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         if (cpumask_empty(&wd_smp_cpus_pending)) {
 136                 wd_smp_last_reset_tb = tb;
 137                 cpumask_andnot(&wd_smp_cpus_pending,
 138                                 &wd_cpus_enabled,
 139                                 &wd_smp_cpus_stuck);
 140         }
 141 }
 142 static void set_cpu_stuck(int cpu, u64 tb)
 143 {
 144         set_cpumask_stuck(cpumask_of(cpu), tb);
 145 }
 146
 147 static void watchdog_smp_panic(int cpu, u64 tb)
 148 {
 149         unsigned long flags;
 150         int c;
 151
 152         wd_smp_lock(&flags);
 153         /* Double check some things under lock */
 154         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 155                 goto out;
 156         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 157                 goto out;
 158         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 159                 goto out;
 160
 161         pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
 162                  cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 163         pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
 164                  cpu, tb, wd_smp_last_reset_tb,
 165                  tb_to_ns(tb - wd_smp_last_reset_tb) / 1000000);
 166
 167         if (!sysctl_hardlockup_all_cpu_backtrace) {
 168                 /*
 169                  * Try to trigger the stuck CPUs, unless we are going to
 170                  * get a backtrace on all of them anyway.
 171                  */
 172                 for_each_cpu(c, &wd_smp_cpus_pending) {
 173                         if (c == cpu)
 174                                 continue;
 175                         smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 176                 }
 177         }
 178
 179         /* Take the stuck CPUs out of the watch group */
 180         set_cpumask_stuck(&wd_smp_cpus_pending, tb);
 181
 182         wd_smp_unlock(&flags);
 183
 184         printk_safe_flush();
 185         /*
 186          * printk_safe_flush() seems to require another print
 187          * before anything actually goes out to console.
 188          */
 189         if (sysctl_hardlockup_all_cpu_backtrace)
 190                 trigger_allbutself_cpu_backtrace();
 191
 192         if (hardlockup_panic)
 193                 nmi_panic(NULL, "Hard LOCKUP");
 194
 195         return;
 196
 197 out:
 198         wd_smp_unlock(&flags);
 199 }
 200
 201 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 202 {
 203         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 204                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 205                         struct pt_regs *regs = get_irq_regs();
 206                         unsigned long flags;
 207
 208                         wd_smp_lock(&flags);
 209
 210                         pr_emerg("CPU %d became unstuck TB:%lld\n",
 211                                  cpu, tb);
 212                         print_irqtrace_events(current);
 213                         if (regs)
 214                                 show_regs(regs);
 215                         else
 216                                 dump_stack();
 217
 218                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 219                         wd_smp_unlock(&flags);
 220                 }
 221                 return;
 222         }
 223         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 224         if (cpumask_empty(&wd_smp_cpus_pending)) {
 225                 unsigned long flags;
 226
 227                 wd_smp_lock(&flags);
 228                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 229                         wd_smp_last_reset_tb = tb;
 230                         cpumask_andnot(&wd_smp_cpus_pending,
 231                                         &wd_cpus_enabled,
 232                                         &wd_smp_cpus_stuck);
 233                 }
 234                 wd_smp_unlock(&flags);
 235         }
 236 }
 237
 238 static void watchdog_timer_interrupt(int cpu)
 239 {
 240         u64 tb = get_tb();
 241
 242         per_cpu(wd_timer_tb, cpu) = tb;
 243
 244         wd_smp_clear_cpu_pending(cpu, tb);
 245
 246         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 247                 watchdog_smp_panic(cpu, tb);
 248 }
 249
 250 void soft_nmi_interrupt(struct pt_regs *regs)
 251 {
 252         unsigned long flags;
 253         int cpu = raw_smp_processor_id();
 254         u64 tb;
 255
 256         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 257                 return;
 258
 259         nmi_enter();
 260
 261         __this_cpu_inc(irq_stat.soft_nmi_irqs);
 262
 263         tb = get_tb();
 264         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 265                 wd_smp_lock(&flags);
 266                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 267                         wd_smp_unlock(&flags);
 268                         goto out;
 269                 }
 270                 set_cpu_stuck(cpu, tb);
 271
 272                 pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
 273                          cpu, (void *)regs->nip);
 274                 pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
 275                          cpu, tb, per_cpu(wd_timer_tb, cpu),
 276                          tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
 277                 print_modules();
 278                 print_irqtrace_events(current);
 279                 show_regs(regs);
 280
 281                 wd_smp_unlock(&flags);
 282
 283                 if (sysctl_hardlockup_all_cpu_backtrace)
 284                         trigger_allbutself_cpu_backtrace();
 285
 286                 if (hardlockup_panic)
 287                         nmi_panic(regs, "Hard LOCKUP");
 288         }
 289         if (wd_panic_timeout_tb < 0x7fffffff)
 290                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 291
 292 out:
 293         nmi_exit();
 294 }
 295
 296 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
 297 {
 298         int cpu = smp_processor_id();
 299
 300         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 301                 return HRTIMER_NORESTART;
 302
 303         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 304                 return HRTIMER_NORESTART;
 305
 306         watchdog_timer_interrupt(cpu);
 307
 308         hrtimer_forward_now(hrtimer, ms_to_ktime(wd_timer_period_ms));
 309
 310         return HRTIMER_RESTART;
 311 }
 312
 313 void arch_touch_nmi_watchdog(void)
 314 {
 315         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 316         int cpu = smp_processor_id();
 317         u64 tb = get_tb();
 318
 319         if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
 320                 per_cpu(wd_timer_tb, cpu) = tb;
 321                 wd_smp_clear_cpu_pending(cpu, tb);
 322         }
 323 }
 324 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 325
 326 static void start_watchdog(void *arg)
 327 {
 328         struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
 329         int cpu = smp_processor_id();
 330         unsigned long flags;
 331
 332         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 333                 WARN_ON(1);
 334                 return;
 335         }
 336
 337         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 338                 return;
 339
 340         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 341                 return;
 342
 343         wd_smp_lock(&flags);
 344         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 345         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 346                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 347                 wd_smp_last_reset_tb = get_tb();
 348         }
 349         wd_smp_unlock(&flags);
 350
 351         *this_cpu_ptr(&wd_timer_tb) = get_tb();
 352
 353         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 354         hrtimer->function = watchdog_timer_fn;
 355         hrtimer_start(hrtimer, ms_to_ktime(wd_timer_period_ms),
 356                       HRTIMER_MODE_REL_PINNED);
 357 }
 358
 359 static int start_watchdog_on_cpu(unsigned int cpu)
 360 {
 361         return smp_call_function_single(cpu, start_watchdog, NULL, true);
 362 }
 363
 364 static void stop_watchdog(void *arg)
 365 {
 366         struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
 367         int cpu = smp_processor_id();
 368         unsigned long flags;
 369
 370         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 371                 return; /* Can happen in CPU unplug case */
 372
 373         hrtimer_cancel(hrtimer);
 374
 375         wd_smp_lock(&flags);
 376         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 377         wd_smp_unlock(&flags);
 378
 379         wd_smp_clear_cpu_pending(cpu, get_tb());
 380 }
 381
 382 static int stop_watchdog_on_cpu(unsigned int cpu)
 383 {
 384         return smp_call_function_single(cpu, stop_watchdog, NULL, true);
 385 }
 386
 387 static void watchdog_calc_timeouts(void)
 388 {
 389         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 390
 391         /* Have the SMP detector trigger a bit later */
 392         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 393
 394         /* 2/5 is the factor that the perf based detector uses */
 395         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 396 }
 397
 398 void watchdog_nmi_stop(void)
 399 {
 400         int cpu;
 401
 402         for_each_cpu(cpu, &wd_cpus_enabled)
 403                 stop_watchdog_on_cpu(cpu);
 404 }
 405
 406 void watchdog_nmi_start(void)
 407 {
 408         int cpu;
 409
 410         watchdog_calc_timeouts();
 411         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 412                 start_watchdog_on_cpu(cpu);
 413 }
 414
 415 /*
 416  * Invoked from core watchdog init.
 417  */
 418 int __init watchdog_nmi_probe(void)
 419 {
 420         int err;
 421
 422         err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 423                                         "powerpc/watchdog:online",
 424                                         start_watchdog_on_cpu,
 425                                         stop_watchdog_on_cpu);
 426         if (err < 0) {
 427                 pr_warn("could not be initialized");
 428                 return err;
 429         }
 430         return 0;
 431 }