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 return data_ns >> 24LL; /* 2^24ns ~= 16.8ms */ 429 } 430 431 static void update_cpustat(void) 432 { 433 int i; 434 u8 util; 435 u16 old_stat, new_stat; 436 struct kernel_cpustat kcpustat; 437 u64 *cpustat = kcpustat.cpustat; 438 u8 tail = __this_cpu_read(cpustat_tail); 439 u16 sample_period_16 = get_16bit_precision(sample_period); 440 441 kcpustat_cpu_fetch(&kcpustat, smp_processor_id()); 442 443 for (i = 0; i < NUM_STATS_PER_GROUP; i++) { 444 old_stat = __this_cpu_read(cpustat_old[i]); 445 new_stat = get_16bit_precision(cpustat[tracked_stats[i]]); 446 util = DIV_ROUND_UP(100 * (new_stat - old_stat), sample_period_16); 447 __this_cpu_write(cpustat_util[tail][i], util); 448 __this_cpu_write(cpustat_old[i], new_stat); 449 } 450 451 __this_cpu_write(cpustat_tail, (tail + 1) % NUM_SAMPLE_PERIODS); 452 } 453 454 static void print_cpustat(void) 455 { 456 int i, group; 457 u8 tail = __this_cpu_read(cpustat_tail); 458 u64 sample_period_second = sample_period; 459 460 do_div(sample_period_second, NSEC_PER_SEC); 461 462 /* 463 * Outputting the "watchdog" prefix on every line is redundant and not 464 * concise, and the original alarm information is sufficient for 465 * positioning in logs, hence here printk() is used instead of pr_crit(). 466 */ 467 printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n", 468 smp_processor_id(), sample_period_second); 469 470 for (i = 0; i < NUM_SAMPLE_PERIODS; i++) { 471 group = (tail + i) % NUM_SAMPLE_PERIODS; 472 printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t" 473 "%3u%% hardirq,\t%3u%% idle\n", i + 1, 474 __this_cpu_read(cpustat_util[group][STATS_SYSTEM]), 475 __this_cpu_read(cpustat_util[group][STATS_SOFTIRQ]), 476 __this_cpu_read(cpustat_util[group][STATS_HARDIRQ]), 477 __this_cpu_read(cpustat_util[group][STATS_IDLE])); 478 } 479 } 480 481 #define HARDIRQ_PERCENT_THRESH 50 482 #define NUM_HARDIRQ_REPORT 5 483 struct irq_counts { 484 int irq; 485 u32 counts; 486 }; 487 488 static DEFINE_PER_CPU(bool, snapshot_taken); 489 490 /* Tabulate the most frequent interrupts. */ 491 static void tabulate_irq_count(struct irq_counts *irq_counts, int irq, u32 counts, int rank) 492 { 493 int i; 494 struct irq_counts new_count = {irq, counts}; 495 496 for (i = 0; i < rank; i++) { 497 if (counts > irq_counts[i].counts) 498 swap(new_count, irq_counts[i]); 499 } 500 } 501 502 /* 503 * If the hardirq time exceeds HARDIRQ_PERCENT_THRESH% of the sample_period, 504 * then the cause of softlockup might be interrupt storm. In this case, it 505 * would be useful to start interrupt counting. 506 */ 507 static bool need_counting_irqs(void) 508 { 509 u8 util; 510 int tail = __this_cpu_read(cpustat_tail); 511 512 tail = (tail + NUM_HARDIRQ_REPORT - 1) % NUM_HARDIRQ_REPORT; 513 util = __this_cpu_read(cpustat_util[tail][STATS_HARDIRQ]); 514 return util > HARDIRQ_PERCENT_THRESH; 515 } 516 517 static void start_counting_irqs(void) 518 { 519 if (!__this_cpu_read(snapshot_taken)) { 520 kstat_snapshot_irqs(); 521 __this_cpu_write(snapshot_taken, true); 522 } 523 } 524 525 static void stop_counting_irqs(void) 526 { 527 __this_cpu_write(snapshot_taken, false); 528 } 529 530 static void print_irq_counts(void) 531 { 532 unsigned int i, count; 533 struct irq_counts irq_counts_sorted[NUM_HARDIRQ_REPORT] = { 534 {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0} 535 }; 536 537 if (__this_cpu_read(snapshot_taken)) { 538 for_each_active_irq(i) { 539 count = kstat_get_irq_since_snapshot(i); 540 tabulate_irq_count(irq_counts_sorted, i, count, NUM_HARDIRQ_REPORT); 541 } 542 543 /* 544 * Outputting the "watchdog" prefix on every line is redundant and not 545 * concise, and the original alarm information is sufficient for 546 * positioning in logs, hence here printk() is used instead of pr_crit(). 547 */ 548 printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n", 549 smp_processor_id(), HARDIRQ_PERCENT_THRESH); 550 551 for (i = 0; i < NUM_HARDIRQ_REPORT; i++) { 552 if (irq_counts_sorted[i].irq == -1) 553 break; 554 555 printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n", 556 i + 1, irq_counts_sorted[i].counts, 557 irq_counts_sorted[i].irq); 558 } 559 560 /* 561 * If the hardirq time is less than HARDIRQ_PERCENT_THRESH% in the last 562 * sample_period, then we suspect the interrupt storm might be subsiding. 563 */ 564 if (!need_counting_irqs()) 565 stop_counting_irqs(); 566 } 567 } 568 569 static void report_cpu_status(void) 570 { 571 print_cpustat(); 572 print_irq_counts(); 573 } 574 #else 575 static inline void update_cpustat(void) { } 576 static inline void report_cpu_status(void) { } 577 static inline bool need_counting_irqs(void) { return false; } 578 static inline void start_counting_irqs(void) { } 579 static inline void stop_counting_irqs(void) { } 580 #endif 581 582 /* 583 * Hard-lockup warnings should be triggered after just a few seconds. Soft- 584 * lockups can have false positives under extreme conditions. So we generally 585 * want a higher threshold for soft lockups than for hard lockups. So we couple 586 * the thresholds with a factor: we make the soft threshold twice the amount of 587 * time the hard threshold is. 588 */ 589 static int get_softlockup_thresh(void) 590 { 591 return watchdog_thresh * 2; 592 } 593 594 /* 595 * Returns seconds, approximately. We don't need nanosecond 596 * resolution, and we don't need to waste time with a big divide when 597 * 2^30ns == 1.074s. 598 */ 599 static unsigned long get_timestamp(void) 600 { 601 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ 602 } 603 604 static void set_sample_period(void) 605 { 606 /* 607 * convert watchdog_thresh from seconds to ns 608 * the divide by 5 is to give hrtimer several chances (two 609 * or three with the current relation between the soft 610 * and hard thresholds) to increment before the 611 * hardlockup detector generates a warning 612 */ 613 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / NUM_SAMPLE_PERIODS); 614 watchdog_update_hrtimer_threshold(sample_period); 615 } 616 617 static void update_report_ts(void) 618 { 619 __this_cpu_write(watchdog_report_ts, get_timestamp()); 620 } 621 622 /* Commands for resetting the watchdog */ 623 static void update_touch_ts(void) 624 { 625 __this_cpu_write(watchdog_touch_ts, get_timestamp()); 626 update_report_ts(); 627 } 628 629 /** 630 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls 631 * 632 * Call when the scheduler may have stalled for legitimate reasons 633 * preventing the watchdog task from executing - e.g. the scheduler 634 * entering idle state. This should only be used for scheduler events. 635 * Use touch_softlockup_watchdog() for everything else. 636 */ 637 notrace void touch_softlockup_watchdog_sched(void) 638 { 639 /* 640 * Preemption can be enabled. It doesn't matter which CPU's watchdog 641 * report period gets restarted here, so use the raw_ operation. 642 */ 643 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT); 644 } 645 646 notrace void touch_softlockup_watchdog(void) 647 { 648 touch_softlockup_watchdog_sched(); 649 wq_watchdog_touch(raw_smp_processor_id()); 650 } 651 EXPORT_SYMBOL(touch_softlockup_watchdog); 652 653 void touch_all_softlockup_watchdogs(void) 654 { 655 int cpu; 656 657 /* 658 * watchdog_mutex cannpt be taken here, as this might be called 659 * from (soft)interrupt context, so the access to 660 * watchdog_allowed_cpumask might race with a concurrent update. 661 * 662 * The watchdog time stamp can race against a concurrent real 663 * update as well, the only side effect might be a cycle delay for 664 * the softlockup check. 665 */ 666 for_each_cpu(cpu, &watchdog_allowed_mask) { 667 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT; 668 wq_watchdog_touch(cpu); 669 } 670 } 671 672 void touch_softlockup_watchdog_sync(void) 673 { 674 __this_cpu_write(softlockup_touch_sync, true); 675 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT); 676 } 677 678 static int is_softlockup(unsigned long touch_ts, 679 unsigned long period_ts, 680 unsigned long now) 681 { 682 if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) { 683 /* 684 * If period_ts has not been updated during a sample_period, then 685 * in the subsequent few sample_periods, period_ts might also not 686 * be updated, which could indicate a potential softlockup. In 687 * this case, if we suspect the cause of the potential softlockup 688 * might be interrupt storm, then we need to count the interrupts 689 * to find which interrupt is storming. 690 */ 691 if (time_after_eq(now, period_ts + get_softlockup_thresh() / NUM_SAMPLE_PERIODS) && 692 need_counting_irqs()) 693 start_counting_irqs(); 694 695 /* 696 * A poorly behaving BPF scheduler can live-lock the system into 697 * soft lockups. Tell sched_ext to try ejecting the BPF 698 * scheduler when close to a soft lockup. 699 */ 700 if (time_after_eq(now, period_ts + get_softlockup_thresh() * 3 / 4)) 701 scx_softlockup(now - touch_ts); 702 703 /* Warn about unreasonable delays. */ 704 if (time_after(now, period_ts + get_softlockup_thresh())) 705 return now - touch_ts; 706 } 707 return 0; 708 } 709 710 /* watchdog detector functions */ 711 static DEFINE_PER_CPU(struct completion, softlockup_completion); 712 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work); 713 714 /* 715 * The watchdog feed function - touches the timestamp. 716 * 717 * It only runs once every sample_period seconds (4 seconds by 718 * default) to reset the softlockup timestamp. If this gets delayed 719 * for more than 2*watchdog_thresh seconds then the debug-printout 720 * triggers in watchdog_timer_fn(). 721 */ 722 static int softlockup_fn(void *data) 723 { 724 update_touch_ts(); 725 stop_counting_irqs(); 726 complete(this_cpu_ptr(&softlockup_completion)); 727 728 return 0; 729 } 730 731 /* watchdog kicker functions */ 732 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) 733 { 734 unsigned long touch_ts, period_ts, now; 735 struct pt_regs *regs = get_irq_regs(); 736 int duration; 737 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; 738 unsigned long flags; 739 740 if (!watchdog_enabled) 741 return HRTIMER_NORESTART; 742 743 watchdog_hardlockup_kick(); 744 745 /* kick the softlockup detector */ 746 if (completion_done(this_cpu_ptr(&softlockup_completion))) { 747 reinit_completion(this_cpu_ptr(&softlockup_completion)); 748 stop_one_cpu_nowait(smp_processor_id(), 749 softlockup_fn, NULL, 750 this_cpu_ptr(&softlockup_stop_work)); 751 } 752 753 /* .. and repeat */ 754 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); 755 756 /* 757 * Read the current timestamp first. It might become invalid anytime 758 * when a virtual machine is stopped by the host or when the watchog 759 * is touched from NMI. 760 */ 761 now = get_timestamp(); 762 /* 763 * If a virtual machine is stopped by the host it can look to 764 * the watchdog like a soft lockup. This function touches the watchdog. 765 */ 766 kvm_check_and_clear_guest_paused(); 767 /* 768 * The stored timestamp is comparable with @now only when not touched. 769 * It might get touched anytime from NMI. Make sure that is_softlockup() 770 * uses the same (valid) value. 771 */ 772 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts)); 773 774 update_cpustat(); 775 776 /* Reset the interval when touched by known problematic code. */ 777 if (period_ts == SOFTLOCKUP_DELAY_REPORT) { 778 if (unlikely(__this_cpu_read(softlockup_touch_sync))) { 779 /* 780 * If the time stamp was touched atomically 781 * make sure the scheduler tick is up to date. 782 */ 783 __this_cpu_write(softlockup_touch_sync, false); 784 sched_clock_tick(); 785 } 786 787 update_report_ts(); 788 return HRTIMER_RESTART; 789 } 790 791 /* Check for a softlockup. */ 792 touch_ts = __this_cpu_read(watchdog_touch_ts); 793 duration = is_softlockup(touch_ts, period_ts, now); 794 if (unlikely(duration)) { 795 #ifdef CONFIG_SYSFS 796 ++softlockup_count; 797 #endif 798 799 /* 800 * Prevent multiple soft-lockup reports if one cpu is already 801 * engaged in dumping all cpu back traces. 802 */ 803 if (softlockup_all_cpu_backtrace) { 804 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn)) 805 return HRTIMER_RESTART; 806 } 807 808 /* Start period for the next softlockup warning. */ 809 update_report_ts(); 810 811 printk_cpu_sync_get_irqsave(flags); 812 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", 813 smp_processor_id(), duration, 814 current->comm, task_pid_nr(current)); 815 report_cpu_status(); 816 print_modules(); 817 print_irqtrace_events(current); 818 if (regs) 819 show_regs(regs); 820 else 821 dump_stack(); 822 printk_cpu_sync_put_irqrestore(flags); 823 824 if (softlockup_all_cpu_backtrace) { 825 trigger_allbutcpu_cpu_backtrace(smp_processor_id()); 826 if (!softlockup_panic) 827 clear_bit_unlock(0, &soft_lockup_nmi_warn); 828 } 829 830 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); 831 if (softlockup_panic) 832 panic("softlockup: hung tasks"); 833 } 834 835 return HRTIMER_RESTART; 836 } 837 838 static void watchdog_enable(unsigned int cpu) 839 { 840 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); 841 struct completion *done = this_cpu_ptr(&softlockup_completion); 842 843 WARN_ON_ONCE(cpu != smp_processor_id()); 844 845 init_completion(done); 846 complete(done); 847 848 /* 849 * Start the timer first to prevent the hardlockup watchdog triggering 850 * before the timer has a chance to fire. 851 */ 852 hrtimer_setup(hrtimer, watchdog_timer_fn, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); 853 hrtimer_start(hrtimer, ns_to_ktime(sample_period), 854 HRTIMER_MODE_REL_PINNED_HARD); 855 856 /* Initialize timestamp */ 857 update_touch_ts(); 858 /* Enable the hardlockup detector */ 859 if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED) 860 watchdog_hardlockup_enable(cpu); 861 } 862 863 static void watchdog_disable(unsigned int cpu) 864 { 865 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); 866 867 WARN_ON_ONCE(cpu != smp_processor_id()); 868 869 /* 870 * Disable the hardlockup detector first. That prevents that a large 871 * delay between disabling the timer and disabling the hardlockup 872 * detector causes a false positive. 873 */ 874 watchdog_hardlockup_disable(cpu); 875 hrtimer_cancel(hrtimer); 876 wait_for_completion(this_cpu_ptr(&softlockup_completion)); 877 } 878 879 static int softlockup_stop_fn(void *data) 880 { 881 watchdog_disable(smp_processor_id()); 882 return 0; 883 } 884 885 static void softlockup_stop_all(void) 886 { 887 int cpu; 888 889 if (!softlockup_initialized) 890 return; 891 892 for_each_cpu(cpu, &watchdog_allowed_mask) 893 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false); 894 895 cpumask_clear(&watchdog_allowed_mask); 896 } 897 898 static int softlockup_start_fn(void *data) 899 { 900 watchdog_enable(smp_processor_id()); 901 return 0; 902 } 903 904 static void softlockup_start_all(void) 905 { 906 int cpu; 907 908 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); 909 for_each_cpu(cpu, &watchdog_allowed_mask) 910 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false); 911 } 912 913 int lockup_detector_online_cpu(unsigned int cpu) 914 { 915 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask)) 916 watchdog_enable(cpu); 917 return 0; 918 } 919 920 int lockup_detector_offline_cpu(unsigned int cpu) 921 { 922 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask)) 923 watchdog_disable(cpu); 924 return 0; 925 } 926 927 static void __lockup_detector_reconfigure(bool thresh_changed) 928 { 929 cpus_read_lock(); 930 watchdog_hardlockup_stop(); 931 932 softlockup_stop_all(); 933 /* 934 * To prevent watchdog_timer_fn from using the old interval and 935 * the new watchdog_thresh at the same time, which could lead to 936 * false softlockup reports, it is necessary to update the 937 * watchdog_thresh after the softlockup is completed. 938 */ 939 if (thresh_changed) 940 watchdog_thresh = READ_ONCE(watchdog_thresh_next); 941 set_sample_period(); 942 lockup_detector_update_enable(); 943 if (watchdog_enabled && watchdog_thresh) 944 softlockup_start_all(); 945 946 watchdog_hardlockup_start(); 947 cpus_read_unlock(); 948 } 949 950 void lockup_detector_reconfigure(void) 951 { 952 mutex_lock(&watchdog_mutex); 953 __lockup_detector_reconfigure(false); 954 mutex_unlock(&watchdog_mutex); 955 } 956 957 /* 958 * Create the watchdog infrastructure and configure the detector(s). 959 */ 960 static __init void lockup_detector_setup(void) 961 { 962 /* 963 * If sysctl is off and watchdog got disabled on the command line, 964 * nothing to do here. 965 */ 966 lockup_detector_update_enable(); 967 968 if (!IS_ENABLED(CONFIG_SYSCTL) && 969 !(watchdog_enabled && watchdog_thresh)) 970 return; 971 972 mutex_lock(&watchdog_mutex); 973 __lockup_detector_reconfigure(false); 974 softlockup_initialized = true; 975 mutex_unlock(&watchdog_mutex); 976 } 977 978 #else /* CONFIG_SOFTLOCKUP_DETECTOR */ 979 static void __lockup_detector_reconfigure(bool thresh_changed) 980 { 981 cpus_read_lock(); 982 watchdog_hardlockup_stop(); 983 if (thresh_changed) 984 watchdog_thresh = READ_ONCE(watchdog_thresh_next); 985 lockup_detector_update_enable(); 986 watchdog_hardlockup_start(); 987 cpus_read_unlock(); 988 } 989 void lockup_detector_reconfigure(void) 990 { 991 __lockup_detector_reconfigure(false); 992 } 993 static inline void lockup_detector_setup(void) 994 { 995 __lockup_detector_reconfigure(false); 996 } 997 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ 998 999 /** 1000 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) 1001 * 1002 * Special interface for parisc. It prevents lockup detector warnings from 1003 * the default pm_poweroff() function which busy loops forever. 1004 */ 1005 void lockup_detector_soft_poweroff(void) 1006 { 1007 watchdog_enabled = 0; 1008 } 1009 1010 #ifdef CONFIG_SYSCTL 1011 1012 /* Propagate any changes to the watchdog infrastructure */ 1013 static void proc_watchdog_update(bool thresh_changed) 1014 { 1015 /* Remove impossible cpus to keep sysctl output clean. */ 1016 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); 1017 __lockup_detector_reconfigure(thresh_changed); 1018 } 1019 1020 /* 1021 * common function for watchdog, nmi_watchdog and soft_watchdog parameter 1022 * 1023 * caller | table->data points to | 'which' 1024 * -------------------|----------------------------------|------------------------------- 1025 * proc_watchdog | watchdog_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED | 1026 * | | WATCHDOG_SOFTOCKUP_ENABLED 1027 * -------------------|----------------------------------|------------------------------- 1028 * proc_nmi_watchdog | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED 1029 * -------------------|----------------------------------|------------------------------- 1030 * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED 1031 */ 1032 static int proc_watchdog_common(int which, const struct ctl_table *table, int write, 1033 void *buffer, size_t *lenp, loff_t *ppos) 1034 { 1035 int err, old, *param = table->data; 1036 1037 mutex_lock(&watchdog_mutex); 1038 1039 old = *param; 1040 if (!write) { 1041 /* 1042 * On read synchronize the userspace interface. This is a 1043 * racy snapshot. 1044 */ 1045 *param = (watchdog_enabled & which) != 0; 1046 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1047 *param = old; 1048 } else { 1049 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1050 if (!err && old != READ_ONCE(*param)) 1051 proc_watchdog_update(false); 1052 } 1053 mutex_unlock(&watchdog_mutex); 1054 return err; 1055 } 1056 1057 /* 1058 * /proc/sys/kernel/watchdog 1059 */ 1060 static int proc_watchdog(const struct ctl_table *table, int write, 1061 void *buffer, size_t *lenp, loff_t *ppos) 1062 { 1063 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED | 1064 WATCHDOG_SOFTOCKUP_ENABLED, 1065 table, write, buffer, lenp, ppos); 1066 } 1067 1068 /* 1069 * /proc/sys/kernel/nmi_watchdog 1070 */ 1071 static int proc_nmi_watchdog(const struct ctl_table *table, int write, 1072 void *buffer, size_t *lenp, loff_t *ppos) 1073 { 1074 if (!watchdog_hardlockup_available && write) 1075 return -ENOTSUPP; 1076 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED, 1077 table, write, buffer, lenp, ppos); 1078 } 1079 1080 #ifdef CONFIG_SOFTLOCKUP_DETECTOR 1081 /* 1082 * /proc/sys/kernel/soft_watchdog 1083 */ 1084 static int proc_soft_watchdog(const struct ctl_table *table, int write, 1085 void *buffer, size_t *lenp, loff_t *ppos) 1086 { 1087 return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED, 1088 table, write, buffer, lenp, ppos); 1089 } 1090 #endif 1091 1092 /* 1093 * /proc/sys/kernel/watchdog_thresh 1094 */ 1095 static int proc_watchdog_thresh(const struct ctl_table *table, int write, 1096 void *buffer, size_t *lenp, loff_t *ppos) 1097 { 1098 int err, old; 1099 1100 mutex_lock(&watchdog_mutex); 1101 1102 watchdog_thresh_next = READ_ONCE(watchdog_thresh); 1103 1104 old = watchdog_thresh_next; 1105 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1106 1107 if (!err && write && old != READ_ONCE(watchdog_thresh_next)) 1108 proc_watchdog_update(true); 1109 1110 mutex_unlock(&watchdog_mutex); 1111 return err; 1112 } 1113 1114 /* 1115 * The cpumask is the mask of possible cpus that the watchdog can run 1116 * on, not the mask of cpus it is actually running on. This allows the 1117 * user to specify a mask that will include cpus that have not yet 1118 * been brought online, if desired. 1119 */ 1120 static int proc_watchdog_cpumask(const struct ctl_table *table, int write, 1121 void *buffer, size_t *lenp, loff_t *ppos) 1122 { 1123 int err; 1124 1125 mutex_lock(&watchdog_mutex); 1126 1127 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); 1128 if (!err && write) 1129 proc_watchdog_update(false); 1130 1131 mutex_unlock(&watchdog_mutex); 1132 return err; 1133 } 1134 1135 static const int sixty = 60; 1136 1137 static const struct ctl_table watchdog_sysctls[] = { 1138 { 1139 .procname = "watchdog", 1140 .data = &watchdog_user_enabled, 1141 .maxlen = sizeof(int), 1142 .mode = 0644, 1143 .proc_handler = proc_watchdog, 1144 .extra1 = SYSCTL_ZERO, 1145 .extra2 = SYSCTL_ONE, 1146 }, 1147 { 1148 .procname = "watchdog_thresh", 1149 .data = &watchdog_thresh_next, 1150 .maxlen = sizeof(int), 1151 .mode = 0644, 1152 .proc_handler = proc_watchdog_thresh, 1153 .extra1 = SYSCTL_ZERO, 1154 .extra2 = (void *)&sixty, 1155 }, 1156 { 1157 .procname = "watchdog_cpumask", 1158 .data = &watchdog_cpumask_bits, 1159 .maxlen = NR_CPUS, 1160 .mode = 0644, 1161 .proc_handler = proc_watchdog_cpumask, 1162 }, 1163 #ifdef CONFIG_SOFTLOCKUP_DETECTOR 1164 { 1165 .procname = "soft_watchdog", 1166 .data = &watchdog_softlockup_user_enabled, 1167 .maxlen = sizeof(int), 1168 .mode = 0644, 1169 .proc_handler = proc_soft_watchdog, 1170 .extra1 = SYSCTL_ZERO, 1171 .extra2 = SYSCTL_ONE, 1172 }, 1173 { 1174 .procname = "softlockup_panic", 1175 .data = &softlockup_panic, 1176 .maxlen = sizeof(int), 1177 .mode = 0644, 1178 .proc_handler = proc_dointvec_minmax, 1179 .extra1 = SYSCTL_ZERO, 1180 .extra2 = SYSCTL_ONE, 1181 }, 1182 #ifdef CONFIG_SMP 1183 { 1184 .procname = "softlockup_all_cpu_backtrace", 1185 .data = &sysctl_softlockup_all_cpu_backtrace, 1186 .maxlen = sizeof(int), 1187 .mode = 0644, 1188 .proc_handler = proc_dointvec_minmax, 1189 .extra1 = SYSCTL_ZERO, 1190 .extra2 = SYSCTL_ONE, 1191 }, 1192 #endif /* CONFIG_SMP */ 1193 #endif 1194 #ifdef CONFIG_HARDLOCKUP_DETECTOR 1195 { 1196 .procname = "hardlockup_panic", 1197 .data = &hardlockup_panic, 1198 .maxlen = sizeof(int), 1199 .mode = 0644, 1200 .proc_handler = proc_dointvec_minmax, 1201 .extra1 = SYSCTL_ZERO, 1202 .extra2 = SYSCTL_ONE, 1203 }, 1204 #ifdef CONFIG_SMP 1205 { 1206 .procname = "hardlockup_all_cpu_backtrace", 1207 .data = &sysctl_hardlockup_all_cpu_backtrace, 1208 .maxlen = sizeof(int), 1209 .mode = 0644, 1210 .proc_handler = proc_dointvec_minmax, 1211 .extra1 = SYSCTL_ZERO, 1212 .extra2 = SYSCTL_ONE, 1213 }, 1214 #endif /* CONFIG_SMP */ 1215 #endif 1216 }; 1217 1218 static struct ctl_table watchdog_hardlockup_sysctl[] = { 1219 { 1220 .procname = "nmi_watchdog", 1221 .data = &watchdog_hardlockup_user_enabled, 1222 .maxlen = sizeof(int), 1223 .mode = 0444, 1224 .proc_handler = proc_nmi_watchdog, 1225 .extra1 = SYSCTL_ZERO, 1226 .extra2 = SYSCTL_ONE, 1227 }, 1228 }; 1229 1230 static void __init watchdog_sysctl_init(void) 1231 { 1232 register_sysctl_init("kernel", watchdog_sysctls); 1233 1234 if (watchdog_hardlockup_available) 1235 watchdog_hardlockup_sysctl[0].mode = 0644; 1236 register_sysctl_init("kernel", watchdog_hardlockup_sysctl); 1237 } 1238 1239 #else 1240 #define watchdog_sysctl_init() do { } while (0) 1241 #endif /* CONFIG_SYSCTL */ 1242 1243 static void __init lockup_detector_delay_init(struct work_struct *work); 1244 static bool allow_lockup_detector_init_retry __initdata; 1245 1246 static struct work_struct detector_work __initdata = 1247 __WORK_INITIALIZER(detector_work, lockup_detector_delay_init); 1248 1249 static void __init lockup_detector_delay_init(struct work_struct *work) 1250 { 1251 int ret; 1252 1253 ret = watchdog_hardlockup_probe(); 1254 if (ret) { 1255 if (ret == -ENODEV) 1256 pr_info("NMI not fully supported\n"); 1257 else 1258 pr_info("Delayed init of the lockup detector failed: %d\n", ret); 1259 pr_info("Hard watchdog permanently disabled\n"); 1260 return; 1261 } 1262 1263 allow_lockup_detector_init_retry = false; 1264 1265 watchdog_hardlockup_available = true; 1266 lockup_detector_setup(); 1267 } 1268 1269 /* 1270 * lockup_detector_retry_init - retry init lockup detector if possible. 1271 * 1272 * Retry hardlockup detector init. It is useful when it requires some 1273 * functionality that has to be initialized later on a particular 1274 * platform. 1275 */ 1276 void __init lockup_detector_retry_init(void) 1277 { 1278 /* Must be called before late init calls */ 1279 if (!allow_lockup_detector_init_retry) 1280 return; 1281 1282 schedule_work(&detector_work); 1283 } 1284 1285 /* 1286 * Ensure that optional delayed hardlockup init is proceed before 1287 * the init code and memory is freed. 1288 */ 1289 static int __init lockup_detector_check(void) 1290 { 1291 /* Prevent any later retry. */ 1292 allow_lockup_detector_init_retry = false; 1293 1294 /* Make sure no work is pending. */ 1295 flush_work(&detector_work); 1296 1297 watchdog_sysctl_init(); 1298 1299 return 0; 1300 1301 } 1302 late_initcall_sync(lockup_detector_check); 1303 1304 void __init lockup_detector_init(void) 1305 { 1306 if (tick_nohz_full_enabled()) 1307 pr_info("Disabling watchdog on nohz_full cores by default\n"); 1308 1309 cpumask_copy(&watchdog_cpumask, 1310 housekeeping_cpumask(HK_TYPE_TIMER)); 1311 1312 if (!watchdog_hardlockup_probe()) 1313 watchdog_hardlockup_available = true; 1314 else 1315 allow_lockup_detector_init_retry = true; 1316 1317 lockup_detector_setup(); 1318 } 1319