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 * osnoise/options entries.
  53 */
  54enum osnoise_options_index {
  55	OSN_DEFAULTS = 0,
  56	OSN_WORKLOAD,
  57	OSN_PANIC_ON_STOP,
  58	OSN_PREEMPT_DISABLE,
  59	OSN_IRQ_DISABLE,
  60	OSN_MAX
  61};
  62
  63static const char * const osnoise_options_str[OSN_MAX] = {
  64							"DEFAULTS",
  65							"OSNOISE_WORKLOAD",
  66							"PANIC_ON_STOP",
  67							"OSNOISE_PREEMPT_DISABLE",
  68							"OSNOISE_IRQ_DISABLE" };
  69
  70#define OSN_DEFAULT_OPTIONS		0x2
  71static unsigned long osnoise_options	= OSN_DEFAULT_OPTIONS;
  72
  73/*
  74 * trace_array of the enabled osnoise/timerlat instances.
  75 */
  76struct osnoise_instance {
  77	struct list_head	list;
  78	struct trace_array	*tr;
  79};
  80
  81static struct list_head osnoise_instances;
  82
  83static bool osnoise_has_registered_instances(void)
  84{
  85	return !!list_first_or_null_rcu(&osnoise_instances,
  86					struct osnoise_instance,
  87					list);
  88}
  89
  90/*
  91 * osnoise_instance_registered - check if a tr is already registered
  92 */
  93static int osnoise_instance_registered(struct trace_array *tr)
  94{
  95	struct osnoise_instance *inst;
  96	int found = 0;
  97
  98	rcu_read_lock();
  99	list_for_each_entry_rcu(inst, &osnoise_instances, list) {
 100		if (inst->tr == tr)
 101			found = 1;
 102	}
 103	rcu_read_unlock();
 104
 105	return found;
 106}
 107
 108/*
 109 * osnoise_register_instance - register a new trace instance
 110 *
 111 * Register a trace_array *tr in the list of instances running
 112 * osnoise/timerlat tracers.
 113 */
 114static int osnoise_register_instance(struct trace_array *tr)
 115{
 116	struct osnoise_instance *inst;
 117
 118	/*
 119	 * register/unregister serialization is provided by trace's
 120	 * trace_types_lock.
 121	 */
 122	lockdep_assert_held(&trace_types_lock);
 123
 124	inst = kmalloc(sizeof(*inst), GFP_KERNEL);
 125	if (!inst)
 126		return -ENOMEM;
 127
 128	INIT_LIST_HEAD_RCU(&inst->list);
 129	inst->tr = tr;
 130	list_add_tail_rcu(&inst->list, &osnoise_instances);
 131
 132	return 0;
 133}
 134
 135/*
 136 *  osnoise_unregister_instance - unregister a registered trace instance
 137 *
 138 * Remove the trace_array *tr from the list of instances running
 139 * osnoise/timerlat tracers.
 140 */
 141static void osnoise_unregister_instance(struct trace_array *tr)
 142{
 143	struct osnoise_instance *inst;
 144	int found = 0;
 145
 146	/*
 147	 * register/unregister serialization is provided by trace's
 148	 * trace_types_lock.
 149	 */
 150	list_for_each_entry_rcu(inst, &osnoise_instances, list,
 151				lockdep_is_held(&trace_types_lock)) {
 152		if (inst->tr == tr) {
 153			list_del_rcu(&inst->list);
 154			found = 1;
 155			break;
 156		}
 157	}
 158
 159	if (!found)
 160		return;
 161
 162	kvfree_rcu(inst);
 163}
 164
 165/*
 166 * NMI runtime info.
 167 */
 168struct osn_nmi {
 169	u64	count;
 170	u64	delta_start;
 171};
 172
 173/*
 174 * IRQ runtime info.
 175 */
 176struct osn_irq {
 177	u64	count;
 178	u64	arrival_time;
 179	u64	delta_start;
 180};
 181
 182#define IRQ_CONTEXT	0
 183#define THREAD_CONTEXT	1
 
 184/*
 185 * sofirq runtime info.
 186 */
 187struct osn_softirq {
 188	u64	count;
 189	u64	arrival_time;
 190	u64	delta_start;
 191};
 192
 193/*
 194 * thread runtime info.
 195 */
 196struct osn_thread {
 197	u64	count;
 198	u64	arrival_time;
 199	u64	delta_start;
 200};
 201
 202/*
 203 * Runtime information: this structure saves the runtime information used by
 204 * one sampling thread.
 205 */
 206struct osnoise_variables {
 207	struct task_struct	*kthread;
 208	bool			sampling;
 209	pid_t			pid;
 210	struct osn_nmi		nmi;
 211	struct osn_irq		irq;
 212	struct osn_softirq	softirq;
 213	struct osn_thread	thread;
 214	local_t			int_counter;
 215};
 216
 217/*
 218 * Per-cpu runtime information.
 219 */
 220DEFINE_PER_CPU(struct osnoise_variables, per_cpu_osnoise_var);
 221
 222/*
 223 * this_cpu_osn_var - Return the per-cpu osnoise_variables on its relative CPU
 224 */
 225static inline struct osnoise_variables *this_cpu_osn_var(void)
 226{
 227	return this_cpu_ptr(&per_cpu_osnoise_var);
 228}
 229
 
 
 
 
 
 230#ifdef CONFIG_TIMERLAT_TRACER
 231/*
 232 * Runtime information for the timer mode.
 233 */
 234struct timerlat_variables {
 235	struct task_struct	*kthread;
 236	struct hrtimer		timer;
 237	u64			rel_period;
 238	u64			abs_period;
 239	bool			tracing_thread;
 240	u64			count;
 
 241};
 242
 243DEFINE_PER_CPU(struct timerlat_variables, per_cpu_timerlat_var);
 244
 245/*
 246 * this_cpu_tmr_var - Return the per-cpu timerlat_variables on its relative CPU
 247 */
 248static inline struct timerlat_variables *this_cpu_tmr_var(void)
 249{
 250	return this_cpu_ptr(&per_cpu_timerlat_var);
 251}
 252
 253/*
 254 * tlat_var_reset - Reset the values of the given timerlat_variables
 255 */
 256static inline void tlat_var_reset(void)
 257{
 258	struct timerlat_variables *tlat_var;
 259	int cpu;
 
 
 
 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		tlat_var = per_cpu_ptr(&per_cpu_timerlat_var, cpu);
 
 
 266		memset(tlat_var, 0, sizeof(*tlat_var));
 267	}
 
 268}
 269#else /* CONFIG_TIMERLAT_TRACER */
 270#define tlat_var_reset()	do {} while (0)
 271#endif /* CONFIG_TIMERLAT_TRACER */
 272
 273/*
 274 * osn_var_reset - Reset the values of the given osnoise_variables
 275 */
 276static inline void osn_var_reset(void)
 277{
 278	struct osnoise_variables *osn_var;
 279	int cpu;
 280
 281	/*
 282	 * So far, all the values are initialized as 0, so
 283	 * zeroing the structure is perfect.
 284	 */
 285	for_each_cpu(cpu, cpu_online_mask) {
 286		osn_var = per_cpu_ptr(&per_cpu_osnoise_var, cpu);
 287		memset(osn_var, 0, sizeof(*osn_var));
 288	}
 289}
 290
 291/*
 292 * osn_var_reset_all - Reset the value of all per-cpu osnoise_variables
 293 */
 294static inline void osn_var_reset_all(void)
 295{
 296	osn_var_reset();
 297	tlat_var_reset();
 298}
 299
 300/*
 301 * Tells NMIs to call back to the osnoise tracer to record timestamps.
 302 */
 303bool trace_osnoise_callback_enabled;
 304
 305/*
 306 * osnoise sample structure definition. Used to store the statistics of a
 307 * sample run.
 308 */
 309struct osnoise_sample {
 310	u64			runtime;	/* runtime */
 311	u64			noise;		/* noise */
 312	u64			max_sample;	/* max single noise sample */
 313	int			hw_count;	/* # HW (incl. hypervisor) interference */
 314	int			nmi_count;	/* # NMIs during this sample */
 315	int			irq_count;	/* # IRQs during this sample */
 316	int			softirq_count;	/* # softirqs during this sample */
 317	int			thread_count;	/* # threads during this sample */
 318};
 319
 320#ifdef CONFIG_TIMERLAT_TRACER
 321/*
 322 * timerlat sample structure definition. Used to store the statistics of
 323 * a sample run.
 324 */
 325struct timerlat_sample {
 326	u64			timer_latency;	/* timer_latency */
 327	unsigned int		seqnum;		/* unique sequence */
 328	int			context;	/* timer context */
 329};
 330#endif
 331
 332/*
 333 * Protect the interface.
 334 */
 335struct mutex interface_lock;
 336
 337/*
 338 * Tracer data.
 339 */
 340static struct osnoise_data {
 341	u64	sample_period;		/* total sampling period */
 342	u64	sample_runtime;		/* active sampling portion of period */
 343	u64	stop_tracing;		/* stop trace in the internal operation (loop/irq) */
 344	u64	stop_tracing_total;	/* stop trace in the final operation (report/thread) */
 345#ifdef CONFIG_TIMERLAT_TRACER
 346	u64	timerlat_period;	/* timerlat period */
 347	u64	print_stack;		/* print IRQ stack if total > */
 348	int	timerlat_tracer;	/* timerlat tracer */
 349#endif
 350	bool	tainted;		/* infor users and developers about a problem */
 351} osnoise_data = {
 352	.sample_period			= DEFAULT_SAMPLE_PERIOD,
 353	.sample_runtime			= DEFAULT_SAMPLE_RUNTIME,
 354	.stop_tracing			= 0,
 355	.stop_tracing_total		= 0,
 356#ifdef CONFIG_TIMERLAT_TRACER
 357	.print_stack			= 0,
 358	.timerlat_period		= DEFAULT_TIMERLAT_PERIOD,
 359	.timerlat_tracer		= 0,
 360#endif
 361};
 362
 363#ifdef CONFIG_TIMERLAT_TRACER
 364static inline bool timerlat_enabled(void)
 365{
 366	return osnoise_data.timerlat_tracer;
 367}
 368
 369static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
 370{
 371	struct timerlat_variables *tlat_var = this_cpu_tmr_var();
 372	/*
 373	 * If the timerlat is enabled, but the irq handler did
 374	 * not run yet enabling timerlat_tracer, do not trace.
 375	 */
 376	if (!tlat_var->tracing_thread) {
 377		osn_var->softirq.arrival_time = 0;
 378		osn_var->softirq.delta_start = 0;
 379		return 0;
 380	}
 381	return 1;
 382}
 383
 384static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
 385{
 386	struct timerlat_variables *tlat_var = this_cpu_tmr_var();
 387	/*
 388	 * If the timerlat is enabled, but the irq handler did
 389	 * not run yet enabling timerlat_tracer, do not trace.
 390	 */
 391	if (!tlat_var->tracing_thread) {
 392		osn_var->thread.delta_start = 0;
 393		osn_var->thread.arrival_time = 0;
 394		return 0;
 395	}
 396	return 1;
 397}
 398#else /* CONFIG_TIMERLAT_TRACER */
 399static inline bool timerlat_enabled(void)
 400{
 401	return false;
 402}
 403
 404static inline int timerlat_softirq_exit(struct osnoise_variables *osn_var)
 405{
 406	return 1;
 407}
 408static inline int timerlat_thread_exit(struct osnoise_variables *osn_var)
 409{
 410	return 1;
 411}
 412#endif
 413
 414#ifdef CONFIG_PREEMPT_RT
 415/*
 416 * Print the osnoise header info.
 417 */
 418static void print_osnoise_headers(struct seq_file *s)
 419{
 420	if (osnoise_data.tainted)
 421		seq_puts(s, "# osnoise is tainted!\n");
 422
 423	seq_puts(s, "#                                _-------=> irqs-off\n");
 424	seq_puts(s, "#                               / _------=> need-resched\n");
 425	seq_puts(s, "#                              | / _-----=> need-resched-lazy\n");
 426	seq_puts(s, "#                              || / _----=> hardirq/softirq\n");
 427	seq_puts(s, "#                              ||| / _---=> preempt-depth\n");
 428	seq_puts(s, "#                              |||| / _--=> preempt-lazy-depth\n");
 429	seq_puts(s, "#                              ||||| / _-=> migrate-disable\n");
 430
 431	seq_puts(s, "#                              |||||| /          ");
 432	seq_puts(s, "                                     MAX\n");
 433
 434	seq_puts(s, "#                              ||||| /                         ");
 435	seq_puts(s, "                    SINGLE      Interference counters:\n");
 436
 437	seq_puts(s, "#                              |||||||               RUNTIME   ");
 438	seq_puts(s, "   NOISE  %% OF CPU  NOISE    +-----------------------------+\n");
 439
 440	seq_puts(s, "#           TASK-PID      CPU# |||||||   TIMESTAMP    IN US    ");
 441	seq_puts(s, "   IN US  AVAILABLE  IN US     HW    NMI    IRQ   SIRQ THREAD\n");
 442
 443	seq_puts(s, "#              | |         |   |||||||      |           |      ");
 444	seq_puts(s, "       |    |            |      |      |      |      |      |\n");
 445}
 446#else /* CONFIG_PREEMPT_RT */
 447static void print_osnoise_headers(struct seq_file *s)
 448{
 449	if (osnoise_data.tainted)
 450		seq_puts(s, "# osnoise is tainted!\n");
 451
 452	seq_puts(s, "#                                _-----=> irqs-off\n");
 453	seq_puts(s, "#                               / _----=> need-resched\n");
 454	seq_puts(s, "#                              | / _---=> hardirq/softirq\n");
 455	seq_puts(s, "#                              || / _--=> preempt-depth\n");
 456	seq_puts(s, "#                              ||| / _-=> migrate-disable     ");
 457	seq_puts(s, "                    MAX\n");
 458	seq_puts(s, "#                              |||| /     delay               ");
 459	seq_puts(s, "                    SINGLE      Interference counters:\n");
 460
 461	seq_puts(s, "#                              |||||               RUNTIME   ");
 462	seq_puts(s, "   NOISE  %% OF CPU  NOISE    +-----------------------------+\n");
 463
 464	seq_puts(s, "#           TASK-PID      CPU# |||||   TIMESTAMP    IN US    ");
 465	seq_puts(s, "   IN US  AVAILABLE  IN US     HW    NMI    IRQ   SIRQ THREAD\n");
 466
 467	seq_puts(s, "#              | |         |   |||||      |           |      ");
 468	seq_puts(s, "       |    |            |      |      |      |      |      |\n");
 469}
 470#endif /* CONFIG_PREEMPT_RT */
 471
 472/*
 473 * osnoise_taint - report an osnoise error.
 474 */
 475#define osnoise_taint(msg) ({							\
 476	struct osnoise_instance *inst;						\
 477	struct trace_buffer *buffer;						\
 478										\
 479	rcu_read_lock();							\
 480	list_for_each_entry_rcu(inst, &osnoise_instances, list) {		\
 481		buffer = inst->tr->array_buffer.buffer;				\
 482		trace_array_printk_buf(buffer, _THIS_IP_, msg);			\
 483	}									\
 484	rcu_read_unlock();							\
 485	osnoise_data.tainted = true;						\
 486})
 487
 488/*
 489 * Record an osnoise_sample into the tracer buffer.
 490 */
 491static void
 492__trace_osnoise_sample(struct osnoise_sample *sample, struct trace_buffer *buffer)
 493{
 494	struct trace_event_call *call = &event_osnoise;
 495	struct ring_buffer_event *event;
 496	struct osnoise_entry *entry;
 497
 498	event = trace_buffer_lock_reserve(buffer, TRACE_OSNOISE, sizeof(*entry),
 499					  tracing_gen_ctx());
 500	if (!event)
 501		return;
 502	entry	= ring_buffer_event_data(event);
 503	entry->runtime		= sample->runtime;
 504	entry->noise		= sample->noise;
 505	entry->max_sample	= sample->max_sample;
 506	entry->hw_count		= sample->hw_count;
 507	entry->nmi_count	= sample->nmi_count;
 508	entry->irq_count	= sample->irq_count;
 509	entry->softirq_count	= sample->softirq_count;
 510	entry->thread_count	= sample->thread_count;
 511
 512	if (!call_filter_check_discard(call, entry, buffer, event))
 513		trace_buffer_unlock_commit_nostack(buffer, event);
 514}
 515
 516/*
 517 * Record an osnoise_sample on all osnoise instances.
 518 */
 519static void trace_osnoise_sample(struct osnoise_sample *sample)
 520{
 521	struct osnoise_instance *inst;
 522	struct trace_buffer *buffer;
 523
 524	rcu_read_lock();
 525	list_for_each_entry_rcu(inst, &osnoise_instances, list) {
 526		buffer = inst->tr->array_buffer.buffer;
 527		__trace_osnoise_sample(sample, buffer);
 528	}
 529	rcu_read_unlock();
 530}
 531
 532#ifdef CONFIG_TIMERLAT_TRACER
 533/*
 534 * Print the timerlat header info.
 535 */
 536#ifdef CONFIG_PREEMPT_RT
 537static void print_timerlat_headers(struct seq_file *s)
 538{
 539	seq_puts(s, "#                                _-------=> irqs-off\n");
 540	seq_puts(s, "#                               / _------=> need-resched\n");
 541	seq_puts(s, "#                              | / _-----=> need-resched-lazy\n");
 542	seq_puts(s, "#                              || / _----=> hardirq/softirq\n");
 543	seq_puts(s, "#                              ||| / _---=> preempt-depth\n");
 544	seq_puts(s, "#                              |||| / _--=> preempt-lazy-depth\n");
 545	seq_puts(s, "#                              ||||| / _-=> migrate-disable\n");
 546	seq_puts(s, "#                              |||||| /\n");
 547	seq_puts(s, "#                              |||||||             ACTIVATION\n");
 548	seq_puts(s, "#           TASK-PID      CPU# |||||||   TIMESTAMP    ID     ");
 549	seq_puts(s, "       CONTEXT                LATENCY\n");
 550	seq_puts(s, "#              | |         |   |||||||      |         |      ");
 551	seq_puts(s, "            |                       |\n");
 552}
 553#else /* CONFIG_PREEMPT_RT */
 554static void print_timerlat_headers(struct seq_file *s)
 555{
 556	seq_puts(s, "#                                _-----=> irqs-off\n");
 557	seq_puts(s, "#                               / _----=> need-resched\n");
 558	seq_puts(s, "#                              | / _---=> hardirq/softirq\n");
 559	seq_puts(s, "#                              || / _--=> preempt-depth\n");
 560	seq_puts(s, "#                              ||| / _-=> migrate-disable\n");
 561	seq_puts(s, "#                              |||| /     delay\n");
 562	seq_puts(s, "#                              |||||            ACTIVATION\n");
 563	seq_puts(s, "#           TASK-PID      CPU# |||||   TIMESTAMP   ID      ");
 564	seq_puts(s, "      CONTEXT                 LATENCY\n");
 565	seq_puts(s, "#              | |         |   |||||      |         |      ");
 566	seq_puts(s, "            |                       |\n");
 567}
 568#endif /* CONFIG_PREEMPT_RT */
 569
 570static void
 571__trace_timerlat_sample(struct timerlat_sample *sample, struct trace_buffer *buffer)
 572{
 573	struct trace_event_call *call = &event_osnoise;
 574	struct ring_buffer_event *event;
 575	struct timerlat_entry *entry;
 576
 577	event = trace_buffer_lock_reserve(buffer, TRACE_TIMERLAT, sizeof(*entry),
 578					  tracing_gen_ctx());
 579	if (!event)
 580		return;
 581	entry	= ring_buffer_event_data(event);
 582	entry->seqnum			= sample->seqnum;
 583	entry->context			= sample->context;
 584	entry->timer_latency		= sample->timer_latency;
 585
 586	if (!call_filter_check_discard(call, entry, buffer, event))
 587		trace_buffer_unlock_commit_nostack(buffer, event);
 588}
 589
 590/*
 591 * Record an timerlat_sample into the tracer buffer.
 592 */
 593static void trace_timerlat_sample(struct timerlat_sample *sample)
 594{
 595	struct osnoise_instance *inst;
 596	struct trace_buffer *buffer;
 597
 598	rcu_read_lock();
 599	list_for_each_entry_rcu(inst, &osnoise_instances, list) {
 600		buffer = inst->tr->array_buffer.buffer;
 601		__trace_timerlat_sample(sample, buffer);
 602	}
 603	rcu_read_unlock();
 604}
 605
 606#ifdef CONFIG_STACKTRACE
 607
 608#define	MAX_CALLS	256
 609
 610/*
 611 * Stack trace will take place only at IRQ level, so, no need
 612 * to control nesting here.
 613 */
 614struct trace_stack {
 615	int		stack_size;
 616	int		nr_entries;
 617	unsigned long	calls[MAX_CALLS];
 618};
 619
 620static DEFINE_PER_CPU(struct trace_stack, trace_stack);
 621
 622/*
 623 * timerlat_save_stack - save a stack trace without printing
 624 *
 625 * Save the current stack trace without printing. The
 626 * stack will be printed later, after the end of the measurement.
 627 */
 628static void timerlat_save_stack(int skip)
 629{
 630	unsigned int size, nr_entries;
 631	struct trace_stack *fstack;
 632
 633	fstack = this_cpu_ptr(&trace_stack);
 634
 635	size = ARRAY_SIZE(fstack->calls);
 636
 637	nr_entries = stack_trace_save(fstack->calls, size, skip);
 638
 639	fstack->stack_size = nr_entries * sizeof(unsigned long);
 640	fstack->nr_entries = nr_entries;
 641
 642	return;
 643
 644}
 645
 646static void
 647__timerlat_dump_stack(struct trace_buffer *buffer, struct trace_stack *fstack, unsigned int size)
 648{
 649	struct trace_event_call *call = &event_osnoise;
 650	struct ring_buffer_event *event;
 651	struct stack_entry *entry;
 652
 653	event = trace_buffer_lock_reserve(buffer, TRACE_STACK, sizeof(*entry) + size,
 654					  tracing_gen_ctx());
 655	if (!event)
 656		return;
 657
 658	entry = ring_buffer_event_data(event);
 659
 660	memcpy(&entry->caller, fstack->calls, size);
 661	entry->size = fstack->nr_entries;
 662
 663	if (!call_filter_check_discard(call, entry, buffer, event))
 664		trace_buffer_unlock_commit_nostack(buffer, event);
 665}
 666
 667/*
 668 * timerlat_dump_stack - dump a stack trace previously saved
 669 */
 670static void timerlat_dump_stack(u64 latency)
 671{
 672	struct osnoise_instance *inst;
 673	struct trace_buffer *buffer;
 674	struct trace_stack *fstack;
 675	unsigned int size;
 676
 677	/*
 678	 * trace only if latency > print_stack config, if enabled.
 679	 */
 680	if (!osnoise_data.print_stack || osnoise_data.print_stack > latency)
 681		return;
 682
 683	preempt_disable_notrace();
 684	fstack = this_cpu_ptr(&trace_stack);
 685	size = fstack->stack_size;
 686
 687	rcu_read_lock();
 688	list_for_each_entry_rcu(inst, &osnoise_instances, list) {
 689		buffer = inst->tr->array_buffer.buffer;
 690		__timerlat_dump_stack(buffer, fstack, size);
 691
 692	}
 693	rcu_read_unlock();
 694	preempt_enable_notrace();
 695}
 696#else /* CONFIG_STACKTRACE */
 697#define timerlat_dump_stack(u64 latency) do {} while (0)
 698#define timerlat_save_stack(a) do {} while (0)
 699#endif /* CONFIG_STACKTRACE */
 700#endif /* CONFIG_TIMERLAT_TRACER */
 701
 702/*
 703 * Macros to encapsulate the time capturing infrastructure.
 704 */
 705#define time_get()	trace_clock_local()
 706#define time_to_us(x)	div_u64(x, 1000)
 707#define time_sub(a, b)	((a) - (b))
 708
 709/*
 710 * cond_move_irq_delta_start - Forward the delta_start of a running IRQ
 711 *
 712 * If an IRQ is preempted by an NMI, its delta_start is pushed forward
 713 * to discount the NMI interference.
 714 *
 715 * See get_int_safe_duration().
 716 */
 717static inline void
 718cond_move_irq_delta_start(struct osnoise_variables *osn_var, u64 duration)
 719{
 720	if (osn_var->irq.delta_start)
 721		osn_var->irq.delta_start += duration;
 722}
 723
 724#ifndef CONFIG_PREEMPT_RT
 725/*
 726 * cond_move_softirq_delta_start - Forward the delta_start of a running softirq.
 727 *
 728 * If a softirq is preempted by an IRQ or NMI, its delta_start is pushed
 729 * forward to discount the interference.
 730 *
 731 * See get_int_safe_duration().
 732 */
 733static inline void
 734cond_move_softirq_delta_start(struct osnoise_variables *osn_var, u64 duration)
 735{
 736	if (osn_var->softirq.delta_start)
 737		osn_var->softirq.delta_start += duration;
 738}
 739#else /* CONFIG_PREEMPT_RT */
 740#define cond_move_softirq_delta_start(osn_var, duration) do {} while (0)
 741#endif
 742
 743/*
 744 * cond_move_thread_delta_start - Forward the delta_start of a running thread
 745 *
 746 * If a noisy thread is preempted by an softirq, IRQ or NMI, its delta_start
 747 * is pushed forward to discount the interference.
 748 *
 749 * See get_int_safe_duration().
 750 */
 751static inline void
 752cond_move_thread_delta_start(struct osnoise_variables *osn_var, u64 duration)
 753{
 754	if (osn_var->thread.delta_start)
 755		osn_var->thread.delta_start += duration;
 756}
 757
 758/*
 759 * get_int_safe_duration - Get the duration of a window
 760 *
 761 * The irq, softirq and thread varaibles need to have its duration without
 762 * the interference from higher priority interrupts. Instead of keeping a
 763 * variable to discount the interrupt interference from these variables, the
 764 * starting time of these variables are pushed forward with the interrupt's
 765 * duration. In this way, a single variable is used to:
 766 *
 767 *   - Know if a given window is being measured.
 768 *   - Account its duration.
 769 *   - Discount the interference.
 770 *
 771 * To avoid getting inconsistent values, e.g.,:
 772 *
 773 *	now = time_get()
 774 *		--->	interrupt!
 775 *			delta_start -= int duration;
 776 *		<---
 777 *	duration = now - delta_start;
 778 *
 779 *	result: negative duration if the variable duration before the
 780 *	interrupt was smaller than the interrupt execution.
 781 *
 782 * A counter of interrupts is used. If the counter increased, try
 783 * to capture an interference safe duration.
 784 */
 785static inline s64
 786get_int_safe_duration(struct osnoise_variables *osn_var, u64 *delta_start)
 787{
 788	u64 int_counter, now;
 789	s64 duration;
 790
 791	do {
 792		int_counter = local_read(&osn_var->int_counter);
 793		/* synchronize with interrupts */
 794		barrier();
 795
 796		now = time_get();
 797		duration = (now - *delta_start);
 798
 799		/* synchronize with interrupts */
 800		barrier();
 801	} while (int_counter != local_read(&osn_var->int_counter));
 802
 803	/*
 804	 * This is an evidence of race conditions that cause
 805	 * a value to be "discounted" too much.
 806	 */
 807	if (duration < 0)
 808		osnoise_taint("Negative duration!\n");
 809
 810	*delta_start = 0;
 811
 812	return duration;
 813}
 814
 815/*
 816 *
 817 * set_int_safe_time - Save the current time on *time, aware of interference
 818 *
 819 * Get the time, taking into consideration a possible interference from
 820 * higher priority interrupts.
 821 *
 822 * See get_int_safe_duration() for an explanation.
 823 */
 824static u64
 825set_int_safe_time(struct osnoise_variables *osn_var, u64 *time)
 826{
 827	u64 int_counter;
 828
 829	do {
 830		int_counter = local_read(&osn_var->int_counter);
 831		/* synchronize with interrupts */
 832		barrier();
 833
 834		*time = time_get();
 835
 836		/* synchronize with interrupts */
 837		barrier();
 838	} while (int_counter != local_read(&osn_var->int_counter));
 839
 840	return int_counter;
 841}
 842
 843#ifdef CONFIG_TIMERLAT_TRACER
 844/*
 845 * copy_int_safe_time - Copy *src into *desc aware of interference
 846 */
 847static u64
 848copy_int_safe_time(struct osnoise_variables *osn_var, u64 *dst, u64 *src)
 849{
 850	u64 int_counter;
 851
 852	do {
 853		int_counter = local_read(&osn_var->int_counter);
 854		/* synchronize with interrupts */
 855		barrier();
 856
 857		*dst = *src;
 858
 859		/* synchronize with interrupts */
 860		barrier();
 861	} while (int_counter != local_read(&osn_var->int_counter));
 862
 863	return int_counter;
 864}
 865#endif /* CONFIG_TIMERLAT_TRACER */
 866
 867/*
 868 * trace_osnoise_callback - NMI entry/exit callback
 869 *
 870 * This function is called at the entry and exit NMI code. The bool enter
 871 * distinguishes between either case. This function is used to note a NMI
 872 * occurrence, compute the noise caused by the NMI, and to remove the noise
 873 * it is potentially causing on other interference variables.
 874 */
 875void trace_osnoise_callback(bool enter)
 876{
 877	struct osnoise_variables *osn_var = this_cpu_osn_var();
 878	u64 duration;
 879
 880	if (!osn_var->sampling)
 881		return;
 882
 883	/*
 884	 * Currently trace_clock_local() calls sched_clock() and the
 885	 * generic version is not NMI safe.
 886	 */
 887	if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
 888		if (enter) {
 889			osn_var->nmi.delta_start = time_get();
 890			local_inc(&osn_var->int_counter);
 891		} else {
 892			duration = time_get() - osn_var->nmi.delta_start;
 893
 894			trace_nmi_noise(osn_var->nmi.delta_start, duration);
 895
 896			cond_move_irq_delta_start(osn_var, duration);
 897			cond_move_softirq_delta_start(osn_var, duration);
 898			cond_move_thread_delta_start(osn_var, duration);
 899		}
 900	}
 901
 902	if (enter)
 903		osn_var->nmi.count++;
 904}
 905
 906/*
 907 * osnoise_trace_irq_entry - Note the starting of an IRQ
 908 *
 909 * Save the starting time of an IRQ. As IRQs are non-preemptive to other IRQs,
 910 * it is safe to use a single variable (ons_var->irq) to save the statistics.
 911 * The arrival_time is used to report... the arrival time. The delta_start
 912 * is used to compute the duration at the IRQ exit handler. See
 913 * cond_move_irq_delta_start().
 914 */
 915void osnoise_trace_irq_entry(int id)
 916{
 917	struct osnoise_variables *osn_var = this_cpu_osn_var();
 918
 919	if (!osn_var->sampling)
 920		return;
 921	/*
 922	 * This value will be used in the report, but not to compute
 923	 * the execution time, so it is safe to get it unsafe.
 924	 */
 925	osn_var->irq.arrival_time = time_get();
 926	set_int_safe_time(osn_var, &osn_var->irq.delta_start);
 927	osn_var->irq.count++;
 928
 929	local_inc(&osn_var->int_counter);
 930}
 931
 932/*
 933 * osnoise_irq_exit - Note the end of an IRQ, sava data and trace
 934 *
 935 * Computes the duration of the IRQ noise, and trace it. Also discounts the
 936 * interference from other sources of noise could be currently being accounted.
 937 */
 938void osnoise_trace_irq_exit(int id, const char *desc)
 939{
 940	struct osnoise_variables *osn_var = this_cpu_osn_var();
 941	s64 duration;
 942
 943	if (!osn_var->sampling)
 944		return;
 945
 946	duration = get_int_safe_duration(osn_var, &osn_var->irq.delta_start);
 947	trace_irq_noise(id, desc, osn_var->irq.arrival_time, duration);
 948	osn_var->irq.arrival_time = 0;
 949	cond_move_softirq_delta_start(osn_var, duration);
 950	cond_move_thread_delta_start(osn_var, duration);
 951}
 952
 953/*
 954 * trace_irqentry_callback - Callback to the irq:irq_entry traceevent
 955 *
 956 * Used to note the starting of an IRQ occurece.
 957 */
 958static void trace_irqentry_callback(void *data, int irq,
 959				    struct irqaction *action)
 960{
 961	osnoise_trace_irq_entry(irq);
 962}
 963
 964/*
 965 * trace_irqexit_callback - Callback to the irq:irq_exit traceevent
 966 *
 967 * Used to note the end of an IRQ occurece.
 968 */
 969static void trace_irqexit_callback(void *data, int irq,
 970				   struct irqaction *action, int ret)
 971{
 972	osnoise_trace_irq_exit(irq, action->name);
 973}
 974
 975/*
 976 * arch specific register function.
 977 */
 978int __weak osnoise_arch_register(void)
 979{
 980	return 0;
 981}
 982
 983/*
 984 * arch specific unregister function.
 985 */
 986void __weak osnoise_arch_unregister(void)
 987{
 988	return;
 989}
 990
 991/*
 992 * hook_irq_events - Hook IRQ handling events
 993 *
 994 * This function hooks the IRQ related callbacks to the respective trace
 995 * events.
 996 */
 997static int hook_irq_events(void)
 998{
 999	int ret;
1000
1001	ret = register_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1002	if (ret)
1003		goto out_err;
1004
1005	ret = register_trace_irq_handler_exit(trace_irqexit_callback, NULL);
1006	if (ret)
1007		goto out_unregister_entry;
1008
1009	ret = osnoise_arch_register();
1010	if (ret)
1011		goto out_irq_exit;
1012
1013	return 0;
1014
1015out_irq_exit:
1016	unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
1017out_unregister_entry:
1018	unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1019out_err:
1020	return -EINVAL;
1021}
1022
1023/*
1024 * unhook_irq_events - Unhook IRQ handling events
1025 *
1026 * This function unhooks the IRQ related callbacks to the respective trace
1027 * events.
1028 */
1029static void unhook_irq_events(void)
1030{
1031	osnoise_arch_unregister();
1032	unregister_trace_irq_handler_exit(trace_irqexit_callback, NULL);
1033	unregister_trace_irq_handler_entry(trace_irqentry_callback, NULL);
1034}
1035
1036#ifndef CONFIG_PREEMPT_RT
1037/*
1038 * trace_softirq_entry_callback - Note the starting of a softirq
1039 *
1040 * Save the starting time of a softirq. As softirqs are non-preemptive to
1041 * other softirqs, it is safe to use a single variable (ons_var->softirq)
1042 * to save the statistics. The arrival_time is used to report... the
1043 * arrival time. The delta_start is used to compute the duration at the
1044 * softirq exit handler. See cond_move_softirq_delta_start().
1045 */
1046static void trace_softirq_entry_callback(void *data, unsigned int vec_nr)
1047{
1048	struct osnoise_variables *osn_var = this_cpu_osn_var();
1049
1050	if (!osn_var->sampling)
1051		return;
1052	/*
1053	 * This value will be used in the report, but not to compute
1054	 * the execution time, so it is safe to get it unsafe.
1055	 */
1056	osn_var->softirq.arrival_time = time_get();
1057	set_int_safe_time(osn_var, &osn_var->softirq.delta_start);
1058	osn_var->softirq.count++;
1059
1060	local_inc(&osn_var->int_counter);
1061}
1062
1063/*
1064 * trace_softirq_exit_callback - Note the end of an softirq
1065 *
1066 * Computes the duration of the softirq noise, and trace it. Also discounts the
1067 * interference from other sources of noise could be currently being accounted.
1068 */
1069static void trace_softirq_exit_callback(void *data, unsigned int vec_nr)
1070{
1071	struct osnoise_variables *osn_var = this_cpu_osn_var();
1072	s64 duration;
1073
1074	if (!osn_var->sampling)
1075		return;
1076
1077	if (unlikely(timerlat_enabled()))
1078		if (!timerlat_softirq_exit(osn_var))
1079			return;
1080
1081	duration = get_int_safe_duration(osn_var, &osn_var->softirq.delta_start);
1082	trace_softirq_noise(vec_nr, osn_var->softirq.arrival_time, duration);
1083	cond_move_thread_delta_start(osn_var, duration);
1084	osn_var->softirq.arrival_time = 0;
1085}
1086
1087/*
1088 * hook_softirq_events - Hook softirq handling events
1089 *
1090 * This function hooks the softirq related callbacks to the respective trace
1091 * events.
1092 */
1093static int hook_softirq_events(void)
1094{
1095	int ret;
1096
1097	ret = register_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1098	if (ret)
1099		goto out_err;
1100
1101	ret = register_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1102	if (ret)
1103		goto out_unreg_entry;
1104
1105	return 0;
1106
1107out_unreg_entry:
1108	unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1109out_err:
1110	return -EINVAL;
1111}
1112
1113/*
1114 * unhook_softirq_events - Unhook softirq handling events
1115 *
1116 * This function hooks the softirq related callbacks to the respective trace
1117 * events.
1118 */
1119static void unhook_softirq_events(void)
1120{
1121	unregister_trace_softirq_entry(trace_softirq_entry_callback, NULL);
1122	unregister_trace_softirq_exit(trace_softirq_exit_callback, NULL);
1123}
1124#else /* CONFIG_PREEMPT_RT */
1125/*
1126 * softirq are threads on the PREEMPT_RT mode.
1127 */
1128static int hook_softirq_events(void)
1129{
1130	return 0;
1131}
1132static void unhook_softirq_events(void)
1133{
1134}
1135#endif
1136
1137/*
1138 * thread_entry - Record the starting of a thread noise window
1139 *
1140 * It saves the context switch time for a noisy thread, and increments
1141 * the interference counters.
1142 */
1143static void
1144thread_entry(struct osnoise_variables *osn_var, struct task_struct *t)
1145{
1146	if (!osn_var->sampling)
1147		return;
1148	/*
1149	 * The arrival time will be used in the report, but not to compute
1150	 * the execution time, so it is safe to get it unsafe.
1151	 */
1152	osn_var->thread.arrival_time = time_get();
1153
1154	set_int_safe_time(osn_var, &osn_var->thread.delta_start);
1155
1156	osn_var->thread.count++;
1157	local_inc(&osn_var->int_counter);
1158}
1159
1160/*
1161 * thread_exit - Report the end of a thread noise window
1162 *
1163 * It computes the total noise from a thread, tracing if needed.
1164 */
1165static void
1166thread_exit(struct osnoise_variables *osn_var, struct task_struct *t)
1167{
1168	s64 duration;
1169
1170	if (!osn_var->sampling)
1171		return;
1172
1173	if (unlikely(timerlat_enabled()))
1174		if (!timerlat_thread_exit(osn_var))
1175			return;
1176
1177	duration = get_int_safe_duration(osn_var, &osn_var->thread.delta_start);
1178
1179	trace_thread_noise(t, osn_var->thread.arrival_time, duration);
1180
1181	osn_var->thread.arrival_time = 0;
1182}
1183
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1184/*
1185 * trace_sched_switch - sched:sched_switch trace event handler
1186 *
1187 * This function is hooked to the sched:sched_switch trace event, and it is
1188 * used to record the beginning and to report the end of a thread noise window.
1189 */
1190static void
1191trace_sched_switch_callback(void *data, bool preempt,
1192			    struct task_struct *p,
1193			    struct task_struct *n,
1194			    unsigned int prev_state)
1195{
1196	struct osnoise_variables *osn_var = this_cpu_osn_var();
1197	int workload = test_bit(OSN_WORKLOAD, &osnoise_options);
1198
1199	if ((p->pid != osn_var->pid) || !workload)
1200		thread_exit(osn_var, p);
1201
1202	if ((n->pid != osn_var->pid) || !workload)
1203		thread_entry(osn_var, n);
1204}
1205
1206/*
1207 * hook_thread_events - Hook the insturmentation for thread noise
1208 *
1209 * Hook the osnoise tracer callbacks to handle the noise from other
1210 * threads on the necessary kernel events.
1211 */
1212static int hook_thread_events(void)
1213{
1214	int ret;
1215
1216	ret = register_trace_sched_switch(trace_sched_switch_callback, NULL);
1217	if (ret)
1218		return -EINVAL;
1219
 
 
 
 
1220	return 0;
 
 
 
 
1221}
1222
1223/*
1224 * unhook_thread_events - *nhook the insturmentation for thread noise
1225 *
1226 * Unook the osnoise tracer callbacks to handle the noise from other
1227 * threads on the necessary kernel events.
1228 */
1229static void unhook_thread_events(void)
1230{
1231	unregister_trace_sched_switch(trace_sched_switch_callback, NULL);
 
1232}
1233
1234/*
1235 * save_osn_sample_stats - Save the osnoise_sample statistics
1236 *
1237 * Save the osnoise_sample statistics before the sampling phase. These
1238 * values will be used later to compute the diff betwneen the statistics
1239 * before and after the osnoise sampling.
1240 */
1241static void
1242save_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1243{
1244	s->nmi_count = osn_var->nmi.count;
1245	s->irq_count = osn_var->irq.count;
1246	s->softirq_count = osn_var->softirq.count;
1247	s->thread_count = osn_var->thread.count;
1248}
1249
1250/*
1251 * diff_osn_sample_stats - Compute the osnoise_sample statistics
1252 *
1253 * After a sample period, compute the difference on the osnoise_sample
1254 * statistics. The struct osnoise_sample *s contains the statistics saved via
1255 * save_osn_sample_stats() before the osnoise sampling.
1256 */
1257static void
1258diff_osn_sample_stats(struct osnoise_variables *osn_var, struct osnoise_sample *s)
1259{
1260	s->nmi_count = osn_var->nmi.count - s->nmi_count;
1261	s->irq_count = osn_var->irq.count - s->irq_count;
1262	s->softirq_count = osn_var->softirq.count - s->softirq_count;
1263	s->thread_count = osn_var->thread.count - s->thread_count;
1264}
1265
1266/*
1267 * osnoise_stop_tracing - Stop tracing and the tracer.
1268 */
1269static __always_inline void osnoise_stop_tracing(void)
1270{
1271	struct osnoise_instance *inst;
1272	struct trace_array *tr;
1273
1274	rcu_read_lock();
1275	list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1276		tr = inst->tr;
1277		trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_,
1278				"stop tracing hit on cpu %d\n", smp_processor_id());
1279
1280		if (test_bit(OSN_PANIC_ON_STOP, &osnoise_options))
1281			panic("tracer hit stop condition on CPU %d\n", smp_processor_id());
1282
1283		tracer_tracing_off(tr);
1284	}
1285	rcu_read_unlock();
1286}
1287
1288/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1289 * notify_new_max_latency - Notify a new max latency via fsnotify interface.
1290 */
1291static void notify_new_max_latency(u64 latency)
1292{
1293	struct osnoise_instance *inst;
1294	struct trace_array *tr;
1295
1296	rcu_read_lock();
1297	list_for_each_entry_rcu(inst, &osnoise_instances, list) {
1298		tr = inst->tr;
1299		if (tr->max_latency < latency) {
1300			tr->max_latency = latency;
1301			latency_fsnotify(tr);
1302		}
1303	}
1304	rcu_read_unlock();
1305}
1306
1307/*
1308 * run_osnoise - Sample the time and look for osnoise
1309 *
1310 * Used to capture the time, looking for potential osnoise latency repeatedly.
1311 * Different from hwlat_detector, it is called with preemption and interrupts
1312 * enabled. This allows irqs, softirqs and threads to run, interfering on the
1313 * osnoise sampling thread, as they would do with a regular thread.
1314 */
1315static int run_osnoise(void)
1316{
1317	bool disable_irq = test_bit(OSN_IRQ_DISABLE, &osnoise_options);
1318	struct osnoise_variables *osn_var = this_cpu_osn_var();
1319	u64 start, sample, last_sample;
1320	u64 last_int_count, int_count;
1321	s64 noise = 0, max_noise = 0;
1322	s64 total, last_total = 0;
1323	struct osnoise_sample s;
1324	bool disable_preemption;
1325	unsigned int threshold;
1326	u64 runtime, stop_in;
1327	u64 sum_noise = 0;
1328	int hw_count = 0;
1329	int ret = -1;
1330
1331	/*
1332	 * Disabling preemption is only required if IRQs are enabled,
1333	 * and the options is set on.
1334	 */
1335	disable_preemption = !disable_irq && test_bit(OSN_PREEMPT_DISABLE, &osnoise_options);
1336
1337	/*
1338	 * Considers the current thread as the workload.
1339	 */
1340	osn_var->pid = current->pid;
1341
1342	/*
1343	 * Save the current stats for the diff
1344	 */
1345	save_osn_sample_stats(osn_var, &s);
1346
1347	/*
1348	 * if threshold is 0, use the default value of 5 us.
1349	 */
1350	threshold = tracing_thresh ? : 5000;
1351
1352	/*
1353	 * Apply PREEMPT and IRQ disabled options.
1354	 */
1355	if (disable_irq)
1356		local_irq_disable();
1357
1358	if (disable_preemption)
1359		preempt_disable();
1360
1361	/*
1362	 * Make sure NMIs see sampling first
1363	 */
1364	osn_var->sampling = true;
1365	barrier();
1366
1367	/*
1368	 * Transform the *_us config to nanoseconds to avoid the
1369	 * division on the main loop.
1370	 */
1371	runtime = osnoise_data.sample_runtime * NSEC_PER_USEC;
1372	stop_in = osnoise_data.stop_tracing * NSEC_PER_USEC;
1373
1374	/*
1375	 * Start timestemp
1376	 */
1377	start = time_get();
1378
1379	/*
1380	 * "previous" loop.
1381	 */
1382	last_int_count = set_int_safe_time(osn_var, &last_sample);
1383
1384	do {
1385		/*
1386		 * Get sample!
1387		 */
1388		int_count = set_int_safe_time(osn_var, &sample);
1389
1390		noise = time_sub(sample, last_sample);
1391
1392		/*
1393		 * This shouldn't happen.
1394		 */
1395		if (noise < 0) {
1396			osnoise_taint("negative noise!");
1397			goto out;
1398		}
1399
1400		/*
1401		 * Sample runtime.
1402		 */
1403		total = time_sub(sample, start);
1404
1405		/*
1406		 * Check for possible overflows.
1407		 */
1408		if (total < last_total) {
1409			osnoise_taint("total overflow!");
1410			break;
1411		}
1412
1413		last_total = total;
1414
1415		if (noise >= threshold) {
1416			int interference = int_count - last_int_count;
1417
1418			if (noise > max_noise)
1419				max_noise = noise;
1420
1421			if (!interference)
1422				hw_count++;
1423
1424			sum_noise += noise;
1425
1426			trace_sample_threshold(last_sample, noise, interference);
1427
1428			if (osnoise_data.stop_tracing)
1429				if (noise > stop_in)
1430					osnoise_stop_tracing();
1431		}
1432
1433		/*
1434		 * In some cases, notably when running on a nohz_full CPU with
1435		 * a stopped tick PREEMPT_RCU has no way to account for QSs.
1436		 * This will eventually cause unwarranted noise as PREEMPT_RCU
1437		 * will force preemption as the means of ending the current
1438		 * grace period. We avoid this problem by calling
1439		 * rcu_momentary_dyntick_idle(), which performs a zero duration
1440		 * EQS allowing PREEMPT_RCU to end the current grace period.
1441		 * This call shouldn't be wrapped inside an RCU critical
1442		 * section.
1443		 *
1444		 * Note that in non PREEMPT_RCU kernels QSs are handled through
1445		 * cond_resched()
1446		 */
1447		if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
1448			if (!disable_irq)
1449				local_irq_disable();
1450
1451			rcu_momentary_dyntick_idle();
1452
1453			if (!disable_irq)
1454				local_irq_enable();
1455		}
1456
1457		/*
1458		 * For the non-preemptive kernel config: let threads runs, if
1459		 * they so wish, unless set not do to so.
1460		 */
1461		if (!disable_irq && !disable_preemption)
1462			cond_resched();
1463
1464		last_sample = sample;
1465		last_int_count = int_count;
1466
1467	} while (total < runtime && !kthread_should_stop());
1468
1469	/*
1470	 * Finish the above in the view for interrupts.
1471	 */
1472	barrier();
1473
1474	osn_var->sampling = false;
1475
1476	/*
1477	 * Make sure sampling data is no longer updated.
1478	 */
1479	barrier();
1480
1481	/*
1482	 * Return to the preemptive state.
1483	 */
1484	if (disable_preemption)
1485		preempt_enable();
1486
1487	if (disable_irq)
1488		local_irq_enable();
1489
1490	/*
1491	 * Save noise info.
1492	 */
1493	s.noise = time_to_us(sum_noise);
1494	s.runtime = time_to_us(total);
1495	s.max_sample = time_to_us(max_noise);
1496	s.hw_count = hw_count;
1497
1498	/* Save interference stats info */
1499	diff_osn_sample_stats(osn_var, &s);
1500
1501	trace_osnoise_sample(&s);
1502
1503	notify_new_max_latency(max_noise);
1504
1505	if (osnoise_data.stop_tracing_total)
1506		if (s.noise > osnoise_data.stop_tracing_total)
1507			osnoise_stop_tracing();
1508
1509	return 0;
1510out:
1511	return ret;
1512}
1513
1514static struct cpumask osnoise_cpumask;
1515static struct cpumask save_cpumask;
 
1516
1517/*
1518 * osnoise_sleep - sleep until the next period
1519 */
1520static void osnoise_sleep(void)
1521{
1522	u64 interval;
1523	ktime_t wake_time;
1524
1525	mutex_lock(&interface_lock);
1526	interval = osnoise_data.sample_period - osnoise_data.sample_runtime;
 
 
 
1527	mutex_unlock(&interface_lock);
1528
1529	/*
1530	 * differently from hwlat_detector, the osnoise tracer can run
1531	 * without a pause because preemption is on.
1532	 */
1533	if (!interval) {
1534		/* Let synchronize_rcu_tasks() make progress */
1535		cond_resched_tasks_rcu_qs();
1536		return;
1537	}
1538
1539	wake_time = ktime_add_us(ktime_get(), interval);
1540	__set_current_state(TASK_INTERRUPTIBLE);
1541
1542	while (schedule_hrtimeout_range(&wake_time, 0, HRTIMER_MODE_ABS)) {
1543		if (kthread_should_stop())
1544			break;
1545	}
1546}
1547
1548/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1549 * osnoise_main - The osnoise detection kernel thread
1550 *
1551 * Calls run_osnoise() function to measure the osnoise for the configured runtime,
1552 * every period.
1553 */
1554static int osnoise_main(void *data)
1555{
 
 
 
 
 
 
 
 
 
 
 
 
 
1556
1557	while (!kthread_should_stop()) {
 
 
 
 
 
 
 
 
 
1558		run_osnoise();
1559		osnoise_sleep();
1560	}
1561
 
1562	return 0;
1563}
1564
1565#ifdef CONFIG_TIMERLAT_TRACER
1566/*
1567 * timerlat_irq - hrtimer handler for timerlat.
1568 */
1569static enum hrtimer_restart timerlat_irq(struct hrtimer *timer)
1570{
1571	struct osnoise_variables *osn_var = this_cpu_osn_var();
1572	struct timerlat_variables *tlat;
1573	struct timerlat_sample s;
1574	u64 now;
1575	u64 diff;
1576
1577	/*
1578	 * I am not sure if the timer was armed for this CPU. So, get
1579	 * the timerlat struct from the timer itself, not from this
1580	 * CPU.
1581	 */
1582	tlat = container_of(timer, struct timerlat_variables, timer);
1583
1584	now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1585
1586	/*
1587	 * Enable the osnoise: events for thread an softirq.
1588	 */
1589	tlat->tracing_thread = true;
1590
1591	osn_var->thread.arrival_time = time_get();
1592
1593	/*
1594	 * A hardirq is running: the timer IRQ. It is for sure preempting
1595	 * a thread, and potentially preempting a softirq.
1596	 *
1597	 * At this point, it is not interesting to know the duration of the
1598	 * preempted thread (and maybe softirq), but how much time they will
1599	 * delay the beginning of the execution of the timer thread.
1600	 *
1601	 * To get the correct (net) delay added by the softirq, its delta_start
1602	 * is set as the IRQ one. In this way, at the return of the IRQ, the delta
1603	 * start of the sofitrq will be zeroed, accounting then only the time
1604	 * after that.
1605	 *
1606	 * The thread follows the same principle. However, if a softirq is
1607	 * running, the thread needs to receive the softirq delta_start. The
1608	 * reason being is that the softirq will be the last to be unfolded,
1609	 * resseting the thread delay to zero.
1610	 *
1611	 * The PREEMPT_RT is a special case, though. As softirqs run as threads
1612	 * on RT, moving the thread is enough.
1613	 */
1614	if (!IS_ENABLED(CONFIG_PREEMPT_RT) && osn_var->softirq.delta_start) {
1615		copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1616				   &osn_var->softirq.delta_start);
1617
1618		copy_int_safe_time(osn_var, &osn_var->softirq.delta_start,
1619				    &osn_var->irq.delta_start);
1620	} else {
1621		copy_int_safe_time(osn_var, &osn_var->thread.delta_start,
1622				    &osn_var->irq.delta_start);
1623	}
1624
1625	/*
1626	 * Compute the current time with the expected time.
1627	 */
1628	diff = now - tlat->abs_period;
1629
1630	tlat->count++;
1631	s.seqnum = tlat->count;
1632	s.timer_latency = diff;
1633	s.context = IRQ_CONTEXT;
1634
1635	trace_timerlat_sample(&s);
1636
1637	if (osnoise_data.stop_tracing) {
1638		if (time_to_us(diff) >= osnoise_data.stop_tracing) {
1639
1640			/*
1641			 * At this point, if stop_tracing is set and <= print_stack,
1642			 * print_stack is set and would be printed in the thread handler.
1643			 *
1644			 * Thus, print the stack trace as it is helpful to define the
1645			 * root cause of an IRQ latency.
1646			 */
1647			if (osnoise_data.stop_tracing <= osnoise_data.print_stack) {
1648				timerlat_save_stack(0);
1649				timerlat_dump_stack(time_to_us(diff));
1650			}
1651
1652			osnoise_stop_tracing();
1653			notify_new_max_latency(diff);
1654
 
 
1655			return HRTIMER_NORESTART;
1656		}
1657	}
1658
1659	wake_up_process(tlat->kthread);
1660
1661	if (osnoise_data.print_stack)
1662		timerlat_save_stack(0);
1663
1664	return HRTIMER_NORESTART;
1665}
1666
1667/*
1668 * wait_next_period - Wait for the next period for timerlat
1669 */
1670static int wait_next_period(struct timerlat_variables *tlat)
1671{
1672	ktime_t next_abs_period, now;
1673	u64 rel_period = osnoise_data.timerlat_period * 1000;
1674
1675	now = hrtimer_cb_get_time(&tlat->timer);
1676	next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1677
1678	/*
1679	 * Save the next abs_period.
1680	 */
1681	tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1682
1683	/*
1684	 * If the new abs_period is in the past, skip the activation.
1685	 */
1686	while (ktime_compare(now, next_abs_period) > 0) {
1687		next_abs_period = ns_to_ktime(tlat->abs_period + rel_period);
1688		tlat->abs_period = (u64) ktime_to_ns(next_abs_period);
1689	}
1690
1691	set_current_state(TASK_INTERRUPTIBLE);
1692
1693	hrtimer_start(&tlat->timer, next_abs_period, HRTIMER_MODE_ABS_PINNED_HARD);
1694	schedule();
1695	return 1;
1696}
1697
1698/*
1699 * timerlat_main- Timerlat main
1700 */
1701static int timerlat_main(void *data)
1702{
1703	struct osnoise_variables *osn_var = this_cpu_osn_var();
1704	struct timerlat_variables *tlat = this_cpu_tmr_var();
1705	struct timerlat_sample s;
1706	struct sched_param sp;
 
1707	u64 now, diff;
1708
1709	/*
1710	 * Make the thread RT, that is how cyclictest is usually used.
1711	 */
1712	sp.sched_priority = DEFAULT_TIMERLAT_PRIO;
1713	sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1714
 
 
 
 
 
 
 
 
 
 
 
 
1715	tlat->count = 0;
1716	tlat->tracing_thread = false;
1717
1718	hrtimer_init(&tlat->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1719	tlat->timer.function = timerlat_irq;
1720	tlat->kthread = current;
1721	osn_var->pid = current->pid;
1722	/*
1723	 * Anotate the arrival time.
1724	 */
1725	tlat->abs_period = hrtimer_cb_get_time(&tlat->timer);
1726
1727	wait_next_period(tlat);
1728
1729	osn_var->sampling = 1;
1730
1731	while (!kthread_should_stop()) {
 
1732		now = ktime_to_ns(hrtimer_cb_get_time(&tlat->timer));
1733		diff = now - tlat->abs_period;
1734
1735		s.seqnum = tlat->count;
1736		s.timer_latency = diff;
1737		s.context = THREAD_CONTEXT;
1738
1739		trace_timerlat_sample(&s);
1740
 
 
1741		timerlat_dump_stack(time_to_us(diff));
1742
1743		tlat->tracing_thread = false;
1744		if (osnoise_data.stop_tracing_total)
1745			if (time_to_us(diff) >= osnoise_data.stop_tracing_total)
1746				osnoise_stop_tracing();
1747
 
 
 
1748		wait_next_period(tlat);
1749	}
1750
1751	hrtimer_cancel(&tlat->timer);
 
1752	return 0;
1753}
1754#else /* CONFIG_TIMERLAT_TRACER */
1755static int timerlat_main(void *data)
1756{
1757	return 0;
1758}
1759#endif /* CONFIG_TIMERLAT_TRACER */
1760
1761/*
1762 * stop_kthread - stop a workload thread
1763 */
1764static void stop_kthread(unsigned int cpu)
1765{
1766	struct task_struct *kthread;
1767
1768	kthread = per_cpu(per_cpu_osnoise_var, cpu).kthread;
1769	if (kthread) {
1770		kthread_stop(kthread);
1771		per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
 
 
 
 
 
 
 
 
 
 
1772	} else {
 
1773		if (!test_bit(OSN_WORKLOAD, &osnoise_options)) {
 
 
 
1774			per_cpu(per_cpu_osnoise_var, cpu).sampling = false;
1775			barrier();
1776			return;
1777		}
1778	}
1779}
1780
1781/*
1782 * stop_per_cpu_kthread - Stop per-cpu threads
1783 *
1784 * Stop the osnoise sampling htread. Use this on unload and at system
1785 * shutdown.
1786 */
1787static void stop_per_cpu_kthreads(void)
1788{
1789	int cpu;
1790
1791	cpus_read_lock();
1792
1793	for_each_online_cpu(cpu)
1794		stop_kthread(cpu);
1795
1796	cpus_read_unlock();
1797}
1798
1799/*
1800 * start_kthread - Start a workload tread
1801 */
1802static int start_kthread(unsigned int cpu)
1803{
1804	struct task_struct *kthread;
1805	void *main = osnoise_main;
1806	char comm[24];
1807
 
 
 
 
1808	if (timerlat_enabled()) {
1809		snprintf(comm, 24, "timerlat/%d", cpu);
1810		main = timerlat_main;
1811	} else {
1812		/* if no workload, just return */
1813		if (!test_bit(OSN_WORKLOAD, &osnoise_options)) {
1814			per_cpu(per_cpu_osnoise_var, cpu).sampling = true;
1815			barrier();
1816			return 0;
1817		}
1818
1819		snprintf(comm, 24, "osnoise/%d", cpu);
1820	}
1821
1822	kthread = kthread_run_on_cpu(main, NULL, cpu, comm);
1823
1824	if (IS_ERR(kthread)) {
1825		pr_err(BANNER "could not start sampling thread\n");
1826		stop_per_cpu_kthreads();
1827		return -ENOMEM;
1828	}
1829
1830	per_cpu(per_cpu_osnoise_var, cpu).kthread = kthread;
 
1831
1832	return 0;
1833}
1834
1835/*
1836 * start_per_cpu_kthread - Kick off per-cpu osnoise sampling kthreads
1837 *
1838 * This starts the kernel thread that will look for osnoise on many
1839 * cpus.
1840 */
1841static int start_per_cpu_kthreads(void)
1842{
1843	struct cpumask *current_mask = &save_cpumask;
1844	int retval = 0;
1845	int cpu;
1846
 
 
 
 
 
1847	cpus_read_lock();
1848	/*
1849	 * Run only on online CPUs in which osnoise is allowed to run.
1850	 */
1851	cpumask_and(current_mask, cpu_online_mask, &osnoise_cpumask);
1852
1853	for_each_possible_cpu(cpu)
1854		per_cpu(per_cpu_osnoise_var, cpu).kthread = NULL;
 
 
 
 
 
 
 
1855
1856	for_each_cpu(cpu, current_mask) {
1857		retval = start_kthread(cpu);
1858		if (retval) {
1859			cpus_read_unlock();
1860			stop_per_cpu_kthreads();
1861			return retval;
1862		}
1863	}
1864
1865	cpus_read_unlock();
1866
1867	return retval;
1868}
1869
1870#ifdef CONFIG_HOTPLUG_CPU
1871static void osnoise_hotplug_workfn(struct work_struct *dummy)
1872{
1873	unsigned int cpu = smp_processor_id();
1874
1875	mutex_lock(&trace_types_lock);
1876
1877	if (!osnoise_has_registered_instances())
1878		goto out_unlock_trace;
1879
1880	mutex_lock(&interface_lock);
1881	cpus_read_lock();
1882
 
 
1883	if (!cpumask_test_cpu(cpu, &osnoise_cpumask))
1884		goto out_unlock;
1885
1886	start_kthread(cpu);
1887
1888out_unlock:
1889	cpus_read_unlock();
1890	mutex_unlock(&interface_lock);
1891out_unlock_trace:
1892	mutex_unlock(&trace_types_lock);
1893}
1894
1895static DECLARE_WORK(osnoise_hotplug_work, osnoise_hotplug_workfn);
1896
1897/*
1898 * osnoise_cpu_init - CPU hotplug online callback function
1899 */
1900static int osnoise_cpu_init(unsigned int cpu)
1901{
1902	schedule_work_on(cpu, &osnoise_hotplug_work);
1903	return 0;
1904}
1905
1906/*
1907 * osnoise_cpu_die - CPU hotplug offline callback function
1908 */
1909static int osnoise_cpu_die(unsigned int cpu)
1910{
1911	stop_kthread(cpu);
1912	return 0;
1913}
1914
1915static void osnoise_init_hotplug_support(void)
1916{
1917	int ret;
1918
1919	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/osnoise:online",
1920				osnoise_cpu_init, osnoise_cpu_die);
1921	if (ret < 0)
1922		pr_warn(BANNER "Error to init cpu hotplug support\n");
1923
1924	return;
1925}
1926#else /* CONFIG_HOTPLUG_CPU */
1927static void osnoise_init_hotplug_support(void)
1928{
1929	return;
1930}
1931#endif /* CONFIG_HOTPLUG_CPU */
1932
1933/*
1934 * seq file functions for the osnoise/options file.
1935 */
1936static void *s_options_start(struct seq_file *s, loff_t *pos)
1937{
1938	int option = *pos;
1939
1940	mutex_lock(&interface_lock);
1941
1942	if (option >= OSN_MAX)
1943		return NULL;
1944
1945	return pos;
1946}
1947
1948static void *s_options_next(struct seq_file *s, void *v, loff_t *pos)
1949{
1950	int option = ++(*pos);
1951
1952	if (option >= OSN_MAX)
1953		return NULL;
1954
1955	return pos;
1956}
1957
1958static int s_options_show(struct seq_file *s, void *v)
1959{
1960	loff_t *pos = v;
1961	int option = *pos;
1962
1963	if (option == OSN_DEFAULTS) {
1964		if (osnoise_options == OSN_DEFAULT_OPTIONS)
1965			seq_printf(s, "%s", osnoise_options_str[option]);
1966		else
1967			seq_printf(s, "NO_%s", osnoise_options_str[option]);
1968		goto out;
1969	}
1970
1971	if (test_bit(option, &osnoise_options))
1972		seq_printf(s, "%s", osnoise_options_str[option]);
1973	else
1974		seq_printf(s, "NO_%s", osnoise_options_str[option]);
1975
1976out:
1977	if (option != OSN_MAX)
1978		seq_puts(s, " ");
1979
1980	return 0;
1981}
1982
1983static void s_options_stop(struct seq_file *s, void *v)
1984{
1985	seq_puts(s, "\n");
1986	mutex_unlock(&interface_lock);
1987}
1988
1989static const struct seq_operations osnoise_options_seq_ops = {
1990	.start		= s_options_start,
1991	.next		= s_options_next,
1992	.show		= s_options_show,
1993	.stop		= s_options_stop
1994};
1995
1996static int osnoise_options_open(struct inode *inode, struct file *file)
1997{
1998	return seq_open(file, &osnoise_options_seq_ops);
1999};
2000
2001/**
2002 * osnoise_options_write - Write function for "options" entry
2003 * @filp: The active open file structure
2004 * @ubuf: The user buffer that contains the value to write
2005 * @cnt: The maximum number of bytes to write to "file"
2006 * @ppos: The current position in @file
2007 *
2008 * Writing the option name sets the option, writing the "NO_"
2009 * prefix in front of the option name disables it.
2010 *
2011 * Writing "DEFAULTS" resets the option values to the default ones.
2012 */
2013static ssize_t osnoise_options_write(struct file *filp, const char __user *ubuf,
2014				     size_t cnt, loff_t *ppos)
2015{
2016	int running, option, enable, retval;
2017	char buf[256], *option_str;
2018
2019	if (cnt >= 256)
2020		return -EINVAL;
2021
2022	if (copy_from_user(buf, ubuf, cnt))
2023		return -EFAULT;
2024
2025	buf[cnt] = 0;
2026
2027	if (strncmp(buf, "NO_", 3)) {
2028		option_str = strstrip(buf);
2029		enable = true;
2030	} else {
2031		option_str = strstrip(&buf[3]);
2032		enable = false;
2033	}
2034
2035	option = match_string(osnoise_options_str, OSN_MAX, option_str);
2036	if (option < 0)
2037		return -EINVAL;
2038
2039	/*
2040	 * trace_types_lock is taken to avoid concurrency on start/stop.
2041	 */
2042	mutex_lock(&trace_types_lock);
2043	running = osnoise_has_registered_instances();
2044	if (running)
2045		stop_per_cpu_kthreads();
2046
2047	mutex_lock(&interface_lock);
2048	/*
2049	 * avoid CPU hotplug operations that might read options.
2050	 */
2051	cpus_read_lock();
2052
2053	retval = cnt;
2054
2055	if (enable) {
2056		if (option == OSN_DEFAULTS)
2057			osnoise_options = OSN_DEFAULT_OPTIONS;
2058		else
2059			set_bit(option, &osnoise_options);
2060	} else {
2061		if (option == OSN_DEFAULTS)
2062			retval = -EINVAL;
2063		else
2064			clear_bit(option, &osnoise_options);
2065	}
2066
2067	cpus_read_unlock();
2068	mutex_unlock(&interface_lock);
2069
2070	if (running)
2071		start_per_cpu_kthreads();
2072	mutex_unlock(&trace_types_lock);
2073
2074	return retval;
2075}
2076
2077/*
2078 * osnoise_cpus_read - Read function for reading the "cpus" file
2079 * @filp: The active open file structure
2080 * @ubuf: The userspace provided buffer to read value into
2081 * @cnt: The maximum number of bytes to read
2082 * @ppos: The current "file" position
2083 *
2084 * Prints the "cpus" output into the user-provided buffer.
2085 */
2086static ssize_t
2087osnoise_cpus_read(struct file *filp, char __user *ubuf, size_t count,
2088		  loff_t *ppos)
2089{
2090	char *mask_str;
2091	int len;
2092
2093	mutex_lock(&interface_lock);
2094
2095	len = snprintf(NULL, 0, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask)) + 1;
2096	mask_str = kmalloc(len, GFP_KERNEL);
2097	if (!mask_str) {
2098		count = -ENOMEM;
2099		goto out_unlock;
2100	}
2101
2102	len = snprintf(mask_str, len, "%*pbl\n", cpumask_pr_args(&osnoise_cpumask));
2103	if (len >= count) {
2104		count = -EINVAL;
2105		goto out_free;
2106	}
2107
2108	count = simple_read_from_buffer(ubuf, count, ppos, mask_str, len);
2109
2110out_free:
2111	kfree(mask_str);
2112out_unlock:
2113	mutex_unlock(&interface_lock);
2114
2115	return count;
2116}
2117
2118/*
2119 * osnoise_cpus_write - Write function for "cpus" entry
2120 * @filp: The active open file structure
2121 * @ubuf: The user buffer that contains the value to write
2122 * @cnt: The maximum number of bytes to write to "file"
2123 * @ppos: The current position in @file
2124 *
2125 * This function provides a write implementation for the "cpus"
2126 * interface to the osnoise trace. By default, it lists all  CPUs,
2127 * in this way, allowing osnoise threads to run on any online CPU
2128 * of the system. It serves to restrict the execution of osnoise to the
2129 * set of CPUs writing via this interface. Why not use "tracing_cpumask"?
2130 * Because the user might be interested in tracing what is running on
2131 * other CPUs. For instance, one might run osnoise in one HT CPU
2132 * while observing what is running on the sibling HT CPU.
2133 */
2134static ssize_t
2135osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
2136		   loff_t *ppos)
2137{
2138	cpumask_var_t osnoise_cpumask_new;
2139	int running, err;
2140	char buf[256];
2141
2142	if (count >= 256)
2143		return -EINVAL;
2144
2145	if (copy_from_user(buf, ubuf, count))
2146		return -EFAULT;
2147
2148	if (!zalloc_cpumask_var(&osnoise_cpumask_new, GFP_KERNEL))
2149		return -ENOMEM;
2150
2151	err = cpulist_parse(buf, osnoise_cpumask_new);
2152	if (err)
2153		goto err_free;
2154
2155	/*
2156	 * trace_types_lock is taken to avoid concurrency on start/stop.
2157	 */
2158	mutex_lock(&trace_types_lock);
2159	running = osnoise_has_registered_instances();
2160	if (running)
2161		stop_per_cpu_kthreads();
2162
2163	mutex_lock(&interface_lock);
2164	/*
2165	 * osnoise_cpumask is read by CPU hotplug operations.
2166	 */
2167	cpus_read_lock();
2168
2169	cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
2170
2171	cpus_read_unlock();
2172	mutex_unlock(&interface_lock);
2173
2174	if (running)
2175		start_per_cpu_kthreads();
2176	mutex_unlock(&trace_types_lock);
2177
2178	free_cpumask_var(osnoise_cpumask_new);
2179	return count;
2180
2181err_free:
2182	free_cpumask_var(osnoise_cpumask_new);
2183
2184	return err;
2185}
2186
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2187/*
2188 * osnoise/runtime_us: cannot be greater than the period.
2189 */
2190static struct trace_min_max_param osnoise_runtime = {
2191	.lock	= &interface_lock,
2192	.val	= &osnoise_data.sample_runtime,
2193	.max	= &osnoise_data.sample_period,
2194	.min	= NULL,
2195};
2196
2197/*
2198 * osnoise/period_us: cannot be smaller than the runtime.
2199 */
2200static struct trace_min_max_param osnoise_period = {
2201	.lock	= &interface_lock,
2202	.val	= &osnoise_data.sample_period,
2203	.max	= NULL,
2204	.min	= &osnoise_data.sample_runtime,
2205};
2206
2207/*
2208 * osnoise/stop_tracing_us: no limit.
2209 */
2210static struct trace_min_max_param osnoise_stop_tracing_in = {
2211	.lock	= &interface_lock,
2212	.val	= &osnoise_data.stop_tracing,
2213	.max	= NULL,
2214	.min	= NULL,
2215};
2216
2217/*
2218 * osnoise/stop_tracing_total_us: no limit.
2219 */
2220static struct trace_min_max_param osnoise_stop_tracing_total = {
2221	.lock	= &interface_lock,
2222	.val	= &osnoise_data.stop_tracing_total,
2223	.max	= NULL,
2224	.min	= NULL,
2225};
2226
2227#ifdef CONFIG_TIMERLAT_TRACER
2228/*
2229 * osnoise/print_stack: print the stacktrace of the IRQ handler if the total
2230 * latency is higher than val.
2231 */
2232static struct trace_min_max_param osnoise_print_stack = {
2233	.lock	= &interface_lock,
2234	.val	= &osnoise_data.print_stack,
2235	.max	= NULL,
2236	.min	= NULL,
2237};
2238
2239/*
2240 * osnoise/timerlat_period: min 100 us, max 1 s
2241 */
2242u64 timerlat_min_period = 100;
2243u64 timerlat_max_period = 1000000;
2244static struct trace_min_max_param timerlat_period = {
2245	.lock	= &interface_lock,
2246	.val	= &osnoise_data.timerlat_period,
2247	.max	= &timerlat_max_period,
2248	.min	= &timerlat_min_period,
2249};
 
 
 
 
 
 
 
2250#endif
2251
2252static const struct file_operations cpus_fops = {
2253	.open		= tracing_open_generic,
2254	.read		= osnoise_cpus_read,
2255	.write		= osnoise_cpus_write,
2256	.llseek		= generic_file_llseek,
2257};
2258
2259static const struct file_operations osnoise_options_fops = {
2260	.open		= osnoise_options_open,
2261	.read		= seq_read,
2262	.llseek		= seq_lseek,
2263	.release	= seq_release,
2264	.write		= osnoise_options_write
2265};
2266
2267#ifdef CONFIG_TIMERLAT_TRACER
2268#ifdef CONFIG_STACKTRACE
2269static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2270{
2271	struct dentry *tmp;
2272
2273	tmp = tracefs_create_file("print_stack", TRACE_MODE_WRITE, top_dir,
2274				  &osnoise_print_stack, &trace_min_max_fops);
2275	if (!tmp)
2276		return -ENOMEM;
2277
2278	return 0;
2279}
2280#else /* CONFIG_STACKTRACE */
2281static int init_timerlat_stack_tracefs(struct dentry *top_dir)
2282{
2283	return 0;
2284}
2285#endif /* CONFIG_STACKTRACE */
2286
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2287/*
2288 * init_timerlat_tracefs - A function to initialize the timerlat interface files
2289 */
2290static int init_timerlat_tracefs(struct dentry *top_dir)
2291{
2292	struct dentry *tmp;
 
2293
2294	tmp = tracefs_create_file("timerlat_period_us", TRACE_MODE_WRITE, top_dir,
2295				  &timerlat_period, &trace_min_max_fops);
2296	if (!tmp)
2297		return -ENOMEM;
 
 
 
 
2298
2299	return init_timerlat_stack_tracefs(top_dir);
2300}
2301#else /* CONFIG_TIMERLAT_TRACER */
2302static int init_timerlat_tracefs(struct dentry *top_dir)
2303{
2304	return 0;
2305}
2306#endif /* CONFIG_TIMERLAT_TRACER */
2307
2308/*
2309 * init_tracefs - A function to initialize the tracefs interface files
2310 *
2311 * This function creates entries in tracefs for "osnoise" and "timerlat".
2312 * It creates these directories in the tracing directory, and within that
2313 * directory the use can change and view the configs.
2314 */
2315static int init_tracefs(void)
2316{
2317	struct dentry *top_dir;
2318	struct dentry *tmp;
2319	int ret;
2320
2321	ret = tracing_init_dentry();
2322	if (ret)
2323		return -ENOMEM;
2324
2325	top_dir = tracefs_create_dir("osnoise", NULL);
2326	if (!top_dir)
2327		return 0;
2328
2329	tmp = tracefs_create_file("period_us", TRACE_MODE_WRITE, top_dir,
2330				  &osnoise_period, &trace_min_max_fops);
2331	if (!tmp)
2332		goto err;
2333
2334	tmp = tracefs_create_file("runtime_us", TRACE_MODE_WRITE, top_dir,
2335				  &osnoise_runtime, &trace_min_max_fops);
2336	if (!tmp)
2337		goto err;
2338
2339	tmp = tracefs_create_file("stop_tracing_us", TRACE_MODE_WRITE, top_dir,
2340				  &osnoise_stop_tracing_in, &trace_min_max_fops);
2341	if (!tmp)
2342		goto err;
2343
2344	tmp = tracefs_create_file("stop_tracing_total_us", TRACE_MODE_WRITE, top_dir,
2345				  &osnoise_stop_tracing_total, &trace_min_max_fops);
2346	if (!tmp)
2347		goto err;
2348
2349	tmp = trace_create_file("cpus", TRACE_MODE_WRITE, top_dir, NULL, &cpus_fops);
2350	if (!tmp)
2351		goto err;
2352
2353	tmp = trace_create_file("options", TRACE_MODE_WRITE, top_dir, NULL,
2354				&osnoise_options_fops);
2355	if (!tmp)
2356		goto err;
2357
2358	ret = init_timerlat_tracefs(top_dir);
2359	if (ret)
2360		goto err;
2361
2362	return 0;
2363
2364err:
2365	tracefs_remove(top_dir);
2366	return -ENOMEM;
2367}
2368
2369static int osnoise_hook_events(void)
2370{
2371	int retval;
2372
2373	/*
2374	 * Trace is already hooked, we are re-enabling from
2375	 * a stop_tracing_*.
2376	 */
2377	if (trace_osnoise_callback_enabled)
2378		return 0;
2379
2380	retval = hook_irq_events();
2381	if (retval)
2382		return -EINVAL;
2383
2384	retval = hook_softirq_events();
2385	if (retval)
2386		goto out_unhook_irq;
2387
2388	retval = hook_thread_events();
2389	/*
2390	 * All fine!
2391	 */
2392	if (!retval)
2393		return 0;
2394
2395	unhook_softirq_events();
2396out_unhook_irq:
2397	unhook_irq_events();
2398	return -EINVAL;
2399}
2400
2401static void osnoise_unhook_events(void)
2402{
2403	unhook_thread_events();
2404	unhook_softirq_events();
2405	unhook_irq_events();
2406}
2407
2408/*
2409 * osnoise_workload_start - start the workload and hook to events
2410 */
2411static int osnoise_workload_start(void)
2412{
2413	int retval;
2414
2415	/*
2416	 * Instances need to be registered after calling workload
2417	 * start. Hence, if there is already an instance, the
2418	 * workload was already registered. Otherwise, this
2419	 * code is on the way to register the first instance,
2420	 * and the workload will start.
2421	 */
2422	if (osnoise_has_registered_instances())
2423		return 0;
2424
2425	osn_var_reset_all();
2426
2427	retval = osnoise_hook_events();
2428	if (retval)
2429		return retval;
2430
2431	/*
2432	 * Make sure that ftrace_nmi_enter/exit() see reset values
2433	 * before enabling trace_osnoise_callback_enabled.
2434	 */
2435	barrier();
2436	trace_osnoise_callback_enabled = true;
2437
2438	retval = start_per_cpu_kthreads();
2439	if (retval) {
2440		trace_osnoise_callback_enabled = false;
2441		/*
2442		 * Make sure that ftrace_nmi_enter/exit() see
2443		 * trace_osnoise_callback_enabled as false before continuing.
2444		 */
2445		barrier();
2446
2447		osnoise_unhook_events();
2448		return retval;
2449	}
2450
2451	return 0;
2452}
2453
2454/*
2455 * osnoise_workload_stop - stop the workload and unhook the events
2456 */
2457static void osnoise_workload_stop(void)
2458{
2459	/*
2460	 * Instances need to be unregistered before calling
2461	 * stop. Hence, if there is a registered instance, more
2462	 * than one instance is running, and the workload will not
2463	 * yet stop. Otherwise, this code is on the way to disable
2464	 * the last instance, and the workload can stop.
2465	 */
2466	if (osnoise_has_registered_instances())
2467		return;
2468
2469	/*
2470	 * If callbacks were already disabled in a previous stop
2471	 * call, there is no need to disable then again.
2472	 *
2473	 * For instance, this happens when tracing is stopped via:
2474	 * echo 0 > tracing_on
2475	 * echo nop > current_tracer.
2476	 */
2477	if (!trace_osnoise_callback_enabled)
2478		return;
2479
2480	trace_osnoise_callback_enabled = false;
2481	/*
2482	 * Make sure that ftrace_nmi_enter/exit() see
2483	 * trace_osnoise_callback_enabled as false before continuing.
2484	 */
2485	barrier();
2486
2487	stop_per_cpu_kthreads();
2488
2489	osnoise_unhook_events();
2490}
2491
2492static void osnoise_tracer_start(struct trace_array *tr)
2493{
2494	int retval;
2495
2496	/*
2497	 * If the instance is already registered, there is no need to
2498	 * register it again.
2499	 */
2500	if (osnoise_instance_registered(tr))
2501		return;
2502
2503	retval = osnoise_workload_start();
2504	if (retval)
2505		pr_err(BANNER "Error starting osnoise tracer\n");
2506
2507	osnoise_register_instance(tr);
2508}
2509
2510static void osnoise_tracer_stop(struct trace_array *tr)
2511{
2512	osnoise_unregister_instance(tr);
2513	osnoise_workload_stop();
2514}
2515
2516static int osnoise_tracer_init(struct trace_array *tr)
2517{
2518	/*
2519	 * Only allow osnoise tracer if timerlat tracer is not running
2520	 * already.
2521	 */
2522	if (timerlat_enabled())
2523		return -EBUSY;
2524
2525	tr->max_latency = 0;
2526
2527	osnoise_tracer_start(tr);
2528	return 0;
2529}
2530
2531static void osnoise_tracer_reset(struct trace_array *tr)
2532{
2533	osnoise_tracer_stop(tr);
2534}
2535
2536static struct tracer osnoise_tracer __read_mostly = {
2537	.name		= "osnoise",
2538	.init		= osnoise_tracer_init,
2539	.reset		= osnoise_tracer_reset,
2540	.start		= osnoise_tracer_start,
2541	.stop		= osnoise_tracer_stop,
2542	.print_header	= print_osnoise_headers,
2543	.allow_instances = true,
2544};
2545
2546#ifdef CONFIG_TIMERLAT_TRACER
2547static void timerlat_tracer_start(struct trace_array *tr)
2548{
2549	int retval;
2550
2551	/*
2552	 * If the instance is already registered, there is no need to
2553	 * register it again.
2554	 */
2555	if (osnoise_instance_registered(tr))
2556		return;
2557
2558	retval = osnoise_workload_start();
2559	if (retval)
2560		pr_err(BANNER "Error starting timerlat tracer\n");
2561
2562	osnoise_register_instance(tr);
2563
2564	return;
2565}
2566
2567static void timerlat_tracer_stop(struct trace_array *tr)
2568{
2569	int cpu;
2570
2571	osnoise_unregister_instance(tr);
2572
2573	/*
2574	 * Instruct the threads to stop only if this is the last instance.
2575	 */
2576	if (!osnoise_has_registered_instances()) {
2577		for_each_online_cpu(cpu)
2578			per_cpu(per_cpu_osnoise_var, cpu).sampling = 0;
2579	}
2580
2581	osnoise_workload_stop();
2582}
2583
2584static int timerlat_tracer_init(struct trace_array *tr)
2585{
2586	/*
2587	 * Only allow timerlat tracer if osnoise tracer is not running already.
2588	 */
2589	if (osnoise_has_registered_instances() && !osnoise_data.timerlat_tracer)
2590		return -EBUSY;
2591
2592	/*
2593	 * If this is the first instance, set timerlat_tracer to block
2594	 * osnoise tracer start.
2595	 */
2596	if (!osnoise_has_registered_instances())
2597		osnoise_data.timerlat_tracer = 1;
2598
2599	tr->max_latency = 0;
2600	timerlat_tracer_start(tr);
2601
2602	return 0;
2603}
2604
2605static void timerlat_tracer_reset(struct trace_array *tr)
2606{
2607	timerlat_tracer_stop(tr);
2608
2609	/*
2610	 * If this is the last instance, reset timerlat_tracer allowing
2611	 * osnoise to be started.
2612	 */
2613	if (!osnoise_has_registered_instances())
2614		osnoise_data.timerlat_tracer = 0;
2615}
2616
2617static struct tracer timerlat_tracer __read_mostly = {
2618	.name		= "timerlat",
2619	.init		= timerlat_tracer_init,
2620	.reset		= timerlat_tracer_reset,
2621	.start		= timerlat_tracer_start,
2622	.stop		= timerlat_tracer_stop,
2623	.print_header	= print_timerlat_headers,
2624	.allow_instances = true,
2625};
2626
2627__init static int init_timerlat_tracer(void)
2628{
2629	return register_tracer(&timerlat_tracer);
2630}
2631#else /* CONFIG_TIMERLAT_TRACER */
2632__init static int init_timerlat_tracer(void)
2633{
2634	return 0;
2635}
2636#endif /* CONFIG_TIMERLAT_TRACER */
2637
2638__init static int init_osnoise_tracer(void)
2639{
2640	int ret;
2641
2642	mutex_init(&interface_lock);
2643
2644	cpumask_copy(&osnoise_cpumask, cpu_all_mask);
2645
2646	ret = register_tracer(&osnoise_tracer);
2647	if (ret) {
2648		pr_err(BANNER "Error registering osnoise!\n");
2649		return ret;
2650	}
2651
2652	ret = init_timerlat_tracer();
2653	if (ret) {
2654		pr_err(BANNER "Error registering timerlat!\n");
2655		return ret;
2656	}
2657
2658	osnoise_init_hotplug_support();
2659
2660	INIT_LIST_HEAD_RCU(&osnoise_instances);
2661
2662	init_tracefs();
2663
2664	return 0;
2665}
2666late_initcall(init_osnoise_tracer);