  78 /*
  79  * Sequence count - we record it when starting a measurement and
  80  * skip the latency if the sequence has changed - some other section
  81  * did a maximum and could disturb our measurement with serial console
  82  * printouts, etc. Truly coinciding maximum latencies should be rare
  83  * and what happens together happens separately as well, so this doesn't
  84  * decrease the validity of the maximum found:
  85  */
  86 static __cacheline_aligned_in_smp       unsigned long max_sequence;
  87
  88 #ifdef CONFIG_FUNCTION_TRACER
  89 /*
  90  * Prologue for the preempt and irqs off function tracers.
  91  *
  92  * Returns 1 if it is OK to continue, and data->disabled is
  93  *            incremented.
  94  *         0 if the trace is to be ignored, and data->disabled
  95  *            is kept the same.
  96  *
  97  * Note, this function is also used outside this ifdef but
  98  *  inside the #ifdef of the function graph tracer below.
  99  *  This is OK, since the function graph tracer is
 100  *  dependent on the function tracer.
 101  */
 102 static int func_prolog_dec(struct trace_array *tr,
 103                            struct trace_array_cpu **data,
 104                            unsigned long *flags)
 105 {
 106         long disabled;
 107         int cpu;
 108
 109         /*
 110          * Does not matter if we preempt. We test the flags
 111          * afterward, to see if irqs are disabled or not.
 112          * If we preempt and get a false positive, the flags
 113          * test will fail.
 114          */
 115         cpu = raw_smp_processor_id();
 116         if (likely(!per_cpu(tracing_cpu, cpu)))
 117                 return 0;
 118
 119         local_save_flags(*flags);
 120         /*
 121          * Slight chance to get a false positive on tracing_cpu,
 122          * although I'm starting to think there isn't a chance.
 123          * Leave this for now just to be paranoid.
 124          */
 125         if (!irqs_disabled_flags(*flags) && !preempt_count())
 126                 return 0;
 127
 128         *data = tr->data[cpu];
 129         disabled = atomic_inc_return(&(*data)->disabled);
 130
 131         if (likely(disabled == 1))
 132                 return 1;
 133
 134         atomic_dec(&(*data)->disabled);
 135
 136         return 0;
 137 }
 138
 139 /*
 140  * irqsoff uses its own tracer function to keep the overhead down:
 141  */
 142 static void
 143 irqsoff_tracer_call(unsigned long ip, unsigned long parent_ip)
 144 {
 145         struct trace_array *tr = irqsoff_trace;
 146         struct trace_array_cpu *data;
 147         unsigned long flags;
 148
 149         if (!func_prolog_dec(tr, &data, &flags))
 150                 return;
 151
 152         trace_function(tr, ip, parent_ip, flags, preempt_count());
 153
 154         atomic_dec(&data->disabled);
 155 }
 156
 157 static struct ftrace_ops trace_ops __read_mostly =
 158 {
 159         .func = irqsoff_tracer_call,
 160         .flags = FTRACE_OPS_FL_GLOBAL,
 161 };
 162 #endif /* CONFIG_FUNCTION_TRACER */
 163
 164 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 165 static int irqsoff_set_flag(u32 old_flags, u32 bit, int set)
 166 {
 167         int cpu;
 168
 169         if (!(bit & TRACE_DISPLAY_GRAPH))
 170                 return -EINVAL;
 171
 172         if (!(is_graph() ^ set))
 173                 return 0;
 174
 175         stop_irqsoff_tracer(irqsoff_trace, !set);
 176
 177         for_each_possible_cpu(cpu)
 178                 per_cpu(tracing_cpu, cpu) = 0;
 179
 180         tracing_max_latency = 0;
 181         tracing_reset_online_cpus(irqsoff_trace);
 182
 183         return start_irqsoff_tracer(irqsoff_trace, set);
 184 }
 185
 186 static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
 187 {
 188         struct trace_array *tr = irqsoff_trace;
 189         struct trace_array_cpu *data;
 190         unsigned long flags;
 191         int ret;
 192         int pc;
 193
 194         if (!func_prolog_dec(tr, &data, &flags))
 195                 return 0;
 196
 197         pc = preempt_count();
 198         ret = __trace_graph_entry(tr, trace, flags, pc);
 199         atomic_dec(&data->disabled);
 200
 201         return ret;
 202 }
 203
 204 static void irqsoff_graph_return(struct ftrace_graph_ret *trace)
 205 {
 206         struct trace_array *tr = irqsoff_trace;
 207         struct trace_array_cpu *data;
 208         unsigned long flags;
 209         int pc;
 210
 211         if (!func_prolog_dec(tr, &data, &flags))
 212                 return;
 213
 214         pc = preempt_count();
 215         __trace_graph_return(tr, trace, flags, pc);
 216         atomic_dec(&data->disabled);
 217 }
 218
 219 static void irqsoff_trace_open(struct trace_iterator *iter)
 220 {
 221         if (is_graph())
 222                 graph_trace_open(iter);
 223
 224 }
 225
 226 static void irqsoff_trace_close(struct trace_iterator *iter)
 227 {
 228         if (iter->private)
 229                 graph_trace_close(iter);
 230 }
 231
 232 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_CPU | \
 233                             TRACE_GRAPH_PRINT_PROC | \
 234                             TRACE_GRAPH_PRINT_ABS_TIME | \
 235                             TRACE_GRAPH_PRINT_DURATION)
 236
 237 static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
 238 {
 239         /*
 240          * In graph mode call the graph tracer output function,
 241          * otherwise go with the TRACE_FN event handler
 242          */
 243         if (is_graph())
 244                 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS);
 245
 246         return TRACE_TYPE_UNHANDLED;
 247 }
 248
 249 static void irqsoff_print_header(struct seq_file *s)
 250 {
 251         if (is_graph())
 252                 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS);
 253         else
 254                 trace_default_header(s);
 255 }
 256
 257 static void
 258 __trace_function(struct trace_array *tr,
 259                  unsigned long ip, unsigned long parent_ip,
 260                  unsigned long flags, int pc)
 261 {
 262         if (is_graph())
 263                 trace_graph_function(tr, ip, parent_ip, flags, pc);
 264         else
 265                 trace_function(tr, ip, parent_ip, flags, pc);
 266 }
 267
 268 #else
 269 #define __trace_function trace_function
 270
 271 static int irqsoff_set_flag(u32 old_flags, u32 bit, int set)
 272 {
 273         return -EINVAL;
 274 }
 275
 276 static int irqsoff_graph_entry(struct ftrace_graph_ent *trace)
 277 {
 278         return -1;
 279 }
 280
 281 static enum print_line_t irqsoff_print_line(struct trace_iterator *iter)
 282 {
 283         return TRACE_TYPE_UNHANDLED;
 284 }
 285
 286 static void irqsoff_graph_return(struct ftrace_graph_ret *trace) { }
 287 static void irqsoff_print_header(struct seq_file *s) { }
 288 static void irqsoff_trace_open(struct trace_iterator *iter) { }
 289 static void irqsoff_trace_close(struct trace_iterator *iter) { }
 290 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
 291
 292 /*
 293  * Should this new latency be reported/recorded?
 294  */
 295 static int report_latency(cycle_t delta)
 296 {
 297         if (tracing_thresh) {
 298                 if (delta < tracing_thresh)
 299                         return 0;
 300         } else {
 301                 if (delta <= tracing_max_latency)
 302                         return 0;
 303         }
 304         return 1;
 305 }
 306
 307 static void
 308 check_critical_timing(struct trace_array *tr,
 309                       struct trace_array_cpu *data,
 310                       unsigned long parent_ip,
 311                       int cpu)
 312 {
 313         cycle_t T0, T1, delta;
 314         unsigned long flags;
 315         int pc;
 316
 317         T0 = data->preempt_timestamp;
 318         T1 = ftrace_now(cpu);
 319         delta = T1-T0;
 320
 321         local_save_flags(flags);
 322
 323         pc = preempt_count();
 324
 325         if (!report_latency(delta))
 326                 goto out;
 327
 328         raw_spin_lock_irqsave(&max_trace_lock, flags);
 329
 330         /* check if we are still the max latency */
 331         if (!report_latency(delta))
 332                 goto out_unlock;
 333
 334         __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
 335         /* Skip 5 functions to get to the irq/preempt enable function */
 336         __trace_stack(tr, flags, 5, pc);
 337
 338         if (data->critical_sequence != max_sequence)
 339                 goto out_unlock;
 340
 341         data->critical_end = parent_ip;
 342
 343         if (likely(!is_tracing_stopped())) {
 344                 tracing_max_latency = delta;
 345                 update_max_tr_single(tr, current, cpu);
 346         }
 347
 348         max_sequence++;
 349
 350 out_unlock:
 351         raw_spin_unlock_irqrestore(&max_trace_lock, flags);
 352
 353 out:
 354         data->critical_sequence = max_sequence;
 355         data->preempt_timestamp = ftrace_now(cpu);
 356         __trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc);
 357 }
 358
 359 static inline void
 360 start_critical_timing(unsigned long ip, unsigned long parent_ip)
 361 {
 362         int cpu;
 363         struct trace_array *tr = irqsoff_trace;
 364         struct trace_array_cpu *data;
 365         unsigned long flags;
 366
 367         if (likely(!tracer_enabled))
 368                 return;
 369
 370         cpu = raw_smp_processor_id();
 371
 372         if (per_cpu(tracing_cpu, cpu))
 373                 return;
 374
 375         data = tr->data[cpu];
 376
 377         if (unlikely(!data) || atomic_read(&data->disabled))
 378                 return;
 379
 380         atomic_inc(&data->disabled);
 381
 382         data->critical_sequence = max_sequence;
 383         data->preempt_timestamp = ftrace_now(cpu);
 384         data->critical_start = parent_ip ? : ip;
 385
 386         local_save_flags(flags);
 387
 388         __trace_function(tr, ip, parent_ip, flags, preempt_count());
 389
 390         per_cpu(tracing_cpu, cpu) = 1;
 391
 392         atomic_dec(&data->disabled);
 393 }
 394
 395 static inline void
 396 stop_critical_timing(unsigned long ip, unsigned long parent_ip)
 397 {
 398         int cpu;
 399         struct trace_array *tr = irqsoff_trace;
 400         struct trace_array_cpu *data;
 401         unsigned long flags;
 402
 403         cpu = raw_smp_processor_id();
 404         /* Always clear the tracing cpu on stopping the trace */
 405         if (unlikely(per_cpu(tracing_cpu, cpu)))
 406                 per_cpu(tracing_cpu, cpu) = 0;
 407         else
 408                 return;
 409
 410         if (!tracer_enabled)
 411                 return;
 412
 413         data = tr->data[cpu];
 414
 415         if (unlikely(!data) ||
 416             !data->critical_start || atomic_read(&data->disabled))
 417                 return;
 418
 419         atomic_inc(&data->disabled);
 420
 421         local_save_flags(flags);
 422         __trace_function(tr, ip, parent_ip, flags, preempt_count());
 423         check_critical_timing(tr, data, parent_ip ? : ip, cpu);
 424         data->critical_start = 0;
 425         atomic_dec(&data->disabled);
 426 }
 427
 428 /* start and stop critical timings used to for stoppage (in idle) */
 429 void start_critical_timings(void)
 430 {
 431         if (preempt_trace() || irq_trace())
 432                 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 433 }
 434 EXPORT_SYMBOL_GPL(start_critical_timings);
 435
 436 void stop_critical_timings(void)
 437 {
 438         if (preempt_trace() || irq_trace())
 439                 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 440 }
 441 EXPORT_SYMBOL_GPL(stop_critical_timings);
 442
 443 #ifdef CONFIG_IRQSOFF_TRACER
 444 #ifdef CONFIG_PROVE_LOCKING
 445 void time_hardirqs_on(unsigned long a0, unsigned long a1)
 446 {
 447         if (!preempt_trace() && irq_trace())
 448                 stop_critical_timing(a0, a1);
 449 }
 450
 451 void time_hardirqs_off(unsigned long a0, unsigned long a1)
 452 {
 453         if (!preempt_trace() && irq_trace())
 454                 start_critical_timing(a0, a1);
 455 }
 456
 457 #else /* !CONFIG_PROVE_LOCKING */
 458
 459 /*
 460  * Stubs:
 461  */
 462
 463 void trace_softirqs_on(unsigned long ip)
 464 {
 465 }
 466
 467 void trace_softirqs_off(unsigned long ip)
 468 {
 469 }
 470
 471 inline void print_irqtrace_events(struct task_struct *curr)
 472 {
 473 }
 474
 475 /*
 476  * We are only interested in hardirq on/off events:
 477  */
 478 void trace_hardirqs_on(void)
 479 {
 480         if (!preempt_trace() && irq_trace())
 481                 stop_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 482 }
 483 EXPORT_SYMBOL(trace_hardirqs_on);
 484
 485 void trace_hardirqs_off(void)
 486 {
 487         if (!preempt_trace() && irq_trace())
 488                 start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
 489 }
 490 EXPORT_SYMBOL(trace_hardirqs_off);
 491
 492 void trace_hardirqs_on_caller(unsigned long caller_addr)
 493 {
 494         if (!preempt_trace() && irq_trace())
 495                 stop_critical_timing(CALLER_ADDR0, caller_addr);
 496 }
 497 EXPORT_SYMBOL(trace_hardirqs_on_caller);
 498
 499 void trace_hardirqs_off_caller(unsigned long caller_addr)
 500 {
 501         if (!preempt_trace() && irq_trace())
 502                 start_critical_timing(CALLER_ADDR0, caller_addr);
 503 }
 504 EXPORT_SYMBOL(trace_hardirqs_off_caller);
 505
 506 #endif /* CONFIG_PROVE_LOCKING */
 507 #endif /*  CONFIG_IRQSOFF_TRACER */
 508
 509 #ifdef CONFIG_PREEMPT_TRACER
 510 void trace_preempt_on(unsigned long a0, unsigned long a1)
 511 {
 512         if (preempt_trace() && !irq_trace())
 513                 stop_critical_timing(a0, a1);
 514 }
 515
 516 void trace_preempt_off(unsigned long a0, unsigned long a1)
 517 {
 518         if (preempt_trace() && !irq_trace())
 519                 start_critical_timing(a0, a1);
 520 }
 521 #endif /* CONFIG_PREEMPT_TRACER */
 522
 523 static int start_irqsoff_tracer(struct trace_array *tr, int graph)
 524 {
 525         int ret = 0;
 526
 527         if (!graph)
 528                 ret = register_ftrace_function(&trace_ops);
 529         else
 530                 ret = register_ftrace_graph(&irqsoff_graph_return,
 531                                             &irqsoff_graph_entry);
 532
 533         if (!ret && tracing_is_enabled())
 534                 tracer_enabled = 1;
 535         else
 536                 tracer_enabled = 0;
 537
 538         return ret;
 539 }
 540
 541 static void stop_irqsoff_tracer(struct trace_array *tr, int graph)
 542 {
 543         tracer_enabled = 0;
 544
 545         if (!graph)
 546                 unregister_ftrace_function(&trace_ops);
 547         else
 548                 unregister_ftrace_graph();
 549 }
 550
 551 static void __irqsoff_tracer_init(struct trace_array *tr)
 552 {
 553         save_flags = trace_flags;
 554
 555         /* non overwrite screws up the latency tracers */
 556         set_tracer_flag(TRACE_ITER_OVERWRITE, 1);
 557         set_tracer_flag(TRACE_ITER_LATENCY_FMT, 1);
 558
 559         tracing_max_latency = 0;
 560         irqsoff_trace = tr;
 561         /* make sure that the tracer is visible */
 562         smp_wmb();
 563         tracing_reset_online_cpus(tr);
 564
 565         if (start_irqsoff_tracer(tr, is_graph()))
 566                 printk(KERN_ERR "failed to start irqsoff tracer\n");
 567 }
 568
 569 static void irqsoff_tracer_reset(struct trace_array *tr)
 570 {
 571         int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT;
 572         int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE;
 573
 574         stop_irqsoff_tracer(tr, is_graph());
 575
 576         set_tracer_flag(TRACE_ITER_LATENCY_FMT, lat_flag);
 577         set_tracer_flag(TRACE_ITER_OVERWRITE, overwrite_flag);
 578 }
 579
 580 static void irqsoff_tracer_start(struct trace_array *tr)
 581 {
 582         tracer_enabled = 1;
 583 }
 584
 585 static void irqsoff_tracer_stop(struct trace_array *tr)
 586 {
 587         tracer_enabled = 0;
 588 }
 589
 590 #ifdef CONFIG_IRQSOFF_TRACER
 591 static int irqsoff_tracer_init(struct trace_array *tr)
 592 {
 593         trace_type = TRACER_IRQS_OFF;
 594
 595         __irqsoff_tracer_init(tr);
 596         return 0;
 597 }
 598 static struct tracer irqsoff_tracer __read_mostly =
 599 {
 600         .name           = "irqsoff",
 601         .init           = irqsoff_tracer_init,
 602         .reset          = irqsoff_tracer_reset,
 603         .start          = irqsoff_tracer_start,
 604         .stop           = irqsoff_tracer_stop,
 605         .print_max      = 1,
 606         .print_header   = irqsoff_print_header,
 607         .print_line     = irqsoff_print_line,
 608         .flags          = &tracer_flags,
 609         .set_flag       = irqsoff_set_flag,
 610         .flag_changed   = trace_keep_overwrite,
 611 #ifdef CONFIG_FTRACE_SELFTEST
 612         .selftest    = trace_selftest_startup_irqsoff,
 613 #endif
 614         .open           = irqsoff_trace_open,
 615         .close          = irqsoff_trace_close,
 616         .use_max_tr     = 1,
 617 };
 618 # define register_irqsoff(trace) register_tracer(&trace)
 619 #else
 620 # define register_irqsoff(trace) do { } while (0)
 621 #endif
 622
 623 #ifdef CONFIG_PREEMPT_TRACER
 624 static int preemptoff_tracer_init(struct trace_array *tr)
 625 {
 626         trace_type = TRACER_PREEMPT_OFF;
 627
 628         __irqsoff_tracer_init(tr);
 629         return 0;
 630 }
 631
 632 static struct tracer preemptoff_tracer __read_mostly =
 633 {
 634         .name           = "preemptoff",
 635         .init           = preemptoff_tracer_init,
 636         .reset          = irqsoff_tracer_reset,
 637         .start          = irqsoff_tracer_start,
 638         .stop           = irqsoff_tracer_stop,
 639         .print_max      = 1,
 640         .print_header   = irqsoff_print_header,
 641         .print_line     = irqsoff_print_line,
 642         .flags          = &tracer_flags,
 643         .set_flag       = irqsoff_set_flag,
 644         .flag_changed   = trace_keep_overwrite,
 645 #ifdef CONFIG_FTRACE_SELFTEST
 646         .selftest    = trace_selftest_startup_preemptoff,
 647 #endif
 648         .open           = irqsoff_trace_open,
 649         .close          = irqsoff_trace_close,
 650         .use_max_tr     = 1,
 651 };
 652 # define register_preemptoff(trace) register_tracer(&trace)
 653 #else
 654 # define register_preemptoff(trace) do { } while (0)
 655 #endif
 656
 657 #if defined(CONFIG_IRQSOFF_TRACER) && \
 658         defined(CONFIG_PREEMPT_TRACER)
 659
 660 static int preemptirqsoff_tracer_init(struct trace_array *tr)
 661 {
 662         trace_type = TRACER_IRQS_OFF | TRACER_PREEMPT_OFF;
 663
 664         __irqsoff_tracer_init(tr);
 665         return 0;
 666 }
 667
 668 static struct tracer preemptirqsoff_tracer __read_mostly =
 669 {
 670         .name           = "preemptirqsoff",
 671         .init           = preemptirqsoff_tracer_init,
 672         .reset          = irqsoff_tracer_reset,
 673         .start          = irqsoff_tracer_start,
 674         .stop           = irqsoff_tracer_stop,
 675         .print_max      = 1,
 676         .print_header   = irqsoff_print_header,
 677         .print_line     = irqsoff_print_line,
 678         .flags          = &tracer_flags,
 679         .set_flag       = irqsoff_set_flag,
 680         .flag_changed   = trace_keep_overwrite,
 681 #ifdef CONFIG_FTRACE_SELFTEST
 682         .selftest    = trace_selftest_startup_preemptirqsoff,
 683 #endif
 684         .open           = irqsoff_trace_open,
 685         .close          = irqsoff_trace_close,
 686         .use_max_tr     = 1,
 687 };
 688
 689 # define register_preemptirqsoff(trace) register_tracer(&trace)
 690 #else
 691 # define register_preemptirqsoff(trace) do { } while (0)
 692 #endif
 693
 694 __init static int init_irqsoff_tracer(void)
 695 {
 696         register_irqsoff(irqsoff_tracer);
 697         register_preemptoff(preemptoff_tracer);
 698         register_preemptirqsoff(preemptirqsoff_tracer);
 699
 700         return 0;
 701 }
 702 device_initcall(init_irqsoff_tracer);