arch/powerpc/kernel/watchdog.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Watchdog support on powerpc systems.
   4  *
   5  * Copyright 2017, IBM Corporation.
   6  *
   7  * This uses code from arch/sparc/kernel/nmi.c and kernel/watchdog.c
   8  */
   9
  10 #define pr_fmt(fmt) "watchdog: " fmt
  11
  12 #include <linux/kernel.h>
  13 #include <linux/param.h>
  14 #include <linux/init.h>
  15 #include <linux/percpu.h>
  16 #include <linux/cpu.h>
  17 #include <linux/nmi.h>
  18 #include <linux/module.h>
  19 #include <linux/export.h>
  20 #include <linux/kprobes.h>
  21 #include <linux/hardirq.h>
  22 #include <linux/reboot.h>
  23 #include <linux/slab.h>
  24 #include <linux/kdebug.h>
  25 #include <linux/sched/debug.h>
  26 #include <linux/delay.h>
  27 #include <linux/smp.h>
  28
  29 #include <asm/paca.h>
  30
  31 /*
  32  * The powerpc watchdog ensures that each CPU is able to service timers.
  33  * The watchdog sets up a simple timer on each CPU to run once per timer
  34  * period, and updates a per-cpu timestamp and a "pending" cpumask. This is
  35  * the heartbeat.
  36  *
  37  * Then there are two systems to check that the heartbeat is still running.
  38  * The local soft-NMI, and the SMP checker.
  39  *
  40  * The soft-NMI checker can detect lockups on the local CPU. When interrupts
  41  * are disabled with local_irq_disable(), platforms that use soft-masking
  42  * can leave hardware interrupts enabled and handle them with a masked
  43  * interrupt handler. The masked handler can send the timer interrupt to the
  44  * watchdog's soft_nmi_interrupt(), which appears to Linux as an NMI
  45  * interrupt, and can be used to detect CPUs stuck with IRQs disabled.
  46  *
  47  * The soft-NMI checker will compare the heartbeat timestamp for this CPU
  48  * with the current time, and take action if the difference exceeds the
  49  * watchdog threshold.
  50  *
  51  * The limitation of the soft-NMI watchdog is that it does not work when
  52  * interrupts are hard disabled or otherwise not being serviced. This is
  53  * solved by also having a SMP watchdog where all CPUs check all other
  54  * CPUs heartbeat.
  55  *
  56  * The SMP checker can detect lockups on other CPUs. A gobal "pending"
  57  * cpumask is kept, containing all CPUs which enable the watchdog. Each
  58  * CPU clears their pending bit in their heartbeat timer. When the bitmask
  59  * becomes empty, the last CPU to clear its pending bit updates a global
  60  * timestamp and refills the pending bitmask.
  61  *
  62  * In the heartbeat timer, if any CPU notices that the global timestamp has
  63  * not been updated for a period exceeding the watchdog threshold, then it
  64  * means the CPU(s) with their bit still set in the pending mask have had
  65  * their heartbeat stop, and action is taken.
  66  *
  67  * Some platforms implement true NMI IPIs, which can be used by the SMP
  68  * watchdog to detect an unresponsive CPU and pull it out of its stuck
  69  * state with the NMI IPI, to get crash/debug data from it. This way the
  70  * SMP watchdog can detect hardware interrupts off lockups.
  71  */
  72
  73 static cpumask_t wd_cpus_enabled __read_mostly;
  74
  75 static u64 wd_panic_timeout_tb __read_mostly; /* timebase ticks until panic */
  76 static u64 wd_smp_panic_timeout_tb __read_mostly; /* panic other CPUs */
  77
  78 static u64 wd_timer_period_ms __read_mostly;  /* interval between heartbeat */
  79
  80 static DEFINE_PER_CPU(struct hrtimer, wd_hrtimer);
  81 static DEFINE_PER_CPU(u64, wd_timer_tb);
  82
  83 /* SMP checker bits */
  84 static unsigned long __wd_smp_lock;
  85 static cpumask_t wd_smp_cpus_pending;
  86 static cpumask_t wd_smp_cpus_stuck;
  87 static u64 wd_smp_last_reset_tb;
  88
  89 static inline void wd_smp_lock(unsigned long *flags)
  90 {
  91         /*
  92          * Avoid locking layers if possible.
  93          * This may be called from low level interrupt handlers at some
  94          * point in future.
  95          */
  96         raw_local_irq_save(*flags);
  97         hard_irq_disable(); /* Make it soft-NMI safe */
  98         while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
  99                 raw_local_irq_restore(*flags);
 100                 spin_until_cond(!test_bit(0, &__wd_smp_lock));
 101                 raw_local_irq_save(*flags);
 102                 hard_irq_disable();
 103         }
 104 }
 105
 106 static inline void wd_smp_unlock(unsigned long *flags)
 107 {
 108         clear_bit_unlock(0, &__wd_smp_lock);
 109         raw_local_irq_restore(*flags);
 110 }
 111
 112 static void wd_lockup_ipi(struct pt_regs *regs)
 113 {
 114         int cpu = raw_smp_processor_id();
 115         u64 tb = get_tb();
 116
 117         pr_emerg("CPU %d Hard LOCKUP\n", cpu);
 118         pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
 119                  cpu, tb, per_cpu(wd_timer_tb, cpu),
 120                  tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
 121         print_modules();
 122         print_irqtrace_events(current);
 123         if (regs)
 124                 show_regs(regs);
 125         else
 126                 dump_stack();
 127
 128         /* Do not panic from here because that can recurse into NMI IPI layer */
 129 }
 130
 131 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
 132 {
 133         cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
 134         cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
 135         /*
 136          * See wd_smp_clear_cpu_pending()
 137          */
 138         smp_mb();
 139         if (cpumask_empty(&wd_smp_cpus_pending)) {
 140                 wd_smp_last_reset_tb = tb;
 141                 cpumask_andnot(&wd_smp_cpus_pending,
 142                                 &wd_cpus_enabled,
 143                                 &wd_smp_cpus_stuck);
 144         }
 145 }
 146 static void set_cpu_stuck(int cpu, u64 tb)
 147 {
 148         set_cpumask_stuck(cpumask_of(cpu), tb);
 149 }
 150
 151 static void watchdog_smp_panic(int cpu, u64 tb)
 152 {
 153         unsigned long flags;
 154         int c;
 155
 156         wd_smp_lock(&flags);
 157         /* Double check some things under lock */
 158         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 159                 goto out;
 160         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 161                 goto out;
 162         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 163                 goto out;
 164
 165         pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
 166                  cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 167         pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
 168                  cpu, tb, wd_smp_last_reset_tb,
 169                  tb_to_ns(tb - wd_smp_last_reset_tb) / 1000000);
 170
 171         if (!sysctl_hardlockup_all_cpu_backtrace) {
 172                 /*
 173                  * Try to trigger the stuck CPUs, unless we are going to
 174                  * get a backtrace on all of them anyway.
 175                  */
 176                 for_each_cpu(c, &wd_smp_cpus_pending) {
 177                         if (c == cpu)
 178                                 continue;
 179                         smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 180                 }
 181         }
 182
 183         /* Take the stuck CPUs out of the watch group */
 184         set_cpumask_stuck(&wd_smp_cpus_pending, tb);
 185
 186         wd_smp_unlock(&flags);
 187
 188         printk_safe_flush();
 189         /*
 190          * printk_safe_flush() seems to require another print
 191          * before anything actually goes out to console.
 192          */
 193         if (sysctl_hardlockup_all_cpu_backtrace)
 194                 trigger_allbutself_cpu_backtrace();
 195
 196         if (hardlockup_panic)
 197                 nmi_panic(NULL, "Hard LOCKUP");
 198
 199         return;
 200
 201 out:
 202         wd_smp_unlock(&flags);
 203 }
 204
 205 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 206 {
 207         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 208                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 209                         struct pt_regs *regs = get_irq_regs();
 210                         unsigned long flags;
 211
 212                         wd_smp_lock(&flags);
 213
 214                         pr_emerg("CPU %d became unstuck TB:%lld\n",
 215                                  cpu, tb);
 216                         print_irqtrace_events(current);
 217                         if (regs)
 218                                 show_regs(regs);
 219                         else
 220                                 dump_stack();
 221
 222                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 223                         wd_smp_unlock(&flags);
 224                 } else {
 225                         /*
 226                          * The last CPU to clear pending should have reset the
 227                          * watchdog so we generally should not find it empty
 228                          * here if our CPU was clear. However it could happen
 229                          * due to a rare race with another CPU taking the
 230                          * last CPU out of the mask concurrently.
 231                          *
 232                          * We can't add a warning for it. But just in case
 233                          * there is a problem with the watchdog that is causing
 234                          * the mask to not be reset, try to kick it along here.
 235                          */
 236                         if (unlikely(cpumask_empty(&wd_smp_cpus_pending)))
 237                                 goto none_pending;
 238                 }
 239                 return;
 240         }
 241
 242         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 243
 244         /*
 245          * Order the store to clear pending with the load(s) to check all
 246          * words in the pending mask to check they are all empty. This orders
 247          * with the same barrier on another CPU. This prevents two CPUs
 248          * clearing the last 2 pending bits, but neither seeing the other's
 249          * store when checking if the mask is empty, and missing an empty
 250          * mask, which ends with a false positive.
 251          */
 252         smp_mb();
 253         if (cpumask_empty(&wd_smp_cpus_pending)) {
 254                 unsigned long flags;
 255
 256 none_pending:
 257                 /*
 258                  * Double check under lock because more than one CPU could see
 259                  * a clear mask with the lockless check after clearing their
 260                  * pending bits.
 261                  */
 262                 wd_smp_lock(&flags);
 263                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 264                         wd_smp_last_reset_tb = tb;
 265                         cpumask_andnot(&wd_smp_cpus_pending,
 266                                         &wd_cpus_enabled,
 267                                         &wd_smp_cpus_stuck);
 268                 }
 269                 wd_smp_unlock(&flags);
 270         }
 271 }
 272
 273 static void watchdog_timer_interrupt(int cpu)
 274 {
 275         u64 tb = get_tb();
 276
 277         per_cpu(wd_timer_tb, cpu) = tb;
 278
 279         wd_smp_clear_cpu_pending(cpu, tb);
 280
 281         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 282                 watchdog_smp_panic(cpu, tb);
 283 }
 284
 285 void soft_nmi_interrupt(struct pt_regs *regs)
 286 {
 287         unsigned long flags;
 288         int cpu = raw_smp_processor_id();
 289         u64 tb;
 290
 291         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 292                 return;
 293
 294         nmi_enter();
 295
 296         __this_cpu_inc(irq_stat.soft_nmi_irqs);
 297
 298         tb = get_tb();
 299         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 300                 wd_smp_lock(&flags);
 301                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 302                         wd_smp_unlock(&flags);
 303                         goto out;
 304                 }
 305                 set_cpu_stuck(cpu, tb);
 306
 307                 pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
 308                          cpu, (void *)regs->nip);
 309                 pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
 310                          cpu, tb, per_cpu(wd_timer_tb, cpu),
 311                          tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
 312                 print_modules();
 313                 print_irqtrace_events(current);
 314                 show_regs(regs);
 315
 316                 wd_smp_unlock(&flags);
 317
 318                 if (sysctl_hardlockup_all_cpu_backtrace)
 319                         trigger_allbutself_cpu_backtrace();
 320
 321                 if (hardlockup_panic)
 322                         nmi_panic(regs, "Hard LOCKUP");
 323         }
 324         if (wd_panic_timeout_tb < 0x7fffffff)
 325                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 326
 327 out:
 328         nmi_exit();
 329 }
 330
 331 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 332 {
 333         int cpu = smp_processor_id();
 334
 335         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 336                 return HRTIMER_NORESTART;
 337
 338         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 339                 return HRTIMER_NORESTART;
 340
 341         watchdog_timer_interrupt(cpu);
 342
 343         hrtimer_forward_now(hrtimer, ms_to_ktime(wd_timer_period_ms));
 344
 345         return HRTIMER_RESTART;
 346 }
 347
 348 void arch_touch_nmi_watchdog(void)
 349 {
 350         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 351         int cpu = smp_processor_id();
 352         u64 tb;
 353
 354         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 355                 return;
 356
 357         tb = get_tb();
 358         if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
 359                 per_cpu(wd_timer_tb, cpu) = tb;
 360                 wd_smp_clear_cpu_pending(cpu, tb);
 361         }
 362 }
 363 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 364
 365 static void start_watchdog(void *arg)
 366 {
 367         struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
 368         int cpu = smp_processor_id();
 369         unsigned long flags;
 370
 371         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 372                 WARN_ON(1);
 373                 return;
 374         }
 375
 376         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 377                 return;
 378
 379         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 380                 return;
 381
 382         wd_smp_lock(&flags);
 383         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 384         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 385                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 386                 wd_smp_last_reset_tb = get_tb();
 387         }
 388         wd_smp_unlock(&flags);
 389
 390         *this_cpu_ptr(&wd_timer_tb) = get_tb();
 391
 392         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 393         hrtimer->function = watchdog_timer_fn;
 394         hrtimer_start(hrtimer, ms_to_ktime(wd_timer_period_ms),
 395                       HRTIMER_MODE_REL_PINNED);
 396 }
 397
 398 static int start_watchdog_on_cpu(unsigned int cpu)
 399 {
 400         return smp_call_function_single(cpu, start_watchdog, NULL, true);
 401 }
 402
 403 static void stop_watchdog(void *arg)
 404 {
 405         struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
 406         int cpu = smp_processor_id();
 407         unsigned long flags;
 408
 409         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 410                 return; /* Can happen in CPU unplug case */
 411
 412         hrtimer_cancel(hrtimer);
 413
 414         wd_smp_lock(&flags);
 415         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 416         wd_smp_unlock(&flags);
 417
 418         wd_smp_clear_cpu_pending(cpu, get_tb());
 419 }
 420
 421 static int stop_watchdog_on_cpu(unsigned int cpu)
 422 {
 423         return smp_call_function_single(cpu, stop_watchdog, NULL, true);
 424 }
 425
 426 static void watchdog_calc_timeouts(void)
 427 {
 428         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 429
 430         /* Have the SMP detector trigger a bit later */
 431         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 432
 433         /* 2/5 is the factor that the perf based detector uses */
 434         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 435 }
 436
 437 void watchdog_nmi_stop(void)
 438 {
 439         int cpu;
 440
 441         for_each_cpu(cpu, &wd_cpus_enabled)
 442                 stop_watchdog_on_cpu(cpu);
 443 }
 444
 445 void watchdog_nmi_start(void)
 446 {
 447         int cpu;
 448
 449         watchdog_calc_timeouts();
 450         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 451                 start_watchdog_on_cpu(cpu);
 452 }
 453
 454 /*
 455  * Invoked from core watchdog init.
 456  */
 457 int __init watchdog_nmi_probe(void)
 458 {
 459         int err;
 460
 461         err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 462                                         "powerpc/watchdog:online",
 463                                         start_watchdog_on_cpu,
 464                                         stop_watchdog_on_cpu);
 465         if (err < 0) {
 466                 pr_warn("could not be initialized");
 467                 return err;
 468         }
 469         return 0;
 470 }