1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Detect hard and soft lockups on a system
   4 *
   5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   6 *
   7 * Note: Most of this code is borrowed heavily from the original softlockup
   8 * detector, so thanks to Ingo for the initial implementation.
   9 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
  10 * to those contributors as well.
  11 */
  12
  13#define pr_fmt(fmt) "watchdog: " fmt
  14
 
  15#include <linux/cpu.h>
 
  16#include <linux/init.h>
  17#include <linux/irq.h>
  18#include <linux/irqdesc.h>
  19#include <linux/kernel_stat.h>
  20#include <linux/kvm_para.h>
  21#include <linux/math64.h>
  22#include <linux/mm.h>
  23#include <linux/module.h>
  24#include <linux/nmi.h>
  25#include <linux/stop_machine.h>
  26#include <linux/sysctl.h>
  27#include <linux/tick.h>
  28
  29#include <linux/sched/clock.h>
  30#include <linux/sched/debug.h>
  31#include <linux/sched/isolation.h>
 
  32
  33#include <asm/irq_regs.h>
 
  34
  35static DEFINE_MUTEX(watchdog_mutex);
  36
  37#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HARDLOCKUP_DETECTOR_SPARC64)
  38# define WATCHDOG_HARDLOCKUP_DEFAULT	1
  39#else
  40# define WATCHDOG_HARDLOCKUP_DEFAULT	0
  41#endif
  42
  43#define NUM_SAMPLE_PERIODS	5
  44
  45unsigned long __read_mostly watchdog_enabled;
  46int __read_mostly watchdog_user_enabled = 1;
  47static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT;
  48static int __read_mostly watchdog_softlockup_user_enabled = 1;
  49int __read_mostly watchdog_thresh = 10;
  50static int __read_mostly watchdog_hardlockup_available;
  51
  52struct cpumask watchdog_cpumask __read_mostly;
  53unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
  54
  55#ifdef CONFIG_HARDLOCKUP_DETECTOR
  56
  57# ifdef CONFIG_SMP
  58int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
  59# endif /* CONFIG_SMP */
  60
  61/*
  62 * Should we panic when a soft-lockup or hard-lockup occurs:
  63 */
  64unsigned int __read_mostly hardlockup_panic =
  65			IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
  66/*
  67 * We may not want to enable hard lockup detection by default in all cases,
  68 * for example when running the kernel as a guest on a hypervisor. In these
  69 * cases this function can be called to disable hard lockup detection. This
  70 * function should only be executed once by the boot processor before the
  71 * kernel command line parameters are parsed, because otherwise it is not
  72 * possible to override this in hardlockup_panic_setup().
  73 */
  74void __init hardlockup_detector_disable(void)
  75{
  76	watchdog_hardlockup_user_enabled = 0;
  77}
  78
  79static int __init hardlockup_panic_setup(char *str)
  80{
  81next:
  82	if (!strncmp(str, "panic", 5))
  83		hardlockup_panic = 1;
  84	else if (!strncmp(str, "nopanic", 7))
  85		hardlockup_panic = 0;
  86	else if (!strncmp(str, "0", 1))
  87		watchdog_hardlockup_user_enabled = 0;
  88	else if (!strncmp(str, "1", 1))
  89		watchdog_hardlockup_user_enabled = 1;
  90	else if (!strncmp(str, "r", 1))
  91		hardlockup_config_perf_event(str + 1);
  92	while (*(str++)) {
  93		if (*str == ',') {
  94			str++;
  95			goto next;
  96		}
  97	}
  98	return 1;
  99}
 100__setup("nmi_watchdog=", hardlockup_panic_setup);
 101
 102#endif /* CONFIG_HARDLOCKUP_DETECTOR */
 103
 104#if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER)
 105
 106static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts);
 107static DEFINE_PER_CPU(int, hrtimer_interrupts_saved);
 108static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned);
 109static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched);
 110static unsigned long hard_lockup_nmi_warn;
 111
 112notrace void arch_touch_nmi_watchdog(void)
 113{
 114	/*
 115	 * Using __raw here because some code paths have
 116	 * preemption enabled.  If preemption is enabled
 117	 * then interrupts should be enabled too, in which
 118	 * case we shouldn't have to worry about the watchdog
 119	 * going off.
 120	 */
 121	raw_cpu_write(watchdog_hardlockup_touched, true);
 122}
 123EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 124
 125void watchdog_hardlockup_touch_cpu(unsigned int cpu)
 126{
 127	per_cpu(watchdog_hardlockup_touched, cpu) = true;
 128}
 129
 130static bool is_hardlockup(unsigned int cpu)
 131{
 132	int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu));
 133
 134	if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
 135		return true;
 136
 137	/*
 138	 * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE
 139	 * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is
 140	 * written/read by a single CPU.
 141	 */
 142	per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
 143
 144	return false;
 145}
 146
 147static void watchdog_hardlockup_kick(void)
 148{
 149	int new_interrupts;
 150
 151	new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts));
 152	watchdog_buddy_check_hardlockup(new_interrupts);
 153}
 154
 155void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
 156{
 157	if (per_cpu(watchdog_hardlockup_touched, cpu)) {
 158		per_cpu(watchdog_hardlockup_touched, cpu) = false;
 159		return;
 160	}
 161
 162	/*
 163	 * Check for a hardlockup by making sure the CPU's timer
 164	 * interrupt is incrementing. The timer interrupt should have
 165	 * fired multiple times before we overflow'd. If it hasn't
 166	 * then this is a good indication the cpu is stuck
 167	 */
 168	if (is_hardlockup(cpu)) {
 169		unsigned int this_cpu = smp_processor_id();
 170		unsigned long flags;
 171
 172		/* Only print hardlockups once. */
 173		if (per_cpu(watchdog_hardlockup_warned, cpu))
 174			return;
 175
 176		/*
 177		 * Prevent multiple hard-lockup reports if one cpu is already
 178		 * engaged in dumping all cpu back traces.
 179		 */
 180		if (sysctl_hardlockup_all_cpu_backtrace) {
 181			if (test_and_set_bit_lock(0, &hard_lockup_nmi_warn))
 182				return;
 183		}
 184
 185		/*
 186		 * NOTE: we call printk_cpu_sync_get_irqsave() after printing
 187		 * the lockup message. While it would be nice to serialize
 188		 * that printout, we really want to make sure that if some
 189		 * other CPU somehow locked up while holding the lock associated
 190		 * with printk_cpu_sync_get_irqsave() that we can still at least
 191		 * get the message about the lockup out.
 192		 */
 193		pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
 194		printk_cpu_sync_get_irqsave(flags);
 195
 196		print_modules();
 197		print_irqtrace_events(current);
 198		if (cpu == this_cpu) {
 199			if (regs)
 200				show_regs(regs);
 201			else
 202				dump_stack();
 203			printk_cpu_sync_put_irqrestore(flags);
 204		} else {
 205			printk_cpu_sync_put_irqrestore(flags);
 206			trigger_single_cpu_backtrace(cpu);
 207		}
 208
 209		if (sysctl_hardlockup_all_cpu_backtrace) {
 210			trigger_allbutcpu_cpu_backtrace(cpu);
 211			if (!hardlockup_panic)
 212				clear_bit_unlock(0, &hard_lockup_nmi_warn);
 213		}
 214
 215		if (hardlockup_panic)
 216			nmi_panic(regs, "Hard LOCKUP");
 217
 218		per_cpu(watchdog_hardlockup_warned, cpu) = true;
 219	} else {
 220		per_cpu(watchdog_hardlockup_warned, cpu) = false;
 221	}
 222}
 223
 224#else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
 225
 226static inline void watchdog_hardlockup_kick(void) { }
 227
 228#endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
 229
 230/*
 231 * These functions can be overridden based on the configured hardlockdup detector.
 232 *
 233 * watchdog_hardlockup_enable/disable can be implemented to start and stop when
 234 * softlockup watchdog start and stop. The detector must select the
 235 * SOFTLOCKUP_DETECTOR Kconfig.
 236 */
 237void __weak watchdog_hardlockup_enable(unsigned int cpu) { }
 238
 239void __weak watchdog_hardlockup_disable(unsigned int cpu) { }
 240
 241/*
 242 * Watchdog-detector specific API.
 243 *
 244 * Return 0 when hardlockup watchdog is available, negative value otherwise.
 245 * Note that the negative value means that a delayed probe might
 246 * succeed later.
 247 */
 248int __weak __init watchdog_hardlockup_probe(void)
 249{
 250	return -ENODEV;
 251}
 252
 253/**
 254 * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration
 255 *
 256 * The reconfiguration steps are:
 257 * watchdog_hardlockup_stop();
 258 * update_variables();
 259 * watchdog_hardlockup_start();
 260 */
 261void __weak watchdog_hardlockup_stop(void) { }
 262
 263/**
 264 * watchdog_hardlockup_start - Start the watchdog after reconfiguration
 265 *
 266 * Counterpart to watchdog_hardlockup_stop().
 267 *
 268 * The following variables have been updated in update_variables() and
 269 * contain the currently valid configuration:
 270 * - watchdog_enabled
 271 * - watchdog_thresh
 272 * - watchdog_cpumask
 273 */
 274void __weak watchdog_hardlockup_start(void) { }
 275
 276/**
 277 * lockup_detector_update_enable - Update the sysctl enable bit
 278 *
 279 * Caller needs to make sure that the hard watchdogs are off, so this
 280 * can't race with watchdog_hardlockup_disable().
 281 */
 282static void lockup_detector_update_enable(void)
 283{
 284	watchdog_enabled = 0;
 285	if (!watchdog_user_enabled)
 286		return;
 287	if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled)
 288		watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED;
 289	if (watchdog_softlockup_user_enabled)
 290		watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED;
 291}
 292
 293#ifdef CONFIG_SOFTLOCKUP_DETECTOR
 294
 295/*
 296 * Delay the soflockup report when running a known slow code.
 297 * It does _not_ affect the timestamp of the last successdul reschedule.
 298 */
 299#define SOFTLOCKUP_DELAY_REPORT	ULONG_MAX
 300
 301#ifdef CONFIG_SMP
 302int __read_mostly sysctl_softlockup_all_cpu_backtrace;
 303#endif
 304
 305static struct cpumask watchdog_allowed_mask __read_mostly;
 306
 307/* Global variables, exported for sysctl */
 308unsigned int __read_mostly softlockup_panic =
 309			IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
 310
 311static bool softlockup_initialized __read_mostly;
 312static u64 __read_mostly sample_period;
 313
 314/* Timestamp taken after the last successful reschedule. */
 315static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
 316/* Timestamp of the last softlockup report. */
 317static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
 318static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
 319static DEFINE_PER_CPU(bool, softlockup_touch_sync);
 320static unsigned long soft_lockup_nmi_warn;
 321
 322static int __init softlockup_panic_setup(char *str)
 323{
 324	softlockup_panic = simple_strtoul(str, NULL, 0);
 325	return 1;
 326}
 327__setup("softlockup_panic=", softlockup_panic_setup);
 328
 329static int __init nowatchdog_setup(char *str)
 330{
 331	watchdog_user_enabled = 0;
 332	return 1;
 333}
 334__setup("nowatchdog", nowatchdog_setup);
 335
 336static int __init nosoftlockup_setup(char *str)
 337{
 338	watchdog_softlockup_user_enabled = 0;
 339	return 1;
 340}
 341__setup("nosoftlockup", nosoftlockup_setup);
 342
 343static int __init watchdog_thresh_setup(char *str)
 344{
 345	get_option(&str, &watchdog_thresh);
 346	return 1;
 347}
 348__setup("watchdog_thresh=", watchdog_thresh_setup);
 349
 350static void __lockup_detector_cleanup(void);
 351
 352#ifdef CONFIG_SOFTLOCKUP_DETECTOR_INTR_STORM
 353enum stats_per_group {
 354	STATS_SYSTEM,
 355	STATS_SOFTIRQ,
 356	STATS_HARDIRQ,
 357	STATS_IDLE,
 358	NUM_STATS_PER_GROUP,
 359};
 360
 361static const enum cpu_usage_stat tracked_stats[NUM_STATS_PER_GROUP] = {
 362	CPUTIME_SYSTEM,
 363	CPUTIME_SOFTIRQ,
 364	CPUTIME_IRQ,
 365	CPUTIME_IDLE,
 366};
 367
 368static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]);
 369static DEFINE_PER_CPU(u8, cpustat_util[NUM_SAMPLE_PERIODS][NUM_STATS_PER_GROUP]);
 370static DEFINE_PER_CPU(u8, cpustat_tail);
 371
 372/*
 373 * We don't need nanosecond resolution. A granularity of 16ms is
 374 * sufficient for our precision, allowing us to use u16 to store
 375 * cpustats, which will roll over roughly every ~1000 seconds.
 376 * 2^24 ~= 16 * 10^6
 377 */
 378static u16 get_16bit_precision(u64 data_ns)
 379{
 380	return data_ns >> 24LL; /* 2^24ns ~= 16.8ms */
 381}
 382
 383static void update_cpustat(void)
 384{
 385	int i;
 386	u8 util;
 387	u16 old_stat, new_stat;
 388	struct kernel_cpustat kcpustat;
 389	u64 *cpustat = kcpustat.cpustat;
 390	u8 tail = __this_cpu_read(cpustat_tail);
 391	u16 sample_period_16 = get_16bit_precision(sample_period);
 392
 393	kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
 394
 395	for (i = 0; i < NUM_STATS_PER_GROUP; i++) {
 396		old_stat = __this_cpu_read(cpustat_old[i]);
 397		new_stat = get_16bit_precision(cpustat[tracked_stats[i]]);
 398		util = DIV_ROUND_UP(100 * (new_stat - old_stat), sample_period_16);
 399		__this_cpu_write(cpustat_util[tail][i], util);
 400		__this_cpu_write(cpustat_old[i], new_stat);
 401	}
 402
 403	__this_cpu_write(cpustat_tail, (tail + 1) % NUM_SAMPLE_PERIODS);
 404}
 405
 406static void print_cpustat(void)
 407{
 408	int i, group;
 409	u8 tail = __this_cpu_read(cpustat_tail);
 410	u64 sample_period_second = sample_period;
 411
 412	do_div(sample_period_second, NSEC_PER_SEC);
 413
 414	/*
 415	 * Outputting the "watchdog" prefix on every line is redundant and not
 416	 * concise, and the original alarm information is sufficient for
 417	 * positioning in logs, hence here printk() is used instead of pr_crit().
 418	 */
 419	printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n",
 420	       smp_processor_id(), sample_period_second);
 421
 422	for (i = 0; i < NUM_SAMPLE_PERIODS; i++) {
 423		group = (tail + i) % NUM_SAMPLE_PERIODS;
 424		printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t"
 425			"%3u%% hardirq,\t%3u%% idle\n", i + 1,
 426			__this_cpu_read(cpustat_util[group][STATS_SYSTEM]),
 427			__this_cpu_read(cpustat_util[group][STATS_SOFTIRQ]),
 428			__this_cpu_read(cpustat_util[group][STATS_HARDIRQ]),
 429			__this_cpu_read(cpustat_util[group][STATS_IDLE]));
 430	}
 431}
 432
 433#define HARDIRQ_PERCENT_THRESH          50
 434#define NUM_HARDIRQ_REPORT              5
 435struct irq_counts {
 436	int irq;
 437	u32 counts;
 438};
 439
 440static DEFINE_PER_CPU(bool, snapshot_taken);
 441
 442/* Tabulate the most frequent interrupts. */
 443static void tabulate_irq_count(struct irq_counts *irq_counts, int irq, u32 counts, int rank)
 444{
 445	int i;
 446	struct irq_counts new_count = {irq, counts};
 447
 448	for (i = 0; i < rank; i++) {
 449		if (counts > irq_counts[i].counts)
 450			swap(new_count, irq_counts[i]);
 451	}
 452}
 453
 454/*
 455 * If the hardirq time exceeds HARDIRQ_PERCENT_THRESH% of the sample_period,
 456 * then the cause of softlockup might be interrupt storm. In this case, it
 457 * would be useful to start interrupt counting.
 458 */
 459static bool need_counting_irqs(void)
 460{
 461	u8 util;
 462	int tail = __this_cpu_read(cpustat_tail);
 463
 464	tail = (tail + NUM_HARDIRQ_REPORT - 1) % NUM_HARDIRQ_REPORT;
 465	util = __this_cpu_read(cpustat_util[tail][STATS_HARDIRQ]);
 466	return util > HARDIRQ_PERCENT_THRESH;
 467}
 468
 469static void start_counting_irqs(void)
 470{
 471	if (!__this_cpu_read(snapshot_taken)) {
 472		kstat_snapshot_irqs();
 473		__this_cpu_write(snapshot_taken, true);
 474	}
 475}
 476
 477static void stop_counting_irqs(void)
 478{
 479	__this_cpu_write(snapshot_taken, false);
 480}
 481
 482static void print_irq_counts(void)
 483{
 484	unsigned int i, count;
 485	struct irq_counts irq_counts_sorted[NUM_HARDIRQ_REPORT] = {
 486		{-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}
 487	};
 488
 489	if (__this_cpu_read(snapshot_taken)) {
 490		for_each_active_irq(i) {
 491			count = kstat_get_irq_since_snapshot(i);
 492			tabulate_irq_count(irq_counts_sorted, i, count, NUM_HARDIRQ_REPORT);
 493		}
 494
 495		/*
 496		 * Outputting the "watchdog" prefix on every line is redundant and not
 497		 * concise, and the original alarm information is sufficient for
 498		 * positioning in logs, hence here printk() is used instead of pr_crit().
 499		 */
 500		printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n",
 501		       smp_processor_id(), HARDIRQ_PERCENT_THRESH);
 502
 503		for (i = 0; i < NUM_HARDIRQ_REPORT; i++) {
 504			if (irq_counts_sorted[i].irq == -1)
 505				break;
 506
 507			printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n",
 508			       i + 1, irq_counts_sorted[i].counts,
 509			       irq_counts_sorted[i].irq);
 510		}
 511
 512		/*
 513		 * If the hardirq time is less than HARDIRQ_PERCENT_THRESH% in the last
 514		 * sample_period, then we suspect the interrupt storm might be subsiding.
 515		 */
 516		if (!need_counting_irqs())
 517			stop_counting_irqs();
 518	}
 519}
 520
 521static void report_cpu_status(void)
 522{
 523	print_cpustat();
 524	print_irq_counts();
 525}
 526#else
 527static inline void update_cpustat(void) { }
 528static inline void report_cpu_status(void) { }
 529static inline bool need_counting_irqs(void) { return false; }
 530static inline void start_counting_irqs(void) { }
 531static inline void stop_counting_irqs(void) { }
 532#endif
 533
 534/*
 535 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
 536 * lockups can have false positives under extreme conditions. So we generally
 537 * want a higher threshold for soft lockups than for hard lockups. So we couple
 538 * the thresholds with a factor: we make the soft threshold twice the amount of
 539 * time the hard threshold is.
 540 */
 541static int get_softlockup_thresh(void)
 542{
 543	return watchdog_thresh * 2;
 544}
 545
 546/*
 547 * Returns seconds, approximately.  We don't need nanosecond
 548 * resolution, and we don't need to waste time with a big divide when
 549 * 2^30ns == 1.074s.
 550 */
 551static unsigned long get_timestamp(void)
 552{
 553	return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
 554}
 555
 556static void set_sample_period(void)
 557{
 558	/*
 559	 * convert watchdog_thresh from seconds to ns
 560	 * the divide by 5 is to give hrtimer several chances (two
 561	 * or three with the current relation between the soft
 562	 * and hard thresholds) to increment before the
 563	 * hardlockup detector generates a warning
 564	 */
 565	sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / NUM_SAMPLE_PERIODS);
 566	watchdog_update_hrtimer_threshold(sample_period);
 567}
 568
 569static void update_report_ts(void)
 570{
 571	__this_cpu_write(watchdog_report_ts, get_timestamp());
 572}
 573
 574/* Commands for resetting the watchdog */
 575static void update_touch_ts(void)
 576{
 577	__this_cpu_write(watchdog_touch_ts, get_timestamp());
 578	update_report_ts();
 579}
 580
 581/**
 582 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
 583 *
 584 * Call when the scheduler may have stalled for legitimate reasons
 585 * preventing the watchdog task from executing - e.g. the scheduler
 586 * entering idle state.  This should only be used for scheduler events.
 587 * Use touch_softlockup_watchdog() for everything else.
 588 */
 589notrace void touch_softlockup_watchdog_sched(void)
 590{
 591	/*
 592	 * Preemption can be enabled.  It doesn't matter which CPU's watchdog
 593	 * report period gets restarted here, so use the raw_ operation.
 594	 */
 595	raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 596}
 597
 598notrace void touch_softlockup_watchdog(void)
 599{
 600	touch_softlockup_watchdog_sched();
 601	wq_watchdog_touch(raw_smp_processor_id());
 602}
 603EXPORT_SYMBOL(touch_softlockup_watchdog);
 604
 605void touch_all_softlockup_watchdogs(void)
 606{
 607	int cpu;
 608
 609	/*
 610	 * watchdog_mutex cannpt be taken here, as this might be called
 611	 * from (soft)interrupt context, so the access to
 612	 * watchdog_allowed_cpumask might race with a concurrent update.
 613	 *
 614	 * The watchdog time stamp can race against a concurrent real
 615	 * update as well, the only side effect might be a cycle delay for
 616	 * the softlockup check.
 617	 */
 618	for_each_cpu(cpu, &watchdog_allowed_mask) {
 619		per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
 620		wq_watchdog_touch(cpu);
 621	}
 622}
 623
 624void touch_softlockup_watchdog_sync(void)
 625{
 626	__this_cpu_write(softlockup_touch_sync, true);
 627	__this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
 628}
 629
 630static int is_softlockup(unsigned long touch_ts,
 631			 unsigned long period_ts,
 632			 unsigned long now)
 633{
 634	if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
 635		/*
 636		 * If period_ts has not been updated during a sample_period, then
 637		 * in the subsequent few sample_periods, period_ts might also not
 638		 * be updated, which could indicate a potential softlockup. In
 639		 * this case, if we suspect the cause of the potential softlockup
 640		 * might be interrupt storm, then we need to count the interrupts
 641		 * to find which interrupt is storming.
 642		 */
 643		if (time_after_eq(now, period_ts + get_softlockup_thresh() / NUM_SAMPLE_PERIODS) &&
 644		    need_counting_irqs())
 645			start_counting_irqs();
 646
 647		/*
 648		 * A poorly behaving BPF scheduler can live-lock the system into
 649		 * soft lockups. Tell sched_ext to try ejecting the BPF
 650		 * scheduler when close to a soft lockup.
 651		 */
 652		if (time_after_eq(now, period_ts + get_softlockup_thresh() * 3 / 4))
 653			scx_softlockup(now - touch_ts);
 654
 655		/* Warn about unreasonable delays. */
 656		if (time_after(now, period_ts + get_softlockup_thresh()))
 657			return now - touch_ts;
 658	}
 659	return 0;
 660}
 661
 662/* watchdog detector functions */
 663static DEFINE_PER_CPU(struct completion, softlockup_completion);
 664static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
 665
 666/*
 667 * The watchdog feed function - touches the timestamp.
 668 *
 669 * It only runs once every sample_period seconds (4 seconds by
 670 * default) to reset the softlockup timestamp. If this gets delayed
 671 * for more than 2*watchdog_thresh seconds then the debug-printout
 672 * triggers in watchdog_timer_fn().
 673 */
 674static int softlockup_fn(void *data)
 675{
 676	update_touch_ts();
 677	stop_counting_irqs();
 678	complete(this_cpu_ptr(&softlockup_completion));
 679
 680	return 0;
 681}
 682
 683/* watchdog kicker functions */
 684static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 685{
 686	unsigned long touch_ts, period_ts, now;
 687	struct pt_regs *regs = get_irq_regs();
 688	int duration;
 689	int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
 690	unsigned long flags;
 691
 692	if (!watchdog_enabled)
 693		return HRTIMER_NORESTART;
 694
 695	watchdog_hardlockup_kick();
 696
 697	/* kick the softlockup detector */
 698	if (completion_done(this_cpu_ptr(&softlockup_completion))) {
 699		reinit_completion(this_cpu_ptr(&softlockup_completion));
 700		stop_one_cpu_nowait(smp_processor_id(),
 701				softlockup_fn, NULL,
 702				this_cpu_ptr(&softlockup_stop_work));
 703	}
 704
 705	/* .. and repeat */
 706	hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
 707
 708	/*
 709	 * Read the current timestamp first. It might become invalid anytime
 710	 * when a virtual machine is stopped by the host or when the watchog
 711	 * is touched from NMI.
 712	 */
 713	now = get_timestamp();
 714	/*
 715	 * If a virtual machine is stopped by the host it can look to
 716	 * the watchdog like a soft lockup. This function touches the watchdog.
 717	 */
 718	kvm_check_and_clear_guest_paused();
 719	/*
 720	 * The stored timestamp is comparable with @now only when not touched.
 721	 * It might get touched anytime from NMI. Make sure that is_softlockup()
 722	 * uses the same (valid) value.
 723	 */
 724	period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
 725
 726	update_cpustat();
 727
 728	/* Reset the interval when touched by known problematic code. */
 729	if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
 730		if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
 731			/*
 732			 * If the time stamp was touched atomically
 733			 * make sure the scheduler tick is up to date.
 734			 */
 735			__this_cpu_write(softlockup_touch_sync, false);
 736			sched_clock_tick();
 737		}
 738
 739		update_report_ts();
 740		return HRTIMER_RESTART;
 741	}
 742
 743	/* Check for a softlockup. */
 744	touch_ts = __this_cpu_read(watchdog_touch_ts);
 745	duration = is_softlockup(touch_ts, period_ts, now);
 746	if (unlikely(duration)) {
 747		/*
 748		 * Prevent multiple soft-lockup reports if one cpu is already
 749		 * engaged in dumping all cpu back traces.
 750		 */
 751		if (softlockup_all_cpu_backtrace) {
 752			if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
 753				return HRTIMER_RESTART;
 754		}
 755
 756		/* Start period for the next softlockup warning. */
 757		update_report_ts();
 758
 759		printk_cpu_sync_get_irqsave(flags);
 760		pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
 761			smp_processor_id(), duration,
 762			current->comm, task_pid_nr(current));
 763		report_cpu_status();
 764		print_modules();
 765		print_irqtrace_events(current);
 766		if (regs)
 767			show_regs(regs);
 768		else
 769			dump_stack();
 770		printk_cpu_sync_put_irqrestore(flags);
 771
 772		if (softlockup_all_cpu_backtrace) {
 773			trigger_allbutcpu_cpu_backtrace(smp_processor_id());
 774			if (!softlockup_panic)
 775				clear_bit_unlock(0, &soft_lockup_nmi_warn);
 776		}
 777
 778		add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
 779		if (softlockup_panic)
 780			panic("softlockup: hung tasks");
 781	}
 782
 783	return HRTIMER_RESTART;
 784}
 785
 786static void watchdog_enable(unsigned int cpu)
 787{
 788	struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 789	struct completion *done = this_cpu_ptr(&softlockup_completion);
 790
 791	WARN_ON_ONCE(cpu != smp_processor_id());
 792
 793	init_completion(done);
 794	complete(done);
 795
 796	/*
 797	 * Start the timer first to prevent the hardlockup watchdog triggering
 798	 * before the timer has a chance to fire.
 799	 */
 800	hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 801	hrtimer->function = watchdog_timer_fn;
 802	hrtimer_start(hrtimer, ns_to_ktime(sample_period),
 803		      HRTIMER_MODE_REL_PINNED_HARD);
 804
 805	/* Initialize timestamp */
 806	update_touch_ts();
 807	/* Enable the hardlockup detector */
 808	if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)
 809		watchdog_hardlockup_enable(cpu);
 810}
 811
 812static void watchdog_disable(unsigned int cpu)
 813{
 814	struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
 815
 816	WARN_ON_ONCE(cpu != smp_processor_id());
 817
 818	/*
 819	 * Disable the hardlockup detector first. That prevents that a large
 820	 * delay between disabling the timer and disabling the hardlockup
 821	 * detector causes a false positive.
 822	 */
 823	watchdog_hardlockup_disable(cpu);
 824	hrtimer_cancel(hrtimer);
 825	wait_for_completion(this_cpu_ptr(&softlockup_completion));
 826}
 827
 828static int softlockup_stop_fn(void *data)
 829{
 830	watchdog_disable(smp_processor_id());
 831	return 0;
 832}
 833
 834static void softlockup_stop_all(void)
 835{
 836	int cpu;
 837
 838	if (!softlockup_initialized)
 839		return;
 840
 841	for_each_cpu(cpu, &watchdog_allowed_mask)
 842		smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
 843
 844	cpumask_clear(&watchdog_allowed_mask);
 845}
 846
 847static int softlockup_start_fn(void *data)
 848{
 849	watchdog_enable(smp_processor_id());
 850	return 0;
 851}
 852
 853static void softlockup_start_all(void)
 854{
 855	int cpu;
 856
 857	cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
 858	for_each_cpu(cpu, &watchdog_allowed_mask)
 859		smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
 860}
 861
 862int lockup_detector_online_cpu(unsigned int cpu)
 863{
 864	if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 865		watchdog_enable(cpu);
 866	return 0;
 867}
 868
 869int lockup_detector_offline_cpu(unsigned int cpu)
 870{
 871	if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
 872		watchdog_disable(cpu);
 873	return 0;
 874}
 875
 876static void __lockup_detector_reconfigure(void)
 877{
 878	cpus_read_lock();
 879	watchdog_hardlockup_stop();
 880
 881	softlockup_stop_all();
 882	set_sample_period();
 883	lockup_detector_update_enable();
 884	if (watchdog_enabled && watchdog_thresh)
 885		softlockup_start_all();
 886
 887	watchdog_hardlockup_start();
 888	cpus_read_unlock();
 889	/*
 890	 * Must be called outside the cpus locked section to prevent
 891	 * recursive locking in the perf code.
 892	 */
 893	__lockup_detector_cleanup();
 894}
 895
 896void lockup_detector_reconfigure(void)
 897{
 898	mutex_lock(&watchdog_mutex);
 899	__lockup_detector_reconfigure();
 900	mutex_unlock(&watchdog_mutex);
 901}
 902
 903/*
 904 * Create the watchdog infrastructure and configure the detector(s).
 905 */
 906static __init void lockup_detector_setup(void)
 907{
 908	/*
 909	 * If sysctl is off and watchdog got disabled on the command line,
 910	 * nothing to do here.
 911	 */
 912	lockup_detector_update_enable();
 913
 914	if (!IS_ENABLED(CONFIG_SYSCTL) &&
 915	    !(watchdog_enabled && watchdog_thresh))
 916		return;
 917
 918	mutex_lock(&watchdog_mutex);
 919	__lockup_detector_reconfigure();
 920	softlockup_initialized = true;
 921	mutex_unlock(&watchdog_mutex);
 922}
 923
 924#else /* CONFIG_SOFTLOCKUP_DETECTOR */
 925static void __lockup_detector_reconfigure(void)
 926{
 927	cpus_read_lock();
 928	watchdog_hardlockup_stop();
 929	lockup_detector_update_enable();
 930	watchdog_hardlockup_start();
 931	cpus_read_unlock();
 932}
 933void lockup_detector_reconfigure(void)
 934{
 935	__lockup_detector_reconfigure();
 936}
 937static inline void lockup_detector_setup(void)
 938{
 939	__lockup_detector_reconfigure();
 940}
 941#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
 942
 943static void __lockup_detector_cleanup(void)
 944{
 945	lockdep_assert_held(&watchdog_mutex);
 946	hardlockup_detector_perf_cleanup();
 947}
 948
 949/**
 950 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
 951 *
 952 * Caller must not hold the cpu hotplug rwsem.
 953 */
 954void lockup_detector_cleanup(void)
 955{
 956	mutex_lock(&watchdog_mutex);
 957	__lockup_detector_cleanup();
 958	mutex_unlock(&watchdog_mutex);
 959}
 960
 961/**
 962 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
 963 *
 964 * Special interface for parisc. It prevents lockup detector warnings from
 965 * the default pm_poweroff() function which busy loops forever.
 966 */
 967void lockup_detector_soft_poweroff(void)
 968{
 969	watchdog_enabled = 0;
 970}
 971
 972#ifdef CONFIG_SYSCTL
 973
 974/* Propagate any changes to the watchdog infrastructure */
 975static void proc_watchdog_update(void)
 976{
 977	/* Remove impossible cpus to keep sysctl output clean. */
 978	cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
 979	__lockup_detector_reconfigure();
 980}
 981
 982/*
 983 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
 984 *
 985 * caller             | table->data points to            | 'which'
 986 * -------------------|----------------------------------|-------------------------------
 987 * proc_watchdog      | watchdog_user_enabled            | WATCHDOG_HARDLOCKUP_ENABLED |
 988 *                    |                                  | WATCHDOG_SOFTOCKUP_ENABLED
 989 * -------------------|----------------------------------|-------------------------------
 990 * proc_nmi_watchdog  | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED
 991 * -------------------|----------------------------------|-------------------------------
 992 * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED
 993 */
 994static int proc_watchdog_common(int which, const struct ctl_table *table, int write,
 995				void *buffer, size_t *lenp, loff_t *ppos)
 996{
 997	int err, old, *param = table->data;
 998
 999	mutex_lock(&watchdog_mutex);
1000
1001	old = *param;
1002	if (!write) {
1003		/*
1004		 * On read synchronize the userspace interface. This is a
1005		 * racy snapshot.
1006		 */
1007		*param = (watchdog_enabled & which) != 0;
1008		err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1009		*param = old;
1010	} else {
 
1011		err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1012		if (!err && old != READ_ONCE(*param))
1013			proc_watchdog_update();
1014	}
1015	mutex_unlock(&watchdog_mutex);
1016	return err;
1017}
1018
1019/*
1020 * /proc/sys/kernel/watchdog
1021 */
1022static int proc_watchdog(const struct ctl_table *table, int write,
1023			 void *buffer, size_t *lenp, loff_t *ppos)
1024{
1025	return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED |
1026				    WATCHDOG_SOFTOCKUP_ENABLED,
1027				    table, write, buffer, lenp, ppos);
1028}
1029
1030/*
1031 * /proc/sys/kernel/nmi_watchdog
1032 */
1033static int proc_nmi_watchdog(const struct ctl_table *table, int write,
1034			     void *buffer, size_t *lenp, loff_t *ppos)
1035{
1036	if (!watchdog_hardlockup_available && write)
1037		return -ENOTSUPP;
1038	return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED,
1039				    table, write, buffer, lenp, ppos);
1040}
1041
1042#ifdef CONFIG_SOFTLOCKUP_DETECTOR
1043/*
1044 * /proc/sys/kernel/soft_watchdog
1045 */
1046static int proc_soft_watchdog(const struct ctl_table *table, int write,
1047			      void *buffer, size_t *lenp, loff_t *ppos)
1048{
1049	return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED,
1050				    table, write, buffer, lenp, ppos);
1051}
1052#endif
1053
1054/*
1055 * /proc/sys/kernel/watchdog_thresh
1056 */
1057static int proc_watchdog_thresh(const struct ctl_table *table, int write,
1058				void *buffer, size_t *lenp, loff_t *ppos)
1059{
1060	int err, old;
1061
1062	mutex_lock(&watchdog_mutex);
1063
1064	old = READ_ONCE(watchdog_thresh);
1065	err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1066
1067	if (!err && write && old != READ_ONCE(watchdog_thresh))
1068		proc_watchdog_update();
1069
1070	mutex_unlock(&watchdog_mutex);
1071	return err;
1072}
1073
1074/*
1075 * The cpumask is the mask of possible cpus that the watchdog can run
1076 * on, not the mask of cpus it is actually running on.  This allows the
1077 * user to specify a mask that will include cpus that have not yet
1078 * been brought online, if desired.
1079 */
1080static int proc_watchdog_cpumask(const struct ctl_table *table, int write,
1081				 void *buffer, size_t *lenp, loff_t *ppos)
1082{
1083	int err;
1084
1085	mutex_lock(&watchdog_mutex);
1086
1087	err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1088	if (!err && write)
1089		proc_watchdog_update();
1090
1091	mutex_unlock(&watchdog_mutex);
1092	return err;
1093}
1094
1095static const int sixty = 60;
1096
1097static struct ctl_table watchdog_sysctls[] = {
1098	{
1099		.procname       = "watchdog",
1100		.data		= &watchdog_user_enabled,
1101		.maxlen		= sizeof(int),
1102		.mode		= 0644,
1103		.proc_handler   = proc_watchdog,
1104		.extra1		= SYSCTL_ZERO,
1105		.extra2		= SYSCTL_ONE,
1106	},
1107	{
1108		.procname	= "watchdog_thresh",
1109		.data		= &watchdog_thresh,
1110		.maxlen		= sizeof(int),
1111		.mode		= 0644,
1112		.proc_handler	= proc_watchdog_thresh,
1113		.extra1		= SYSCTL_ZERO,
1114		.extra2		= (void *)&sixty,
1115	},
1116	{
1117		.procname	= "watchdog_cpumask",
1118		.data		= &watchdog_cpumask_bits,
1119		.maxlen		= NR_CPUS,
1120		.mode		= 0644,
1121		.proc_handler	= proc_watchdog_cpumask,
1122	},
1123#ifdef CONFIG_SOFTLOCKUP_DETECTOR
1124	{
1125		.procname       = "soft_watchdog",
1126		.data		= &watchdog_softlockup_user_enabled,
1127		.maxlen		= sizeof(int),
1128		.mode		= 0644,
1129		.proc_handler   = proc_soft_watchdog,
1130		.extra1		= SYSCTL_ZERO,
1131		.extra2		= SYSCTL_ONE,
1132	},
1133	{
1134		.procname	= "softlockup_panic",
1135		.data		= &softlockup_panic,
1136		.maxlen		= sizeof(int),
1137		.mode		= 0644,
1138		.proc_handler	= proc_dointvec_minmax,
1139		.extra1		= SYSCTL_ZERO,
1140		.extra2		= SYSCTL_ONE,
1141	},
1142#ifdef CONFIG_SMP
1143	{
1144		.procname	= "softlockup_all_cpu_backtrace",
1145		.data		= &sysctl_softlockup_all_cpu_backtrace,
1146		.maxlen		= sizeof(int),
1147		.mode		= 0644,
1148		.proc_handler	= proc_dointvec_minmax,
1149		.extra1		= SYSCTL_ZERO,
1150		.extra2		= SYSCTL_ONE,
1151	},
1152#endif /* CONFIG_SMP */
1153#endif
1154#ifdef CONFIG_HARDLOCKUP_DETECTOR
1155	{
1156		.procname	= "hardlockup_panic",
1157		.data		= &hardlockup_panic,
1158		.maxlen		= sizeof(int),
1159		.mode		= 0644,
1160		.proc_handler	= proc_dointvec_minmax,
1161		.extra1		= SYSCTL_ZERO,
1162		.extra2		= SYSCTL_ONE,
1163	},
1164#ifdef CONFIG_SMP
1165	{
1166		.procname	= "hardlockup_all_cpu_backtrace",
1167		.data		= &sysctl_hardlockup_all_cpu_backtrace,
1168		.maxlen		= sizeof(int),
1169		.mode		= 0644,
1170		.proc_handler	= proc_dointvec_minmax,
1171		.extra1		= SYSCTL_ZERO,
1172		.extra2		= SYSCTL_ONE,
1173	},
1174#endif /* CONFIG_SMP */
1175#endif
 
1176};
1177
1178static struct ctl_table watchdog_hardlockup_sysctl[] = {
1179	{
1180		.procname       = "nmi_watchdog",
1181		.data		= &watchdog_hardlockup_user_enabled,
1182		.maxlen		= sizeof(int),
1183		.mode		= 0444,
1184		.proc_handler   = proc_nmi_watchdog,
1185		.extra1		= SYSCTL_ZERO,
1186		.extra2		= SYSCTL_ONE,
1187	},
 
1188};
1189
1190static void __init watchdog_sysctl_init(void)
1191{
1192	register_sysctl_init("kernel", watchdog_sysctls);
1193
1194	if (watchdog_hardlockup_available)
1195		watchdog_hardlockup_sysctl[0].mode = 0644;
1196	register_sysctl_init("kernel", watchdog_hardlockup_sysctl);
1197}
1198
1199#else
1200#define watchdog_sysctl_init() do { } while (0)
1201#endif /* CONFIG_SYSCTL */
1202
1203static void __init lockup_detector_delay_init(struct work_struct *work);
1204static bool allow_lockup_detector_init_retry __initdata;
1205
1206static struct work_struct detector_work __initdata =
1207		__WORK_INITIALIZER(detector_work, lockup_detector_delay_init);
1208
1209static void __init lockup_detector_delay_init(struct work_struct *work)
1210{
1211	int ret;
1212
1213	ret = watchdog_hardlockup_probe();
1214	if (ret) {
1215		if (ret == -ENODEV)
1216			pr_info("NMI not fully supported\n");
1217		else
1218			pr_info("Delayed init of the lockup detector failed: %d\n", ret);
1219		pr_info("Hard watchdog permanently disabled\n");
1220		return;
1221	}
1222
1223	allow_lockup_detector_init_retry = false;
1224
1225	watchdog_hardlockup_available = true;
1226	lockup_detector_setup();
1227}
1228
1229/*
1230 * lockup_detector_retry_init - retry init lockup detector if possible.
1231 *
1232 * Retry hardlockup detector init. It is useful when it requires some
1233 * functionality that has to be initialized later on a particular
1234 * platform.
1235 */
1236void __init lockup_detector_retry_init(void)
1237{
1238	/* Must be called before late init calls */
1239	if (!allow_lockup_detector_init_retry)
1240		return;
1241
1242	schedule_work(&detector_work);
1243}
1244
1245/*
1246 * Ensure that optional delayed hardlockup init is proceed before
1247 * the init code and memory is freed.
1248 */
1249static int __init lockup_detector_check(void)
1250{
1251	/* Prevent any later retry. */
1252	allow_lockup_detector_init_retry = false;
1253
1254	/* Make sure no work is pending. */
1255	flush_work(&detector_work);
1256
1257	watchdog_sysctl_init();
1258
1259	return 0;
1260
1261}
1262late_initcall_sync(lockup_detector_check);
1263
1264void __init lockup_detector_init(void)
1265{
1266	if (tick_nohz_full_enabled())
1267		pr_info("Disabling watchdog on nohz_full cores by default\n");
1268
1269	cpumask_copy(&watchdog_cpumask,
1270		     housekeeping_cpumask(HK_TYPE_TIMER));
1271
1272	if (!watchdog_hardlockup_probe())
1273		watchdog_hardlockup_available = true;
1274	else
1275		allow_lockup_detector_init_retry = true;
1276
1277	lockup_detector_setup();
1278}