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 35 static 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 45 unsigned long __read_mostly watchdog_enabled; 46 int __read_mostly watchdog_user_enabled = 1; 47 static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT; 48 static int __read_mostly watchdog_softlockup_user_enabled = 1; 49 int __read_mostly watchdog_thresh = 10; 50 static int __read_mostly watchdog_thresh_next; 51 static int __read_mostly watchdog_hardlockup_available; 52 53 struct cpumask watchdog_cpumask __read_mostly; 54 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); 55 56 #ifdef CONFIG_HARDLOCKUP_DETECTOR 57 58 # ifdef CONFIG_SMP 59 int __read_mostly sysctl_hardlockup_all_cpu_backtrace; 60 # endif /* CONFIG_SMP */ 61 62 /* 63 * Should we panic when a soft-lockup or hard-lockup occurs: 64 */ 65 unsigned int __read_mostly hardlockup_panic = 66 IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC); 67 68 #ifdef CONFIG_SYSFS 69 70 static unsigned int hardlockup_count; 71 72 static ssize_t hardlockup_count_show(struct kobject *kobj, struct kobj_attribute *attr, 73 char *page) 74 { 75 return sysfs_emit(page, "%u\n", hardlockup_count); 76 } 77 78 static struct kobj_attribute hardlockup_count_attr = __ATTR_RO(hardlockup_count); 79 80 static __init int kernel_hardlockup_sysfs_init(void) 81 { 82 sysfs_add_file_to_group(kernel_kobj, &hardlockup_count_attr.attr, NULL); 83 return 0; 84 } 85 86 late_initcall(kernel_hardlockup_sysfs_init); 87 88 #endif // CONFIG_SYSFS 89 90 /* 91 * We may not want to enable hard lockup detection by default in all cases, 92 * for example when running the kernel as a guest on a hypervisor. In these 93 * cases this function can be called to disable hard lockup detection. This 94 * function should only be executed once by the boot processor before the 95 * kernel command line parameters are parsed, because otherwise it is not 96 * possible to override this in hardlockup_panic_setup(). 97 */ 98 void __init hardlockup_detector_disable(void) 99 { 100 watchdog_hardlockup_user_enabled = 0; 101 } 102 103 static int __init hardlockup_panic_setup(char *str) 104 { 105 next: 106 if (!strncmp(str, "panic", 5)) 107 hardlockup_panic = 1; 108 else if (!strncmp(str, "nopanic", 7)) 109 hardlockup_panic = 0; 110 else if (!strncmp(str, "0", 1)) 111 watchdog_hardlockup_user_enabled = 0; 112 else if (!strncmp(str, "1", 1)) 113 watchdog_hardlockup_user_enabled = 1; 114 else if (!strncmp(str, "r", 1)) 115 hardlockup_config_perf_event(str + 1); 116 while (*(str++)) { 117 if (*str == ',') { 118 str++; 119 goto next; 120 } 121 } 122 return 1; 123 } 124 __setup("nmi_watchdog=", hardlockup_panic_setup); 125 126 #endif /* CONFIG_HARDLOCKUP_DETECTOR */ 127 128 #if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER) 129 130 static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts); 131 static DEFINE_PER_CPU(int, hrtimer_interrupts_saved); 132 static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned); 133 static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched); 134 static unsigned long hard_lockup_nmi_warn; 135 136 notrace void arch_touch_nmi_watchdog(void) 137 { 138 /* 139 * Using __raw here because some code paths have 140 * preemption enabled. If preemption is enabled 141 * then interrupts should be enabled too, in which 142 * case we shouldn't have to worry about the watchdog 143 * going off. 144 */ 145 raw_cpu_write(watchdog_hardlockup_touched, true); 146 } 147 EXPORT_SYMBOL(arch_touch_nmi_watchdog); 148 149 void watchdog_hardlockup_touch_cpu(unsigned int cpu) 150 { 151 per_cpu(watchdog_hardlockup_touched, cpu) = true; 152 } 153 154 static bool is_hardlockup(unsigned int cpu) 155 { 156 int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu)); 157 158 if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint) 159 return true; 160 161 /* 162 * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE 163 * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is 164 * written/read by a single CPU. 165 */ 166 per_cpu(hrtimer_interrupts_saved, cpu) = hrint; 167 168 return false; 169 } 170 171 static void watchdog_hardlockup_kick(void) 172 { 173 int new_interrupts; 174 175 new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts)); 176 watchdog_buddy_check_hardlockup(new_interrupts); 177 } 178 179 void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs) 180 { 181 if (per_cpu(watchdog_hardlockup_touched, cpu)) { 182 per_cpu(watchdog_hardlockup_touched, cpu) = false; 183 return; 184 } 185 186 /* 187 * Check for a hardlockup by making sure the CPU's timer 188 * interrupt is incrementing. The timer interrupt should have 189 * fired multiple times before we overflow'd. If it hasn't 190 * then this is a good indication the cpu is stuck 191 */ 192 if (is_hardlockup(cpu)) { 193 unsigned int this_cpu = smp_processor_id(); 194 unsigned long flags; 195 196 #ifdef CONFIG_SYSFS 197 ++hardlockup_count; 198 #endif 199 200 /* Only print hardlockups once. */ 201 if (per_cpu(watchdog_hardlockup_warned, cpu)) 202 return; 203 204 /* 205 * Prevent multiple hard-lockup reports if one cpu is already 206 * engaged in dumping all cpu back traces. 207 */ 208 if (sysctl_hardlockup_all_cpu_backtrace) { 209 if (test_and_set_bit_lock(0, &hard_lockup_nmi_warn)) 210 return; 211 } 212 213 /* 214 * NOTE: we call printk_cpu_sync_get_irqsave() after printing 215 * the lockup message. While it would be nice to serialize 216 * that printout, we really want to make sure that if some 217 * other CPU somehow locked up while holding the lock associated 218 * with printk_cpu_sync_get_irqsave() that we can still at least 219 * get the message about the lockup out. 220 */ 221 pr_emerg("CPU%u: Watchdog detected hard LOCKUP on cpu %u\n", this_cpu, cpu); 222 printk_cpu_sync_get_irqsave(flags); 223 224 print_modules(); 225 print_irqtrace_events(current); 226 if (cpu == this_cpu) { 227 if (regs) 228 show_regs(regs); 229 else 230 dump_stack(); 231 printk_cpu_sync_put_irqrestore(flags); 232 } else { 233 printk_cpu_sync_put_irqrestore(flags); 234 trigger_single_cpu_backtrace(cpu); 235 } 236 237 if (sysctl_hardlockup_all_cpu_backtrace) { 238 trigger_allbutcpu_cpu_backtrace(cpu); 239 if (!hardlockup_panic) 240 clear_bit_unlock(0, &hard_lockup_nmi_warn); 241 } 242 243 if (hardlockup_panic) 244 nmi_panic(regs, "Hard LOCKUP"); 245 246 per_cpu(watchdog_hardlockup_warned, cpu) = true; 247 } else { 248 per_cpu(watchdog_hardlockup_warned, cpu) = false; 249 } 250 } 251 252 #else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */ 253 254 static inline void watchdog_hardlockup_kick(void) { } 255 256 #endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */ 257 258 /* 259 * These functions can be overridden based on the configured hardlockdup detector. 260 * 261 * watchdog_hardlockup_enable/disable can be implemented to start and stop when 262 * softlockup watchdog start and stop. The detector must select the 263 * SOFTLOCKUP_DETECTOR Kconfig. 264 */ 265 void __weak watchdog_hardlockup_enable(unsigned int cpu) { } 266 267 void __weak watchdog_hardlockup_disable(unsigned int cpu) { } 268 269 /* 270 * Watchdog-detector specific API. 271 * 272 * Return 0 when hardlockup watchdog is available, negative value otherwise. 273 * Note that the negative value means that a delayed probe might 274 * succeed later. 275 */ 276 int __weak __init watchdog_hardlockup_probe(void) 277 { 278 return -ENODEV; 279 } 280 281 /** 282 * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration 283 * 284 * The reconfiguration steps are: 285 * watchdog_hardlockup_stop(); 286 * update_variables(); 287 * watchdog_hardlockup_start(); 288 */ 289 void __weak watchdog_hardlockup_stop(void) { } 290 291 /** 292 * watchdog_hardlockup_start - Start the watchdog after reconfiguration 293 * 294 * Counterpart to watchdog_hardlockup_stop(). 295 * 296 * The following variables have been updated in update_variables() and 297 * contain the currently valid configuration: 298 * - watchdog_enabled 299 * - watchdog_thresh 300 * - watchdog_cpumask 301 */ 302 void __weak watchdog_hardlockup_start(void) { } 303 304 /** 305 * lockup_detector_update_enable - Update the sysctl enable bit 306 * 307 * Caller needs to make sure that the hard watchdogs are off, so this 308 * can't race with watchdog_hardlockup_disable(). 309 */ 310 static void lockup_detector_update_enable(void) 311 { 312 watchdog_enabled = 0; 313 if (!watchdog_user_enabled) 314 return; 315 if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled) 316 watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED; 317 if (watchdog_softlockup_user_enabled) 318 watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED; 319 } 320 321 #ifdef CONFIG_SOFTLOCKUP_DETECTOR 322 323 /* 324 * Delay the soflockup report when running a known slow code. 325 * It does _not_ affect the timestamp of the last successdul reschedule. 326 */ 327 #define SOFTLOCKUP_DELAY_REPORT ULONG_MAX 328 329 #ifdef CONFIG_SMP 330 int __read_mostly sysctl_softlockup_all_cpu_backtrace; 331 #endif 332 333 static struct cpumask watchdog_allowed_mask __read_mostly; 334 335 /* Global variables, exported for sysctl */ 336 unsigned int __read_mostly softlockup_panic = 337 IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC); 338 339 static bool softlockup_initialized __read_mostly; 340 static u64 __read_mostly sample_period; 341 342 #ifdef CONFIG_SYSFS 343 344 static unsigned int softlockup_count; 345 346 static ssize_t softlockup_count_show(struct kobject *kobj, struct kobj_attribute *attr, 347 char *page) 348 { 349 return sysfs_emit(page, "%u\n", softlockup_count); 350 } 351 352 static struct kobj_attribute softlockup_count_attr = __ATTR_RO(softlockup_count); 353 354 static __init int kernel_softlockup_sysfs_init(void) 355 { 356 sysfs_add_file_to_group(kernel_kobj, &softlockup_count_attr.attr, NULL); 357 return 0; 358 } 359 360 late_initcall(kernel_softlockup_sysfs_init); 361 362 #endif // CONFIG_SYSFS 363 364 /* Timestamp taken after the last successful reschedule. */ 365 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); 366 /* Timestamp of the last softlockup report. */ 367 static DEFINE_PER_CPU(unsigned long, watchdog_report_ts); 368 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); 369 static DEFINE_PER_CPU(bool, softlockup_touch_sync); 370 static unsigned long soft_lockup_nmi_warn; 371 372 static int __init softlockup_panic_setup(char *str) 373 { 374 softlockup_panic = simple_strtoul(str, NULL, 0); 375 return 1; 376 } 377 __setup("softlockup_panic=", softlockup_panic_setup); 378 379 static int __init nowatchdog_setup(char *str) 380 { 381 watchdog_user_enabled = 0; 382 return 1; 383 } 384 __setup("nowatchdog", nowatchdog_setup); 385 386 static int __init nosoftlockup_setup(char *str) 387 { 388 watchdog_softlockup_user_enabled = 0; 389 return 1; 390 } 391 __setup("nosoftlockup", nosoftlockup_setup); 392 393 static int __init watchdog_thresh_setup(char *str) 394 { 395 get_option(&str, &watchdog_thresh); 396 return 1; 397 } 398 __setup("watchdog_thresh=", watchdog_thresh_setup); 399 400 #ifdef CONFIG_SOFTLOCKUP_DETECTOR_INTR_STORM 401 enum stats_per_group { 402 STATS_SYSTEM, 403 STATS_SOFTIRQ, 404 STATS_HARDIRQ, 405 STATS_IDLE, 406 NUM_STATS_PER_GROUP, 407 }; 408 409 static const enum cpu_usage_stat tracked_stats[NUM_STATS_PER_GROUP] = { 410 CPUTIME_SYSTEM, 411 CPUTIME_SOFTIRQ, 412 CPUTIME_IRQ, 413 CPUTIME_IDLE, 414 }; 415 416 static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]); 417 static DEFINE_PER_CPU(u8, cpustat_util[NUM_SAMPLE_PERIODS][NUM_STATS_PER_GROUP]); 418 static DEFINE_PER_CPU(u8, cpustat_tail); 419 420 /* 421 * We don't need nanosecond resolution. A granularity of 16ms is 422 * sufficient for our precision, allowing us to use u16 to store 423 * cpustats, which will roll over roughly every ~1000 seconds. 424 * 2^24 ~= 16 * 10^6 425 */ 426 static u16 get_16bit_precision(u64 data_ns) 427 { 428 /* 429 * 2^24ns ~= 16.8ms 430 * Round to the nearest multiple of 16.8 milliseconds. 431 */ 432 return (data_ns + (1 << 23)) >> 24LL; 433 } 434 435 static void update_cpustat(void) 436 { 437 int i; 438 u8 util; 439 u16 old_stat, new_stat; 440 struct kernel_cpustat kcpustat; 441 u64 *cpustat = kcpustat.cpustat; 442 u8 tail = __this_cpu_read(cpustat_tail); 443 u16 sample_period_16 = get_16bit_precision(sample_period); 444 445 kcpustat_cpu_fetch(&kcpustat, smp_processor_id()); 446 447 for (i = 0; i < NUM_STATS_PER_GROUP; i++) { 448 old_stat = __this_cpu_read(cpustat_old[i]); 449 new_stat = get_16bit_precision(cpustat[tracked_stats[i]]); 450 util = DIV_ROUND_UP(100 * (new_stat - old_stat), sample_period_16); 451 /* 452 * Since we use 16-bit precision, the raw data will undergo 453 * integer division, which may sometimes result in data loss, 454 * and then result might exceed 100%. To avoid confusion, 455 * we enforce a 100% display cap when calculations exceed this threshold. 456 */ 457 if (util > 100) 458 util = 100; 459 __this_cpu_write(cpustat_util[tail][i], util); 460 __this_cpu_write(cpustat_old[i], new_stat); 461 } 462 463 __this_cpu_write(cpustat_tail, (tail + 1) % NUM_SAMPLE_PERIODS); 464 } 465 466 static void print_cpustat(void) 467 { 468 int i, group; 469 u8 tail = __this_cpu_read(cpustat_tail); 470 u64 sample_period_msecond = sample_period; 471 472 do_div(sample_period_msecond, NSEC_PER_MSEC); 473 474 /* 475 * Outputting the "watchdog" prefix on every line is redundant and not 476 * concise, and the original alarm information is sufficient for 477 * positioning in logs, hence here printk() is used instead of pr_crit(). 478 */ 479 printk(KERN_CRIT "CPU#%d Utilization every %llums during lockup:\n", 480 smp_processor_id(), sample_period_msecond); 481 482 for (i = 0; i < NUM_SAMPLE_PERIODS; i++) { 483 group = (tail + i) % NUM_SAMPLE_PERIODS; 484 printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t" 485 "%3u%% hardirq,\t%3u%% idle\n", i + 1, 486 __this_cpu_read(cpustat_util[group][STATS_SYSTEM]), 487 __this_cpu_read(cpustat_util[group][STATS_SOFTIRQ]), 488 __this_cpu_read(cpustat_util[group][STATS_HARDIRQ]), 489 __this_cpu_read(cpustat_util[group][STATS_IDLE])); 490 } 491 } 492 493 #define HARDIRQ_PERCENT_THRESH 50 494 #define NUM_HARDIRQ_REPORT 5 495 struct irq_counts { 496 int irq; 497 u32 counts; 498 }; 499 500 static DEFINE_PER_CPU(bool, snapshot_taken); 501 502 /* Tabulate the most frequent interrupts. */ 503 static void tabulate_irq_count(struct irq_counts *irq_counts, int irq, u32 counts, int rank) 504 { 505 int i; 506 struct irq_counts new_count = {irq, counts}; 507 508 for (i = 0; i < rank; i++) { 509 if (counts > irq_counts[i].counts) 510 swap(new_count, irq_counts[i]); 511 } 512 } 513 514 /* 515 * If the hardirq time exceeds HARDIRQ_PERCENT_THRESH% of the sample_period, 516 * then the cause of softlockup might be interrupt storm. In this case, it 517 * would be useful to start interrupt counting. 518 */ 519 static bool need_counting_irqs(void) 520 { 521 u8 util; 522 int tail = __this_cpu_read(cpustat_tail); 523 524 tail = (tail + NUM_HARDIRQ_REPORT - 1) % NUM_HARDIRQ_REPORT; 525 util = __this_cpu_read(cpustat_util[tail][STATS_HARDIRQ]); 526 return util > HARDIRQ_PERCENT_THRESH; 527 } 528 529 static void start_counting_irqs(void) 530 { 531 if (!__this_cpu_read(snapshot_taken)) { 532 kstat_snapshot_irqs(); 533 __this_cpu_write(snapshot_taken, true); 534 } 535 } 536 537 static void stop_counting_irqs(void) 538 { 539 __this_cpu_write(snapshot_taken, false); 540 } 541 542 static void print_irq_counts(void) 543 { 544 unsigned int i, count; 545 struct irq_counts irq_counts_sorted[NUM_HARDIRQ_REPORT] = { 546 {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0} 547 }; 548 549 if (__this_cpu_read(snapshot_taken)) { 550 for_each_active_irq(i) { 551 count = kstat_get_irq_since_snapshot(i); 552 tabulate_irq_count(irq_counts_sorted, i, count, NUM_HARDIRQ_REPORT); 553 } 554 555 /* 556 * Outputting the "watchdog" prefix on every line is redundant and not 557 * concise, and the original alarm information is sufficient for 558 * positioning in logs, hence here printk() is used instead of pr_crit(). 559 */ 560 printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n", 561 smp_processor_id(), HARDIRQ_PERCENT_THRESH); 562 563 for (i = 0; i < NUM_HARDIRQ_REPORT; i++) { 564 if (irq_counts_sorted[i].irq == -1) 565 break; 566 567 printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n", 568 i + 1, irq_counts_sorted[i].counts, 569 irq_counts_sorted[i].irq); 570 } 571 572 /* 573 * If the hardirq time is less than HARDIRQ_PERCENT_THRESH% in the last 574 * sample_period, then we suspect the interrupt storm might be subsiding. 575 */ 576 if (!need_counting_irqs()) 577 stop_counting_irqs(); 578 } 579 } 580 581 static void report_cpu_status(void) 582 { 583 print_cpustat(); 584 print_irq_counts(); 585 } 586 #else 587 static inline void update_cpustat(void) { } 588 static inline void report_cpu_status(void) { } 589 static inline bool need_counting_irqs(void) { return false; } 590 static inline void start_counting_irqs(void) { } 591 static inline void stop_counting_irqs(void) { } 592 #endif 593 594 /* 595 * Hard-lockup warnings should be triggered after just a few seconds. Soft- 596 * lockups can have false positives under extreme conditions. So we generally 597 * want a higher threshold for soft lockups than for hard lockups. So we couple 598 * the thresholds with a factor: we make the soft threshold twice the amount of 599 * time the hard threshold is. 600 */ 601 static int get_softlockup_thresh(void) 602 { 603 return watchdog_thresh * 2; 604 } 605 606 /* 607 * Returns seconds, approximately. We don't need nanosecond 608 * resolution, and we don't need to waste time with a big divide when 609 * 2^30ns == 1.074s. 610 */ 611 static unsigned long get_timestamp(void) 612 { 613 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ 614 } 615 616 static void set_sample_period(void) 617 { 618 /* 619 * convert watchdog_thresh from seconds to ns 620 * the divide by 5 is to give hrtimer several chances (two 621 * or three with the current relation between the soft 622 * and hard thresholds) to increment before the 623 * hardlockup detector generates a warning 624 */ 625 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / NUM_SAMPLE_PERIODS); 626 watchdog_update_hrtimer_threshold(sample_period); 627 } 628 629 static void update_report_ts(void) 630 { 631 __this_cpu_write(watchdog_report_ts, get_timestamp()); 632 } 633 634 /* Commands for resetting the watchdog */ 635 static void update_touch_ts(void) 636 { 637 __this_cpu_write(watchdog_touch_ts, get_timestamp()); 638 update_report_ts(); 639 } 640 641 /** 642 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls 643 * 644 * Call when the scheduler may have stalled for legitimate reasons 645 * preventing the watchdog task from executing - e.g. the scheduler 646 * entering idle state. This should only be used for scheduler events. 647 * Use touch_softlockup_watchdog() for everything else. 648 */ 649 notrace void touch_softlockup_watchdog_sched(void) 650 { 651 /* 652 * Preemption can be enabled. It doesn't matter which CPU's watchdog 653 * report period gets restarted here, so use the raw_ operation. 654 */ 655 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT); 656 } 657 658 notrace void touch_softlockup_watchdog(void) 659 { 660 touch_softlockup_watchdog_sched(); 661 wq_watchdog_touch(raw_smp_processor_id()); 662 } 663 EXPORT_SYMBOL(touch_softlockup_watchdog); 664 665 void touch_all_softlockup_watchdogs(void) 666 { 667 int cpu; 668 669 /* 670 * watchdog_mutex cannpt be taken here, as this might be called 671 * from (soft)interrupt context, so the access to 672 * watchdog_allowed_cpumask might race with a concurrent update. 673 * 674 * The watchdog time stamp can race against a concurrent real 675 * update as well, the only side effect might be a cycle delay for 676 * the softlockup check. 677 */ 678 for_each_cpu(cpu, &watchdog_allowed_mask) { 679 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT; 680 wq_watchdog_touch(cpu); 681 } 682 } 683 684 void touch_softlockup_watchdog_sync(void) 685 { 686 __this_cpu_write(softlockup_touch_sync, true); 687 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT); 688 } 689 690 static int is_softlockup(unsigned long touch_ts, 691 unsigned long period_ts, 692 unsigned long now) 693 { 694 if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) { 695 /* 696 * If period_ts has not been updated during a sample_period, then 697 * in the subsequent few sample_periods, period_ts might also not 698 * be updated, which could indicate a potential softlockup. In 699 * this case, if we suspect the cause of the potential softlockup 700 * might be interrupt storm, then we need to count the interrupts 701 * to find which interrupt is storming. 702 */ 703 if (time_after_eq(now, period_ts + get_softlockup_thresh() / NUM_SAMPLE_PERIODS) && 704 need_counting_irqs()) 705 start_counting_irqs(); 706 707 /* 708 * A poorly behaving BPF scheduler can live-lock the system into 709 * soft lockups. Tell sched_ext to try ejecting the BPF 710 * scheduler when close to a soft lockup. 711 */ 712 if (time_after_eq(now, period_ts + get_softlockup_thresh() * 3 / 4)) 713 scx_softlockup(now - touch_ts); 714 715 /* Warn about unreasonable delays. */ 716 if (time_after(now, period_ts + get_softlockup_thresh())) 717 return now - touch_ts; 718 } 719 return 0; 720 } 721 722 /* watchdog detector functions */ 723 static DEFINE_PER_CPU(struct completion, softlockup_completion); 724 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work); 725 726 /* 727 * The watchdog feed function - touches the timestamp. 728 * 729 * It only runs once every sample_period seconds (4 seconds by 730 * default) to reset the softlockup timestamp. If this gets delayed 731 * for more than 2*watchdog_thresh seconds then the debug-printout 732 * triggers in watchdog_timer_fn(). 733 */ 734 static int softlockup_fn(void *data) 735 { 736 update_touch_ts(); 737 stop_counting_irqs(); 738 complete(this_cpu_ptr(&softlockup_completion)); 739 740 return 0; 741 } 742 743 /* watchdog kicker functions */ 744 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) 745 { 746 unsigned long touch_ts, period_ts, now; 747 struct pt_regs *regs = get_irq_regs(); 748 int duration; 749 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; 750 unsigned long flags; 751 752 if (!watchdog_enabled) 753 return HRTIMER_NORESTART; 754 755 watchdog_hardlockup_kick(); 756 757 /* kick the softlockup detector */ 758 if (completion_done(this_cpu_ptr(&softlockup_completion))) { 759 reinit_completion(this_cpu_ptr(&softlockup_completion)); 760 stop_one_cpu_nowait(smp_processor_id(), 761 softlockup_fn, NULL, 762 this_cpu_ptr(&softlockup_stop_work)); 763 } 764 765 /* .. and repeat */ 766 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); 767 768 /* 769 * Read the current timestamp first. It might become invalid anytime 770 * when a virtual machine is stopped by the host or when the watchog 771 * is touched from NMI. 772 */ 773 now = get_timestamp(); 774 /* 775 * If a virtual machine is stopped by the host it can look to 776 * the watchdog like a soft lockup. This function touches the watchdog. 777 */ 778 kvm_check_and_clear_guest_paused(); 779 /* 780 * The stored timestamp is comparable with @now only when not touched. 781 * It might get touched anytime from NMI. Make sure that is_softlockup() 782 * uses the same (valid) value. 783 */ 784 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts)); 785 786 update_cpustat(); 787 788 /* Reset the interval when touched by known problematic code. */ 789 if (period_ts == SOFTLOCKUP_DELAY_REPORT) { 790 if (unlikely(__this_cpu_read(softlockup_touch_sync))) { 791 /* 792 * If the time stamp was touched atomically 793 * make sure the scheduler tick is up to date. 794 */ 795 __this_cpu_write(softlockup_touch_sync, false); 796 sched_clock_tick(); 797 } 798 799 update_report_ts(); 800 return HRTIMER_RESTART; 801 } 802 803 /* Check for a softlockup. */ 804 touch_ts = __this_cpu_read(watchdog_touch_ts); 805 duration = is_softlockup(touch_ts, period_ts, now); 806 if (unlikely(duration)) { 807 #ifdef CONFIG_SYSFS 808 ++softlockup_count; 809 #endif 810 811 /* 812 * Prevent multiple soft-lockup reports if one cpu is already 813 * engaged in dumping all cpu back traces. 814 */ 815 if (softlockup_all_cpu_backtrace) { 816 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn)) 817 return HRTIMER_RESTART; 818 } 819 820 /* Start period for the next softlockup warning. */ 821 update_report_ts(); 822 823 printk_cpu_sync_get_irqsave(flags); 824 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", 825 smp_processor_id(), duration, 826 current->comm, task_pid_nr(current)); 827 report_cpu_status(); 828 print_modules(); 829 print_irqtrace_events(current); 830 if (regs) 831 show_regs(regs); 832 else 833 dump_stack(); 834 printk_cpu_sync_put_irqrestore(flags); 835 836 if (softlockup_all_cpu_backtrace) { 837 trigger_allbutcpu_cpu_backtrace(smp_processor_id()); 838 if (!softlockup_panic) 839 clear_bit_unlock(0, &soft_lockup_nmi_warn); 840 } 841 842 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); 843 if (softlockup_panic) 844 panic("softlockup: hung tasks"); 845 } 846 847 return HRTIMER_RESTART; 848 } 849 850 static void watchdog_enable(unsigned int cpu) 851 { 852 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); 853 struct completion *done = this_cpu_ptr(&softlockup_completion); 854 855 WARN_ON_ONCE(cpu != smp_processor_id()); 856 857 init_completion(done); 858 complete(done); 859 860 /* 861 * Start the timer first to prevent the hardlockup watchdog triggering 862 * before the timer has a chance to fire. 863 */ 864 hrtimer_setup(hrtimer, watchdog_timer_fn, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); 865 hrtimer_start(hrtimer, ns_to_ktime(sample_period), 866 HRTIMER_MODE_REL_PINNED_HARD); 867 868 /* Initialize timestamp */ 869 update_touch_ts(); 870 /* Enable the hardlockup detector */ 871 if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED) 872 watchdog_hardlockup_enable(cpu); 873 } 874 875 static void watchdog_disable(unsigned int cpu) 876 { 877 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); 878 879 WARN_ON_ONCE(cpu != smp_processor_id()); 880 881 /* 882 * Disable the hardlockup detector first. That prevents that a large 883 * delay between disabling the timer and disabling the hardlockup 884 * detector causes a false positive. 885 */ 886 watchdog_hardlockup_disable(cpu); 887 hrtimer_cancel(hrtimer); 888 wait_for_completion(this_cpu_ptr(&softlockup_completion)); 889 } 890 891 static int softlockup_stop_fn(void *data) 892 { 893 watchdog_disable(smp_processor_id()); 894 return 0; 895 } 896 897 static void softlockup_stop_all(void) 898 { 899 int cpu; 900 901 if (!softlockup_initialized) 902 return; 903 904 for_each_cpu(cpu, &watchdog_allowed_mask) 905 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false); 906 907 cpumask_clear(&watchdog_allowed_mask); 908 } 909 910 static int softlockup_start_fn(void *data) 911 { 912 watchdog_enable(smp_processor_id()); 913 return 0; 914 } 915 916 static void softlockup_start_all(void) 917 { 918 int cpu; 919 920 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); 921 for_each_cpu(cpu, &watchdog_allowed_mask) 922 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false); 923 } 924 925 int lockup_detector_online_cpu(unsigned int cpu) 926 { 927 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask)) 928 watchdog_enable(cpu); 929 return 0; 930 } 931 932 int lockup_detector_offline_cpu(unsigned int cpu) 933 { 934 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask)) 935 watchdog_disable(cpu); 936 return 0; 937 } 938 939 static void __lockup_detector_reconfigure(bool thresh_changed) 940 { 941 cpus_read_lock(); 942 watchdog_hardlockup_stop(); 943 944 softlockup_stop_all(); 945 /* 946 * To prevent watchdog_timer_fn from using the old interval and 947 * the new watchdog_thresh at the same time, which could lead to 948 * false softlockup reports, it is necessary to update the 949 * watchdog_thresh after the softlockup is completed. 950 */ 951 if (thresh_changed) 952 watchdog_thresh = READ_ONCE(watchdog_thresh_next); 953 set_sample_period(); 954 lockup_detector_update_enable(); 955 if (watchdog_enabled && watchdog_thresh) 956 softlockup_start_all(); 957 958 watchdog_hardlockup_start(); 959 cpus_read_unlock(); 960 } 961 962 void lockup_detector_reconfigure(void) 963 { 964 mutex_lock(&watchdog_mutex); 965 __lockup_detector_reconfigure(false); 966 mutex_unlock(&watchdog_mutex); 967 } 968 969 /* 970 * Create the watchdog infrastructure and configure the detector(s). 971 */ 972 static __init void lockup_detector_setup(void) 973 { 974 /* 975 * If sysctl is off and watchdog got disabled on the command line, 976 * nothing to do here. 977 */ 978 lockup_detector_update_enable(); 979 980 if (!IS_ENABLED(CONFIG_SYSCTL) && 981 !(watchdog_enabled && watchdog_thresh)) 982 return; 983 984 mutex_lock(&watchdog_mutex); 985 __lockup_detector_reconfigure(false); 986 softlockup_initialized = true; 987 mutex_unlock(&watchdog_mutex); 988 } 989 990 #else /* CONFIG_SOFTLOCKUP_DETECTOR */ 991 static void __lockup_detector_reconfigure(bool thresh_changed) 992 { 993 cpus_read_lock(); 994 watchdog_hardlockup_stop(); 995 if (thresh_changed) 996 watchdog_thresh = READ_ONCE(watchdog_thresh_next); 997 lockup_detector_update_enable(); 998 watchdog_hardlockup_start(); 999 cpus_read_unlock(); 1000 } 1001 void lockup_detector_reconfigure(void) 1002 { 1003 __lockup_detector_reconfigure(false); 1004 } 1005 static inline void lockup_detector_setup(void) 1006 { 1007 __lockup_detector_reconfigure(false); 1008 } 1009 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ 1010 1011 /** 1012 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) 1013 * 1014 * Special interface for parisc. It prevents lockup detector warnings from 1015 * the default pm_poweroff() function which busy loops forever. 1016 */ 1017 void lockup_detector_soft_poweroff(void) 1018 { 1019 watchdog_enabled = 0; 1020 } 1021 1022 #ifdef CONFIG_SYSCTL 1023 1024 /* Propagate any changes to the watchdog infrastructure */ 1025 static void proc_watchdog_update(bool thresh_changed) 1026 { 1027 /* Remove impossible cpus to keep sysctl output clean. */ 1028 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); 1029 __lockup_detector_reconfigure(thresh_changed); 1030 } 1031 1032 /* 1033 * common function for watchdog, nmi_watchdog and soft_watchdog parameter 1034 * 1035 * caller | table->data points to | 'which' 1036 * -------------------|----------------------------------|------------------------------- 1037 * proc_watchdog | watchdog_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED | 1038 * | | WATCHDOG_SOFTOCKUP_ENABLED 1039 * -------------------|----------------------------------|------------------------------- 1040 * proc_nmi_watchdog | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED 1041 * -------------------|----------------------------------|------------------------------- 1042 * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED 1043 */ 1044 static int proc_watchdog_common(int which, const struct ctl_table *table, int write, 1045 void *buffer, size_t *lenp, loff_t *ppos) 1046 { 1047 int err, old, *param = table->data; 1048 1049 mutex_lock(&watchdog_mutex); 1050 1051 old = *param; 1052 if (!write) { 1053 /* 1054 * On read synchronize the userspace interface. This is a 1055 * racy snapshot. 1056 */ 1057 *param = (watchdog_enabled & which) != 0; 1058 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1059 *param = old; 1060 } else { 1061 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1062 if (!err && old != READ_ONCE(*param)) 1063 proc_watchdog_update(false); 1064 } 1065 mutex_unlock(&watchdog_mutex); 1066 return err; 1067 } 1068 1069 /* 1070 * /proc/sys/kernel/watchdog 1071 */ 1072 static int proc_watchdog(const struct ctl_table *table, int write, 1073 void *buffer, size_t *lenp, loff_t *ppos) 1074 { 1075 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED | 1076 WATCHDOG_SOFTOCKUP_ENABLED, 1077 table, write, buffer, lenp, ppos); 1078 } 1079 1080 /* 1081 * /proc/sys/kernel/nmi_watchdog 1082 */ 1083 static int proc_nmi_watchdog(const struct ctl_table *table, int write, 1084 void *buffer, size_t *lenp, loff_t *ppos) 1085 { 1086 if (!watchdog_hardlockup_available && write) 1087 return -ENOTSUPP; 1088 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED, 1089 table, write, buffer, lenp, ppos); 1090 } 1091 1092 #ifdef CONFIG_SOFTLOCKUP_DETECTOR 1093 /* 1094 * /proc/sys/kernel/soft_watchdog 1095 */ 1096 static int proc_soft_watchdog(const struct ctl_table *table, int write, 1097 void *buffer, size_t *lenp, loff_t *ppos) 1098 { 1099 return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED, 1100 table, write, buffer, lenp, ppos); 1101 } 1102 #endif 1103 1104 /* 1105 * /proc/sys/kernel/watchdog_thresh 1106 */ 1107 static int proc_watchdog_thresh(const struct ctl_table *table, int write, 1108 void *buffer, size_t *lenp, loff_t *ppos) 1109 { 1110 int err, old; 1111 1112 mutex_lock(&watchdog_mutex); 1113 1114 watchdog_thresh_next = READ_ONCE(watchdog_thresh); 1115 1116 old = watchdog_thresh_next; 1117 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1118 1119 if (!err && write && old != READ_ONCE(watchdog_thresh_next)) 1120 proc_watchdog_update(true); 1121 1122 mutex_unlock(&watchdog_mutex); 1123 return err; 1124 } 1125 1126 /* 1127 * The cpumask is the mask of possible cpus that the watchdog can run 1128 * on, not the mask of cpus it is actually running on. This allows the 1129 * user to specify a mask that will include cpus that have not yet 1130 * been brought online, if desired. 1131 */ 1132 static int proc_watchdog_cpumask(const struct ctl_table *table, int write, 1133 void *buffer, size_t *lenp, loff_t *ppos) 1134 { 1135 int err; 1136 1137 mutex_lock(&watchdog_mutex); 1138 1139 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); 1140 if (!err && write) 1141 proc_watchdog_update(false); 1142 1143 mutex_unlock(&watchdog_mutex); 1144 return err; 1145 } 1146 1147 static const int sixty = 60; 1148 1149 static const struct ctl_table watchdog_sysctls[] = { 1150 { 1151 .procname = "watchdog", 1152 .data = &watchdog_user_enabled, 1153 .maxlen = sizeof(int), 1154 .mode = 0644, 1155 .proc_handler = proc_watchdog, 1156 .extra1 = SYSCTL_ZERO, 1157 .extra2 = SYSCTL_ONE, 1158 }, 1159 { 1160 .procname = "watchdog_thresh", 1161 .data = &watchdog_thresh_next, 1162 .maxlen = sizeof(int), 1163 .mode = 0644, 1164 .proc_handler = proc_watchdog_thresh, 1165 .extra1 = SYSCTL_ZERO, 1166 .extra2 = (void *)&sixty, 1167 }, 1168 { 1169 .procname = "watchdog_cpumask", 1170 .data = &watchdog_cpumask_bits, 1171 .maxlen = NR_CPUS, 1172 .mode = 0644, 1173 .proc_handler = proc_watchdog_cpumask, 1174 }, 1175 #ifdef CONFIG_SOFTLOCKUP_DETECTOR 1176 { 1177 .procname = "soft_watchdog", 1178 .data = &watchdog_softlockup_user_enabled, 1179 .maxlen = sizeof(int), 1180 .mode = 0644, 1181 .proc_handler = proc_soft_watchdog, 1182 .extra1 = SYSCTL_ZERO, 1183 .extra2 = SYSCTL_ONE, 1184 }, 1185 { 1186 .procname = "softlockup_panic", 1187 .data = &softlockup_panic, 1188 .maxlen = sizeof(int), 1189 .mode = 0644, 1190 .proc_handler = proc_dointvec_minmax, 1191 .extra1 = SYSCTL_ZERO, 1192 .extra2 = SYSCTL_ONE, 1193 }, 1194 #ifdef CONFIG_SMP 1195 { 1196 .procname = "softlockup_all_cpu_backtrace", 1197 .data = &sysctl_softlockup_all_cpu_backtrace, 1198 .maxlen = sizeof(int), 1199 .mode = 0644, 1200 .proc_handler = proc_dointvec_minmax, 1201 .extra1 = SYSCTL_ZERO, 1202 .extra2 = SYSCTL_ONE, 1203 }, 1204 #endif /* CONFIG_SMP */ 1205 #endif 1206 #ifdef CONFIG_HARDLOCKUP_DETECTOR 1207 { 1208 .procname = "hardlockup_panic", 1209 .data = &hardlockup_panic, 1210 .maxlen = sizeof(int), 1211 .mode = 0644, 1212 .proc_handler = proc_dointvec_minmax, 1213 .extra1 = SYSCTL_ZERO, 1214 .extra2 = SYSCTL_ONE, 1215 }, 1216 #ifdef CONFIG_SMP 1217 { 1218 .procname = "hardlockup_all_cpu_backtrace", 1219 .data = &sysctl_hardlockup_all_cpu_backtrace, 1220 .maxlen = sizeof(int), 1221 .mode = 0644, 1222 .proc_handler = proc_dointvec_minmax, 1223 .extra1 = SYSCTL_ZERO, 1224 .extra2 = SYSCTL_ONE, 1225 }, 1226 #endif /* CONFIG_SMP */ 1227 #endif 1228 }; 1229 1230 static struct ctl_table watchdog_hardlockup_sysctl[] = { 1231 { 1232 .procname = "nmi_watchdog", 1233 .data = &watchdog_hardlockup_user_enabled, 1234 .maxlen = sizeof(int), 1235 .mode = 0444, 1236 .proc_handler = proc_nmi_watchdog, 1237 .extra1 = SYSCTL_ZERO, 1238 .extra2 = SYSCTL_ONE, 1239 }, 1240 }; 1241 1242 static void __init watchdog_sysctl_init(void) 1243 { 1244 register_sysctl_init("kernel", watchdog_sysctls); 1245 1246 if (watchdog_hardlockup_available) 1247 watchdog_hardlockup_sysctl[0].mode = 0644; 1248 register_sysctl_init("kernel", watchdog_hardlockup_sysctl); 1249 } 1250 1251 #else 1252 #define watchdog_sysctl_init() do { } while (0) 1253 #endif /* CONFIG_SYSCTL */ 1254 1255 static void __init lockup_detector_delay_init(struct work_struct *work); 1256 static bool allow_lockup_detector_init_retry __initdata; 1257 1258 static struct work_struct detector_work __initdata = 1259 __WORK_INITIALIZER(detector_work, lockup_detector_delay_init); 1260 1261 static void __init lockup_detector_delay_init(struct work_struct *work) 1262 { 1263 int ret; 1264 1265 ret = watchdog_hardlockup_probe(); 1266 if (ret) { 1267 if (ret == -ENODEV) 1268 pr_info("NMI not fully supported\n"); 1269 else 1270 pr_info("Delayed init of the lockup detector failed: %d\n", ret); 1271 pr_info("Hard watchdog permanently disabled\n"); 1272 return; 1273 } 1274 1275 allow_lockup_detector_init_retry = false; 1276 1277 watchdog_hardlockup_available = true; 1278 lockup_detector_setup(); 1279 } 1280 1281 /* 1282 * lockup_detector_retry_init - retry init lockup detector if possible. 1283 * 1284 * Retry hardlockup detector init. It is useful when it requires some 1285 * functionality that has to be initialized later on a particular 1286 * platform. 1287 */ 1288 void __init lockup_detector_retry_init(void) 1289 { 1290 /* Must be called before late init calls */ 1291 if (!allow_lockup_detector_init_retry) 1292 return; 1293 1294 schedule_work(&detector_work); 1295 } 1296 1297 /* 1298 * Ensure that optional delayed hardlockup init is proceed before 1299 * the init code and memory is freed. 1300 */ 1301 static int __init lockup_detector_check(void) 1302 { 1303 /* Prevent any later retry. */ 1304 allow_lockup_detector_init_retry = false; 1305 1306 /* Make sure no work is pending. */ 1307 flush_work(&detector_work); 1308 1309 watchdog_sysctl_init(); 1310 1311 return 0; 1312 1313 } 1314 late_initcall_sync(lockup_detector_check); 1315 1316 void __init lockup_detector_init(void) 1317 { 1318 if (tick_nohz_full_enabled()) 1319 pr_info("Disabling watchdog on nohz_full cores by default\n"); 1320 1321 cpumask_copy(&watchdog_cpumask, 1322 housekeeping_cpumask(HK_TYPE_TIMER)); 1323 1324 if (!watchdog_hardlockup_probe()) 1325 watchdog_hardlockup_available = true; 1326 else 1327 allow_lockup_detector_init_retry = true; 1328 1329 lockup_detector_setup(); 1330 } 1331