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 by 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 timer_list, wd_timer);
  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         pr_emerg("CPU %d Hard LOCKUP\n", raw_smp_processor_id());
 115         print_modules();
 116         print_irqtrace_events(current);
 117         if (regs)
 118                 show_regs(regs);
 119         else
 120                 dump_stack();
 121
 122         /* Do not panic from here because that can recurse into NMI IPI layer */
 123 }
 124
 125 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
 126 {
 127         cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
 128         cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
 129         if (cpumask_empty(&wd_smp_cpus_pending)) {
 130                 wd_smp_last_reset_tb = tb;
 131                 cpumask_andnot(&wd_smp_cpus_pending,
 132                                 &wd_cpus_enabled,
 133                                 &wd_smp_cpus_stuck);
 134         }
 135 }
 136 static void set_cpu_stuck(int cpu, u64 tb)
 137 {
 138         set_cpumask_stuck(cpumask_of(cpu), tb);
 139 }
 140
 141 static void watchdog_smp_panic(int cpu, u64 tb)
 142 {
 143         unsigned long flags;
 144         int c;
 145
 146         wd_smp_lock(&flags);
 147         /* Double check some things under lock */
 148         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 149                 goto out;
 150         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 151                 goto out;
 152         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 153                 goto out;
 154
 155         pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
 156                  cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 157
 158         if (!sysctl_hardlockup_all_cpu_backtrace) {
 159                 /*
 160                  * Try to trigger the stuck CPUs, unless we are going to
 161                  * get a backtrace on all of them anyway.
 162                  */
 163                 for_each_cpu(c, &wd_smp_cpus_pending) {
 164                         if (c == cpu)
 165                                 continue;
 166                         smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 167                 }
 168                 smp_flush_nmi_ipi(1000000);
 169         }
 170
 171         /* Take the stuck CPUs out of the watch group */
 172         set_cpumask_stuck(&wd_smp_cpus_pending, tb);
 173
 174         wd_smp_unlock(&flags);
 175
 176         printk_safe_flush();
 177         /*
 178          * printk_safe_flush() seems to require another print
 179          * before anything actually goes out to console.
 180          */
 181         if (sysctl_hardlockup_all_cpu_backtrace)
 182                 trigger_allbutself_cpu_backtrace();
 183
 184         if (hardlockup_panic)
 185                 nmi_panic(NULL, "Hard LOCKUP");
 186
 187         return;
 188
 189 out:
 190         wd_smp_unlock(&flags);
 191 }
 192
 193 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 194 {
 195         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 196                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 197                         unsigned long flags;
 198
 199                         pr_emerg("CPU %d became unstuck\n", cpu);
 200                         wd_smp_lock(&flags);
 201                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 202                         wd_smp_unlock(&flags);
 203                 }
 204                 return;
 205         }
 206         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 207         if (cpumask_empty(&wd_smp_cpus_pending)) {
 208                 unsigned long flags;
 209
 210                 wd_smp_lock(&flags);
 211                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 212                         wd_smp_last_reset_tb = tb;
 213                         cpumask_andnot(&wd_smp_cpus_pending,
 214                                         &wd_cpus_enabled,
 215                                         &wd_smp_cpus_stuck);
 216                 }
 217                 wd_smp_unlock(&flags);
 218         }
 219 }
 220
 221 static void watchdog_timer_interrupt(int cpu)
 222 {
 223         u64 tb = get_tb();
 224
 225         per_cpu(wd_timer_tb, cpu) = tb;
 226
 227         wd_smp_clear_cpu_pending(cpu, tb);
 228
 229         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 230                 watchdog_smp_panic(cpu, tb);
 231 }
 232
 233 void soft_nmi_interrupt(struct pt_regs *regs)
 234 {
 235         unsigned long flags;
 236         int cpu = raw_smp_processor_id();
 237         u64 tb;
 238
 239         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 240                 return;
 241
 242         nmi_enter();
 243
 244         __this_cpu_inc(irq_stat.soft_nmi_irqs);
 245
 246         tb = get_tb();
 247         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 248                 per_cpu(wd_timer_tb, cpu) = tb;
 249
 250                 wd_smp_lock(&flags);
 251                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 252                         wd_smp_unlock(&flags);
 253                         goto out;
 254                 }
 255                 set_cpu_stuck(cpu, tb);
 256
 257                 pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n", cpu, (void *)regs->nip);
 258                 print_modules();
 259                 print_irqtrace_events(current);
 260                 show_regs(regs);
 261
 262                 wd_smp_unlock(&flags);
 263
 264                 if (sysctl_hardlockup_all_cpu_backtrace)
 265                         trigger_allbutself_cpu_backtrace();
 266
 267                 if (hardlockup_panic)
 268                         nmi_panic(regs, "Hard LOCKUP");
 269         }
 270         if (wd_panic_timeout_tb < 0x7fffffff)
 271                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 272
 273 out:
 274         nmi_exit();
 275 }
 276
 277 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 278 {
 279         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 280         if (wd_timer_period_ms > 1000)
 281                 t->expires = __round_jiffies_up(t->expires, cpu);
 282         add_timer_on(t, cpu);
 283 }
 284
 285 static void wd_timer_fn(struct timer_list *t)
 286 {
 287         int cpu = smp_processor_id();
 288
 289         watchdog_timer_interrupt(cpu);
 290
 291         wd_timer_reset(cpu, t);
 292 }
 293
 294 void arch_touch_nmi_watchdog(void)
 295 {
 296         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 297         int cpu = smp_processor_id();
 298         u64 tb = get_tb();
 299
 300         if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
 301                 per_cpu(wd_timer_tb, cpu) = tb;
 302                 wd_smp_clear_cpu_pending(cpu, tb);
 303         }
 304 }
 305 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 306
 307 static void start_watchdog_timer_on(unsigned int cpu)
 308 {
 309         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 310
 311         per_cpu(wd_timer_tb, cpu) = get_tb();
 312
 313         timer_setup(t, wd_timer_fn, TIMER_PINNED);
 314         wd_timer_reset(cpu, t);
 315 }
 316
 317 static void stop_watchdog_timer_on(unsigned int cpu)
 318 {
 319         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 320
 321         del_timer_sync(t);
 322 }
 323
 324 static int start_wd_on_cpu(unsigned int cpu)
 325 {
 326         unsigned long flags;
 327
 328         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 329                 WARN_ON(1);
 330                 return 0;
 331         }
 332
 333         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 334                 return 0;
 335
 336         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 337                 return 0;
 338
 339         wd_smp_lock(&flags);
 340         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 341         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 342                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 343                 wd_smp_last_reset_tb = get_tb();
 344         }
 345         wd_smp_unlock(&flags);
 346
 347         start_watchdog_timer_on(cpu);
 348
 349         return 0;
 350 }
 351
 352 static int stop_wd_on_cpu(unsigned int cpu)
 353 {
 354         unsigned long flags;
 355
 356         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 357                 return 0; /* Can happen in CPU unplug case */
 358
 359         stop_watchdog_timer_on(cpu);
 360
 361         wd_smp_lock(&flags);
 362         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 363         wd_smp_unlock(&flags);
 364
 365         wd_smp_clear_cpu_pending(cpu, get_tb());
 366
 367         return 0;
 368 }
 369
 370 static void watchdog_calc_timeouts(void)
 371 {
 372         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 373
 374         /* Have the SMP detector trigger a bit later */
 375         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 376
 377         /* 2/5 is the factor that the perf based detector uses */
 378         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 379 }
 380
 381 void watchdog_nmi_stop(void)
 382 {
 383         int cpu;
 384
 385         for_each_cpu(cpu, &wd_cpus_enabled)
 386                 stop_wd_on_cpu(cpu);
 387 }
 388
 389 void watchdog_nmi_start(void)
 390 {
 391         int cpu;
 392
 393         watchdog_calc_timeouts();
 394         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 395                 start_wd_on_cpu(cpu);
 396 }
 397
 398 /*
 399  * Invoked from core watchdog init.
 400  */
 401 int __init watchdog_nmi_probe(void)
 402 {
 403         int err;
 404
 405         err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 406                                         "powerpc/watchdog:online",
 407                                         start_wd_on_cpu, stop_wd_on_cpu);
 408         if (err < 0) {
 409                 pr_warn("could not be initialized");
 410                 return err;
 411         }
 412         return 0;
 413 }