1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * OS Noise Tracer: computes the OS Noise suffered by a running thread. 4 * Timerlat Tracer: measures the wakeup latency of a timer triggered IRQ and thread. 5 * 6 * Based on "hwlat_detector" tracer by: 7 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com> 8 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com> 9 * With feedback from Clark Williams <williams@redhat.com> 10 * 11 * And also based on the rtsl tracer presented on: 12 * DE OLIVEIRA, Daniel Bristot, et al. Demystifying the real-time linux 13 * scheduling latency. In: 32nd Euromicro Conference on Real-Time Systems 14 * (ECRTS 2020). Schloss Dagstuhl-Leibniz-Zentrum fur Informatik, 2020. 15 * 16 * Copyright (C) 2021 Daniel Bristot de Oliveira, Red Hat, Inc. <bristot@redhat.com> 17 */ 18 19 #include <linux/kthread.h> 20 #include <linux/tracefs.h> 21 #include <linux/uaccess.h> 22 #include <linux/cpumask.h> 23 #include <linux/delay.h> 24 #include <linux/sched/clock.h> 25 #include <uapi/linux/sched/types.h> 26 #include <linux/sched.h> 27 #include "trace.h" 28 29 #ifdef CONFIG_X86_LOCAL_APIC 30 #include <asm/trace/irq_vectors.h> 31 #undef TRACE_INCLUDE_PATH 32 #undef TRACE_INCLUDE_FILE 33 #endif /* CONFIG_X86_LOCAL_APIC */ 34 35 #include <trace/events/irq.h> 36 #include <trace/events/sched.h> 37 38 #define CREATE_TRACE_POINTS 39 #include <trace/events/osnoise.h> 40 41 /* 42 * Default values. 43 */ 44 #define BANNER "osnoise: " 45 #define DEFAULT_SAMPLE_PERIOD 1000000 /* 1s */ 46 #define DEFAULT_SAMPLE_RUNTIME 1000000 /* 1s */ 47 48 #define DEFAULT_TIMERLAT_PERIOD 1000 /* 1ms */ 49 #define DEFAULT_TIMERLAT_PRIO 95 /* FIFO 95 */ 50 51 /* 52 * trace_array of the enabled osnoise/timerlat instances. 53 */ 54 struct osnoise_instance { 55 struct list_head list; 56 struct trace_array *tr; 57 }; 58 struct list_head osnoise_instances; 59 60 static bool osnoise_has_registered_instances(void) 61 { 62 return !!list_first_or_null_rcu(&osnoise_instances, 63 struct osnoise_instance, 64 list); 65 } 66 67 /* 68 * osnoise_instance_registered - check if a tr is already registered 69 */ 70 static int osnoise_instance_registered(struct trace_array *tr) 71 { 72 struct osnoise_instance *inst; 73 int found = 0; 74 75 rcu_read_lock(); 76 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 77 if (inst->tr == tr) 78 found = 1; 79 } 80 rcu_read_unlock(); 81 82 return found; 83 } 84 85 /* 86 * osnoise_register_instance - register a new trace instance 87 * 88 * Register a trace_array *tr in the list of instances running 89 * osnoise/timerlat tracers. 90 */ 91 static int osnoise_register_instance(struct trace_array *tr) 92 { 93 struct osnoise_instance *inst; 94 95 /* 96 * register/unregister serialization is provided by trace's 97 * trace_types_lock. 98 */ 99 lockdep_assert_held(&trace_types_lock); 100 101 inst = kmalloc(sizeof(*inst), GFP_KERNEL); 102 if (!inst) 103 return -ENOMEM; 104 105 INIT_LIST_HEAD_RCU(&inst->list); 106 inst->tr = tr; 107 list_add_tail_rcu(&inst->list, &osnoise_instances); 108 109 return 0; 110 } 111 112 /* 113 * osnoise_unregister_instance - unregister a registered trace instance 114 * 115 * Remove the trace_array *tr from the list of instances running 116 * osnoise/timerlat tracers. 117 */ 118 static void osnoise_unregister_instance(struct trace_array *tr) 119 { 120 struct osnoise_instance *inst; 121 int found = 0; 122 123 /* 124 * register/unregister serialization is provided by trace's 125 * trace_types_lock. 126 */ 127 lockdep_assert_held(&trace_types_lock); 128 129 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 130 if (inst->tr == tr) { 131 list_del_rcu(&inst->list); 132 found = 1; 133 break; 134 } 135 } 136 137 if (!found) 138 return; 139 140 synchronize_rcu(); 141 kfree(inst); 142 } 143 144 /* 145 * NMI runtime info. 146 */ 147 struct osn_nmi { 148 u64 count; 149 u64 delta_start; 150 }; 151 152 /* 153 * IRQ runtime info. 154 */ 155 struct osn_irq { 156 u64 count; 157 u64 arrival_time; 158 u64 delta_start; 159 }; 160 161 #define IRQ_CONTEXT 0 162 #define THREAD_CONTEXT 1 163 /* 164 * sofirq runtime info. 165 */ 166 struct osn_softirq { 167 u64 count; 168 u64 arrival_time; 169 u64 delta_start; 170 }; 171 172 /* 173 * thread runtime info. 174 */ 175 struct osn_thread { 176 u64 count; 177 u64 arrival_time; 178 u64 delta_start; 179 }; 180 181 /* 182 * Runtime information: this structure saves the runtime information used by 183 * one sampling thread. 184 */ 185 struct osnoise_variables { 186 struct task_struct *kthread; 187 bool sampling; 188 pid_t pid; 189 struct osn_nmi nmi; 190 struct osn_irq irq; 191 struct osn_softirq softirq; 192 struct osn_thread thread; 193 local_t int_counter; 194 }; 195 196 /* 197 * Per-cpu runtime information. 198 */ 199 DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var); 200 201 /* 202 * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU 203 */ 204 static inline struct osnoise_variables *this_cpu_osn_var(void) 205 { 206 return this_cpu_ptr(&per_cpu_osnoise_var); 207 } 208 209 #ifdef CONFIG_TIMERLAT_TRACER 210 /* 211 * Runtime information for the timer mode. 212 */ 213 struct timerlat_variables { 214 struct task_struct *kthread; 215 struct hrtimer timer; 216 u64 rel_period; 217 u64 abs_period; 218 bool tracing_thread; 219 u64 count; 220 }; 221 222 DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var); 223 224 /* 225 * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU 226 */ 227 static inline struct timerlat_variables *this_cpu_tmr_var(void) 228 { 229 return this_cpu_ptr(&per_cpu_timerlat_var); 230 } 231 232 /* 233 * tlat_var_reset - Reset the values of the given timerlat_variables 234 */ 235 static inline void tlat_var_reset(void) 236 { 237 struct timerlat_variables *tlat_var; 238 int cpu; 239 /* 240 * So far, all the values are initialized as 0, so 241 * zeroing the structure is perfect. 242 */ 243 for_each_cpu(cpu, cpu_online_mask) { 244 tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu); 245 memset(tlat_var, 0, sizeof(*tlat_var)); 246 } 247 } 248 #else /* CONFIG_TIMERLAT_TRACER */ 249 #define tlat_var_reset() do {} while (0) 250 #endif /* CONFIG_TIMERLAT_TRACER */ 251 252 /* 253 * osn_var_reset - Reset the values of the given osnoise_variables 254 */ 255 static inline void osn_var_reset(void) 256 { 257 struct osnoise_variables *osn_var; 258 int cpu; 259 260 /* 261 * So far, all the values are initialized as 0, so 262 * zeroing the structure is perfect. 263 */ 264 for_each_cpu(cpu, cpu_online_mask) { 265 osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu); 266 memset(osn_var, 0, sizeof(*osn_var)); 267 } 268 } 269 270 /* 271 * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables 272 */ 273 static inline void osn_var_reset_all(void) 274 { 275 osn_var_reset(); 276 tlat_var_reset(); 277 } 278 279 /* 280 * Tells NMIs to call back to the osnoise tracer to record timestamps. 281 */ 282 bool trace_osnoise_callback_enabled; 283 284 /* 285 * osnoise sample structure definition. Used to store the statistics of a 286 * sample run. 287 */ 288 struct osnoise_sample { 289 u64 runtime; /* runtime */ 290 u64 noise; /* noise */ 291 u64 max_sample; /* max single noise sample */ 292 int hw_count; /* # HW (incl. hypervisor) interference */ 293 int nmi_count; /* # NMIs during this sample */ 294 int irq_count; /* # IRQs during this sample */ 295 int softirq_count; /* # softirqs during this sample */ 296 int thread_count; /* # threads during this sample */ 297 }; 298 299 #ifdef CONFIG_TIMERLAT_TRACER 300 /* 301 * timerlat sample structure definition. Used to store the statistics of 302 * a sample run. 303 */ 304 struct timerlat_sample { 305 u64 timer_latency; /* timer_latency */ 306 unsigned int seqnum; /* unique sequence */ 307 int context; /* timer context */ 308 }; 309 #endif 310 311 /* 312 * Protect the interface. 313 */ 314 struct mutex interface_lock; 315 316 /* 317 * Tracer data. 318 */ 319 static struct osnoise_data { 320 u64 sample_period; /* total sampling period */ 321 u64 sample_runtime; /* active sampling portion of period */ 322 u64 stop_tracing; /* stop trace in the internal operation (loop/irq) */ 323 u64 stop_tracing_total; /* stop trace in the final operation (report/thread) */ 324 #ifdef CONFIG_TIMERLAT_TRACER 325 u64 timerlat_period; /* timerlat period */ 326 u64 print_stack; /* print IRQ stack if total > */ 327 int timerlat_tracer; /* timerlat tracer */ 328 #endif 329 bool tainted; /* infor users and developers about a problem */ 330 } osnoise_data = { 331 .sample_period = DEFAULT_SAMPLE_PERIOD, 332 .sample_runtime = DEFAULT_SAMPLE_RUNTIME, 333 .stop_tracing = 0, 334 .stop_tracing_total = 0, 335 #ifdef CONFIG_TIMERLAT_TRACER 336 .print_stack = 0, 337 .timerlat_period = DEFAULT_TIMERLAT_PERIOD, 338 .timerlat_tracer = 0, 339 #endif 340 }; 341 342 #ifdef CONFIG_TIMERLAT_TRACER 343 static inline bool timerlat_enabled(void) 344 { 345 return osnoise_data.timerlat_tracer; 346 } 347 348 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var) 349 { 350 struct timerlat_variables *tlat_var = this_cpu_tmr_var(); 351 /* 352 * If the timerlat is enabled, but the irq handler did 353 * not run yet enabling timerlat_tracer, do not trace. 354 */ 355 if (!tlat_var->tracing_thread) { 356 osn_var->softirq.arrival_time = 0; 357 osn_var->softirq.delta_start = 0; 358 return 0; 359 } 360 return 1; 361 } 362 363 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var) 364 { 365 struct timerlat_variables *tlat_var = this_cpu_tmr_var(); 366 /* 367 * If the timerlat is enabled, but the irq handler did 368 * not run yet enabling timerlat_tracer, do not trace. 369 */ 370 if (!tlat_var->tracing_thread) { 371 osn_var->thread.delta_start = 0; 372 osn_var->thread.arrival_time = 0; 373 return 0; 374 } 375 return 1; 376 } 377 #else /* CONFIG_TIMERLAT_TRACER */ 378 static inline bool timerlat_enabled(void) 379 { 380 return false; 381 } 382 383 static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var) 384 { 385 return 1; 386 } 387 static inline int timerlat_thread_exit(struct osnoise_variables *osn_var) 388 { 389 return 1; 390 } 391 #endif 392 393 #ifdef CONFIG_PREEMPT_RT 394 /* 395 * Print the osnoise header info. 396 */ 397 static void print_osnoise_headers(struct seq_file *s) 398 { 399 if (osnoise_data.tainted) 400 seq_puts(s, "# osnoise is tainted!\n"); 401 402 seq_puts(s, "# _-------=> irqs-off\n"); 403 seq_puts(s, "# / _------=> need-resched\n"); 404 seq_puts(s, "# | / _-----=> need-resched-lazy\n"); 405 seq_puts(s, "# || / _----=> hardirq/softirq\n"); 406 seq_puts(s, "# ||| / _---=> preempt-depth\n"); 407 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n"); 408 seq_puts(s, "# ||||| / _-=> migrate-disable\n"); 409 410 seq_puts(s, "# |||||| / "); 411 seq_puts(s, " MAX\n"); 412 413 seq_puts(s, "# ||||| / "); 414 seq_puts(s, " SINGLE Interference counters:\n"); 415 416 seq_puts(s, "# ||||||| RUNTIME "); 417 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n"); 418 419 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP IN US "); 420 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n"); 421 422 seq_puts(s, "# | | | ||||||| | | "); 423 seq_puts(s, " | | | | | | | |\n"); 424 } 425 #else /* CONFIG_PREEMPT_RT */ 426 static void print_osnoise_headers(struct seq_file *s) 427 { 428 if (osnoise_data.tainted) 429 seq_puts(s, "# osnoise is tainted!\n"); 430 431 seq_puts(s, "# _-----=> irqs-off\n"); 432 seq_puts(s, "# / _----=> need-resched\n"); 433 seq_puts(s, "# | / _---=> hardirq/softirq\n"); 434 seq_puts(s, "# || / _--=> preempt-depth\n"); 435 seq_puts(s, "# ||| / _-=> migrate-disable "); 436 seq_puts(s, " MAX\n"); 437 seq_puts(s, "# |||| / delay "); 438 seq_puts(s, " SINGLE Interference counters:\n"); 439 440 seq_puts(s, "# ||||| RUNTIME "); 441 seq_puts(s, " NOISE %% OF CPU NOISE +-----------------------------+\n"); 442 443 seq_puts(s, "# TASK-PID CPU# ||||| TIMESTAMP IN US "); 444 seq_puts(s, " IN US AVAILABLE IN US HW NMI IRQ SIRQ THREAD\n"); 445 446 seq_puts(s, "# | | | ||||| | | "); 447 seq_puts(s, " | | | | | | | |\n"); 448 } 449 #endif /* CONFIG_PREEMPT_RT */ 450 451 /* 452 * osnoise_taint - report an osnoise error. 453 */ 454 #define osnoise_taint(msg) ({ \ 455 struct osnoise_instance *inst; \ 456 struct trace_buffer *buffer; \ 457 \ 458 rcu_read_lock(); \ 459 list_for_each_entry_rcu(inst, &osnoise_instances, list) { \ 460 buffer = inst->tr->array_buffer.buffer; \ 461 trace_array_printk_buf(buffer, _THIS_IP_, msg); \ 462 } \ 463 rcu_read_unlock(); \ 464 osnoise_data.tainted = true; \ 465 }) 466 467 /* 468 * Record an osnoise_sample into the tracer buffer. 469 */ 470 static void 471 __trace_osnoise_sample(struct osnoise_sample *sample, struct trace_buffer *buffer) 472 { 473 struct trace_event_call *call = &event_osnoise; 474 struct ring_buffer_event *event; 475 struct osnoise_entry *entry; 476 477 event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry), 478 tracing_gen_ctx()); 479 if (!event) 480 return; 481 entry = ring_buffer_event_data(event); 482 entry->runtime = sample->runtime; 483 entry->noise = sample->noise; 484 entry->max_sample = sample->max_sample; 485 entry->hw_count = sample->hw_count; 486 entry->nmi_count = sample->nmi_count; 487 entry->irq_count = sample->irq_count; 488 entry->softirq_count = sample->softirq_count; 489 entry->thread_count = sample->thread_count; 490 491 if (!call_filter_check_discard(call, entry, buffer, event)) 492 trace_buffer_unlock_commit_nostack(buffer, event); 493 } 494 495 /* 496 * Record an osnoise_sample on all osnoise instances. 497 */ 498 static void trace_osnoise_sample(struct osnoise_sample *sample) 499 { 500 struct osnoise_instance *inst; 501 struct trace_buffer *buffer; 502 503 rcu_read_lock(); 504 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 505 buffer = inst->tr->array_buffer.buffer; 506 __trace_osnoise_sample(sample, buffer); 507 } 508 rcu_read_unlock(); 509 } 510 511 #ifdef CONFIG_TIMERLAT_TRACER 512 /* 513 * Print the timerlat header info. 514 */ 515 #ifdef CONFIG_PREEMPT_RT 516 static void print_timerlat_headers(struct seq_file *s) 517 { 518 seq_puts(s, "# _-------=> irqs-off\n"); 519 seq_puts(s, "# / _------=> need-resched\n"); 520 seq_puts(s, "# | / _-----=> need-resched-lazy\n"); 521 seq_puts(s, "# || / _----=> hardirq/softirq\n"); 522 seq_puts(s, "# ||| / _---=> preempt-depth\n"); 523 seq_puts(s, "# |||| / _--=> preempt-lazy-depth\n"); 524 seq_puts(s, "# ||||| / _-=> migrate-disable\n"); 525 seq_puts(s, "# |||||| /\n"); 526 seq_puts(s, "# ||||||| ACTIVATION\n"); 527 seq_puts(s, "# TASK-PID CPU# ||||||| TIMESTAMP ID "); 528 seq_puts(s, " CONTEXT LATENCY\n"); 529 seq_puts(s, "# | | | ||||||| | | "); 530 seq_puts(s, " | |\n"); 531 } 532 #else /* CONFIG_PREEMPT_RT */ 533 static void print_timerlat_headers(struct seq_file *s) 534 { 535 seq_puts(s, "# _-----=> irqs-off\n"); 536 seq_puts(s, "# / _----=> need-resched\n"); 537 seq_puts(s, "# | / _---=> hardirq/softirq\n"); 538 seq_puts(s, "# || / _--=> preempt-depth\n"); 539 seq_puts(s, "# ||| / _-=> migrate-disable\n"); 540 seq_puts(s, "# |||| / delay\n"); 541 seq_puts(s, "# ||||| ACTIVATION\n"); 542 seq_puts(s, "# TASK-PID CPU# ||||| TIMESTAMP ID "); 543 seq_puts(s, " CONTEXT LATENCY\n"); 544 seq_puts(s, "# | | | ||||| | | "); 545 seq_puts(s, " | |\n"); 546 } 547 #endif /* CONFIG_PREEMPT_RT */ 548 549 static void 550 __trace_timerlat_sample(struct timerlat_sample *sample, struct trace_buffer *buffer) 551 { 552 struct trace_event_call *call = &event_osnoise; 553 struct ring_buffer_event *event; 554 struct timerlat_entry *entry; 555 556 event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry), 557 tracing_gen_ctx()); 558 if (!event) 559 return; 560 entry = ring_buffer_event_data(event); 561 entry->seqnum = sample->seqnum; 562 entry->context = sample->context; 563 entry->timer_latency = sample->timer_latency; 564 565 if (!call_filter_check_discard(call, entry, buffer, event)) 566 trace_buffer_unlock_commit_nostack(buffer, event); 567 } 568 569 /* 570 * Record an timerlat_sample into the tracer buffer. 571 */ 572 static void trace_timerlat_sample(struct timerlat_sample *sample) 573 { 574 struct osnoise_instance *inst; 575 struct trace_buffer *buffer; 576 577 rcu_read_lock(); 578 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 579 buffer = inst->tr->array_buffer.buffer; 580 __trace_timerlat_sample(sample, buffer); 581 } 582 rcu_read_unlock(); 583 } 584 585 #ifdef CONFIG_STACKTRACE 586 587 #define MAX_CALLS 256 588 589 /* 590 * Stack trace will take place only at IRQ level, so, no need 591 * to control nesting here. 592 */ 593 struct trace_stack { 594 int stack_size; 595 int nr_entries; 596 unsigned long calls[MAX_CALLS]; 597 }; 598 599 static DEFINE_PER_CPU(struct trace_stack, trace_stack); 600 601 /* 602 * timerlat_save_stack - save a stack trace without printing 603 * 604 * Save the current stack trace without printing. The 605 * stack will be printed later, after the end of the measurement. 606 */ 607 static void timerlat_save_stack(int skip) 608 { 609 unsigned int size, nr_entries; 610 struct trace_stack *fstack; 611 612 fstack = this_cpu_ptr(&trace_stack); 613 614 size = ARRAY_SIZE(fstack->calls); 615 616 nr_entries = stack_trace_save(fstack->calls, size, skip); 617 618 fstack->stack_size = nr_entries * sizeof(unsigned long); 619 fstack->nr_entries = nr_entries; 620 621 return; 622 623 } 624 625 static void 626 __timerlat_dump_stack(struct trace_buffer *buffer, struct trace_stack *fstack, unsigned int size) 627 { 628 struct trace_event_call *call = &event_osnoise; 629 struct ring_buffer_event *event; 630 struct stack_entry *entry; 631 632 event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size, 633 tracing_gen_ctx()); 634 if (!event) 635 return; 636 637 entry = ring_buffer_event_data(event); 638 639 memcpy(&entry->caller, fstack->calls, size); 640 entry->size = fstack->nr_entries; 641 642 if (!call_filter_check_discard(call, entry, buffer, event)) 643 trace_buffer_unlock_commit_nostack(buffer, event); 644 } 645 646 /* 647 * timerlat_dump_stack - dump a stack trace previously saved 648 */ 649 static void timerlat_dump_stack(u64 latency) 650 { 651 struct osnoise_instance *inst; 652 struct trace_buffer *buffer; 653 struct trace_stack *fstack; 654 unsigned int size; 655 656 /* 657 * trace only if latency > print_stack config, if enabled. 658 */ 659 if (!osnoise_data.print_stack || osnoise_data.print_stack > latency) 660 return; 661 662 preempt_disable_notrace(); 663 fstack = this_cpu_ptr(&trace_stack); 664 size = fstack->stack_size; 665 666 rcu_read_lock(); 667 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 668 buffer = inst->tr->array_buffer.buffer; 669 __timerlat_dump_stack(buffer, fstack, size); 670 671 } 672 rcu_read_unlock(); 673 preempt_enable_notrace(); 674 } 675 #else /* CONFIG_STACKTRACE */ 676 #define timerlat_dump_stack(u64 latency) do {} while (0) 677 #define timerlat_save_stack(a) do {} while (0) 678 #endif /* CONFIG_STACKTRACE */ 679 #endif /* CONFIG_TIMERLAT_TRACER */ 680 681 /* 682 * Macros to encapsulate the time capturing infrastructure. 683 */ 684 #define time_get() trace_clock_local() 685 #define time_to_us(x) div_u64(x, 1000) 686 #define time_sub(a, b) ((a) - (b)) 687 688 /* 689 * cond_move_irq_delta_start - Forward the delta_start of a running IRQ 690 * 691 * If an IRQ is preempted by an NMI, its delta_start is pushed forward 692 * to discount the NMI interference. 693 * 694 * See get_int_safe_duration(). 695 */ 696 static inline void 697 cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration) 698 { 699 if (osn_var->irq.delta_start) 700 osn_var->irq.delta_start += duration; 701 } 702 703 #ifndef CONFIG_PREEMPT_RT 704 /* 705 * cond_move_softirq_delta_start - Forward the delta_start of a running softirq. 706 * 707 * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed 708 * forward to discount the interference. 709 * 710 * See get_int_safe_duration(). 711 */ 712 static inline void 713 cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration) 714 { 715 if (osn_var->softirq.delta_start) 716 osn_var->softirq.delta_start += duration; 717 } 718 #else /* CONFIG_PREEMPT_RT */ 719 #define cond_move_softirq_delta_start(osn_var, duration) do {} while (0) 720 #endif 721 722 /* 723 * cond_move_thread_delta_start - Forward the delta_start of a running thread 724 * 725 * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start 726 * is pushed forward to discount the interference. 727 * 728 * See get_int_safe_duration(). 729 */ 730 static inline void 731 cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration) 732 { 733 if (osn_var->thread.delta_start) 734 osn_var->thread.delta_start += duration; 735 } 736 737 /* 738 * get_int_safe_duration - Get the duration of a window 739 * 740 * The irq, softirq and thread varaibles need to have its duration without 741 * the interference from higher priority interrupts. Instead of keeping a 742 * variable to discount the interrupt interference from these variables, the 743 * starting time of these variables are pushed forward with the interrupt's 744 * duration. In this way, a single variable is used to: 745 * 746 * - Know if a given window is being measured. 747 * - Account its duration. 748 * - Discount the interference. 749 * 750 * To avoid getting inconsistent values, e.g.,: 751 * 752 * now = time_get() 753 * ---> interrupt! 754 * delta_start -= int duration; 755 * <--- 756 * duration = now - delta_start; 757 * 758 * result: negative duration if the variable duration before the 759 * interrupt was smaller than the interrupt execution. 760 * 761 * A counter of interrupts is used. If the counter increased, try 762 * to capture an interference safe duration. 763 */ 764 static inline s64 765 get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start) 766 { 767 u64 int_counter, now; 768 s64 duration; 769 770 do { 771 int_counter = local_read(&osn_var->int_counter); 772 /* synchronize with interrupts */ 773 barrier(); 774 775 now = time_get(); 776 duration = (now - *delta_start); 777 778 /* synchronize with interrupts */ 779 barrier(); 780 } while (int_counter != local_read(&osn_var->int_counter)); 781 782 /* 783 * This is an evidence of race conditions that cause 784 * a value to be "discounted" too much. 785 */ 786 if (duration < 0) 787 osnoise_taint("Negative duration!\n"); 788 789 *delta_start = 0; 790 791 return duration; 792 } 793 794 /* 795 * 796 * set_int_safe_time - Save the current time on *time, aware of interference 797 * 798 * Get the time, taking into consideration a possible interference from 799 * higher priority interrupts. 800 * 801 * See get_int_safe_duration() for an explanation. 802 */ 803 static u64 804 set_int_safe_time(struct osnoise_variables *osn_var, u64 *time) 805 { 806 u64 int_counter; 807 808 do { 809 int_counter = local_read(&osn_var->int_counter); 810 /* synchronize with interrupts */ 811 barrier(); 812 813 *time = time_get(); 814 815 /* synchronize with interrupts */ 816 barrier(); 817 } while (int_counter != local_read(&osn_var->int_counter)); 818 819 return int_counter; 820 } 821 822 #ifdef CONFIG_TIMERLAT_TRACER 823 /* 824 * copy_int_safe_time - Copy *src into *desc aware of interference 825 */ 826 static u64 827 copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src) 828 { 829 u64 int_counter; 830 831 do { 832 int_counter = local_read(&osn_var->int_counter); 833 /* synchronize with interrupts */ 834 barrier(); 835 836 *dst = *src; 837 838 /* synchronize with interrupts */ 839 barrier(); 840 } while (int_counter != local_read(&osn_var->int_counter)); 841 842 return int_counter; 843 } 844 #endif /* CONFIG_TIMERLAT_TRACER */ 845 846 /* 847 * trace_osnoise_callback - NMI entry/exit callback 848 * 849 * This function is called at the entry and exit NMI code. The bool enter 850 * distinguishes between either case. This function is used to note a NMI 851 * occurrence, compute the noise caused by the NMI, and to remove the noise 852 * it is potentially causing on other interference variables. 853 */ 854 void trace_osnoise_callback(bool enter) 855 { 856 struct osnoise_variables *osn_var = this_cpu_osn_var(); 857 u64 duration; 858 859 if (!osn_var->sampling) 860 return; 861 862 /* 863 * Currently trace_clock_local() calls sched_clock() and the 864 * generic version is not NMI safe. 865 */ 866 if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) { 867 if (enter) { 868 osn_var->nmi.delta_start = time_get(); 869 local_inc(&osn_var->int_counter); 870 } else { 871 duration = time_get() - osn_var->nmi.delta_start; 872 873 trace_nmi_noise(osn_var->nmi.delta_start, duration); 874 875 cond_move_irq_delta_start(osn_var, duration); 876 cond_move_softirq_delta_start(osn_var, duration); 877 cond_move_thread_delta_start(osn_var, duration); 878 } 879 } 880 881 if (enter) 882 osn_var->nmi.count++; 883 } 884 885 /* 886 * osnoise_trace_irq_entry - Note the starting of an IRQ 887 * 888 * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs, 889 * it is safe to use a single variable (ons_var->irq) to save the statistics. 890 * The arrival_time is used to report... the arrival time. The delta_start 891 * is used to compute the duration at the IRQ exit handler. See 892 * cond_move_irq_delta_start(). 893 */ 894 void osnoise_trace_irq_entry(int id) 895 { 896 struct osnoise_variables *osn_var = this_cpu_osn_var(); 897 898 if (!osn_var->sampling) 899 return; 900 /* 901 * This value will be used in the report, but not to compute 902 * the execution time, so it is safe to get it unsafe. 903 */ 904 osn_var->irq.arrival_time = time_get(); 905 set_int_safe_time(osn_var, &osn_var->irq.delta_start); 906 osn_var->irq.count++; 907 908 local_inc(&osn_var->int_counter); 909 } 910 911 /* 912 * osnoise_irq_exit - Note the end of an IRQ, sava data and trace 913 * 914 * Computes the duration of the IRQ noise, and trace it. Also discounts the 915 * interference from other sources of noise could be currently being accounted. 916 */ 917 void osnoise_trace_irq_exit(int id, const char *desc) 918 { 919 struct osnoise_variables *osn_var = this_cpu_osn_var(); 920 int duration; 921 922 if (!osn_var->sampling) 923 return; 924 925 duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start); 926 trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration); 927 osn_var->irq.arrival_time = 0; 928 cond_move_softirq_delta_start(osn_var, duration); 929 cond_move_thread_delta_start(osn_var, duration); 930 } 931 932 /* 933 * trace_irqentry_callback - Callback to the irq:irq_entry traceevent 934 * 935 * Used to note the starting of an IRQ occurece. 936 */ 937 static void trace_irqentry_callback(void *data, int irq, 938 struct irqaction *action) 939 { 940 osnoise_trace_irq_entry(irq); 941 } 942 943 /* 944 * trace_irqexit_callback - Callback to the irq:irq_exit traceevent 945 * 946 * Used to note the end of an IRQ occurece. 947 */ 948 static void trace_irqexit_callback(void *data, int irq, 949 struct irqaction *action, int ret) 950 { 951 osnoise_trace_irq_exit(irq, action->name); 952 } 953 954 /* 955 * arch specific register function. 956 */ 957 int __weak osnoise_arch_register(void) 958 { 959 return 0; 960 } 961 962 /* 963 * arch specific unregister function. 964 */ 965 void __weak osnoise_arch_unregister(void) 966 { 967 return; 968 } 969 970 /* 971 * hook_irq_events - Hook IRQ handling events 972 * 973 * This function hooks the IRQ related callbacks to the respective trace 974 * events. 975 */ 976 static int hook_irq_events(void) 977 { 978 int ret; 979 980 ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL); 981 if (ret) 982 goto out_err; 983 984 ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL); 985 if (ret) 986 goto out_unregister_entry; 987 988 ret = osnoise_arch_register(); 989 if (ret) 990 goto out_irq_exit; 991 992 return 0; 993 994 out_irq_exit: 995 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL); 996 out_unregister_entry: 997 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL); 998 out_err: 999 return -EINVAL; 1000 } 1001 1002 /* 1003 * unhook_irq_events - Unhook IRQ handling events 1004 * 1005 * This function unhooks the IRQ related callbacks to the respective trace 1006 * events. 1007 */ 1008 static void unhook_irq_events(void) 1009 { 1010 osnoise_arch_unregister(); 1011 unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL); 1012 unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL); 1013 } 1014 1015 #ifndef CONFIG_PREEMPT_RT 1016 /* 1017 * trace_softirq_entry_callback - Note the starting of a softirq 1018 * 1019 * Save the starting time of a softirq. As softirqs are non-preemptive to 1020 * other softirqs, it is safe to use a single variable (ons_var->softirq) 1021 * to save the statistics. The arrival_time is used to report... the 1022 * arrival time. The delta_start is used to compute the duration at the 1023 * softirq exit handler. See cond_move_softirq_delta_start(). 1024 */ 1025 static void trace_softirq_entry_callback(void *data, unsigned int vec_nr) 1026 { 1027 struct osnoise_variables *osn_var = this_cpu_osn_var(); 1028 1029 if (!osn_var->sampling) 1030 return; 1031 /* 1032 * This value will be used in the report, but not to compute 1033 * the execution time, so it is safe to get it unsafe. 1034 */ 1035 osn_var->softirq.arrival_time = time_get(); 1036 set_int_safe_time(osn_var, &osn_var->softirq.delta_start); 1037 osn_var->softirq.count++; 1038 1039 local_inc(&osn_var->int_counter); 1040 } 1041 1042 /* 1043 * trace_softirq_exit_callback - Note the end of an softirq 1044 * 1045 * Computes the duration of the softirq noise, and trace it. Also discounts the 1046 * interference from other sources of noise could be currently being accounted. 1047 */ 1048 static void trace_softirq_exit_callback(void *data, unsigned int vec_nr) 1049 { 1050 struct osnoise_variables *osn_var = this_cpu_osn_var(); 1051 int duration; 1052 1053 if (!osn_var->sampling) 1054 return; 1055 1056 if (unlikely(timerlat_enabled())) 1057 if (!timerlat_softirq_exit(osn_var)) 1058 return; 1059 1060 duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start); 1061 trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration); 1062 cond_move_thread_delta_start(osn_var, duration); 1063 osn_var->softirq.arrival_time = 0; 1064 } 1065 1066 /* 1067 * hook_softirq_events - Hook softirq handling events 1068 * 1069 * This function hooks the softirq related callbacks to the respective trace 1070 * events. 1071 */ 1072 static int hook_softirq_events(void) 1073 { 1074 int ret; 1075 1076 ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL); 1077 if (ret) 1078 goto out_err; 1079 1080 ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL); 1081 if (ret) 1082 goto out_unreg_entry; 1083 1084 return 0; 1085 1086 out_unreg_entry: 1087 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL); 1088 out_err: 1089 return -EINVAL; 1090 } 1091 1092 /* 1093 * unhook_softirq_events - Unhook softirq handling events 1094 * 1095 * This function hooks the softirq related callbacks to the respective trace 1096 * events. 1097 */ 1098 static void unhook_softirq_events(void) 1099 { 1100 unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL); 1101 unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL); 1102 } 1103 #else /* CONFIG_PREEMPT_RT */ 1104 /* 1105 * softirq are threads on the PREEMPT_RT mode. 1106 */ 1107 static int hook_softirq_events(void) 1108 { 1109 return 0; 1110 } 1111 static void unhook_softirq_events(void) 1112 { 1113 } 1114 #endif 1115 1116 /* 1117 * thread_entry - Record the starting of a thread noise window 1118 * 1119 * It saves the context switch time for a noisy thread, and increments 1120 * the interference counters. 1121 */ 1122 static void 1123 thread_entry(struct osnoise_variables *osn_var, struct task_struct *t) 1124 { 1125 if (!osn_var->sampling) 1126 return; 1127 /* 1128 * The arrival time will be used in the report, but not to compute 1129 * the execution time, so it is safe to get it unsafe. 1130 */ 1131 osn_var->thread.arrival_time = time_get(); 1132 1133 set_int_safe_time(osn_var, &osn_var->thread.delta_start); 1134 1135 osn_var->thread.count++; 1136 local_inc(&osn_var->int_counter); 1137 } 1138 1139 /* 1140 * thread_exit - Report the end of a thread noise window 1141 * 1142 * It computes the total noise from a thread, tracing if needed. 1143 */ 1144 static void 1145 thread_exit(struct osnoise_variables *osn_var, struct task_struct *t) 1146 { 1147 int duration; 1148 1149 if (!osn_var->sampling) 1150 return; 1151 1152 if (unlikely(timerlat_enabled())) 1153 if (!timerlat_thread_exit(osn_var)) 1154 return; 1155 1156 duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start); 1157 1158 trace_thread_noise(t, osn_var->thread.arrival_time, duration); 1159 1160 osn_var->thread.arrival_time = 0; 1161 } 1162 1163 /* 1164 * trace_sched_switch - sched:sched_switch trace event handler 1165 * 1166 * This function is hooked to the sched:sched_switch trace event, and it is 1167 * used to record the beginning and to report the end of a thread noise window. 1168 */ 1169 static void 1170 trace_sched_switch_callback(void *data, bool preempt, struct task_struct *p, 1171 struct task_struct *n) 1172 { 1173 struct osnoise_variables *osn_var = this_cpu_osn_var(); 1174 1175 if (p->pid != osn_var->pid) 1176 thread_exit(osn_var, p); 1177 1178 if (n->pid != osn_var->pid) 1179 thread_entry(osn_var, n); 1180 } 1181 1182 /* 1183 * hook_thread_events - Hook the insturmentation for thread noise 1184 * 1185 * Hook the osnoise tracer callbacks to handle the noise from other 1186 * threads on the necessary kernel events. 1187 */ 1188 static int hook_thread_events(void) 1189 { 1190 int ret; 1191 1192 ret = register_trace_sched_switch(trace_sched_switch_callback, NULL); 1193 if (ret) 1194 return -EINVAL; 1195 1196 return 0; 1197 } 1198 1199 /* 1200 * unhook_thread_events - *nhook the insturmentation for thread noise 1201 * 1202 * Unook the osnoise tracer callbacks to handle the noise from other 1203 * threads on the necessary kernel events. 1204 */ 1205 static void unhook_thread_events(void) 1206 { 1207 unregister_trace_sched_switch(trace_sched_switch_callback, NULL); 1208 } 1209 1210 /* 1211 * save_osn_sample_stats - Save the osnoise_sample statistics 1212 * 1213 * Save the osnoise_sample statistics before the sampling phase. These 1214 * values will be used later to compute the diff betwneen the statistics 1215 * before and after the osnoise sampling. 1216 */ 1217 static void 1218 save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s) 1219 { 1220 s->nmi_count = osn_var->nmi.count; 1221 s->irq_count = osn_var->irq.count; 1222 s->softirq_count = osn_var->softirq.count; 1223 s->thread_count = osn_var->thread.count; 1224 } 1225 1226 /* 1227 * diff_osn_sample_stats - Compute the osnoise_sample statistics 1228 * 1229 * After a sample period, compute the difference on the osnoise_sample 1230 * statistics. The struct osnoise_sample *s contains the statistics saved via 1231 * save_osn_sample_stats() before the osnoise sampling. 1232 */ 1233 static void 1234 diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s) 1235 { 1236 s->nmi_count = osn_var->nmi.count - s->nmi_count; 1237 s->irq_count = osn_var->irq.count - s->irq_count; 1238 s->softirq_count = osn_var->softirq.count - s->softirq_count; 1239 s->thread_count = osn_var->thread.count - s->thread_count; 1240 } 1241 1242 /* 1243 * osnoise_stop_tracing - Stop tracing and the tracer. 1244 */ 1245 static __always_inline void osnoise_stop_tracing(void) 1246 { 1247 struct osnoise_instance *inst; 1248 struct trace_array *tr; 1249 1250 rcu_read_lock(); 1251 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 1252 tr = inst->tr; 1253 trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, 1254 "stop tracing hit on cpu %d\n", smp_processor_id()); 1255 1256 tracer_tracing_off(tr); 1257 } 1258 rcu_read_unlock(); 1259 } 1260 1261 /* 1262 * notify_new_max_latency - Notify a new max latency via fsnotify interface. 1263 */ 1264 static void notify_new_max_latency(u64 latency) 1265 { 1266 struct osnoise_instance *inst; 1267 struct trace_array *tr; 1268 1269 rcu_read_lock(); 1270 list_for_each_entry_rcu(inst, &osnoise_instances, list) { 1271 tr = inst->tr; 1272 if (tr->max_latency < latency) { 1273 tr->max_latency = latency; 1274 latency_fsnotify(tr); 1275 } 1276 } 1277 rcu_read_unlock(); 1278 } 1279 1280 /* 1281 * run_osnoise - Sample the time and look for osnoise 1282 * 1283 * Used to capture the time, looking for potential osnoise latency repeatedly. 1284 * Different from hwlat_detector, it is called with preemption and interrupts 1285 * enabled. This allows irqs, softirqs and threads to run, interfering on the 1286 * osnoise sampling thread, as they would do with a regular thread. 1287 */ 1288 static int run_osnoise(void) 1289 { 1290 struct osnoise_variables *osn_var = this_cpu_osn_var(); 1291 u64 start, sample, last_sample; 1292 u64 last_int_count, int_count; 1293 s64 noise = 0, max_noise = 0; 1294 s64 total, last_total = 0; 1295 struct osnoise_sample s; 1296 unsigned int threshold; 1297 u64 runtime, stop_in; 1298 u64 sum_noise = 0; 1299 int hw_count = 0; 1300 int ret = -1; 1301 1302 /* 1303 * Considers the current thread as the workload. 1304 */ 1305 osn_var->pid = current->pid; 1306 1307 /* 1308 * Save the current stats for the diff 1309 */ 1310 save_osn_sample_stats(osn_var, &s); 1311 1312 /* 1313 * if threshold is 0, use the default value of 5 us. 1314 */ 1315 threshold = tracing_thresh ? : 5000; 1316 1317 /* 1318 * Make sure NMIs see sampling first 1319 */ 1320 osn_var->sampling = true; 1321 barrier(); 1322 1323 /* 1324 * Transform the *_us config to nanoseconds to avoid the 1325 * division on the main loop. 1326 */ 1327 runtime = osnoise_data.sample_runtime * NSEC_PER_USEC; 1328 stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC; 1329 1330 /* 1331 * Start timestemp 1332 */ 1333 start = time_get(); 1334 1335 /* 1336 * "previous" loop. 1337 */ 1338 last_int_count = set_int_safe_time(osn_var, &last_sample); 1339 1340 do { 1341 /* 1342 * Get sample! 1343 */ 1344 int_count = set_int_safe_time(osn_var, &sample); 1345 1346 noise = time_sub(sample, last_sample); 1347 1348 /* 1349 * This shouldn't happen. 1350 */ 1351 if (noise < 0) { 1352 osnoise_taint("negative noise!"); 1353 goto out; 1354 } 1355 1356 /* 1357 * Sample runtime. 1358 */ 1359 total = time_sub(sample, start); 1360 1361 /* 1362 * Check for possible overflows. 1363 */ 1364 if (total < last_total) { 1365 osnoise_taint("total overflow!"); 1366 break; 1367 } 1368 1369 last_total = total; 1370 1371 if (noise >= threshold) { 1372 int interference = int_count - last_int_count; 1373 1374 if (noise > max_noise) 1375 max_noise = noise; 1376 1377 if (!interference) 1378 hw_count++; 1379 1380 sum_noise += noise; 1381 1382 trace_sample_threshold(last_sample, noise, interference); 1383 1384 if (osnoise_data.stop_tracing) 1385 if (noise > stop_in) 1386 osnoise_stop_tracing(); 1387 } 1388 1389 /* 1390 * For the non-preemptive kernel config: let threads runs, if 1391 * they so wish. 1392 */ 1393 cond_resched(); 1394 1395 last_sample = sample; 1396 last_int_count = int_count; 1397 1398 } while (total < runtime && !kthread_should_stop()); 1399 1400 /* 1401 * Finish the above in the view for interrupts. 1402 */ 1403 barrier(); 1404 1405 osn_var->sampling = false; 1406 1407 /* 1408 * Make sure sampling data is no longer updated. 1409 */ 1410 barrier(); 1411 1412 /* 1413 * Save noise info. 1414 */ 1415 s.noise = time_to_us(sum_noise); 1416 s.runtime = time_to_us(total); 1417 s.max_sample = time_to_us(max_noise); 1418 s.hw_count = hw_count; 1419 1420 /* Save interference stats info */ 1421 diff_osn_sample_stats(osn_var, &s); 1422 1423 trace_osnoise_sample(&s); 1424 1425 notify_new_max_latency(max_noise); 1426 1427 if (osnoise_data.stop_tracing_total) 1428 if (s.noise > osnoise_data.stop_tracing_total) 1429 osnoise_stop_tracing(); 1430 1431 return 0; 1432 out: 1433 return ret; 1434 } 1435 1436 static struct cpumask osnoise_cpumask; 1437 static struct cpumask save_cpumask; 1438 1439 /* 1440 * osnoise_main - The osnoise detection kernel thread 1441 * 1442 * Calls run_osnoise() function to measure the osnoise for the configured runtime, 1443 * every period. 1444 */ 1445 static int osnoise_main(void *data) 1446 { 1447 u64 interval; 1448 1449 while (!kthread_should_stop()) { 1450 1451 run_osnoise(); 1452 1453 mutex_lock(&interface_lock); 1454 interval = osnoise_data.sample_period - osnoise_data.sample_runtime; 1455 mutex_unlock(&interface_lock); 1456 1457 do_div(interval, USEC_PER_MSEC); 1458 1459 /* 1460 * differently from hwlat_detector, the osnoise tracer can run 1461 * without a pause because preemption is on. 1462 */ 1463 if (interval < 1) { 1464 /* Let synchronize_rcu_tasks() make progress */ 1465 cond_resched_tasks_rcu_qs(); 1466 continue; 1467 } 1468 1469 if (msleep_interruptible(interval)) 1470 break; 1471 } 1472 1473 return 0; 1474 } 1475 1476 #ifdef CONFIG_TIMERLAT_TRACER 1477 /* 1478 * timerlat_irq - hrtimer handler for timerlat. 1479 */ 1480 static enum hrtimer_restart timerlat_irq(struct hrtimer *timer) 1481 { 1482 struct osnoise_variables *osn_var = this_cpu_osn_var(); 1483 struct timerlat_variables *tlat; 1484 struct timerlat_sample s; 1485 u64 now; 1486 u64 diff; 1487 1488 /* 1489 * I am not sure if the timer was armed for this CPU. So, get 1490 * the timerlat struct from the timer itself, not from this 1491 * CPU. 1492 */ 1493 tlat = container_of(timer, struct timerlat_variables, timer); 1494 1495 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer)); 1496 1497 /* 1498 * Enable the osnoise: events for thread an softirq. 1499 */ 1500 tlat->tracing_thread = true; 1501 1502 osn_var->thread.arrival_time = time_get(); 1503 1504 /* 1505 * A hardirq is running: the timer IRQ. It is for sure preempting 1506 * a thread, and potentially preempting a softirq. 1507 * 1508 * At this point, it is not interesting to know the duration of the 1509 * preempted thread (and maybe softirq), but how much time they will 1510 * delay the beginning of the execution of the timer thread. 1511 * 1512 * To get the correct (net) delay added by the softirq, its delta_start 1513 * is set as the IRQ one. In this way, at the return of the IRQ, the delta 1514 * start of the sofitrq will be zeroed, accounting then only the time 1515 * after that. 1516 * 1517 * The thread follows the same principle. However, if a softirq is 1518 * running, the thread needs to receive the softirq delta_start. The 1519 * reason being is that the softirq will be the last to be unfolded, 1520 * resseting the thread delay to zero. 1521 * 1522 * The PREEMPT_RT is a special case, though. As softirqs run as threads 1523 * on RT, moving the thread is enough. 1524 */ 1525 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && osn_var->softirq.delta_start) { 1526 copy_int_safe_time(osn_var, &osn_var->thread.delta_start, 1527 &osn_var->softirq.delta_start); 1528 1529 copy_int_safe_time(osn_var, &osn_var->softirq.delta_start, 1530 &osn_var->irq.delta_start); 1531 } else { 1532 copy_int_safe_time(osn_var, &osn_var->thread.delta_start, 1533 &osn_var->irq.delta_start); 1534 } 1535 1536 /* 1537 * Compute the current time with the expected time. 1538 */ 1539 diff = now - tlat->abs_period; 1540 1541 tlat->count++; 1542 s.seqnum = tlat->count; 1543 s.timer_latency = diff; 1544 s.context = IRQ_CONTEXT; 1545 1546 trace_timerlat_sample(&s); 1547 1548 notify_new_max_latency(diff); 1549 1550 if (osnoise_data.stop_tracing) 1551 if (time_to_us(diff) >= osnoise_data.stop_tracing) 1552 osnoise_stop_tracing(); 1553 1554 wake_up_process(tlat->kthread); 1555 1556 if (osnoise_data.print_stack) 1557 timerlat_save_stack(0); 1558 1559 return HRTIMER_NORESTART; 1560 } 1561 1562 /* 1563 * wait_next_period - Wait for the next period for timerlat 1564 */ 1565 static int wait_next_period(struct timerlat_variables *tlat) 1566 { 1567 ktime_t next_abs_period, now; 1568 u64 rel_period = osnoise_data.timerlat_period * 1000; 1569 1570 now = hrtimer_cb_get_time(&tlat->timer); 1571 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period); 1572 1573 /* 1574 * Save the next abs_period. 1575 */ 1576 tlat->abs_period = (u64) ktime_to_ns(next_abs_period); 1577 1578 /* 1579 * If the new abs_period is in the past, skip the activation. 1580 */ 1581 while (ktime_compare(now, next_abs_period) > 0) { 1582 next_abs_period = ns_to_ktime(tlat->abs_period + rel_period); 1583 tlat->abs_period = (u64) ktime_to_ns(next_abs_period); 1584 } 1585 1586 set_current_state(TASK_INTERRUPTIBLE); 1587 1588 hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD); 1589 schedule(); 1590 return 1; 1591 } 1592 1593 /* 1594 * timerlat_main- Timerlat main 1595 */ 1596 static int timerlat_main(void *data) 1597 { 1598 struct osnoise_variables *osn_var = this_cpu_osn_var(); 1599 struct timerlat_variables *tlat = this_cpu_tmr_var(); 1600 struct timerlat_sample s; 1601 struct sched_param sp; 1602 u64 now, diff; 1603 1604 /* 1605 * Make the thread RT, that is how cyclictest is usually used. 1606 */ 1607 sp.sched_priority = DEFAULT_TIMERLAT_PRIO; 1608 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); 1609 1610 tlat->count = 0; 1611 tlat->tracing_thread = false; 1612 1613 hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD); 1614 tlat->timer.function = timerlat_irq; 1615 tlat->kthread = current; 1616 osn_var->pid = current->pid; 1617 /* 1618 * Anotate the arrival time. 1619 */ 1620 tlat->abs_period = hrtimer_cb_get_time(&tlat->timer); 1621 1622 wait_next_period(tlat); 1623 1624 osn_var->sampling = 1; 1625 1626 while (!kthread_should_stop()) { 1627 now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer)); 1628 diff = now - tlat->abs_period; 1629 1630 s.seqnum = tlat->count; 1631 s.timer_latency = diff; 1632 s.context = THREAD_CONTEXT; 1633 1634 trace_timerlat_sample(&s); 1635 1636 timerlat_dump_stack(time_to_us(diff)); 1637 1638 tlat->tracing_thread = false; 1639 if (osnoise_data.stop_tracing_total) 1640 if (time_to_us(diff) >= osnoise_data.stop_tracing_total) 1641 osnoise_stop_tracing(); 1642 1643 wait_next_period(tlat); 1644 } 1645 1646 hrtimer_cancel(&tlat->timer); 1647 return 0; 1648 } 1649 #else /* CONFIG_TIMERLAT_TRACER */ 1650 static int timerlat_main(void *data) 1651 { 1652 return 0; 1653 } 1654 #endif /* CONFIG_TIMERLAT_TRACER */ 1655 1656 /* 1657 * stop_kthread - stop a workload thread 1658 */ 1659 static void stop_kthread(unsigned int cpu) 1660 { 1661 struct task_struct *kthread; 1662 1663 kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread; 1664 if (kthread) 1665 kthread_stop(kthread); 1666 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL; 1667 } 1668 1669 /* 1670 * stop_per_cpu_kthread - Stop per-cpu threads 1671 * 1672 * Stop the osnoise sampling htread. Use this on unload and at system 1673 * shutdown. 1674 */ 1675 static void stop_per_cpu_kthreads(void) 1676 { 1677 int cpu; 1678 1679 cpus_read_lock(); 1680 1681 for_each_online_cpu(cpu) 1682 stop_kthread(cpu); 1683 1684 cpus_read_unlock(); 1685 } 1686 1687 /* 1688 * start_kthread - Start a workload tread 1689 */ 1690 static int start_kthread(unsigned int cpu) 1691 { 1692 struct task_struct *kthread; 1693 void *main = osnoise_main; 1694 char comm[24]; 1695 1696 if (timerlat_enabled()) { 1697 snprintf(comm, 24, "timerlat/%d", cpu); 1698 main = timerlat_main; 1699 } else { 1700 snprintf(comm, 24, "osnoise/%d", cpu); 1701 } 1702 1703 kthread = kthread_create_on_cpu(main, NULL, cpu, comm); 1704 1705 if (IS_ERR(kthread)) { 1706 pr_err(BANNER "could not start sampling thread\n"); 1707 stop_per_cpu_kthreads(); 1708 return -ENOMEM; 1709 } 1710 1711 per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread; 1712 wake_up_process(kthread); 1713 1714 return 0; 1715 } 1716 1717 /* 1718 * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads 1719 * 1720 * This starts the kernel thread that will look for osnoise on many 1721 * cpus. 1722 */ 1723 static int start_per_cpu_kthreads(void) 1724 { 1725 struct cpumask *current_mask = &save_cpumask; 1726 int retval = 0; 1727 int cpu; 1728 1729 cpus_read_lock(); 1730 /* 1731 * Run only on online CPUs in which osnoise is allowed to run. 1732 */ 1733 cpumask_and(current_mask, cpu_online_mask, &osnoise_cpumask); 1734 1735 for_each_possible_cpu(cpu) 1736 per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL; 1737 1738 for_each_cpu(cpu, current_mask) { 1739 retval = start_kthread(cpu); 1740 if (retval) { 1741 stop_per_cpu_kthreads(); 1742 break; 1743 } 1744 } 1745 1746 cpus_read_unlock(); 1747 1748 return retval; 1749 } 1750 1751 #ifdef CONFIG_HOTPLUG_CPU 1752 static void osnoise_hotplug_workfn(struct work_struct *dummy) 1753 { 1754 unsigned int cpu = smp_processor_id(); 1755 1756 mutex_lock(&trace_types_lock); 1757 1758 if (!osnoise_has_registered_instances()) 1759 goto out_unlock_trace; 1760 1761 mutex_lock(&interface_lock); 1762 cpus_read_lock(); 1763 1764 if (!cpumask_test_cpu(cpu, &osnoise_cpumask)) 1765 goto out_unlock; 1766 1767 start_kthread(cpu); 1768 1769 out_unlock: 1770 cpus_read_unlock(); 1771 mutex_unlock(&interface_lock); 1772 out_unlock_trace: 1773 mutex_unlock(&trace_types_lock); 1774 } 1775 1776 static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn); 1777 1778 /* 1779 * osnoise_cpu_init - CPU hotplug online callback function 1780 */ 1781 static int osnoise_cpu_init(unsigned int cpu) 1782 { 1783 schedule_work_on(cpu, &osnoise_hotplug_work); 1784 return 0; 1785 } 1786 1787 /* 1788 * osnoise_cpu_die - CPU hotplug offline callback function 1789 */ 1790 static int osnoise_cpu_die(unsigned int cpu) 1791 { 1792 stop_kthread(cpu); 1793 return 0; 1794 } 1795 1796 static void osnoise_init_hotplug_support(void) 1797 { 1798 int ret; 1799 1800 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online", 1801 osnoise_cpu_init, osnoise_cpu_die); 1802 if (ret < 0) 1803 pr_warn(BANNER "Error to init cpu hotplug support\n"); 1804 1805 return; 1806 } 1807 #else /* CONFIG_HOTPLUG_CPU */ 1808 static void osnoise_init_hotplug_support(void) 1809 { 1810 return; 1811 } 1812 #endif /* CONFIG_HOTPLUG_CPU */ 1813 1814 /* 1815 * osnoise_cpus_read - Read function for reading the "cpus" file 1816 * @filp: The active open file structure 1817 * @ubuf: The userspace provided buffer to read value into 1818 * @cnt: The maximum number of bytes to read 1819 * @ppos: The current "file" position 1820 * 1821 * Prints the "cpus" output into the user-provided buffer. 1822 */ 1823 static ssize_t 1824 osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count, 1825 loff_t *ppos) 1826 { 1827 char *mask_str; 1828 int len; 1829 1830 mutex_lock(&interface_lock); 1831 1832 len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1; 1833 mask_str = kmalloc(len, GFP_KERNEL); 1834 if (!mask_str) { 1835 count = -ENOMEM; 1836 goto out_unlock; 1837 } 1838 1839 len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)); 1840 if (len >= count) { 1841 count = -EINVAL; 1842 goto out_free; 1843 } 1844 1845 count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len); 1846 1847 out_free: 1848 kfree(mask_str); 1849 out_unlock: 1850 mutex_unlock(&interface_lock); 1851 1852 return count; 1853 } 1854 1855 /* 1856 * osnoise_cpus_write - Write function for "cpus" entry 1857 * @filp: The active open file structure 1858 * @ubuf: The user buffer that contains the value to write 1859 * @cnt: The maximum number of bytes to write to "file" 1860 * @ppos: The current position in @file 1861 * 1862 * This function provides a write implementation for the "cpus" 1863 * interface to the osnoise trace. By default, it lists all CPUs, 1864 * in this way, allowing osnoise threads to run on any online CPU 1865 * of the system. It serves to restrict the execution of osnoise to the 1866 * set of CPUs writing via this interface. Why not use "tracing_cpumask"? 1867 * Because the user might be interested in tracing what is running on 1868 * other CPUs. For instance, one might run osnoise in one HT CPU 1869 * while observing what is running on the sibling HT CPU. 1870 */ 1871 static ssize_t 1872 osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count, 1873 loff_t *ppos) 1874 { 1875 cpumask_var_t osnoise_cpumask_new; 1876 int running, err; 1877 char buf[256]; 1878 1879 if (count >= 256) 1880 return -EINVAL; 1881 1882 if (copy_from_user(buf, ubuf, count)) 1883 return -EFAULT; 1884 1885 if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL)) 1886 return -ENOMEM; 1887 1888 err = cpulist_parse(buf, osnoise_cpumask_new); 1889 if (err) 1890 goto err_free; 1891 1892 /* 1893 * trace_types_lock is taken to avoid concurrency on start/stop. 1894 */ 1895 mutex_lock(&trace_types_lock); 1896 running = osnoise_has_registered_instances(); 1897 if (running) 1898 stop_per_cpu_kthreads(); 1899 1900 mutex_lock(&interface_lock); 1901 /* 1902 * osnoise_cpumask is read by CPU hotplug operations. 1903 */ 1904 cpus_read_lock(); 1905 1906 cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new); 1907 1908 cpus_read_unlock(); 1909 mutex_unlock(&interface_lock); 1910 1911 if (running) 1912 start_per_cpu_kthreads(); 1913 mutex_unlock(&trace_types_lock); 1914 1915 free_cpumask_var(osnoise_cpumask_new); 1916 return count; 1917 1918 err_free: 1919 free_cpumask_var(osnoise_cpumask_new); 1920 1921 return err; 1922 } 1923 1924 /* 1925 * osnoise/runtime_us: cannot be greater than the period. 1926 */ 1927 static struct trace_min_max_param osnoise_runtime = { 1928 .lock = &interface_lock, 1929 .val = &osnoise_data.sample_runtime, 1930 .max = &osnoise_data.sample_period, 1931 .min = NULL, 1932 }; 1933 1934 /* 1935 * osnoise/period_us: cannot be smaller than the runtime. 1936 */ 1937 static struct trace_min_max_param osnoise_period = { 1938 .lock = &interface_lock, 1939 .val = &osnoise_data.sample_period, 1940 .max = NULL, 1941 .min = &osnoise_data.sample_runtime, 1942 }; 1943 1944 /* 1945 * osnoise/stop_tracing_us: no limit. 1946 */ 1947 static struct trace_min_max_param osnoise_stop_tracing_in = { 1948 .lock = &interface_lock, 1949 .val = &osnoise_data.stop_tracing, 1950 .max = NULL, 1951 .min = NULL, 1952 }; 1953 1954 /* 1955 * osnoise/stop_tracing_total_us: no limit. 1956 */ 1957 static struct trace_min_max_param osnoise_stop_tracing_total = { 1958 .lock = &interface_lock, 1959 .val = &osnoise_data.stop_tracing_total, 1960 .max = NULL, 1961 .min = NULL, 1962 }; 1963 1964 #ifdef CONFIG_TIMERLAT_TRACER 1965 /* 1966 * osnoise/print_stack: print the stacktrace of the IRQ handler if the total 1967 * latency is higher than val. 1968 */ 1969 static struct trace_min_max_param osnoise_print_stack = { 1970 .lock = &interface_lock, 1971 .val = &osnoise_data.print_stack, 1972 .max = NULL, 1973 .min = NULL, 1974 }; 1975 1976 /* 1977 * osnoise/timerlat_period: min 100 us, max 1 s 1978 */ 1979 u64 timerlat_min_period = 100; 1980 u64 timerlat_max_period = 1000000; 1981 static struct trace_min_max_param timerlat_period = { 1982 .lock = &interface_lock, 1983 .val = &osnoise_data.timerlat_period, 1984 .max = &timerlat_max_period, 1985 .min = &timerlat_min_period, 1986 }; 1987 #endif 1988 1989 static const struct file_operations cpus_fops = { 1990 .open = tracing_open_generic, 1991 .read = osnoise_cpus_read, 1992 .write = osnoise_cpus_write, 1993 .llseek = generic_file_llseek, 1994 }; 1995 1996 #ifdef CONFIG_TIMERLAT_TRACER 1997 #ifdef CONFIG_STACKTRACE 1998 static int init_timerlat_stack_tracefs(struct dentry *top_dir) 1999 { 2000 struct dentry *tmp; 2001 2002 tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir, 2003 &osnoise_print_stack, &trace_min_max_fops); 2004 if (!tmp) 2005 return -ENOMEM; 2006 2007 return 0; 2008 } 2009 #else /* CONFIG_STACKTRACE */ 2010 static int init_timerlat_stack_tracefs(struct dentry *top_dir) 2011 { 2012 return 0; 2013 } 2014 #endif /* CONFIG_STACKTRACE */ 2015 2016 /* 2017 * init_timerlat_tracefs - A function to initialize the timerlat interface files 2018 */ 2019 static int init_timerlat_tracefs(struct dentry *top_dir) 2020 { 2021 struct dentry *tmp; 2022 2023 tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir, 2024 &timerlat_period, &trace_min_max_fops); 2025 if (!tmp) 2026 return -ENOMEM; 2027 2028 return init_timerlat_stack_tracefs(top_dir); 2029 } 2030 #else /* CONFIG_TIMERLAT_TRACER */ 2031 static int init_timerlat_tracefs(struct dentry *top_dir) 2032 { 2033 return 0; 2034 } 2035 #endif /* CONFIG_TIMERLAT_TRACER */ 2036 2037 /* 2038 * init_tracefs - A function to initialize the tracefs interface files 2039 * 2040 * This function creates entries in tracefs for "osnoise" and "timerlat". 2041 * It creates these directories in the tracing directory, and within that 2042 * directory the use can change and view the configs. 2043 */ 2044 static int init_tracefs(void) 2045 { 2046 struct dentry *top_dir; 2047 struct dentry *tmp; 2048 int ret; 2049 2050 ret = tracing_init_dentry(); 2051 if (ret) 2052 return -ENOMEM; 2053 2054 top_dir = tracefs_create_dir("osnoise", NULL); 2055 if (!top_dir) 2056 return 0; 2057 2058 tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir, 2059 &osnoise_period, &trace_min_max_fops); 2060 if (!tmp) 2061 goto err; 2062 2063 tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir, 2064 &osnoise_runtime, &trace_min_max_fops); 2065 if (!tmp) 2066 goto err; 2067 2068 tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir, 2069 &osnoise_stop_tracing_in, &trace_min_max_fops); 2070 if (!tmp) 2071 goto err; 2072 2073 tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir, 2074 &osnoise_stop_tracing_total, &trace_min_max_fops); 2075 if (!tmp) 2076 goto err; 2077 2078 tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops); 2079 if (!tmp) 2080 goto err; 2081 2082 ret = init_timerlat_tracefs(top_dir); 2083 if (ret) 2084 goto err; 2085 2086 return 0; 2087 2088 err: 2089 tracefs_remove(top_dir); 2090 return -ENOMEM; 2091 } 2092 2093 static int osnoise_hook_events(void) 2094 { 2095 int retval; 2096 2097 /* 2098 * Trace is already hooked, we are re-enabling from 2099 * a stop_tracing_*. 2100 */ 2101 if (trace_osnoise_callback_enabled) 2102 return 0; 2103 2104 retval = hook_irq_events(); 2105 if (retval) 2106 return -EINVAL; 2107 2108 retval = hook_softirq_events(); 2109 if (retval) 2110 goto out_unhook_irq; 2111 2112 retval = hook_thread_events(); 2113 /* 2114 * All fine! 2115 */ 2116 if (!retval) 2117 return 0; 2118 2119 unhook_softirq_events(); 2120 out_unhook_irq: 2121 unhook_irq_events(); 2122 return -EINVAL; 2123 } 2124 2125 /* 2126 * osnoise_workload_start - start the workload and hook to events 2127 */ 2128 static int osnoise_workload_start(void) 2129 { 2130 int retval; 2131 2132 /* 2133 * Instances need to be registered after calling workload 2134 * start. Hence, if there is already an instance, the 2135 * workload was already registered. Otherwise, this 2136 * code is on the way to register the first instance, 2137 * and the workload will start. 2138 */ 2139 if (osnoise_has_registered_instances()) 2140 return 0; 2141 2142 osn_var_reset_all(); 2143 2144 retval = osnoise_hook_events(); 2145 if (retval) 2146 return retval; 2147 2148 /* 2149 * Make sure that ftrace_nmi_enter/exit() see reset values 2150 * before enabling trace_osnoise_callback_enabled. 2151 */ 2152 barrier(); 2153 trace_osnoise_callback_enabled = true; 2154 2155 retval = start_per_cpu_kthreads(); 2156 if (retval) { 2157 unhook_irq_events(); 2158 return retval; 2159 } 2160 2161 return 0; 2162 } 2163 2164 /* 2165 * osnoise_workload_stop - stop the workload and unhook the events 2166 */ 2167 static void osnoise_workload_stop(void) 2168 { 2169 /* 2170 * Instances need to be unregistered before calling 2171 * stop. Hence, if there is a registered instance, more 2172 * than one instance is running, and the workload will not 2173 * yet stop. Otherwise, this code is on the way to disable 2174 * the last instance, and the workload can stop. 2175 */ 2176 if (osnoise_has_registered_instances()) 2177 return; 2178 2179 trace_osnoise_callback_enabled = false; 2180 /* 2181 * Make sure that ftrace_nmi_enter/exit() see 2182 * trace_osnoise_callback_enabled as false before continuing. 2183 */ 2184 barrier(); 2185 2186 stop_per_cpu_kthreads(); 2187 2188 unhook_irq_events(); 2189 unhook_softirq_events(); 2190 unhook_thread_events(); 2191 } 2192 2193 static void osnoise_tracer_start(struct trace_array *tr) 2194 { 2195 int retval; 2196 2197 /* 2198 * If the instance is already registered, there is no need to 2199 * register it again. 2200 */ 2201 if (osnoise_instance_registered(tr)) 2202 return; 2203 2204 retval = osnoise_workload_start(); 2205 if (retval) 2206 pr_err(BANNER "Error starting osnoise tracer\n"); 2207 2208 osnoise_register_instance(tr); 2209 } 2210 2211 static void osnoise_tracer_stop(struct trace_array *tr) 2212 { 2213 osnoise_unregister_instance(tr); 2214 osnoise_workload_stop(); 2215 } 2216 2217 static int osnoise_tracer_init(struct trace_array *tr) 2218 { 2219 /* 2220 * Only allow osnoise tracer if timerlat tracer is not running 2221 * already. 2222 */ 2223 if (timerlat_enabled()) 2224 return -EBUSY; 2225 2226 tr->max_latency = 0; 2227 2228 osnoise_tracer_start(tr); 2229 return 0; 2230 } 2231 2232 static void osnoise_tracer_reset(struct trace_array *tr) 2233 { 2234 osnoise_tracer_stop(tr); 2235 } 2236 2237 static struct tracer osnoise_tracer __read_mostly = { 2238 .name = "osnoise", 2239 .init = osnoise_tracer_init, 2240 .reset = osnoise_tracer_reset, 2241 .start = osnoise_tracer_start, 2242 .stop = osnoise_tracer_stop, 2243 .print_header = print_osnoise_headers, 2244 .allow_instances = true, 2245 }; 2246 2247 #ifdef CONFIG_TIMERLAT_TRACER 2248 static void timerlat_tracer_start(struct trace_array *tr) 2249 { 2250 int retval; 2251 2252 /* 2253 * If the instance is already registered, there is no need to 2254 * register it again. 2255 */ 2256 if (osnoise_instance_registered(tr)) 2257 return; 2258 2259 retval = osnoise_workload_start(); 2260 if (retval) 2261 pr_err(BANNER "Error starting timerlat tracer\n"); 2262 2263 osnoise_register_instance(tr); 2264 2265 return; 2266 } 2267 2268 static void timerlat_tracer_stop(struct trace_array *tr) 2269 { 2270 int cpu; 2271 2272 osnoise_unregister_instance(tr); 2273 2274 /* 2275 * Instruct the threads to stop only if this is the last instance. 2276 */ 2277 if (!osnoise_has_registered_instances()) { 2278 for_each_online_cpu(cpu) 2279 per_cpu(per_cpu_osnoise_var, cpu).sampling = 0; 2280 } 2281 2282 osnoise_workload_stop(); 2283 } 2284 2285 static int timerlat_tracer_init(struct trace_array *tr) 2286 { 2287 /* 2288 * Only allow timerlat tracer if osnoise tracer is not running already. 2289 */ 2290 if (osnoise_has_registered_instances() && !osnoise_data.timerlat_tracer) 2291 return -EBUSY; 2292 2293 /* 2294 * If this is the first instance, set timerlat_tracer to block 2295 * osnoise tracer start. 2296 */ 2297 if (!osnoise_has_registered_instances()) 2298 osnoise_data.timerlat_tracer = 1; 2299 2300 tr->max_latency = 0; 2301 timerlat_tracer_start(tr); 2302 2303 return 0; 2304 } 2305 2306 static void timerlat_tracer_reset(struct trace_array *tr) 2307 { 2308 timerlat_tracer_stop(tr); 2309 2310 /* 2311 * If this is the last instance, reset timerlat_tracer allowing 2312 * osnoise to be started. 2313 */ 2314 if (!osnoise_has_registered_instances()) 2315 osnoise_data.timerlat_tracer = 0; 2316 } 2317 2318 static struct tracer timerlat_tracer __read_mostly = { 2319 .name = "timerlat", 2320 .init = timerlat_tracer_init, 2321 .reset = timerlat_tracer_reset, 2322 .start = timerlat_tracer_start, 2323 .stop = timerlat_tracer_stop, 2324 .print_header = print_timerlat_headers, 2325 .allow_instances = true, 2326 }; 2327 2328 __init static int init_timerlat_tracer(void) 2329 { 2330 return register_tracer(&timerlat_tracer); 2331 } 2332 #else /* CONFIG_TIMERLAT_TRACER */ 2333 __init static int init_timerlat_tracer(void) 2334 { 2335 return 0; 2336 } 2337 #endif /* CONFIG_TIMERLAT_TRACER */ 2338 2339 __init static int init_osnoise_tracer(void) 2340 { 2341 int ret; 2342 2343 mutex_init(&interface_lock); 2344 2345 cpumask_copy(&osnoise_cpumask, cpu_all_mask); 2346 2347 ret = register_tracer(&osnoise_tracer); 2348 if (ret) { 2349 pr_err(BANNER "Error registering osnoise!\n"); 2350 return ret; 2351 } 2352 2353 ret = init_timerlat_tracer(); 2354 if (ret) { 2355 pr_err(BANNER "Error registering timerlat!\n"); 2356 return ret; 2357 } 2358 2359 osnoise_init_hotplug_support(); 2360 2361 INIT_LIST_HEAD_RCU(&osnoise_instances); 2362 2363 init_tracefs(); 2364 2365 return 0; 2366 } 2367 late_initcall(init_osnoise_tracer); 2368