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 #include <linux/kernel.h>
  10 #include <linux/param.h>
  11 #include <linux/init.h>
  12 #include <linux/percpu.h>
  13 #include <linux/cpu.h>
  14 #include <linux/nmi.h>
  15 #include <linux/module.h>
  16 #include <linux/export.h>
  17 #include <linux/kprobes.h>
  18 #include <linux/hardirq.h>
  19 #include <linux/reboot.h>
  20 #include <linux/slab.h>
  21 #include <linux/kdebug.h>
  22 #include <linux/sched/debug.h>
  23 #include <linux/delay.h>
  24 #include <linux/smp.h>
  25
  26 #include <asm/paca.h>
  27
  28 /*
  29  * The watchdog has a simple timer that runs on each CPU, once per timer
  30  * period. This is the heartbeat.
  31  *
  32  * Then there are checks to see if the heartbeat has not triggered on a CPU
  33  * for the panic timeout period. Currently the watchdog only supports an
  34  * SMP check, so the heartbeat only turns on when we have 2 or more CPUs.
  35  *
  36  * This is not an NMI watchdog, but Linux uses that name for a generic
  37  * watchdog in some cases, so NMI gets used in some places.
  38  */
  39
  40 static cpumask_t wd_cpus_enabled __read_mostly;
  41
  42 static u64 wd_panic_timeout_tb __read_mostly; /* timebase ticks until panic */
  43 static u64 wd_smp_panic_timeout_tb __read_mostly; /* panic other CPUs */
  44
  45 static u64 wd_timer_period_ms __read_mostly;  /* interval between heartbeat */
  46
  47 static DEFINE_PER_CPU(struct timer_list, wd_timer);
  48 static DEFINE_PER_CPU(u64, wd_timer_tb);
  49
  50 /*
  51  * These are for the SMP checker. CPUs clear their pending bit in their
  52  * heartbeat. If the bitmask becomes empty, the time is noted and the
  53  * bitmask is refilled.
  54  *
  55  * All CPUs clear their bit in the pending mask every timer period.
  56  * Once all have cleared, the time is noted and the bits are reset.
  57  * If the time since all clear was greater than the panic timeout,
  58  * we can panic with the list of stuck CPUs.
  59  *
  60  * This will work best with NMI IPIs for crash code so the stuck CPUs
  61  * can be pulled out to get their backtraces.
  62  */
  63 static unsigned long __wd_smp_lock;
  64 static cpumask_t wd_smp_cpus_pending;
  65 static cpumask_t wd_smp_cpus_stuck;
  66 static u64 wd_smp_last_reset_tb;
  67
  68 static inline void wd_smp_lock(unsigned long *flags)
  69 {
  70         /*
  71          * Avoid locking layers if possible.
  72          * This may be called from low level interrupt handlers at some
  73          * point in future.
  74          */
  75         raw_local_irq_save(*flags);
  76         hard_irq_disable(); /* Make it soft-NMI safe */
  77         while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
  78                 raw_local_irq_restore(*flags);
  79                 spin_until_cond(!test_bit(0, &__wd_smp_lock));
  80                 raw_local_irq_save(*flags);
  81                 hard_irq_disable();
  82         }
  83 }
  84
  85 static inline void wd_smp_unlock(unsigned long *flags)
  86 {
  87         clear_bit_unlock(0, &__wd_smp_lock);
  88         raw_local_irq_restore(*flags);
  89 }
  90
  91 static void wd_lockup_ipi(struct pt_regs *regs)
  92 {
  93         pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", raw_smp_processor_id());
  94         print_modules();
  95         print_irqtrace_events(current);
  96         if (regs)
  97                 show_regs(regs);
  98         else
  99                 dump_stack();
 100
 101         if (hardlockup_panic)
 102                 nmi_panic(regs, "Hard LOCKUP");
 103 }
 104
 105 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
 106 {
 107         cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
 108         cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
 109         /*
 110          * See wd_smp_clear_cpu_pending()
 111          */
 112         smp_mb();
 113         if (cpumask_empty(&wd_smp_cpus_pending)) {
 114                 wd_smp_last_reset_tb = tb;
 115                 cpumask_andnot(&wd_smp_cpus_pending,
 116                                 &wd_cpus_enabled,
 117                                 &wd_smp_cpus_stuck);
 118         }
 119 }
 120 static void set_cpu_stuck(int cpu, u64 tb)
 121 {
 122         set_cpumask_stuck(cpumask_of(cpu), tb);
 123 }
 124
 125 static void watchdog_smp_panic(int cpu, u64 tb)
 126 {
 127         unsigned long flags;
 128         int c;
 129
 130         wd_smp_lock(&flags);
 131         /* Double check some things under lock */
 132         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 133                 goto out;
 134         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 135                 goto out;
 136         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 137                 goto out;
 138
 139         pr_emerg("Watchdog CPU:%d detected Hard LOCKUP other CPUS:%*pbl\n",
 140                         cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 141
 142         /*
 143          * Try to trigger the stuck CPUs.
 144          */
 145         for_each_cpu(c, &wd_smp_cpus_pending) {
 146                 if (c == cpu)
 147                         continue;
 148                 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 149         }
 150         smp_flush_nmi_ipi(1000000);
 151
 152         /* Take the stuck CPUs out of the watch group */
 153         set_cpumask_stuck(&wd_smp_cpus_pending, tb);
 154
 155         wd_smp_unlock(&flags);
 156
 157         printk_safe_flush();
 158         /*
 159          * printk_safe_flush() seems to require another print
 160          * before anything actually goes out to console.
 161          */
 162         if (sysctl_hardlockup_all_cpu_backtrace)
 163                 trigger_allbutself_cpu_backtrace();
 164
 165         if (hardlockup_panic)
 166                 nmi_panic(NULL, "Hard LOCKUP");
 167
 168         return;
 169
 170 out:
 171         wd_smp_unlock(&flags);
 172 }
 173
 174 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 175 {
 176         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 177                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 178                         unsigned long flags;
 179
 180                         pr_emerg("Watchdog CPU:%d became unstuck\n", cpu);
 181                         wd_smp_lock(&flags);
 182                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 183                         wd_smp_unlock(&flags);
 184                 } else {
 185                         /*
 186                          * The last CPU to clear pending should have reset the
 187                          * watchdog so we generally should not find it empty
 188                          * here if our CPU was clear. However it could happen
 189                          * due to a rare race with another CPU taking the
 190                          * last CPU out of the mask concurrently.
 191                          *
 192                          * We can't add a warning for it. But just in case
 193                          * there is a problem with the watchdog that is causing
 194                          * the mask to not be reset, try to kick it along here.
 195                          */
 196                         if (unlikely(cpumask_empty(&wd_smp_cpus_pending)))
 197                                 goto none_pending;
 198                 }
 199                 return;
 200         }
 201
 202         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 203
 204         /*
 205          * Order the store to clear pending with the load(s) to check all
 206          * words in the pending mask to check they are all empty. This orders
 207          * with the same barrier on another CPU. This prevents two CPUs
 208          * clearing the last 2 pending bits, but neither seeing the other's
 209          * store when checking if the mask is empty, and missing an empty
 210          * mask, which ends with a false positive.
 211          */
 212         smp_mb();
 213         if (cpumask_empty(&wd_smp_cpus_pending)) {
 214                 unsigned long flags;
 215
 216 none_pending:
 217                 /*
 218                  * Double check under lock because more than one CPU could see
 219                  * a clear mask with the lockless check after clearing their
 220                  * pending bits.
 221                  */
 222                 wd_smp_lock(&flags);
 223                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 224                         wd_smp_last_reset_tb = tb;
 225                         cpumask_andnot(&wd_smp_cpus_pending,
 226                                         &wd_cpus_enabled,
 227                                         &wd_smp_cpus_stuck);
 228                 }
 229                 wd_smp_unlock(&flags);
 230         }
 231 }
 232
 233 static void watchdog_timer_interrupt(int cpu)
 234 {
 235         u64 tb = get_tb();
 236
 237         per_cpu(wd_timer_tb, cpu) = tb;
 238
 239         wd_smp_clear_cpu_pending(cpu, tb);
 240
 241         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 242                 watchdog_smp_panic(cpu, tb);
 243 }
 244
 245 void soft_nmi_interrupt(struct pt_regs *regs)
 246 {
 247         unsigned long flags;
 248         int cpu = raw_smp_processor_id();
 249         u64 tb;
 250
 251         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 252                 return;
 253
 254         nmi_enter();
 255
 256         __this_cpu_inc(irq_stat.soft_nmi_irqs);
 257
 258         tb = get_tb();
 259         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 260                 per_cpu(wd_timer_tb, cpu) = tb;
 261
 262                 wd_smp_lock(&flags);
 263                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 264                         wd_smp_unlock(&flags);
 265                         goto out;
 266                 }
 267                 set_cpu_stuck(cpu, tb);
 268
 269                 pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", cpu);
 270                 print_modules();
 271                 print_irqtrace_events(current);
 272                 if (regs)
 273                         show_regs(regs);
 274                 else
 275                         dump_stack();
 276
 277                 wd_smp_unlock(&flags);
 278
 279                 if (sysctl_hardlockup_all_cpu_backtrace)
 280                         trigger_allbutself_cpu_backtrace();
 281
 282                 if (hardlockup_panic)
 283                         nmi_panic(regs, "Hard LOCKUP");
 284         }
 285         if (wd_panic_timeout_tb < 0x7fffffff)
 286                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 287
 288 out:
 289         nmi_exit();
 290 }
 291
 292 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 293 {
 294         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 295         if (wd_timer_period_ms > 1000)
 296                 t->expires = __round_jiffies_up(t->expires, cpu);
 297         add_timer_on(t, cpu);
 298 }
 299
 300 static void wd_timer_fn(unsigned long data)
 301 {
 302         struct timer_list *t = this_cpu_ptr(&wd_timer);
 303         int cpu = smp_processor_id();
 304
 305         watchdog_timer_interrupt(cpu);
 306
 307         wd_timer_reset(cpu, t);
 308 }
 309
 310 void arch_touch_nmi_watchdog(void)
 311 {
 312         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 313         int cpu = smp_processor_id();
 314         u64 tb;
 315
 316         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 317                 return;
 318
 319         tb = get_tb();
 320         if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
 321                 per_cpu(wd_timer_tb, cpu) = tb;
 322                 wd_smp_clear_cpu_pending(cpu, tb);
 323         }
 324 }
 325 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 326
 327 static void start_watchdog_timer_on(unsigned int cpu)
 328 {
 329         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 330
 331         per_cpu(wd_timer_tb, cpu) = get_tb();
 332
 333         setup_pinned_timer(t, wd_timer_fn, 0);
 334         wd_timer_reset(cpu, t);
 335 }
 336
 337 static void stop_watchdog_timer_on(unsigned int cpu)
 338 {
 339         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 340
 341         del_timer_sync(t);
 342 }
 343
 344 static int start_wd_on_cpu(unsigned int cpu)
 345 {
 346         unsigned long flags;
 347
 348         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 349                 WARN_ON(1);
 350                 return 0;
 351         }
 352
 353         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 354                 return 0;
 355
 356         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 357                 return 0;
 358
 359         wd_smp_lock(&flags);
 360         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 361         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 362                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 363                 wd_smp_last_reset_tb = get_tb();
 364         }
 365         wd_smp_unlock(&flags);
 366
 367         start_watchdog_timer_on(cpu);
 368
 369         return 0;
 370 }
 371
 372 static int stop_wd_on_cpu(unsigned int cpu)
 373 {
 374         unsigned long flags;
 375
 376         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 377                 return 0; /* Can happen in CPU unplug case */
 378
 379         stop_watchdog_timer_on(cpu);
 380
 381         wd_smp_lock(&flags);
 382         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 383         wd_smp_unlock(&flags);
 384
 385         wd_smp_clear_cpu_pending(cpu, get_tb());
 386
 387         return 0;
 388 }
 389
 390 static void watchdog_calc_timeouts(void)
 391 {
 392         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 393
 394         /* Have the SMP detector trigger a bit later */
 395         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 396
 397         /* 2/5 is the factor that the perf based detector uses */
 398         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 399 }
 400
 401 void watchdog_nmi_stop(void)
 402 {
 403         int cpu;
 404
 405         for_each_cpu(cpu, &wd_cpus_enabled)
 406                 stop_wd_on_cpu(cpu);
 407 }
 408
 409 void watchdog_nmi_start(void)
 410 {
 411         int cpu;
 412
 413         watchdog_calc_timeouts();
 414         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 415                 start_wd_on_cpu(cpu);
 416 }
 417
 418 /*
 419  * Invoked from core watchdog init.
 420  */
 421 int __init watchdog_nmi_probe(void)
 422 {
 423         int err;
 424
 425         err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 426                                         "powerpc/watchdog:online",
 427                                         start_wd_on_cpu, stop_wd_on_cpu);
 428         if (err < 0) {
 429                 pr_warn("Watchdog could not be initialized");
 430                 return err;
 431         }
 432         return 0;
 433 }
 434
 435 static void handle_backtrace_ipi(struct pt_regs *regs)
 436 {
 437         nmi_cpu_backtrace(regs);
 438 }
 439
 440 static void raise_backtrace_ipi(cpumask_t *mask)
 441 {
 442         unsigned int cpu;
 443
 444         for_each_cpu(cpu, mask) {
 445                 if (cpu == smp_processor_id())
 446                         handle_backtrace_ipi(NULL);
 447                 else
 448                         smp_send_nmi_ipi(cpu, handle_backtrace_ipi, 1000000);
 449         }
 450 }
 451
 452 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
 453 {
 454         nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_backtrace_ipi);
 455 }