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         if (cpumask_empty(&wd_smp_cpus_pending)) {
 110                 wd_smp_last_reset_tb = tb;
 111                 cpumask_andnot(&wd_smp_cpus_pending,
 112                                 &wd_cpus_enabled,
 113                                 &wd_smp_cpus_stuck);
 114         }
 115 }
 116 static void set_cpu_stuck(int cpu, u64 tb)
 117 {
 118         set_cpumask_stuck(cpumask_of(cpu), tb);
 119 }
 120
 121 static void watchdog_smp_panic(int cpu, u64 tb)
 122 {
 123         unsigned long flags;
 124         int c;
 125
 126         wd_smp_lock(&flags);
 127         /* Double check some things under lock */
 128         if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
 129                 goto out;
 130         if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
 131                 goto out;
 132         if (cpumask_weight(&wd_smp_cpus_pending) == 0)
 133                 goto out;
 134
 135         pr_emerg("Watchdog CPU:%d detected Hard LOCKUP other CPUS:%*pbl\n",
 136                         cpu, cpumask_pr_args(&wd_smp_cpus_pending));
 137
 138         /*
 139          * Try to trigger the stuck CPUs.
 140          */
 141         for_each_cpu(c, &wd_smp_cpus_pending) {
 142                 if (c == cpu)
 143                         continue;
 144                 smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
 145         }
 146         smp_flush_nmi_ipi(1000000);
 147
 148         /* Take the stuck CPUs out of the watch group */
 149         set_cpumask_stuck(&wd_smp_cpus_pending, tb);
 150
 151         wd_smp_unlock(&flags);
 152
 153         printk_safe_flush();
 154         /*
 155          * printk_safe_flush() seems to require another print
 156          * before anything actually goes out to console.
 157          */
 158         if (sysctl_hardlockup_all_cpu_backtrace)
 159                 trigger_allbutself_cpu_backtrace();
 160
 161         if (hardlockup_panic)
 162                 nmi_panic(NULL, "Hard LOCKUP");
 163
 164         return;
 165
 166 out:
 167         wd_smp_unlock(&flags);
 168 }
 169
 170 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
 171 {
 172         if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
 173                 if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
 174                         unsigned long flags;
 175
 176                         pr_emerg("Watchdog CPU:%d became unstuck\n", cpu);
 177                         wd_smp_lock(&flags);
 178                         cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
 179                         wd_smp_unlock(&flags);
 180                 }
 181                 return;
 182         }
 183         cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
 184         if (cpumask_empty(&wd_smp_cpus_pending)) {
 185                 unsigned long flags;
 186
 187                 wd_smp_lock(&flags);
 188                 if (cpumask_empty(&wd_smp_cpus_pending)) {
 189                         wd_smp_last_reset_tb = tb;
 190                         cpumask_andnot(&wd_smp_cpus_pending,
 191                                         &wd_cpus_enabled,
 192                                         &wd_smp_cpus_stuck);
 193                 }
 194                 wd_smp_unlock(&flags);
 195         }
 196 }
 197
 198 static void watchdog_timer_interrupt(int cpu)
 199 {
 200         u64 tb = get_tb();
 201
 202         per_cpu(wd_timer_tb, cpu) = tb;
 203
 204         wd_smp_clear_cpu_pending(cpu, tb);
 205
 206         if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
 207                 watchdog_smp_panic(cpu, tb);
 208 }
 209
 210 void soft_nmi_interrupt(struct pt_regs *regs)
 211 {
 212         unsigned long flags;
 213         int cpu = raw_smp_processor_id();
 214         u64 tb;
 215
 216         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 217                 return;
 218
 219         nmi_enter();
 220
 221         __this_cpu_inc(irq_stat.soft_nmi_irqs);
 222
 223         tb = get_tb();
 224         if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
 225                 per_cpu(wd_timer_tb, cpu) = tb;
 226
 227                 wd_smp_lock(&flags);
 228                 if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
 229                         wd_smp_unlock(&flags);
 230                         goto out;
 231                 }
 232                 set_cpu_stuck(cpu, tb);
 233
 234                 pr_emerg("Watchdog CPU:%d Hard LOCKUP\n", cpu);
 235                 print_modules();
 236                 print_irqtrace_events(current);
 237                 if (regs)
 238                         show_regs(regs);
 239                 else
 240                         dump_stack();
 241
 242                 wd_smp_unlock(&flags);
 243
 244                 if (sysctl_hardlockup_all_cpu_backtrace)
 245                         trigger_allbutself_cpu_backtrace();
 246
 247                 if (hardlockup_panic)
 248                         nmi_panic(regs, "Hard LOCKUP");
 249         }
 250         if (wd_panic_timeout_tb < 0x7fffffff)
 251                 mtspr(SPRN_DEC, wd_panic_timeout_tb);
 252
 253 out:
 254         nmi_exit();
 255 }
 256
 257 static void wd_timer_reset(unsigned int cpu, struct timer_list *t)
 258 {
 259         t->expires = jiffies + msecs_to_jiffies(wd_timer_period_ms);
 260         if (wd_timer_period_ms > 1000)
 261                 t->expires = __round_jiffies_up(t->expires, cpu);
 262         add_timer_on(t, cpu);
 263 }
 264
 265 static void wd_timer_fn(unsigned long data)
 266 {
 267         struct timer_list *t = this_cpu_ptr(&wd_timer);
 268         int cpu = smp_processor_id();
 269
 270         watchdog_timer_interrupt(cpu);
 271
 272         wd_timer_reset(cpu, t);
 273 }
 274
 275 void arch_touch_nmi_watchdog(void)
 276 {
 277         unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
 278         int cpu = smp_processor_id();
 279         u64 tb = get_tb();
 280
 281         if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
 282                 per_cpu(wd_timer_tb, cpu) = tb;
 283                 wd_smp_clear_cpu_pending(cpu, tb);
 284         }
 285 }
 286 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
 287
 288 static void start_watchdog_timer_on(unsigned int cpu)
 289 {
 290         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 291
 292         per_cpu(wd_timer_tb, cpu) = get_tb();
 293
 294         setup_pinned_timer(t, wd_timer_fn, 0);
 295         wd_timer_reset(cpu, t);
 296 }
 297
 298 static void stop_watchdog_timer_on(unsigned int cpu)
 299 {
 300         struct timer_list *t = per_cpu_ptr(&wd_timer, cpu);
 301
 302         del_timer_sync(t);
 303 }
 304
 305 static int start_wd_on_cpu(unsigned int cpu)
 306 {
 307         unsigned long flags;
 308
 309         if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
 310                 WARN_ON(1);
 311                 return 0;
 312         }
 313
 314         if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
 315                 return 0;
 316
 317         if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 318                 return 0;
 319
 320         wd_smp_lock(&flags);
 321         cpumask_set_cpu(cpu, &wd_cpus_enabled);
 322         if (cpumask_weight(&wd_cpus_enabled) == 1) {
 323                 cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
 324                 wd_smp_last_reset_tb = get_tb();
 325         }
 326         wd_smp_unlock(&flags);
 327
 328         start_watchdog_timer_on(cpu);
 329
 330         return 0;
 331 }
 332
 333 static int stop_wd_on_cpu(unsigned int cpu)
 334 {
 335         unsigned long flags;
 336
 337         if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
 338                 return 0; /* Can happen in CPU unplug case */
 339
 340         stop_watchdog_timer_on(cpu);
 341
 342         wd_smp_lock(&flags);
 343         cpumask_clear_cpu(cpu, &wd_cpus_enabled);
 344         wd_smp_unlock(&flags);
 345
 346         wd_smp_clear_cpu_pending(cpu, get_tb());
 347
 348         return 0;
 349 }
 350
 351 static void watchdog_calc_timeouts(void)
 352 {
 353         wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
 354
 355         /* Have the SMP detector trigger a bit later */
 356         wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
 357
 358         /* 2/5 is the factor that the perf based detector uses */
 359         wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
 360 }
 361
 362 void watchdog_nmi_stop(void)
 363 {
 364         int cpu;
 365
 366         for_each_cpu(cpu, &wd_cpus_enabled)
 367                 stop_wd_on_cpu(cpu);
 368 }
 369
 370 void watchdog_nmi_start(void)
 371 {
 372         int cpu;
 373
 374         watchdog_calc_timeouts();
 375         for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
 376                 start_wd_on_cpu(cpu);
 377 }
 378
 379 /*
 380  * Invoked from core watchdog init.
 381  */
 382 int __init watchdog_nmi_probe(void)
 383 {
 384         int err;
 385
 386         err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 387                                         "powerpc/watchdog:online",
 388                                         start_wd_on_cpu, stop_wd_on_cpu);
 389         if (err < 0) {
 390                 pr_warn("Watchdog could not be initialized");
 391                 return err;
 392         }
 393         return 0;
 394 }
 395
 396 static void handle_backtrace_ipi(struct pt_regs *regs)
 397 {
 398         nmi_cpu_backtrace(regs);
 399 }
 400
 401 static void raise_backtrace_ipi(cpumask_t *mask)
 402 {
 403         unsigned int cpu;
 404
 405         for_each_cpu(cpu, mask) {
 406                 if (cpu == smp_processor_id())
 407                         handle_backtrace_ipi(NULL);
 408                 else
 409                         smp_send_nmi_ipi(cpu, handle_backtrace_ipi, 1000000);
 410         }
 411 }
 412
 413 void arch_trigger_cpumask_backtrace(const cpumask_t *mask, bool exclude_self)
 414 {
 415         nmi_trigger_cpumask_backtrace(mask, exclude_self, raise_backtrace_ipi);
 416 }