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 Nadia Yvette Chambers 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 #include <linux/sched/rt.h> 19 #include <linux/sched/deadline.h> 20 #include <trace/events/sched.h> 21 #include "trace.h" 22 23 static struct trace_array *wakeup_trace; 24 static int __read_mostly tracer_enabled; 25 26 static struct task_struct *wakeup_task; 27 static int wakeup_cpu; 28 static int wakeup_current_cpu; 29 static unsigned wakeup_prio = -1; 30 static int wakeup_rt; 31 static int wakeup_dl; 32 static int tracing_dl = 0; 33 34 static arch_spinlock_t wakeup_lock = 35 (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED; 36 37 static void wakeup_reset(struct trace_array *tr); 38 static void __wakeup_reset(struct trace_array *tr); 39 static int wakeup_graph_entry(struct ftrace_graph_ent *trace); 40 static void wakeup_graph_return(struct ftrace_graph_ret *trace); 41 42 static int save_flags; 43 static bool function_enabled; 44 45 #define TRACE_DISPLAY_GRAPH 1 46 47 static struct tracer_opt trace_opts[] = { 48 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 49 /* display latency trace as call graph */ 50 { TRACER_OPT(display-graph, TRACE_DISPLAY_GRAPH) }, 51 #endif 52 { } /* Empty entry */ 53 }; 54 55 static struct tracer_flags tracer_flags = { 56 .val = 0, 57 .opts = trace_opts, 58 }; 59 60 #define is_graph() (tracer_flags.val & TRACE_DISPLAY_GRAPH) 61 62 #ifdef CONFIG_FUNCTION_TRACER 63 64 /* 65 * Prologue for the wakeup function tracers. 66 * 67 * Returns 1 if it is OK to continue, and preemption 68 * is disabled and data->disabled is incremented. 69 * 0 if the trace is to be ignored, and preemption 70 * is not disabled and data->disabled is 71 * kept the same. 72 * 73 * Note, this function is also used outside this ifdef but 74 * inside the #ifdef of the function graph tracer below. 75 * This is OK, since the function graph tracer is 76 * dependent on the function tracer. 77 */ 78 static int 79 func_prolog_preempt_disable(struct trace_array *tr, 80 struct trace_array_cpu **data, 81 int *pc) 82 { 83 long disabled; 84 int cpu; 85 86 if (likely(!wakeup_task)) 87 return 0; 88 89 *pc = preempt_count(); 90 preempt_disable_notrace(); 91 92 cpu = raw_smp_processor_id(); 93 if (cpu != wakeup_current_cpu) 94 goto out_enable; 95 96 *data = per_cpu_ptr(tr->trace_buffer.data, cpu); 97 disabled = atomic_inc_return(&(*data)->disabled); 98 if (unlikely(disabled != 1)) 99 goto out; 100 101 return 1; 102 103 out: 104 atomic_dec(&(*data)->disabled); 105 106 out_enable: 107 preempt_enable_notrace(); 108 return 0; 109 } 110 111 /* 112 * wakeup uses its own tracer function to keep the overhead down: 113 */ 114 static void 115 wakeup_tracer_call(unsigned long ip, unsigned long parent_ip, 116 struct ftrace_ops *op, struct pt_regs *pt_regs) 117 { 118 struct trace_array *tr = wakeup_trace; 119 struct trace_array_cpu *data; 120 unsigned long flags; 121 int pc; 122 123 if (!func_prolog_preempt_disable(tr, &data, &pc)) 124 return; 125 126 local_irq_save(flags); 127 trace_function(tr, ip, parent_ip, flags, pc); 128 local_irq_restore(flags); 129 130 atomic_dec(&data->disabled); 131 preempt_enable_notrace(); 132 } 133 #endif /* CONFIG_FUNCTION_TRACER */ 134 135 static int register_wakeup_function(struct trace_array *tr, int graph, int set) 136 { 137 int ret; 138 139 /* 'set' is set if TRACE_ITER_FUNCTION is about to be set */ 140 if (function_enabled || (!set && !(trace_flags & TRACE_ITER_FUNCTION))) 141 return 0; 142 143 if (graph) 144 ret = register_ftrace_graph(&wakeup_graph_return, 145 &wakeup_graph_entry); 146 else 147 ret = register_ftrace_function(tr->ops); 148 149 if (!ret) 150 function_enabled = true; 151 152 return ret; 153 } 154 155 static void unregister_wakeup_function(struct trace_array *tr, int graph) 156 { 157 if (!function_enabled) 158 return; 159 160 if (graph) 161 unregister_ftrace_graph(); 162 else 163 unregister_ftrace_function(tr->ops); 164 165 function_enabled = false; 166 } 167 168 static void wakeup_function_set(struct trace_array *tr, int set) 169 { 170 if (set) 171 register_wakeup_function(tr, is_graph(), 1); 172 else 173 unregister_wakeup_function(tr, is_graph()); 174 } 175 176 static int wakeup_flag_changed(struct trace_array *tr, u32 mask, int set) 177 { 178 struct tracer *tracer = tr->current_trace; 179 180 if (mask & TRACE_ITER_FUNCTION) 181 wakeup_function_set(tr, set); 182 183 return trace_keep_overwrite(tracer, mask, set); 184 } 185 186 static int start_func_tracer(struct trace_array *tr, int graph) 187 { 188 int ret; 189 190 ret = register_wakeup_function(tr, graph, 0); 191 192 if (!ret && tracing_is_enabled()) 193 tracer_enabled = 1; 194 else 195 tracer_enabled = 0; 196 197 return ret; 198 } 199 200 static void stop_func_tracer(struct trace_array *tr, int graph) 201 { 202 tracer_enabled = 0; 203 204 unregister_wakeup_function(tr, graph); 205 } 206 207 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 208 static int 209 wakeup_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set) 210 { 211 212 if (!(bit & TRACE_DISPLAY_GRAPH)) 213 return -EINVAL; 214 215 if (!(is_graph() ^ set)) 216 return 0; 217 218 stop_func_tracer(tr, !set); 219 220 wakeup_reset(wakeup_trace); 221 tr->max_latency = 0; 222 223 return start_func_tracer(tr, set); 224 } 225 226 static int wakeup_graph_entry(struct ftrace_graph_ent *trace) 227 { 228 struct trace_array *tr = wakeup_trace; 229 struct trace_array_cpu *data; 230 unsigned long flags; 231 int pc, ret = 0; 232 233 if (!func_prolog_preempt_disable(tr, &data, &pc)) 234 return 0; 235 236 local_save_flags(flags); 237 ret = __trace_graph_entry(tr, trace, flags, pc); 238 atomic_dec(&data->disabled); 239 preempt_enable_notrace(); 240 241 return ret; 242 } 243 244 static void wakeup_graph_return(struct ftrace_graph_ret *trace) 245 { 246 struct trace_array *tr = wakeup_trace; 247 struct trace_array_cpu *data; 248 unsigned long flags; 249 int pc; 250 251 if (!func_prolog_preempt_disable(tr, &data, &pc)) 252 return; 253 254 local_save_flags(flags); 255 __trace_graph_return(tr, trace, flags, pc); 256 atomic_dec(&data->disabled); 257 258 preempt_enable_notrace(); 259 return; 260 } 261 262 static void wakeup_trace_open(struct trace_iterator *iter) 263 { 264 if (is_graph()) 265 graph_trace_open(iter); 266 } 267 268 static void wakeup_trace_close(struct trace_iterator *iter) 269 { 270 if (iter->private) 271 graph_trace_close(iter); 272 } 273 274 #define GRAPH_TRACER_FLAGS (TRACE_GRAPH_PRINT_PROC | \ 275 TRACE_GRAPH_PRINT_ABS_TIME | \ 276 TRACE_GRAPH_PRINT_DURATION) 277 278 static enum print_line_t wakeup_print_line(struct trace_iterator *iter) 279 { 280 /* 281 * In graph mode call the graph tracer output function, 282 * otherwise go with the TRACE_FN event handler 283 */ 284 if (is_graph()) 285 return print_graph_function_flags(iter, GRAPH_TRACER_FLAGS); 286 287 return TRACE_TYPE_UNHANDLED; 288 } 289 290 static void wakeup_print_header(struct seq_file *s) 291 { 292 if (is_graph()) 293 print_graph_headers_flags(s, GRAPH_TRACER_FLAGS); 294 else 295 trace_default_header(s); 296 } 297 298 static void 299 __trace_function(struct trace_array *tr, 300 unsigned long ip, unsigned long parent_ip, 301 unsigned long flags, int pc) 302 { 303 if (is_graph()) 304 trace_graph_function(tr, ip, parent_ip, flags, pc); 305 else 306 trace_function(tr, ip, parent_ip, flags, pc); 307 } 308 #else 309 #define __trace_function trace_function 310 311 static int 312 wakeup_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set) 313 { 314 return -EINVAL; 315 } 316 317 static int wakeup_graph_entry(struct ftrace_graph_ent *trace) 318 { 319 return -1; 320 } 321 322 static enum print_line_t wakeup_print_line(struct trace_iterator *iter) 323 { 324 return TRACE_TYPE_UNHANDLED; 325 } 326 327 static void wakeup_graph_return(struct ftrace_graph_ret *trace) { } 328 static void wakeup_trace_open(struct trace_iterator *iter) { } 329 static void wakeup_trace_close(struct trace_iterator *iter) { } 330 331 #ifdef CONFIG_FUNCTION_TRACER 332 static void wakeup_print_header(struct seq_file *s) 333 { 334 trace_default_header(s); 335 } 336 #else 337 static void wakeup_print_header(struct seq_file *s) 338 { 339 trace_latency_header(s); 340 } 341 #endif /* CONFIG_FUNCTION_TRACER */ 342 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ 343 344 /* 345 * Should this new latency be reported/recorded? 346 */ 347 static int report_latency(struct trace_array *tr, cycle_t delta) 348 { 349 if (tracing_thresh) { 350 if (delta < tracing_thresh) 351 return 0; 352 } else { 353 if (delta <= tr->max_latency) 354 return 0; 355 } 356 return 1; 357 } 358 359 static void 360 probe_wakeup_migrate_task(void *ignore, struct task_struct *task, int cpu) 361 { 362 if (task != wakeup_task) 363 return; 364 365 wakeup_current_cpu = cpu; 366 } 367 368 static void notrace 369 probe_wakeup_sched_switch(void *ignore, 370 struct task_struct *prev, struct task_struct *next) 371 { 372 struct trace_array_cpu *data; 373 cycle_t T0, T1, delta; 374 unsigned long flags; 375 long disabled; 376 int cpu; 377 int pc; 378 379 tracing_record_cmdline(prev); 380 381 if (unlikely(!tracer_enabled)) 382 return; 383 384 /* 385 * When we start a new trace, we set wakeup_task to NULL 386 * and then set tracer_enabled = 1. We want to make sure 387 * that another CPU does not see the tracer_enabled = 1 388 * and the wakeup_task with an older task, that might 389 * actually be the same as next. 390 */ 391 smp_rmb(); 392 393 if (next != wakeup_task) 394 return; 395 396 pc = preempt_count(); 397 398 /* disable local data, not wakeup_cpu data */ 399 cpu = raw_smp_processor_id(); 400 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled); 401 if (likely(disabled != 1)) 402 goto out; 403 404 local_irq_save(flags); 405 arch_spin_lock(&wakeup_lock); 406 407 /* We could race with grabbing wakeup_lock */ 408 if (unlikely(!tracer_enabled || next != wakeup_task)) 409 goto out_unlock; 410 411 /* The task we are waiting for is waking up */ 412 data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu); 413 414 __trace_function(wakeup_trace, CALLER_ADDR0, CALLER_ADDR1, flags, pc); 415 tracing_sched_switch_trace(wakeup_trace, prev, next, flags, pc); 416 417 T0 = data->preempt_timestamp; 418 T1 = ftrace_now(cpu); 419 delta = T1-T0; 420 421 if (!report_latency(wakeup_trace, delta)) 422 goto out_unlock; 423 424 if (likely(!is_tracing_stopped())) { 425 wakeup_trace->max_latency = delta; 426 update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu); 427 } 428 429 out_unlock: 430 __wakeup_reset(wakeup_trace); 431 arch_spin_unlock(&wakeup_lock); 432 local_irq_restore(flags); 433 out: 434 atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled); 435 } 436 437 static void __wakeup_reset(struct trace_array *tr) 438 { 439 wakeup_cpu = -1; 440 wakeup_prio = -1; 441 tracing_dl = 0; 442 443 if (wakeup_task) 444 put_task_struct(wakeup_task); 445 446 wakeup_task = NULL; 447 } 448 449 static void wakeup_reset(struct trace_array *tr) 450 { 451 unsigned long flags; 452 453 tracing_reset_online_cpus(&tr->trace_buffer); 454 455 local_irq_save(flags); 456 arch_spin_lock(&wakeup_lock); 457 __wakeup_reset(tr); 458 arch_spin_unlock(&wakeup_lock); 459 local_irq_restore(flags); 460 } 461 462 static void 463 probe_wakeup(void *ignore, struct task_struct *p, int success) 464 { 465 struct trace_array_cpu *data; 466 int cpu = smp_processor_id(); 467 unsigned long flags; 468 long disabled; 469 int pc; 470 471 if (likely(!tracer_enabled)) 472 return; 473 474 tracing_record_cmdline(p); 475 tracing_record_cmdline(current); 476 477 /* 478 * Semantic is like this: 479 * - wakeup tracer handles all tasks in the system, independently 480 * from their scheduling class; 481 * - wakeup_rt tracer handles tasks belonging to sched_dl and 482 * sched_rt class; 483 * - wakeup_dl handles tasks belonging to sched_dl class only. 484 */ 485 if (tracing_dl || (wakeup_dl && !dl_task(p)) || 486 (wakeup_rt && !dl_task(p) && !rt_task(p)) || 487 (!dl_task(p) && (p->prio >= wakeup_prio || p->prio >= current->prio))) 488 return; 489 490 pc = preempt_count(); 491 disabled = atomic_inc_return(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled); 492 if (unlikely(disabled != 1)) 493 goto out; 494 495 /* interrupts should be off from try_to_wake_up */ 496 arch_spin_lock(&wakeup_lock); 497 498 /* check for races. */ 499 if (!tracer_enabled || tracing_dl || 500 (!dl_task(p) && p->prio >= wakeup_prio)) 501 goto out_locked; 502 503 /* reset the trace */ 504 __wakeup_reset(wakeup_trace); 505 506 wakeup_cpu = task_cpu(p); 507 wakeup_current_cpu = wakeup_cpu; 508 wakeup_prio = p->prio; 509 510 /* 511 * Once you start tracing a -deadline task, don't bother tracing 512 * another task until the first one wakes up. 513 */ 514 if (dl_task(p)) 515 tracing_dl = 1; 516 else 517 tracing_dl = 0; 518 519 wakeup_task = p; 520 get_task_struct(wakeup_task); 521 522 local_save_flags(flags); 523 524 data = per_cpu_ptr(wakeup_trace->trace_buffer.data, wakeup_cpu); 525 data->preempt_timestamp = ftrace_now(cpu); 526 tracing_sched_wakeup_trace(wakeup_trace, p, current, flags, pc); 527 528 /* 529 * We must be careful in using CALLER_ADDR2. But since wake_up 530 * is not called by an assembly function (where as schedule is) 531 * it should be safe to use it here. 532 */ 533 __trace_function(wakeup_trace, CALLER_ADDR1, CALLER_ADDR2, flags, pc); 534 535 out_locked: 536 arch_spin_unlock(&wakeup_lock); 537 out: 538 atomic_dec(&per_cpu_ptr(wakeup_trace->trace_buffer.data, cpu)->disabled); 539 } 540 541 static void start_wakeup_tracer(struct trace_array *tr) 542 { 543 int ret; 544 545 ret = register_trace_sched_wakeup(probe_wakeup, NULL); 546 if (ret) { 547 pr_info("wakeup trace: Couldn't activate tracepoint" 548 " probe to kernel_sched_wakeup\n"); 549 return; 550 } 551 552 ret = register_trace_sched_wakeup_new(probe_wakeup, NULL); 553 if (ret) { 554 pr_info("wakeup trace: Couldn't activate tracepoint" 555 " probe to kernel_sched_wakeup_new\n"); 556 goto fail_deprobe; 557 } 558 559 ret = register_trace_sched_switch(probe_wakeup_sched_switch, NULL); 560 if (ret) { 561 pr_info("sched trace: Couldn't activate tracepoint" 562 " probe to kernel_sched_switch\n"); 563 goto fail_deprobe_wake_new; 564 } 565 566 ret = register_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL); 567 if (ret) { 568 pr_info("wakeup trace: Couldn't activate tracepoint" 569 " probe to kernel_sched_migrate_task\n"); 570 return; 571 } 572 573 wakeup_reset(tr); 574 575 /* 576 * Don't let the tracer_enabled = 1 show up before 577 * the wakeup_task is reset. This may be overkill since 578 * wakeup_reset does a spin_unlock after setting the 579 * wakeup_task to NULL, but I want to be safe. 580 * This is a slow path anyway. 581 */ 582 smp_wmb(); 583 584 if (start_func_tracer(tr, is_graph())) 585 printk(KERN_ERR "failed to start wakeup tracer\n"); 586 587 return; 588 fail_deprobe_wake_new: 589 unregister_trace_sched_wakeup_new(probe_wakeup, NULL); 590 fail_deprobe: 591 unregister_trace_sched_wakeup(probe_wakeup, NULL); 592 } 593 594 static void stop_wakeup_tracer(struct trace_array *tr) 595 { 596 tracer_enabled = 0; 597 stop_func_tracer(tr, is_graph()); 598 unregister_trace_sched_switch(probe_wakeup_sched_switch, NULL); 599 unregister_trace_sched_wakeup_new(probe_wakeup, NULL); 600 unregister_trace_sched_wakeup(probe_wakeup, NULL); 601 unregister_trace_sched_migrate_task(probe_wakeup_migrate_task, NULL); 602 } 603 604 static bool wakeup_busy; 605 606 static int __wakeup_tracer_init(struct trace_array *tr) 607 { 608 save_flags = trace_flags; 609 610 /* non overwrite screws up the latency tracers */ 611 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, 1); 612 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, 1); 613 614 tr->max_latency = 0; 615 wakeup_trace = tr; 616 ftrace_init_array_ops(tr, wakeup_tracer_call); 617 start_wakeup_tracer(tr); 618 619 wakeup_busy = true; 620 return 0; 621 } 622 623 static int wakeup_tracer_init(struct trace_array *tr) 624 { 625 if (wakeup_busy) 626 return -EBUSY; 627 628 wakeup_dl = 0; 629 wakeup_rt = 0; 630 return __wakeup_tracer_init(tr); 631 } 632 633 static int wakeup_rt_tracer_init(struct trace_array *tr) 634 { 635 if (wakeup_busy) 636 return -EBUSY; 637 638 wakeup_dl = 0; 639 wakeup_rt = 1; 640 return __wakeup_tracer_init(tr); 641 } 642 643 static int wakeup_dl_tracer_init(struct trace_array *tr) 644 { 645 if (wakeup_busy) 646 return -EBUSY; 647 648 wakeup_dl = 1; 649 wakeup_rt = 0; 650 return __wakeup_tracer_init(tr); 651 } 652 653 static void wakeup_tracer_reset(struct trace_array *tr) 654 { 655 int lat_flag = save_flags & TRACE_ITER_LATENCY_FMT; 656 int overwrite_flag = save_flags & TRACE_ITER_OVERWRITE; 657 658 stop_wakeup_tracer(tr); 659 /* make sure we put back any tasks we are tracing */ 660 wakeup_reset(tr); 661 662 set_tracer_flag(tr, TRACE_ITER_LATENCY_FMT, lat_flag); 663 set_tracer_flag(tr, TRACE_ITER_OVERWRITE, overwrite_flag); 664 ftrace_reset_array_ops(tr); 665 wakeup_busy = false; 666 } 667 668 static void wakeup_tracer_start(struct trace_array *tr) 669 { 670 wakeup_reset(tr); 671 tracer_enabled = 1; 672 } 673 674 static void wakeup_tracer_stop(struct trace_array *tr) 675 { 676 tracer_enabled = 0; 677 } 678 679 static struct tracer wakeup_tracer __read_mostly = 680 { 681 .name = "wakeup", 682 .init = wakeup_tracer_init, 683 .reset = wakeup_tracer_reset, 684 .start = wakeup_tracer_start, 685 .stop = wakeup_tracer_stop, 686 .print_max = true, 687 .print_header = wakeup_print_header, 688 .print_line = wakeup_print_line, 689 .flags = &tracer_flags, 690 .set_flag = wakeup_set_flag, 691 .flag_changed = wakeup_flag_changed, 692 #ifdef CONFIG_FTRACE_SELFTEST 693 .selftest = trace_selftest_startup_wakeup, 694 #endif 695 .open = wakeup_trace_open, 696 .close = wakeup_trace_close, 697 .allow_instances = true, 698 .use_max_tr = true, 699 }; 700 701 static struct tracer wakeup_rt_tracer __read_mostly = 702 { 703 .name = "wakeup_rt", 704 .init = wakeup_rt_tracer_init, 705 .reset = wakeup_tracer_reset, 706 .start = wakeup_tracer_start, 707 .stop = wakeup_tracer_stop, 708 .print_max = true, 709 .print_header = wakeup_print_header, 710 .print_line = wakeup_print_line, 711 .flags = &tracer_flags, 712 .set_flag = wakeup_set_flag, 713 .flag_changed = wakeup_flag_changed, 714 #ifdef CONFIG_FTRACE_SELFTEST 715 .selftest = trace_selftest_startup_wakeup, 716 #endif 717 .open = wakeup_trace_open, 718 .close = wakeup_trace_close, 719 .allow_instances = true, 720 .use_max_tr = true, 721 }; 722 723 static struct tracer wakeup_dl_tracer __read_mostly = 724 { 725 .name = "wakeup_dl", 726 .init = wakeup_dl_tracer_init, 727 .reset = wakeup_tracer_reset, 728 .start = wakeup_tracer_start, 729 .stop = wakeup_tracer_stop, 730 .print_max = true, 731 .print_header = wakeup_print_header, 732 .print_line = wakeup_print_line, 733 .flags = &tracer_flags, 734 .set_flag = wakeup_set_flag, 735 .flag_changed = wakeup_flag_changed, 736 #ifdef CONFIG_FTRACE_SELFTEST 737 .selftest = trace_selftest_startup_wakeup, 738 #endif 739 .open = wakeup_trace_open, 740 .close = wakeup_trace_close, 741 .use_max_tr = true, 742 }; 743 744 __init static int init_wakeup_tracer(void) 745 { 746 int ret; 747 748 ret = register_tracer(&wakeup_tracer); 749 if (ret) 750 return ret; 751 752 ret = register_tracer(&wakeup_rt_tracer); 753 if (ret) 754 return ret; 755 756 ret = register_tracer(&wakeup_dl_tracer); 757 if (ret) 758 return ret; 759 760 return 0; 761 } 762 core_initcall(init_wakeup_tracer); 763